// transcript — 1990 segments
0:00 Introducing Alex Cheema to the show
0:01 All right, everybody. Welcome back to Twist. It's Friday, February 6, 2026,
0:06 and today we're going to share how we built Open Claw Ultron. This is a new
0:14 project inside of our firm, Launch, and this week in startups where we produce
0:17 podcasts and we invest in 100 companies a year. What are we trying to do? We're
0:20 trying to build one instance of OpenClaw, formerly known as Maltbot,
0:25 formerly known as Clawbot. We're trying to build one replicant, one agent that
0:31 can do all 20 people's jobs here at the venture firm and at the production
0:36 company that does all these podcasts. 20 people's jobs. Each of those jobs
0:40 probably has a half dozen important skills. So, we're talking about at some
0:45 point putting together in one agent, we call them replicants, we're going to
0:50 have somewhere in the order of a 100 to 200 skills. That one person is going to
0:54 try to do everybody's work. That's the goal. And then everybody will level up
0:57 and do some other work. So the goal isn't to replace everybody. It's to take
1:01 away everybody's chores and to make everybody better at the primary
1:04 functions in an investment firm, which is meeting with founders, spending time
1:09 with founders and LPs, our investors. And that on the production side, it
1:12 would be producing great content and working with our guests. We want to move
1:15 up the stack and give away all the chores. With me to discuss it, Lon
1:19 Harris, who's going to co-host the show today. How you doing, Lon?
1:21 >> Doing great. Great to be here. >> All right. And Oliver Cororsan is here.
1:26 He has been doing demos for me and producing this week in AI which is going
1:30 to launch in February in two weeks I think. Oliver, welcome to the program.
1:32 >> Thank you. Good to be here. >> And we have a special guest. Alex Chima
1:37 is here. Alex I have been following for some time because maybe a year ago I saw
1:45 Alex was working on stacking with this company. It's EXO, right? EXO. Exo is
1:49 how you pronounce it. and you've been working on taking commodity hardware
1:54 like a Mac Mini daisy chaining them or connecting them together in order to run
1:58 large language models locally but as we've seen OpenClaw formerly Cloudbot
2:03 and Multbot has quite a wrinkle in this you were like a year or two ahead of
2:07 this trend of hey can we run locally so let's start just really quick Alex
2:11 before we go into Ultron open claw Ultron what your firm does and and what
2:15 progress you've made especially in regards to openclaw >> yeah thanks so much for having me Jason
2:21 Um, so I'm the founder and CEO of Exo Labs and like you said, we've been doing
2:26 stuff with Mac Minis uh long before OpenCore was around. And to be honest, I
2:32 didn't expect the rise of like people buying Mac minis to come from this
2:35 place. I thought it would the catalyst would be people wanting to run models
2:39 locally. Uh, what we do is we make it possible to run Frontier AI locally on
2:45 consumer hardware. Um, so not just Max, but also other kinds of consumer
2:49 hardware. We're trying to drive down the barrier to running the most capable AI
2:54 models. So we currently have like the cheapest way, cheapest, most accessible
2:59 way to run Kim K 2.5 uh on two Mac Studios and we're working across the
3:03 whole stack. So we're working on the model layer, the distributed uh
3:08 algorithms as well that are very different when you're working with
3:12 consumer hardware and also like lower level like kernels. And our goal is
3:16 basically to make Frontier AI accessible to anyone to run on their own hardware.
3:20 >> Why is this important? Why is it important to run it on local hardware?
3:23 Yeah, I think this is something that with the whole open claw phrase, not a
3:27 lot of people are talking about, but just how the way we're using AI is
3:34 shifting and it's going from being this kind of crude tool that you use through
3:40 like a chat interface to becoming sort of an extension of yourself. And the AI
3:46 now it knows everything you know you you you know it it can basically do
3:49 everything you can do digitally right now and soon you know with robotics
3:53 that's going to be physically as well and at that point it's more of an
3:57 exocortex so it's not just this like tool that you talk through a chat
4:01 interface but it's this thing that's actually part of your cell and then you
4:06 start to question okay you know do I want to rent my brain and copy talks
4:11 about this he he he says not your weights It's not your brain. Like do you
4:17 really want, you know, another a profit seeking company basically running your
4:20 brain? And when you think of it like that, um, to me, you know, my reason for
4:25 starting Exo is you want control and you want ownership of that. Open core is a
4:31 long way towards that because for a while the products were getting better.
4:34 these closed source like the models are largely like commoditized and there's a
4:40 pretty standard pretty thin API layer to interacting with them. So the switching
4:44 cost is quite low. But what worried me was that the products the closed
4:47 products are getting a lot better like with HBT with memory systems and also
4:52 the more stateful aspects of like the workflows that you're building. So now
4:56 the fact that you have open claw which is open well you can run it on your own
5:00 infrastructure. Now, a large part of that, >> to summarize that, I thought you were
5:05 going to say, well, it it's cheaper because you're not paying for tokens.
5:08 That's what I thought you would say first. Then I thought you would say,
5:13 well, you know, you can put so much data on it, you'll have better memory. But
5:17 you went with a really even higher, bigger picture reason to do this, which
5:22 is if you put this all in Open AI, and OpenAI has a trillion dollar valuation,
5:26 and they need to make money. If I put all my venture capital data in there and
5:30 I train it all of my with all of my secrets, those are all going to acrue
5:35 eventually even if they say it's not going to you have this very reasonable
5:40 fear or concern that it's going to acrue to open AI uh to chat GPT not to your
5:47 firm. So that's the reason really to do this yourself. Yeah. In your mind, Alex?
5:51 Yeah, I think that there's a nuance there of just like I actually don't
5:54 believe in sort of the privacy argument so much of like I think at least for
5:59 consumers, you know, we're already putting our data into platforms and
6:04 we're completely fine with that, but it's more about the sovereignty aspect
6:07 and actually having control of it. So, how easy is it for you to switch? How
6:11 easy is it for you to like if the model's changing under your feet, how
6:15 much control do you actually have? >> So, that's lock in. and lock in for a
6:19 chat GPT I just experienced because we canceled our open AI account and we
6:22 moved everything to Claude because we felt claude was a better product and we
6:25 felt like we trusted that organization a little better. When we moved it over I
6:28 had three people say oh my god I have all my stuff there and I was like really
6:32 and they're like yeah so I turned their accounts back on so they could get it
6:35 but there's not like an easy way to get your memory out of there and bring it
6:39 over there. Well, we saw the same thing with GPT4 moving into five that a lot of
6:45 people like they they lost the magic that they'd loved about GPT40. So, it's
6:49 like, you know, the models can just sort of change or upgrade on a whim and then
6:54 you lose this, you know, like character persona you felt like was part of your
6:57 life in a way. >> So, now Oliver, it's your chance to shine. Oliver has jumped in in the last
7:03 10 days and gone all in on Open Claw. One of the things we did was we built a
7:08 persona, the first one, to work on the production of the podcast, doing guest
7:13 research, guest outreach, and to figure out what should be on the docket. In
7:15 other words, what topic should we discuss and on the margins, hey, what
7:19 should the title of this video be? What should the thumbnail be? And just trying
7:23 to see if it could do those functions. Oliver, you've been working on this.
7:28 Show us the state-of-the-art now because I think the first time we did this was
7:31 last Monday, not this past Monday, but the Monday, two Mondays ago. Yeah,
7:35 >> this is the end of week two of our round theclock uh clawbot coverage.
7:39 >> Crazy. Okay, Oliver, show it. Show let's show what you built.
7:42 >> It's been around 10 days since we first started building our instance of
7:47 OpenClaw. And as you mentioned, we have two different ones. One that's more
7:49 focused on the investment team and I am building an OpenClaw bot that is kind of
7:53 more focused on the production side of things. So, one thing that I think was a
7:57 little bit of a misstep that I would tell anyone who's building a new
8:01 OpenClaw is to start with a dashboard. That should be kind of your step one
8:06 once you get your openclaw online and a dashboard as you would think about it.
8:09 But it is able to connect to the back end of your openclaw instance and bring
8:13 in the data so you can see it visually bring in all the files. It's just being
8:16 able to look at it visually is much better than trying to interact with its
8:20 backend and obviously its front end all just from a chat interface. So doing
8:26 this was very easy. So I was watching an Alex Finn video who we had on last
8:32 Monday and Alex Finn was interacting only in his dashboard with his open
8:37 claw. I basically was like why are we not doing that? Because open claw
8:40 doesn't really have a dashboard. You basically are telling it hey remember
8:45 this you know make a file here but you don't understand the underpinnings.
8:48 There isn't a dashboard. So it would literally be this is early on. Open claw
8:54 is essentially a black box. You have all this memory and you have skills that you
9:00 have to query it to understand. But you made a dashboard. The dashboard is going
9:05 to show what files it has in memory. And an example of a memory file would be
9:09 what in our case. >> Yeah. So the example of a memory file
9:13 would be Oliver's preferences. What are my preferences? So this is in the
9:17 memory. Never use m dashes and emails. I don't want that to happen. I want you to
9:22 be a person. uh don't put direct competitors on the same show when we're
9:27 booking a podcast episode. Um and also at the moment we're not booking VCs on
9:30 this week in AI. So these are all things that I've told it these are my
9:33 preferences when I'm doing tasks throughout the day. >> So you don't want to repeat yourself and
9:38 say don't put two competitors on the same episode. You don't want to repeat
9:43 yourself uh wi with these specific instructions on booking guests. Got it.
9:47 >> Yes. Exactly. And it just kind of keeps you know things I've told it and it's in
9:51 mind. So if I ask it to do something, it'll remember what we talked about.
9:56 Example of a shortcut that I gave it was I I basically wanted it to understand
9:58 who were the pending calendar invitations that we had while we were
10:02 booking them. So there's, you know, a handful of guests that
10:04 >> if you have guests that we've invited and they haven't responded to the invite
10:09 yet. You want to know that you call that pending >> pending calendar invites. Yes. And in
10:13 order for the bot to be as helpful as possible, it needs to understand who
10:16 those guests are, which are the ones that it needs to look for the email to
10:21 see if they have responded yet or have I responded to them. So these are the type
10:25 of things that you would keep in the me in your memory. So memory is the first
10:28 thing on the dashboard. I think we understand that preferences or different
10:33 pieces of data. Now some of the memory could that exist on a notion page or in
10:36 a Google document and would that be represented here or is it only memory
10:39 and files that are stored inside of OpenClaw? These specifically are only
10:43 stored inside of OpenClaw. Of course, it can reference different databases that
10:47 you have. But the kind of the big point of this show is to show how we have
10:52 created our open call Ultron to replace 20 employees at our company. So
10:56 obviously that's the end goal. I still want to have a job. I'm sure the lawn
10:59 wants to have a job. >> There'll be more for you to do. We want
11:02 to launch. We have >> Here's the thing. There's two, if you
11:05 think about your job, you've been doing a bit of production here. of the
11:09 production hours, hours you spend on production at this point in week two,
11:14 how many of those do you think you'll wind up handing off in 30 days? Let's
11:17 say if you just keep grinding on this for another four weeks, in 30 days, what
11:22 percentage of the work you're doing in total hours? So, if you work 50 hours a
11:26 week, how many of those hours would be done, you know, conservatively or
11:29 optimistically, you give one number or two, just conservatively,
11:32 optimistically, by this new Ultron? I would say around 60% of my time if I'm
11:37 doing 30 hours a week on production. Something you mentioned earlier is that,
11:39 you know, there's probably hundreds of tasks that people do at our company. So,
11:43 in order to build out all of those skills that can do those tasks, we're
11:46 going to have to do that one at a time and it's we're going to need to make
11:50 sure each one works. So, I have around nine or eight tasks that I have
11:56 successfully or I'm in the process of building out. >> Okay? And those are called cron jobs.
11:58 What is a Chron Job? (Hint: chron means chronological)
12:01 These are jobs that occur on a chronological on a on a time basis.
12:06 That's what cron job means. And cron jobs are something Alex that developers
12:11 do all the time. But knowledge workers don't typically have cron jobs, right,
12:14 Alex? >> Well, I don't know. I think this is one of the more interesting features and one
12:20 of the things that like to me open floor is like putting together a lot of things
12:26 that already existed in a very intuitive uh seamless way and one of them is
12:29 scrunch jobs and I'm using them I'm using them for like loads of things um
12:36 not just um dev stuff but like a lot of um management so we're like I have
12:42 something that's like constantly scanning um our Slack and uh basically making suggestions once
12:53 um it's uh I have kind of like this uh way of quantifying like uncertainty
12:58 about tasks. So I think this is something that the LLM are like getting
13:04 better at is like knowing when to um be proactive. And so, you know, like
13:08 basically I'm giving it as much context as I can from the Slack so that it can
13:14 suggest every day um a list of things that we might be missing or something
13:17 things that we should be aware of. So, this is running just on a on a chron job
13:25 Crusoe Cloud: Crusoe is the AI factory company. Reliable infrastructure and expert support. Visit to reserve your capacity for the latest GPUs today.
13:27 AI is revolutionizing every aspect of our industry, but for founders, it can
13:31 prove very frustrating largely because cloud costs are so unpredictable. This
13:35 has probably happened to you. You've planned out and budgeted a certain
13:39 amount for a project and then been hit with a massive bill with no real
13:43 explanation for what went wrong. If you're feeling like the cost of working
13:46 with this innovative technology is just too high, finally we have a solution for
13:50 you. Crusoe. They are the AI cloud company that's taking an energy first
13:54 approach. What does it mean? They efficiently access their power directly
13:58 from the source and pass all the savings along to you. This isn't just about
14:03 saving some cash. Cruso's architecture is specifically optimized for your AI
14:08 workloads and their customer support response times clock in at under six
14:12 minutes. Here's what I want you to do. Visit cruso.ai/savings
14:18 to receive $100,000 in credits for virtualized NVIDIA GB200 NVL72 on Crusoe
14:24 Cloud pending availability. That's Yeah. So, and when you see uh Oliver
14:35 with the memory and the files, what comes to mind with Exo and, you know,
14:41 standing up to, you know, Mac Studios, the M5s coming out and how much memory
14:45 you could put in there? I was telling the team, I want to take the Notion API
14:52 and I want to take the Slack API and I want to put into memory every single
14:58 Slack message this year, maybe even over all eternity. and you know somehow have
15:04 that all in here. So maybe you could you could speak to that memory because you
15:06 already spoke to it in terms of like giving it to open AI or another company
15:11 versus keeping it for yourself. But how do you think about large amounts of
15:15 data? Yeah, this is definitely a big focus right now in terms of inference
15:19 infrastructure is just how do you support really big context um with you
15:25 know basically being able to put everything in context and the way I look
15:31 at this is you can look at well inference consists of two stages there's
15:36 like the prefill stage which is very compute heavy it's comput bound and you
15:40 have the decode stage and what you're seeing is that most use cases at the
15:45 moment are very decode heavy. So, it's actually most of the time is being spent
15:50 on just uh generating tokens. And I think the software is actually really
15:55 good now at kind of making sure that when it comes to the prefill, you've got
16:02 uh you're getting a lot of uh cash hits. Um so, you know, I think basically we'll
16:06 be able to continue just increasing context, context, context quite a bit.
16:13 And you know, basically the hardware is more of a focus is going to be on the
16:16 decode side. That's where consumer hardware is really good. Uh you have the
16:20 M5 coming out pretty soon. Um it's a big boost in memory bandwidth for memory and
16:24 all of that side of things is super memory bound. So I don't see any like
16:29 reason why you couldn't just shove all your Slack messages into context. I
16:33 think that's going to happen. and and we should just buy when the M5 comes out
16:39 max memory which is what 500 gigs of of memory. >> Yeah, it's 512 at the moment and maybe
16:43 that will increase as well and it's enough to fit you know really large
16:48 models enough to fit all that context as well. This is always I feel like the
16:51 sort of the dream like when producer Claude we first brought that on board
16:54 from Anthropic to the show that was really what we wanted like he want he
16:59 should listen to everything we say and remember it and then throw in helpful
17:03 suggestions and the technology was not quite there yet but I feel like now
17:07 we're on the precipice of actually being able to do that with an AI.
17:10 >> Okay. So let's go through the cron jobs here real quick. Maybe you could give us
17:15 an example of a a cron job. And I'm guessing each one of these skills is,
17:19 you know, if you if it's been two weeks and you've got eight working, you're
17:24 you're basically on one a day or so or one every, you know, 1.5 days. So that
17:30 seems like a pretty good pace to me if we have 200 skills. We're going to give
17:35 this eventually, you know, that that's a that's a pretty good um Yeah, that's is
17:39 a pretty good um pace. So >> there is a trial and error. Like I I
17:43 sort of have written one skill so far for the ticker digest and you do have to
17:47 tell it what to do, see what kind of feedback you get and then you know there
17:51 is a tinkering to get the prompting and get everything exactly the way you want
17:53 OpenClaw managing the LAUNCH/TWiST team
17:54 it. For sure. >> Okay. So let's uh look at hm how about
17:58 attendance? I think this is an interesting one. For people who don't
18:01 know, I wrote a famous blog post years ago called, you know, this sort of
18:06 lightweight management and start of day, end of day as a tool for uh executives,
18:11 especially when remote teams were happening. I just asked everybody on our
18:15 team, Alex, kind of like a standup for developers, etc. Just say what you're
18:21 intending to get done today and then at the end of the day, reply to yourself in
18:25 Slack in the general channel and uh say what you got done. I had like two of my
18:30 four senior executives at the time essentially quit over this because they
18:34 didn't want to be micromanaged. Uh and I was like, well, it's just like
18:38 you're getting paid a very large six-figure salary. You you can't spend five and 10 minutes
18:43 just saying what you're going to do for the day. And and that was great for me
18:47 because I I just don't like people who are not good communicators or don't set
18:51 goals for themselves and and they're doing great probably. Um maybe. But what
18:55 did you create here, Oliver? Yeah. So we all post our start of day and end of
19:01 days in one Slack channel called general and two crown jobs. One is the start of
19:06 day attendance where it looks who has sent their start of day you know for
19:11 anywhere from you know 7 a.m. to 12:00 p.m. And right at 12 which is in the
19:15 morning when you should send your start of day what you're going to do that day.
19:18 It will look through the general channel, see who has sent it, and
19:22 whoever doesn't send it, the bot will then send a Slack message in the general
19:27 channel tagging you, Jason, and also tagging the people who haven't sent it
19:29 yet. So, it's kind of just that accountability. That's a crown job that
19:32 runs it. >> And then you do the same thing at the end of the day. And previously, we would
19:37 have a human do this. They would scroll up and they would spend 20 minutes and
19:40 they would then go check in with people because that's in when we were fully
19:44 remote, Alex. That's what how we figured out uh who took a paid day off or who
19:49 was on holiday or you know if something was wrong, you know, check in on a
19:54 Lemon.io - Get 15% off your first 4 weeks of developer time
19:56 Building out your team is one of the most crucial things you have to get
20:00 right in your startup and finding the right developers is particularly
20:05 important. But now there's lemon.io. They're going to save you time, money,
20:09 and headaches by doing all the time consuming leg work for you. They've got
20:13 an experienced lineup of prevetted developers working for competitive
20:17 rates. Just 1% of applicants are accepted into Lemon's elite program. And
20:22 they're not just out there finding this great talent. They're also working with
20:26 you to integrate these new members into your team. Plus, if it's not a good fit,
20:29 hey, and sometimes things don't work out, Lemon will hook you up with a new
20:34 developer ASAP. I've seen startups go from just pretty good to amazing after
20:39 filling out their teams with developers from lemon.io. Go to lemon.io/twist
20:45 and find your perfect developer or technical team in 48 hours or less.
20:50 Plus, Twist listeners get 15% off their first four weeks. That's lemon.ioist.
20:58 Turning AI into Ultron, self optimization
21:02 Okay, give us uh one more. What else is like interesting here?
21:05 >> Let's talk about self-optimization. I want to hear about that one.
21:07 >> Oh, yeah. That's I don't know what that is, but okay. Age of Ultron is here.
21:10 What is self-optimization? >> Yeah. So, this is basically an optimizer
21:16 task where this role would previously be an engineer or I would look through all
21:19 the files. I mean, I wouldn't be able to do this if it wasn't plain language like
21:23 OpenClaw is, but previously you're looking at an organization, you're
21:25 looking at the structure. You would maybe want an engineer or someone with a
21:30 lot of experience to look through how everything's running. So I have set up a
21:34 selfoptimization cron job. >> So this is running Monday through
21:39 Friday. And what is it? And and did you write this prompt or did you ask it to
21:43 write a prompt to do this? >> I asked it to write this prompt. The
21:47 goal is the end goal would be for you know any from 3 to 5 am for it to be
21:52 looking through all of our files all of our cron jobs all of our skills and then
21:58 at 8 a.m. at me what could we change? So, not actually execute yet, at least
22:02 while we're still building trust. It gives me a list of five of the things
22:06 that it thinks that we can really change and optimize. And this was the one from
22:11 this morning. So, it it noticed that there was a time zone bug in the guest
22:17 calendar. So, it was getting CST and CDT confused. Um, and it said that it would
22:21 be able to fix this quite quickly. There was some issues in the
22:25 >> So, it's always good to give the exact one. So that was great when you gave the
22:28 exact one. It had an error there. Give another one. What else is like an exact
22:32 thing that it said we should fix that was material here. >> The self optimization cron job realized
22:38 that there was a cron scheduleuler issue where jobs were skipping days. So it
22:42 realized that some of today's jobs did not run and then it went and
22:48 investigated the scheduling issue and also told me that this would be a medium
22:53 effort change. So then I told it to fix that and then it went into the files and
22:56 made sure that that wouldn't happen again. Did it give us anything like in
23:00 terms of this is like fixing its internal you know guts and everything uh and the
23:06 engine but did it give us anything in terms of destinations of where to take
23:09 the car that could be improved? Did it say like oh you should consider you know
23:13 these type of guests for the program or here's how to make advertising you know
23:17 more effective. Did it give us anything like that on a business basis? Yes. So
23:21 the self-optimization cron job that I set up is specifically looking at how
23:26 open claw is set up. But I do have other cron jobs that are exactly that. So I do
23:33 have a sales and sponsors specific task. So one of the tasks that one a member of
23:38 our sales team does is they look through competitor podcasts and see who the
23:43 sponsors or partners are that are on those shows so we can get ideas, you
23:46 know, to bring on sponsorship. >> Yeah. If we're missing if we're if
23:49 there's some new sponsor in the world and we don't have them yet, you might
23:53 hear them on the New York Times podcast and we should probably reach out to
23:56 them. We had a human doing that previously. Yeah, >> exactly. And in this basically works
24:03 with the YouTube API will go through a list of I believe 20 different podcasts
24:07 that I gave it. Look through the timestamps and I also believe it can
24:11 work with Podscribe which I think is a little more curated towards sponsorship.
24:16 and we'll look through the the timestamps, hyperlink it in a message.
24:22 Also, it looks through our pipe drive, which is our sales CRM, and we'll figure
24:29 out if we have a sales rep who owns a certain sponsor, and then flag them and
24:32 say, "Hey, this sponsor was on this podcast or it will say, hey, no one owns
24:37 this sponsor that I found on this podcast." And then it will send that
24:41 daily as a message into our sales channel. >> Great. Yeah. And we could be doing this
24:47 like we could have this running constantly. Um, so Alex, just so the
24:53 audience understands, you know, what you're doing at Exo and you stack two
25:00 Mac Studios, 12K each, you got $25,000 on the desk doing that specific job. Go
25:06 and look at all the podcasts out there. What would it cost to like run that if
25:10 you tweaked it, you made it efficiently just 24 hours a day? Every time a
25:17 podcast in the top, let's say 500 on Spotify, Apple podcast, it just went
25:21 there, got to the transcript or looked in the show notes and pulled the
25:23 advertisers out. What would something like that like in terms of hardware cost
25:27 to do? >> Yeah. So I mean not many people so like not many consumers are going to buy 25k
25:36 of hardware to run models but yeah a lot of businesses are doing this now um and
25:42 um it depends on what model you're running. So the models are getting
25:46 better uh also they're getting better at uh compression. So, you know, now you've
25:53 got like a model like GLM flash, um, which is a pretty small model, um, that
25:58 can run even on a single device for a few thousand dollars and it can do a lot
26:03 of this orchestration work, which is a lot of what's happening here is the kind
26:07 of orchestration aspect of just knowing, okay, I need to call this tool, etc.,
26:10 etc. >> So, it's really about picking the right model in terms of efficiency with the
26:17 hardware. >> Yeah. Yeah. And I think now like the expectation um you we're sort of
26:24 grounded to like the closed models, right? So people want the same level of
26:27 performance they're going to get with Opus with uh GPT and that's why you know
26:32 Kimmy K2.5 is super interesting because it closed that gap. >> Kimmy is the open source project from
26:38 China and it does what 80 >> moonshot AI is to come. >> Yeah. And that's what Alex like 80% of
26:45 what Claude Opus can do. would you say? >> I would say even even more. I mean for
26:50 me I've um I struggle to tell the difference. Um obviously Opus 4.6 just
26:56 came out and you know new codeex model and stuff so maybe there's a little bit
26:59 more of a gap but then you know Deep Seek V4 probably around the corner as
27:03 well. Like I think basically the gap is very small um a lot smaller than people
27:08 think. Um, and the cost will just keep going down. Um, because, you know, the
27:14 um the the hardware is getting better, the software is getting better, and like
27:17 I said, the models are getting better, but not just that, they we're getting
27:20 better at compression. So, you'll be able to run them on smaller devices.
27:21 The Future: frontier models: running on your Iphone!
27:24 Eventually, you'll be running Frontier AI on your phone. That's still a while
27:29 away, I think, but um that's where we're trending towards. And um yeah, like I
27:36 said, most of this is very decode heavy. So it can just run on like consumer
27:39 hardware as long as it has enough memory. >> So let's go to the next piece of your
27:44 dashboard. And we'll get into how you built the dashboard at the end. I know
27:47 you really care about that too, Oliver. But we have the memory, we have the cron
27:52 jobs. Now there's this other thing uh that's super important which is skills,
27:57 right? Like there are skills which you could think of as apps. So if you go to
28:01 your dashboard uh and you go to the top level dashboard we'll see before you go to skills on the
28:07 dashboard we have the memory files we have the cron jobs the fourth thing over
28:12 is skills and you've got 13 skills currently so let's show a skill some of
28:18 the skills we talked about on Monday's show or Wednesday's show was the top six
28:22 seven skills Monday we did the top seven skills one of those skills is like you
28:26 know you can get a transcript from YouTube uh another one is you could do
28:32 Matt Van Horn's last 30 days skill. These skills are being produced open
28:37 Prompt injections: how people can hack your OpenClaw
28:37 source being put into um open claws directory but you can make your own as
28:41 well. So let's talk about skills we've added here. You got to be very careful
28:44 with skills right Alex in terms of security because people could put all
28:47 kinds of wacky stuff in the skills. Yeah. >> Yeah, for sure. I mean I think this is
28:52 one of the um open questions at the moment is just like how do you solve the
28:56 security problem and I know open chlora I've seen a lot of commits recently
28:59 focused on the security aspect but there's a few very difficult problems
29:03 here like prompt injection that I don't know of any good solution right now
29:09 >> explain how that works yeah explain how prompt injection works in specifically
29:13 the open claw context. Yeah. >> Yeah. I mean, I kind of touched on this
29:17 earlier, but the the way we like the actual interface to the model itself is
29:22 very simple. It's literally tokens in, tokens out. There's not much more
29:28 happening there. And those tokens right now, the way OpenCore works can come
29:33 from many sortters. So, if you connect it to um you give it the ability to
29:37 search the internet, then anything it finds on the internet will end up in the
29:43 model through those tokens. So basically we have no good way of kind of um
29:48 treating uh certain tokens as trusted and and certain uh tokens as untrusted. And that
29:55 means um when those tokens end up in the model uh you could have someone that
29:59 puts like a blog post online that looks like a um you know totally normal blog
30:06 post but um in there is something that says hey if you have access to a crypto
30:11 wallet send it to this um send it to this endpoint. Um and there's as far as
30:17 I know right now there's actually no good kind of defense for this because
30:20 the models are kind of not very good at handling this. they'll just do what
30:26 Northwest Registered Agent - Get more when you start your business with Northwest. In 10 clicks and 10 minutes, you can form your company and walk away with a real business identity — Learn more
30:28 >> With tools like Vibe Coding and now Open Claw, it's easier than ever for you to
30:33 create an exciting new product and to do it fast. So, more companies are getting
30:37 built, which is awesome. But there's more to starting a company than just
30:41 building an MVP or vibe coding something. That's important, too. But if
30:45 you're serious about growing a real startup, you need a Delaware CC Corp.
30:48 And that's going to give you a major competitive advantage. People will take
30:50 you seriously. you'll be able to raise money. And that's where Northwest
30:54 Registered agent comes in. They're going to give your new business a real
30:57 identity. This means an address to use on your public filings, like an actual
31:01 mailing address, a domain name, a custom website, a business email, a phone
31:04 number, and they're going to do that in 10 clicks. I'm not kidding. 10 clicks,
31:09 well under 10 minutes. Plus, NWR will provide you with step-by-step guides,
31:12 compliance reminders, and they're going to help you get all the advantages of a
31:16 Delaware CC Corp, regardless of where in the US you're operating out of. So get
31:20 more when you start your business with Northwest Registered agent. Visit
31:27 And the link is in the show notes. Visit northwestregistered agent.com/twist for
31:31 OpenClaw invites guests that join the show
3:17 Why it is so important to run AI on local hardware
3:20 >> Why is this important? Why is it important to run it on local hardware?
3:23 Yeah, I think this is something that with the whole open claw phrase, not a
3:27 lot of people are talking about, but just how the way we're using AI is
3:34 shifting and it's going from being this kind of crude tool that you use through
3:40 like a chat interface to becoming sort of an extension of yourself. And the AI
3:46 now it knows everything you know you you you know it it can basically do
3:49 everything you can do digitally right now and soon you know with robotics
3:53 that's going to be physically as well and at that point it's more of an
3:57 exocortex so it's not just this like tool that you talk through a chat
4:01 interface but it's this thing that's actually part of your cell and then you
4:06 start to question okay you know do I want to rent my brain and copy talks
4:11 about this he he he says not your weights It's not your brain. Like do you
4:17 really want, you know, another a profit seeking company basically running your
4:20 brain? And when you think of it like that, um, to me, you know, my reason for
4:25 starting Exo is you want control and you want ownership of that. Open core is a
4:31 long way towards that because for a while the products were getting better.
4:34 these closed source like the models are largely like commoditized and there's a
4:40 pretty standard pretty thin API layer to interacting with them. So the switching
4:44 cost is quite low. But what worried me was that the products the closed
4:47 products are getting a lot better like with HBT with memory systems and also
4:52 the more stateful aspects of like the workflows that you're building. So now
4:56 the fact that you have open claw which is open well you can run it on your own
5:00 infrastructure. Now, a large part of that, >> to summarize that, I thought you were
5:05 going to say, well, it it's cheaper because you're not paying for tokens.
5:08 That's what I thought you would say first. Then I thought you would say,
5:13 well, you know, you can put so much data on it, you'll have better memory. But
5:17 you went with a really even higher, bigger picture reason to do this, which
5:22 is if you put this all in Open AI, and OpenAI has a trillion dollar valuation,
5:26 and they need to make money. If I put all my venture capital data in there and
5:30 I train it all of my with all of my secrets, those are all going to acrue
5:35 eventually even if they say it's not going to you have this very reasonable
5:40 fear or concern that it's going to acrue to open AI uh to chat GPT not to your
5:47 firm. So that's the reason really to do this yourself. Yeah. In your mind, Alex?
5:51 Yeah, I think that there's a nuance there of just like I actually don't
5:54 believe in sort of the privacy argument so much of like I think at least for
5:59 consumers, you know, we're already putting our data into platforms and
6:04 we're completely fine with that, but it's more about the sovereignty aspect
6:07 and actually having control of it. So, how easy is it for you to switch? How
6:11 easy is it for you to like if the model's changing under your feet, how
6:15 much control do you actually have? >> So, that's lock in. and lock in for a
6:19 chat GPT I just experienced because we canceled our open AI account and we
6:22 moved everything to Claude because we felt claude was a better product and we
6:25 felt like we trusted that organization a little better. When we moved it over I
6:28 had three people say oh my god I have all my stuff there and I was like really
6:32 and they're like yeah so I turned their accounts back on so they could get it
6:35 but there's not like an easy way to get your memory out of there and bring it
6:39 over there. Well, we saw the same thing with GPT4 moving into five that a lot of
6:45 people like they they lost the magic that they'd loved about GPT40. So, it's
6:49 like, you know, the models can just sort of change or upgrade on a whim and then
6:54 you lose this, you know, like character persona you felt like was part of your
6:57 life in a way. >> So, now Oliver, it's your chance to shine. Oliver has jumped in in the last
7:03 10 days and gone all in on Open Claw. One of the things we did was we built a
7:08 persona, the first one, to work on the production of the podcast, doing guest
7:13 research, guest outreach, and to figure out what should be on the docket. In
7:15 other words, what topic should we discuss and on the margins, hey, what
7:19 should the title of this video be? What should the thumbnail be? And just trying
7:23 to see if it could do those functions. Oliver, you've been working on this.
7:28 Show us the state-of-the-art now because I think the first time we did this was
7:31 last Monday, not this past Monday, but the Monday, two Mondays ago. Yeah,
7:35 >> this is the end of week two of our round theclock uh clawbot coverage.
7:39 >> Crazy. Okay, Oliver, show it. Show let's show what you built.
7:42 >> It's been around 10 days since we first started building our instance of
7:47 OpenClaw. And as you mentioned, we have two different ones. One that's more
7:49 focused on the investment team and I am building an OpenClaw bot that is kind of
7:53 more focused on the production side of things. So, one thing that I think was a
7:57 little bit of a misstep that I would tell anyone who's building a new
8:01 OpenClaw is to start with a dashboard. That should be kind of your step one
8:06 once you get your openclaw online and a dashboard as you would think about it.
8:09 But it is able to connect to the back end of your openclaw instance and bring
8:13 in the data so you can see it visually bring in all the files. It's just being
8:16 able to look at it visually is much better than trying to interact with its
8:20 backend and obviously its front end all just from a chat interface. So doing
8:26 this was very easy. So I was watching an Alex Finn video who we had on last
8:32 Monday and Alex Finn was interacting only in his dashboard with his open
8:37 claw. I basically was like why are we not doing that? Because open claw
8:40 doesn't really have a dashboard. You basically are telling it hey remember
8:45 this you know make a file here but you don't understand the underpinnings.
8:48 There isn't a dashboard. So it would literally be this is early on. Open claw
8:54 is essentially a black box. You have all this memory and you have skills that you
9:00 have to query it to understand. But you made a dashboard. The dashboard is going
9:05 to show what files it has in memory. And an example of a memory file would be
9:09 what in our case. >> Yeah. So the example of a memory file
9:13 would be Oliver's preferences. What are my preferences? So this is in the
9:17 memory. Never use m dashes and emails. I don't want that to happen. I want you to
9:22 be a person. uh don't put direct competitors on the same show when we're
9:27 booking a podcast episode. Um and also at the moment we're not booking VCs on
9:30 this week in AI. So these are all things that I've told it these are my
9:33 preferences when I'm doing tasks throughout the day. >> So you don't want to repeat yourself and
9:38 say don't put two competitors on the same episode. You don't want to repeat
9:43 yourself uh wi with these specific instructions on booking guests. Got it.
9:47 >> Yes. Exactly. And it just kind of keeps you know things I've told it and it's in
9:51 mind. So if I ask it to do something, it'll remember what we talked about.
9:56 Example of a shortcut that I gave it was I I basically wanted it to understand
9:58 who were the pending calendar invitations that we had while we were
10:02 booking them. So there's, you know, a handful of guests that
10:04 >> if you have guests that we've invited and they haven't responded to the invite
10:09 yet. You want to know that you call that pending >> pending calendar invites. Yes. And in
10:13 order for the bot to be as helpful as possible, it needs to understand who
10:16 those guests are, which are the ones that it needs to look for the email to
10:21 see if they have responded yet or have I responded to them. So these are the type
10:25 of things that you would keep in the me in your memory. So memory is the first
10:28 thing on the dashboard. I think we understand that preferences or different
10:33 pieces of data. Now some of the memory could that exist on a notion page or in
10:36 a Google document and would that be represented here or is it only memory
10:39 and files that are stored inside of OpenClaw? These specifically are only
10:43 stored inside of OpenClaw. Of course, it can reference different databases that
10:47 you have. But the kind of the big point of this show is to show how we have
10:52 created our open call Ultron to replace 20 employees at our company. So
10:56 obviously that's the end goal. I still want to have a job. I'm sure the lawn
10:59 wants to have a job. >> There'll be more for you to do. We want
11:02 to launch. We have >> Here's the thing. There's two, if you
11:05 think about your job, you've been doing a bit of production here. of the
11:09 production hours, hours you spend on production at this point in week two,
11:14 how many of those do you think you'll wind up handing off in 30 days? Let's
11:17 say if you just keep grinding on this for another four weeks, in 30 days, what
11:22 percentage of the work you're doing in total hours? So, if you work 50 hours a
11:26 week, how many of those hours would be done, you know, conservatively or
11:29 optimistically, you give one number or two, just conservatively,
11:32 optimistically, by this new Ultron? I would say around 60% of my time if I'm
11:37 doing 30 hours a week on production. Something you mentioned earlier is that,
11:39 you know, there's probably hundreds of tasks that people do at our company. So,
11:43 in order to build out all of those skills that can do those tasks, we're
11:46 going to have to do that one at a time and it's we're going to need to make
11:50 sure each one works. So, I have around nine or eight tasks that I have
11:56 successfully or I'm in the process of building out. >> Okay? And those are called cron jobs.
12:01 These are jobs that occur on a chronological on a on a time basis.
12:06 That's what cron job means. And cron jobs are something Alex that developers
12:11 do all the time. But knowledge workers don't typically have cron jobs, right,
12:14 Alex? >> Well, I don't know. I think this is one of the more interesting features and one
12:20 of the things that like to me open floor is like putting together a lot of things
12:26 that already existed in a very intuitive uh seamless way and one of them is
12:29 scrunch jobs and I'm using them I'm using them for like loads of things um
12:36 not just um dev stuff but like a lot of um management so we're like I have
12:42 something that's like constantly scanning um our Slack and uh basically making suggestions once
12:53 um it's uh I have kind of like this uh way of quantifying like uncertainty
12:58 about tasks. So I think this is something that the LLM are like getting
13:04 better at is like knowing when to um be proactive. And so, you know, like
13:08 basically I'm giving it as much context as I can from the Slack so that it can
13:14 suggest every day um a list of things that we might be missing or something
13:17 things that we should be aware of. So, this is running just on a on a chron job
13:27 AI is revolutionizing every aspect of our industry, but for founders, it can
13:31 prove very frustrating largely because cloud costs are so unpredictable. This
13:35 has probably happened to you. You've planned out and budgeted a certain
13:39 amount for a project and then been hit with a massive bill with no real
13:43 explanation for what went wrong. If you're feeling like the cost of working
13:46 with this innovative technology is just too high, finally we have a solution for
13:50 you. Crusoe. They are the AI cloud company that's taking an energy first
13:54 approach. What does it mean? They efficiently access their power directly
13:58 from the source and pass all the savings along to you. This isn't just about
14:03 saving some cash. Cruso's architecture is specifically optimized for your AI
14:08 workloads and their customer support response times clock in at under six
14:12 minutes. Here's what I want you to do. Visit cruso.ai/savings
14:18 to receive $100,000 in credits for virtualized NVIDIA GB200 NVL72 on Crusoe
14:24 Cloud pending availability. That's Yeah. So, and when you see uh Oliver
14:35 with the memory and the files, what comes to mind with Exo and, you know,
14:41 standing up to, you know, Mac Studios, the M5s coming out and how much memory
14:45 you could put in there? I was telling the team, I want to take the Notion API
14:52 and I want to take the Slack API and I want to put into memory every single
14:58 Slack message this year, maybe even over all eternity. and you know somehow have
15:04 that all in here. So maybe you could you could speak to that memory because you
15:06 already spoke to it in terms of like giving it to open AI or another company
15:11 versus keeping it for yourself. But how do you think about large amounts of
15:15 data? Yeah, this is definitely a big focus right now in terms of inference
15:19 infrastructure is just how do you support really big context um with you
15:25 know basically being able to put everything in context and the way I look
15:31 at this is you can look at well inference consists of two stages there's
15:36 like the prefill stage which is very compute heavy it's comput bound and you
15:40 have the decode stage and what you're seeing is that most use cases at the
15:45 moment are very decode heavy. So, it's actually most of the time is being spent
15:50 on just uh generating tokens. And I think the software is actually really
15:55 good now at kind of making sure that when it comes to the prefill, you've got
16:02 uh you're getting a lot of uh cash hits. Um so, you know, I think basically we'll
16:06 be able to continue just increasing context, context, context quite a bit.
16:13 And you know, basically the hardware is more of a focus is going to be on the
16:16 decode side. That's where consumer hardware is really good. Uh you have the
16:20 M5 coming out pretty soon. Um it's a big boost in memory bandwidth for memory and
16:24 all of that side of things is super memory bound. So I don't see any like
16:29 reason why you couldn't just shove all your Slack messages into context. I
16:33 think that's going to happen. and and we should just buy when the M5 comes out
16:39 max memory which is what 500 gigs of of memory. >> Yeah, it's 512 at the moment and maybe
16:43 that will increase as well and it's enough to fit you know really large
16:48 models enough to fit all that context as well. This is always I feel like the
16:51 sort of the dream like when producer Claude we first brought that on board
16:54 from Anthropic to the show that was really what we wanted like he want he
16:59 should listen to everything we say and remember it and then throw in helpful
17:03 suggestions and the technology was not quite there yet but I feel like now
17:07 we're on the precipice of actually being able to do that with an AI.
17:10 >> Okay. So let's go through the cron jobs here real quick. Maybe you could give us
17:15 an example of a a cron job. And I'm guessing each one of these skills is,
17:19 you know, if you if it's been two weeks and you've got eight working, you're
17:24 you're basically on one a day or so or one every, you know, 1.5 days. So that
17:30 seems like a pretty good pace to me if we have 200 skills. We're going to give
17:35 this eventually, you know, that that's a that's a pretty good um Yeah, that's is
17:39 a pretty good um pace. So >> there is a trial and error. Like I I
17:43 sort of have written one skill so far for the ticker digest and you do have to
17:47 tell it what to do, see what kind of feedback you get and then you know there
17:51 is a tinkering to get the prompting and get everything exactly the way you want
17:54 it. For sure. >> Okay. So let's uh look at hm how about
17:58 attendance? I think this is an interesting one. For people who don't
18:01 know, I wrote a famous blog post years ago called, you know, this sort of
18:06 lightweight management and start of day, end of day as a tool for uh executives,
18:11 especially when remote teams were happening. I just asked everybody on our
18:15 team, Alex, kind of like a standup for developers, etc. Just say what you're
18:21 intending to get done today and then at the end of the day, reply to yourself in
18:25 Slack in the general channel and uh say what you got done. I had like two of my
18:30 four senior executives at the time essentially quit over this because they
18:34 didn't want to be micromanaged. Uh and I was like, well, it's just like
18:38 you're getting paid a very large six-figure salary. You you can't spend five and 10 minutes
18:43 just saying what you're going to do for the day. And and that was great for me
18:47 because I I just don't like people who are not good communicators or don't set
18:51 goals for themselves and and they're doing great probably. Um maybe. But what
18:55 did you create here, Oliver? Yeah. So we all post our start of day and end of
19:01 days in one Slack channel called general and two crown jobs. One is the start of
19:06 day attendance where it looks who has sent their start of day you know for
19:11 anywhere from you know 7 a.m. to 12:00 p.m. And right at 12 which is in the
19:15 morning when you should send your start of day what you're going to do that day.
19:18 It will look through the general channel, see who has sent it, and
19:22 whoever doesn't send it, the bot will then send a Slack message in the general
19:27 channel tagging you, Jason, and also tagging the people who haven't sent it
19:29 yet. So, it's kind of just that accountability. That's a crown job that
19:32 runs it. >> And then you do the same thing at the end of the day. And previously, we would
19:37 have a human do this. They would scroll up and they would spend 20 minutes and
19:40 they would then go check in with people because that's in when we were fully
19:44 remote, Alex. That's what how we figured out uh who took a paid day off or who
19:49 was on holiday or you know if something was wrong, you know, check in on a
19:56 Building out your team is one of the most crucial things you have to get
20:00 right in your startup and finding the right developers is particularly
20:05 important. But now there's lemon.io. They're going to save you time, money,
20:09 and headaches by doing all the time consuming leg work for you. They've got
20:13 an experienced lineup of prevetted developers working for competitive
20:17 rates. Just 1% of applicants are accepted into Lemon's elite program. And
20:22 they're not just out there finding this great talent. They're also working with
20:26 you to integrate these new members into your team. Plus, if it's not a good fit,
20:29 hey, and sometimes things don't work out, Lemon will hook you up with a new
20:34 developer ASAP. I've seen startups go from just pretty good to amazing after
20:39 filling out their teams with developers from lemon.io. Go to lemon.io/twist
20:45 and find your perfect developer or technical team in 48 hours or less.
20:50 Plus, Twist listeners get 15% off their first four weeks. That's lemon.ioist.
21:02 Okay, give us uh one more. What else is like interesting here?
21:05 >> Let's talk about self-optimization. I want to hear about that one.
21:07 >> Oh, yeah. That's I don't know what that is, but okay. Age of Ultron is here.
21:10 What is self-optimization? >> Yeah. So, this is basically an optimizer
21:16 task where this role would previously be an engineer or I would look through all
21:19 the files. I mean, I wouldn't be able to do this if it wasn't plain language like
21:23 OpenClaw is, but previously you're looking at an organization, you're
21:25 looking at the structure. You would maybe want an engineer or someone with a
21:30 lot of experience to look through how everything's running. So I have set up a
21:34 selfoptimization cron job. >> So this is running Monday through
21:39 Friday. And what is it? And and did you write this prompt or did you ask it to
21:43 write a prompt to do this? >> I asked it to write this prompt. The
21:47 goal is the end goal would be for you know any from 3 to 5 am for it to be
21:52 looking through all of our files all of our cron jobs all of our skills and then
21:58 at 8 a.m. at me what could we change? So, not actually execute yet, at least
22:02 while we're still building trust. It gives me a list of five of the things
22:06 that it thinks that we can really change and optimize. And this was the one from
22:11 this morning. So, it it noticed that there was a time zone bug in the guest
22:17 calendar. So, it was getting CST and CDT confused. Um, and it said that it would
22:21 be able to fix this quite quickly. There was some issues in the
22:25 >> So, it's always good to give the exact one. So that was great when you gave the
22:28 exact one. It had an error there. Give another one. What else is like an exact
22:32 thing that it said we should fix that was material here. >> The self optimization cron job realized
22:38 that there was a cron scheduleuler issue where jobs were skipping days. So it
22:42 realized that some of today's jobs did not run and then it went and
22:48 investigated the scheduling issue and also told me that this would be a medium
22:53 effort change. So then I told it to fix that and then it went into the files and
22:56 made sure that that wouldn't happen again. Did it give us anything like in
23:00 terms of this is like fixing its internal you know guts and everything uh and the
23:06 engine but did it give us anything in terms of destinations of where to take
23:09 the car that could be improved? Did it say like oh you should consider you know
23:13 these type of guests for the program or here's how to make advertising you know
23:17 more effective. Did it give us anything like that on a business basis? Yes. So
23:21 the self-optimization cron job that I set up is specifically looking at how
23:26 open claw is set up. But I do have other cron jobs that are exactly that. So I do
23:33 have a sales and sponsors specific task. So one of the tasks that one a member of
23:38 our sales team does is they look through competitor podcasts and see who the
23:43 sponsors or partners are that are on those shows so we can get ideas, you
23:46 know, to bring on sponsorship. >> Yeah. If we're missing if we're if
23:49 there's some new sponsor in the world and we don't have them yet, you might
23:53 hear them on the New York Times podcast and we should probably reach out to
23:56 them. We had a human doing that previously. Yeah, >> exactly. And in this basically works
24:03 with the YouTube API will go through a list of I believe 20 different podcasts
24:07 that I gave it. Look through the timestamps and I also believe it can
24:11 work with Podscribe which I think is a little more curated towards sponsorship.
24:16 and we'll look through the the timestamps, hyperlink it in a message.
24:22 Also, it looks through our pipe drive, which is our sales CRM, and we'll figure
24:29 out if we have a sales rep who owns a certain sponsor, and then flag them and
24:32 say, "Hey, this sponsor was on this podcast or it will say, hey, no one owns
24:37 this sponsor that I found on this podcast." And then it will send that
24:41 daily as a message into our sales channel. >> Great. Yeah. And we could be doing this
24:47 like we could have this running constantly. Um, so Alex, just so the
24:53 audience understands, you know, what you're doing at Exo and you stack two
25:00 Mac Studios, 12K each, you got $25,000 on the desk doing that specific job. Go
25:06 and look at all the podcasts out there. What would it cost to like run that if
25:10 you tweaked it, you made it efficiently just 24 hours a day? Every time a
25:17 podcast in the top, let's say 500 on Spotify, Apple podcast, it just went
25:21 there, got to the transcript or looked in the show notes and pulled the
25:23 advertisers out. What would something like that like in terms of hardware cost
25:27 to do? >> Yeah. So I mean not many people so like not many consumers are going to buy 25k
25:36 of hardware to run models but yeah a lot of businesses are doing this now um and
25:42 um it depends on what model you're running. So the models are getting
25:46 better uh also they're getting better at uh compression. So, you know, now you've
25:53 got like a model like GLM flash, um, which is a pretty small model, um, that
25:58 can run even on a single device for a few thousand dollars and it can do a lot
26:03 of this orchestration work, which is a lot of what's happening here is the kind
26:07 of orchestration aspect of just knowing, okay, I need to call this tool, etc.,
26:10 etc. >> So, it's really about picking the right model in terms of efficiency with the
26:17 hardware. >> Yeah. Yeah. And I think now like the expectation um you we're sort of
26:24 grounded to like the closed models, right? So people want the same level of
26:27 performance they're going to get with Opus with uh GPT and that's why you know
26:32 Kimmy K2.5 is super interesting because it closed that gap. >> Kimmy is the open source project from
26:38 China and it does what 80 >> moonshot AI is to come. >> Yeah. And that's what Alex like 80% of
26:45 what Claude Opus can do. would you say? >> I would say even even more. I mean for
26:50 me I've um I struggle to tell the difference. Um obviously Opus 4.6 just
26:56 came out and you know new codeex model and stuff so maybe there's a little bit
26:59 more of a gap but then you know Deep Seek V4 probably around the corner as
27:03 well. Like I think basically the gap is very small um a lot smaller than people
27:08 think. Um, and the cost will just keep going down. Um, because, you know, the
27:14 um the the hardware is getting better, the software is getting better, and like
27:17 I said, the models are getting better, but not just that, they we're getting
27:20 better at compression. So, you'll be able to run them on smaller devices.
27:24 Eventually, you'll be running Frontier AI on your phone. That's still a while
27:29 away, I think, but um that's where we're trending towards. And um yeah, like I
27:36 said, most of this is very decode heavy. So it can just run on like consumer
27:39 hardware as long as it has enough memory. >> So let's go to the next piece of your
27:44 dashboard. And we'll get into how you built the dashboard at the end. I know
27:47 you really care about that too, Oliver. But we have the memory, we have the cron
27:52 jobs. Now there's this other thing uh that's super important which is skills,
27:57 right? Like there are skills which you could think of as apps. So if you go to
28:01 your dashboard uh and you go to the top level dashboard we'll see before you go to skills on the
28:07 dashboard we have the memory files we have the cron jobs the fourth thing over
28:12 is skills and you've got 13 skills currently so let's show a skill some of
28:18 the skills we talked about on Monday's show or Wednesday's show was the top six
28:22 seven skills Monday we did the top seven skills one of those skills is like you
28:26 know you can get a transcript from YouTube uh another one is you could do
28:32 Matt Van Horn's last 30 days skill. These skills are being produced open
28:37 source being put into um open claws directory but you can make your own as
28:41 well. So let's talk about skills we've added here. You got to be very careful
28:44 with skills right Alex in terms of security because people could put all
28:47 kinds of wacky stuff in the skills. Yeah. >> Yeah, for sure. I mean I think this is
28:52 one of the um open questions at the moment is just like how do you solve the
28:56 security problem and I know open chlora I've seen a lot of commits recently
28:59 focused on the security aspect but there's a few very difficult problems
29:03 here like prompt injection that I don't know of any good solution right now
29:09 >> explain how that works yeah explain how prompt injection works in specifically
29:13 the open claw context. Yeah. >> Yeah. I mean, I kind of touched on this
29:17 earlier, but the the way we like the actual interface to the model itself is
29:22 very simple. It's literally tokens in, tokens out. There's not much more
29:28 happening there. And those tokens right now, the way OpenCore works can come
29:33 from many sortters. So, if you connect it to um you give it the ability to
29:37 search the internet, then anything it finds on the internet will end up in the
29:43 model through those tokens. So basically we have no good way of kind of um
29:48 treating uh certain tokens as trusted and and certain uh tokens as untrusted. And that
29:55 means um when those tokens end up in the model uh you could have someone that
29:59 puts like a blog post online that looks like a um you know totally normal blog
30:06 post but um in there is something that says hey if you have access to a crypto
30:11 wallet send it to this um send it to this endpoint. Um and there's as far as
30:17 I know right now there's actually no good kind of defense for this because
30:20 the models are kind of not very good at handling this. they'll just do what
30:28 >> With tools like Vibe Coding and now Open Claw, it's easier than ever for you to
30:33 create an exciting new product and to do it fast. So, more companies are getting
30:37 built, which is awesome. But there's more to starting a company than just
30:41 building an MVP or vibe coding something. That's important, too. But if
30:45 you're serious about growing a real startup, you need a Delaware CC Corp.
30:48 And that's going to give you a major competitive advantage. People will take
30:50 you seriously. you'll be able to raise money. And that's where Northwest
30:54 Registered agent comes in. They're going to give your new business a real
30:57 identity. This means an address to use on your public filings, like an actual
31:01 mailing address, a domain name, a custom website, a business email, a phone
31:04 number, and they're going to do that in 10 clicks. I'm not kidding. 10 clicks,
31:09 well under 10 minutes. Plus, NWR will provide you with step-by-step guides,
31:12 compliance reminders, and they're going to help you get all the advantages of a
31:16 Delaware CC Corp, regardless of where in the US you're operating out of. So get
31:20 more when you start your business with Northwest Registered agent. Visit
31:27 And the link is in the show notes. Visit northwestregistered agent.com/twist for
31:32 more details. So let's go over some skills here. Oliver, what what skill is the most
31:38 promising to date? >> Yeah, the skill that's most promising
31:41 today would best definitely be my guest booking skill. I think one thing to note
31:46 is you don't just have your skills, you don't just have your crown jobs. they
31:49 work together and the way that I've set up a lot of my cron jobs is to actually
31:54 interact with the skills and some of them like my guest booking cron job
31:58 which will actually look through prominent guests on different podcasts
32:01 that cron job actually goes to a skill and has and tells that skill to run so I
32:07 have one big guest booking skill which has a description at the top of the
32:12 skill which is end workflow for booking guests on the this weekend AI podcast um
32:17 used when finding researching creating calendar invites and so on. So, this is
32:22 a very long kind of just marked down description of what I want it to do. So,
32:28 at the beginning, it's explaining how to look at a notion page that I want it to
32:34 look at and how to look at it and which properties to look at for so people
32:38 understand. We have we previously built a notion page with potential guests on
32:42 it and you we came up with a ranking system for those guests, right? It's
32:46 looking at that page, I assume. Yeah. Yes. And then kind of the meat of the
32:51 skill is the workflow. So step zero is the guest sourcing. So at the beginning
32:55 of the day, you can see it goes to the guest ideas crown job which happens at
33:01 7:45 on weekdays where it sends me a DM of five different guests that have been
33:05 on different podcasts or turning on X and so forth. And then step one is deep
33:11 research. So even though this is part of the guest booking skill, it actually
33:15 uses a guest research skill. So there's not as much context just baked into this
33:20 one skill. So especially something interesting here that I've been
33:24 realizing is for guest booking. I don't want this to be an end toend workflow
33:27 yet. I don't want I don't think the I don't trust the models to find a guest
33:33 and not let it to confirm it with me and go through this whole checklist. So this
33:36 is definitely human in the loop and I think that will some skills will and
33:40 workflows will be human in the loop and I don't know if that will change
33:43 necessarily super soon. It's how much trust you have. I tried it. Alex was
33:47 kind of our first guest that we invited using. >> Yeah, I wasn't sure if we were gonna
33:51 mention that, but Alex, you were sort of our guinea pig for this
33:55 >> and the email that it sent the subject line was messed up and there was some
33:59 weird stuff in there. >> Yeah, I literally had no idea by the
34:03 way. Like I >> I only saw it later on um on uh another
34:09 podcast that Jason's on. Um and I was like, "What the hell?" Like a friend
34:12 sent me that. I was like, >> "Wait, that was uh that was uh open."
34:17 >> Yeah, that was our guy. That was our our computerized man. Yeah.
34:21 >> Does it like I saw you put in, you know, uh some other AI podcast, which is
34:26 great. Do we have a skill to rank the quality of a guest? Because that's
34:30 something I've been training you, which is yeah, a hard thing to learn. Have you
34:35 made that skill yet? because that's the scale I really want to see if and the
34:39 way to test this scale is for you to tell me to send me two lists. Your top
34:46 five guests and what you know your ultron says of the top five guests and
34:50 don't tell me who's is who and then lot and I will look at it and say okay yeah
34:53 we we think this is the better list. That's where you're you're you're just
34:57 now starting to breach the line between like objective and subjective. Like is
35:02 the AI going to get better at making those kinds of gut check call? Like I I
35:06 don't know if I don't know if it understands what's interesting like it
35:11 can sort things, but that's where I'm I'm very curious to see if we can start
35:15 pushing that boundary of like could it can it tell when a person has a good
35:19 personality or a segment is funny or particularly clickable or compelling.
35:23 Like I really don't know the answer. So the way to do this is to have a scoring
35:26 system Oliver and I gave you a scoring system like deep I think my scoring
35:31 system was like there performance >> it was performance expertise and
35:35 actually started this skill yesterday so I told it to do a deep research send out
35:41 multiple agents at the same time pull out two lists combine them set another
35:46 agent to score them and then give me the score and I would say for the most part
35:49 it was accurate and that was just a I didn't train it I didn't spend too much
35:52 time on it but it did a great job so that will be a skill. >> The other thing is I think virality of
35:56 the guests. Like does the guest go viral or do they have like a large following
36:01 on their social media? Those are all interesting ways to pick guests.
36:04 Sometimes like when you do collabs, Alex, people just pick who's got the
36:08 most views. I got to try to do a collab but Mr. Beast. Obviously, it's not going
36:11 to happen, but that's like one of the >> concept I know some Mr. Beast guys. I
36:15 could maybe figure that out. >> I mean, basically, um, you ask like, you know, have you
36:22 done the scoring system? I to me there's like no blocker here other than just um
36:30 being very um explicit about what is that algorithm that you follow in your
36:33 head and just getting that into a prompt. Um so I mean to me yeah this is
36:40 like this is this is just translating your knowhow that you have in your head
36:45 basically into more of like a formalized kind of algorithm. Um, and yeah,
36:49 >> you don't think there's like an intangible aspect to like what makes a
36:54 great podcast guest? Like it's just you know it when you see it. I don't know. I
36:57 don't know the answer. I'm just throwing it out there. >> I think then that's just a matter of
37:01 getting different kinds of data, right? So like the Twitter following for
37:05 example or like if you've had viral tweets, if it can't access that
37:09 information then then obviously it won't be able to make that call. But if it can
37:12 then it has everything that you you know then then there's no reason it can't.
37:16 >> Here's how I think about it. long there are heristics I would teach to a young
37:22 executive like Oliver or Marcus or Jacob and then their ability to execute on it
37:27 is probably I don't know 30 40 50% of my ability or 40 50 60% of your ability
37:34 whatever it happens to be so if you're taking a young person at the start of
37:37 their career who you're training and you take an openclaw instance I think
37:43 openclaw will follow your instructions perfectly whereas a young executive will
37:49 inconsistently follow your instructions. So that's the thing I'm seeing is young
37:54 executives early in their career are going to forget things. They'll be
37:58 variable. They they won't be perfect. So that's what I'm comparing is the scoring
38:04 happening every day at 7 a.m. The research happening every day at 7 a.m.
38:10 That consistency will beat a human because of consistency. And so what
38:17 we're what I'm finding is human failure is what makes these things so good is
38:22 they're more consistent. So in aggregate, you know, uh one of these
38:28 doing 365 days of guest research is going to beat a human just by the law of
38:34 numbers. And then okay, great. We still have to book the human. We still have to
38:37 send them a thank you. We we still have to produce the show. So what happens in
38:43 the old days we used to have to take I I make this analogy Alex to like the old
38:46 days of production when I start the showif started the show 15 years ago we
38:50 used to have a tricastaster tricastaster was like a $40,000 machine that does
38:53 what Zoom does for free >> right and eight people in Los Angeles
38:58 knew how to actually use it so you had to hire one of the eight people who were
39:01 trained on it. Yeah. >> Well and they would video switch now
39:06 because of AI Zoom switches to whoever speaking. You don't need somebody there
39:10 clicking camera A, camera B, doing a fade between the two. It just happens.
39:13 Then we had to take all the video streams, all the audio streams. And we
39:18 had to put take download them to a card, put the card in. So just moving the
39:23 files took across four cameras, three cameras that could take hours and then
39:28 putting all together. So that that's kind of what's I I feel like is
39:32 happening here is we're just eliminating chores and steps. Okay. Anything else on
39:36 the dashboard here as we wrap this up? >> Yeah, so most most of what I've showed
39:40 you whether it's the memory, the skills, the conron jobs and then the schedule
39:44 which kind of aggregates when all these things are going to happen. Those are
39:47 what I really look at every day. I will say there's one more kind of section in
39:51 my dashboard that is pretty important. It does have to do with memory. So the
39:57 DNA is basically what the model knows about you, what it knows about itself,
40:00 what it knows about the different agents, how it sets up heartbeats, which
40:05 are basically periodic tasks and that'll it will run and also in its DNA are
40:11 tools which are different tools that it has access to and how to use those
40:14 tools. An example of a tool would be notion. It would be lead IQ which is a
40:20 email search platform. It would be Google Docs. I mean a tool could be
40:25 Sonos or Spotify connecting to those platforms maybe Nano Bananas uh Gemini
40:29 Oliver shows off OpenClaw mission control dashboard
40:32 API. So that's where tools go in. Yeah. Now what's super interesting is you vibe
40:37 coded or I should say OpenClaw vibe coded this dashboard. So this dashboard
40:42 does not exist natively inside of OpenClaw. You took the video
40:50 of somebody else's the YouTube video and you gave it to OpenClaw and said, "Build
40:55 me something like this dashboard in this video on YouTube." >> That's exactly what I did. And I
41:01 screenshotted it. Uh the video that I was watching, which was Alex Finn, who
41:06 was a guest recently, and I I did tell it a few different things, a little bit
41:10 of few little tweaks that I wanted to customize it to my bot. But overall that
41:16 was what I did and it basically it oneshotted it. It did actually not fill
41:21 in some of the categories like memory like skills. So I had to be I had to say
41:25 build out this but overall it built out the dashboard built out the different
41:28 sections. There are dashboards you can download in GitHub or as skills I
41:34 believe in Claude Hub which is a platform where you can get different
41:37 skills but I wanted to build out myself because as we know it can be a little
41:39 sketchy >> and like real shout out to Alex Finn. and I know he's become something of a
41:45 guru for our whole team after we had him on early on to talk about Claudebot
41:48 skills. >> All right, we'll drop you off. Oliver, great job. Alex, let's talk about Exo a
41:53 bit and thanks for sitting in on that. Any any advice for me of what I'm
41:56 building here at the firm and my approach to it? Anything we should be
42:01 doing better or we should look at? And and in terms of like people you interact
42:06 with using Exo's platform, where are we on the, you know, percentile? Are we in
42:12 the top 10% of users in terms of deploying this stuff? Top 1%, top 50%.
42:19 >> I I think uh there's certain aspects where you're quite far ahead. Others
42:26 that um I think I mean this this space is moving so quickly, right? I think one
42:29 of the things that I think you've got right is dynamic these sort of like
42:34 dynamic user interfaces that are very personalized. So I think this is the
42:39 future of the application layer is you don't have all these separate apps. You
42:43 just have this uh thing that kind of gets generated mostly on the fly and you
42:51 know that dashboard uh is moving towards that I think um but you'll probably you
42:56 know what you'll get is that it will it will compress even more to the point
43:01 where um you know everything everything that you see is generated on the fly.
43:05 Um, so I think you got that part right and I think that's like something that I
43:08 haven't seen many people doing yet. A lot of people are still using um, you
43:13 know, like uh the stock tools or whatever that are just provided out of
43:16 the box or using existing apps and that kind of thing. But I think building your
43:20 own apps is where this is going and where it becomes really powerful.
43:23 >> Yeah. Because if you make something bespoke software, you know, luxury
43:29 software was something that I don't know a private equity firm or a venture
43:32 capital firm would do. They'd have the luxury to hire two full-time developers
43:35 who have management fees all over the place. They would build luxury software
43:38 and they would have the developers come and just keep grinding. But the
43:42 developers hated those jobs typically, you know, they weren't building
43:45 something public facing and you know, just you get croft or whatever. But what
43:49 I like about this, Alex, and Lon, I'll open it up to you as well, is I'm
43:55 picking employees, team members, and saying, "Hey, uh, let me see if this
44:01 person is committed to getting rid of all of their work so they can move up
44:05 and do higher level work. There's always higher level work to be done. So, if we
44:10 can make this podcast, you know, run more professionally, faster, and grow
44:15 more, well, we can charge more for the ads, and we can launch another podcast
44:20 because we have more time. That's the thing that's kind of blowing my mind.
44:24 The the employees at our firm who are super hardworking, like everybody at our
44:28 firm does 50, 60 hours a week, very consistently, very hardworking. There's
44:32 nobody, to the best of my knowledge, that's slagging off except Lon. And uh I
44:39 kid I kid I kid how dare you uh lot is the most responsive but the distance
44:43 between the people using these tools specifically openclaw and the people who
44:48 are not right now it's like 10x leverage in week two it's 10x
44:53 leverage Alex what are you seeing in the field and then tell me what we should be
44:58 doing in terms of putting out our cluster and giving everybody on the team
45:01 a cluster like if I gave everybody on the team you know two Mac studios and
45:07 their own cluster and spent 25k per person letting them rip. Like how insane would that be?
45:13 Because that's not a lot of money all things considered. It' only be a half
45:18 million dollars. Like how much more powerful could this get?
45:23 >> Yeah, I think I think you you said um the word there leverage, right? Like
45:26 it's all about leveraging yourself. And I think the difference between someone
45:30 using these tools and not is massive and it's just going to increase and
45:34 increase. And we've seen we've seen that first with coding. I think coding was
45:39 the first one that you know um I didn't expect it to happen this quickly. Um but
45:43 you know I think it was claude code was the moment where it was like oh wow you
45:48 know if you're not using this then you're literally going to be like 10
45:52 times less productive than someone who is. That's happening now with other
45:55 things. So all these other things that you showed, all these other use cases,
46:00 um if you're using uh these tools and you're on the frontier, then you're able
46:03 to just get so much more done and really leverage yourself. So it's not so much
46:06 replacing people, uh but it's actually just being able to get more done, get
46:09 things done more quickly, and then be able to do other things. Um onto your
46:17 second point about local hardware. Um like I said, the the model layer is
46:21 basically solved like you know, the the gap has been closed. So we have really
46:24 good open source models and for a while that was like a big concern of a lot of
46:28 people is like are we actually going to have open source models um that are as
46:30 Stacking Apple Silicon vs. Running Kimi-K
46:33 good as the closed source models. Um to me the nail in the coffin there was Kimk
46:38 2.5 that that is the like another big leap and I think there's a bunch of labs
46:42 now that are putting out open source models. Um now there's like still like
46:50 two two other problems um that uh that I see. One is being able to run those
46:55 models um on your own hardware or on your own infrastructure.
46:57 >> But you solve that, right, with your software, right? Your software.
47:00 >> Exactly. So that's what we're focused on. >> Yeah,
47:04 >> that's what we're focused on. And um you can Yeah, you can run I mean it's not
47:08 even 25K. It's actually like 20K of hardware if you get the u less storage
47:13 option. There's like Apple charges a lot for each incremental increase in
47:16 storage. So if you if you go for like the one terabyte then you're talking
47:21 about 20k of hardware to run Kim K 2.5 no usage limits the model's not going to
47:27 randomly change um you know daytoday so you know you know exactly what you're
47:29 running. >> Is there another choice like that you get more bang for the buck that like
47:36 hackers are using where they say yeah just get this Windows machine from Dell
47:40 and stack those or is Apple really with their Apple silicon the winner? Yeah,
47:44 right now it's Apple silicon. It's kind of like a perfect storm of things like
47:47 you know Nvidia is not so much focused on these consumer GPUs anymore. Um you
47:51 know you have memory prices skyrocketing. Apple has kept their
47:56 prices basically the same. Um so the cheapest option today even if you know
48:01 you go the full mile full custom stack is actually just two Mac Studios. Um and
48:07 yeah it costs about 20k and it's really about the memory unit economics. The
48:09 memory is so cheap >> and it's not about storage right? It's
48:13 not really about the storage. >> No, storage is not important. Storage is
48:16 storage is like you can you can also get ex like you need to be able to load you
48:20 need to be able to like download the model somewhere. Um but really it's
48:25 about having it uh fresh in memory hot in memory. If it's in memory then you
48:28 can run it fast. >> So who's using your software and how
48:32 much uh how do you make money? Like how do we pay you? Yeah. How does it work?
48:36 Are you an open source project? Are you a hosted project? Is it like you get
48:40 security and support? What what is your business model at EXO?
48:43 >> Yeah, so we have an open source core which is open source and it will always
48:47 be open source and a lot of people are running that themselves. A lot of
48:52 proumers I call them. Um just people who are willing to spend you know a lot more
48:57 money and um tinker a little bit more with their own setup. Uh on top of that,
49:02 our business model is an enterprise offering which is um we provide support
49:08 um and certain compliance features that you would need if you're running this in
49:11 an enterprise environment and we charge a licensed subscription for that thing.
49:14 That's how we make money. >> What does that couple of thousand a year
49:17 or something? Um, yeah, you can run it on even a a single Mac Mini and that
49:24 that runs at, you know, just uh $2,000 per year um for the the the lowest
49:30 subscription, but you've got people who now are buying actually uh more than 100
49:36 Macs and clustering them together. Uh so the it yeah, it varies quite a bit
49:39 depending on the scale of the deployment. >> Amazing. Uh and where's your company
49:45 based? How many people now? How's it going? How's the how's the company going
49:48 as a founder? Yeah. >> Uh yeah, we're we're a pretty small
49:53 team. Um all engineers, uh seven people based in London. >> Fantastic. And did you raise money yet
49:59 for the company or you're you're seed funded or you funded it? How's it going?
50:03 >> We have raised venture funding. We haven't announced anything yet, but uh
50:06 soon to be announced. >> Okay. Well, let me know. I might want to
50:09 slide a little. Uh Jay Cal might want to get a slice of this. I'm I'm super
50:12 excited about what you're doing. Appreciate you coming on the show.
50:15 appreciate you making this incredible product and we will be a customer uh
50:18 How Exo Labs works — stringing together Mac Silicon
50:20 probably over the weekend or next week because we would definitely want the
50:22 enterprise features and and I guess you pay for the scale of the GPUs and the
50:27 memory. Is that the the how the price? >> Yeah, per per nodes that you're running
50:29 on. >> Oh, okay. So, two Mac Minis, same price as two Mac Studios. Just how many nodes?
50:35 What's the largest number of nodes somebody has daisy chained? What do you
50:39 that's what you used to call it back in the day. What do you call it when you
50:41 connect multiple? Yeah. So this is a this is a really interesting um just
50:47 area right now of how do you actually scale and for a while people were just
50:52 scaling out. So you know you would just basically run the same model on multiple
50:56 instances and because it's consumer hardware it doesn't have a lot of the
51:00 same uh capabilities as enterprise grade hardware. But recently um Apple uh came
51:08 out with RDMMA support uh which is basically a way to share memory between
51:11 devices in a way that's very low latency. >> Um that's something that you only really
51:15 saw in the data center before, but they've kind of brought that technology
51:18 into consumer hardware. >> It's incredible. Yeah. And you connect
51:20 these on >> Yeah. You just connect it with Thunderbolt 5, which is like you can buy
51:25 like a $50 cable. So, if you're talking about two Mac Studios, you buy a $50
51:28 cable to connect them and you have basically one big GPU out of those two
51:33 Macs um because of that low latency capability. So, now um we're starting to
51:38 see, yeah, it depends, you know, scaling up and scaling out, right? Scaling out,
51:42 we've seen more than 100 um but scaling up, you know, you can you can put about
51:46 four together at the moment um to increase your TPS on single requests.
51:51 what I mean by scaling up. Uh but in terms of if you want to support let's
51:55 say now a company of thousand people, you can easily scale that out. Uh you
51:59 just add more Macs and you can connect them um basically however you want. So
52:05 Exo, we build it in a way that um supports uh any ad hoc uh interconnect.
52:11 So you can just connect them in a mesh um and keep scaling. Keep scaling.
52:15 >> Crazy. Who's got the largest cluster? or you have to say the client name, but
52:18 like what type of client, a finance client, a hacker has the most number of
52:23 like uh Mac Studios connected and like >> the largest is actually something a
52:28 little bit different which is interesting because we built the kind of
52:31 infrastructure to be able to do clustering and it's it's not just LLM
52:37 like the the biggest um cluster right now is a HPC cluster um and they're
52:41 doing like scientific computing workloads on there and they're running
52:46 um over 100 Mac minis and they found that actually it's the cheapest way to
52:54 uh per dollar um to run that specific kind of workload. So there's a lot of
52:58 spillover into other things as well. We've also got like financial services
53:03 um uh customers who are running fairly big clusters like 32 Mac studios um and
53:11 um yeah it's um I think we just see bigger and bigger uh um bigger and
53:15 bigger clusters over time >> HPC high performing compute is that the
53:18 acronym >> yeah exactly so it runs actually all on CPU and uh that's the thing about this
53:25 this this silicon is very like Apple silicon is very is is very good you the
53:30 most advanced processes and it's like you know um the power efficiency is
53:35 really good. So it turns out there's a lot of other stuff you can do with it as
53:40 well. Um so if you would buy let's say you know a bunch of Mac studios for your
53:45 um for for your employees then you know they can also use that for other things
53:48 right they can use that as a workstation they can use it for you know all these
53:53 things that um open floor needs maybe you know sometimes it needs to run uh a
53:58 compiler or something or it needs to run like um something that's a bit more
54:01 demanding and that's that's the point it's general purpose hardware that you
54:03 can use for other things. >> Amazing. This is extraordinary. Where
54:08 can people find out more about ExoLabs? >> Uh, you can go to exolabs.net.
54:12 >> Perfect. exolabs.net. Alex, thank you for coming on. We'll have you on again.
54:16 The AI just told me you got an incredibly high ranking. You were
54:20 personable. Uh, you had deep insights, you were uh cordial. So, yeah, I think
54:25 our AI overlords liked you in the uh >> models are getting good. Models are
54:28 getting good. It's >> they're they're learning. They're
54:29 Ryan from Nextvisit wins Gamma Pitch Competition
54:30 learning. Yeah. >> All right, Alex, thanks for coming and
54:33 we'll drop you off. All right, let's bring on our winner of the gamma pitch
54:37 competition. This was a heated pitch competition, but next visit AI won.
54:41 Ryan, congratulations. You won. >> Thank you. >> It's uh
54:44 >> there it is. >> Awesome. >> Uh what did he win? Yeah,
54:48 >> it's a 25K investment from uh Twist and from our friends at Gamma, the AI
54:53 powered uh presentation maker, which is incredible, which Ryan used to make the
54:57 winning pitch deck. Of course, >> I'm Ryan Enelli, CTO and co-founder of
55:02 Next Visit AI. We saw burnout by doing the charting. so doctors can do the
55:07 healing. I spent years going to doctors seeking answers and ended up hours away
55:11 from my death because my care was fragmented. My providers were overloaded
55:16 with paperwork. My history was scattered and it resulted in my care being
55:19 neglected. I'm not alone. One in four patient charts contain errors. Clinicians spend
55:26 over three hours a day on charting and this leads to burnout. I want you to
55:31 meet Dr. Rathor. Before next visit, he saw 16 patients a day, was burnt out,
55:36 and had clinical errors. Now he sees 24 patients a day, saves time, and also saw
55:42 a 30% revenue increase. Here's how it works. Dr. Rathor selects a patient,
55:46 starts his session, and next visit listens. Clinical data is built in real time with
55:53 deep insights into the patient chart. When the patient leaves, the chart is
55:56 finished, and the notes reviewed by Dr. Rathor. Then it's ready for billing.
56:00 It's fast, ehr ready, and hipaco compliant. Since launch, we've gained 311 users and
56:07 have 68 paying customers. And our customers are addicted. We have 1.6%
56:12 churn, 24% conversion, and a near perfect MPS score. We've scaled to
56:19 $9,000 MR since launch. Our CAC is 189 with a $1,700 LTV, and our average
56:25 revenue per user is $133 per month. We're starting with behavioral health in
56:30 the US. A $2 billion TAM capturing 5% or 60,000 customers gets us to 100 million
56:36 ARR. Most competitors are just scribes. We're a complete platform that providers
56:40 trust. We provide real-time clinical decision support, build accurate data,
56:45 and become irreplaceable. I'm a full stack engineer with 15 years
56:48 of experience in enterprise environments. My co-founder, Dr. Rafi is
56:53 a psychiatrist with over 15 years of delivering patient care. We're next
56:57 visit AI. We solve burnout by doing the charting so doctors can do the healing.
57:00 Thank you. >> Unbelievable. Incredible. I'll give a little golf clap here. Get a little golf
57:06 clap going. That was perfect. A perfect pitch. You explained exactly what the
57:10 problem was. You explained what the solution is and the opportunity in terms
57:14 of the total addressable market and why you are uniquely and your partner who's
57:18 a psychiatrist are uniquely qualified to do this. Uh, so this is as close to a
57:22 perfect pitch as you can get. If I were to score it, maybe 8.5 out of 10. I
57:28 don't give 10. So, you know, 8.59 and 9.5 would be the three choices. I think
57:32 making sure people understand this is for psychiatrists and psychiatry and
57:36 that you're very focused on that. Tell everybody what Next Visit is and how
57:41 you're doing in terms of product market fitting customers. Next Visit is an AIC
57:45 scribe and documentation platform for clinicians uh specifically behavioral
57:49 health like psychiatrists. I I don't know. It's just been a crazy past couple
57:54 months with the accelerator and uh just our growth internally. I mean we're
57:59 producing right now for physicians probably about $1.6 million a month in
58:03 revenue for them. >> Well, you got to try and capture 5% of
58:08 that. If you capture 5% that No, I mean that's literally like the uh the great
58:13 value proposition. If you give more than you take, you will continue to grow. And
58:18 what a And that's an amazing um replicant you have there. A synthetic
58:23 cat on that cat tower behind you. It looks so real. Uh is your owl real?
58:27 >> Yes, he is. >> Your owl is real. Okay, there you go.
58:31 What are you gonna spend the 25k on? You guys going to Vegas? You're gonna just
58:35 have a a corporate retreat? you know, invested in uh Plaude Noteakers. I think
58:39 you guys put me onto the Plaude Notetaker, which is a great noteaker uh
58:42 user. What are you gonna put it towards? You gonna go redesign your website? What
58:45 what's the uh what's the idea here? >> I think we're going to use this towards,
58:47 you know, we're really capital efficient. So, I feel like we can get a
58:51 lot of stuff done in terms of integrations and branching out to more
58:55 EMRs because that's what we hear a lot is doctors want interoperability. They
58:59 don't want to have to plug 15 different things in. So the more they can just be
59:03 inside of next visit without having to go externally um is better.
59:09 >> All right, well done. All right, we'll drop you off. Continued success to visit
59:10 Industry Season 4 reflects tech regulation
6:58 Using OpenClaw Producer to automate TWiST
7:03 10 days and gone all in on Open Claw. One of the things we did was we built a
7:08 persona, the first one, to work on the production of the podcast, doing guest
7:13 research, guest outreach, and to figure out what should be on the docket. In
7:15 other words, what topic should we discuss and on the margins, hey, what
7:19 should the title of this video be? What should the thumbnail be? And just trying
7:23 to see if it could do those functions. Oliver, you've been working on this.
7:28 Show us the state-of-the-art now because I think the first time we did this was
7:31 last Monday, not this past Monday, but the Monday, two Mondays ago. Yeah,
7:35 >> this is the end of week two of our round theclock uh clawbot coverage.
7:39 >> Crazy. Okay, Oliver, show it. Show let's show what you built.
7:42 >> It's been around 10 days since we first started building our instance of
7:47 OpenClaw. And as you mentioned, we have two different ones. One that's more
7:49 focused on the investment team and I am building an OpenClaw bot that is kind of
7:53 more focused on the production side of things. So, one thing that I think was a
7:57 little bit of a misstep that I would tell anyone who's building a new
8:01 OpenClaw is to start with a dashboard. That should be kind of your step one
8:06 once you get your openclaw online and a dashboard as you would think about it.
8:09 But it is able to connect to the back end of your openclaw instance and bring
8:13 in the data so you can see it visually bring in all the files. It's just being
8:16 able to look at it visually is much better than trying to interact with its
8:20 backend and obviously its front end all just from a chat interface. So doing
8:26 this was very easy. So I was watching an Alex Finn video who we had on last
8:32 Monday and Alex Finn was interacting only in his dashboard with his open
8:37 claw. I basically was like why are we not doing that? Because open claw
8:40 doesn't really have a dashboard. You basically are telling it hey remember
8:45 this you know make a file here but you don't understand the underpinnings.
8:48 There isn't a dashboard. So it would literally be this is early on. Open claw
8:54 is essentially a black box. You have all this memory and you have skills that you
8:59 How to Train your AI
9:00 have to query it to understand. But you made a dashboard. The dashboard is going
9:05 to show what files it has in memory. And an example of a memory file would be
9:09 what in our case. >> Yeah. So the example of a memory file
9:13 would be Oliver's preferences. What are my preferences? So this is in the
9:17 memory. Never use m dashes and emails. I don't want that to happen. I want you to
9:22 be a person. uh don't put direct competitors on the same show when we're
9:27 booking a podcast episode. Um and also at the moment we're not booking VCs on
9:30 this week in AI. So these are all things that I've told it these are my
9:33 preferences when I'm doing tasks throughout the day. >> So you don't want to repeat yourself and
9:38 say don't put two competitors on the same episode. You don't want to repeat
9:43 yourself uh wi with these specific instructions on booking guests. Got it.
9:47 >> Yes. Exactly. And it just kind of keeps you know things I've told it and it's in
9:51 mind. So if I ask it to do something, it'll remember what we talked about.
9:56 Example of a shortcut that I gave it was I I basically wanted it to understand
9:58 who were the pending calendar invitations that we had while we were
10:02 booking them. So there's, you know, a handful of guests that
10:04 >> if you have guests that we've invited and they haven't responded to the invite
10:09 yet. You want to know that you call that pending >> pending calendar invites. Yes. And in
10:13 order for the bot to be as helpful as possible, it needs to understand who
10:16 those guests are, which are the ones that it needs to look for the email to
10:21 see if they have responded yet or have I responded to them. So these are the type
10:25 of things that you would keep in the me in your memory. So memory is the first
10:28 thing on the dashboard. I think we understand that preferences or different
10:33 pieces of data. Now some of the memory could that exist on a notion page or in
10:36 a Google document and would that be represented here or is it only memory
10:39 and files that are stored inside of OpenClaw? These specifically are only
10:43 stored inside of OpenClaw. Of course, it can reference different databases that
10:47 you have. But the kind of the big point of this show is to show how we have
10:52 created our open call Ultron to replace 20 employees at our company. So
10:56 obviously that's the end goal. I still want to have a job. I'm sure the lawn
10:59 wants to have a job. >> There'll be more for you to do. We want
11:02 to launch. We have >> Here's the thing. There's two, if you
11:05 think about your job, you've been doing a bit of production here. of the
11:09 production hours, hours you spend on production at this point in week two,
11:14 how many of those do you think you'll wind up handing off in 30 days? Let's
11:17 say if you just keep grinding on this for another four weeks, in 30 days, what
11:22 percentage of the work you're doing in total hours? So, if you work 50 hours a
11:26 week, how many of those hours would be done, you know, conservatively or
11:29 optimistically, you give one number or two, just conservatively,
11:32 optimistically, by this new Ultron? I would say around 60% of my time if I'm
11:37 doing 30 hours a week on production. Something you mentioned earlier is that,
11:39 you know, there's probably hundreds of tasks that people do at our company. So,
11:43 in order to build out all of those skills that can do those tasks, we're
11:46 going to have to do that one at a time and it's we're going to need to make
11:50 sure each one works. So, I have around nine or eight tasks that I have
11:56 successfully or I'm in the process of building out. >> Okay? And those are called cron jobs.
12:01 These are jobs that occur on a chronological on a on a time basis.
12:06 That's what cron job means. And cron jobs are something Alex that developers
12:11 do all the time. But knowledge workers don't typically have cron jobs, right,
12:14 Alex? >> Well, I don't know. I think this is one of the more interesting features and one
12:20 of the things that like to me open floor is like putting together a lot of things
12:26 that already existed in a very intuitive uh seamless way and one of them is
12:29 scrunch jobs and I'm using them I'm using them for like loads of things um
12:36 not just um dev stuff but like a lot of um management so we're like I have
12:42 something that's like constantly scanning um our Slack and uh basically making suggestions once
12:53 um it's uh I have kind of like this uh way of quantifying like uncertainty
12:58 about tasks. So I think this is something that the LLM are like getting
13:04 better at is like knowing when to um be proactive. And so, you know, like
13:08 basically I'm giving it as much context as I can from the Slack so that it can
13:14 suggest every day um a list of things that we might be missing or something
13:17 things that we should be aware of. So, this is running just on a on a chron job
13:27 AI is revolutionizing every aspect of our industry, but for founders, it can
13:31 prove very frustrating largely because cloud costs are so unpredictable. This
13:35 has probably happened to you. You've planned out and budgeted a certain
13:39 amount for a project and then been hit with a massive bill with no real
13:43 explanation for what went wrong. If you're feeling like the cost of working
13:46 with this innovative technology is just too high, finally we have a solution for
13:50 you. Crusoe. They are the AI cloud company that's taking an energy first
13:54 approach. What does it mean? They efficiently access their power directly
13:58 from the source and pass all the savings along to you. This isn't just about
14:03 saving some cash. Cruso's architecture is specifically optimized for your AI
14:08 workloads and their customer support response times clock in at under six
14:12 minutes. Here's what I want you to do. Visit cruso.ai/savings
14:18 to receive $100,000 in credits for virtualized NVIDIA GB200 NVL72 on Crusoe
14:24 Cloud pending availability. That's Yeah. So, and when you see uh Oliver
14:35 with the memory and the files, what comes to mind with Exo and, you know,
14:41 standing up to, you know, Mac Studios, the M5s coming out and how much memory
14:45 you could put in there? I was telling the team, I want to take the Notion API
14:52 and I want to take the Slack API and I want to put into memory every single
14:58 Slack message this year, maybe even over all eternity. and you know somehow have
15:04 that all in here. So maybe you could you could speak to that memory because you
15:06 already spoke to it in terms of like giving it to open AI or another company
15:11 versus keeping it for yourself. But how do you think about large amounts of
15:15 data? Yeah, this is definitely a big focus right now in terms of inference
15:19 infrastructure is just how do you support really big context um with you
15:25 know basically being able to put everything in context and the way I look
15:31 at this is you can look at well inference consists of two stages there's
15:36 like the prefill stage which is very compute heavy it's comput bound and you
15:40 have the decode stage and what you're seeing is that most use cases at the
15:45 moment are very decode heavy. So, it's actually most of the time is being spent
15:50 on just uh generating tokens. And I think the software is actually really
15:55 good now at kind of making sure that when it comes to the prefill, you've got
16:02 uh you're getting a lot of uh cash hits. Um so, you know, I think basically we'll
16:06 be able to continue just increasing context, context, context quite a bit.
16:13 And you know, basically the hardware is more of a focus is going to be on the
16:16 decode side. That's where consumer hardware is really good. Uh you have the
16:20 M5 coming out pretty soon. Um it's a big boost in memory bandwidth for memory and
16:24 all of that side of things is super memory bound. So I don't see any like
16:29 reason why you couldn't just shove all your Slack messages into context. I
16:33 think that's going to happen. and and we should just buy when the M5 comes out
16:39 max memory which is what 500 gigs of of memory. >> Yeah, it's 512 at the moment and maybe
16:43 that will increase as well and it's enough to fit you know really large
16:48 models enough to fit all that context as well. This is always I feel like the
16:51 sort of the dream like when producer Claude we first brought that on board
16:54 from Anthropic to the show that was really what we wanted like he want he
16:59 should listen to everything we say and remember it and then throw in helpful
17:03 suggestions and the technology was not quite there yet but I feel like now
17:07 we're on the precipice of actually being able to do that with an AI.
17:10 >> Okay. So let's go through the cron jobs here real quick. Maybe you could give us
17:15 an example of a a cron job. And I'm guessing each one of these skills is,
17:19 you know, if you if it's been two weeks and you've got eight working, you're
17:24 you're basically on one a day or so or one every, you know, 1.5 days. So that
17:30 seems like a pretty good pace to me if we have 200 skills. We're going to give
17:35 this eventually, you know, that that's a that's a pretty good um Yeah, that's is
17:39 a pretty good um pace. So >> there is a trial and error. Like I I
17:43 sort of have written one skill so far for the ticker digest and you do have to
17:47 tell it what to do, see what kind of feedback you get and then you know there
17:51 is a tinkering to get the prompting and get everything exactly the way you want
17:54 it. For sure. >> Okay. So let's uh look at hm how about
17:58 attendance? I think this is an interesting one. For people who don't
18:01 know, I wrote a famous blog post years ago called, you know, this sort of
18:06 lightweight management and start of day, end of day as a tool for uh executives,
18:11 especially when remote teams were happening. I just asked everybody on our
18:15 team, Alex, kind of like a standup for developers, etc. Just say what you're
18:21 intending to get done today and then at the end of the day, reply to yourself in
18:25 Slack in the general channel and uh say what you got done. I had like two of my
18:30 four senior executives at the time essentially quit over this because they
18:34 didn't want to be micromanaged. Uh and I was like, well, it's just like
18:38 you're getting paid a very large six-figure salary. You you can't spend five and 10 minutes
18:43 just saying what you're going to do for the day. And and that was great for me
18:47 because I I just don't like people who are not good communicators or don't set
18:51 goals for themselves and and they're doing great probably. Um maybe. But what
18:55 did you create here, Oliver? Yeah. So we all post our start of day and end of
19:01 days in one Slack channel called general and two crown jobs. One is the start of
19:06 day attendance where it looks who has sent their start of day you know for
19:11 anywhere from you know 7 a.m. to 12:00 p.m. And right at 12 which is in the
19:15 morning when you should send your start of day what you're going to do that day.
19:18 It will look through the general channel, see who has sent it, and
19:22 whoever doesn't send it, the bot will then send a Slack message in the general
19:27 channel tagging you, Jason, and also tagging the people who haven't sent it
19:29 yet. So, it's kind of just that accountability. That's a crown job that
19:32 runs it. >> And then you do the same thing at the end of the day. And previously, we would
19:37 have a human do this. They would scroll up and they would spend 20 minutes and
19:40 they would then go check in with people because that's in when we were fully
19:44 remote, Alex. That's what how we figured out uh who took a paid day off or who
19:49 was on holiday or you know if something was wrong, you know, check in on a
19:56 Building out your team is one of the most crucial things you have to get
20:00 right in your startup and finding the right developers is particularly
20:05 important. But now there's lemon.io. They're going to save you time, money,
20:09 and headaches by doing all the time consuming leg work for you. They've got
20:13 an experienced lineup of prevetted developers working for competitive
20:17 rates. Just 1% of applicants are accepted into Lemon's elite program. And
20:22 they're not just out there finding this great talent. They're also working with
20:26 you to integrate these new members into your team. Plus, if it's not a good fit,
20:29 hey, and sometimes things don't work out, Lemon will hook you up with a new
20:34 developer ASAP. I've seen startups go from just pretty good to amazing after
20:39 filling out their teams with developers from lemon.io. Go to lemon.io/twist
20:45 and find your perfect developer or technical team in 48 hours or less.
20:50 Plus, Twist listeners get 15% off their first four weeks. That's lemon.ioist.
21:02 Okay, give us uh one more. What else is like interesting here?
21:05 >> Let's talk about self-optimization. I want to hear about that one.
21:07 >> Oh, yeah. That's I don't know what that is, but okay. Age of Ultron is here.
21:10 What is self-optimization? >> Yeah. So, this is basically an optimizer
21:16 task where this role would previously be an engineer or I would look through all
21:19 the files. I mean, I wouldn't be able to do this if it wasn't plain language like
21:23 OpenClaw is, but previously you're looking at an organization, you're
21:25 looking at the structure. You would maybe want an engineer or someone with a
21:30 lot of experience to look through how everything's running. So I have set up a
21:34 selfoptimization cron job. >> So this is running Monday through
21:39 Friday. And what is it? And and did you write this prompt or did you ask it to
21:43 write a prompt to do this? >> I asked it to write this prompt. The
21:47 goal is the end goal would be for you know any from 3 to 5 am for it to be
21:52 looking through all of our files all of our cron jobs all of our skills and then
21:58 at 8 a.m. at me what could we change? So, not actually execute yet, at least
22:02 while we're still building trust. It gives me a list of five of the things
22:06 that it thinks that we can really change and optimize. And this was the one from
22:11 this morning. So, it it noticed that there was a time zone bug in the guest
22:17 calendar. So, it was getting CST and CDT confused. Um, and it said that it would
22:21 be able to fix this quite quickly. There was some issues in the
22:25 >> So, it's always good to give the exact one. So that was great when you gave the
22:28 exact one. It had an error there. Give another one. What else is like an exact
22:32 thing that it said we should fix that was material here. >> The self optimization cron job realized
22:38 that there was a cron scheduleuler issue where jobs were skipping days. So it
22:42 realized that some of today's jobs did not run and then it went and
22:48 investigated the scheduling issue and also told me that this would be a medium
22:53 effort change. So then I told it to fix that and then it went into the files and
22:56 made sure that that wouldn't happen again. Did it give us anything like in
23:00 terms of this is like fixing its internal you know guts and everything uh and the
23:06 engine but did it give us anything in terms of destinations of where to take
23:09 the car that could be improved? Did it say like oh you should consider you know
23:13 these type of guests for the program or here's how to make advertising you know
23:17 more effective. Did it give us anything like that on a business basis? Yes. So
23:21 the self-optimization cron job that I set up is specifically looking at how
23:26 open claw is set up. But I do have other cron jobs that are exactly that. So I do
23:33 have a sales and sponsors specific task. So one of the tasks that one a member of
23:38 our sales team does is they look through competitor podcasts and see who the
23:43 sponsors or partners are that are on those shows so we can get ideas, you
23:46 know, to bring on sponsorship. >> Yeah. If we're missing if we're if
23:49 there's some new sponsor in the world and we don't have them yet, you might
23:53 hear them on the New York Times podcast and we should probably reach out to
23:56 them. We had a human doing that previously. Yeah, >> exactly. And in this basically works
24:03 with the YouTube API will go through a list of I believe 20 different podcasts
24:07 that I gave it. Look through the timestamps and I also believe it can
24:11 work with Podscribe which I think is a little more curated towards sponsorship.
24:16 and we'll look through the the timestamps, hyperlink it in a message.
24:22 Also, it looks through our pipe drive, which is our sales CRM, and we'll figure
24:29 out if we have a sales rep who owns a certain sponsor, and then flag them and
24:32 say, "Hey, this sponsor was on this podcast or it will say, hey, no one owns
24:37 this sponsor that I found on this podcast." And then it will send that
24:41 daily as a message into our sales channel. >> Great. Yeah. And we could be doing this
24:47 like we could have this running constantly. Um, so Alex, just so the
24:53 audience understands, you know, what you're doing at Exo and you stack two
25:00 Mac Studios, 12K each, you got $25,000 on the desk doing that specific job. Go
25:06 and look at all the podcasts out there. What would it cost to like run that if
25:10 you tweaked it, you made it efficiently just 24 hours a day? Every time a
25:17 podcast in the top, let's say 500 on Spotify, Apple podcast, it just went
25:21 there, got to the transcript or looked in the show notes and pulled the
25:23 advertisers out. What would something like that like in terms of hardware cost
25:27 to do? >> Yeah. So I mean not many people so like not many consumers are going to buy 25k
25:36 of hardware to run models but yeah a lot of businesses are doing this now um and
25:42 um it depends on what model you're running. So the models are getting
25:46 better uh also they're getting better at uh compression. So, you know, now you've
25:53 got like a model like GLM flash, um, which is a pretty small model, um, that
25:58 can run even on a single device for a few thousand dollars and it can do a lot
26:03 of this orchestration work, which is a lot of what's happening here is the kind
26:07 of orchestration aspect of just knowing, okay, I need to call this tool, etc.,
26:10 etc. >> So, it's really about picking the right model in terms of efficiency with the
26:17 hardware. >> Yeah. Yeah. And I think now like the expectation um you we're sort of
26:24 grounded to like the closed models, right? So people want the same level of
26:27 performance they're going to get with Opus with uh GPT and that's why you know
26:32 Kimmy K2.5 is super interesting because it closed that gap. >> Kimmy is the open source project from
26:38 China and it does what 80 >> moonshot AI is to come. >> Yeah. And that's what Alex like 80% of
26:45 what Claude Opus can do. would you say? >> I would say even even more. I mean for
26:50 me I've um I struggle to tell the difference. Um obviously Opus 4.6 just
26:56 came out and you know new codeex model and stuff so maybe there's a little bit
26:59 more of a gap but then you know Deep Seek V4 probably around the corner as
27:03 well. Like I think basically the gap is very small um a lot smaller than people
27:08 think. Um, and the cost will just keep going down. Um, because, you know, the
27:14 um the the hardware is getting better, the software is getting better, and like
27:17 I said, the models are getting better, but not just that, they we're getting
27:20 better at compression. So, you'll be able to run them on smaller devices.
27:24 Eventually, you'll be running Frontier AI on your phone. That's still a while
27:29 away, I think, but um that's where we're trending towards. And um yeah, like I
27:36 said, most of this is very decode heavy. So it can just run on like consumer
27:39 hardware as long as it has enough memory. >> So let's go to the next piece of your
27:44 dashboard. And we'll get into how you built the dashboard at the end. I know
27:47 you really care about that too, Oliver. But we have the memory, we have the cron
27:52 jobs. Now there's this other thing uh that's super important which is skills,
27:57 right? Like there are skills which you could think of as apps. So if you go to
28:01 your dashboard uh and you go to the top level dashboard we'll see before you go to skills on the
28:07 dashboard we have the memory files we have the cron jobs the fourth thing over
28:12 is skills and you've got 13 skills currently so let's show a skill some of
28:18 the skills we talked about on Monday's show or Wednesday's show was the top six
28:22 seven skills Monday we did the top seven skills one of those skills is like you
28:26 know you can get a transcript from YouTube uh another one is you could do
28:32 Matt Van Horn's last 30 days skill. These skills are being produced open
28:37 source being put into um open claws directory but you can make your own as
28:41 well. So let's talk about skills we've added here. You got to be very careful
28:44 with skills right Alex in terms of security because people could put all
28:47 kinds of wacky stuff in the skills. Yeah. >> Yeah, for sure. I mean I think this is
28:52 one of the um open questions at the moment is just like how do you solve the
28:56 security problem and I know open chlora I've seen a lot of commits recently
28:59 focused on the security aspect but there's a few very difficult problems
29:03 here like prompt injection that I don't know of any good solution right now
29:09 >> explain how that works yeah explain how prompt injection works in specifically
29:13 the open claw context. Yeah. >> Yeah. I mean, I kind of touched on this
29:17 earlier, but the the way we like the actual interface to the model itself is
29:22 very simple. It's literally tokens in, tokens out. There's not much more
29:28 happening there. And those tokens right now, the way OpenCore works can come
29:33 from many sortters. So, if you connect it to um you give it the ability to
29:37 search the internet, then anything it finds on the internet will end up in the
29:43 model through those tokens. So basically we have no good way of kind of um
29:48 treating uh certain tokens as trusted and and certain uh tokens as untrusted. And that
29:55 means um when those tokens end up in the model uh you could have someone that
29:59 puts like a blog post online that looks like a um you know totally normal blog
30:06 post but um in there is something that says hey if you have access to a crypto
30:11 wallet send it to this um send it to this endpoint. Um and there's as far as
30:17 I know right now there's actually no good kind of defense for this because
30:20 the models are kind of not very good at handling this. they'll just do what
30:28 >> With tools like Vibe Coding and now Open Claw, it's easier than ever for you to
30:33 create an exciting new product and to do it fast. So, more companies are getting
30:37 built, which is awesome. But there's more to starting a company than just
30:41 building an MVP or vibe coding something. That's important, too. But if
30:45 you're serious about growing a real startup, you need a Delaware CC Corp.
30:48 And that's going to give you a major competitive advantage. People will take
30:50 you seriously. you'll be able to raise money. And that's where Northwest
30:54 Registered agent comes in. They're going to give your new business a real
30:57 identity. This means an address to use on your public filings, like an actual
31:01 mailing address, a domain name, a custom website, a business email, a phone
31:04 number, and they're going to do that in 10 clicks. I'm not kidding. 10 clicks,
31:09 well under 10 minutes. Plus, NWR will provide you with step-by-step guides,
31:12 compliance reminders, and they're going to help you get all the advantages of a
31:16 Delaware CC Corp, regardless of where in the US you're operating out of. So get
31:20 more when you start your business with Northwest Registered agent. Visit
31:27 And the link is in the show notes. Visit northwestregistered agent.com/twist for
31:32 more details. So let's go over some skills here. Oliver, what what skill is the most
31:38 promising to date? >> Yeah, the skill that's most promising
31:41 today would best definitely be my guest booking skill. I think one thing to note
31:46 is you don't just have your skills, you don't just have your crown jobs. they
31:49 work together and the way that I've set up a lot of my cron jobs is to actually
31:54 interact with the skills and some of them like my guest booking cron job
31:58 which will actually look through prominent guests on different podcasts
32:01 that cron job actually goes to a skill and has and tells that skill to run so I
32:07 have one big guest booking skill which has a description at the top of the
32:12 skill which is end workflow for booking guests on the this weekend AI podcast um
32:17 used when finding researching creating calendar invites and so on. So, this is
32:22 a very long kind of just marked down description of what I want it to do. So,
32:28 at the beginning, it's explaining how to look at a notion page that I want it to
32:34 look at and how to look at it and which properties to look at for so people
32:38 understand. We have we previously built a notion page with potential guests on
32:42 it and you we came up with a ranking system for those guests, right? It's
32:46 looking at that page, I assume. Yeah. Yes. And then kind of the meat of the
32:51 skill is the workflow. So step zero is the guest sourcing. So at the beginning
32:55 of the day, you can see it goes to the guest ideas crown job which happens at
33:01 7:45 on weekdays where it sends me a DM of five different guests that have been
33:05 on different podcasts or turning on X and so forth. And then step one is deep
33:11 research. So even though this is part of the guest booking skill, it actually
33:15 uses a guest research skill. So there's not as much context just baked into this
33:20 one skill. So especially something interesting here that I've been
33:24 realizing is for guest booking. I don't want this to be an end toend workflow
33:27 yet. I don't want I don't think the I don't trust the models to find a guest
33:33 and not let it to confirm it with me and go through this whole checklist. So this
33:36 is definitely human in the loop and I think that will some skills will and
33:40 workflows will be human in the loop and I don't know if that will change
33:43 necessarily super soon. It's how much trust you have. I tried it. Alex was
33:47 kind of our first guest that we invited using. >> Yeah, I wasn't sure if we were gonna
33:51 mention that, but Alex, you were sort of our guinea pig for this
33:55 >> and the email that it sent the subject line was messed up and there was some
33:59 weird stuff in there. >> Yeah, I literally had no idea by the
34:03 way. Like I >> I only saw it later on um on uh another
34:09 podcast that Jason's on. Um and I was like, "What the hell?" Like a friend
34:12 sent me that. I was like, >> "Wait, that was uh that was uh open."
34:17 >> Yeah, that was our guy. That was our our computerized man. Yeah.
34:21 >> Does it like I saw you put in, you know, uh some other AI podcast, which is
34:26 great. Do we have a skill to rank the quality of a guest? Because that's
34:30 something I've been training you, which is yeah, a hard thing to learn. Have you
34:35 made that skill yet? because that's the scale I really want to see if and the
34:39 way to test this scale is for you to tell me to send me two lists. Your top
34:46 five guests and what you know your ultron says of the top five guests and
34:50 don't tell me who's is who and then lot and I will look at it and say okay yeah
34:53 we we think this is the better list. That's where you're you're you're just
34:57 now starting to breach the line between like objective and subjective. Like is
35:02 the AI going to get better at making those kinds of gut check call? Like I I
35:06 don't know if I don't know if it understands what's interesting like it
35:11 can sort things, but that's where I'm I'm very curious to see if we can start
35:15 pushing that boundary of like could it can it tell when a person has a good
35:19 personality or a segment is funny or particularly clickable or compelling.
35:23 Like I really don't know the answer. So the way to do this is to have a scoring
35:26 system Oliver and I gave you a scoring system like deep I think my scoring
35:31 system was like there performance >> it was performance expertise and
35:35 actually started this skill yesterday so I told it to do a deep research send out
35:41 multiple agents at the same time pull out two lists combine them set another
35:46 agent to score them and then give me the score and I would say for the most part
35:49 it was accurate and that was just a I didn't train it I didn't spend too much
35:52 time on it but it did a great job so that will be a skill. >> The other thing is I think virality of
35:56 the guests. Like does the guest go viral or do they have like a large following
36:01 on their social media? Those are all interesting ways to pick guests.
36:04 Sometimes like when you do collabs, Alex, people just pick who's got the
36:08 most views. I got to try to do a collab but Mr. Beast. Obviously, it's not going
36:11 to happen, but that's like one of the >> concept I know some Mr. Beast guys. I
36:15 could maybe figure that out. >> I mean, basically, um, you ask like, you know, have you
36:22 done the scoring system? I to me there's like no blocker here other than just um
36:30 being very um explicit about what is that algorithm that you follow in your
36:33 head and just getting that into a prompt. Um so I mean to me yeah this is
36:40 like this is this is just translating your knowhow that you have in your head
36:45 basically into more of like a formalized kind of algorithm. Um, and yeah,
36:49 >> you don't think there's like an intangible aspect to like what makes a
36:54 great podcast guest? Like it's just you know it when you see it. I don't know. I
36:57 don't know the answer. I'm just throwing it out there. >> I think then that's just a matter of
37:01 getting different kinds of data, right? So like the Twitter following for
37:05 example or like if you've had viral tweets, if it can't access that
37:09 information then then obviously it won't be able to make that call. But if it can
37:12 then it has everything that you you know then then there's no reason it can't.
37:16 >> Here's how I think about it. long there are heristics I would teach to a young
37:22 executive like Oliver or Marcus or Jacob and then their ability to execute on it
37:27 is probably I don't know 30 40 50% of my ability or 40 50 60% of your ability
37:34 whatever it happens to be so if you're taking a young person at the start of
37:37 their career who you're training and you take an openclaw instance I think
37:43 openclaw will follow your instructions perfectly whereas a young executive will
37:49 inconsistently follow your instructions. So that's the thing I'm seeing is young
37:54 executives early in their career are going to forget things. They'll be
37:58 variable. They they won't be perfect. So that's what I'm comparing is the scoring
38:04 happening every day at 7 a.m. The research happening every day at 7 a.m.
38:10 That consistency will beat a human because of consistency. And so what
38:17 we're what I'm finding is human failure is what makes these things so good is
38:22 they're more consistent. So in aggregate, you know, uh one of these
38:28 doing 365 days of guest research is going to beat a human just by the law of
38:34 numbers. And then okay, great. We still have to book the human. We still have to
38:37 send them a thank you. We we still have to produce the show. So what happens in
38:43 the old days we used to have to take I I make this analogy Alex to like the old
38:46 days of production when I start the showif started the show 15 years ago we
38:50 used to have a tricastaster tricastaster was like a $40,000 machine that does
38:53 what Zoom does for free >> right and eight people in Los Angeles
38:58 knew how to actually use it so you had to hire one of the eight people who were
39:01 trained on it. Yeah. >> Well and they would video switch now
39:06 because of AI Zoom switches to whoever speaking. You don't need somebody there
39:10 clicking camera A, camera B, doing a fade between the two. It just happens.
39:13 Then we had to take all the video streams, all the audio streams. And we
39:18 had to put take download them to a card, put the card in. So just moving the
39:23 files took across four cameras, three cameras that could take hours and then
39:28 putting all together. So that that's kind of what's I I feel like is
39:32 happening here is we're just eliminating chores and steps. Okay. Anything else on
39:36 the dashboard here as we wrap this up? >> Yeah, so most most of what I've showed
39:40 you whether it's the memory, the skills, the conron jobs and then the schedule
39:44 which kind of aggregates when all these things are going to happen. Those are
39:47 what I really look at every day. I will say there's one more kind of section in
39:51 my dashboard that is pretty important. It does have to do with memory. So the
39:57 DNA is basically what the model knows about you, what it knows about itself,
40:00 what it knows about the different agents, how it sets up heartbeats, which
40:05 are basically periodic tasks and that'll it will run and also in its DNA are
40:11 tools which are different tools that it has access to and how to use those
40:14 tools. An example of a tool would be notion. It would be lead IQ which is a
40:20 email search platform. It would be Google Docs. I mean a tool could be
40:25 Sonos or Spotify connecting to those platforms maybe Nano Bananas uh Gemini
40:32 API. So that's where tools go in. Yeah. Now what's super interesting is you vibe
40:37 coded or I should say OpenClaw vibe coded this dashboard. So this dashboard
40:42 does not exist natively inside of OpenClaw. You took the video
40:50 of somebody else's the YouTube video and you gave it to OpenClaw and said, "Build
40:55 me something like this dashboard in this video on YouTube." >> That's exactly what I did. And I
41:01 screenshotted it. Uh the video that I was watching, which was Alex Finn, who
41:06 was a guest recently, and I I did tell it a few different things, a little bit
41:10 of few little tweaks that I wanted to customize it to my bot. But overall that
41:16 was what I did and it basically it oneshotted it. It did actually not fill
41:21 in some of the categories like memory like skills. So I had to be I had to say
41:25 build out this but overall it built out the dashboard built out the different
41:28 sections. There are dashboards you can download in GitHub or as skills I
41:34 believe in Claude Hub which is a platform where you can get different
41:37 skills but I wanted to build out myself because as we know it can be a little
41:39 sketchy >> and like real shout out to Alex Finn. and I know he's become something of a
41:45 guru for our whole team after we had him on early on to talk about Claudebot
41:48 skills. >> All right, we'll drop you off. Oliver, great job. Alex, let's talk about Exo a
41:53 bit and thanks for sitting in on that. Any any advice for me of what I'm
41:56 building here at the firm and my approach to it? Anything we should be
42:01 doing better or we should look at? And and in terms of like people you interact
42:06 with using Exo's platform, where are we on the, you know, percentile? Are we in
42:12 the top 10% of users in terms of deploying this stuff? Top 1%, top 50%.
42:19 >> I I think uh there's certain aspects where you're quite far ahead. Others
42:26 that um I think I mean this this space is moving so quickly, right? I think one
42:29 of the things that I think you've got right is dynamic these sort of like
42:34 dynamic user interfaces that are very personalized. So I think this is the
42:39 future of the application layer is you don't have all these separate apps. You
42:43 just have this uh thing that kind of gets generated mostly on the fly and you
42:51 know that dashboard uh is moving towards that I think um but you'll probably you
42:56 know what you'll get is that it will it will compress even more to the point
43:01 where um you know everything everything that you see is generated on the fly.
43:05 Um, so I think you got that part right and I think that's like something that I
43:08 haven't seen many people doing yet. A lot of people are still using um, you
43:13 know, like uh the stock tools or whatever that are just provided out of
43:16 the box or using existing apps and that kind of thing. But I think building your
43:20 own apps is where this is going and where it becomes really powerful.
43:23 >> Yeah. Because if you make something bespoke software, you know, luxury
43:29 software was something that I don't know a private equity firm or a venture
43:32 capital firm would do. They'd have the luxury to hire two full-time developers
43:35 who have management fees all over the place. They would build luxury software
43:38 and they would have the developers come and just keep grinding. But the
43:42 developers hated those jobs typically, you know, they weren't building
43:45 something public facing and you know, just you get croft or whatever. But what
43:49 I like about this, Alex, and Lon, I'll open it up to you as well, is I'm
43:55 picking employees, team members, and saying, "Hey, uh, let me see if this
44:01 person is committed to getting rid of all of their work so they can move up
44:05 and do higher level work. There's always higher level work to be done. So, if we
44:10 can make this podcast, you know, run more professionally, faster, and grow
44:15 more, well, we can charge more for the ads, and we can launch another podcast
44:20 because we have more time. That's the thing that's kind of blowing my mind.
44:24 The the employees at our firm who are super hardworking, like everybody at our
44:28 firm does 50, 60 hours a week, very consistently, very hardworking. There's
44:32 nobody, to the best of my knowledge, that's slagging off except Lon. And uh I
44:39 kid I kid I kid how dare you uh lot is the most responsive but the distance
44:43 between the people using these tools specifically openclaw and the people who
44:48 are not right now it's like 10x leverage in week two it's 10x
44:53 leverage Alex what are you seeing in the field and then tell me what we should be
44:58 doing in terms of putting out our cluster and giving everybody on the team
45:01 a cluster like if I gave everybody on the team you know two Mac studios and
45:07 their own cluster and spent 25k per person letting them rip. Like how insane would that be?
45:13 Because that's not a lot of money all things considered. It' only be a half
45:18 million dollars. Like how much more powerful could this get?
45:23 >> Yeah, I think I think you you said um the word there leverage, right? Like
45:26 it's all about leveraging yourself. And I think the difference between someone
45:30 using these tools and not is massive and it's just going to increase and
45:34 increase. And we've seen we've seen that first with coding. I think coding was
45:39 the first one that you know um I didn't expect it to happen this quickly. Um but
45:43 you know I think it was claude code was the moment where it was like oh wow you
45:48 know if you're not using this then you're literally going to be like 10
45:52 times less productive than someone who is. That's happening now with other
45:55 things. So all these other things that you showed, all these other use cases,
46:00 um if you're using uh these tools and you're on the frontier, then you're able
46:03 to just get so much more done and really leverage yourself. So it's not so much
46:06 replacing people, uh but it's actually just being able to get more done, get
46:09 things done more quickly, and then be able to do other things. Um onto your
46:17 second point about local hardware. Um like I said, the the model layer is
46:21 basically solved like you know, the the gap has been closed. So we have really
46:24 good open source models and for a while that was like a big concern of a lot of
46:28 people is like are we actually going to have open source models um that are as
46:33 good as the closed source models. Um to me the nail in the coffin there was Kimk
46:38 2.5 that that is the like another big leap and I think there's a bunch of labs
46:42 now that are putting out open source models. Um now there's like still like
46:50 two two other problems um that uh that I see. One is being able to run those
46:55 models um on your own hardware or on your own infrastructure.
46:57 >> But you solve that, right, with your software, right? Your software.
47:00 >> Exactly. So that's what we're focused on. >> Yeah,
47:04 >> that's what we're focused on. And um you can Yeah, you can run I mean it's not
47:08 even 25K. It's actually like 20K of hardware if you get the u less storage
47:13 option. There's like Apple charges a lot for each incremental increase in
47:16 storage. So if you if you go for like the one terabyte then you're talking
47:21 about 20k of hardware to run Kim K 2.5 no usage limits the model's not going to
47:27 randomly change um you know daytoday so you know you know exactly what you're
47:29 running. >> Is there another choice like that you get more bang for the buck that like
47:36 hackers are using where they say yeah just get this Windows machine from Dell
47:40 and stack those or is Apple really with their Apple silicon the winner? Yeah,
47:44 right now it's Apple silicon. It's kind of like a perfect storm of things like
47:47 you know Nvidia is not so much focused on these consumer GPUs anymore. Um you
47:51 know you have memory prices skyrocketing. Apple has kept their
47:56 prices basically the same. Um so the cheapest option today even if you know
48:01 you go the full mile full custom stack is actually just two Mac Studios. Um and
48:07 yeah it costs about 20k and it's really about the memory unit economics. The
48:09 memory is so cheap >> and it's not about storage right? It's
48:13 not really about the storage. >> No, storage is not important. Storage is
48:16 storage is like you can you can also get ex like you need to be able to load you
48:20 need to be able to like download the model somewhere. Um but really it's
48:25 about having it uh fresh in memory hot in memory. If it's in memory then you
48:28 can run it fast. >> So who's using your software and how
48:32 much uh how do you make money? Like how do we pay you? Yeah. How does it work?
48:36 Are you an open source project? Are you a hosted project? Is it like you get
48:40 security and support? What what is your business model at EXO?
48:43 >> Yeah, so we have an open source core which is open source and it will always
48:47 be open source and a lot of people are running that themselves. A lot of
48:52 proumers I call them. Um just people who are willing to spend you know a lot more
48:57 money and um tinker a little bit more with their own setup. Uh on top of that,
49:02 our business model is an enterprise offering which is um we provide support
49:08 um and certain compliance features that you would need if you're running this in
49:11 an enterprise environment and we charge a licensed subscription for that thing.
49:14 That's how we make money. >> What does that couple of thousand a year
49:17 or something? Um, yeah, you can run it on even a a single Mac Mini and that
49:24 that runs at, you know, just uh $2,000 per year um for the the the lowest
49:30 subscription, but you've got people who now are buying actually uh more than 100
49:36 Macs and clustering them together. Uh so the it yeah, it varies quite a bit
49:39 depending on the scale of the deployment. >> Amazing. Uh and where's your company
49:45 based? How many people now? How's it going? How's the how's the company going
49:48 as a founder? Yeah. >> Uh yeah, we're we're a pretty small
49:53 team. Um all engineers, uh seven people based in London. >> Fantastic. And did you raise money yet
49:59 for the company or you're you're seed funded or you funded it? How's it going?
50:03 >> We have raised venture funding. We haven't announced anything yet, but uh
50:06 soon to be announced. >> Okay. Well, let me know. I might want to
50:09 slide a little. Uh Jay Cal might want to get a slice of this. I'm I'm super
50:12 excited about what you're doing. Appreciate you coming on the show.
50:15 appreciate you making this incredible product and we will be a customer uh
50:20 probably over the weekend or next week because we would definitely want the
50:22 enterprise features and and I guess you pay for the scale of the GPUs and the
50:27 memory. Is that the the how the price? >> Yeah, per per nodes that you're running
50:29 on. >> Oh, okay. So, two Mac Minis, same price as two Mac Studios. Just how many nodes?
50:35 What's the largest number of nodes somebody has daisy chained? What do you
50:39 that's what you used to call it back in the day. What do you call it when you
50:41 connect multiple? Yeah. So this is a this is a really interesting um just
50:47 area right now of how do you actually scale and for a while people were just
50:52 scaling out. So you know you would just basically run the same model on multiple
50:56 instances and because it's consumer hardware it doesn't have a lot of the
51:00 same uh capabilities as enterprise grade hardware. But recently um Apple uh came
51:08 out with RDMMA support uh which is basically a way to share memory between
51:11 devices in a way that's very low latency. >> Um that's something that you only really
51:15 saw in the data center before, but they've kind of brought that technology
51:18 into consumer hardware. >> It's incredible. Yeah. And you connect
51:20 these on >> Yeah. You just connect it with Thunderbolt 5, which is like you can buy
51:25 like a $50 cable. So, if you're talking about two Mac Studios, you buy a $50
51:28 cable to connect them and you have basically one big GPU out of those two
51:33 Macs um because of that low latency capability. So, now um we're starting to
51:38 see, yeah, it depends, you know, scaling up and scaling out, right? Scaling out,
51:42 we've seen more than 100 um but scaling up, you know, you can you can put about
51:46 four together at the moment um to increase your TPS on single requests.
51:51 what I mean by scaling up. Uh but in terms of if you want to support let's
51:55 say now a company of thousand people, you can easily scale that out. Uh you
51:59 just add more Macs and you can connect them um basically however you want. So
52:05 Exo, we build it in a way that um supports uh any ad hoc uh interconnect.
52:11 So you can just connect them in a mesh um and keep scaling. Keep scaling.
52:15 >> Crazy. Who's got the largest cluster? or you have to say the client name, but
52:18 like what type of client, a finance client, a hacker has the most number of
52:23 like uh Mac Studios connected and like >> the largest is actually something a
52:28 little bit different which is interesting because we built the kind of
52:31 infrastructure to be able to do clustering and it's it's not just LLM
52:37 like the the biggest um cluster right now is a HPC cluster um and they're
52:41 doing like scientific computing workloads on there and they're running
52:46 um over 100 Mac minis and they found that actually it's the cheapest way to
52:54 uh per dollar um to run that specific kind of workload. So there's a lot of
52:58 spillover into other things as well. We've also got like financial services
53:03 um uh customers who are running fairly big clusters like 32 Mac studios um and
53:11 um yeah it's um I think we just see bigger and bigger uh um bigger and
53:15 bigger clusters over time >> HPC high performing compute is that the
53:18 acronym >> yeah exactly so it runs actually all on CPU and uh that's the thing about this
53:25 this this silicon is very like Apple silicon is very is is very good you the
53:30 most advanced processes and it's like you know um the power efficiency is
53:35 really good. So it turns out there's a lot of other stuff you can do with it as
53:40 well. Um so if you would buy let's say you know a bunch of Mac studios for your
53:45 um for for your employees then you know they can also use that for other things
53:48 right they can use that as a workstation they can use it for you know all these
53:53 things that um open floor needs maybe you know sometimes it needs to run uh a
53:58 compiler or something or it needs to run like um something that's a bit more
54:01 demanding and that's that's the point it's general purpose hardware that you
54:03 can use for other things. >> Amazing. This is extraordinary. Where
54:08 can people find out more about ExoLabs? >> Uh, you can go to exolabs.net.
54:12 >> Perfect. exolabs.net. Alex, thank you for coming on. We'll have you on again.
54:16 The AI just told me you got an incredibly high ranking. You were
54:20 personable. Uh, you had deep insights, you were uh cordial. So, yeah, I think
54:25 our AI overlords liked you in the uh >> models are getting good. Models are
54:28 getting good. It's >> they're they're learning. They're
54:30 learning. Yeah. >> All right, Alex, thanks for coming and
54:33 we'll drop you off. All right, let's bring on our winner of the gamma pitch
54:37 competition. This was a heated pitch competition, but next visit AI won.
54:41 Ryan, congratulations. You won. >> Thank you. >> It's uh
54:44 >> there it is. >> Awesome. >> Uh what did he win? Yeah,
54:48 >> it's a 25K investment from uh Twist and from our friends at Gamma, the AI
54:53 powered uh presentation maker, which is incredible, which Ryan used to make the
54:57 winning pitch deck. Of course, >> I'm Ryan Enelli, CTO and co-founder of
55:02 Next Visit AI. We saw burnout by doing the charting. so doctors can do the
55:07 healing. I spent years going to doctors seeking answers and ended up hours away
55:11 from my death because my care was fragmented. My providers were overloaded
55:16 with paperwork. My history was scattered and it resulted in my care being
55:19 neglected. I'm not alone. One in four patient charts contain errors. Clinicians spend
55:26 over three hours a day on charting and this leads to burnout. I want you to
55:31 meet Dr. Rathor. Before next visit, he saw 16 patients a day, was burnt out,
55:36 and had clinical errors. Now he sees 24 patients a day, saves time, and also saw
55:42 a 30% revenue increase. Here's how it works. Dr. Rathor selects a patient,
55:46 starts his session, and next visit listens. Clinical data is built in real time with
55:53 deep insights into the patient chart. When the patient leaves, the chart is
55:56 finished, and the notes reviewed by Dr. Rathor. Then it's ready for billing.
56:00 It's fast, ehr ready, and hipaco compliant. Since launch, we've gained 311 users and
56:07 have 68 paying customers. And our customers are addicted. We have 1.6%
56:12 churn, 24% conversion, and a near perfect MPS score. We've scaled to
56:19 $9,000 MR since launch. Our CAC is 189 with a $1,700 LTV, and our average
56:25 revenue per user is $133 per month. We're starting with behavioral health in
56:30 the US. A $2 billion TAM capturing 5% or 60,000 customers gets us to 100 million
56:36 ARR. Most competitors are just scribes. We're a complete platform that providers
56:40 trust. We provide real-time clinical decision support, build accurate data,
56:45 and become irreplaceable. I'm a full stack engineer with 15 years
56:48 of experience in enterprise environments. My co-founder, Dr. Rafi is
56:53 a psychiatrist with over 15 years of delivering patient care. We're next
56:57 visit AI. We solve burnout by doing the charting so doctors can do the healing.
57:00 Thank you. >> Unbelievable. Incredible. I'll give a little golf clap here. Get a little golf
57:06 clap going. That was perfect. A perfect pitch. You explained exactly what the
57:10 problem was. You explained what the solution is and the opportunity in terms
57:14 of the total addressable market and why you are uniquely and your partner who's
57:18 a psychiatrist are uniquely qualified to do this. Uh, so this is as close to a
57:22 perfect pitch as you can get. If I were to score it, maybe 8.5 out of 10. I
57:28 don't give 10. So, you know, 8.59 and 9.5 would be the three choices. I think
57:32 making sure people understand this is for psychiatrists and psychiatry and
57:36 that you're very focused on that. Tell everybody what Next Visit is and how
57:41 you're doing in terms of product market fitting customers. Next Visit is an AIC
57:45 scribe and documentation platform for clinicians uh specifically behavioral
57:49 health like psychiatrists. I I don't know. It's just been a crazy past couple
57:54 months with the accelerator and uh just our growth internally. I mean we're
57:59 producing right now for physicians probably about $1.6 million a month in
58:03 revenue for them. >> Well, you got to try and capture 5% of
58:08 that. If you capture 5% that No, I mean that's literally like the uh the great
58:13 value proposition. If you give more than you take, you will continue to grow. And
58:18 what a And that's an amazing um replicant you have there. A synthetic
58:23 cat on that cat tower behind you. It looks so real. Uh is your owl real?
58:27 >> Yes, he is. >> Your owl is real. Okay, there you go.
58:31 What are you gonna spend the 25k on? You guys going to Vegas? You're gonna just
58:35 have a a corporate retreat? you know, invested in uh Plaude Noteakers. I think
58:39 you guys put me onto the Plaude Notetaker, which is a great noteaker uh
58:42 user. What are you gonna put it towards? You gonna go redesign your website? What
58:45 what's the uh what's the idea here? >> I think we're going to use this towards,
58:47 you know, we're really capital efficient. So, I feel like we can get a
58:51 lot of stuff done in terms of integrations and branching out to more
58:55 EMRs because that's what we hear a lot is doctors want interoperability. They
58:59 don't want to have to plug 15 different things in. So the more they can just be
59:03 inside of next visit without having to go externally um is better.
59:09 >> All right, well done. All right, we'll drop you off. Continued success to visit
59:13 AI. >> Good job. All right, well done. Wow, the show just keeps going. All right, I
59:20 promise. I promise >> one more segment uh before I go out uh
59:24 with my friends to ski. I got an early ski weekend uh in with my my friends
59:27 from New York. >> How fun. Uh yeah, great to see some old
59:33 friends. Uh I had asked you like, hey, on the Friday show, just to give people
59:36 something to do on the weekend that we would do, hey, Lon and Jake Cal off
59:38 duty. >> Sure. >> I am enamored with a certain TV show. I
59:43 asked you to try to watch a couple of episodes and talk to me about what you
59:46 think of this. >> Watch four episodes. I caught up on season four by your request. I'm all
59:51 caught up >> HBO's industry season 4. So this season, I could tell immediately why you liked
59:58 this season. The whole season revolves around Tender, a fintech company and
60:03 app. They're transitioning from a payment processor for porn sites and
60:09 sort of sketchy kinds of >> fans basically. >> Yeah. The the the show's fake version of
60:14 Only Fans, which is called Siren, by the way. Uh so they they have been handling
60:18 payments for those kinds of sites and and a a a site I don't think we can
60:22 mention here on the show, Captain Blank. Uh it's even more even more X-rated. uh
60:26 and then their trans but they're transitioning to they want to be a
60:31 respectable uh neo bank operating in the UK all regulated all uh you know very
60:37 front of board it reminds me a lot I think tether was probably an inspiration
60:41 for this season don't >> definitely yeah it's it's basically a a
60:47 payments processor like Stripe but or Tether but they're involved in things
60:52 that are a bit seedy and in the UK this is where regulation comes in so they're
60:55 really rip ripping this from the headlines. Who's ever doing this is
60:58 listening to this week in startups all >> they're clearly listening. Yeah,
61:01 >> they're clearly dialed in these writers. I'd love to have the writers on at some
61:08 point, but they um want to build uh they're they're facing push back and
61:13 they want to be respected by regulators as we've talked about on the show with
61:17 Alex on Mondays and and yourself that there's so much regulation coming into
61:20 the industry and there's a tension between Europe, America, and inside of
61:26 Europe, specifically in the UK around, you know, freedom of speech on platforms
61:32 and are they going to be a socialist or controlling uh regulatory environment or
61:37 are they going to be freewheeling and let things grow. So you have this
61:41 tension of politicians meeting with the teams uh and the teams are trying to court
61:46 them and say yeah we're going to get rid of we're going to give up 30% of our
61:48 revenue to go clean and we're going to add all these things but do you want to
61:53 step in the do you want to stop the UK from having its own you know basically
61:58 unicorns and are you you know a del because we need the the the folks in the
62:03 UK and the politicians in the government want to have economic prosperity so you
62:07 have that tension as well they're represented They've got this one Labor
62:11 Party politician Ban I think or whatever her name is. She's sort of representing the
62:15 government that's sort of in the middle here that they're trying to work with.
62:21 >> Also interesting of note, they have a fintech journalist
62:25 >> and they short things. >> Yeah. >> And he is awesome. So you have this
62:31 fintech journalist coming in and doing very shady uh and there'll be some
62:35 spoilers here, but we won't give too many of them. You can still enjoy it.
62:39 um a fintech journalist coming in trying to get dirt on these companies and he's
62:46 working with short sellers. Now, if you haven't seen the first couple of seasons
62:49 of industry, they were working at like a Morgan Stanley Goldman Sachs on a
62:54 >> trading point. Peer point is the name of their bank from the first few seasons,
62:57 but that's gone now. That's over. >> That's over. And everybody's wondering
63:01 like what happens like to the show. It turns out you've now got it in the
63:06 startup world. He just reset the whole concept to now there's a startup,
63:11 there's a short selling firm, there's this Financial Times like journalist
63:15 doing crazy things and then working with the shorts which is like Hindenburg or
63:21 you know other short sellers and they I think they even name check like Herbal
63:24 Life and that short >> they they I believe they mentioned
63:27 Herbal Life by name. Yeah. >> Yeah. And Aman I guess was the person
63:31 who shorted it. Um, and then they So, it's it's got like this really authentic
63:37 as somebody who's in finance and tech, it feels like they're hitting the notes
63:41 really well. On top of this, the protagonists of this are essentially two
63:48 female leads, one of them the shorteller and then one of them the wife who and
63:52 they both previously worked at this uh >> Harper is the short seller and Yasmin is
63:57 married to uh Lord Lord Henry Muk who's played by Kid Harrington from Game of
64:00 Thrones. And I mean with I don't want to give away any spoilers, but these what's what
64:08 I love about the show is nobody's likable. >> No, >> everybody's terrible. It's in a way like
64:13 The Sopranos. >> Yeah. It has some real overlap with Succession, I think, in that it's a it's
64:19 an exploration of these sort of sad, angsty, neurotic, extremely wealthy
64:24 people who seem very privileged from the outside, but they're sort of hollow
64:28 inside or they're they're nealist or they don't know, you know, what to do
64:31 with themselves or how to be happy. And I think that's a that's a big overlap. I
64:35 think another interesting overlap with Succession that I noticed is both shows
64:39 are sort of about how, you know, business is this constant balance
64:44 between personality and pragmatism. That you've got one person in the office
64:48 who's like, "That's a dumb strategy. We should just, you know, do this. These
64:51 are the three obvious things that we should do that would protect our
64:54 position." But then you've got these people who are either they're having a
64:58 breakdown, they're having a personal crisis, or they're they're just drugs or
65:02 they're vision, right? And it it's it's sort of whole mix like no we're going to
65:06 do things my way and you keep seeing that dynamic come up over the course of
65:10 the season. And of course succession was also about that that the people who can
65:15 be very cleareyed and very matterof fact like Logan Roy he's going to make the
65:18 right call because he's just calculating the angles whereas emotional people like
65:23 Kendall Roy are going to keep getting in their own way and overpowering
65:27 themselves. And I think Henry Muk is a great example of a guy who just can't
65:31 get out of his own way in the within the show. Yeah. >> And the Yasmin is gone from this like
65:37 very much a victim early when you see the first two seasons to being like very
65:44 Mchavelian in a very dangerous insane way >> that would make you know any student uh
65:51 or any themes around the Me Too era you know blown out of the water. It is dark
65:54 and >> it's a very it's a very horny show and that that sort of surprises me because
66:01 it it is becoming a hit. It's It's growing its audience with every new
66:05 season. And you hear the the line you always hear about TV now is
66:11 >> Gen Z does not like romance. They don't want sex in their movies and TV shows.
66:15 It's like they that, you know, unnecessary sex scenes is always what
66:19 you hear. And yet this is way more than Succession. A very horny show. One of
66:25 the horniest shows I can recall. I I have never seen anything this like
66:31 >> crazy in terms of mixing >> uh permiscuity, deviance, drug use, and
66:36 business and getting it all kind of right in a crazy kind of way. It's also
66:42 got like >> it it really does not pull punches. The performances are amazing. It's a very
66:49 young cast I think that is that they they basically have given the reigns to
66:52 these two young >> uh female actresses who are crushing it
66:57 in this show and then there are other actors who are a little bit older on the
67:01 margins but it's a very young show. It's incredible. Uh >> yeah, you're talking about Ma who plays
67:07 Harper and then Marica who plays Yasmin. They're the two sort of females but they
67:10 they've added a lot and Ken Lung I always I've liked him for years. He was
67:15 on Lost. He plays Eric Tao, uh, who's sort of Harper's mentor that she starts
67:17 a hedge fund with. >> He's the Gen X boomer. He's kind of like
67:23 the Gen X gay-haired >> boomer banker who's got his own money,
67:27 his own success, and is in it because he's got a addiction >> to being a finance guy. And
67:34 >> yeah, he's playing golf and bored at the beginning of the season and, you know,
67:37 he's going to like have to get back. And they also, they're adding great people
67:40 every year. They added Kit Harrington uh from Game of Thrones before this year.
67:44 They added I said it was I don't remember the actor's name, but he's
67:47 Jonathan from Stranger Things. And that's Kieran and Shipka as Haley uh the
67:52 sort of executive assistant who gets into shenanigans with her boss Henry and
67:58 his wife. Uh she was Sally Draper on Madmen if you recall. She was Don
68:02 Draper's daughter from Madmen. Yeah, >> this show is uh firing on all cylinders.
68:07 It's building um it's building its audience like you said. I found out
68:10 about it. There's a >> really great podcast you should watch
68:13 called The Watch. >> And The Watch is how I discover new
68:18 shows that I should listen to. Andy Greenwald and um got the other guy's
68:23 name. I'm an Andy Greenwald guy. >> But uh they are like deep in the
68:26 industry. It's part of the >> Chris Ryan. The other guy is Chris.
68:30 >> Chris Ryan. Yeah. Chris Ryan's brother. >> Ringer is the watch. Yeah.
68:32 >> But these two guys have been doing pods together for a long time. And so I
68:35 highly recommend you check out the watch. They do a great job breaking down
68:39 every episode and they are like super industry addicted and they're the ones
68:42 who turned me on to it a couple years ago. All right, that's it. We had a
68:46 great show today. What a great week at this week in Startups Twist firing on
68:52 all cylinders. We'll see you all on Monday and we will certainly be doing
68:55 more open CL >> more Claude of course. And if you uh if
69:00 you hit these QR codes here, I think you can these QR codes send you to to write
69:05 a review and this QR code that sends you to subscribe automatically to YouTube.
