ThursdAI - Oct 30 - From ASI in a Decade to Home Humanoids: MiniMax M2's Speed Demon, OpenAI's Bold Roadmap, and 2026 Robot Revolution

Whoa.

Haha. That was close. Hey everyone. Welcome to the Halloween edition of Thursday I for October 30th. My name is Alex Volkov. I'm an AI evangelist with Weights, & Biases, and I am very excited to bring you the show this week because this week there's been just an incredible amount of updates around all aspects of ai from open source to. OpenAI restructuring to scary fast agents. We're gonna have a very spooky, scary Halloween, including the potential, addition of robots in your house with the new humanoid robots that we saw that was announced. And so to help me cover this, we have our friends here. Let me add them to the stage. Happy Halloween. What's happening? You look great, Alex. Wow. Thank you. I've been working on my looking great. So I appreciate you calling. Yes. there's been, today was my daughter's, and son's, Halloween parade. And so I went and joined that and I had to look apart and then decided, this is Halloween. This is the third Halloween we're doing, by the way. so we're gonna have a replacement. Alex, at some point actually deliver some of the news to you. Hopefully that will be ready in a minute. but meanwhile, wanted to chime in with you folks and ask, first of all, how are you? But also I think the most important thing is what is the one thing? One thing, and I insist yam you. I know you have many things always, about AI this week that we must absolutely talk about that blew your mind and people cannot, ignore, I would say yam first and then in the run,

MinMax drop a bomb.

Like seriously an open weight model that is. I know the best we saw and, quite interesting architecture actually. I have

Halloween surprise for you.

Let's go.

we have in the studio with us, Skylar, who is working

on Minimax from Hulu, really. And we're gonna ask Skylar in a minute, folks. so please stay with us, to have a conversation about Minimax because I think it is a banker. It is a big one for sure. we're gonna have to do the TLDR first, so folks, you're gonna have to wait with us. but Skylar, is gonna join us in just a minute, to talk about Minimax, from Lio. I always say Hallelu and I think I'm saying this incorrectly. Ryan, what about you? what's your one main big update this week?

cursor 2.0 is pretty big deal.

yes, I love amp. it's fascinating. And I think the more meta thing that's interesting is their composer model. I think we're seeing the rise of Agent Labs ship their own models, and I think this is a good thing. I cheer 'em on, even though we compete, I'm

cheering 'em on.

Yeah. And, cursor 2.0 launch this week, including a full rebrand. Not a rebrand, but reevaluation of the AI agent wise, that now Cursor and Atlas look the same. I found it very interesting. The cursor now is like the big prompt and say, Hey, what do you wanna do? And basically like multiple agents running. yeah. my big thing this week, by the way, was the OpenAI live stream with Sam Altman and Jakub and, voyage, who was also a co-founder of Voicage. And I think that live stream flew above people's heads. Folks, they dropped so many, just bombs in this live stream. And because it was live, it wasn't on YouTube, it wasn't transcribed, there was so many things I went outta my way to find a way to download that live stream It is really good at not making that possible, transcribed it. And I will bring some of the updates here to you. One of them is their chief scientist, Jacob, who replaced Ilia after Ilia party. For SSI, thinks that a SI is achievable within a decade. A SI, not a GI. they also talked about an autonomous. Research engineer by Marshall 2028, fully autonomous research engineer, the likes of which OpenAI pace, north of a half a million dollars to a year to just do their stuff. So basically they dropped a lot of stuff. Oh, Sam Alman talk about $1.4 trillion commitment to energy. And I was like, where's the other 400 billion come from? we accounted only, 1 trillion so far with starlink and the Nvidia deal. a lot of stuff in there, just a lot of stuff in there was a dense live stream. They also answered a bunch of questions, that they took from the audience. this was probably my highlight of the week because, it's insane. Obviously opening, also finished their recapitalization, the non-profit, which is profit. this was mine. IY Fox, time for A-T-L-D-R. And then we're gonna get to a conversation about Minimax M two with Skyler Meow from the Minimax team, and, a bunch of other Halloween stuff. Just a lot of stuff. Humanoids self-driving car. There's, there's, so, there's so much every week folks. Let's go to TLDR.

All right, here's it. Here's the TLDR R for ThursdAI, for this week, here's everything we're gonna talk about, and we're gonna run through this super quick so you know, all of the stuff that, we're gonna mention and then, you'll be up to date even if you don't follow the rest of the show where we're gonna have a lot of folks. All righty. let's start with open source LLMs folks. Obviously Minimax M two from Lio ai, is the new King Ryan. I don't know when was the last time we had the new open source king, but it is here right now. We go and, it's a agen model, 8% of CLOs price with twice the speed and the same pretty much performance in open source. This is crazy. This is like son at home, pretty much. OpenAI, folks OpenAI. Brought us Another open source model. This is crazy. so OpenAI open sourced GPT, OSS safeguard. And this one is, the first open weight safety reasoning from OpenAI from classification. there has been other, guardrail models, but OpenAI has fine tuned their GPT OSS models for safety and specific things, so that's great. we also had IBM, granite. And granite is like a whole series of models, And this series of models has been great. these ones are nano. These ones like, tiny models. I think one of them is like 300 million and artificial analysis. Tested them. One of them is 350 million, When ton's gonna get here. He's gonna tell you this runs on a toaster. This runs on an AirPod. This runs everywhere. so we're gonna cover that a little bit. Ming Flash released, Ming is the third model from, from these folks. we covered ring before and link before, so this is getting interesting. this one is a nominee model. This one, like only model, like understands video, understands the images on exit. And I think this one also, does image outputs as well. So that's gonna be very interesting. Eh, the first, for the longest time, we will cover robotics and focus. this is it. This is where a ton of people now are, on our timelines getting this robot. This is a humanoid robot. We, I think we covered this previously. It's a very tiny one. It weighs only like 70, 66 pounds. it's one X neo and the pre-order is available right now. You can get this humanoid robot in your home probably by March. as far as they promised. They had a keynote, they had a bunch of people. It's tele operated. There's a whole conversation that we have to do there. this is the spooky stuff on Halloween. this is the robot that in then. Okay, let me, zoom in here. Let me zoom in and talk to you directly, folks. Here's my prediction. within a year from this Halloween, next Halloween, we'll have a bunch of these robots and robots from the other companies dressed up as Halloween puppets, like helping folks, in their house. I'm hoping that I will also buy them, have one of my own. so we're gonna chat about the one X Neo and what this means to be able to pre-order human robot, to your home in the end of 2025. Ryan, you wanna help me cover this, big companies, segment.

so the OpenAI is like fascinating because everyone was up

in arms about, is it a not-for-profit? Is it for-profit? How does this work? and they recap the company and change the structure. And here comes ASI so fascinating and it's gonna be a very valuable company. So here we go. And all of our friends, OpenAI are very happy about their options.

Yes, sir.

and next up

Cursor 2.0 is a big deal. and the new composer model is very exciting to see the agent labs, training, releasing their own models, getting us. out from underneath the boot of, the big labs. So I think this is, and I've tried it. it's very fast and honestly it's good. excited about that. And then we saw, cognition come right in and punch right back. SWE 1.5 and it looks very good. So this is good for all of us, y'all. We want this competition. We don't want to be, restricted to the big smart models from big Labs.

the higher there for me was both of these releases.

Same day yesterday, curer started the day with Cursor 2.0, and then Composer and their own model for the first time. I think they releasing their own model and then boom, we, cognition, probably not sure if they scheduled it for that day. And, they probably rush their market team and say, Hey, we're ready. Let's go. Let's freeze this. And it's powered by Cerebral, so it's like super duper fast. In this week's buzz, the category where I talk about everything, Weights, & Biases, and CoreWeave related, I only had these notes because there's some announcement from today that we haven't talked about. obviously fully connected in London is coming up next week, the one in Japan is happening. tomorrow. So if you're in Japan, hey. Hi. Hello. but, if you're going to London, please join us in fully connected, fully connected.com. And, if you need tickets, please let me know. We're also going to a AI engineer code Summit in ny and we're sponsoring, and we would love to see you there. Come to our booth. And the news that we buried and we, we now can talk about, CoreWeave has a announced another acquisition of Marmo Notebooks. This is a new, super exciting company. I'm looking forward to tell you all about this, CoreWeave, acquires marmo. If you've noticed, we're announcing around one acquisition a month lately, so that's super, super exciting because we're building out this essential cloud for ai. And I would love to tell you about this vision. And video has not been quiet. Ody, CML launched the V two A real time interactive AI video. And when I say real time interactive, I mean it like you can go in there and we'll definitely do this demo. You go in there, you basically, type whatever you want in the prompt. The video starts streaming and then you can live, change the video. The video will adjust live to whatever you typing. It's nuts. Honestly. It's nuts. I recorded the video on this and I will definitely show this here as well. we have updates from Sora. Sora character Cameos is here. Leaderboards are here as well. Character cameras basically allow you to add your pets, which I prompted it. and also whatever pumpkin your kids carved, you can add that as a cameo so then everybody else can use this as a character. It's really cool as a concept.

how is your SOA influencer campaign going?

I'm working on this.

I think I'm mentioning over 3000 followers. Some people tag me. I'll give an update after the TDR, but also SOA is invite free. That's it. You can just download Sora if you waited for an invite. we gave out invites here about Sore is invite free in US and Japan. And it's growing, like crazy. And also from, the same folks from Minimax, they are known for their, Lio, I think this is how we pronounce it. We're gonna ask Sclar in a second. Uh, LIO, released Helio 2.3, which is an update to their video model, cinematic professional grade, and the fast version as well. they're all over the place. so I'm very happy to have Sclar here because Minimax speech also launched and it's, 2.6. I've, for the longest time, this to me was the best hidden kind of speech model. the one that closest matches my audio and the one that like is cheaper. so they're presenting ultra human voice ai. I need to clone my voice once again to test it. and I think the biggest news not the biggest news, but like definitely, the biggest news Cartesia launched Sonic Three including a hundred million dollar investment. We talked to you about Cartesia back in June of last year. Cartesia is gonna join us and talk to us how the hell they're getting, their voice generation in 50 milliseconds. 50 milli. This is faster than you blink. I think they say 49. Once you click, the voice starts streaming. It's really crazy. And so we're gonna chat about this for sure, because I think it's crazy and it sounds really good as well. I had the, they gave me some credits to like voice clone my voice with their pro version. And it's nuts, honestly. It's nuts. I think that's pretty much it. Invoicing audio, and we have some tools. pokey, we told you about pokey last week with their research agent. They launched something called the Genetic Workflow Builder. And, this tool from Google is really cool. It's like an ad creator for your, it's called Ply AI Marketing Agent Targets, brands Analysis and Campaigns. So you send it to a website and it analyzes the website and it's pretty cool. Folks, this is it. I think everything. Anything else here or should we just jump? All right, let's jump to open source, our favorite corner, and plus an interview with Skylar from minimax. Let's go. Open source.

And for that, we have Skyler here, from the Minimax team. Skyler, how about you introduce yourself first to the folks, and then we're gonna talk about the announcement that you guys released this week, at least one of them.

Hi Alex.

And hi everyone. I'm Sky Meow. I'm the head of engineering at Minimax. our team in charge of the product development like, MIAX agents. Hello Video and Miax audio. Yeah. Also API platform. And, what's more, we are also, took part in the M two post training, especially in the code and the agent's, capabilities. it's really nice to meet you actually I'm a typical engineer, being on a podcast is a very new thing for me, first time for me. Thanks. You're gonna do great, man.

I promise you're gonna go do great.

I already hear that you're gonna do, okay. So first all I congrats on the releases. There's three releases on our weekly update show. So all of the three that you guys released was released this week. we're gonna obviously talk about minimax two because we're in the open source LLMs, kinda segment of our show. But it's great to hear from you that you also involved in the video model, because we've covered the video model multiple times, and the update there was released, I think, 2.3, And we also Minimax speech, which I used for a long time. first question for you is, how do you pronounce the company name? Lio. Julio. Like what? What's the actual pronunciation?

Minimax is our company name.

Actually, so what is Helio or Julio?

Is that just like a product?

I would pronounce is as a

HAI means C, and the law means share in Chinese.

I see.

Okay. So Halo, and that is a brand thing, but you guys are going by Minimax and, M two is the main announcement. could you tell us about what is M two and why, are we excited about this?

okay.

as you have mentioned, M two has made the top five model Yeah. So it suppress, Opus 4.1, and it is just a little behind the sonet, 4.5. And, from the LM Arena, they will release soon. Our performance on web dev is, less the sonet 4.5. Maybe they will release, but actually it's just 8% price of the, sonet and twice speed and, it is all open sourced. and actually, you may know, it's now, the fifth most called model on the open looter today. And, actually it's the first in the hack and face training and why we build this model, I think because, we used a lot of agent in our company. As well. We are not only building the AI model, we also want to build a AI native company. And, actually we can't live, with, agent in our job today, even our HR team, they are sourcing, potential candidates with agents. And, we find, what we call is there is an impossible triangle in the, in available models. it's performance, cost, and speed. you can maybe have only two, for example, SONET, experience, but it is, too expensive to common people to afford it. but for the other models like, open source models, they may be cheaper, but their performance wasn't there, especially in our internal use. It's industry requirements. So we guess they should be a model can, resolve this problems. Then we trained M two. I don't know if this, answer your questions.

Yeah, no, this absolutely answers my question.

This is great. the thing that, we first of all congratulate you guys released this is Apache two, right? Apache two license open source.

it is all open source and the license is

Yeah.

Yeah.

so congratulations on topping the open source,

leaderboards in the model. this is absolutely stunning. Like we're seeing the Thank you. I think it's fifth overall on artificial analysis, right? yeah. Just behind cloud four point size sonnet as well. Artificial analysis does this intelligence index, which like has 10 evals, like ML, pro GPQA diamond, humanities, last exam and even au bench. And you guys have all of these things in the evals here that you posted. and you were in charge of some of the post training for this. And this is not your first model, right? Minimax and one obviously was there also was a great model. could you talk about a little bit of like what went into achieving, this, the size? we should probably mention the size, like it's only 200 billion parameters, so it's smaller by three times than deep CR one, for example. it's smaller than the bigger Qwen, but also it's an MOE and only 10 billion active. is that correct?

10 billion active, we designed it.

Yeah.

So this is a very unique design that we see, like a high ratio. Could you just speak about the MOE approach and, how you're achieving this much with only 200 billion parameters?

Yeah.

Okay. First of all, I'm an engineer, not a researcher, so I'm not expert. I can just, talk about this in my side. Yeah. at the, very beginning, we designed M two, we think we could achieve the, performance of Sonet with, 10% of its, prior actually we did it. that's what was our insight. That's what our algorithm team, discussed and the research for a long time. we think all the models on the market, we think it's not, the best for the agent tasks because, not just the price, even the speed, as you all know, typical agent task may be around for 20, 30, maybe more rounds. So if the TPS is low, it would take a long time. And it will lower the efficiency. you'll have to wait for a long time for the job to be done. we do a lot of experience in our pre-train, compared to M1, we do a lot of job. We almost, reconstruct the pre-train data, to make it higher quality and, even more data compared to M1. so I think this may be the, most important point. one of the most important point, the quality of the pre-train. we see that it's, on, just used, I think half the data, its performance in our internal benchmark has exceeded M1. it's very important. the second one is, our post trend. for example, in the web dev, we deeply involve our senior from, frontier, engineer. And we design the task, design, the data design, how we produce the thinking content to combine it. And, we also design the, reinforcement, learning reward. overall I think it's a really important insight That's, we think we could find a good, a sweet spot in the triangle. that's, we can get brilliant performance, but with, a lower cost and, a fast speed and we actually made it. Let's talk about what I want mention more is, I think it's, about our vision, of our company. Our slogan is intelligence With everyone. We want, our model can be, accessible and affordable for other people, not just a very expensive one.

This is great, dude.

Thank you so much for covering this. And like we, we obviously are very excited about, you guys are committment to open source, at least in the LLMs, the video stuff, I think it's proprietary. and then the voice stuff is definitely not, I haven't seen open source, but the LLM stuff is an open source. It's an open weight model, a hundred TPS throughput on your servers. and it's only, yeah, it's possible to host it on something four H one hundreds. I saw, at FPA Precision. let's talk about context length. if you can address this, it's 200,000 tokens per prompt and previously, like the big thing about M1, I think was 1 million context length. Could you speak as to why was that like a conscious decision

Okay.

I think the, most important, topic on the question is linear attention. and the context dance. I think first of all, in the aging task, we found, we can use memory to, avoid the, context problem, just like sonic do. And the performance will be the most important thing. and I think we will extend to 1 million contacts, in the, future, but, we choose to, get a bad performance first. And, it's about the choice between lightning and attention, linear attention, and full attention. And, we have a tech blog about it. in short of it is that, in the small size experiment you can, hard to find some, minor but important, problems for the linear attention. So as the computer power is going up, we think the most important thing is give, people, best performance and, cost efficiency model.

Nisten, go ahead.

Yeah.

we've seen with the mixture of expert models, it allows for continual training as you go along. And with code data, as you get better and better code, it often becomes nicer to continually train the model. Is that something you guys are planning to continue doing? Because this is, for example, what we saw from, deep seek and so on, and do you find it easier with the mixture of experts, models versus the Moes to do this? Or is it a lot harder? I think people find training Moes quite a bit harder, these days have those been figured out? Yeah.

Yeah.

let me make sure I understand your questions. are you asking, the training, MOE model? Is it, difficult

we've seen companies that they release an MOE and, it makes it

easier sometimes to continue to release updated, versions or fixes as you go on. do you find this architecture good for that, and do you plan to do that?

first off, MOEI think it's a metro architect for us, actually.

We have, trend, MOE architects, since almost two years ago. So I think it's a mature, acted for us and, not, our infrastructure and our team. were really familiar with it and, we were going to continue, training, the MOE. we think it will enhance the performance of the model, especially if you do some things like, reinforcement learning. I do think it will, improve the performance in the field. You want to improve. Actually, we are working hard on, M 2.1 version and, we hope to do this in a few months.

Speaking of the next model, Skylar, the way that I found you, you

were very active on Twitter while people tried the M two model and, I saw in a bunch of threads being very active. So first of all, the community always great to have folks like you just from actual the engineering team going and helping people thank you and seeing what they do. and also you guys released, support for both, the Tropic API and also the OpenAI. completions, API, but not responses, API. And I started talking about the difference between the two, and I started talk about the difference in the context of how the model actually behaves in between those different APIs. Could you speak about this a little bit? 'cause I think some people, may have tried them to, and that they notice that the performance is not that great because of the different responses. Could you speak about this to clarify, this for folks who do wanna try, what's the best way to try this model?

Okay.

it's a complex situation. the problem is simple. That's, I think, people gets to know inter lead the thinking from Sonet four is, the most important feature of this model, and it enhance a lot of its performance compared to, SONET 3.7. And it means, every time, the model responds with a tool core, it also with a thinking. So it is somehow the influence at time scaling to use more test time, computer, powers. it actually important for the multi wrong reasoning and the complex tasks.

the model itself at the next step?

Yeah. when you ask it, it's important for it to receive what the user ask. Plus it's thinking tokens themselves, right? The reasoning tokens, yeah.

yeah.

like the model search first, and then it can think, from the result of the searching and, decide what's to do next. can think or it can see what it has been thinking. it will significantly, influence the performance of the model across all the, thinking models, including sonet, including GP five. I think that's the base of the, all the discussing. But, now in the community, the apps, didn't support the thinking. as we found out this week, and we are surprised, a lot of coding tool, like low code kilo codes, they improvement the API without syncing pass by. So I think they don't really use the, sonics file five points, all performance. we have talked with them and they will, support it. so that's why I, talked. many times, on X, that's, inter thinking is very important. We like mine and Slack. A p 'cause open. IAPI, the chat, chat completion. API, didn't support, don't support, thinking pass back. Yeah, it only support in a responsive API, but the community, a lot of apps don't support it.

so the best way to test it out right now is.

Via API that passes back the thinking tokens. And you said the tropic API that you guys also support, which we saw also from the folks at Kimi. we see more and more, labs releasing open source models, but also hosting on their servers and also supporting the entropic API. and it seems for agent stuff, it's a little bit better and it doesn't seem like anybody's picking up the responses API from OpenAI yet. They do pick up the completions API, which standard, but not that great. Skylar, I wanna say, thank you so much for joining us. we need to move on, but we really appreciate you guys releases. Just to make sure that I'm not missing two other releases that you guys also have. The video model was updated with like high realism in the fast model as well. and that looks great. So we're gonna talk about this at the end of the show as well, and the audio model. Anything important you wanna say about one of those two? I'm assuming that's also a crazy engineering challenge to serve a video model at scale and also an audio model. Anything important you wanna mention about those two releases before we let you go?

Yeah, I think you have mentioned the most important things,

hello Video is best, in this version, in the, dynamic and, complex, realistic videos creation and especially the complex movements and the different visual styles and, the Mixtral expressions. and, we will also list, several, versions I think, in the future. But I can talk more and I think we will have big versions, later. And, I think that's all I can say. And the speech model, I think it is actually the best, speech model now and, a lot of our customer say that. And, I think you have mentioned, or I want to say the, latency and the ultra human, voice.

Skylar, thank you so much for joining us.

Should be aware, of mini Maxim two should give it a try. And thank you for sharing the best practice with us. We look forward to invite you guys back to the show once you guys release some more updates as well. you guys are now considered friends of the pod. Thank you for open sourcing as well. Cheers.

Thank you.

Thank you all.

righty, folks.

Now we're gonna move on and we're gonna skip some of the open source stuff super quick, but first reactions to M two. from the team here. Reactions to what Skylar just talked to us about.

The benchmarks are, they're pretty out there.

Yeah.

that, that's like the perfect architecture for something that

you can run in a single H 200 in four bit. you'd probably, run maybe like in five bid or something. it would be like the fastest, most intelligent thing that you could run on that GPU as a single instance because it only has 10 billion active parameters. You might even get 15 tokens per second on the CPU.

Dude, we haven't seen this ratio that much.

200 billion, model to tell the parameters.

Alex, do we need this on Weave?

yes.

But it's a very good, architecture.

I have a feeling in the beginning of this week

that this is gonna be a banger. And I was right. I was very happy to see Skylar come in here and talk to us about this. Ryan, any comments from the AMP team about these open source model? I know you guys are very selective about your models. Do you guys test it? Did you test it? what's going on?

Yes.

All of us Agent Labs are very excited about these models, right? Because, we are over the age of where you have one model selector for agents, that is irrelevant. Now what is happening at the cutting edge is that every agent lab is choosing specific models, even for specific sub-agents. and so we use several of these. This is, again, putting the power back in the hands of the devs, right? so very excited about what's going on. We're just gonna see more and more of this. We're eventually gonna see these models, fine tuned and specified for specific frameworks and languages and workflows. so this is a big deal and this is good for all of us.

A hundred percent.

And while we are on the topic of open source, we now have breaking news that I just noticed just now. So I definitely wanna, hit this announcement super quick. AI breaking news coming at you only on Thursday. Our breaking news right now is that Kimi, moonshot AI Kimi that is very famous for their like, unique model called Kimmi K two, just with a new model. And this model is called kimmi Linear, five one. It's a 48 billion parameter, 3 billion parameter, MOE. So they call it kimmi Linear. They did some, some adjustments there. but achieves 51% performance with similar speed in full attention. so significant speedups at long sequence length. So it looks like they're innovating at the sequence length, area. Contact lengths of 1 million parameters. so qmi linear release, the base model, and then instruct model with only 3 billion active parameters. This looks, this one kind of sick. key features delta attention. This is I think, a new one. Qmi delta. Attention. KDA superior performance outperforms full attention and variety of tasks, including long context and a rail style benchmarks.

20 pages on archive at the middle on unlike linear attention.

We don't have time to read, we don't have time to read. I want to review this, but I'm like, oh, wait. That's complex. Wait, this is a lot. You went for gated internet and okay, I wanna know what I'm reviewing, but really cool. Seriously? Yeah. really cool. This

literally just dropped, five minutes ago, 10 minutes ago.

we don't have time to talk about this because this week was insane and we're in the business of covering the whole week and not just releases from this day, even just yesterday was like five releases. unfortunately, as much as we're super excited, we'll have to test this out and talk to you guys about this model tomorrow. Maybe we'll bring some Kimi folks, to talk about this, next week, but 1 million contact length with 3 billion parameters. Go ahead.

Do they have any benchmarks?

can we quickly see, I haven't seen posted

benchmarks Okay.

I, we'll

see.

Oh, that's just a seller. Okay. This might be crazy. Yeah, we'll

so this is breaking news from right now folks.

And then we're back to our, updates about open source. Maxim two was the biggest one in open source for sure. the other big one was OpenAI is open source folks, did you see this? The OpenAI continued open sourcing. OpenAI dropped GPT, OSS Safeguard, which is a Finetune of the GPT OSS models, 120 billion parameter and 20 billion parameter. they have focused their models on safeguard and guardrails, right? So if you want to use, custom policies for your safeguards for your LMS to be safe, previously you had to build specific safeguard models that were fast enough to do some of these tasks with specific, policies in mind. What they're claiming is because of these models, you can specify policies on the fly. You don't have to train models for these policies, and they show a significant performance boost on their internal moderation eval. Now we should mention that open

Sorry I keep interrupt, but this is so useful

if, if you have actual customers. No, you can just fire up a Chinese model and put this like little tiny 20 B with two B active parameters super fast too. It'll make it safe. I OpenAI to filter it.

I, I'm building a legal AI assistant, and this is gonna be

killer, to verify citations and results

If there's one thing that hope AI knows how

to do is how to censor stuff.

If there's one thing the opener knows how to

do well is to censor stuff. And I think I even read on their, blog post about this, that they trained, SOA is very good at censoring, different outputs, et cetera. The models that are running are these models. So basically this OSS safeguard has a policy iteration. It does chain of thought, and then it makes a policy decision based on the chain of thought. And then you can apply this policy. But look at these stats, folks. Look at these stats. these models are fine tuned on there. This is an internal moderation eval. So open, I didn't share this eval with us, for obvious reasons. they don't want to tell us which type of stuff they're evaluating, but on their internal stuff, their safety reason or which they have, they don't expose to us. I think it's exposed at the, API level is getting like 52%, so getting half of the coverage of their eval. But the bigger one gets 46% on this benchmark. Whereas the regular GPT OSS, the base model, this is fine tuned from, gets only 32. So this is like a very significant jump in the ability to do moderation. You get, flexible guardrails, obviously a plug here, with Weights, & Biases Weave. We offer you to build your systems full of evals as much as possible, right? we also have the ability to give you guardrails. we haven't connected the dots yet between the inference service and the guardrails, but we will definitely, and these are the type of models that we're definitely gonna use because these are, great models, absolutely great models. this is very important for actual builders, of folks who are building, and not just like coding with ai, who actually put stuff on production and care about customers. so shout out to OpenAI for this, and hey, who would've, th the OpenAI is gonna be back on hug and face that soon after, just like releasing these two models.

presume, people can get this served from CoreWeave,

Not today, but, this presumption is generally

the right presumption. I'm gonna work really hard for us to test these models, to also put them on the waiting devices, inference with CoreWeave because I think this is very important, but also because we have a guardrail product. So we will integrate those and we'll test them out. super quick updates in the open source. before we move on. IBM Granite 4.0, nano, I don't know, Nisten, I thought you would be very excited about these models. Granite 4.0, 350 million. Beats Gemma three 1 billion in artificial analysis just by a little bit. So this is a model three times the, less the size, like one third, the size of the 1 billion. I think it's crazy of how tiny these models get. four is 1 billion beats, LFM and Qwen three, almost beats three. But yeah, these are tiny models. We have covered granite for a long time. They is a bunch of models. We don't have time to dive into them Super quick. If

you have a, just check out the Enova or Joshes

from hugging face, because these are the ideal ones to put on web. GPU on the web as well. And he has a whole bunch of pages. So just XE, NOVA on Twitter.

Alright folks, so we have to move on for open source.

this segment is gonna talk about robotics folks and this week, one x the company that's been in our radar for a while. I think we mentioned they have this like little humanoid robot that only weights 66 pounds. I think one of my kids weighed 66 pounds, right? Fucking Tesla hu the Optimus, when you walk to it, it's your height and size and probably can beat you up if it was d dextrous one X, you can pick up and throw in the garbage if it, if it misbehaves. This is the whole point. This robot, this human robot, was announced with a wait list that you can purchase right now and it only cost like $20,000 and they promised to have it in your home By what? Next year? Somewhere in March. That's insane. obviously we talk about humanoids are coming folks, but I, have you seen the announcement, this robot again, 66 pounds weight, of five feet, six inches. Works on four hours, walks itself to the charging point and charges itself. Like the video shows him, grabs the charging pack and attaches it to its belt. And it has a very soft body to actually focus on the fear that people have from robots. That it's not gonna fall on you, on your kids or your pet or whatever. And it's very easy to pick up and beat up if you really want to because it's not gonna I go all our robot on you with their own AI model. And I think the kicker that bothers a lot of people in this release is that, they have mentioned that they gonna have a fleet of tele operator folks, managing this robot. Four more advanced tasks. So stuff like, hey. Please clean my dishwasher, put the dishes in my dishwasher, put the dishwasher dishes that are clean in my thing. they mentioned that some of these tasks are gonna be like the more complex task and they will be tele operator, which means a human somewhere will put on VR goggles and headsets and controllers and will tell it port inside your robot when you want 'em to, and then we'll do those tasks for you. they do some of the privacy stuff. They blur out people's faces. They only do it on demand. So you will decide when the stellar operator, thing comes, but humanoid $20,000 or $500 a month, rent in your home. Tell operator for the task. What do we think guys who already purchased this thing? What do we think?

So I've already had a big conversation with

my wife about this morning and

yes

and it's interesting 'cause I think I see a

exciting business opportunity. 'cause imagine you buy a fleet of these things and you start a business that looks after people's pets while they're on holiday and waters their plants. you could deploy this thing, it's probably got a lifetime, I hope, of 15 years. So you're talking a little more than a grand a year, for the asset. And honestly, I think this is gonna be huge. Like it, I think there's a lot of people like, I want my plants watered and I want my pets not to die.

The, I think, watering plants is one of the

definite, things they showed. They also showed that the gloves they have are plasticy, so it can like, it wash itself, whatever. I don't know about the sweater thing, but, 'cause that's probably gonna get dirty. But I can absolutely see the pet sitter scenario for sure. But I want more than this dude. I want my laundry folded while I'm here talking to you guys actively. And it seems like I will have to pay them to pay somebody else to actually perform this via VR versus paying somebody who comes to my house, like cleaners that I have. so that's a very interesting thing. I would just say a year ago as a single, AI forward tech person, I would absolutely click that buy button. Like I would smash that buy button. But now that I'm not alone in this house, there's privacy concerns, et cetera. And they're real, for many folks, they're real. They don't want like a random person teleporting into the house. Nisten, what do you think? Are you buying one

look of all people Who's worried about privacy is me.

And I think people are thinking of this wrong. This is like a telepresence device. if I can't go somewhere, can I just sending this thing and have it drag around, a luggage of battery Also, if you have, if you have parents, grandparents to, to take care of, like you are already into their system and stuff and they can't do something or they can't hook something up and then you can just go in remotely through that. I think it's huge. I think people are just focusing wrongly on the, it's rightful, but I don't think that's the main business case. the main business case is, we're gonna have more seniors than younger people to take care of. Like now you can actually go and do that and fix their laptop and hook up the, I don't know, dishwasher, maybe the. It's, this is huge, like $20,000 just for that. and it's already on sale. This is a big, that's a big deal achievement. And, I like the fabric. I think the cats are gonna love it.

Yeah.

the fabric is dope, don't get me wrong. people walking Caty. they're taking the bet that people don't want the futuristic, stylized humanoid that's like sleek metal. Their whole UI is like wooden base, whatever. it's really nice. LDJ, you're joining us as well. have you ordered one for yourself?

I have not yet.

I think I may wait few generations, but I have a friend that said he is actually gonna buy it. He actually. Said if it gets around 20 K and made by a US company, and it looks pretty decent, I'll get it. and I said, Hey, you said it. You said you're gonna get it. Are you gonna get it? And he said, yeah, he is. He is gonna pull the trigger. Yeah. So yeah, I think it'll be really fun. And when I visit some friends in San Francisco, I'm sure I'll see a bunch of people's houses with it.

I can't wait for one of those to just walk with me out of a Uber in San

Francisco while I'm wearing a Vision Pro. And the Uber is actually a very mode dress itself. Like San Francisco's definitely the city where everything's happened. We're gonna see a bunch of those in offices. We're actually probably gonna buy one in our office. Honestly, I think we've bought, enough robotic dogs at the worth of one humanoid so we can just get one humanoid to walk around and bring drinks for people. It's carrying capacity is 55 pounds.

Look, this is what I'm asking you.

You guys wanna buy it because your developers, you wanna developer kit, that's absolutely legit. But what I'm asking is for the average person, how autonomous,

clothes.

I wanna fold my clothes. I'm tired.

it depends whether or not it, how autonomous is it or

whether or not we are the first step where it's only like remote controlled, which is absolutely fine. It's the first step to collect data and then make it autonomous. But yeah, you have to have someone driving, like holding the wheel and yes, who this person is gonna be while you don't want to fold your laundry. So you know, it's your laundry, it's your in your bedroom. So

I think they're very transparent about as well.

they're very transparent about

The type of folks who buy this type of product at this point are not

folks who are, disillusioned by the fact, that this is gonna be fully autonomous. Folks understand that in order to have it learn the task, somebody needs to perform the task in your home. your home has a very specific distance from where your dishwasher is to where you put the dishes. Your home is very specific type of clothes, type of laundry, type of where you put them, et cetera, the different closets. it has to learn how it'll learn. It will learn by people totally operating. I also think that like people would care less at some point. Like I think we'll get to a point where if those folks are vetted, and one x claims that all those folks are US based, then it wouldn't like care as much. Like people wouldn't care as much. And we're gonna show the robot itself. so 1 54 pounds lift, so it can lift stuff. it can lift the stuff load three times. Its size crazy and it can carry stuff that weighs almost as much as this. if they add just 10 pounds more to carry, one of these robots can carry another one of these robots. It's really funny. Early adopters order today? First delivery is expected in 2026. we're gonna talk about 2026 in a bit. all right folks, we're moving on. But we are very looking forward very much to when these robots are among us. and definitely LDJ had one last comment, I think.

Yeah, I was going to comment on the situation with, when they say the one

X experts are controlling the humanoid, they say that's specifically for complex tasks where you would have to schedule a session with an expert and then they would go in at a specific time and tell operate. But it's not like randomly you're gonna have people tele operating your robot at random times.

Yeah.

So obviously fears aside, and obviously Hot Takes on X aside, you will have a Teleoperated presence when you say it's like a schedule. I want you on Thursday morning while I'm on whatever, and you walk in, you have no walk-in zones, so you can go into bathrooms or whatever, and you're gonna do dishes or you're gonna do laundry. Recapitalization announcement. this is a big one, folks. I'm very excited to talk about this. is Wolfram also here? Hey, Wolfram. Welcome. Ah, there we go. Happy

Halloween.

Oh my God.

Happy Halloween.

Wolfram. Wait, hold on. Let's zoom on you to show the photo. What's going on over there?

put on my makeup.

this is how wolf really looks, folks.

Yes. Other days I

wear a mask.

I'll just say all of the cohost got a, got an invite to

show up in costumes this Halloween. And, Yam has the deal with the glasses.

Come on.

Yes. Leave me alone.

All of the hosts got an invite, to celebrate.

All right, folks, with this introduction, let's move on to talk about, the biggest news in the world of big companies. the biggest company of them all, maybe the fourth total biggest company. has finalized, or at least so they say the recapitalization structure. So OpenAI now has, ownership. It's a, OpenAI group and OpenAI, foundation. OpenAI Foundation is the nonprofit that controls the OpenAI Group. the group is called PBC, public Benefit Company. It's worth, evaluation of $500 billion, I think is the biggest private company in this way. here's the split between the ownership of OpenAI. Microsoft controls around 27%. Just so you think, OpenAI is working with this. Microsoft owns 27% of OpenAI. The nonprofit arm has around 26% and is dedicated to actual, like public, benefit and goods. We're gonna talk about this in a bit. OpenAI employees, control like 26% investors, whatever, blah, blah, blah. So this is the breakdown. Sam Altman, Yaku, pki and the, VO Reba, who is a co-founder have all joined the livestream and talked about this ization effort. and, here's the highlights of that, of that conversation. I can probably pull up some clips, but I wish I brought you clips because some of this is just mind boggling. so first of all, the new structure open a foundation governs the open a group, public benefit company. It starts with 26% equity and it can go larger. So like it can grow based on their performance. the foundation can get more equity from the main company. one of the coolest things that they've announced that literally nobody talked about is $25 billion to use AI for curing diseases and launch AI resilience ecosystem. I think that's where they brought out, voyage, to talk about this. they will invest up to $25 billion into bio risk mental health and job shifts. like cybersecurity layer for AI era, basically is how my AI summarize this. But, the job shifts thing is they talked about, like people will lose jobs because, based on what they plan, based on a SI, this is gonna be a big deal. I wanna talk about this one at length, including a discussion on stage, please. They had the chief scientist of OpenAI, Jakub pki I think is the last name. Yes. PKI Yakob, basically said that, Hey, at the rate we're going, at the rate we're seeing things improve, we are going to have an AI research intern, quote unquote intern by September of next year, and by, March, 2028, we're gonna have an AI researcher that essentially is comparable to a researcher that works on OpenAI, that currently the salaries range from. If you're not in San Francisco, this price will not make sense to you at all. AI researcher by March, 2028, This is like their announcement that, hey, we're gonna replace ourselves, or not replace ourselves, but we're gonna build, tools and we're gonna build, AI are gonna help us in delivering what we need to deliver, by next year as an intern. And by two years into the future, like a full on AI researcher, which will then be able to get scaled by a thousand percent just based on GPUs and achieve way, way more at the rate we're going right now with human researchers. Getting AI researchers at that level isn't in insane. What do we think? is this realistic? bro, tell me.

I just wanna say, this is marketing talk media talk,

but not in the way you think,

bro.

Chief Scientist,

have you guys tested Codex right now?

Have you played with Codex right now and saw what it can do? have you tried giving it a very hard task and just walk away and see it, do what it does? I am running Codex with GPT five high on. I dunno, really expensive machines with GPUs and it controls the GPUs and does what it wants. Go online, do a lot of AI research. I dunno about intern or not, this is why I'm saying it's media talk because I'm not sure what the definition of a, an intern capability is, but it definitely can do a lot of AI research at the moment. The things you do on daily basis. And so it's not farfetched at all. I don't know about intern PhD level and so on, but yeah, it definitely will go autonomously researching, advances in ai, because it already does a lot of this at the moment.

think for folks who are just listening, just a second, Ryan,

just for folks who are just listening, Yum has his, let's fucking go acceleration glasses on. Nisten has them on. I just put mine on because this is the type of stuff we're talking about. OpenAI is sharing transparently their plans for the future, including their most ambitions plans, which we're gonna talk about some gigawatt stuff as well. Ryan, go ahead.

I think we're going to see the beginning of specialized vertical

agents completely take off, right? I speak from experience, like I am building a specialist agent for a specific job role right now, and it's real, right? if you're hyper-focused on a specific role, specific, goals and outcomes and guardrails, you can build this now I think this is just the beginning and we're gonna see thousands of companies build vertical specific agents that will be very effective at jobs.

My point here to interject is it's our job.

It's my job for sure, but like you guys are joining as well, it's our job to follow everything that happens in AI for the past two and a half years, coming up on three years, very soon since GPT four. It's getting harder and harder to keep up with whatever the heck is going on, and the speed is insane. We're now seeing open source models, tiny open source models with 10 billion active parameters beating the crap out of whatever G PT four was. G PT four was like huge back then. And now, the model that you can probably run in your house, is beating that by wide margins, open source. And we've seen the scale like a speed of acceleration from human researchers scaling up their human capabilities while getting sick, while getting moving within companies famously to MSL, whatever. Not that they did anything so far, but we'll see. And the thing that people don't understand, I think Sean, one of the commenters here, highlighted this in the comment as well. People have a really hard time to understand exponential. When an AI researcher, the launches from OpenAI, that's gonna be AI researcher, that's gonna be more like in infinitely scalable as much as computer allows. We're gonna talk about computer in a second. People are talking

people are talking about, what's gonna happen in the future.

They are right now shipping, social networks, browsers, app stores, video gen models, code editors, agent Frontier Ls, open weights, models, CLI tools, the same company. Was this even possible for any company couple of years ago? Like of course there is automation already. it's accelerating because of this. They're using it themselves, like for sure, a

hundred percent.

How many people use spreadsheets list

with the breaking news will not let Tropic step in into

the big announcement about OpenAI. We're gonna wait a little bit and we're gonna talk about on topic. but I think this is huge because what you're saying is correct. OpenAI is shipping an incredible pace. OpenAI themselves, we have a corner dedicated to OpenAI almost every week. and this is people assisted by ai. This is not AI scalable. And when we're gonna get to a point where it's like fully autonomous folks, it's gonna be absolutely insane. So insane already, but hey, that's not more that's not all. This is not all they talked about they had, LDJ, I promise to get to you before I move on. about the AI researcher and about AI intern.

Yeah, so on that point, yeah, I do think it's quite realistic and there is,

there's the whole aspect of, I feel like there is a lot of people that feel like, oh, even in the next 12, 18 months, we might have fully automated researchers and this fast takeoff and whatever. And while I think that's possible, I don't think it's the most likely and, but I like people's enthusiasm and their optimism in that sense. But I think this does seem quite realistic though, Brown has recently said he had a tweet that said something along the lines of, I know people say Sam hypes things a lot, or that they feel like he vague posts or things like that, but his views that are publicly stated often match the median of what I see internally at OpenAI. And I imagine these public roadmap milestones that they're describing here, this does seem like it's a good, or maybe even slightly conservative to the median of people at OpenAI and in the Frontier Labs,

they brought out Jacob, who's like chief scientist,

and here's what he said. I'm gonna read this quote verbatim, One consequence we discuss a lot, there is a GI, artificial intelligence, artificial general intelligence, but we find that in some way, even this maybe understates a bit the magnitude of the possible progress and change there. And in particular, we believe that it is possible the deep learning systems are less than a decade away from super intelligence, So systems that are smarter than all of us, a large number on a large number of critical axes. This is not like high P post, This is like chief scientist of the biggest public company, the top AI company, who's generally very conservative. and everything else that the people ask him, he was like, evasive. This is like a straight discussion. That's saying a SI is when reach a GI we don't care about anymore. We've seen this from other companies. We've seen this from meta, superintelligence labs. They talk about si and the shift to SI is very interesting, but we have to move on because all of this si is gonna have to run somewhere. and there's two more things that they mentioned. One of them, OpenAI is moving into a platform company from a product, company from just in intelligence company. J GPT moves into AI cloud and platform. we also obviously saw the OpenAI account is now becoming a thing. You can log in with that account into soa. You can log in into Atlas, the browser. Atlas browser itself a browser. They have devices that are coming, multiple devices they mentioned, they didn't mention which, but they mentioned it's gonna be part of their stack. and then, they will definitely move into the app store era with productization. So like basically they're gonna be become platform. they also will have user freedom and privacy stance. that's gonna be interesting. And then all of this is gonna be powered by infrastructure. And currently they were very transparent about this. Like it's very interesting. Already over 30 gigawatts of compute are, invested. Okay? So OpenAI right now is invested right now commitment of up to 30 gigawatts of compute disclosure. I work for CoreWeave, that's part of this compute, as well. it's a small disclosure there, but it's around $1.4 trillion in obligations for compute. I wanted to quickly understand with folks here on stage, where the hell is the 1.4 trillion comes from? 'cause we only tracked the 500 billion for, Stargate, and then another $500 billion with Nvidia and SoftBank. And like all of these, I don't know where the additional $400 million for compute came from LDJ have a hand up from before or is this a Oh no,

Okay.

so on this point, the original Stargate announcement mentioning $500 billion, that was for the initial, it was, five or 10 gigawatts of compute that they announced. I wanna say 10 gigawatts. Yeah. Yeah, around 10. but they haven't fully given the total dollar amount for the extra gigawatts that they've committed since. But since that initial 10 gigawatt announcement, they've now mentioned total commitments adding up to, I think it's at least 26 gigawatts at this point. 30, yep. yeah. Closer to 30 probably. And so at that point, when you add up all that, it gets 1.4 trillion. Makes sense for that 30 gigawatts plus of compute.

And then Sam Altman sets there, thanks LDJ.

And basically says, Hey, but this is not gonna be enough. And so we have aspirations to get to a compute factory, to a gigawatt a week at 20 billion, at the special price of 20 billion per gigawatt. Because right now it's like way more than this one gigawatt a week. They want to get to a point where things make general things. He said, we may need to rethink our robotics strategy in order to achieve this. Again, I'll say after we just saw the one x neo humanoid robot in your home. And after we heard space uncle talk about the only way to achieve what they want to achieve is Optimus is building factories, is of Optimus is the build factors of Optimus is the dent can build some other stuff. Sam Alman mentioned that they aspiration wanna get to one gigawatt of compute, stand up in a week. I gig. Gigawatt of compute is a lot of compute. The magnitudes for people are not like calibrating. I went to data center not far from here. We had eight megawatts here, which is what, two orders of magnitude, less than a gigawatt. that can power up three hospitals. This is the type of what, eight megawatts of computers. And that's around, I dunno, 64 H 200 or something like this. One gigawatt is an insane number and they want get to one gigawatt a week.

gigawatt can roughly power a million homes.

Yes.

This is there, there's countries that have less like power than this. It's absolutely mind blowing. the mentioned way to get there is we're gonna have to rethink our robotics strategy. That's directly from Sam Altman, small product notes as well. They're talking about the adult mode that's coming in December. Sam Altman started and said, Hey, unfortunately I misspoke on Twitter. I mentioned the article. Everybody lost their minds. It's not like we're gonna give you porn, but adult mode is coming. If you age, verify yourself, OpenAI will give you a I GPT. That's like better. focused, for you as well. no plans to sunset GBT four, 4.0, and until much better. Rider ships 4.5 will stay on as well. So there's some answers from, from, from Sam Altman. Essentially, folks ask the questions about IPO and then the bill. hundreds of billions dollar a year per revenue is needed. Hundreds of billions of dollar per revenue is needed. we just saw yesterday Google, including, Gemini raise, talked about their first a hundred billion dollar quarter. so that's Google with all the, whatever, 150 plus thousand people in Google's employee and whatever. so OpenAI is aiming really high. Basically, this is the update. What do we think?

Yeah, I think, the short term financial forecast, that they've been

putting out or I think it's maybe like semicon confirmed, 'cause I think it's somewhat leaked to an extent, but it seems they're aiming for around a hundred, $150 billion in revenue per year by 2029. And I think that's not even that crazy considering, as you just said that Google right now. Yes, Google's a very big company. Of course it's been around for longer, but they're already doing around $400 billion per year in revenue right now. It's current state. And that's even like without that significant of a boost from ai. And it seems entirely possible of OpenAI like being this new powerhouse and American industry that I think it could also have definitely reach like 200 billion revenue per year by the end of this decade. That seems realistic

All right, folks, we gotta move on.

We gotta land some of this down to realistics because some of this is like mind blowing numbers. the gigawatt numbers are just like, they're not connecting. Like honestly, it's really hard to impress how much. Of outsized power, this is, and what it will take to get to this power, including probably breakthroughs in energy, which Samon refuses to keep talking about folks, I think that's coming. I literally asked him directly Hey Sam, where's this? The powers gonna come from? He said, we're not ready to talk about this yet. On the livestream, the, they had a outline of how things are happening. Energy was a big piece there. I wish I had the visuals to show you. I don't. they said we're not ready to talk about energy yet. But like we should remember, Sammo is invested in Heal the Company, that does fusion, for energy. And, the current energy grids in the US cannot contain this growth. definitely not with this, level of gigabytes. maybe there's some stuff that we dunno. Let's move on. Let's move on to actually touch Grass Cursor 2.0 folks. I'm really excited to talk about Cursor, who wants to cover this? Ryan? I think you got excited about this first. Go ahead.

we're starting to see Agent Labs.

train their own models, do rl and they're designed, for swe and we're seeing great results. and I think the new composer model is very fast and good. I'll always be honest and straightforward about these things, like this is clearly a competitor to amp. but it's good. and competition is good. so that's exciting. The other interesting thing is that they have an agent, view, which you're showing us right now. Yes. And the idea here is that you're really, managing a fleet of agents. and they've got some cool features here, right? So you can spin up, an agent with work trees out of the box, so you don't have to set that up, which is great 'cause that's hard. and, this is exciting. So excited about this. I will say, the reason why I switched off of Cursor is because the UI is a lot, right? I love working from the CLI, right? So I have a TUI and I have Neo Vim, and I just, to me, that is easier. but some people like all the UI here. that's my drive by thoughts on it.

thank you Ryan.

great thoughts for sure. like two main things, course 2.0, the new kind of course or whatever focused on multiple agent that you can run. While this one runs, I can kick off another agent. This is new to Courser. Second of all is this UI looks like Chat GPT folks, this is literally for folks who are just listening. The new one does not even look at code like, like what the new one looks like. Chat GPT and because they understand that many people into Vibe coding, they don't really care about code anymore. they care about code vendors. There's changes and they have this like changes implemented here. but basically this is how the UI is gonna look. They have the agent and then you're gonna click and then you're gonna show the editor kind of changes on the right and that's it. but also this is just on the cursor side. They also released a composer, which is their new model. Ryan, as you mentioned, they mentioned a composer combines strong coding intelligence with best in class speed. And they show the comparison. I don't love evals like this. Shout out to Cursor of course, for what they're doing in building, but evals that don't show me comparisons to other models are effectively useless to me. So these are like marketing graphs, because they're showing, best open source fast Frontier models Frontier to July of 2025. I don't know why they chose July. And then the best frontier models that include like GPD five and Cursor three, 4.5, and they show composer somewhere just behind best Frontier models, but significantly faster than best frontier models. the speed is great. They're showing themselves to be best at speed. but yeah, this comparison is not great. And Apple to Apple comparison is also impossible when you post like this. but hey. Both, of course, sir and Wind, which gonna cover in a second, are moving to, building their own models as well. Any other comments from folks here?

Yeah, I just wanna say, their entire site, just got a,

remake and you can, feel the new, UI through the, weather profile site. I really like that. we see this shift of all the major agent product. The product is an agent system. You can strap the agent system into an IDE for sure, but it's first and foremost an agent that does stuff with your computer and you can launch now. Now it's absolutely obvious you have a UI that just launch agents, swarm of agents to do stuff. And this is separate from the IDE, that's the main thing I just wanna highlight. we had an announcement from them a month ago that they do on policy online, RL on the users themselves while they're running. And now a month later we get a new model, which I don't have, a solid benchmark. that's not nice out of cursor not to provide, not to be transparent about the actual benchmark just aggregate. Frontier or like we are faster. Okay. But that's, so I don't know how it performs, but it's really cool to see that small players are now, fine tuning or even online are ing their models, which is wild by the way. And, we see the production. It's really cool. I didn't have time to test it yet, so I don't say anything to say.

say I agree with everything you said besides the

small player, of course there, but the evaluation is insane. That making $200 million in revenue. I just

wanna say people are really mad online about these benchmarks.

Oh, you gotta keep in mind it's very competitive now. So there's gonna be a lot of online be between different, providers, but, people are pretty mad that, there's no public benchmark. The results are private. no one else can test it, and no one else can see the results or run their own test on it. So it's just like a made up test to get more sales. But as far as the model itself, it did show some pretty promising things. A lot of people feel like it's very small. People are suspecting it might even be like a que 30 B three active, Finetune, but they guess, oh, you think it's

a Finetune?

yeah, of course it's a Finetune I would bet like 99 times

out of a hundred that it is a Finetune.

but you're acting like that's bad Nisten.

that's actually good. That means that we'll see more. Great.

It's gonna be better.

Yeah. but, people tried it. So shout out to Alpen Dale. He tried to get it to build a whole operating system in Kernel from scratch, and he is yeah, it feels like it's pretty smart with the tool calling, but, he had to go in and use Saturn and Opus to actually fix stuff in the end. So it is very well trained for the tool calling. I think these are best paired with other models. People feel like, cursors is like very overpriced for, it's more profit play than anything else. But it is fast. it does get stuff done. So yeah, this is more of an enterprise good sales play rather than, actually that good. That's my take on it.

All right.

we're gonna do a super quick, this week's buzz to tell you about some exciting stuff about, weights and by cor wheel. And then we're gonna talk about, C cognition's model that came out the same day yesterday as coursers model. we're gonna chat about some tools cover, multimodal a lot. So stay with us. We're gonna have video and voice and very exciting, examples of some of those. for now we're gonna move to, this week's buzz.

All righty, folks. Super quick, announcement about this week's buzz. This week's buzz is a category on Thursday, either we talk about, everything that happens in Weights, & Biases, but also our parent company, CoreWeave. And, this week has, breaking news, actually breaking news like literally an hour ago, news from the parent company, CoreWeave, as we've announced that we have, made an acquisition decision of this company called mimo. If you haven't heard about mimo, MIMO is basically a, notebook, open source notebook competitor to like Python notebooks. and it's been growing in popularity among multiple, folks in the areas of machine learning and data science specifically for ai, ML development, stem, et cetera. it's, so shout out to the folks on joining CoreWeave. I'm very, like excited, to have them join because I think, this again highlights the way that, CoreWeave is thinking about the essential cloud for ai. it's not only GPUs that they serve and save to bigger companies, there's now Weights, & Biases for machine learning. There's Weave for AI engineers. There is. Open pipe for a rail training. There's WB infr, four engineers, and now we're getting to notebooks as well. They will run on this hardware, but also we'll be able to use some of the other tools in conjunction. So I'm very excited. welcome to the Marmo team that I haven't talked with yet. I'll definitely hope to bring them on here to talk to you about mimo Wyatt. Cool. And why you should use this against other notebooks in additional news. fully connected.com in London Connected is our, premier conference and you're gonna get to see from the ways and biases folks, a bunch of folks from Miss Trial gonna show there. a bunch of folks from different European countries as well. we have tons of like speakers there and we are inviting you to both days and multiple tracks. We have, keynotes from GSK, we have folks from, Google and Mr. And Lama Index and a bunch of other folks, to come and talk to us about what's going on. And I think the keynote, the closing keynote is from Wavy, the, autonomous driving company. so definitely worth. checking that out and coming. If you need tickets, find me fully connected.com for the registration. we hope to see you there if you are in Europe. this is the update from the last thing that I'll say. Ai Engineer Code Summit is in New York, November 19 to 22. And, I'm gonna be there. We have a big booth, and Ryan's gonna be there. they have a booth as well. we're gonna be very excited to welcome you guys and talk about our offerings for the AI engineers who code with AI intelligence. So very looking forward to meet some of the listeners there. As always, the engineers is a place where, a lot of us who listen to the show and a lot of us who enjoy the show are, meeting up. very much looking forward to that with this week's buzz. I'm very excited, to welcome the folks. And, let's move on to video and audio folks. I think there's a lot to talk about there.

All righty. So we are back and we're gonna have a few more guests, hopefully. I think Swyx was supposed to join some point to talk about, cognition. We'll see if that's gonna happen. I will just mention this in case it doesn't show it. the same day yesterday that, cursor released their update and, they told us a new model, cognition that now owns Windsurf. The other competitor in the, Gentech IDs, which I don't even know what to call Cursor anymore because it's an IDE. It's an agent platform. I don't know Ryan, call it an agent lab. Agent lab. They're all converging, basically. and so Windsurf is one of those folks who are converging as well. And, they have announced their own, model. Basically this model is called, SWE 1.5. If you guys remember a few weeks ago, SWIX was here. Swyx our friend from NT Space, and he now works with recognition and he said, that he talked to us about their SWE grip, product that's like super fast. That was power by CEUs. Cognition's agent model is also now powered by cereus. And you can see SWE 1.5 is around 950 tokens per second while achieving a whopping 40% on Swyx Bench Pro. It gets very close to sonnet 4.5 at 43%. 950 tokens per second versus 69. Obviously a bunch of the speedup comes from the cereus, integration. and also they have a harness there that's called Cascade, which is their, agentic, sidebar. And, that one is achieving a very high SWE inch Pro score with 40%, which is, yeah, it's quite crazy. I think the speed here is the craziest one. they trained this model hundreds of billions of parameters built on the latest GB 200 cluster. I really wanna get Swix in here to talk about this, model optimized with end-to-end RL real coding environments and recognition, as we've talked about before, multiple times. reinforcement learning is big and especially for these models, especially for verifiable outputs, where you can run things in environments and have them actually achieve something. Test that something, say, okay, this is the right way to go about Agentic approach to code. Send it back to the training and keep training on this. These loops are what causing these, like the huge speed ups, but obviously again, the speed is the highlight here. So shout out to cognition about the model. And we're moving on to video. Wolfram, I think. I think Should we show folks Odyssey? I think he mentioned Odyssey.

We definitely should.

It was so cool to see that. It made me smile. And anything that makes me smile, we should show, I think. Yeah,

Odyssey was like insane.

highlights of the week

it was, and the, thank you Wolfram for reminding me, folks, we talk

to you about video models all the time. We're gonna mention, high Luo, this is, how to say this, high Lua, 2.3, which is like a updated realism. and these models, most of them take five to 10 to 15 to 20 minutes to generate, right? So we can take a look at the new hia, for example. And, you're probably not gonna hear this, but here's like a video. we've come to expect this level of motion from AI videos. Maybe there's like some improvements in the motion here. definitely a crazy release from hia, with the turbo model as well. But even the turbo model here, will obviously look the level of detail here is absolutely insane. So maybe I should pause here in a second and talk about, actually Haloa. but they released, halo and Halo fast and faster, like 50% faster than the previous. shout out to them for this release, but. Odyssey is a completely different beast. So this model and the SOA model, all these models, they take a while to generate videos. Odyssey, let me go to Odyssey super quick. I, hopefully they will let me in because they have a queue. we're in line. We're first in line. Okay? Odyssey is a completely different beast. Folks, when you go to Odyssey, you get like a little bit of a time and you get this like TikTok ish interface of the videos. But, here's the kicker. You can imagine anything that you want. And I will let the co-host here shout out different wild ideas of what we're gonna render.

Army of robots in a Starship corridor

of robots in a Starship corridor.

Alright, Starship, let me make sure that I'm typing so it will get it, once I hit enter, you can see this start streaming immediately. Do you see this happening immediately,

folks?

fast.

Now, obviously the fidelity of this is not as much as what we just saw in high low or soa, but it's compared to the stable effusion ones, for example, it's really good. This is the fidelity, but heres the kicker. this is live streaming video that shows something like, it shows suggestions of what to do. And here's the kicker folks. This is a real time interactive video. What does interactive video even mean? You can say turn these, robots into fluffy, covered cat robots. And you hit enter and they will start streaming this change almost immediately. And you see the cat robots. Now again, you can see the little glitches folks. you can see these robots turning into cat robots like immediately and then you can like completely change the scene. Here's an additional thing that I love about this thing. When you click let's imagine another, model. Let's look at, this lady back, for example. Okay. This is a ladybug on the leaf with some rain. then we can say, there's suggestions here. When you click on the leaf, it understands that point and shows me suggestions about grow the plant or splash the water. When you click on the ladybug, it shows me, ladybug flies away. Let's click on ladybug flies away. And supposedly the ladybug will now fly away. Folks, this is interactive video. This is mind blowing.

what is this thing, man?

What do we know about this?

I don't know if we know a lot.

I've asked, the Oliver Cameron and one of the folks to come and join us, and, Oliver will hopefully join in the future session. But I just wanna like show again, how fun it is to control videos. These are short videos. They're looped, but still this is a pride of Lions in the African savanna. And we can just say, you can either type or click. We can say, turn the lions into sheep.

Joseph has a nice comment and he says, Odyssey speed makes it

highly useful for a tutorial scenario or fabricating a story to explain compost top

topics.

Yeah, a hundred percent. And I just wanna highlight that the fact that okay, we're seeing this, the fidelity is here, is pretty good. Look at this doggy. And then we can add this dog to get him a robo dog friend. and we'll see like another dog step in the fidelity of the dog is not too bad at all. It's not. It's not as high quality as other like big labs, but this is live streaming video. This is absolutely insane, folks. So I'm just like very excited about this. Wolfram. any other comments about Odyssey?

Yeah, so the quality, when you do it, you see it's not as

usable at the other bigger models like Sova or vu, but it's so much fun. Just show it to your kids and have them make a story and, create the video for that. So it, it's a glimpse at what we'll have in a couple of years when you have really interactive shows or something like that. Yeah, you can change anything just by prompting it.

I'm still dumbfound about this.

I'm like, this is so fun to play around folks. Wow, it's free for now. You can just join this, it's free and then you get seven minutes a piece and then you can just keep going. You can keep scrolling between videos and try them, in a different setting. It's really fun. I recommend you all to go ahead and try this. and yeah, this is just absolutely, this is RCML. This is absolutely stunning. Iy folks. We're gonna move on, because I think this is the highlight of the video. Sora two is now free without invites. you have to pay for J GPT, but Sora is invite free and character cameos are really big there. Lemme actually show you character cameos super quick. And then we're gonna chat about voice and we have another guest here to talk about voice with. So in SOA, you can now, have character cameos. What does that mean? before, you had to scan your face and then create a cameo like I did. but now you have this button that says create a character and you can create a character of anything that's not human. you can upload, different AI generated images. You can, upload your pets, which I promptly did. This is Sonya by the way. I can show you the resemblance if you guys see Sonia in the middle. Hold on one second. I'll show you actual Sonya

Sonya's gonna be famous now.

Yes.

So this is my actual cat. This is Sonya making her first appearance on ThursdAI can say hi. She's cute. Oh, she's cute. She's super sleepy. So I just woke her up. She doesn't understand what's going on, but this is her cameo. let me show you her cameo super quick. this is her cameo and the pumpkin that she eats is also an actual pumpkin that I scanned and turned into a cameo. So this pumpkin can show up on its own, on different places. Can you make a coat, do the cat theme, just like ramming at the keyboard? Yes, I'm pretty sure. Pretty sure I can. and I think the highlight there is the, these like cameos will turn into. Potential IP for people. Many folks will add their own cameos and then people can use them, and there will not just be people. I think it's like super, super close. This is an update from the SOA team. Alright, folks, so this is an update from SOA team, and now I want to, invite, our June, gimme a second folks, for, to talk to us about the latest release as well. So I'm gonna, I'm gonna add Arjun here to the stage and welcome, welcome, Arjun. Hold on, let me, switch the setting here so that it's gonna be great. I don't think we can hear you though. Let's test this out. How about now? Yes. There we go. Awesome. All right, folks. I wanna welcome Arjun, from Cartesia for the second time on the pod. I think you guys have been there before a while ago with Cartesia just launched. and would love for you to give us the kind of this week's news about what you guys released, what launched, and to talk about some exciting stuff about Cartia.

Yeah.

Yeah. no thanks Alex. You appreciate you having me on again. yeah. this week, Cartia, we launched, Sonic Three. so it's the most natural, most controllable, realtime voice AI model. really the bread and butter of building this model was really about how do we give developers the level of quality control and speed that, you need for interactive applications. so I think, yeah, this demo's super cool, and play around with it on our playground, as well as like on our website. maybe Alex, if you wanna, I don't know if it'll play the audio, but Yeah, I think it'll play. Hold on. Let's

see.

Oh, wow.

Valentine's Day snuck up on you. Don't worry. We'll get you a table. No problem. Let's make it special.

Yeah.

So I think it's, really cool to be able to get to a point where, everyone is trying to build, invoice ai. I think it's took off huge, over the last year and a half. and a lot of folks are really trying to build, in the voice agent space, very naturally. and I think, the piece that was really missing is, when you have really fast models, you don't really often get the same level of quality or controllability that you want, right? For, building these kinds of agents. I think we, solved that problem with Sonic Three. I think, Cartesia has always been known for building incredibly fast models. our team is responsible for building state space models, which was the, alternative to Transformers a while back. it's taking off a lot over the last, two, three years. Mamba being, maybe one of the most popular ones. we also wanted to show that, hey, look, latency is not all of it. there's much more beyond that, things that people care about and you don't need the trade off, You can, build. with all of the characteristics. so it's pretty good.

Arjun, you jumped on here a little bit, but I wanted,

you thankfully helped me to create my own professional clone, which is Yeah, you guys offer an instant clone, which is 10 seconds of audio. for the professional one, you need to upload two hours or whatever. luckily I'm a professional yapper, so I had two hours on my first to go and send it, and so I actually generated my own voice and I wanted to have it ask a question from you, folks from now on the ai, Alex is talking, or at least for the next sentence. Alright, Arjun, thanks for that. Can you speak about why it's so fast?

Yeah.

that's awesome. how do you like it?

was pretty good.

one of the things that we introduced with Sonic Two was

the, super high quality, instant voice cloning feature, and then we extended it to pro voice cloning, which is, what Alex is using here. to answer that question, going back to, state space models, fundamentally when you think about transformers, especially in cases of long context, your, compute and memory scale drastically, similar to a database you're remembering every single token, you've ever seen. And then you're trying to do some kind of computation between all previous tokens, in order to figure out what token you should generate next. state-based models operate a little bit differently. actually they're inspired by, recurrent models which have this. fundamental idea of memory, right? So if you think about the human brain, it's not it's remembering every single token that ever existed in your life, but you've compressed it into these memories that you're able to pull from, reconstruct from. and I think that's, a huge reason why you get this. there's a primary reason why you don't scale that, drastically. It's like a linear scaling. yeah, I think you can definitely, geek out about some of the details there.

Let's geek out about this next thing.

Hold on. yeah. Again, ai, Alex is gonna speak. Yeah, that's awesome. Can it also speak Russian? He code. This is very interesting because when I trained this model, I trained it on a bunch of English text. and then I think you guys mentioned, multilingual support as well was extended. and also the code switching thing, which I, I'm gonna be like, I'm gonna joke a little bit here. I don't love and the reason why I don't love this Arjun. Yeah. And then, I will have, have your thoughts on this as well.

Yeah.

Is as somebody who speaks multiple languages, three interchangeably

with multiple folks in our household, the way that I told them to test whether or not it's a real Alex or not. To not trust the voice anymore is to kind of code switch immediately. So if I ask a question in one language, they would answer another. And then we both understand that hey, this is like some AI bullshit. it, it doesn't, and now you guys added code switching, and now this technique of mine does not work anymore. So two, two things about this. One, uh, al support code switching. Please talk about this. how do you guys see the move away from what's real in terms of voices of everything that can be said can be cloned. And you guys always, obviously we're professionally uppers. We have a public podcast, but you guys also have some audio of yours out there. And yours can also be cloned very well based on your technology. So we'd love, to response to this.

Yeah.

you said it really well, Sonic Three, we expanded the set of, languages that we support. We have, 42 languages now, 43 including English, that we support and, hundreds of dialects, right? I think one of the beautiful things about language is that, even if the syntactical grammar may be shared, it's very different in things like pronunciation, intonation, prosody, depending on, where in the world you go and speak. I think what we really found quite exciting, especially from, a research problem, which, often gets me excited, is how do you really generalize? these different, concepts of language and, Interactions that, we have, across, different ways that people are actually trying to the systems that we're building. multilingual was actually one of the ones that stood out very clearly because everyone, around the world is not just speaking English. in order to reach the large population that exists, outside of English speakers, we have to generalize across multiple languages and multiple dialects. but on top of that, you also don't wanna take away from the essence of how certain languages are spoken, right? So folks that you know, are multilingual or are able to speak in multiple languages have a very particular way of speaking it. And we want to give the level of flexibility and control to say, Hey, you wanna preserve that accent, right? You wanna make sure that it's not sounding super you've anglicized the accent just 'cause you're speaking English or something. so that's, part of what we wanted to preserve, in the experience as well. pretty interesting research problem. I think there's a very different way of thinking about where is the signal coming from the prompts that you're giving it. and so it was pretty cool to be able to see that in action.

Yeah.

you guys also introduced laughter control And emotion control via tags and laughter is very interesting, because, a lot of the stuff that we say, you can hear the intonation of how I say this, whether or not I just left before this or about left after this. as I'm smiling, you can literally, like folks who analyze speech, can hear the differences when somebody's happy versus sad, et cetera. So emotion control, laughter. Talk about this a little bit and how this applies, how this shows up, how this is trained for.

yeah.

So I think when, like when we think about, again, you said it really well, which is when we think about emotion or these like nonverbal cues, it's not just the cue that like, you're, you're like trying to speak out, right? It's like you're changing the full intonation of the sentence. the way that you ask a question might change if you're angry versus if you're actually curious. and so being able to like, almost deconstruct, what does it mean to be like sad and happy? And being able to qualify that and allow it to work with a bunch of different transcripts, not just quote unquote the ones that were like used at train time. I think is like a very interesting problem. and so like this disentanglement almost is like really critical for being able to give people the level of control that they want. I think that also, warrants how do we actually think about Building a system that, allows people to have the kinds of levers, right? emotion control is great. These laughter nonverbal cues are great, but there's a bunch of other controls that I think, we're really excited about building, into Sonic. that, will be really useful for a lot of voice applications. I think, it's interesting a lot of people are trying to build in this direction of how do I get my voice to sound so natural, right? there's a lot of customer support companies, right? the Sierra Decagon, cresses of the world or, doing amazing stuff, right? When we're thinking about voice and how they're, how, they're able to like, build these experiences around customer support and what they talk about is we're really trying to like, build the level of control so that our developers can also be able to use these systems in the way that reflects their brand. so we're like, we'll give you that power, right? We build the models, we're happy to take that ownership. and, we're excited to see kinda what the world builds with it. That's perfect. Ryan, go ahead,

Arjun.

It's cool that you're here. 'cause I was literally, looking at the Cartia and Sonic three. I'm like, oh, wow. Cool. so Alex props to you for getting all the right people in the show. so I'm gonna talk as a developer, less about the tech, And how you train the models and more okay, I wanna use this stuff. I'm building an agent. we're specializing in the legal tech space. So make your pitch like as a dev, like why should I switch? I was looking at your TypeScript, SDK. Just talk to me about how easy it is to set up an agent that is pure voice. if I'm on, next ask. Yeah. and then the cost comparison.

No, that's a great question.

I think, my background is of course in the research side, but I care a lot about the engineering. I'd say, at cartia that's in our blood as well. we have a really strong product team whose entire ethos is how do we build the right experience for the developer. you have, the TypeScript library, right? we have a bunch of SDKs that, we've built, for our, APIs. I think getting started with TTS honestly probably should not take you more than five minutes. If it does ping me, we need to fix that, right? Oh, I think, it should not take more than that. we work with a lot of great partners to make sure that, folks are able to build, with our TTS like we have integrations right into daily pipe, live kit, et cetera. yeah, if you're finding it hard to build, I think we have not done our job. So just let me know.

Yeah, I'll just shout out, daily and pipet both, Kwindla Kramer,

great friend of the show, one of the top real time voice folks that we have. and it's it's great testament to the library quality that build in open source, like every model maker like integrates into them, once they release.

I did not know about you all until I feel like it was yesterday.

I was on X I'm like, oh my God, this looks interesting, exciting. it's hard for me to believe you're this good and this affordable when I've never heard of you. So help me understand why. why should we all trust you then?

Yeah.

we've been building this technology, for over a year and a half. particularly in audio, like for Sonic specifically. we've been building this tech since our PhD. So all of the fruits of the labor that you're seeing in Sonic Today, we started about eight years ago, to build the right tech, to power these kinds of applications. I'm glad you finally heard of us. I think that means we did our launch, hopefully. really excited. I think it's just really exciting to see people build in this space. I think that's what we really wanna empower. and so making that at a price point that people can afford and being able to build systems that, actually feel like what they wanna do. I think that's what we're after.

I think worth mentioning also, Albert and his, access to

all this, I think he's one of the co-founders of, sorry, one of the authors of, transformer Alternatives.

right.

So Albert, is our megamind, he's done, incredible amount of research. I call him the godfather of state-based models in modern machine learning. he's the one that actually, spent years of his PhD building out, the fundamentals, behind the technology that we have. I think it's super exciting to see, and he's also gets really excited about this is, a lot of the fundamental first principles way that we, think about research actually has downstream impacts on the way that we build the systems and things that people can use, and being able to see that we've already had a large number of, customers trust us, that we're able to build with and partner with to build amazing, experiences. we're really just excited to, keep growing that, and then also keep pushing the research that will enable new experiences moving forward.

So I think just to, Arjun, thanks so much for joining us again.

it is great to have you here. you guys are doing great work. We all played with some of these models. I think you're talking about like 49 milliseconds latency to the first token or something. Yeah. Yeah. That's the turbo. it's insane. It's insane for, it really is something, it's very similar, like we're just talking about Odyssey and looks live streaming video to you. hitting that button and getting instant response is just like unbelievable. and building bots with some of this intelligence plus some of the faster intelligences that we have, really makes you think about the stuff that McQuin claimer talks to us about where there's realtime models that do everything, but they're like definitely smaller and stupider than the production models. And now with this TTS text to speech on the other side, you can get, with this latency, you can get. As close to real time as possible, while also getting the benefit of switching the models, adjusting, fine tuning, et cetera. So really cool. I wanna highlight, Arjun, thank you so much for joining us. I wanna highlight that. What's possible now to use with these models? I wanna show a little segment, I've prepared using Argil. So shout out to Argil. I think they used vo. this is a segment that I prepared, basically had the LM summarize our show here. fed it to Cartia with my voice that Argent helped me create, and then, led to ar. I haven't seen this yet. It's a minute and a half. Let's take a look together. it's probably gonna be wonky, it just shows what's possible without spending that much time into it. shout out to our goal as well. Happy Halloween. I am Alex Oloff, and tonight on Thursday ai. October 30th. I'm joined by Wolfram Yam, Nisten and LDJ and Ryan on a candy band. June updates we'll unwrap Minimax M two and Agent Tick slash code model. That's about two sonnet speed at roughly 8% of the cost. Then open AI's GPT off Safeguard and IBM granite 4.0 nano for tiny but mighty efficiency. We'll even peek into robots with one X's mini OA consumer humanoid. You can actually pre-order and recap open AI's restructuring and Roadmap foundation governance, the a SI within a decade claim and timelines from an AI research intern in 2026 to an autonomous researcher by 2028 plus platform shifts like Atlas devices and products. On the dev side, cursor 2.0 with composer for parallel identity coating cognition as Bay one five. Now in windsurf and perplexed the email assistant for a privacy forward AI inbox in vision and video. Sora goes in bike tree with character cameos and leaderboards and minimax. Helio 2.3 pushes realism reimagined in voice, minimax speech 2.6 hits sub 250 MS. Conversations and Cartia Sonic three brings low latency emotion and even laughter for truly natural delivery. Yes, really plus quick hits on Pokeys Agentic Workflow builder and Google's Pelli marketing agent and our this week's buzz on fully connected London and AI engineer and mamo. It's dense, fast, and a little spooky.

was one shot,

Yes.

That was pretty great. Wow. Here's the process of what happened here. I took the notes that I prepared for the show. I dumped it into chat g bt to give me a script. I then asked GBT to go to Cartia, figure out their prompting techniques. 'cause Cartia has specific, pausing things. It's very important to pause when you speak something I keep forgetting, but it's really important when you write a script to keep pausing. and so I asked her to go find Cartesia, Sonic Three model, then dumped this into Cartesia with my voice that, that was created I think last night for a couple hours. and then I took that audio, dumped it into Aril, which was trained on my face. So not one shot in terms of just provide one image. I prepared some of these assets beforehand, but the whole generation was one shot. And I think it's incredible. And the, it sounds like me, not entirely, but it's very close to sounding like me. Sounding like me is hard. I have three mixes of, I think it's a, one of the highest level fake alexes that I've been able to achieve so far. For sure. This has been done again with Argi, nobody pays me to do this and Esia no, none of these folks sponsor me besides giving me some access to their systems and some credits. So this was really great. very spooky for very spooky Halloween folks. So with this, I think we've covered. Almost everything besides the fact that folks, it is Halloween and we just showed the ability to clo your voice, clone your video, have videos, clone your pets, clo your pumpkins. We showed, coding companies getting into releasing their own models. We also showed new models releasing that are like half the size of the previous models that outperformed them, As we look at Wolfram here we realize that this is the Halloween type thing, but we are not here to spread, scare about ai. We're here to spread positivity. I think everybody else here, we believe that like AI is built and is gonna achieve great number of things, including, powering humanity, to achieve things we haven't achieved before. So on this very spooky Halloween, I wanna say like nothing that we saw today frightens me to the point. It was like, oh no, let's pause. And not that it's possible to pause anyway, but I did wanna, wish you a great Halloween for those of you who celebrate and mention that we'll be here next week as well to cover the next week of the show. So the last comment that I have is if you missed any part of the show, which we had a great show today with two conversations with the folks who are actually building the stuff. Cartesia raised a hundred million dollars round, and this is part of the reason why they were in news. Congratulations, by the way, to Cartesia, OpenAIs talking about multiple billions of dollars in revenue, multiple hundreds, billions of dollars in revenue. Talking about gigawatts of compute. where we're going is absolutely insane, and we're very happy to be here to cover this for you, to condense this noise into signal, to send it directly to you. So if you missed any part of the show, some folks were asking about show notes, if you missed any part of the show, show notes and everything we've talked about is gonna be published on Thursday, I that, so please subscribe to that. Share it with your friends. Thursday at the news is where we publish our podcast and our news, and basically the edited form of this. With that, if you missed anything, subscribe, give us five stars and we'll see you here next week. Bye-bye everyone, and bye-bye folks.