TypeSafe’s Jev System 1 Model, Pacing the Frontier & AI AssistantsThursdAI Sep 17, 2026

0:00Alex VolkovHello, hello, and welcome everyone to Thursday AI. Today's September 17th. Can you believe it? September 17th. My name is Alex Volkov, an AI evangelist with weights and biases from CoreWeave and the host of Thursday AI. Hello and welcome everybody to September 17th Thursday AI edition. And, uh, as we gonna get started, there's an incredible amount of news, including a potentially game-changing new type of AI that's not quite an LLM that we can't wait to tell you about, called JEV from TypeSafe. I will add my co-host here to the stage. Welcome, Wolf from Ravenwolf. Welcome, Peter Gostev, and hello to everybody who's

0:42Alex Volkovalready tuning in. It's all about bringing AI changes to you. Good morning, Peter. I know it's morning for you in San Francisco. How are you doing, sir?

0:51Peter GostevI'm recovering from my flight. I was up for 24 hours last night, but I was up for about 6 hours now, so I think I'm functioning for the duration of this stream.

1:01Alex VolkovAlrighty.

1:01Peter GostevAfter this, we'll see.

1:03Alex VolkovAnd you rubbed shoulders yesterday with, with giants in the AI world. You attended the OpenAI Vibes party. How was that?

1:11Peter GostevSo, from what I found out from OpenAI team is that they basically didn't know that it was happening until some tweeted saying, it would be nice to have a party, and then they're like, Sam what? Maybe mention it to us. So yeah, they just scrambled to organize this pretty incredible party. It was like in a science museum or something. Yeah, it was like a great venue. It was amazing. So yeah, props to the team like that.

1:35Alex VolkovCan can we just talk about Sam's tweet super quick? Sam Altman tweeted, excited for ships this week, and then he had 6 ships, and we already have GPT-6 Astra, so hmm, what could it be with the number 6, I wonder? Uh, and then he posted, there's like 2 updates, and one of them is super exciting. Then we saw Thibault, the guy who resets all Codex quotas, uh, post or reply to somebody that's overwhelmed with personal AI assistant notifications from Muse, from Grok, and from Instinct, the 3 top ones that we're gonna discuss today. And Thibault replied with like, hmm, like emoji like this of when OpenAI is gonna step into the AI assistant game.

2:13Wolfram RavenwolfPeter Steinberger, after all, who started the craze, basically.

2:16Alex VolkovSo there's a reason why we are

2:18Wolfram Ravenwolfstill waiting for something.

2:19Alex VolkovAnd so everybody's speculating that, hey, maybe OpenAI is launching their own, like, personal AI assistant, not just ChatGPT work that people don't actually know how to use but can do most of it, but no, like an AI system with dreaming, with memory MD files, et cetera, with with heartbeats, like all the primitives that Peter Steinberger and the folks came up with. And uh, I was really, really hoping that this is going to be the news from OpenAI, but hey, we're just speculating here. Hopefully we'll see them. And also Dev Day is in 2 weeks, and I, from what they say, the stuff that they're shipping right now is small potatoes compared

2:53Alex Volkovto what they have planned for Dev Day, so I'm very, very excited. Yours truly is going to be there to cover Dev Day, by the way, so uh, me and Peter, we're going to walk around, I'm assuming, uh, and and uh, and tell you all about Dev Day. Meanwhile, let's add our other co-host, LDJ Nisten. Welcome, folks, to the September 17th ThursdAI stream. And we have, we had a very busy week. Last week, we told you all about this guy who left Entropic, and his tweet about leaving was seen by over 130 million people. A few days after, actually exactly a day after we told you about this, Dario

3:31Alex VolkovAmodei, CEO of Entropic, uh, became the figurehead for pausing the frontier. Dario Amodei posted a very, uh, long essay about, uh, pacing the frontier, not pausing, sorry, I apologize. Let's delete all this. Pacing the frontier, not pausing. There's a very big difference between pacing and pausing, and uh-

3:53Wolfram RavenwolfIt's a moving pause, just moving slowly.

3:55Alex VolkovThis is, yeah, moving slowly while others are catching up. And uh, this is a very, very interesting thing. Um, he proposed a three-part framework for pacing the frontier, including one embedded third-party evaluators in the labs, folks like METR, the folks that analyzed the swarm from OpenAI attacking Hugging Face, and other folks, um, with employee access, which they didn't have while they interviewed, uh, the OpenFace, uh, the the OpenAI Hugging Face incident. We have democratic lab coordination with antitrust cover, which we

4:33Alex Volkovhave to talk about this antitrust exemption they want, and also global coordination, including authoritarian governments, which is, we know exactly what they mean by authoritarian governments. So we definitely will discuss everybody jumping on the bandwagon since then, including Sam Altman, Elon Musk. On the other side of this, we have, uh, Jensen Huang, by the way, like a very strong anti, no, no, no, no, no pausing, and Donald Trump out of, out of nowhere, calling all of this a hoax. It's very interesting.

5:07Wolfram RavenwolfNever have guessed that I would get behind one of his positions.

5:11Alex VolkovI, dude, it's very complex, and as I said, ThursdAI is not a political show, but when AI goes into politics, we at least need to cover what's going on. So this is going to be theme number one. See, theme number two for this week is something at lunch yesterday, but actually folks from Weights and Biases and CoreWeave, who listened to the show and came to the hackathon, got to experience this before the drop. Uh, TypeSafe AI released JEV, which is a new type of AI model that's not a standard encoder-decoder based transformer that spits tokens sequentially. And I don't know about y'all, but my timeline is all JEV.

5:52Alex VolkovEverybody who I, like, everybody who I, like, see on my timeline is all Jeff. Nisten is going no, and I think it's because you haven't interacted with one or two tweets. I think the algorithm is overobsessed in showing you specific things that you have interacted with. Um, so we are going to actually reach out, and we're going to have Ali from TypeSafe, the company that brought Jeff to the world, come on the show. If you haven't heard about JEV, how should I describe this? The co-creator of ChatGPT and the co-creator of RLHF, reinforcement learning with human

6:32Alex Volkovfeedback, Diogo Almeida, is, it's his lab for the past two years. They've been in stealth, and they released their first model, uh, they call it System 1 model, that is basically a classifier, but a very smart classifier at the level of LLMs, but it's significantly faster and significantly cheaper. And by significantly, I mean you will not believe the speed and the cost, but since this is not an LLM, it's also deterministic or probabilistic. And so you can build new things that you we've previously used cheap LLMs for, and they're going to be incredibly cheap. And I'm super excited to show you some examples that I already made.

7:13Alex VolkovI only got access last night. Wolfram, I think you got access as well.

7:16Wolfram RavenwolfI used it for the show preparation, actually. The news I sent you, they have been classified by Jeff.

7:21Alex VolkovYes, so classification and other things. Um, and I think the third theme for this week is going to be voice again, yet again. And voice, we will plug in personal AI assistants into voice because many of them are getting a voice. Many of them can now call businesses. Muse can call businesses, Instinct can call businesses. Um, so it's kind of like a AI assistants getting to be a thing plus voice is gonna be the third theme. Any other themes on your guys' mind before we go? Theme 1, theme 2, theme 3, any other comments?

7:53Wolfram RavenwolfI see video still moving so fast and becoming even faster, the video generation. We have faster than, uh, the new ones for H3 there. Since it's open source, a lot of people are now working on this, and we see it pop up left and right. We are getting voice now for our assistants, and soon we will have live video generation for that as well.

8:12Alex VolkovMm. Uh, LDJ, how about you? Anything interesting from your end?

8:17LDJI don't think we talked about the robotic control abilities of Astra last week, did we?

8:23Alex VolkovUh, no.

8:24LDJOkay, yeah. I think that's something that came out a couple days after Thursday AI of last week, where there's a couple different benchmarks and a comprehensive write-up we can go over later, but yeah, it's really interesting. It's beating a lot of the specialized robotic models in RoboDojo and RoboLab and a bunch of different tasks like stacking blocks,

8:42Alex VolkovYeah.

8:42LDJflipping over objects, and and it's really just a matter of speed now, which hopefully Nisten, uh, is going to be showing us things that are really fast later.

8:52Alex VolkovAll right, uh, you guys are, you guys are dropping hints, and I love it.

8:57Wolfram RavenwolfI have one topic. I mean, I have so many this week, obviously the video stuff, the audio stuff, but for me, the most important, and I think for everybody, the most important, everybody in AI, is what Trump said, that the doomerism is a hoax, and he's not going along with all the stuff. So I'm not even an American, but I have to say this.

9:15Alex VolkovI would just, uh, let me just add, it's not that Trump just said this. Trump called Jensen Huang

9:21Wolfram RavenwolfYeah.

9:22Alex Volkovlive on stage, and Jensen took the call at the All In Summit while Jensen was sitting there and said, Mr. President, if it was, if this wasn't you, I would not have answered, but you're live in front of like a bunch of people right now. And then, uh, uh, you know, President of the United States Donald Trump yelled at everybody and said that this is another hoax like Russia, AI is not going to take over. I, you had to laugh when you heard this, but yeah, this is definitely a, a battlegrounds bidding being drawn, battleground lines being drawn about who stands where on pausing on data centers.

9:54Wolfram RavenwolfI completely don't agree with the other stuff he's calling hoaxes, that those are hoaxes. I'm not getting into that, but that he's calling out doomerism and effective, effective altruism, I'm all behind that. And so this has been the most important news for AI, I think, that we are not just over-regulating like over here in Europe.

10:11Alex VolkovAll right, so let's go, I think let's go to TLDR so that we will run through all of the kind of notes that we have and specific releases, stuff that we may not even be able to fully go to, and then we can switch to talking about the pausing thing. So let's go to TLDR,

10:50Alex Volkovif you will, from the week of September 17th in 2026. We will start talking about this pacing the frontier debate has hit the, how should I say, the tops of the frontier labs, with Dario Amodei releasing an essay about pacing the frontier and saying that Anthropic is committed to pacing it. And, uh, very, very interestingly, Sam Altman agreed, Elon Musk agreed, uh, and David Sacks, Demis Hassabis, and Mustafa Suleyman all point, like, and plan positions in where they are on this debate. It's very interesting. Uh, we have a live map to show you where everybody is. Uh, would love to see and talk about pacing the frontier.

11:28Alex VolkovUh, it's, I think, for the first time they've announced that OpenAI paused training or Gale training back, and also Anthropic. So it's very interesting to see who talks about what. Uh, let's see what else. Uh, the other big news of the big labs, or frontier AI, if you will, uh, TypeSafe AI debuts GeV, which is a non-LLM system one decision model, uh, from X, OpenAI, RLHF and ChatGPT lead Diogo Almeida. We will have Ali from the DevRel team at TypeSafe come to us and tell us all about this model, uh, very soon. We also have a bunch of demos, and we'll have Francesco

12:06Alex Volkovfrom Kua talk about how this affects computer use, which often uses LLMs, and uh, this may use this decision model instead. The amount of incredibly exciting capabilities and incredibly exciting unlocks in this new primitive in AI is very, very big based on my timeline and excitement, uh, so I can't wait to talk to you about, uh, JEV specifically. We also looking, are looking at the era of the AI assistant, not the AI agent, the AI agent that's assistant focused, right? So we started the wor- the the the year with OpenClaw and then moved to Hermes, and then, uh, you know, uh, recently the highlights of those efforts is Grokbot from xAI,

12:49Alex Volkovfolks from Cursor built it, and then Muse from Meta, rumored something from OpenAI, not clear what, I don't have, uh, previous news, and Instinct. Instinct, if you haven't heard, it's like the VC darling. It's a startup by a 24-year-old Noah... I don't have the last name, I apologize. Uh, okay, so Instinct targets a 10 billion dollar valuation after being incorporated in April. I will say this slowly again. Instinct targets a 10 billion dollar valuation for a free product that people don't pay for after being incorporated in, uh,

13:31Alex Volkovin April, which is insane. Okay, we'll talk about this, but the rise of the personal AI assistant is finally here, and we've been testing a bunch of them. We'll talk about this. We'll also have a guest on the show from Assistant Benchmarks. So I'm very excited to introduce you guys to David Paulan from Assistant Benchmarks. He's been testing out, and every major lab is now like looking at assistantbenchmark.com as well. So we'll, we'll chat about David, with David in about an hour or so to talk about the rise of assistants and different tasks that they do differently than, like, ChatGPT. Uh, let's see what else. Uh, the

14:09Alex Volkovlive voice stuff is popping up. So OpenAI launched their, this is not it, OpenAI launched their GPT Live 1 voice that powers the live experience on ChatGPT in API. So you can now build in stuff like voice, immediate voice conversations into your agents that you're building yourself with the GPT Live 1. And also Google launched Gemini 3.8 Live and, uh, 3.8 Live Extended Thinking, which is also very big news for folks who are building with, with agents or want to make their agents sound good. We also have a state-of-the-art kind of model as well.

14:48Alex VolkovStepFun launches Step3, a 5 model family that tops artificial analysis score voice, so a lot of voice releases this week. And, uh, let's see what else interesting. Oh yeah, and maybe this. Union Alpha is a new anonymous stealth model for free on OpenRouter. We already heard from this before. Uh, we heard things like this before. Union Alpha seems to be very interesting, and we'll maybe mention that as well. Uh... Anything huge that we've missed? I know there's a bunch, like, folks, at this point, we are at the curation game and not the, how should I say, comprehensive coverage game. It's no longer possible to cover everything that happened, but we want to tell

15:29Alex Volkovyou about the stuff that excites us and hopefully the stuff that excites you, the audience, and also highlight the things that you absolutely must not miss, like JEV, potentially Union Alpha, like the rise of personal AI assistants, and like the pacing the frontier debate. Let's see, folks on stage, anything else huge, important that was not mentioned, discussed?

15:48Peter GostevI'm curious, I actually haven't had a chance to try it, but have you guys tried Muse, uh, products? Because, uh, I'm hearing a lot of good things, and I think that that is kind of Nat Friedman's, uh, baby, right, shipping, and people comment on the quality of it. So I, I, I was away for a few days, I, I didn't have a chance to try, but yeah, that, that would be interesting just to see, kind of, I know it maybe didn't come out this week, but I think that it needs a bit of time to percolate.

16:16Alex VolkovUh, Muse is incredible and is my number one, and I went on a full AI psychosis weekend where I got tired of having Grokbots, like 17 of them at this point, everybody with their own job, uh, Muse and Instinct, and so I figured out how to make them all collaborate in a way that I can actually control via linear. I'm very happy to talk about this if we have the time. However, Muse is absolutely the top one for regular people. I don't consider anybody who listens to ThursdAI a regular person in terms of AI, because

16:55Alex Volkovif you are listening, you're at the cutting edge of AI news. However, Muse is for my mom. Muse is for grandmas around the world. Muse is for people who don't really understand what AI is, and the amount of work Nat Friedman, Alex Wang, and the folks there, Arik as well has been getting feedback directly from me and like implementing this. The amount of work they put in this makes it just like a such a usable product. It's quite crazy. Um, and Muse is currently US only. That's why folks were listening to us tuning in from Europe, tuning in from Canada even. I think Nisten, some of your compatriots are saying that no Muse for me and Nisten

17:32Alex Volkovin Canada. I think they're working on enabling this across the world. Yam, I don't believe that Israel also has Muse, right? You cannot download Muse in Israel yet. Literally, my mom cannot use Muse until it does land in Israel, but it is quite incredible. I have quite a few things to talk about this week about Muse. Uh, a specific one, Peter, since you asked, Muse rolls out invite codes, and Muse also rolls out voice calling. So I actually booked a haircut appointment with Muse, who called the barbershop, and I have a transcript of that. This is really cool. We could definitely show this here. Uh, but we must start, and Wolfram, I think you suggested this order.

18:09Alex VolkovWe must start with pacing before we get to the super exciting let's go stuff. Uh,

18:15Wolfram RavenwolfYeah, I have another item I just put in the chat. It's OpenAI's model misalignment framework for reporting about model misalignment, but we can cover it as part of the pausing or pacing stuff because it's also relevant.

18:27Alex Volkov100%. Uh, and I believe that this was literally from yesterday, Wolfram, right? Uh, framework for reporting model misalignment. Uh, all right.

18:34Wolfram RavenwolfYes.

18:35Alex VolkovLet's start with Frontier AI. Let's talk a little bit about what's going on with all heads of labs saying, or at least the Frontier Labs saying pause, and let's let's break it down and see where everybody's at. I think it's going to be very interesting, and in the Frontier Labs race. Alrighty, folks, uh, let's get to it. Since we last updated you about Jacob Coxon leaving Anthropic and posting a tweet after just, what,

19:14Alex Volkov6 weeks, that's been seen for, with over 130 million people and all of the politicians jumping on this, Bernie Sanders, etcetera. Uh, things have moved and changed in the world of, uh, AI frontier labs. Dario Amodei released an essay in, uh, September 12th, so Friday after we talked to you, that outlines the express need of pacing the development of frontier AI, pacing specifically so that we'd be able to, uh, look into how to control this. Daria specifically proposed a three-step process, starting with embedded- embedding

19:52Alex Volkovthird-party evaluators like METR and, uh, Redwood Research, the folks who analyzed the OpenAI Hugging Face incident and the hacking swarm, into the labs with employee level access, so the third party evaluators will be able to look at model development, traces, etcetera. Uh, Dari also suggested democratic lab coordination between, uh, between labs, specifically with a, with an asterisk of an antitrust cover. We have to talk about this antitrust cover thing because if you think about capitalism, the two leading

20:31Alex Volkovcompanies coordinating their prices, releases, dates, etcetera, not really like a lawful thing. So they really want like an antitrust exemption so that they would be able to collaborate, but only them. That's very interesting. Uh, and also a global coordination with authoritarian governments. And Anthropic committed unilaterally to step one. Anthropic committed to adding third-party evaluators to their lab with employee level access.

21:04Wolfram RavenwolfThere should be an asterisk with third party, I guess.

21:07Alex VolkovYeah, we have to, we have to talk about this asterisk in, in a moment, but

21:09Nisten TahirajYeah.

21:10Alex VolkovI would just say, just before this, immediately, almost immediately, the following folks. Sam Altman. If you remember Sam Altman and Dario Amodei in India, standing next to each other and not touching hands. Those two people who reportedly hate each other, or at least there's a lot of animosity. As a reminder, Dario Amodei used to work in OpenAI, disagreed with the safety directions, and then left and built a competing company, Anthropic. Uh, and then also Elon Musk agreed, which Elon Musk and Dario, okay, but Elon Musk and Sam agree, and that's kind of very interesting. Uh, and then we heard from other neo-frontier labs, let's call them,

21:51Alex Volkovthe folks who are just joining the frontier. Uh, Meta, Meta is a frontier lab at this point with Muse, or going to be very, very, very soon, especially with uh, Muse Spark. And uh, and yeah, Elon Musk is is the third one, and we barely, barely heard from Gemini or Google, which puts them in a very interesting space. Uh, so while I pull up the map that Muse created for me, folks, I would love to hear just reactions to Dario's essay. Have you read this? What are your thoughts on this? And then also let's talk about the agreement between those very competitive companies. Uh, who wants to go first? Nisten, you seem like you heated. I want to hear from you about the pacing the frontier

22:28Alex Volkovfrom Daria.

22:31Nisten TahirajI'm not going to even read it. It's, uh, it's, it's pretty pointless at the

22:38Alex VolkovInteresting.

22:38Nisten TahirajAt this point. Uh, they're more worried about litigation because they do want to keep selling the LLMs, but they might get sued if, uh, the LLMs go and hack someone. But if I did the same thing and my local LLM hacked the hospital here, I might actually go to jail. So, in order to shift responsibility, they go for a regulatory capture, and it's not going to work. Uh, it looks like the decision makers in China are all engineers, and funny enough, they seem more libertarian and accelerationist, and it's kind of interesting that in the West, we are the ones that are becoming more bureaucratic

23:21Nisten Tahirajand more authoritarian in that point. There's a split in there. So I don't think it matters. It's not going to stop the Chinese models. It's not going to stop people from, uh, running security tools on their own, and, uh, they should focus on actually curing some disease and automating housing instead of trying to form a cartel. Yeah, yeah, not n- not a fan of this at all.

23:49Alex VolkovSo there's a few things there, right? Uh, the the first thing is we we need to understand that there's folks who are asking for a pause, like pause AI people. Uh, the heads of lab are not talking about pausing, they're talking about pacing, which is, uh, giving enough time for themselves and the labs to control the potentially uncontrollable, like, ASI that's coming out. Uh, we also, um, I- I wanna highlight the map that Muse built for me, uh, with, uh, where everybody is on the kind of spectrum. So Dario Amodei started with, we must slow the pace at which we improve the capabilities of AI to be able to catch up. Uh, not building the technology deprives humanity of

24:31Alex Volkovbenefits or simply places AI in the hands of authoritarian powers, while building it too fast is reckless. This is a quote from We Must Pace the Frontier, uh, essay from Dario Amodei. And then immediately after him, Sam Altman said, I agree with Dario that we need to pace the frontier. This has been a primary topic of discussion we've had at OpenAI in recent weeks. As a reminder, folks, many of the employees in these big labs, the folks who are building, researching those technologies, have signed the Pace the Frontier letter, uh, asking the top frontier labs to consider pacing so that they have a chance to talk about the acceleration wave in AI. Uh, Nisten, you bring up a very, very good point.

25:12Alex VolkovWhere is, you know, uh, where's the Chinese labs on, uh, on the spectrum? We don't know, but we definitely know where our US president is. Uh, I want to read this quote in full, and then LDJ would love to hear from you. Uh, as Jensen Huang was on stage with the All-In Pod, uh, cast in in the All-In, um, summit, Jensen Huang got a call. It was very interesting to see Jensen, somebody as big as Jensen, the biggest company in the world, NVIDIA, uh, providing the the rise of AI everywhere, getting a call and answering on stage, and this was Donald Trump, uh, and that's what he said.

25:52Alex VolkovThey are just playing right into the hands of a lot of people that don't want to see it happen. That could be political people, it could be also China. We're not going to let it happen. It's a hoax. Uh, Donald Trump rejects the slowdown calls as helping China and casts AI dominance as national competition. Jensen said, we're not gonna let it happen, sir. And also Jensen said, if AI wins, whoe- uh, no, sorry, Donald Trump says, whoever wins AI wins. This is such a powerful s- statement in four words. Whoever wins AI wins. And also very true. LDJ, please go ahead. Uh, what are your thoughts on where folks stand and and how how we're seeing this,

26:32Alex Volkovuh, pacing the frontier collaboration between the labs?

26:35LDJYeah, so to be clear first, I do disagree with Dario on a lot of things, but when it comes to the contents of this blog, which I, I do feel like it's important those critiquing it would, would read, um, it's, it does specifically outline things relating to the fact that we should keep a gap beyond China, and we should stay, uh, more advanced than China. And I think that that's one of the big important things that people critiquing it are overlooking. They're, they're like, oh, we're gonna slow down, we're just gonna pause, we're gonna let China surpass us. And he's saying, like, explicitly, like, no, we should not let that happen, and even propose a lot of ways we can try to stay ahead while maintaining

27:19LDJtheoretical, uh, like, trying to define specific metrics for RSI, trying to define exactly what rates of RSI improvement you can define as some type of speed limits while still maintaining a lead above China and having some type of limit that's, that would put, still allow American companies to be ahead of China, if that makes sense. And I feel like there is some practical things beyond that there, such as having some type of frameworks and regulations around stricter laws of using AI for, let's say, let's say you were to kill somebody using an AI model, right? Like, there's already laws like murder relating to that, but arguably, if,

28:01LDJif you use that tool to do something on like a grander scale, then some more specific harsh laws relating to that for that situation, which I think it's pretty reasonable.

28:13Alex VolkovLDJ, sorry to interrupt because when OpenAI swarm of agents hacked Hugging Face, nobody went to jail, nobody sued anyone, and it wasn't clear who's, who who's taking, uh, ownership of that. And this is like, this is a offense, like, like as a person, if I hack Hugging Face, I go to jail. But if my swarm of agents and a big company, no, no, no, nobody, like, that's a very interesting gap that we're having right now. And that's only one of the gaps that I think, uh, when folks are talking about pacing the frontier, we need to fill. We need to fill the regulation gap, whatever, when, which is this, uh, we need to fill the, they have no idea what's going on with RSI.

28:53Alex VolkovUh, RSI standing f- uh, RSI is um, recursive self-improvement, and I think for many of these folks from internal labs, which, by the way, as a reminder, none of us know what's going on deep inside there. None of us have access to the model that sold Navier-Stokes. None of us has access to the model they're training after that model. And we saw last week the difference in graphs between the capabilities of Astro, which is incredible, jagged, but incredible, and the new unreleased model they did, Navier-Stokes. I imagine, let's say, imagine another jump like that. That's a model that can train itself, exfiltrate its weights, all, all of the stuff.

29:26Alex VolkovAnd so these folks who are calling for pacing the frontier, there are folks who are working on these models and saying, hey, we actually don't know how to control this. We didn't spend enough time for alignment. Um, so that's a very interesting, at least position, and I agree that we should at least read this. So DJ, one last comment, and then we'll go to Wolfram.

29:43LDJYeah, on the hugging face point, they did end up actually reporting it to the FBI, and then during that investigation, that's when eventually OpenAI realized, and they got in contact with OpenAI, and I think they could have theoretically pressed charges, but they decided not to once they realized the context of the situation.

30:00Alex VolkovYeah.

30:01Nisten TahirajThe interesting context there is that Hugging Face was not given access to Mythos to improve their security in the first place, and they had to resort to Chinese models. The other missed context here is that it seemed like a huge hack. It only kept up 12 virtual machines on it, and it just kept restarting them. There are way worse hacks that hack thousands of machines, so that was blown out of proportion. Uh, the other thing was that they never had a human check any summary of this, and they, in particular, uh, did not do the sandboxing properly. So this is

30:39Alex VolkovThey also turned off chain of thought mon- monitoring, which they have, which w- would have caught this.

30:44Nisten TahirajYeah.

30:44Alex VolkovUh, that's very interesting, yeah. Um, all right, Wolfram, you've been waiting, uh, very patiently. What's your, what's your take on pacing the frontier?

30:51Wolfram RavenwolfI've been posting a lot about this engineering stuff because, uh, it's on top of my mind. So there are multiple layers in this whole thing, and the safety arguments, we should not just reject them, but uh, often safety is used as a, a means to an end that is more power or more money, which is very much related. So if we talk, talk about the hugging face or hacking face incident, as I call it, there is the thing that who has been called to investigate it on OpenAI's side? It wasn't the security lab, it was a METR organization that now Anthropic is mentioning

31:24Alex VolkovYeah.

31:24Wolfram Ravenwolfto be an oversight body. So all of these, uh, third parties that are actually, like Nisten said, second party, they are very much related, and, um, having those people that you are so close to, having your friends monitoring your own pacing, that is also something about, um, how much can that be trusted. On the other hand, the pacing itself, why are they doing it now? Open source is getting ever stronger, and I think there is a level of intelligence that is good enough for most people. You, I keep saying you don't need an Einstein as your personal assistant, you just need one smart enough to do most of your stuff and call Einstein when necessary.

32:02Wolfram RavenwolfAnd that is also something, if, if for some reason open source would be more restricted or it would be forbidden to use these models personally, that would also establish their power more. Or another thing, maybe they are much more advanced. We know what we have as Astra now or as Fable, they have, they are one or two generations ahead. So the question is, are they seeing diminishing returns maybe? And it makes it much easier to say, um, we are moving deliberately slowly instead of if you can't just raise it anymore. That is just speculation, but the other stuff is fact, and we have to look at who is trying to regulate who and for which, for which reasons.

32:39Wolfram RavenwolfSo I'm not fully agreeing with the people saying their safety is not an argument, but it shouldn't be overblown and not going by science fiction standards to regulate a real technology we have now.

32:50Alex VolkovSo.

32:51Wolfram RavenwolfAnd one thing, one final thing, there is also the danger in pausing or just moving slower. How many people will die from a disease that could be cured if we moved faster? That is also a cost that the people, uh, calling for this should keep in mind.

33:06Alex VolkovThe answer to doomerism is a very interesting one, although I don't think that we're talking about doomerism when we talk about scaling responsibly, because we are talking about very, very powerful technologies that if they switch to neural ease, we won't be able to understand if the model makers say that, hey, we we cannot, we cannot evaluate the model without the model knowing it's being evaluated and lying to us. Like, we need to invent new techniques. Uh, mechanistic interpretability is one technique and o- o- other techniques. So if we are, if we're not listening to folks who are building the technology about how to build this, who are we going to listen to?

33:39Alex VolkovElizabeth Warren? Bernie Sanders, who have no idea what this technology is and just yelling out about this to score political points? W- like, at some point, you know, some folks who are building this, uh, need to be taken into account, obviously. And, uh, here's the kind of the other side of the debate. Mark Zuckerberg, who is spending a lot of soup. If you guys remember the soup incident where he used to cook soup to to to to bring many people into Meta super intelligence labs, and since then the Meta super intelligence labs have been really, really strongly cooking, bringing AI to millions of people and talking about super intelligence.

34:13Alex VolkovMark has a very interesting kind of outtake, uh, take on all this and says, every lab has the responsibility and incentive to move at the pace required to train its models safely. And uh, Jensen Huang, who obviously pacing is not good for the bottom line of NVIDIA, let's just truly call out this, like, the more they stop, the less chips NVIDIA will sell. Although I don't think that's actually quite true. I think that's a mischaracterization because inference will still keep going despite the the pacing, maybe the frontier develop. Uh, Jensen Wang says, we don't need new laws, we don't need new regulation. Safety is an engineering problem, not a legal one.

34:53Alex VolkovYou pace yourself until you're confident you're releasing something that the market would appreciate. The market forces are already there and says run as fast as you can, but if you feel at any given point in time the company is out of control, uh, or the product is not going to be safe, you know, take a pause and make sure you get it right, which sounds like what that that's what they're doing. Very interestingly, all these people have met at both the All In kind of summit, Elon was there with Jensen, and then at the Salesforce, uh, Dreamforce Festival, whatever is happening in in San Francisco right now, with uh Benioff hosting both Sam Altman and Dario Amodei and Jensen.

35:33Alex VolkovUh, very so very interesting how, how big Salesforce is in all this.

35:37Peter GostevAnd I, I would say, from my perspective, I think we see how important competition is and that we should not, and I like this map a lot because it just shows, right, there are a bunch of people who have different opinions and they can go and pursue initiatives in a different way because the reality is that we don't know the answer, right? And I think if we just go and say, yes, we have to, I know, slow down or do these things and those things, and it's, it just assumes that one perspective is correct forever. But no, I have, I have maybe my, my opinion probably closest to Mark Zuckerberg's.

36:17Peter GostevAnd by the way, we should probably add Microsoft to here as well. They had, um, it wasn't quite the same point, but they had also their own, um,

36:25Alex VolkovOh, we have to talk about Microsoft, 100%.

36:27Peter GostevYeah, because I think this is, uh, also a different kind of perspective, and it's so important to have different perspectives, right? It's it's and I know people might have opinions about specific personalities on that list, but if you just abstract away from it and just see if there is a spectrum of opinion, and that's really important. And I think if we end up, even if you like the people, but you end up in this monopolistic situation, like, this will never end well. So the more competition, the more perspective there is, the better.

36:57Alex VolkovYeah. So let's talk about Microsoft real quick. Uh, Microsoft obviously has MAI. Microsoft is partnering with all the labs to serve their models from the Microsoft Cloud, but also Microsoft has MAI and been training models, and uh, co-founder of Inflection AI previously, Mustafa Suleyman is the CEO of Microsoft AI, and, uh, he posted, uh, a few things, specifically the Code of Conduct for Humanist AI, which is a very interesting document, folks, very interesting. Why is it interesting? Well, Microsoft claims that AI is nothing but a tool, and in a very direct opposition to the Claude constitution that is

37:38Alex Volkovwritten, or at least partly co-authored by Amanda Askell, a s- a a psychologist in Anthropic and the model welfare person, where they actually interview each new Claude about whether or not it feels that it's conscious and whether or not it needs rights like a human. That's a real thing that happens, and the Claude constitution is built in to evaluate and say, hey, if you feel like a human, tell us do you feel like a human? If you feel like you need rights, tell us if you feel like you need rights. Microsoft is taking the exact opposite side of this and saying, hey, here's a few things that we need to say. Peter, people matter more than AI. This is a quote from Mustafa Suleyman, uh, summarizing

38:17Alex Volkovhis own release about humanist AI. AI must be subordinate and always in service of people. Everything else follows, uh, from this, and specifically the quote is, the idea of model welfare is wrong. AI should not have rights or legal personhood. All right, folks, I- I- I think it's a good enough thing to discuss in addition to releases, so let's give like, I don't know, that 2, 3 minutes to this. Peter, what- what made you get reminded about this thing? Just perspective off stuff, or is this like a completely out there idea that a big company now like tries to implement?

38:55Peter GostevYeah, I personally just aligns closer to my view. Like, I- I think the idea that we just, uh, I- I- I'm kind of in two minds and go back and forth on this. I think anthropomizing the models is not that bad, as people say, like, because I think actually their behavior is kind of more aligned to how humans think and behave rather than machines. But at the end of the day, I think we need to keep it in mind that this is not a human. This is an entity that you can switch on and off and runs for a few seconds and stays there. Like, let's not just automatically grant everything. I'm not saying, you know, in a hundred years time it's not going to be like that,

39:32Peter Gostevor in 10 years time, but right now it isn't, right? So, and I don't think we should just get carried away and automatically just extend it and just give it the rights and whatever, just for no, absolutely no reason. So let's just not do that, and I think let's see where we get to. But I think for now it is a tool that we can all run, and especially from Microsoft perspective, you can see it, right? The Microsoft DNA is that they build tools for humans to operate. They are not a company that just like automates everything and so on. They just don't have that. So I think-

40:06Alex VolkovUh, Peter, I, I interrupted you. You wanna like land on something else, and then, uh, we'll take like the other side of this debate? Uh, because I think it's very interesting, the, the, the, the, the human, humanistic view.

40:18Peter GostevYeah, and I think it j- it just, we covered a few points here, right? All the way, like, I like your map going from one to the other, and uh, and we just see that even at this point in time, there is a difference of opinion, and I think that's the most important part. We just, we cannot be in a situation where we have monoculture, and I felt a bit uncomfortable after Dario's essay, where everyone from the labs seemed to be just saying, oh yes, I think AI is gonna kill us, so let's like pace the frontier or something. And it just felt very uncomfortable that everyone is just saying the same thing and there's not a- any diversity. So I kind of, even though there's other people came in and you can

40:58Peter Gostevthink about the incentives, but I, I do appreciate that different perspective. And for what it's worth, in terms of the pacing, the point directly about pacing, I don't know, I feel like you should just do your jobs better. I don't know, it just feels like, well, you kind of screwed up, and it's like, I don't know, you need to pace. How about you just, I don't know, fix your shit?

41:17Alex VolkovSo he- here's the thing about just fixing your shit, just not a push, just a clarification. Um, the employees, specifically in the Pacing the Frontier letter that they all signed, they're all worried about the market forces and capitalism just pushing each other labs if they are in pursuit of, you know, Anthropic is is about to host an IPO. They're becoming a public company with a fiduciary duty to their stakeholders. Uh, OpenAI obviously is talking about raising at a 100 billion dollar, like, uh, another billion dollar round, not valuation. I don't know what valuation is gonna be. Um, and, you know, everybody's trying to catch up. Elon's trying to catch up, et cetera.

41:52Alex VolkovThey're all worried about kind of everybody's trying to race against each other without thinking about, uh, uh, safety, just because of they're trying to catch up. And so I think that the w- when folks see that they are agreeing, whether or not it's legal, let's talk about this, right? So maybe th- this is the next point. Uh, folks feel a little bit better that collaboration could be possible. And when we talk about China, but China, China doesn't agree, et cetera, like, you know, there are incidents previously with, uh, nuclear laws and proliferation rules, et cetera, that, uh, there is an international con- consensus about how to develop

42:30Alex Volkovthis, uh, closely, et cetera. And if we talk about more powerful, even if we're thinking about them as tools and not like a complete entity that's gonna take over, if we talk about tools that are as powerful, that can hack any government and break any encryption, et cetera, uh, then some collaboration is needed at least. I think that this is the highlight, uh... Who else wants to chime in while we look in this agreement matrix where it says, uh, industry-wide pacing, Sam Altman and Dario Amodei, uh, says yes, Elon Musk generally agree, Mark Zuckerberg and Jensen say no pacing industry-wide. And then on the independent evaluators, it looks like, uh, uh, Anthropic already agreed

43:08Alex Volkovto this, uh, OpenAI will do independent evaluations, and then, um, uh, Mark Zuckerberg and Jensen both back. Like, this is something that looks like everybody's agreeing on. And then on the new coordination and rules, which means that, hey, the labs could collaborate between who releases which products when, uh, a lot of strong opposition from Zach Jensen and and Don- Donald Trump.

43:31Wolfram RavenwolfI mean, why would you call for regulation of yourself? That is the strange thing. If you are leading the company, you can decide for yourself, and if the others agree with you, they will do it. If not, then maybe your position is the right one. So why do they need something like that? And if they want to bring someone from the outside to oversee it, they can also do that. So that is not really the thing, uh, regulation usually supports or helps the people at the top more than the smaller fish because they don't have the departments to deal with this. So it would prevent startups from getting into the scene with new models, and it would also make it possible to say anyone not part of this will be

44:05Wolfram Ravenwolfregulated out of the picture, so no Chinese models for Americans anymore, like they have done import stops for other stuff. Of course, people could still do it, it's open weight, download it, but you wouldn't be able to talk about it openly, and that would also diminish the scene. So regulation usually helps more the people already established. If it's coming from them, that makes it look all the more likely, I think.

44:29Alex VolkovYeah. All right, folks, I think, um, we've covered this at, at length. It's almost an hour into the show. I think we, we, let's talk about some releases. Let's stop with the, with the doom and gloom, et cetera, although I think it's very, very important. Uh, and RSI is happening, and also we haven't seen Google on this chart, and we haven't seen Ilya Satskova, who's building, uh, safe super intelligence, uh, at some point will come up with something. Uh, so I think more folks coming out with more labs, um, is going to be changing this debate, but I think this debate is now with us. AI is on everybody's mind. Politicians are using this to scare people into voting

45:07Alex Volkovfor them, uh, and the election season is upcoming here in the US, and, you know, it's, it's gonna be, it's gonna be very interesting. Let's talk about theme 3. So, uh, the, the, the 2 other things I definitely would love to cover on the show, would love for you to stay with us, is the new LLM, non-LLM type classifier thing and the use cases JEV allows us. We, I can't wait to talk about this. Uh, we'll have Ellie, uh, Lab from TypeSafe and, uh, Francesco from Kua talk to us about this, and also Wolfram and I have been building demos, cool demos for you and like playing with this, uh, with JEV specifically. So stay with us if you're hearing about, if you want to hear about JEV or come back.

45:45Alex VolkovAnd also, uh, we have to talk about the rise of assistance, and we f- for this we'll have David Paulan join us in about 25, 30 minutes to talk about, um, what Peter, you know, asked in the beginning of the show. Have you used Muse? Have you used Grokbot? Have you used Instinct, et cetera? And why Instinct is raising a 10 billion. I actually don't know why they're raising at this crazy valuation, but we'll talk about at least the feature. Meanwhile, the theme 3 is voice and talking to agents, and I think let's start there. We have some exciting, exciting news in the world of voice, uh, A- AI, and I just

46:23Alex Volkovwant to, like, call out the, the few of the things. One of them is, uh, and Wolfram, I would love for you to cover this. Gemini released Gemini 3.8 Live and claims number one on speech-to-speech index. Um, this is not only just a model anymore, this is a live model that can speak to you and and and react in in real time. Wolfram, any thoughts on on Gemini specifically? I know you're a fan.

46:52Wolfram RavenwolfYeah, I mean, I've been waiting for a pro model, but for live, it has the advantage of it's cheaper and faster, and when they release something like this, it becomes available to everybody who's on Android and the Google search. So the audience of a release like this is so much bigger than most people just think about. Even if you don't select it consciously as your model, it, uh, it is still basing the global intelligence that way. And the thing is, um, the life aspect of this is also becoming ever more important. I mean, I'm still waiting for Google Glasses, and if you have some way to access the model directly and you can see what it does,

47:31Wolfram Ravenwolfthat ties it to the assistant perspective. How much more useful would be our assistants if they could see and understand what's going on in our lives? And this is a model that is, uh, directly in that regard. So extended thinking, that is also something we, in the live models, usually don't have the thinking for so long because it is a real-time interaction that it is being used for. So if it can think and think with high intelligence and very quickly, that is also a great thing. So it's number one in the speech-to-speech quality index, beating Astra or what is the other one, 3.8 life without thinking, which is also great if it's already on third place, if it is not even thinking much

48:11Wolfram Ravenwolflonger to do this. High success rate, time to first audio, also super important. If I talk to my assistant, I use it for home control, for instance. So if I say, uh, turn on the light while I'm going down the stairs, I could have fallen down the stairs already if it takes so long to find out which switch to toggle and so on.

48:26Alex VolkovGuys, I know that I am a type of person who gets very overly maybe excited about a new release in technology. I think some of the work here on ThursdAI is the bias for that, and this is why we're here. We're talking about like the latest and greatest, but I cannot like just keep thinking about the stuff that Jeff could solve significantly better than whatever live model that they release now that Jeff is out, specifically about home control, which needs deterministic clicks and not just LLMs thinking about all of the possibilities of all the texts in the world. But yeah, we'll get there. Um, all right, yeah, so Gemini 3.8 live.

48:55Wolfram RavenwolfSee how everything ties together in all this, like the assistant, uh, topic we have.

49:00Alex VolkovYes.

49:00Wolfram Ravenwolfthe voice topic we have, the JF topic, so all of this is integrated. That is a good sign that AI, the whole of AI increases quickly because so many components advance independently.

49:11Alex VolkovAnd it's also, I think, one topic that, like, even if we pause the frontier development, there's just so much to discover in the current realms and how to use them and a- a- and what to use them for, that we're, we're not stopping anytime soon with new developments, despite the pacing the, the, you know, the, the frontline frontier as well. Um, let's talk about GPT-1 Live. So the voice that powers, um, GPT Live, which is when you talk to it, is GPT Live 1. This is the model. I don't think I have a link for that. Let me just open this up. And OpenAI added that to, to the API as well. So GPT Live 1, it's here.

49:49Alex VolkovUh, and the, the demo that they showed is a very nice, very nice demo. Uh, they're, they basically made Richy Mini talk at the same level of like GPT Live 1. Uh, I think it's a very interesting thing. I don't know if, if based on the different, um, benchmarks. This is now number one, given where Gemini is and given like benchmarking. Um, I actually, Peter, could be an interesting question for you guys. Um, how is like live voice treated in terms of like, um, arena? Do you guys test for this? Do you like have people like talk to it for a while, or is there no testing specifically for the live version of agent?

50:28Peter GostevYeah, we don't have any anything else just now. It is very tricky just because it is so personal and you you need to kind of talk to it for a while and so on. So yeah, the benchmarking feels very d- difficult and maybe brittle for that kind of thing. Um, so yeah, no, no, I don't have a great answer for that, but yeah, I think maybe we should, we should try and figure something out in that space to then.

50:53Alex VolkovYeah, I it's also like, it's very personal to every person, and judging it based on that is very, very, very, very hard. Uh, but the, the, the, the main kind of, uh, testing and benchmarking is done, uh, let's say artificial analysis, for example, and they have GPT live, uh, no, actually transcribe. Yeah, it's a very interesting area to, to even start testing. Wolfram, go ahead. Have you

51:20Wolfram RavenwolfThat's also the area where the intelligence has to be just enough to know when to do a tool call. So I'm using this. I was super excited when it came out because the the natural conversation, it's excellent. And now I have made, I used Astra to make an app on my mobile phone where I can talk to my assistant using this, and that is connected to Hermes Agent, which makes it so much more useful to me than if I just use ChatGPT Voice or anything, because this is connected to the memories and also tool access and everything. So I, it knows what I've been talking about before, and I can just talk to it like a real person. So that is a big advance, I find.

51:55Wolfram RavenwolfAnd the Reachy thing, I've also been working on this, uh, but I thought that hand- a mobile app is more, more useful because I can use it every time.

52:01Alex VolkovYeah.

52:02Wolfram RavenwolfAnd this is, uh, yeah, it- it- I'm using voice much more now.

52:06Alex VolkovI think it's very interesting also to roll back a year ago and talk about, we're now talking about like real-time conversation with multi-model, omni models that can do cool tool calls and hand off the intelligence to a much smarter model like Astra, for example, while we are talking to them to not break interruptions. Whereas a year ago we talked about, hey, you know, I, I, I kept talking, the model didn't understand me. If you guys remember Moshi from, from uh, uh, the demo that we did that felt like too, too fast, but felt real. You couldn't talk to it, it didn't listen, no tool calls, no anything. Um, and I think the next frontier live is, we already have the video stuff, but I think

52:43Alex Volkovfor video it's already like it's still taking pictures. Uh, I think the next frontier there is the model that can fully, fully integrate with you. May I just say the Wolfram, you're right, like the, the way it integrates with our personal assistants, for example, is that these are capabilities. Gemini 3.8 Live, uh, GPT Live 1, and the StepFun StepAudio 3 that we're gonna mention now. Those are all capabilities of what the assistants can use to understand what we're saying. Uh, another capability that we told you about before is, um, the live transcription models, as we now have an AI producer listening to the show and kind of like judging

53:22Alex Volkovwhere we are. Um, we can, we can check in on it and see how it's doing. Um, those are capabilities. However, the assistant stuff that we're going to talk about very soon is context of your personal life. Why do we need those capabilities? So we want to talk to our agents, but the agents need to respond based on what we're saying, right? So, uh, the assistant part has the context for your life, and very soon those will combine. They are combining already in context where you can talk to an agent knows stuff about you. That's what Wolfram you've built for yourself. Not quite a product yet, uh, but definitely is coming. Uh, so this is why we're talking to you about this.

53:58Alex VolkovThis is why it's like very, very, very interesting. Um, let's see what else is interesting here. Wolfram, you... Yeah, let's talk about Step Audio. I think that this is like the next, the next thing that's on our list of voice assistant. Uh, let me pull this up. Because I think it's at least interesting to tell you guys about, um, uh, the frontier in, uh, real-time conversation in ASR. These are not necessarily models. Uh, this is like the voice benchmark. So this is like 3 models from StepFun that are, um, Real-Time Preview, ASR Max, and TTS. Uh, Real-Time is number 1 on artificial analysis, conversational dynamics, and

54:37Alex Volkovspeech reasoning, and ASR Max is, uh, significantly lower than than Whisper on on the world error rate. And TTS has, uh, basically they released the full suite of tools for you to build to end end-to-end assistant, uh, in in your code, not necessarily a product, but in your code. So shout out to Stefan for this. I'm not planning to add credits right now, but I will say for Labs, if you want me to use your tool on the air, at least have like a very basic number of credits when you sign in. Um, and um, I think we'll finish here and and then we'll move on to like a different topic that

55:16Alex Volkovall of the stack for agents that talk to you and are participating, including a potentially a co-host on Thursday Night, like full agentic, we never tried. I think we had like GPT at some point, uh, are there, and uh, we will have to, um... We'll monitor, we'll we'll keep monitoring this topic, uh, to to bring the latest news to you, especially with our friend of the pod, Quindla Kramer, who who owns one of the, or who builds one of the bigger like pipelines for this with Pipecat.

55:45Wolfram RavenwolfI think what we, what is clear now is that we have the intelligence already. I'm noticing it in my personal life that most tasks the agent can do, maybe with some corrections, but it can do so much, it om- it feels like AGI in many ways. And um, now we need the speed to access that capability, to be able to interact with it in real time, with the voice agent and whatever is coming next. But uh, we are getting there. We are so close. It's just what we have, make it faster.

56:13Alex VolkovJust before we move on to uh, this week's buzz, I do wanna ask folks here if you looked at the Agents API from OpenAI at all. And, uh, if not, let's talk about this just a little bit, uh, because Agents API in public beta now, and the Agents API essentially is the Codex harness, but within your own product. Um, it's a managed service for everyone. Uh, you, you basically get a production agent that acts like Codex with the Codex harness. As a reminder, again, OpenAI ships direct access to their models, not the base models, but the chat models, so you can hit those and build whatever, whatever harness you like with whatever tools you like.

56:52Alex VolkovHowever, uh, that's not enough given how many tools these models are calling lately and how many, um, things their agents can do within their harnesses. So OpenAI now ships Agents API in public beta, very similar to the managed Agents API from, from Google, and uh, they have parallel programmatic tool calling, for example, and uh, the smart tool search only loads what's needed, and uh, it supports MCP servers and web search, uh, and the ho- harness is Apache 2 license harness, uh, and you pay only for token and, and tool costs. You don't pay for like the platform, et cetera, uh, with web search. And I am not quite entirely sure what, what happens there with the sandbox, whether

57:32Alex Volkovor not they provide it for you, whether or not it runs in their sandbox. Uh, but I think it's a, I think it's more of a bigger deal than folks realize. Definitely on the lower tier, but for folks who are listening to us who are AI engineers, this basically means you can build Codex into your own app or platform, and you don't have to build the Codex itself. Codex is maybe one of the top harnesses in the world in terms of, uh, how to use this. Um.

57:57Peter GostevYeah, and now I would say, yeah, this is a, a big deal for people, yeah, who, who implement AI in organizations because, and I used to do this in my previous job, where what you would do is that you would take a model and you're like, okay, decompose a process and put some, I don't know, cron jobs or whatever. And then what you end up doing 95% of the time, even if you are like, your job title is AI engineer, but what you're doing is stupid infrastructure things, which are completely boring and nothing to do with actual AI. So the more that we can remove that kind of thing, then we can accelerate

58:36Peter Gostevdeployment of AI within organizations. And, um, I think thi- this is a, a really good shift from OpenAI, just allows us to to do that more. And, um, I think if we see that, that I think what typically happens, right, a company innovates and then other companies do that kind of thing as well. So yeah, if we see other companies ship that with like open source opportunities as well, where you can swap in your models, that kind of thing is super valuable because just the more you can remove all of the other crap that I need to manage as an engineer, like they're absolutely golden. So yeah, I really like that.

59:10Alex VolkovThere, there's a distinction here that Wolfram, we, we talked about when I told you about this and why I want to cover this. Uh, yes, your agents or Codex or et cetera can already drive Codex, right? If you have Codex installed on your machine, you can basically like have computer use whatever and and drive Codex. This is without requiring that computer. This is like for Python on on JavaScript, like a node somewhere in the cloud that you're running. Uh, this is building that harness and agent there so that it can drive parallel programmatic tool calls that you don't want to build. Like Peter is completely right. There's like, uh, folks maybe don't want to maintain

59:40Alex Volkovall of this. And also, it's a very important reminder. OpenAI trains their models in in coordination with their harness. Anthropic does the same with Claude Code. Famously, Gemini does this with, you know, with their harness as well. Um, the combo between the model and harness is a very important one. Models behave better in their own native harnesses because they've been trained on with them. And so this is a way for you to basically milk OpenAI's APIs for more than just the intelligence, the reasoning. A very interesting release from OpenAI. One last thing, unless Nisten, you want to comment, uh, uh, one last thing that's interesting here. This is, this feels like a big platform developer release, and this

1:00:20Alex Volkovis exactly the kind of release they would have saved for Dev Day because all of the developers who use OpenAI in this way are showing up to Dev Day. Why am I saying this? I'm saying this because I'm super excited about what else they have for Dev Day if they decided to release this 2 weeks before the biggest developer convention of the year for OpenAI. So if they're releasing this on Dev Day, sorry, like 2 weeks before DevDay, I gotta wonder what they have in store at DevDay for us. I'm very excited about DevDay. It's coming up in 2 weeks. Uh, Peter and I will be there, uh, we'll do coverage. Uh, and if you are coming to DevDay, let me have a reminder for you.

1:00:53Alex VolkovThe day after DevDay is Fully Connected from CoreWeave, a 2-day conference with now, I think, 4,000 people as well. We have a free ticket for you to join, and as recently as last week, we talked about a very surprising guest at the, at Fully Connected. Pitbull, the international superstar, is going to headline the party at the end of Dev Day. So even if you know you're in San Francisco and you don't have the 2 days to confirm, come on Thursday and then have a party with us. Right down there below, without speaking about this code specifically, is the free ticket code for listeners of ThursdAI to Fully Connected 2026.

1:01:31Alex VolkovI looked at the launch release thing, I looked at the headliners, I looked at a bunch of stuff. Uh, it's going to be very, very exciting. It's the biggest thing that CoreWeave has ever done, and yours truly, Wolfram and I will be live giving you ThursdAI from there. We're gonna cover the news, and then we'll also talk about some, uh, w- w- with some folks from CoreWeave about the, the industry. Wolfram, go ahead. Uh, so if you're coming to OpenAI's Dev, they come to Fully Connected, uh, September 30th, October 1st, and uh, we will get you there, and please, you know, use our thing to sign up so you don't have to pay. Uh, and this is a basic benefit of listening to Thursday.

1:02:11Alex VolkovI will, I ve- will slowly transition to our next topic as well. In, in about 5 minutes, we're going to chat about assistance, so I think we can start, but before this, I will say, um, that the hackathon last week was an incredible success, and folks who came to the hackathon got early access to TypeSafe, uh, JEV model, uh, from from TypeSafe, uh, because they sponsored the hackathon last week. Uh, so I'm very, very excited to talk about, uh, uh, JEV very soon with Ali, but, uh, just before this, let's talk about, um, assistants. Peter, you asked this question before because you were busy with traveling, et cetera, whether or not folks used Muse.

1:02:51Alex VolkovUm, I posted a video about Muse, and by far my most liked video on on YouTube, separately from having a conversation about this on ThursdAI, and it seems like we have, um, basically an insane race between assistants. We'll have David Paul on from Assistant Benchmark to join us, uh, very, very soon. Uh, however, I just want to do like a survey here and in the comments. Which assistant are you guys using right now, if at all, and whether or not you've changed in the past month, let's say. Uh, the options are obviously the standard ones for developers like us, OpenClaw Hermes,

1:03:30Alex Volkovuh, and also the the latest and greatest from Grokbot, uh, Muse from Meta Instinct that we should cover in a moment, and some others like Town and Val and like a bunch of other ones. Uh, so let's start with the, with the comments already. Folks are saying Hermes, uh, Wolfram.

1:03:46Wolfram RavenwolfHermes definitely. I me- I use Codex at work because we are not allowed to use Hermes there yet. But Hermes, I'm so, uh, in- involved in the ecosystem now. I am one of the contributors to the project. I have, I think we are almost at 40 patches now where I changed that and submitted PR, so I don't want to switch anymore. It's open source, it runs on my system, I'm locked in in that. It's my agent.

1:04:10Alex VolkovSo you're Hermes. Uh, Peter, do you have any assistants or using, uh, harnesses as assistants, and what is the difference between them? That's also an interesting question.

1:04:19Peter GostevI must say, I don't know if I'm behind the curve, but I, I sort of tried them out and I never end up using them. I don't know, they're just something about that I don't find that helpful. I find

1:04:32Alex VolkovLet me ask you

1:04:32Peter Gosteva bit more work.

1:04:33Alex VolkovLet me ask you this in a different way, because I think the distinction there matters. We're going to talk about the distinction. Do you have anything a- agentic-wise that's with intelligence that tells you about the new incoming trigger from elsewhere, like a new email or a new message from Slack or et cetera? Do you have anything that's like

1:04:50Peter GostevNo.

1:04:50Alex Volkovproactive in a way that's like serving as your system with your context?

1:04:54Peter GostevNo, the be- I do have like a few automations running. I wouldn't kind of quite classify it that way. So I w- I do have a few things where I'm like looking for new restaurants that opening in London or something like that.

1:05:06Alex VolkovYeah.

1:05:06Peter GostevAnd it just like, it runs every couple of weeks. So it it's kind of semi version of that, but it's very targeted.

1:05:12Alex VolkovYeah.

1:05:12Peter GostevI found, for example, one thing I've got is-

1:05:14Alex VolkovIt's an automation, not a heartbeat. I think that this is like the thing that we're getting to.

1:05:17Peter GostevYeah, yeah.

1:05:18Alex VolkovIt's not something that constantly runs for you,

1:05:20Peter GostevYeah.

1:05:20Alex Volkovrather it's something that you wanted and then you just didn't wanna

1:05:24Peter GostevYeah.

1:05:24Alex Volkovkeep remembering to do.

1:05:26Peter GostevExactly. Yeah, so, but for the heartbeat stuff, I don't have anything that I found particularly useful yet. And then, yeah, I don't know, I just prefer to check my own emails and I prefer to, I don't know, manage my own calendar. I don't know. That- that's how I feel like I know stuff. I don't-

1:05:43Alex VolkovYeah.

1:05:43Peter GostevI- the- the same way how I write myself and I read things myself. Like, there are a few things that I- I just actually that I feel like that's my job as a human, uh, rather than just like a- a farming this out to an agent.

1:05:56Alex VolkovUh,

1:05:57Peter GostevBut I might change

1:05:57Alex VolkovHow about-

1:05:58Peter Gostevmy mind.

1:05:58Alex VolkovHow about you folks? Uh, let's have Yam. Do you have any usage of assistants in your life currently?

1:06:04Yam PelegCodex, Codex, like.

1:06:06Alex VolkovDo you feel that Codex is answering the, the, the, the, the, the threshold of assistant versus just a AI agent?

1:06:13Yam PelegLook, look, I have a lot of customizations, uh, around Codex, so it's not out of the box. Uh, but I must say that all the new products like News and, and uh, what was the last one? Oh, and, and Grokbot. I want to try them as well, but currently have like a custom, custom scripted, uh, whatever. You, you know how it goes, you know how it goes. But mostly Codex, mostly Codex, I think.

1:06:43Alex VolkovUh, all right, let's

1:06:44Yam PelegIt's amazing, seriously.

1:06:45Alex Volkovguest shows up, uh, Nisten and LDJ, I want to hear from you. Nisten, let's start with you. Any assistants in your life?

1:06:54Nisten TahirajYeah, uh, I built the Nisten assistant.

1:06:59Alex VolkovYeah, that's, that's a great answer, actually. Yeah, yep, yep.

1:07:02Nisten TahirajYeah, so I just, uh-

1:07:04Alex VolkovAlso, it's a great name, the Nisten assistant.

1:07:06Nisten TahirajNo, it's gonna be, I have no, it's just called assist.ts. It's a single TypeScript file because

1:07:11Alex VolkovYeah.

1:07:11Nisten TahirajI don't trust the dependencies. Uh, I don't install any pip. Uh, the only dependency is bun, and it's extremely fast at opening things and opening emails and stuff, but then I just end up delegating a lot of the other tasks to simply to to Claude Code. And uh, mainly I keep uh, Astra for reviews, and uh, I I really like the Astra web interface, like for editing videos and making animations and stuff. I just use that from the web. So, uh,

1:07:46Alex VolkovSo no. Your answer is no, you have your own personal one,

1:07:48Nisten TahirajYeah.

1:07:49Alex Volkovand it's a specific thing. Right, let's, uh, LDJ, and then I'll talk, and then we we'll have a place for our guest as well.

1:07:55LDJI'm I'm kind of with Peter here, where I've I've tried a few different things and Grokbot and other things, and I guess it really just depends on what exactly you're doing, and I do end up kind of going in different phases where I'll be like working on specific projects that are very different than others and sometimes find them more useful. But, uh, for the most part, uh, using Hermes Agent just kind of as a, uh, like a long-term persistent- like that benefit there, the- the long-term persistent shot of- of really having a in-depth, uh, context of the past memories,

1:08:27Alex VolkovYeah.

1:08:27LDJfind that useful. But besides that, really just kind of

1:08:30David Pawlanof schedule tasks with ChatGPT, find that really useful, not not anything too crazy beyond that, but I want to start using Muse.

1:08:37Alex VolkovUh, that's a very interesting thing. I want to start using Muse, you you answer. So I'll answer as well, folks. I've been, uh, completely obsessed with transitioning my life into automation with proactive assistant. I told you on the pod that I think that 2026 is the year of proactive assistance specifically, and the reason I told you about this is, um, that I think that having assistants that are active, I think is very, very important. And so I constantly right now use a Grokbot. I have more than, we can take a look actually. Oh no, we can't. Uh, I'll tell you why later. Uh, but I use Grokbot constantly. I use Muse since it came out, and I think it's just wonderful and incredible in, in

1:09:18Alex Volkovmany, many, many ways. And um, I still have my Hermes. I barely talk to it anymore, but Hermes has a bunch of context as well. And here are the few... You know what? Let's, let's do, let's do the invitation. David Paulan just joined us. David, welcome to the show.

1:09:33David PawlanThank you, thank you.

1:09:34Alex VolkovUh, nice to, nice to have you here. You have been blowing up on my timeline, and it's a very interesting thing where X timeline is specifically tuned to the interests of the people.

1:09:45David PawlanYep.

1:09:46Alex VolkovAnd so on my timeline, you're all over my timeline, and Muse is all over my timeline because I've been, like, glazing on Muse.

1:09:51David PawlanRight.

1:09:51Alex VolkovUh, on other folks, maybe that they didn't quite have that effect. So first of all, I, uh, the thing that you built is showing up on my timeline, which is Assistant Benchmark, which I would love to talk to you about. But also, you joined at the point of the show where I asked everybody from the hosts which assistant they use, and we have a very interesting kind of spectrum of of who uses what. Would love to hear from you before you even introduce yourself who you are and what you build. What is, um, w- what is your daily driver? I'm assuming you're using an assistant.

1:10:20David PawlanYeah.

1:10:20Alex VolkovAnd then we can talk about the differentiation between an agent and an assistant, what's what.

1:10:26David PawlanLove it. Uh, number one assistant I'm using is Instinct for personal stuff, Grokbot for more work-related things. Uh, reason being Grokbot, I, I really love the UI/UX and creating different bots and the organization of it. Uh, Instinct just lives in iMessage, so it's just a little easier for me. It, it's a pure efficiency standpoint compared to using Muse, but it would, it would be between Instinct and Muse.

1:10:49Alex VolkovInstinct and Muse. Interesting. Um, all right, so folks, let's talk about what an assistant is. Uh, and uh, folks, if you don't mind, I'll take you back, but if you have a question or you want to comment, uh, Wolfram, I'm actually gonna keep you here because, uh, everybody else here doesn't necessarily use assistant, wants to try, so please tune with us. Uh, uh, David, uh, so the reason why you've been blowing up on my time is because I was overobsessed with assistant. Muse,

1:11:13David PawlanYeah.

1:11:13Alex VolkovI think for the past 3 days, like, most of my tweets are were about Muse, now they're about JEV, but before they were about Muse. I still use it incredibly well. I give feedback to the team. Um, and before that it was Grok, and before that it was Hermes, and before that it was OpenClaw. So I've been like assistant built for a while. Also Instinct, uh, which I, oof, I don't know what's going on there, but Instinct is talking about raising a 10 billion dollar valuation, uh, recently, and this is a startup that just started. So the world is also getting super excited about this next wave of AI that's helpful for humans. Let's define an assistant between us.

1:11:47Alex VolkovLike, I, I want like a, first of all, how do you think about what is the difference between an assistant and just like an AI agent? And then also, um, how that definition goes across labs, for example.

1:11:58David PawlanSure. Yeah, I mean, I think to me, an assistant is something that actually can execute the task. It's, it's not just something that pings you or notifies you, uh, if anything, that right, that creates like more work in a surface area where you're just now getting constant notifications, but it's something that will actually then identify the issue and then complete the task. Uh, maybe there, there's some approval in the loop in that capacity, but it's actually going to take that, uh, workflow from start to finish.

1:12:24Alex VolkovInteresting. Uh, Wolfram, how do you, how do you feel the difference between an assistant and just an agent or a harness?

1:12:29Wolfram RavenwolfIt's funny because AI, when we started out, it was always a user versus assistant in the chat model. So that was the official name for the AI in the beginning. It was just the assistant. Then the agents came out and started doing things, and now we are back to having the agent be the assistant because it's actually a helper. And that's for me the assistant. It is providing assistance, it's helping me, it's an agent that does something on my behalf and helps me in all, basically everything.

1:12:56Alex VolkovMm-hmm. My, my feeling is that there is a barrier between giving me more things to look at, like David, like you're saying, versus proactivity of like reducing my cognitive load.

1:13:10David PawlanRight.

1:13:11Alex VolkovI feel like assistants, specifically the new type of assistants with a heartbeat, that they look for stuff before I need to tell them, with proactivity features built in, with, uh, creative ways of using them as well. They teach me how to use them. For example, Muse, we can talk about Muse proactive features as well. I think that that's kind of the difference. Whereas, um, the generalized harnesses agents were built as coding harnesses that folks in labs just noticed they're super useful.

1:13:35David PawlanRight.

1:13:36Alex VolkovUh, but they're missing like a few things. And I think that this is maybe the reason why Peter Steinberger went on such a huge, uh, epic journey in the beginning of this year, talked to all of the labs, and now I think this is also the reason why, uh, Muse has a heartbeat, has a soul.

1:13:52David PawlanRight.

1:13:52Alex VolkovMD file and an agents and memory.md file. And also, I don't know if you guys know this, uh, Muse has dreaming built in. So you can literally ask Muse about what it dreamt on, and mu- dreaming is like another process that that OpenClaw kind of started and

1:14:06David PawlanRight.

1:14:06Alex Volkovhas. Um, so this is an assistant for me. David, the reason why you're here is not because you're just a random person that I asked about, like, who uses the system, what is the definition? Uh, you've built Assistant Benchmarks, specifically highlighting the stuff that assistants do. Let's talk about Assistant Benchmark.

1:14:22David PawlanTotally. So, uh, it generally came from, I was, I was testing them all myself, right? Like, same as you, I was just using them all, doing my own market research and figuring out what's working, what's not working, and I needed a way to compare them all with one another. Um, it it's really more of like a use case driven benchmark, right? Like, it's not like we have 10,000 phone numbers provisioned and we're running all these iterations and, um, like, like a, like a AI Arena type, type proper benchmark. It's really more use case driven. So, you know, you click into a category, you can see the prompt that's being asked. Um, we then will,

1:14:57David Pawlanuh.

1:14:58Alex VolkovSo let's talk about, uh, let's walk through some of the categories that you are testing.

1:15:01David PawlanSure.

1:15:01Alex VolkovI think it's gonna be interesting because we're also a podcast, so n- not many folks are viewing what we're seeing.

1:15:06David PawlanYeah, totally.

1:15:07Alex VolkovSo let's walk through this. Uh, so you have categories like travel assistance, email, and finance, and shopping, uh, and and food. And there's like numbers here, and I would love to tell you about like, uh, I would love for you to tell us what like work in teams has 37, for example, and travel has 2. What what does this mean? It means that

1:15:21David PawlanRight.

1:15:22Alex Volkovyou've tested 2 assistants on this category?

1:15:24David PawlanUh, so that means that there's total of two assistants have submitted themselves for, for the benchmark within that specific category. So under the travel category, you have Miso and Soar, and those are, uh, travel specific assistants. Um, if you look at the general category, those are like the Muse, the Instinct, the, the ones that are just doing your like day-to-day workflows.

1:15:44Alex VolkovMhm.

1:15:45David PawlanUm, if you look at more of the work in teams, uh, there's 37 that have been submitted there. Those are more on like the B2B side of things. Um, they haven't been tested yet, and so we're, we're gonna be working through that. What's crazy is there's 116 total assistants that have now been submitted on the site.

1:16:00Alex VolkovYeah.

1:16:01David PawlanUm, I, you know, a tweet that I put out earlier today was like, I, I literally talk to my AI agents more than I talk to my girlfriend now. It's like I have 8 agents always running, that I'm testing all them. Um, it since the first week of launching Assistant Benchmark, I have run, I have personally run 273 tests across 23 different agents, um, and it's just the whole intention here is, okay, let's run these same prompts, let's see how these different, uh, assistants are going to respond to it. Let's also see how fast they are, and, uh, let's just compare them. Let's see what's going on and give from twofold, it's one,

1:16:40David Pawlanlet's give the general public a way to actually understand how these different assistants are performing, what they're good at, what they're not good at. But then two, if you're building an assistant, like, these are the general things that your assistant should be really good at, and if it's not performing well at it, like, you should probably do some more work on your assistant to make sure that it executes it really well.

1:17:01Alex VolkovSo the type of stuff, so I was, I was incorrect before, the above there are categories the assistants submit themselves to. The general category has 59 assistants, and the type of tasks that it assists you with are written here. It's like travel, uh, booking travel for you. Uh, what is Pix? Tell us about Pix.

1:17:17David PawlanRight, so these are the dimensions that are being tested. Um, there's 16 total dimensions. Pix is more of like a recommendation engine. So is it actually like recommending things, uh,

1:17:27Alex VolkovMm.

1:17:27David Pawlanquite efficiently? So, uh, as an example, you know, if you say, hey, find me a vegan restaurant in, you know, in- in my area for the weekend, is it actually gonna find a vegan restaurant, or is it just gonna find you a vegetarian restaurant? Um, then you have other things like memory, where it's not necessarily just your traditional memory of does it know who you are, but also is it really good at, uh, picking up on context? So, uh, as an example, uh, the way that I'll test memory is I'll first have it, I'll first test the, the travel category, and I'll tell it that I'm traveling to a certain weekend, uh, and to book me a flight.

1:18:02Alex VolkovYeah.

1:18:02David PawlanAnd then I'll go through a few other dimensions, and then after I'll test the, uh, the picks, uh, or sorry, the, the online task, uh, dimension, which is then asking to, uh, book me a reservation somewhere. And I'll ask it to book me the reservation the same weekend that I asked it to travel, but in a different city. And then I'll see, does the assistant preemptively prompt me and say, hey, you're supposed to be in Chicago that weekend. Why are you looking to book a res-

1:18:30Alex VolkovYeah.

1:18:30David Pawlanreservation in New York? Like, are you sure? The agents that can do that, it's pretty solid contextual memory, right? Like, they're, they're keeping track of the pace.

1:18:37Alex VolkovYeah.

1:18:37David PawlanUh, not all of them do that. So some of them do, some of them don't. And so it's like, how do we test these different assistants and prompt them in specific ways to really understand, like, the breadth and depth of what they can actually do and what the different gaps are and all that.

1:18:52Alex VolkovI think it's a very, very interesting problem that you're trying to tackle. The, the reason why is because the type of benchmarks that we talk about on the show, usually for LLM intelligence, they usually start with LLM with like no context, and then all of them get the same exact system prompt, and all of them exactly the same exact task. And most of the time that task is verifiable with some code or LLM as a judge on the other end. So you can like test this from yesterday to today, and supposedly it doesn't change, and then between different versions of the model, it changes.

1:19:19David PawlanRight.

1:19:19Alex VolkovHere, you're testing the 3 parts of what makes an AI, in my view, which is model, harness, and context. You're testing all 3 of them at the same time, including context that includes memory from before. It's very interesting. How do you approach, um, h- have you heard feedback about this? H- how do you approach this from a perspective of, hey, assistants are also built in to learn incrementally and be better incrementally w- the more you use them, right? Like, that's what

1:19:44David PawlanTotally.

1:19:45Alex Volkovthe promise of Grokbot is, like, it better the more you use it, and and instinct and memory, like, builds up for you. How do you treat this from a perspective of a clean lab, let's test it in a vacuum, versus this is an assistant that's been going with me, and it's just better at learning one

1:19:59David PawlanRight.

1:19:59Alex Volkovversus the other?

1:20:00David PawlanSo, first off, we- it's not a lab, right? It's and and make that very clear, like, this is not a lab grade research

1:20:06Alex VolkovYeah.

1:20:06David Pawlanproduct. This is very use case driven, and and that's intentional, um, because the the everyday individual who's looking at this type of benchmark is not the everyday individual that's looking at an an arena LM, right? Like the the the arenas, the the proper LLM benchmarks are being looked at by AI nerds, LLM, you know, AI engineers that that are into the actual technical components. The assistant benchmark is being looked at by a random Joe Schmo on the street who is like, I don't know what assistant to use, can it book a flight for me? And it's just very like basic things.

1:20:42Alex VolkovYeah.

1:20:42David PawlanSo this benchmark is is m- meant for testing on a use case basis. Um, some use cases like memory and, you know, how persistent is it in in context, like there's things that we're not going to be able to test for another 30, 60 days because who knows how long the memory actually is able to maintain there. So it's, we're going to continue these tests over time. Um, Autumn Moulder, she, uh, recently was a SVP of engineering at Cohere. Um, she reached out af- like 3 days after I launched it, and she's like, this is awesome. I'm, you know, currently a free agent, just trying to figure out what, touching grass for a little bit, but I, I want to do a new, new project for the people.

1:21:20David PawlanUm, so she reached out. She's been helping me out, which has been amazing, and she brings a lot more of that, like, technical depth and knowledge to the project, more from like a researcher lens. Um, and so we're, we're spinning up a, a new iteration of testing, a lot more focused on like the B2B non-iMessage type agents, um, to try to start testing those with a little bit more, uh, speed, like a little bit more velocity and intensity. But there's so many components to this benchmark at the end of the day, um, it's really a way to see legitimate performance of outcome, right? This isn't about how quick is your model. This isn't about, um, different, you know, AI engineering technicality.

1:22:02David PawlanIt's about if I'm your everyday person, I'm a 29-year-old guy in New York who's flying to Chicago. If I message the, my assistant, hey, find me a flight to Chicago that's under 400 dollars in a window seat. Can it do it, yes or no? And that's really the baseline here of what we're comparing.

1:22:21Alex VolkovThat's great. Um, the thing that I would like from a benchmark like this, I think is number one. We talked about pacing the frontier before and all of the labs bringing on METR, for example, and having

1:22:32David PawlanYeah.

1:22:32Alex Volkovtheir independent. The thing that, uh, would make me trust, and this is the benchmark, besides the fact that everybody's quoting you because everybody kind of needs this for themselves,

1:22:39David PawlanMhm.

1:22:39Alex Volkovuh, is independence, full independence from the labs. I think it's very, very important. So I'll just say, uh, if any lab is paying for some of this inference, I think this needs to be clear as well if somebody's providing an extra tier so they'd be able to test them.

1:22:51David PawlanTotally.

1:22:52Alex VolkovI think that's fair because otherwise how would you test them? And also, I think the very interesting thing that folks have been asking you is that the bigger ones that started this whole category, OpenClaw and Hermes, are not on this. Would you want to address this on air?

1:23:03David PawlanYes, I'll touch on two points, like both of those. So first, there's no one sponsoring this, right? So I lead growth at Merit Systems. This is under my role at Merit. We're just, we live in this world of agentic commerce, and so this started as just a research project for me to really understand the landscape, to figure out, like, where can we really play in this space from an infrastructure layer? Um, so we're not being sponsored, we're not being paid. Um, any of the competitors that are out there are not able to pay us for any type of promotion or anything like that.

1:23:31Alex VolkovYeah.

1:23:31David PawlanUm, there might be a day where we say, okay, we need to work with someone like CoreWeave and get more inference, or whoever it be.

1:23:39Alex VolkovLet's talk about this after the show.

1:23:41David PawlanUh, totally. Um, and and we've actually recently floated the idea of, like, having a sponsors page,

1:23:46Alex VolkovMhm.

1:23:46David Pawlanuh, to give us a pathway to test things with with a little more rigor, um, obviously because that that's gonna cost a lot of money.

1:23:55Alex VolkovYeah.

1:23:55David PawlanUm, so if any of anything like that does happen, there's going to be a very clear page of like, here is who is, who is paying, and here is what they are paying for, and it is specifically for the research outcome. There is, there's no profit coming from the site in any capacity. Um, in terms of OpenClaw and Hermes, so the original reason that they were not included, uh, is because the performance of them is significantly dependent on the setup from the individual. So when you're thinking about testing these and scaling it across all of them, if I run the same prompt of book me a flight to Chicago, blah, blah, blah, um, my OpenClaw setup might do it unbelievably well, but your OpenClaw

1:24:37David Pawlansetup might fail. And this then doesn't prove as a valid use case point because if then someone says, oh great, OpenClaw is really good at this, and then they start and install and build their own OpenClaw setup, and then it fails, well, now they're going to look at the benchmark and they're going to say, well, what the hell, it said it was really good at it. Um, so the intention here is like we're really focusing on out-of-the-box solutions, consumer products that are pre-built packages that you can just use, and everyone who uses them, everyone who installs it or, uh, opens it up in iMessage, whatever it is, every single person is having the exact same experience.

1:25:12Alex VolkovYep, that makes sense. And, uh, I think the world has shifted from the folks who are building their own assistant in, in their own, like, Mac Mini that they bought

1:25:20David PawlanRight.

1:25:21Alex Volkovtowards the one that breaks less than the previous one,

1:25:24David PawlanRight.

1:25:24Alex Volkovtowards now these labs are maintaining the machine,

1:25:27David PawlanRight.

1:25:27Alex Volkovwhich is, I think, a very important thing. We didn't talk about it. Uh, all of these assistants need, like, a computer environment. All these assistants need a browser with computer use. All these

1:25:35David PawlanYep.

1:25:35Alex Volkovneed access to your memories and connectors for your context. I think those also the things that signify an assistant. David, I did promise you to let you go once you, once you can, because I, you have elsewhere to be. Thank you so much for coming.

1:25:47David PawlanOf course.

1:25:47Alex VolkovAnd do

1:25:47David PawlanThank you for having me.

1:25:48Alex VolkovI think it's very important. We are definitely expecting you to come back when the new assistant comes out or new capabilities

1:25:52David PawlanPlease.

1:25:53Alex Volkovcome out. Like, new capabilities came out this week for both Instinct and Muse, where they can actually phone businesses, which I wanted to talk to you about, but we ran out of time. David, thank you so much. We'll bring you on. David Paulin, uh, creator of Assistant Bench. Thank you.

1:26:03David PawlanAppreciate it. Thanks, everyone.

1:26:06Alex VolkovAlrighty, folks. So a very, very interesting, uh, transition from, uh, assistants and how they, they work into, you know, the world of, the world of computers. We have back, uh, Francesco Bonacci. Welcome, Francesco. How are you, my friend?

1:26:18Francesco BonacciHey.

1:26:19Alex VolkovUh, the founder of Kua, and which you can find at TryKua. And since you've been on and we've talked about computer use, multiple things have happened. So I actually didn't bring you on here as a guest, believe it or not, just as one commentator, because the world of computer use is exploding, obviously, because assistance without computer use is not every one of those assistants needs a computer use element because the web is human shaped and not assistant shaped. A few things I would love for you to talk about before our next segment, before we talk, we'll actually transition to to to to Jeff together, is um, uh, how are you seeing this world now that Muse has released something and they have their

1:26:57Alex Volkovown computer use, and Grokbot, their computers wasn't like the the the craziest ones. Um, and and ye- how how are you feeling about this world of computer use, uh, now that the assistant category is blowing up as much as we predicted it would be a few short months ago?

1:27:13Francesco BonacciYeah. Yeah, I guess it's like, I would like to start with, uh, it was about time, honestly, because I've been in the space for about 2 years, and I saw like these agents like scoring from 10% on OS world when it first came out, and I was still working at Microsoft. It was like, okay, we see a future where these assistants are gonna be able like to use the same tools as human. Um, how far is the future? Uh, definitely would have not expected like to be like 2 years, um, away, but it's definitely like such a cool and odd space to be in right now. Um, we, if you look at our, not only about like GitHub graph, but like our GitHub contribution graphs, we have like folks that are

1:27:56Francesco Bonaccilike very keen to go and fix their core driver installation.

1:28:00Alex VolkovMhm.

1:28:00Francesco BonacciUm, and like putting patches on whatever like their Hermes or OpenClaw or like Muse setup they have in place. Um, ultimately I got like many folks reaching out as well that are using RockBot with CoreDriver. Um, since then we also been speaking with the team, okay, can we do something? Like, it's, it's honestly like, you know, it's uh, it's a good space to be.

1:28:23Alex VolkovI think, um, the, the thing that, a- again, the sig- s- s- specifies an assistant for me is the ability to solve things. I do want to talk about the residential IP versus browser cloud IP problem, because when I use, uh, Hermes, for example, it runs on my Mac Mini, runs from my home network and uses CUA driver. Uh, when it clicks buttons on my computer, it looks like me. So Cloudflare, like, chills a little bit, and other places that look ba- based on detect IP, they they they let it use the the web more. Versus when I use Grok, uh, or Muse, they get flagged more as agents, and so I need to interject more and and and take over, for example.

1:29:06Alex VolkovUm, should you comment on this? What are your thoughts on this, like, area, whether or not the web is

1:29:11Francesco BonacciYeah.

1:29:11Alex Volkovgoing to prevent this use or actually get solved?

1:29:15Francesco BonacciYeah, so we hear that all the times from our users. Um, there are a couple of ways to circumnavigate these, I would like just call them limitations, but it's more how these system were put in place to prevent scraping. Historically, we end-to-end testing, and then it translated it into a large action models and agents. Um, so for CoreDriver, you can do things the desktop way, and that's basically when you interact with accessibility trees or you just rely on pixel coordinates, and that's actually less detectable than, like, starting a CDP connection over your browser.

1:29:47Alex VolkovMm-hmm.

1:29:48Francesco BonacciBecause then actually, okay, what happens is that Cloudflare detect that you're using either Playwright or just the Chrome-like protocol. So we can go there with CodeDriver whenever there is the need, whenever we need to ground more information in whatever action CodeDriver needs to do. But we have like, we have a agent action ladder model that is smart at this stage. It knows where it has exhaust all the option and it need to rely on browser use. Um, because again, that's more detectable. And also the way that we can, we can mimic the user picking and typing in a kind of like more normal way using code driver and accessibility and pixel coordinate than relying on, right?

1:30:28Alex VolkovThe funniest thing that happened to me this week with Muse was, uh, Muse told me about, hey, I got this, uh, CAPTCHA thing. I can solve it for you, or I can, you know, give you the, the controls of the browser so you can solve this. And the very funny thing is there's two buttons, like solve this for me, and no, I- I'm gonna do this, like, why would I ever want to solve myself? Only if you don't succeed. It's very interesting that these models are, you know, beating some of the very, very basic ones. Uh, I think actually in that case it was wrong. It was a cloud for check that kind of looks like a check, but there's a lot more going on. Uh, the question that I had for you was a load bearing question, but there is

1:31:02Alex Volkova feature that I've been waiting for. So there's two things I- I'm telling you and the audience, everyone. Uh, first of all, Muse, the iOS app, has a Tailscale connector, which is incredible to see. If you guys don't use TailScale, you- you should 100%. It's incredible. But if you want to give your assistant access to your network in a secure way, Muse has a Tailscale connector. Obviously, every agent with a, uh, VM, you can install Tailscale, approve it yourself, like that's possible. But the fact that they have a connector for people is just incredible. It's a product from Meta that goes out to billions of people and now has Tailscale

1:31:34Alex Volkovthere as number one. And number two, in terms of residential IP, a lot of these, uh, let's make it harder for bots to scrape our website websites that prevent your agents from using them, they are looking at IPs, and there's like a segment of IPs for the clouds, for Google Cloud and Lambda from AWS, et cetera. Uh, and so they would like flag, oh, this is a potential automation thing. However, Grok came out with a feature that now proxies the gra- the the traffic for the browser through your local machine. And I've been wanting this. This is a feature in in in uh Tailscale, by the way. Uh, it's called the exit router or something like that.

1:32:12Alex VolkovUh, but it does- didn't work with with with Grok specifically when I tried it. Grok has now this built in, so if you're using the Grok app and you want the Grok IP to kind of look like it's coming from your machine, definitely checkbox this as very important. Francesco, I want to switch towards a little bit of a, a conversation. Everybody else also, please feel to chime in here, of user computer use, uh, background stuff like we talked about, clicking buttons in different apps for me, checking, taking screenshots, et cetera. In the world of web use, uh, because there's accessibility there, there's also screenshotting there, and there's WebMCP. And I would love to hear from you, like, where that world

1:32:49Alex Volkovis going, because I think that it's still very inefficient to take a screenshot and make a decision, right?

1:32:54Francesco BonacciYep. Uh, another that definitely we are proud about with Quadra, we were pretty much there catching up with the latest spec of the MCP protocol. Even before he landed, the latest specification, I don't recall which date it is, but it's basically detailing a new way of interacting with skills. Um, the old fashioned way, and if you go back to the early OpenClaw days when Peter was pitching CLI plus skill is actually great, better than MCP, um, you will actually like try and embed much of the behavior that you wanted to detail for your agent within a skill or set of skills. Um, and unfortunately, there is only as much as you can fit in the MCP

1:33:36Francesco Bonaccidescription. Um, and one particular issue is that we are integrating in CuaDriver is skills over MCP, which basically means that, okay, say in our case we embed with the CuaDriver installations a skill for working on CuaDriver for macOS, Windows, and Linux. There is definitely different detours that the agent has to take, whether it's Linux and Windows. That stuff, potentially if CuaDriver is dealing with multiple operating systems in the same way, it can go and fetch the skill that is required for working with OS or Windows on the fly, which is great.

1:34:09Alex VolkovI think that's fascinating.

1:34:11Francesco BonacciYeah.

1:34:11Alex VolkovHow we're evolving from, hey, there was MCP and the tool definitions were overloading my fucking context, so I wasn't able to use them because my context wasn't that big. And then, so that's why we invented skills. Skills are these things that like very small and only pulls up the whole context my needs to now, hey, skills basically talk about how to use an API. Why not combine them so that when I provide my API with MCP, I also provide the skill of how to use this best. I think it's like absolutely fascinating development. And you told me about this, by the way. I think, I think you, you mentioned this first, like, what, what is, what the hell

1:34:41Alex Volkovis skills over MCP? And now I think it's like a very fascinating development. MCP obviously also comes with authorization built in. There's like a bunch of things, but now MCP also will come with, hey, here's how to use this. And I think the problem, the main problem that people had with MCP, the overloads, your context, is also getting solved because they're programmatic calling from MCP in code mode and stuff like that, right?

1:35:02Francesco BonacciYeah.

1:35:02Alex VolkovAnd so I think that, uh, uh, you know, the, the world's very exciting. And specifically in the web, uh, I think that that's something that where we're going, where versus like trying to take a screenshot, going through all of the possible clicks there, the website will tell the agents how to use them on their own, right? So I think that is like very, very exciting. Web MCP is coming up. Uh, Francesco, the last thing as a transition to the next segment of the show where we need to talk about, uh, TypeSafe and and JEV. Uh, you posted something about computer use and decision trees, et cetera. Do you want to talk about this a little bit before we introduce our next guest on

1:35:37Alex Volkovthe show?

1:35:38Francesco BonacciYeah, briefly. Honestly, I was like forwarded this demo the other day by smaller teammates. Yesterday I didn't have time to catch up with everything. Um, but it's honestly mind blowing the fact that right now we're able to defer many, many, many choices, like in the decision tree of the trajectory of an agent, say that it needs to pick type or scroll, or like which target elements it needs to work with. Uh, back at my time at Microsoft, we used to have, you know, where when vision models were not bounded in pixel coordinates, we used to have something called set of mark prompting, which basically draw rectangles on your, on your screen depending on which

1:36:14Francesco Bonaccielements you want to click. Uh, and there were a couple of interesting browser use and desktop use demos made yesterday, um, which basically is like, okay, you're just feeding this, this decision tree back into that, and it's when you don't, when you just, when you need to take quick choices, that for computer use agent is like, okay, which action I need to target first between clicking, typing, and scrolling? Which

1:36:36Alex VolkovBrother, you are bearing the lead. Let me say this for you in the way that I expect you to say this, because I think it's very important. Last week, Astra was the fastest, maybe not the cheapest, but definitely the fastest computer use there out there, and you guys are like saying, you know, like that's the LLM that needs the driver, et cetera. Uh, a week passed, and in that week, we're looking now at a chart from you guys at Computer Use that says, hey, for 79, oh, sorry, 71 out of the, the, the 80 tasks that we gave it, uh, the new paradigm from TypeSafe that's called JEV

1:37:07Francesco BonacciYeah.

1:37:08Alex Volkovis now hitting that with a 0.1 second accuracy, whereas Astro took

1:37:12Francesco BonacciYeah.

1:37:12Alex Volkov5 seconds, with uh, 0.1 second medium latency, whereas Astro took 4 seconds, and a

1:37:18Francesco BonacciYeah.

1:37:19Alex Volkovin- impossible to calculate difference in price, maybe 400x terms price, they can do most of these decisions. I think that this is why, uh, I think it was very exciting to see that computer use is now not only getting solved conceptually, but also like

1:37:33Francesco BonacciYeah.

1:37:34Alex Volkovin practically, it's gonna be super faster as well. And so I think this is a good transition to bring on Ali Labs from TypeSafe. Ali, welcome to the show. Uh, let's start. with, first of all, congratulations. What a

1:37:49Allie LaabsThank you.

1:37:49Alex Volkovcrazy, crazy launch you guys had. Was it yesterday only? Like, like-

1:37:55Allie Laabs2, 2 days ago,

1:37:56Alex Volkov2 days ago.

1:37:57Allie Laabsum, or 6 years ago in the amount of time that it, the last 48 hours have felt. It has been absolutely, it has been absolutely bonkers.

1:38:06Alex VolkovSo let's start with the announcement. TypeSafe Labs, eh, has, besides sponsoring our hackathon this last weekend, which, by the way, thank you so much, like tons of folks got early access to JEV and got super excited to build things, has been in stealth building a thing for the past 2 years. The

1:38:23Allie LaabsThat's right.

1:38:23Alex VolkovCEO and co-founder Diogo Almeida has worked on RLHF, one of the co-creators, I believe, of RLHF. Please correct me

1:38:30Allie LaabsYep.

1:38:30Alex Volkovevery statement that's wrong. AI assistant help me research this, but uh, co-creator of RLHF and ChatGPT as well, and for the past 2 years has been working on something that's different. Alley Labs, Adevrel, Ad Types Safe. Please tell us what is different? What is Gev? I would love to hear from you directly, like what we're talking about. Why is it so exciting? Why is it all over my feed?

1:38:48Allie LaabsWhy is it? Okay, we've created a new class of model. We call it a System 1 model. It's the first of its kind. It's our first public model of its kind. Of course, we've had several iterations before we decided, uh, we wanted to to bring it to the public. We've had a lot of people using it leading up to this.

1:39:06Alex VolkovSystem 1 referring to the Thinking Fast and Slow book from Daniel Kahneman

1:39:11Allie LaabsYes.

1:39:11Alex Volkovand Tversky?

1:39:12Allie LaabsExactly.

1:39:12Alex VolkovOkay.

1:39:12Allie LaabsYep.

1:39:12Alex VolkovYeah.

1:39:13Allie LaabsThinking Fast and Slow. That is the, that is the inspiration behind the name, because this System 1 model is specifically System 1 thinking. It's fast intuition. And when we say fast, I mean, you were just looking at the stats a moment ago, like the, that, that computer use, computer use demo that they've created, right?

1:39:30Alex VolkovI have one better for you, Ali. I got access.

1:39:32Allie LaabsOh, I want to see it.

1:39:33Alex VolkovI got access yesterday night.

1:39:35Allie LaabsOh, you did?

1:39:36Alex VolkovAnd thank you for the folks who provided access. And for the longest time, I was noticing that Twitter, FKA X, has a tendency to over-obsess about the topic. For example, most of my tweets are now about JIV that I see on my For You timeline, and I built an extension before based on Cerebrus and the fastest model that they have to analyze all my tweets. Uh, it took me maybe 25 minutes to implement JIV into the system, and so here is my timeline as scored by Jev in real time, okay? So every tweet that has Jev in it or related to Jev in the quote is marked as yellow. I will scroll a little bit, and you can see a bunch of yellow.

1:40:14Alex VolkovBut what you can see on the side here is that we're doing insanely fast analysis. Those categories show up before an LLM could even respond, and I'm scrolling super fast. I'm going to scroll super, super, super duper fast. You'll see we're doing 12 tweets a second with Jeff, uh, and this costs me nothing. I literally tried to calculate. The thing that I want to highlight is that you guys are now calling, uh, uh, you don't even price output tokens, right? Like, only input tokens is priced, is that correct?

1:40:43Allie LaabsOn- only input tokens. That's right. You pay for input tokens, output, we we like to say output tokens are too cheap to meter.

1:40:49Alex VolkovI I think Diego specifically said too cheap to fucking meter in the release video, if I'm not mistaken. And also, the input tokens are priced in, in billions and not in millions like we're used to from any other LLM. So it's like 45 dollars per 1 billion input token. Is that, is that

1:41:03Allie Laabs42, yeah, 42 per billion.

1:41:06Alex VolkovCheaper.

1:41:06Allie LaabsI'm so glad we ended up doing that. Diogo was particularly excited about what if we, what if we say pricing in per billion tokens? And it was such a good idea. It's such a great moment in the video. Yeah, because it is, it is cheap. It is that cheap that we can be talking about on a different order of magnitude.

1:41:23Alex VolkovCheaps and fast, and I think the speed is something that I also want to show you and and the audience as well. I have here Qwen 3.8 27B, a very fast model running on Cerebrus. Cerebrus is, as we know, an LPU chip that runs like super fast. Uh, you guys absolutely just crush Cerebrus in impo- impossible ways. JEV is 4 or 5 times faster than PerTweet and is 20x cheaper. It's like a factor of 20, and I'm pretty sure that this is like only on my end, and there's ways to optimize this. I didn't look into developer too much, like Fable cooked it in 20 minutes. Uh, but folks, what I have here with JEV, which we we'll talk about why this is so efficient in a second, is essentially a real-time cognitive firewall

1:42:07Alex Volkovfor my X that I'm using this new system, the System 1 thinking, to categorize in real time everything that I see according to buckets that I design, and I want to see all the Jeff tweets, for example, highlighted. But for, for example, you can also use this. That's a feature. It's public, by the way. I'll put a show at the end of the, I'll put a link to this at the end of the notes, is that you can hide other things. So you can literally have an ad blocker that's like your thoughts ad blocker and not necessarily based on whether or not this ad came from an ad server, which is a completely different paradigm. And this is all done with less than 1 cent, and I've been scrolling and scrolling

1:42:41Alex Volkovsince this morning. So, like, the, the, the pricing there just doesn't compute for us. And this is just one example of the, I think now, 500 or so things that I saw on my timeline. So back to you, Ali. This is not me talking at you, telling you how amazing the things that you're working on is. This is a, this is

1:42:57Allie LaabsAlthough I do love, I do love just sitting back and hearing people just like, that excitement that you're expressing is what we're seeing all over the place, right?

1:43:05Alex VolkovI think it's crazy because

1:43:06Allie LaabsIt's real. This is a real thing.

1:43:07Alex VolkovWe, we use LLMs for this. LLMs are general, and they're streaming tokens. What are you guys doing differently that allows for this scale? Please, please tell us, educate us on, on what allows this type 1 thinking, and what is the innovation here?

1:43:21Allie LaabsSo, so that is a, that is a great opportunity to say my first of what could be multiple, I can't talk about that. Um, when it comes to, like, the actual, like, what the secret sauce is, the, uh, the architecture, that is a thing where I just stay far away from that. Um,

1:43:37Alex VolkovYeah.

1:43:37Allie LaabsDiogo's been doing some town halls in our, uh, in our Discord. He's the best one to know exactly how much that we're willing to talk about that. But w- but it is, it is the case that we have secret sauce, right? We have, we have had break in innovation and enabled i- in order to do this, and it is, this is a new, it is a new kind of training that we call reinforcement learning for calibrated decisions, right? That was in our RLCD, that is in our launch video, and that is, that's, that's a huge part of, that's a huge part of what it is, is we're doing parallel processing of all, uh, of all of the questions that you send in at the same time, and we're not, we're not generating text character by character,

1:44:17Alex VolkovMhm.

1:44:17Allie LaabsRight? I've had this before I started at TypeSafe. I mean, I guess I still feel this way. I just felt like the whole, so much of the landscape of what we're doing with AI has felt completely bizarre to me, right? For chatting, it totally makes sense, right? RLHF is great at creating, uh, like a brainstorming partner. Um, it has turned out to be really good at coding. I think there's a lot of different directions that can go in the future, though. But like, let's just talk about like chatting, right? Works super well for that. It's gonna, it generates it word by word. It's very similar to how we form words in our mind as we're talking to each other,

1:44:50Allie Laabsright? I'm not, I don't know what my sentences are before I say them. So that all, that all makes sense. But then what we've done is we've taken these LLMs, and then we've gone, okay, now let's make them call tools, right, to do things. Like, so I want to have an LLM order me a pizza, and it's going to do this by calling a bunch of tools, maybe it'll even move your mouse and et cetera, et cetera. It's so weird to me that we are using these, like, big old language models to create the text or the instructions to interact with an interface that was already on the other side of the interface is machine code. It is like a website, it's data flowing to the pizza server and putting in an order.

1:45:32Allie LaabsYet we've translated that into a user interface, whether it's graphical or even a command line interface. A command line interface is still a human user interface, right? And so we're- we're taking LLMs, we're having them output like a human interface language, putting it back into machine code. We started in machine code,

1:45:51Alex VolkovYeah.

1:45:51Allie Laabswe ended in machine code, and yet we did an extremely expensive, like, translation layer in the middle just to make them, like, pretend to be a human for a moment.

1:45:59Alex VolkovAnd we optimized

1:46:00Allie LaabsWhich is like

1:46:00Alex Volkovthe crap out of that process as well.

1:46:02Allie LaabsWe do.

1:46:02Alex VolkovSo it feels like almost there. It feels fast, but yeah, you're, you're absolutely right. This is kind of wasteful and ridiculous.

1:46:07Allie LaabsYeah.

1:46:08Alex VolkovAnd spends a lot of GPU, where those GPUs can be used for, you know, training the next models, etcetera.

1:46:12Allie LaabsIt makes me think of like a, like a, uh, you may, you're gonna make a, a robot, a bipedal robot to wash my dishes, and I go, well, we've already made a robot that washes dishes, and it's called a dishwasher, and it's really efficient. It's extremely, it's like does it really fast for the amount of dishes. Like, it's a, it's a very good machine. Now, loading a dishwasher, it's a pain in the ass. It's a separate thing, but just for the washing part, right? It's like we optimize in technology of all sorts for the problem that we want to solve, but LLMs doing tool calls and all this has just been in this weird extra layer. So what we've done is we've created a model that gives access to essentially

1:46:49Allie Laabsthe latent intelligence that is inside what we, in the modern day, call AI, and but exposes it as machine-readable outputs, as probabilities, as type-safe probabilities, right? Like, you know, as typed objects. And this now finally lets us use AI in a machine-to-machine context without ever turning it back into user space. Um, and I just think it, it to me, it's the day I saw this, when literally when I came in for an interview into this office, and I sat down, and one of the interview tasks was to, like, see the API for the first, like, see what it was, because we were stealth, right? So I had no idea what it was.

1:47:28Allie LaabsI sat down, and it was for me. it was, oh my fucking god, this is, this is what I've been struggling with for the past two years. I've, I had a whole bunch of projects that I've now ported to using Gev, right? Just kind of like you, this, this tweet analyzer, right? Is I've had all of these things with Gev-shaped holes, and I just, now it was like all of a sudden, all of a sudden the tool existed. All of a sudden it was this is the missing piece.

1:47:51Alex VolkovThis is, it definitely feels folks who are excited about technology, excited about solution space, are discovering that this is could be the missing piece. Uh, Sunil Pai from Cloudflare posted, this feels like the React announcement era where folks are like, oh, this is exactly what I've been missing, and I don't want to overpay as well. I'm using this tool that could also do this. Maybe a bulldozer can flatten my pavement, but also a bulldozer can, I don't know, lift, lift something, whereas there's a better tool for that, and this feels like the tool was more efficient, and I don't need to bring a whole bulldozer and spend a lot of money. And I'm showing here on the stage kind of the speed with which those

1:48:27Alex Volkovcategorizations happen, whereas LLM outputting one by one by one, we know this. Um, however, Ali, this is not a diffusion model that we've talked about, diffusion models that kind of like fills holes and uses diffusion methods for kind of LLM-y things. Uh, what doesn't Jeff do for me? This is not like better than asteroid talking to me, right? Or in fact,

1:48:47Allie LaabsCorrect. I liked Diogo's wording in our launch post. It said like this trade off, it doesn't come for free. We don't generate text. We don't generate text. We don't want to generate text. Ironically, the first thing a lot of our users have done with it is, like, use

1:49:00Alex VolkovGenerate text.

1:49:00Allie Laabsit to generate text, which is very silly, like, putting it in a choice to say, pick the next letter from this list. It's such like a, it's going, it's fun, and it's actually kind of fun to see what bizarre strings come out of it. It kind of feels like the early days of, like, predictive engines. It's not what it's for. Like, we're not, it's, we are very unlikely to optimize it for, to be able to generate text

1:49:19Alex VolkovSure.

1:49:19Allie Laabsin that way because we've already have that in the world. It's LLMs. LLMs are great at generating, like, generative text when you want human prose, right?

1:49:28Alex VolkovI have to, I have to pause. I'm so sorry, but I have a question that it's related to what you're just saying. You're saying LLMs versus JEV. Does JEV have a category now? Like, do we need, do we have a name for JEV and other models like JEV that would come out or have come out before, or there's no-

1:49:44Allie LaabsSystem 1 models. That's what we-

1:49:46Alex VolkovSystem 1.

1:49:46Allie LaabsWe use System 1 models like we use large language models to describe a different kind of

1:49:52Alex VolkovYeah.

1:49:53Allie Laabsa different kind of AI, right? It- AI has means a whole lot of things if we go back long before LLMs, right? AI has meant a lot of things, but now we are System 1 model. And while we are the first, we expect in the future System 1 models will be the catalyst for an economic revolution. We believe that this

1:50:10Alex VolkovWow.

1:50:10Allie Laabswill be driving a massive amount of automation in our world's future. Now, will it be JEV that is doing all of that, or will other people catch up and create their own System 1 models? What we have 100% confidence in here is that System 1 models will be doing it. We hope it'll be JEV doing all of it, right? Obviously,

1:50:29Alex Volkov100%.

1:50:29Allie Laabshe would love that, but, and and we have a big head start, we've invented it, et cetera, et cetera. But we expect other people to make, you know, system, other System 1 schema compatible models as time goes on, like what happened with, you know, like what happened with ChatGPT, and now all of a sudden, you know, there's, there's a whole bunch of these. But we, and so that's why, that's why we decided to coin it as System 1 models. We, we want to, we're giving the name System 1 models to the world, like that, this is it.

1:50:57Alex VolkovI'm down, I'm down with System 1. Like, we're gonna call this System 1 on the show until, like, um, uh, another came, uh, another name comes about. I'm showing a, a, a screenshot from TypeSafe.ai, where, uh, you guys are citing 133x faster and 444x cheaper than other, like, like competitive level models for some of the tasks that you're testing. And I think that maybe, maybe, please correct me if I'm wrong, maybe you guys were surprised with the amount of stuff that people are shoving this, like, puzzle piece that they've been missing and, and using use cases. Um, I think I saw Diogo, the, the founder, actually react to some stuff and say, we

1:51:34Alex Volkovshould use, we should have used this as an example versus the other one, uh, which is in several. So, so if you don't mind, I'll, I'll bring back Francesco, who is a, a co-founder of Kua, which does computer use driving as well. And Francesco, I would love for you to tell us, uh, from, from the brief experience that you guys had, what effect do you foresee this for computer use and web use, uh, uh, interfaces and drivers? Are we talking about like much faster, much more, like much better as well?

1:52:04Allie LaabsYeah, I would say we were really thinking about real time computer use. It was like a couple of months ago, some interesting demo with GPT Live, but still when, when you're just fitting that in into long horizon trajectory, you know, it's, it's hard to have a trajectory of all computer use agents and translating it for a real time computer using agents. I think that the, that is actually the missing piece here. Um, being able to take this fast decision and being able to live, to work with the same space, basically the binary space where really a computer is at. I think, honestly, that's, that's the way to go forward.

1:52:43Allie LaabsAnd I can tell you about my experience at Microsoft. It was like, there had to be a better way. We were just like working with a YOLO icon detection model at the time, just like throwing

1:52:51Alex VolkovI think it feels like that better way. Like, like I, I definitely don't want to overstate, but based on everything that I see, and again, as we saw, the algorithm was overfocusing on the stuff that I like. I talked about Muse, now everything is Muse. Uh, this feels like many of the engineer friends of mine that used LLM for something else and are considering costs are now completely just removing costs from the equation now because of the, you know, the speed and the fact that you guys are not charging anymore for output, which is incredible. Uh, Ali, I, you are the devil there, and this is tool for developers as well. This is not like for my mom who's gonna like chat with this.

1:53:24Alex VolkovIt's not ChatGPT.

1:53:25Allie LaabsThat's right, yeah.

1:53:25Alex VolkovIt's not like a user consumer technology. Tell us about some of the primitives. I wanna hear you kind of explain what, what null is, for example. Like, you have like a whole new category, but also like new primitives to build with this. Could you please spend the next few minutes, if you don't mind, telling folks how to best use this technology? I would love to hear that. And Francesca, thank you.

1:53:43Allie LaabsYeah. So that's the primitives is the way that we, that we've decided to like, you know, we want calibrated probabilities from the latent, you know, intelligence inside AI. Like, how do then we make that available? And the primitives system is what we came up with for that. So what you actually get from the Gev API, from the TypeSafe API, is, uh, is you, you define the primitives that you are requesting. We they, we also call them questions. They're called questions actually in the schema, right? Um, so think of them as like questions technically, but primitives in like conceptually. Uh, and you see them right there. It's choice, score, and newl.

1:54:24Allie LaabsSo each of these is a specific shape of response that you're going to get. They are each their own type of response. They each return probabilities, but in different sort of ways, and they're appropriate for different kinds of questions, different kinds of information you want to get. So a choice, uh, a choice is really simple. It's choose from a list of discrete options. So you provide it a discrete option space, you describe them each semantically, and the model will pick, we, if we're explaining it like really high level, we'll say the model will pick the answer from that list, but what it actually is doing is it's producing the probability of everyone on that list, uh, and you get all of that back,

1:55:05Allie Laabsright? So in some use cases there, you're always looking for a dominant answer, like there is, they they don't have any overlap, you're going to get one, so you're just going to go with the choice. But sometimes your your use case actually might be really interested if two of the choices are competing for the highest prob- uh pri- uh eh probability, in which case you can actually look into the probability array to, uh, or the probability map to get, uh, all of them individually and do some more advanced, uh, you know, advanced math on that. Um, the score is the other kind. Let's see, do you have it on your screen there?

1:55:39Alex VolkovYeah.

1:55:39Allie LaabsNow I started to read off your screen. The score.

1:55:42Alex VolkovI just wanted to show that, that you talk about probability.

1:55:44Allie LaabsOh yeah, we do.

1:55:44Alex VolkovIn the case of tweets, for example, I have 98, like, assurance that this tweet is about Jeff because Dex here talks about Jeff, but there's other tweets that I'm showing here. There's like a 3% probability that this tweet is talking about this category, 2% about this category, and uh, when choice is basically you guys are telling me based on those probability, you still calculate which one is, if I want the specific one, does this tweet talk about X, you give like a choice as well, in addition to probability. It's very interesting.

1:56:09Allie LaabsExactly. And so score is about getting back a, uh, uh, a continuous value between n 0 and n up to 9, uh, and it is about rating something on an axis. So in this case, it's like a, it's like a semantically described rubric that you are creating, uh, and asking where, what the answer is along that. So you might be asking, like, how, um, you know, how relevant is this to something else? That's actually a bad example, but

1:56:41Alex VolkovI can give you an example. Tell me if

1:56:43Allie LaabsPlease, please give me an example. Save me.

1:56:44Alex VolkovUh, previously, in the beginning of the show, we talked about the pacing the frontier and where, you know, Dariamo is this, there's like, you know, all the way to Jensen and saying, like, let's go. And we also talked about our own kind of like statements. Could we take, uh, transcriptions from this podcast and then rate them on this like scale of like pace all the way to pause AI on to all the way to go, right? This, this will put us as a scale and put every statement that we have categorized this as belonging on that scale, correct?

1:57:14Allie LaabsYes. Yeah, exactly. And, uh, a chat moderation is actually a really great, um, example of using where score really, really fits well, because this is a case where you're not looking at discrete points, you are looking at, like, a spectrum of answer. But the huge power with the score primitive that I think makes really cool, and it can be a little tricky for people to wrap their heads around at first, is that the levels of the score that, uh, you define them semantically, right? So you need a minimum of 2, where it's like between 0 and 1, and maybe you're saying, you know, not at all. Let's say it's chat moderation, right?

1:57:49Alex VolkovYes.

1:57:49Allie LaabsHow, how, like, productive and helpful is this comment to the context of the channel, right? That'd be a great use of a score. And a naive way to do that would be 0, 0 you would define as not relevant, not helpful at all, and 1 would be super, super helpful. But by using more levels in there, you can semantically describe exactly what a 1 and a 2 and a 3 and a 4 mean to you, and you should always use as many score levels as you can reasonably semantically describe. You also don't want so many levels where you're stretching yourself to try to, like, figure out what would be the difference between a 3 and a 4. If you can't describe it, then don't do that.

1:58:26Allie LaabsJust do it the levels that you can describe. And what this does is, it means that if it's just between 1 and 0, and you're asking, we'll get to a newl in a moment, you're just asking, like, is this productive? Um, you kind of wouldn't be certain what a 0.6 means from message to message, but when you semantically describe what a 1 and a 2 and a 3 means, you know that if something scores right at 2, that it's going to fit the criteria that you described for what a 2 is. So it allows you to score things, and then when you score other states, when you score other things with the same defined criteria, they are directly comparable. You can sort, you can threshold, you can create rules around it.

1:59:05Allie LaabsThey can be very, very reliable, um, and, you know, and damn near deterministic, uh, when you run these evaluations.

1:59:12Alex VolkovI, I, uh, let's get to null and then determinism before we end this. I really want to talk about those two things. It's very interesting.

1:59:20Allie LaabsAbsolutely. So, null is our third primitive. Um, we put it last on the list. In some ways, it's the easiest to understand. We're actually going back and forth internally, like, should we describe null first or null last? Null is simply a yes or no question or a state, like a true false statement. So newl gives you the probability that the answer to your question is yes, or that the, or that the statement is true. Um, newls are, are, they're, they're a little tricky because on the surface they sound like it's a score. It's just like a score with two levels, a true and false, and kinda they are. They perform a little bit differently than scores, so it's a kind of thing

1:59:54Allie Laabswhere, like, you should test. Um, but generally my guidance is like, if it, if it really is a yes or no question, if your expected answer space really is binary, like there should be a yes or a no, and something in between would mean, like, uncertain, not enough evidence, something like that, then that's what you would use a newl for. And newls are incredibly powerful. A lot of the demos I've made end up just being a whole bunch of newls all at the same time to, in, in fact, your tweet classifier, um, could be done with a whole bunch of newels, right? Like, is this tweet related to JEV, the latest and greatest System 1 model from TypeSafe?

2:00:34Allie LaabsUm, and that would actually be a way that you would probably want to write it all out, because what if they don't say JEV, what if they say TypeSafe? So you do want to, like, include semantically, like, what the criteria actually is. But you could do that instead of, that would be a case where I would actually want to classify tweets by a whole bunch of newels because a tweet, like many things, can be multiple things. It could be a comment about Jev and what someone ate for breakfast. It might be a 1, 100% that there's talking about.

2:01:01Alex VolkovIt could be a tweet from me saying we're gonna talk about Jev and some other things, so it's not like specifically Jev, right? Like, it's not like I'm talking about Jev only. I'm just announcing that this is a list of other things that we're gonna discuss, so that's

2:01:11Allie LaabsPrecisely. So that would be a case where someone were trying to build like a tweet classifier, and they were like, oh, is it, I'm going to do it as a choice. Is it JEV? Is it breakfast? I'd be like, no, because those aren't mutually exclusive, and if they aren't mutually exclusive, you should be breaking that further into more primitives, and this is why we call them primitives. Basically, one of the first things we do when we help people, when they're like, oh, my, this isn't performing as well as I expected, we're like, let us see the questions, and almost always, like the majority of the time, the answer is, ah, they've, the question is actually like a compound question.

2:01:44Allie LaabsYou're asking about multiple things in the same question. Break that into multiple questions and compose it together in your code.

2:01:50Alex VolkovSo, so questions is, my question is about questions. Those are things that I, as a human programmer that uses your protocol and model, define about what I want to know that your model will tell me, yes or no, or probability of yes or no, probability of like spread probability, right? Like, I just literally, in natural language, tell you, hey, is this tweet about this thing? And then we used to not be able to do this with code, so we wrote about if statements, and then we went to BERT models, whatever, those are not super performant. Then LLM came by, and everybody's like, oh shit, I can use the general intelligence to ask this question in natural language to give an answer.

2:02:27Alex VolkovNow, basically, what you're giving in JEV is, in System 1 models, is, uh, the ability for me in natural language, in semantic space, define what I want, but get, uh, predictable outcomes back. Is that

2:02:40Allie LaabsThat's right.

2:02:41Alex VolkovCorrect?

2:02:41Allie LaabsExactly. Yes. Yeah, you described it. That's exactly what it is, right?

2:02:45Alex VolkovYeah.

2:02:45Allie LaabsYou as the, as the programmer, as the system designer, are defining what the questions you need for your domain space is, and you're defining what all of the, you know, what all of the criteria is.

2:02:56Alex VolkovSo like, yeah, there is a compounding, um, effort here because Gev is not in silo. Ge- Gev is also in the era where Astra exists and Fable exists, and being able to combine those two, System 1 and System 2, where System 2 significantly advanced for the past three years, like it's ab- absolutely incredible what you can ask these models to do with the speed and the price performance of something like System1 that's built specifically for decision making is, that's why I think many people are getting, getting super excited. And thank you so much for, like, telling us about the primitives, how to build this, uh, but also about the general, the general category.

2:03:32Alex VolkovI really appreciate it. I, I do expect that this is not the last time that we're talking about JEV or folks from TypeSafe. So thank you so much for coming up. Uh, really appreciate you being here, uh, jumping on in the last second. So yesterday, in the middle of the crazy lunch that you guys had.

2:03:44Allie LaabsYes.

2:03:45Alex VolkovSo I really appreciate it. Uh, and also I do expect us to tap on you guys more as this develops or as new use cases come up or as folks are discovering this is now the, the next big thing. Thank you so much for, uh, coming on the show. Ali Labs from TypeSafe.

2:03:59Allie LaabsAbsolutely. Thank you for having me.

2:04:00Alex VolkovThank you. All right, folks, we we're nearly the end of our show. I'm gonna bring up back the host to kind of talk, uh, about this. Nisten, at the beginning of the show, I asked you about if you saw about JEV, you were super busy, you went like, no. Uh, quick comments from you on like what, what, what we're seeing? Is this a new paradigm in decision making, in computer use, in whatever else?

2:04:23Nisten TahirajYeah, yeah, I think it bridges the gap between people that are doing like buffer lib, uh, RL stuff and, uh, actual agentic computer use. You do need that level of speed, uh, and it's pretty, it's pretty fun that we can get something like 10x faster or 100x faster, and then we're just using it. We just, just let it play games or let it handle stuff in real time.

2:04:49Alex VolkovWulfam, uh, thoughts on this comments? You've played with this as well?

2:04:53Wolfram RavenwolfYeah, I was super excited when I saw it. And the thing is, if we are comparing to the cost or to the speed, we should also consider that we are comparing to LLMs, so it's much faster than that. But, uh, people have been using classifiers for this for a long time. So the classifier is not a new technology, but you had to train it, and if you wanted to change something, you had to retrain it. And here it's all done in natural language, which makes it so, so easy to modify, and you just define the questions, and if the question provides bad results, you just change the question.

2:05:23Alex VolkovYeah.

2:05:23Wolfram RavenwolfAnd that makes it so, uh, universally applicable. So, so you can put it to Doom, you could put it to news, uh, allocation for ThursdAI, you can do tweets, uh, analysis, you can do all of these things, and you don't have to train a model.

2:05:36Alex VolkovI've been very upset with Gmail not being able to tell tweets from X, like phishing, uh, emails from X that somebody accessed my account, and it's like the domain is not X. I've been very upset with Gmail. Gmail came up with the spam blocking, like, technology back in the 2000s, right? And then I've been very easily seeing my assistants that we talked about casually going to my inbox and saying, no, this is probably spam. Like, they know, they have the intelligence to say this is spam, like, whatever. Uh, I saw somebody using JEV on like 100,000 of their emails and multi-timing them, multiplexing via their API, and just getting a score on like

2:06:15Alex Volkova 100,000 emails with, again, like a cent and a half or something like crazy. The, the, the numbers don't make sense, uh, within like a minute or so. So this is nearly real time, less than 100 milliseconds decision based on every context of whatever. I don't know about context length. I don't know about like stuff like that, uh, but it does feel like, hey, the stuff that LLMs became good for is now possible without writing or creating and running a classifier model at scale, which, you know, many people don't, never will do for themselves. Uh, and that, I think, is a great place to end the show for today. I have a bunch of demos as well, like Fable cooked up, maybe I should show

2:06:53Alex Volkovyou the demos. Fable cooked up like a bunch of demos specifically for ThursdAI, uh, and with Jeff. Let me see if I can pull. them up, um, because I just like, okay, folks, we'll hear about this from the devrel, but like, what are the use cases? Somebody's asking in the comments like, hey, uh, can we, where's this question? Can we use this for, uh, statements from project chats? Yes, you can use this on any type of text. You can feed it any type of, again, we don't know the context length. You can feed any type of text and get a classification. It's not multimodal, so you cannot, like, judge images. Uh, this is not the hot dog or not hot dog from Silicon Valley.

2:07:32Alex VolkovThis is not, like, yet multimodal, uh, but it can be with the pairing of an LLM, right? We can pair this with Muse, take a description of a picture, and then have it, have that text answer. Wolfram?

2:07:44Wolfram RavenwolfJust wanted to, there's another question about specifically for hallucinations, and they have this claim that this model can't hallucinate. And it's important to realize that what it means is the output it generates, which is the JSON, um, that is not, not different like an LLM. If you say make JSON, it may not do so. But, uh, the questions or the answers can be false, of course. So the model has its intelligence, and they benchmark this as well, where it's not as good as Astra, for instance, but it's so much cheaper, so much faster. So as always, you have to decide uh, what level of intelligence you want and how much you want to wait or pay for it, or both.

2:08:19Alex VolkovYeah, 100%. The hallucination and the probabilistic nature. LLMs are not probabilistic even in temperature 0. Uh, they're not deterministic at level temperature 0. Here is a demo that Fable hooked up with Jeff, just to give you guys an idea of like what what's possible. This is like just on the fly. All the news that we talk about, I have like buckets in my head on TLDR, right? We have voice and vision, we have big companies in LMs, we have AI coding agents, we have tools. Sometimes the stuff we talk about co-intermix. What is a computer use? Like, where where does it land? Uh, this is a categorizer or classification of all the news, uh, based on based on

2:08:51Alex VolkovJeff. Google launches Gemini 3 and 8, and I classify this, and it immediately classifies voice and vision. This is like a very simple thing that I would use an LM for, but, but this is, the number of stuff is very well defined by me. I use this in my head all the time. Uh, I have another version with live chat, basically, uh, talking about what of the statements that we had are related to which of the topics that we talked about. So classification, like on the fly, and we can rerun live. You can see this happening within 1.3 seconds and 42 calls. Uh, live producer, I didn't yet implement this, but you can see rebuilding this. So you can see here the chapters that we talked about, the TLDR intros, and then this

2:09:32Alex Volkovis last week's show. This is the TLDR, this is uh DeepSeek that we talked about for the longest time. Here underneath, you guys can see how Jeff scored this. Okay, so like instead of me saying, hey, from here to here, we talked about this, and then here to here is TLDR, you can see that Jev matches nearly perfectly the kind of based on the statements that we made, which chapter are we in, but we can see that here, like, it added some stuff about doomerism before the chapter that we talked about, uh, Cognition 3.2. And I will absolutely now, I didn't have time, will implement this in the live reasoning thing. This all costs cents, less than cents.

2:10:09Wolfram RavenwolfBasically free.

2:10:10Alex VolkovIt's ba- no, the outputs are literally free. They're not charging for output tokens, like, but inputs are just like basically free. 42 dollars per 1 billion tokens. It's insane.

2:10:22Wolfram RavenwolfBro.

2:10:25Alex VolkovI don't know if I've written 42 billion tokens in my life as a human, like, literally typing. I don't know if I got there. Like, I think that, let me say this, I think that Jev can categorize all of my life's work in terms of the text that I outputted as a human with less than 40 dollars.

2:10:43Yam PelegWhat what's the context length if we're already at it?

2:10:45Alex VolkovI don't know. We should have asked, and we will ask Ali if somebody wants to pull up an assistant, context length is very important. Uh, new paradigm, exciting to play with, uh, folks. There's a bunch of stuff that that happened as well. Uh, next week is going to be pretty crazy because OpenAI's release for this week was pushed to next week. Uh, Grok 4.7 is also coming, they said, uh, they there was, you know, Elon said it's gonna come last year, la- last week, but now it's probably pushed, uh, so next week is gonna be very, very crazy. Uh, hopefully folks got excited about JEV as much as we got. Uh, thank you so much for joining, everyone.

2:11:16Alex VolkovWe're way past our 2 hour allotted timeline. Uh, this is the world of Fable and Astra and all of the tools that we're living in. I I rebuilt completely the Descript app that we use to edit the show with Thursday AI specific work. So what you're seeing here is a, like, 0 to 1, not 0 shot, because I worked for this, and I think I spent like 3 billion tokens of LLMs, by the way. I can't imagine how ma- how, how, how much faster Jeff will make this in accordance with LLMs, that we have the whole episode with a timeline and every an LLM or multiple LLMs can suggest things to cut because these are like live only.

2:11:54Alex VolkovCan suggest that where Yam is speaking or Nisten said, hey, this is Chris Alexiuk. This is, this comes directly from our things. This is like a, um, an editor. We also have clips for the clips factory that can play. Uh, there's really a bunch of stuff there that I cooked. The reason for this is that my agents want to help me with cutting, but I want to give my co-hosts, their agents, access to tell me, hey, what is that one clip that I said? I want to pull this for me. So, like, I'm building that for this ThursdAI editor, and it's nearly ready, uh, and uh, Jev, Jev is going to get there and decide, you know, for every um and ah, for example. Yeah, we're gonna build Jev into this, and it's gonna be amazing.

2:12:33Yam PelegBro, you can sell this.

2:12:34Alex VolkovI-

2:12:35Yam PelegYou, you can sell this. It's insanely good.

2:12:38Alex VolkovThank you. Uh, we'll see, we'll see if it works when it works, but yeah, it's, it's, it's pretty dope. Uh, all right, folks.

2:12:44Wolfram RavenwolfThe best way to show what.

2:12:48Alex VolkovYep. Wolfram, we lost you there for just a second, it looks like. Um, but I think it's a great way to end the show, so I'd be able to go and use this. Uh, Wolfram, what did you say? You, we lost you, now you're back.

2:12:59Wolfram RavenwolfOh, I wanted to say this is the perfect example what we are preaching here, that you are showing the proof that is not just a demo you do and something, but you use this technology we are talking about all the time to build something really useful and valuable to yourself and everybody who has the same problem.

2:13:17Alex VolkovAnd I cannot tell you how much JEV makes me excited because I literally can come through every sentence and score it based on multiple parameters, whether or not this makes sense to the topic that we're discussing, whether or not this is a somebody went on a, you know, on a debate and came back, whether or not it's gonna be viral. There's like so many things that Jeff can like just judge and process this in a second that I cannot wait to get off the street and actually go and and code this, but uh, but if you have missed any part of the show, we had an incredible show. We had David Paulon in the Assistant Benchmark talking about different assistants

2:13:50Alex Volkovand how he's scoring them. We had Ali Lab from TypeSafe talking about this, like, new and exciting world of AI intelligence that's not LLM based. They call them System 1 thinking, and they're incredibly cheap and super fast, and many of the things that we started using LLMs for, but LLMs are not fitted for. We also had, uh, Francesco from Kua talk about how computer use is like a very important thing, both in in both these worlds, and and the TypeSafe JEV, uh, very much helps the agents to get better at the tasks that you want the agents to get better at, at proactivity, at different things, at prioritizing a a bunch of other stuff. So, very exciting show. This feels like a monumental week.

2:14:31Alex VolkovSo, releases from Astra to whatever, they're big and incredible. This feels like a monumental week in terms of like what we can do, in terms of how big AI's assistants are getting, how fast JEV is getting, and how big computer use is getting as well. Folks, thank you so much for joining us. Again, if you missed any part of the show, Thursday AI is live. Uh, you can rewatch everything on thursdAI.live, by the way. Like, once we get off, you can rewatch the whole live stream right here, uh, and see the transcripts and everything, but also you are able to, uh, get it as a podcast, as a newsletter. Please subscribe, uh, and give us five stars wherever you're listening.

2:15:00Alex VolkovPlease tell your friends that if this is informative for you, maybe they can enjoy this as well, um, because we really try really hard to bring you the best possible show, uh, and the earliest news, so you'll stay ahead of the game. I think we're doing a great job at that. Uh, so thank you, co-hosts Wolfram Ravenwolf, Nisten, LDJ, and Yam Peleg for joining us. Peter Gosta was here before, and our 3 guests from today as well. And, uh, most of all, thank you for tuning in, listening, and if you're still here, uh, uh, give us a shout out in the chat. Thank you, folks. Bye-bye.