📆 ThursdAI - Sep 4 - Codex Rises, Anthropic Raises $13B, Nous plays poker, Apple speeds up VLMs & more AI news

Hello, hello, hello and welcome everyone to

Thursday I for September 4th. so fun to be back. I missed this, I missed the show. I wasn't here with you guys last week, and I'm very, very happy to come back here and talk about everything that happened in the world of AI for the past week. and maybe a week and a half or two because you know, some of the stuff I missed and I wanna maybe discuss, we have a great show for you today. So, my name is Alex Volkov. I'm in the AI evangelist with Weights, & Biases for those of you who forgot after one week. but with me, I have co-hosts, I have Yam Peleg, and LDJ Nisten. Peleg gonna join us and we have a few great guests today. So very, very much a great show. I'm looking forward to. we'll have folks from News research. Roger Jin is gonna join us, to talk about their new release. And we have our own Quila Kramer, One of our experts on voice and audio, everything, voice and audio, founder of Daily and the maintainer of Pipe. with that, I wanna say hi to my friends. Yam Peleg. LDJ. How was last week folks? How was this week for you in ai? Yam, how are you doing? I'm

doing great.

Coding agents are blowing up. Yeah, and, everyone is competing and we are the one to benefit out of it. We get better agents, we get better models and I just wanna say, yeah, I had a post that, took two about Codex because it's good. It is really, really good. And there is a hype, there was hype building up, so I was curious. I went to check whether or not it's real or it's just hype. I know it's real. Codex with GPT five is really good. It's a solid option and,

you're really looking forward to hear more on that because I think that

people are always looking for some stuff. and while co codex is. You know, it came out with a bang and then people was like, oh, no, clock call is better. But they kept working on it and there's some great, great people on there. so definitely looking forward, to hear some tips and tricks for you for Codex. I know one thing and I'll be able to contribute that conversation, because they did update some stuff, in Codex, LDJ. What's up man? How are you? How was your week? What's the most interesting things that stood out to you, this week?

Yeah, my week was good.

I noticed what seems really interesting is a Rock Fast Code, which I think it was maybe released or maybe released today after our last Thursday, but it seems to be one of the fastest scoring models on Open Router now. It seems to be in top three places, actually, at least last I checked. And, what Yam was saying about Codex. Yeah, it does seem to be better now. Sam actually tweeted yesterday that usage is up 10 x in the past two weeks, which is pretty insane. And he said improvements are coming, which if people remember there's the Codex one model that they have with it, and maybe they're hinting at Codex two, and I guess like maybe coding Finetune of GPT five. We'll see

that's very interesting.

definitely Codex blowing up. Sonic was, if I'm not mistaken, the code name for, grok Code Fast one or whatever they call it now. And I think it was tested in multiple places. people mentioned that like super fast. And yeah, I think it was released within the last episode or just a little bit afterwards. And, yeah, taken off, space UNCLE is very proud of that achievement. and very interesting that Claude, sorry, the GR is now part of the coding world, which seems to be one of the more hitting areas of ai, at least lately. and one of the areas where AI seems to get the most adoption and the most excitement and, the most improvements for productivity, which is great. my week was a little bit off. I am coming back from a very special place and still recovering, but I think that, I really. Was very happy to see you guys covered nano banana last week, and Nano Banana is the Gemini image one now, but everybody still calls it Nano banana. Just such a cool name.

And they change it back to nano banana.

No, on, yeah, Google, Google AI studio. There is a model called Nano Banana. Ah, that's great. That's great. Yeah, absolutely. Great. Yeah,

so shout out to the person in Google who, who came up with this

incredible naming, because it's stuck. so Google has been definitely killing it this summer. Nana Banana is just the latest and greatest Kwindla update and yeah, I'm catching up on all of the creative uses of it where people take the Nana banana image model and then they put it, the outputs as a last frame on some of the more. Interesting video models lately that really support for first name. I think Van released their support for first name, last, last frame. Julio Cling, like all these things. And then people do some incredible stuff, like exploding objects. So I'm really excited to catch up on some of the more creative uses of stuff like the responsible nano banana. And it's fast as well. I love that it's fast. So this was Kwindla my highlight as well. if you guys are tuning in on X, on YouTube, on LinkedIn, high LinkedIn, five people from LinkedIn, everywhere via my stream upstream or, Wolf Froms, who is not with us today taking a a well deserved break. then please comment with whatever, it was the highlight in the world of AI for you this week. We would love to know in case we missed it. we absolutely are trying to cover everything that that happened to the world of ai. I would say no huge model releases this week, but there's a few. And as I said before, we'll have, the folks from News research join us, shortly to talk about imas four. They released another model today, so they had a few fine tunes and now they had another model, actually yesterday. And then we'll chat with Queen La Kramer about, GPT real time, which is also very, very exciting. if we have some time, we will also Maybe have breaking news. I hear our friends from our favorite Chinese lab are, up to something. But with that said, let's get into the TLDR. Let's chat about what we have, going on the show this week and we'll dive into open source immediately after that. for those who are unfamiliar, TLDR is the corner where we basically run through super fast with focus, on every bit of news that we have that we're gonna chat about. in that case that you don't have the time to sit with us for a few hours, you will know everything that happened. let's go to TLDR.

TLDR of course stands for too long. Didn't read in this case. It's too long. Didn't listen because the show is long. But let's, let's get to it. So today on the show, the host is Alex Ov. That's me. I'm very happy to be back. I have co-host. LDJ is here. Yam Peleg is here. Nisten is gonna join in a bit. Wolfram is get take the world. Deserve break. the guest for today is Roger Jin. and I think he's bringing another friend that I didn't write down yet, from News Research and, Quila Kramer, the founder and CEO of Daily and the maintainer of podcast. Which is incredible. Qwen has been multiple times on the show before, and if you've seen him before and you're into the voice stuff, you know how, incredible Qwen is. So I'm very excited to have him back. and then we're gonna talk about open source at the start, and I think we're gonna talk about news Irma, It just launched last week, so folks already covered it on the show last week. However, they kept launching and they released the 4 0 5 B and I think the 70 B last week. And this week they released the 14 b Finetune of Qwen. And this is a compact reasoning model with tool calling for local cloud use. We also have some news that this air, MS four is getting some incredible results on the EQ bench from Sam Peach, benchmark. Not only that, the news folks released another thing. they released a benchmark called Husky Holden Bench. we will chat with them about this as well. Healthcare Holden Bench is open source poker bot eval. So LLMs create poker bots that play poker and then, strategically play against each other and have some ELO score. Very, interested in hearing about that. We will welcome Nisten, and then we'll talk about the Tencent. Open source. Ian, mt. So this is a translation model, specific translation model, that won the WMT 2025. very interesting from them. Alibaba has like multiple ai, areas in it. so Tonge is one of them and they open source the vision language, deep research agent with a new state of the art, with a VQA style retrieval and planning on benchmarks. It's called Web Watcher. It's very interesting. Aptus eight B and 70 B from Switzerland, fully open, multilingual LLM. it's ran on 15 trillion tokens and it should across 1800 languages. What this, this is news to me as well. I'm very excited to talk about this 1800 languages. I wasn't aware there's so many languages in the world generally. it looks like a lot of tokens went into this, for this like multilingual model. So a bunch of stuff to talk about in open source. Anything from hosts here, anything that we should talk about that I missed?

Kimmy K two update that.

They release a little update to their model. it's not like too major, but some preferences, coding capabilities, and all around little bumps.

We love Kimmy.

let's go to big companies and a and LLMs and APIs next. we will start with MS Trial adding 20 MCP power connectors and memories in Leche, for enterprise, Anthropic. I think with the big news of the big labs this week raised $13 billion in series F funding. Those GPUs are not cheap folks. They are not cheap. And so $13 billion and a 183 billion post money valuation for Anthropic people. Do seem to be buying and tropic stuff. The revenue grew incredibly from just a few months ago. I think in the multiple billions of dollars growth, not just like total, just growth is multiple billions of dollars. and so shout out to Tropic. Also shout out to the guy who's in jail. Sam Bankman Freed, who's investment in Tropic would've been worth like, I don't know, like $30 billion at this point. last he, you know, didn't screw up with some crypto stuff and then went to jail. but yeah, a very pernicious investment from SBF Open AI ships, GPT Real Time, which folks talked about on the show last week. And, realtime APIs now in GA with remote MCP tool support and image input and c phone calling. And we'll have Qwen talk to us about some updates there. I'm very, very looking forward to that. openly also did some updates for free users, so free users not get projects. so if you're not paying the 20 bucks or 200 bucks, and we don't blame you. There's so many AI tools to choose from. you will now get projects and there's a little bit of updates, to projects as well as, OpenAI buys the company Sig. the folks from STAT six join as like VPs. In some application building or another within OpenAI. this means a lot for the coding industry as well. And I think the last thing up to update in the big companies and APIs, Guac, code one, FKA, formerly known as Sonic, is now taking 50% of traffic on open router. It helps that it's given out for free on multiple places. so the pricing is flexible there, but definitely a new contender for coding. And we will talk about why and how. Folks, I have a very exciting this week's buzz. very, very exciting this week's buzz because CoreWeave announced a purchase of a new company, just like they purchased Weights, & Biases. We have announced that, the great folks from Open Pipe are joining. the great folks from Open Pipe, Kyle Corbet, David Corbett, and a bunch of other folks who work on the Corwell platform, on the art, training RL platform is our joining CoreWeave as well. So, CoreWeave, went into an agreement and buys out open Pipe. Kyle and David and a bunch of folks are joining, the Wisdom Biases team, the CoreWeave team, to build out a full complete infrastructure metal to model, including tracing availability, including training for r rl, including everything. and I'm really excited to talk about this a little bit, about the implications of this and how one platform is gonna fit everything. we now have, tracing, we have model training, we have inference. we're really gearing up to be an all around company. So I'm very excited about this and again, shout out to, with Biases and CoreWeave for allowing this podcast to happen for two and a half years already, which is very, very exciting. So shout out, and also congratulations to Kyle and David about this LDJ, go ahead.

Yeah, so a couple of things on big Co LLMs.

Yes. that's reported on within the past couple days, which is $10 billion raise at $500 billion. Post money valuation. but it's actually interesting in that it's actually a sale of employee share. So it's allowing the employees to basically sell about $10 billion of shares, and that's basically the employees get getting $10 billion, essentially.

So it is like a share buyback for employees.

Wow. Let's go, let's shouts. Go guys. Let's go. shout out to multiple of AI millionaires that were born and are able to realize their gains, which is not usual for a private company. Anything else in the big companies looks like? No. All righty, folks. So we're moving forward. apple released Fast V LLM seven B. It's a speed first vision encoder 85 times faster times the first token versus peers. And it's like, it's really super fast, is blowing up on hug and face right now. so shout out to Apple and really looking forward to see if Apple's gonna mention any AI next week on Monday, where Apple's, the the new Apple event is gonna happen. and also last thing that we should definitely talk about is Codex versus Cloud code discussion. would love to hear from, Yam and Nisten about this, folks. I think that's most of the show. And I think with the interviews we have a full show and I think this time to get started with open source. Let's go

open source ai. Let's get it started. Alrighty, folks, it's time. Let's get it started. I think we'll start with, with smaller potatoes for a little bit, and then we'll get to the interviews as well, as we have folks joining us, today. let me see if I am, yeah, I'm sharing my screen so I can talk to you about this one. Tencent, we'll start with Tencent. Tencent, Everywhere. The video models everywhere, they had like multiple language models. At this point, it's getting harder to follow the Chinese labs, but hey, that's why we're here for, they announced the open source release of Huon t MT stands for machine translation, obviously seven B, laser translation model that they won the WMT 25, 20 25 awards. Now, very interesting to chat about, between us super quick. why would a specific translation model is needed when these models are great in translation anyway? and I think like with everything, fine tuning models for a specific use case first, and then rolling those updates into a big model has been a pattern that we've seen so far everywhere. Like the Codex model started out as this coding model, then they rolled into GPT. I think Elon talks about, grok code also getting rolled into this. We've seen the same with Gwen. it seems like this is the pattern. but, seven B for a translation model is Kwindla chunky. I think. Nisten we always talk about what's the use case for smaller models running on device. And we always mention translation's. One of them, like, you can run a local model for translation and most of the time it will be correct 'cause translation is like very easy. what do you think about this one?

Because yeah, you gotta pull in data from multiple sources and

you do have often to do translation, even if you're just benchmarking. So a seven B to run at scale is pretty good if you go a bit lower, it doesn't have enough like meat in it to just pick up the, the different languages. But at, at seven beats pretty good when, yeah. when I, I've tried it, so I think this is like the smallest good translator that they could make and then they can run really fast for, I would say it's more for data work rather than personal inference. For personal inference. Just pick something for one or two languages and maybe Finetune it but it's good. Right.

On the floors 200, which is a dataset, a benchmark I guess of

200 languages, they look like they're getting very, very close to GPT 4.1 and, this is a seven B local model, so expect this to be useful for, local use cases. they're comparing themselves to all the major labs models. So we see, GPT 2.5, GPT G 2.5 Pro G GPT 4.1 CLOs on it. And this seven B model stands with the shoulders of giants. I really like the fact that, when we see a new model lease with a specific use case, like a translation model, they all compare to the larger models. All the larger models are well-rounded. They're multimodal as well. So comparing with this one niche use case is like, it is great, but like it's Kwindla emitting everything else. The other model does. But for a specific use case, I think it makes sense. so they also posted a technical report and folks who are into translation, definitely should check it out for local, local use cases. 33 languages support and fine minority language support. very, very high coverage. they also open sourced Mt. Kymera, which is, open source integrated translation model intentionally refines translations from multiple models to deliver more accurate professional results. Oh, that's interesting. So this takes translation from other models and Kwindla refines them. So Kwindla like a pipeline model in the mix.

But I, I don't, I don't think this is, that this is more

that it speaks different languages. But anyway, I, I might actually use it 'cause I like generating medical data sets in other people's languages. Like you're a French doctor for medicine stuff from, and like Dr. And, you do want to have something that that translates it back to, to English because then you get a more varied, varied dataset. So. I wouldn't really consider this for on device, something you want a three B or lower on device. It's just a little bit chunkier, but it is great for, doing data work. that's how I see it.

Yeah.

So shout out to Tencent for releasing this. let's just take a look at the license before we move on. I believe it's a PA chain. Oh,

it's probably MIT or it might be there.

We are looking at, I don't see license here.

License,

license list.

Nobody.

I just, I just wanna say the reason, this is important, it's important for this model to be this small is because you generate, when you train a model in a new language or extend, languages, for models. You really need to translate a huge amount of data. So you need the smaller model as small as possible because you're going to translate, inference of trillions of tokens, maybe from one language to the other just for your training, like synthetic data. So the smaller the model that you can have, it's better. It's just gonna be cheaper or magnitude cheaper. seven beans. Can be really good translation, models. I tried only for two languages, but I can see how it works well with more languages if you have more compute and Tencent definitely has more compute. Yeah. So, yeah.

All right.

we'll cover the last thing before Kwindla, the main coverage of open source. we are looking at Operatus eight B and upward to 70 B. This is a 15 trillion talking train on open data from 1800 languages, which is way more language than I thought existed generally, academia to work in train LLM across the full stack. they did, pre-training with some innovations. They posted a technical report soon to be posted on archive. this is a Switzerland large scale multilingual language model. they say it's fully open source. So shout out to folks, on this. I wanna see the model on HA face as well. So this is Central Switzer Road, the Calculus Scientifico, CAS Cs, national Supercomputer Center from Switzerland. so it, it's, it's great to see Europe Kwindla contributing to the open source. Ms. Charles is Kwindla, like a little bit lagging behind in open source. sometimes there really cool stuff, and they were like, clap. But generally Europe has been a little bit lagging behind and

there are really good people from Switzerland specifically.

there, there are many, many notable people in some not notable academics. And,

in Zurich is very well received, like in

multiple conference as well. Like, there's tons of people in ICML and CVPR and, you know, everywhere basically. So, the folks posted, the link to hug and face. So let me go find it. the link on the page is just the link to hug and face. Not necessarily their hack face. So here's the collection of models, upper two LLM. let's take a look at this collection. We have two 70 B versions, so I'm assuming one is base, one is instruct, and then two eight B versions. One is base, one is instruct 25 0 9, and the license will be Apache two. So yeah, definitely worth shouting out.

Oh yeah, Apache two.

Thank you very much.

Every time Apache license is with us, we clap because this

is what we want from open source. everything else is not really open source. I need to ask to the

paper.

The paper is really interesting. they reveal everything and anything you just wanted to know, very insightful, all the experiments, all decisions. highly, highly recommend. We don't have time to go through everything in the paper, obviously, but highly recommend, if you are, into these Kwindla things, training models, fine tuning, it's a really, really good, end-to-end paper release. One of the few that we have. So

we credit Allen AI always for, fully open models.

So we should credit these folks as well. Not only do they open Weights, they did open data as well, which is incredible. And then the full training details, including all data and training recipes as well. So if you want to train a model such as this, these releases are incredible. alright folks, what else is interesting here? It's using, you can use this in v LLM in, fest transformers, 15 trillion tokens and supports an incredible 1800 languages. I don't see any variations, so I'm not assuming they like trained the soda model fully open source. they did post some. Compare comparisons to like Omo and Euro LLM and smaller them. So usually these folks that do like fully open, they don't hit soda necessarily, but they do compare themselves to other, like soda, open source models. Shout out to idu, releases from, Swiss ai and we are now moving forward to cover the awesome news research releases. The folks who have been following Thursday, I for a while know that we have a particular, love in our hearts for news research. We've been chatting with news folks, way before they were even a company. I remember announcing that they Kwindla like founded the company and since then they went on to do incredible stuff. Obviously, one of the highlights of news research, this for me is that they Finetune open models and usually create significant jumps in the those fine tunes. So, to chat with us about this, we have Roger Jin and Bove. Bo Bavesh. I would love to for you to add your last name here, to chat about this last week's and this week's releases. So welcome folks. Welcome Roger. Welcome Bavesh. How are you guys doing?

Hey, I'm doing great.

Thank you

for having us.

Great meeting you guys, and thank you for hopping on.

we have a busy show, but like, I'm very, very, very happy that you guys came on because, last week obviously Herman Sport was released and big, big fanfare, 4 0 5 B and 70 B. And this week you guys followed up with another release of a 14 B Finetune of, I wanna say Finetune of Qwen. Yeah, that's right. Awesome. And then you guys also release a technical report that looks like a paper, but it's a technical report because the news girl is on top. And then all your names are in there, including techno who, whose first time I think is like first name show shows up there. But also all your ex handles are underneath, which is the most important part of the thing in AI research for group such as yours. So Roger, how about we start with a brief intro to you and then Vesh, and then we're gonna talk about the release and then follow up. We're gonna talk about the new bench you guys released as well.

yeah, sure.

So, I've been at News for about a year. I'm on the post-training team under Technium, machine learning engineer. I work on data, I work on evals. I work a little bit on rl. and yeah, I was involved in the Hermes four, project. Awesome.

And, Bavesh, how about you?

hi.

I just joined Noose two months ago. It was a very exciting opportunity for me and, I was not as involved in the Hermes four release. I just was involved in the benchmark.

Awesome.

So, we'll, we're gonna cover both and I think we'll start with the announcement of es Roger, let's treat this Kwindla release conversation about all the models as well. I know the folks covered a little bit of the MS four, release before and last week, but for folks who missed it, can you give us like a brief overview of the releases of MS four, the work that went into this and Kwindla, like what the fine tunes? I think there's a mix of fine tunes, right? Like the one is LAMA 4 0 5 B, a few of them is Qwen. how about you give us like a holistic overview of Hermas four, the release and we'll celebrate? Yeah, yeah,

exactly.

yeah, so we released three models. in size is 4 0 5 B and incentive B. These are fine tunes of, LAMA 3.1, and we also release a 14 D, which is a Finetune of, Quin Que three. and for all of these, we started from the base model. So, not from the instruction Finetune model. and, the goal of this, our fine tunes is, or these models is always Kwindla to like, it's Kwindla like to create a, generally aligned, neutrally aligned model. so a model that can be shaped by the user that doesn't Kwindla like, like Kwindla, like lecture the user about ethics that the model company may have baked into the model. but it is Kwindla, like a neutral tool, just like a computer or something that the user can shape, to their own liking. we can see that, Kwindla did some evals. this is one of our first, flagship reasoning models. We had this deep Hermes preview release, but this is our first flagship, reasoning model. just like other reasoning models, we find I can do math and code, but when we evaluate the model on high school Olympiad problems, it actually doesn't perform quite as well as other open source competitors. but we find that we see interesting generalization to other domains because along with math and code, we didn't just focus on these areas. We actually, mix in quite a bit of chaotic data from all sorts of different domains. we actually get some nice performance gains in other non-math, code benchmark. so for example, it seems like we're, fifth place on the leaderboard for YouTube bench three. This, these complicated interpersonal relationship scenarios. benchmark, in B'S like Husky, Holdem, like poker bots, eval. It seems like we're actually the leading open source model. so we're seeing some gains on like maybe these non-traditional benchmarks I think some interesting, facet of this model is we might have actually undertrained it in terms of reasoning benchmarks, Well, basically we stopped training the model, not because the reasoning benchmark stopped going up, but because, our more spiritually, enlightened, model behavior, people like decided, like yeah, this model of like definitely losing its model soul in, in the, in the chain of thought. so this is,

oh, train of thought.

That's interesting.

Yeah.

So, we're looking for base modelly chaotic behaviors in the chain of thought. and when we started observing that the entropy was going down, that's when we started deciding to terminate the training. It wasn't actually that the reasoning benchmarks saturated or something. it seems like we could still potentially get more gains from, a longer training run. but we wanted to preserve the entropy of the base model. that was the cutoff point for this project.

I think following the news folks, on your discord,

you have multiple models that are trying to get unlocked as well. Like Opus for the longest time was a very, unlocked model, on the discord as well. And there's multiple folks there that smell the soul. I know, like, shout out to Koran for example. And like the bigger models are definitely, more soulful, like 4 0 5 B for example. so could you talk about comparing, those not Moes that you train, those that you Finetune, you Finetune, like big, large, dense models. just talk about a little bit of a data mix, like how do you get four, five B to the top of EQ bench, for example, was a direct question I think on X that somebody asked, and I would love to hear also. what, what do you think, contributes to these like specific results on the, the, the less, hill climbed benchmarks, for example?

yeah, yeah.

So, so I think we, we put a lot of effort into the, into the data for Hermes four. So, so Hermes three dataset was, 1 million samples, and then around like 1.2 billion tokens. and, Hermes four dataset was around 5 million samples and, and 56 billion tokens. So it's like five x, more samples and then like around like 45 x more tokens. and into this data mix, I think there was plenty of like, traditional stem, like math coding, like, data, reasoning data. and a lot of that is actually just like open source already. but I think we put substantial effort into generating interesting data for, these like all sorts of domains. And, we have like various techniques to do this. So, one of these is what we call data forge. So this is based on a paper, called Agent Instruct. where we take pre-training data as almost a random seed. We take a piece of pre-training data, and then let's say it's a news article. we then might transform it into some, other document. Let's say we, we transform the news article into a, rap song. and then we take this rap song, and then we generate an instruction about it. like, you know, what's the beat of the rap song or give me a rhetorical analysis of the rap song, and, we do this over a large, random subset of pre-training data. so that's where we get some of our diversity. and then we have other techniques where maybe for specific use cases we really want the model to perform well. So for example, we really care that the model can output things in specific formats, complicated nested JSO or whatever, kinds of formats. so for these kinds of things where we want to list out enumerate entire space, we Kwindla have the model generate a whole taxonomy of, all the formats Kwindla, like list some domain, have the model partition it, produce a partition of the domain to sub domains and then subdomains, within those et cetera. And then when it gets to something that feels as like indivisible, you can kinda just have it list out examples. and we found this as helpful for when we want to completely cover some, area. And then another thing that might be related to this, is we actually generate some data from personas. so from persona hub, like find personas, we see, use these as seeds, to simulate synthetic users. so this might be another source of where the understanding of different kinds of humans comes from. but the data mix is quite chaotic. So I think we were also quite surprised that, we did so well on Eki bench three. and so I think we still have yet to do the proper ablations to figure out like, yeah. what is the source of the magic here?

Yeah, just like one or two more questions about the model.

Super quick. any changes to context Link? I wanna shout out, news folks, because literally since we chatted with I Mozilla. And, a bunch of other folks about rope like, I don't know, two years ago. And, and then yarn and some of the work that you guys did for this, like literally every other lab just like uses those techniques. Some don't even like expose that they use them. any, any interesting updates to contact center or the same as base models, license changes would love to hear as well. And Kwindla, like what are you gearing hermis, models for folks to use for, I think that's also very interesting. Like folks who listening to our show, they're like getting excited about all these fine tunes. some different benchmark evil, et cetera. the Chinese models are open source and the, they, they want like rebuild the image of China within the open source community. So there's like multiple reasons to release. We would love to hear where Nous Research stands right now in terms of like, why and for whom are you releasing these models. And then absolutely we'll talk about the bench as well. And the, yeah,

so I think, starting off with the context length changes,

like we trained, Hermes three at context length, like 8,000 and Hermes four, we actually train at 16,000. But we actually see that it can reason up to like 40,000, tokens. the max context length for the LAMA model 70 B and 45 B is, 1 31 K, same as the base model and for, Qwen. the base models were trained at 30 2K, but then they extended it to 49 60. so we also released ours at 49 60. and they recommend like in the Qwen three dogs, like you can use yarn to extend it up to 1 28 K. I didn't actually run evals like that for our model at 1 28 K, but, that is something one can do, and then, I guess regarding what is the model useful for? I think one of our main goals for this model is that it's primed to do RL on not just math and coding, but on, a multitude of domains. if people are interested in doing reinforcement learning for creative writing, for role play, et cetera, I think this is a very good, choice, of model. we've tried to, if we believe that RL takes pass at K and turns into pass at one, we've seeded the model with plenty of niche data, in all sorts of domains. So it'd be really cool to see maybe the model, it might take a hundred tries to do something, but, but yeah, eventually it gets, that's not answer.

Just really quick, I have a practical question because I do make like

actual commercial bots that customers have to use and, one of the main complaints. While developing them is that they always sound too much like a report. I always try to pick something that, doesn't matter if it's deep seek Kimmy or even if you sonnet, although sounds usually good, but it's too expensive. what went into the training, 'cause I see you guys did pretty high on EQ bench, like how do you balance out that, entire checklist for an answer? Like if it's a very academic question, you wanna hold checklist with it, being able to carry a conversation on naturally. 'cause that does feel like it's pretty hard.

so I guess one of the techniques is rather simple.

it's to just Kwindla. Like we have, a different system prompt for each task category basically like, and like I, I guess that there is like, hundreds to thousands of like these task categories each with a specific system prompt that, describes how the model should respond. and so we Kwindla have, answers in, different formats that we believe are, suitable for each Kwindla task category. of course the user can shape the model. if they disagree with our, default behavior, that's okay. Like you just, edit the system prompt in, it'll respond in your desired format.

Okay.

So that's a very good tip. Just dynamically change the system problem based on what the use case is. Okay. All right. Thank you. All righty.

Roger.

Yeah. As part of Kwindla, the places where this model did well, you mentioned Husky bench, and this is also, one of the things that you guys released this week. would love to hear. first of all, we're always for new evals, new benchmarks, especially creative ones. Especially like stuff that is not very easy to hill climb. I'm very much interested in evals that are not the same as other evals. Most other evals have the top models on top, but we will very interested in Husky bench. Tell me about Husky Bench. I want to hear about, why testing models on playing poker means any type of sense. Bavesh, feel free to to chime in here as well. Roger, I think you worked on this as well. what is Husky bench? Why should we care? And who worked on it and what is it for?

Bavesh please go ahead.

so Husky Holden Bench.

essentially we came up with this interesting idea. I'm a student at U-Dub, so I hosted this competition called Husky Holden, where we had people come in and build their poker bots and we had them play against each other. we wanted to make this open source just as a tournament, and I brought it to Roger who gave me this great idea of turning this into a benchmark. I think if you look at the history of poker bots or any competitive, bots, a lot of it is focused and geared towards really computationally heavy, bots. LLMs on the other hand are more of, thinking. So we Kwindla wanted to flip the script and instead of it being LLM playing humans in poker, we Kwindla wanted to see how well they can develop the poker bots themselves and how well they do against each other. So that's Kwindla like the idea behind, Husky Holden bench. So just to clarify,

this is not the LLM deciding on the next move.

And how's the cards? This is the LL m's coding up a poker model. So we're, you're testing for like coding ability to an extent, but also some understanding of poker, I guess. That's very, very interesting.

Yeah, so the interesting part is that the LLM has to

understand the entirety, like poker rules, Texas Holden rules. Then they have to understand our code format. They, they have to follow specific interfaces, and we have memory and time constraint as well. And it has to translate all of that into code, make sure it runs and syntactically correct. And then finally produce like this final bot that we push out here.

And, do they produce multiple bots, to get, do you get like

average or something, or is this like one model creates one bot and then you run this bot against other models? because of the variability of coding, sometimes the bots they can create may not even run Like, how do you treating the stochasticity of like building a software as such as this?

So if you go into our games, we, we've run six batches

and in each batch, each model creates five iterations of the bot. And we take the best one and we put it in for, as that batch is final bot. And then we run this for six different batches where each batch you have 13, 2, 6, so a thousand plus games happening all, all of with all of each other.

That's awesome.

So you have a bunch of batches and the models that I see here that are competing Gnet 2.5, pro Coin three, code Gnet 2.5 slash deep Seeker one O Pro, this is the first one. And then also you got IMIS 4, 4 0 5 B also playing. and let's take a look at some of the results as well. You wanna give us the results of who's better at this coding?

It's more, it's more, it's pairwise, it's,

one against the other Yeah. Like to compare models, one, against the other. by the way, specifically, poker is extremely hard. It, it, the rabbit hole theory is about poker, about the, un you know, games that you don't have, all information and so on. I'm just saying it's not just coding the poker, boat. It's extremely hard and, and yeah, it's really cool. By the way, the pair wives one playing against the other. We need more benchmarks like this. This is really, really, really cool. Yeah. To style.

But that's question before LDJ joins as well.

the poker bots themselves are not ai. Those is like just code, right? Just logic and code. the AI part comes into like, I'm gonna code up the bot. The bot itself is straight up software.

Yeah.

it's just straight up software. I think eventually we hope that the AI can code its own AI poker bot that Kwindla changes across the game. But right now, it seems like it's just basic logic code

LDJ.

Go ahead and then we'll talk about the leaderboard

a little bit and tell you guys what's interesting results here.

Yeah.

So, I think this is a good Kwindla new emerging category of benchmarks where a lot of benchmarks have like just a bunch of questions where you have to get, like there's a theoretical max limit where there's a hundred percent accuracy on all the questions, and you can't really go past that, and thus you get saturation. But here, when you have things involving elo, like this could virtually go on like infinitely, right? Like there's not really a cap here of how good it could be, right? Yeah. And, yeah,

it's like evergreen in that sense.

yeah, as the top models get better and better. naturally the benchmark gets harder. you have, more and more skilled adversaries, until a point where like maybe poke like maybe this is like actually solve, like there's like some optimal thing.

how much time do the ai, get for the poker about building?

So when we ran this competition at University of Washington, each

person got like a month, each team got a month to build this poker robot. And then similar, the final competition.

Yeah.

And then, and then here, like these uhis get, like five, iterations to like five iterations of improvement. although in general this is just a minimal agent that we use this alpha launch of the benchmark. But, what we want, our vision for this benchmark is for this to be a general model suite, agent scaffold, like a pair benchmark. So, we do want to add more, suite agents to, test out cloud code, et cetera. That is awesome.

So we always appreciate the benches.

We appreciate you guys coming in and talk to us about like the novel approach. I very interested the bench itself is for the outputs of the model and not the models', Kwindla thinking or reasoning. This is like a novel approach that I wish we will see some more of as well, IMAS, for as all imas releases, first of all, they're beautiful. The art is incredible, always on the release, which is very important to capture the mind share, but also those models are always found to talk to folks can try 'em out, download them on hugging and face, with the, I wanna say the same license as the original models. Roger, I'm, yes, I'm mistaken. Yes. so you can download them on hack and face and, chat with them on the news research chat. Roger and bsh, thank you so much for joining us for the first segment for open source, of this. Anything else you wanna mention, shout out to the team super quick. Feel free to do so.

Yeah.

Thank, thank you so much for, for having us. Yeah, I guess like quick chats and like Tech is the leader of the post train team. He was really the mastermind of the Hermes Sport project. Jay Sofa also worked a lot on the synthetic data with us. a lot of the data came from rejection example, playing against an RL environment framework, that, that was pioneered by, Dakota. and then, yeah, like, I guess like this is all under like our, our fearless leader, like e Mozilla, Jeffrey Nel. and then a lot of the, model like chain of thought analysis came from Karin, a whole ultra. and, some of the game data from Shannon,

one of the co-founders and one of the more soulful skill testers.

so shout out to Karen as well. All right folks, thank you so much for joining us. we we're moving on because we have breaking news, but feel free to stick around and listen to the rest of the show. And then we're gonna chat with Queen about GPT real time. So get ready for that folks. And as we move on to breaking news, I've missed this so much. I have missed breaking news so much. AI breaking news coming at you only on Thursday. I.

love breaking news. I love breaking the news on the show as well. so folks, we have some breaking news from Google, from a friend Omar San Viro there, who works at Gemma. they have released Embedding Gemma, and it's been a minute since we talk about the embedding models here on the show. the state of where seems to be always moving in the bigger models. And the huge model it is that like the better embeddings are embeddings obviously are useful for like rack purposes and just like turning anything into machine, understandable, searchable, Kwindla information. and so they launched a bedding Gemma, I want to say this is, open source. Yeah, it looks like open source. only 300 million parameters, which is great for embedding models. You want them running quick, you want them running on device. they have released the top open model under 500 million parameters, so no longer. I remember at some point the MTB, the, the massive. Embedding, benchmark was, controlled by like Mistral, backed embedding models, and those were huge. So we need like smaller embedding models, a hundred plus languages also incredible because, sometimes the top embedding models are only English, flexible embeddings with different dimensions, with Maka dimension. We had a chat with folks about Maka, embeddings before. and if you're interested, definitely look into archives of what that means. We have a deep dive about Maka embeddings, and, you only need 200 megabytes to run this. So shout out to the folks about embedding Gemma taking the top on the MTB multilingual benchmark. what this feels like a new soda is something you would use.

yeah, because it's great for the phone.

yes. Right. it's small and fast, and that's exactly what you want in an embedding model. I've been using the 0.6 B, Qwen and, I felt like, yeah, you could go smaller with that, especially if you do need some Kwindla on-device, rag, which Apple really needs. Yes. so yeah.

All right, so this is our breaking news.

Shout out to the Gemma folks, in mobile first rec pipelines enabling. So definitely, they're getting to release some incredible stuff on the phone and, very happy that they're open sourcing some of the stuff. And folks, it's been almost exactly an hour since we are here on Thursday. I, which is usually the time allotted for open source and discussions. and at which point we would love to introduce a friend of ours, Qwen Halman Kramer, welcome back to the show friend. How are you? How are you doing? it's been a minute, very exciting topics for you to discuss with us.

Good to see you.

Glad you're back from vacation. Yes.

Back with a lot of energy as well.

and the actual thing that we're gonna talk about was actually released Kwindla last week-ish. But you had some times to talk about this, to test it out. And, who better of a person to talk about a new voice release than, Qwen, who I consider a resident expert. Qwen is the founder of Daily and the maintainer of Pipe Cat, which is an open source, library to use all these incredible voice releases that we have. so Qwen, with that introduction, folks should already know you've been here, for a minute. we talked about advanced voice mode releases and real time stuff. last week, open the ai, I think on Thursday, Yam, you guys did like a stream on Thursday and included that stream in the show. They, updated, GPT real time. So Qwen would love to hear from you about Kwindla, what did they release, and then some of your experiences, why is this important and why folks should care.

So two big things.

it's the GA of the realtime speech to speech model and the web RTC and web socket, API that sits in front of that model. So they released it, late last year, have released improvements along the way. Now they're both, taking it to ga. So it's just called GPT real time, Suddenly something preview, which indicates I think a lot of, confidence in that model from open ai. And it's a really good step function improvement in model intelligence. the main problem with, speech to speech models today is that you have to compromise on instruction following accuracy function, calling accuracy, contextual intelligence if you want to use a speech to speech model, rather than use like a soda LLM in text mode and put a transcription model in front of it and a voice generation model behind it. And what that means is that like 99% of voice AI use cases today, not using the speech to speech models yet, that will change, but it won't change just because of this release. So this model's big improvement, still not quite as good on lots of real world benchmarks at things you really need to use in production, but actually clears my bar for one of my vibe check things that I do every day, which is talk to the computer as much as I can and I can now use this speech to speech model for a bunch of my, daily talking to the computer stuff that I couldn't do with the previous releases.

So just to clarify, for me, and for some listeners, most of the

chat bots that they talk to, they own the Kwindla, the highest intelligence. So the highest intelligence still exist in the textual streaming world and token streaming world. so they have to, Kwindla build the infrastructure, which is Kwindla. Like what pipe cat also helps with significantly where you use a different model, like maybe a translation model to turn, speech into text, then input those textual tokens into the model. And then on the other side, you do some voicing, like 11 labs or whatever. There's like multiple companies, Cartia and a bunch of companies, to then animate this model to feel like it talks to you. But essentially it's a very like textual thing. speech to speech is a full and better model, but also because of that, it doesn't get up to the level of like, reasoning models.

So what we're seeing from the big labs is that they're

sort of forking their post-training and the way they, they think about the model. They're going to reasoning models for. All the stuff that's not intended to be realtime. And then they're training the realtime models to be low latency in a slightly different, you know, path of, of, of, of post training. I think in some ways that's really good because latency matters so much for voice stuff. And if you, you know, if you focus on making the thinking model really great, almost inevitably you're not thinking a lot about time to first token. And for most of the stuff I build, I care a lot about time to first token and like having that alignment with how people are training the models matters. In some ways it's not so great though, because I don't want the real time use cases to feel like second class citizens. I think the single most interesting thing I'm seeing in the voice AI space, and it mirrors a lot of what we're seeing in like the coding agent stuff these days is we don't live in a one model world anymore. We live in like a multi-agent world or whatever you want to call it. And in the voice AI space, that means we often have like a conversational loop. There's an LLM that's like talking to you and then you kick off or you run parallel inference processes that you don't expect to, you know, finish in a hundred milliseconds. And those might take. Two seconds, or they might take an hour and you want to figure out how to coordinate those multiple LLM driven processes, feed data back into the conversational loop, let the conversational loop control them. And the architecture for that stuff is new. Super interesting. Like, I'm actually spending a lot of our time in Pipe Kat, like building that layer of the architecture now. And I actually think the sort of talk to the computer use case is like a great example of that because you wanna talk conversationally, but then you also want to control lots of different things that are going on in the software you're using. it feels like you're living in the future a little bit if you're trying to build these apps where you can like talk to them, but also you're controlling a bunch of different apps. It's like, what's the next layer of the operating system that you want for like a voice enabled future? You want this sort of multi-agent distributed systems control thing.

And that's very interesting 'cause you don't wanna, keep Waiting

for it and have it go silent while it does this computer can multitask. Like why wouldn't it go off, do some stuff, and then I can ask a bunch of questions about research, about, Hey, go online, do this. And at some point it's like, Hey, actually I finished that task. You wanted me like, you want to hear about this now? Or like, maybe later. it shouldn't get stuck, as a processing thing. Waiting for this one thread, to finish. So, let's get back to real time. Any new changes? Could you give us a little bit of a factual update about what they released and they changed

the shapes of the A PIA little bit.

Mm-hmm. So they took the GA as an opportunity to, you know, all the experimentation everybody in the community has done around these new realtime APIs. They, they sort of finalized the API and made it a ga. So if you've got code, you probably have to do a little bit of updating of the API shapes. They also added a bunch of context engineering stuff along the way that, the GA is like the capstone of, and as we all know, context engineering is like the most important thing for agent reliability. You gotta give the LLM the tokens it needs at the moment. Not, the whole conversation history to get from, the 80 percent-ish reliability and a lot of real world tests, like 99%. So the real-time API added a bunch of context engineering stuff. You can do context surgery. it's still much harder to do that with the realtime APIs than with the traditional APIs, which is again another reason why a lot of the enterprise use cases are sort of slow to move to speech to speech models and APIs. But to their credit, I think they heard a lot of feedback from, people trying to use these models for hard real world tasks and they added the context engineering stuff. It's still not like I would like to see the voices be even more steerable. They've added some stuff there and they've actually separated out these models that are audio focused into the real-time API, which is speech to speech. And then they also have a steerable transcription, GPT four and a steerable voice generation, GPT four. Those are really flexible transcription and voice generation models. I'd actually like to see sort of all of that stuff in the real time. API. Yeah. But you know, sort of different models for different use cases. Everybody's experimenting. We're all like pushing the frontier in different ways. So I think you can see the sort of audio focus going in these experimental directions I think is really helpful.

That's great.

I, read about a few features that the folks asked for. Obviously MCP has been Kwindla like on a hype down train a little bit, but it looks like people are like heads down working on MCP. I think, Yam, you guys talked about MCP last time and how much usage actually we get from like MCP stuff. so that added remote MCP server support for voice. So assuming that you could, like in the case that you wanna talk to your computer, now, your computer can like do some stuff with easy connectors as well. I wanted to ask you about image inputs. how does that work? So this is a multimodal model, not only in oil use, so it is also sees what you wanted to see.

Yeah.

Sorry, I forgot about that. 'cause I've had this checkpoint for a long time, so I did the image generation stuff a while ago, but yeah, you're right. It was just released, Kwindla taking the GPT four oh omni idea that you can pass it. Text images, Audio and the model will just do, the right thing and making that available in the speech to speech model in the real-time api. So there's a new like API Surface area for just giving it images and it works pretty much the way you would expect if you've been doing vision stuff with LLMs. It's good.

So last year, as a reminder for folks who are listening,

I built a Halloween bot thingy, and I think GPT real time just came out back then like the speech to speech model. way before GA I had to pipeline a bunch of stuff. I had to take a screenshot, go to Gemini separately, get a description, then fit this between into the text, have the text talk about this, et cetera. this now seems like a holistic approach. if I can spend the money to actually do this in real time, because one of the things that held me back, back then was it was very, very expensive as well. I think the prices since then came out like 10 x and

then it's more than 10 x since you built that and it's

now if you do the math, it depends on your use case, but it's now about four times as expensive as using the separate models. Mm-hmm. Which is for many use cases, that price difference just doesn't matter. Like if, if what you really want is the lowest possible latency and the highest possible sort of natural voice input and output vibes, yeah. Then you, there are a lot of use cases where that, you know, three x to five x. Price Delta is totally fine. 4 cents a minute, very loosely speaking, depending on your use case. and a bunch of factors on the image stuff. That's sort of back to the multi-agent and the context engineering things though, I had some conversations with the real time API team about this because you can only give it so many images before, like you have to do a bunch of context engineering or the model performance is just not gonna be what you expect. So for a lot of the stuff, I think you would build like the Halloween costume thing, you may still want to use like a parallel inference process. It's focused on images.

And

use the realtime API maybe a little bit for vision, but

if you're feeding it like a video stream, it's not gonna be the best model for that. So maybe you want to feed it a couple of recent frames. You wanna have a separate video inference model. You want to be doing some coordination between those streams so that like the realtime API always knows what the video inference. Stream is doing, but you're not taxing the real-time API with, you know, 10 frames a second for an hour. Sure,

for sure.

It's not, it's not for that. And I think the last thing is CIP for phone flows. Would love to hear from you about that. That I think is very important for many folks in enterprise.

Super, super big on the enterprise side.

So this, acronym SIPP is the, like core telephony, inter integration INTEROPERATION standard. So if you're making a digital telephony connection to a call center or to like the big phone system in like a hospital or whatever, you're talking sip If you're taking a phone number in from Twilio, like somebody, you, you buy a phone number from Twilio. If somebody calls that phone number, Twilio is terminating that phone number. And then you can talk to that Twilio service by like, sip. So it is the lingua franca for like old school telephony stuff. So the fact that OpenAI has native SIP support now means you can directly wire up the real-time API to basically any phone system in the world and 99% of voice AI monetizable use cases today touch a phone system somewhere. So I think that's a really nice recognition from open AI teams. Part of what those of us building these applications for commercial use need, they're really leaning in. they want to be the full stack. They've got the agents SDK, they've got the real-time API, they've got sip, they've got web RTC, they've got web sockets. They're really trying to like give you everything you need.

Yep.

It looks like they're gearing towards like, Hey folks, this is the, the one-stop, solution. So, last two thing, new two voices, CDAR and Marine, specifically exclusively for the real time, and the existing ones also upgraded to an extent. I saw our friend, Swyx test out on his karaoke bench or whatever when he tries to get these models to sing. very famously, like last year, I like asked the folks at Open Air like, why wouldn't it sing? and they said we also wanted to sing. At some point there was a lapse in the release of, of, dance O Mode. It sang and then they pulled it back. but I don't think that he got them to sing this time as well. So while these voices are malleable, they still are like under restrictions as, as something you would expect from opening eye.

It's partly restrictions.

It's partly just there's not enough data and that's why the instruction following and function calling aren't as good in the multi turn and audio domain. Like we got a lot of really good tokens now for text models. When you're going to multi turn conversation. the data's Kwindla out of distribution relative to what the big models are trained on. And it's even worse in audio multi turns. So, you know, it's not surprising that they haven't tried to, label a whole bunch of music data so that these are like full bore music models. 'cause like they don't even have enough speech data yet.

That's true.

But we know they can sing. Like we've heard them sing Shakira or whatever it is, just like, you know, maybe some sound effects. But

I think if you don't lock that down in post

training, you probably, this is my theory 'cause I'm obviously not inside these big labs, but you probably end up with a lot more surface area for hallucinations and other things that would be not ideal behavior from the model. A hundred percent Yam, you had a question, a follow up

question

as

well?

Yeah. my question is you can, build your own, speech to speech system by using some, model as an engine and then having a speech to text and text to speech on the sides. And then you can switch the engine. Everyone understands the benefits of that. You can switch the engine and so on. but you know very well that. it hurts the quality of it. you got delays and so on and time to speech and you need to talk and then transcribe. only when you're done talking, you can start transcribing because you know how it goes. So speech to speech, but speech to speech is also very expensive today. when do you think about the price, and where it's going? Because it's an interesting question about whether or not speech to speech, is the right way to go. If this is the price is a completely new system and you, we might just wanna work on, efficiency of the current systems. what do you think about this?

So I think there's gonna be for a long time room for both,

because as models get, as big models get better, small models get better too. And as all the models get better, there are more things we all wanna do, which sometimes they're actually best done by a collection of small models that you've done the engineering to. Glue together in a specialized way versus just using one big model. a lot of use cases are gonna stay in the model fusion approach. Because if you need multiple models anyway and you need to specialize and you need to context engineer and latency engineer and cost engineer, and you need great Observability, you're still better off with multiple models. I mean, the future is just gonna be multi-model, I think. I believe that now. I think there's no more question. Are we going to a future where there's like just big models? I think we all probably who do this stuff every day think that's not what the future's gonna look like. I wanna push back just a little bit on your framing the speech to speech models today. The best speech to speech models today have good latency, good general intelligence performance, great mixed language performance. So if you're doing something like language teaching and learning and you need to switch languages in the middle of a sentence, the speech to speech models are so much better than any other approach you can glue together, but they're actually significantly worse. Not just on price, but on how well you can actually get them to do what you want them to do, which is a big deal. Like the reason enterprises aren't moving to them en mass is because I can do benchmarks and show you that I can build a more reliable multi-model system than I can build a speech tope model system. And that matters like you're making an outbound call to qualify a sales lead. It's not the price that's the barrier to using the realtime API is the fact that if you are 10% worse on converting those leads, like that's a massive, massive hit economically for your application. The other thing I'd say about these speech to speech APIs is Observability is really hard. So like the, the speech mode and the transcription you get back from these models are sort of decoupled enough that you actually can't really use the transcription to know what the model heard. Whereas if I build a transcription LLM voice generation pipeline, I know exactly what I fed to the model because I did it in my code, I got that transcription, I modified it or processed it however I want, and then I fed that thing to the model. So there's like no gap in Observability and until the open AI API or the Gemini API completely close that Observability gap. It's actually really hard to build enterprise applications if you have Observability gaps.

And right now they do this by also transcribing.

Just to send it back to you as a developer. Just for you to like understand, okay, the model may have heard this, but because this model understands nonverbal cues like laughs, et cetera, many translation models don't catch that. Many translation models don't catch even the pauses that we have or different, so that's what you mean by like what you get as a developer backwards. It's not necessarily what the model heard. Yeah. And so it's hard to like, you know, square this, together into like, okay, this, this is why it's at this thing. If you do a lot of

like Observability with these models, you'll find that

in the realtime mode, you very often know that the model heard something different than you read in the transcript. Like the gap is fairly large and also you still see in, in the realtime API, the model will sometimes transcribe what I say per turn in like Japanese. Like I, I really, really didn't speak Japanese that term. Like I wish I could, but no I didn't.

Well, I think we've covered this, ga release.

Definitely there's still gaps for enterprises to adopt this, but it seems that for some use cases, real time. Could be very, very interesting, I think, specifically. So

it's, it's gaining market share and it is

the future for a lot of use cases. It's just, you know, like everything else we're doing, There's a jagged frontier. Absolutely.

Qwen, thank you so much for helping us cover this.

I really, really appreciate your insights. As always, folks should definitely check out, pipe for the ability to plug in, very simple, models like this and switch around them and also build the alternative if you need to into multimodal pipeline. feel free to stick with us as we're moving forward and we're gonna discuss some other open AI stuff. big company, API and, related as well. I think you mentioned this. Let's go through some OpenAI updates super quick. there's a few acquisitions and a few folks that were posted about, being, VPs of apps stat Sig and Alex, those are two companies that OpenAI have purchased. I think the folks from Stat Sig become the VP of applications now, Yam Stat Sig, should be familiar to folks

Anthropic and Claude specifically.

Claude Code uses static to track all metrics and, everyone is now, Suspicious about this, telemetry recording service that OpenAI, just acquired. it can completely be, because I know it's a great product. these are great people. it can be the reason, but people know that, and Tropic is using the servers, pretty extensively.

So in, the, data usage Kwindla section of a tropic, they

say cloud code connects from users machines to set six servers to log operational metrics such as latency, reliability, and usage patterns. this login does not include any code or file pass, et cetera, and static security documentation. So basically you can disable telemetry, but they are absolutely using some type of the service to improve that code. and now OpenAI has. At least some of the data in a 1.1 billion transaction, which is very, very impressive. and Alex, I believe, was the Xcode coding Kwindla co-pilot for Xcode that they started like as a, as a plugin. LDJ you correctly, corrected me at the beginning that those are two different companies open produced, and some of these folks are joining Open the Eye at key positions, to, V Vij, I believe is joining, one of the top folks, to strengthen Open the eyes. thing specifically also, we'll get to this, oh yeah. v Vij Raji is joining as, CTO of applications to lead engineering for JGBT and Codex. This is, this is Kwindla the outcome. So it's Kwindla, like a Acquihire plus, execute, execute hire type, that we've seen before in meta and other places. what else is, with Open the Eye? I think a small update. There is open air research projects for free users and they have updated some of the project, settings. So like larger file uploads for free. you can upload up to five files and everybody else can upload way more files as well. and then you can customize the projects as well. I think what's more important is projects only memory control. So Chat GPT is notoriously good at product. So besides the LLM memory is very important. And project only memory is something that people have been asking for, for a long time because you want to encapsulate and maybe within the project to work about some stuff, but maybe not have it leak towards your general chats. So project only memory improvements have come to, you can control whether or not the project memory leaks into general memory. I think, memory is very important and, GPT projects will add this as well.

I think that you have some, Delicate control about, what exact memory,

is going in to the context of the moment. You can just have like, some control panel. This is from the project. This is my memory. this is from a previous chat and so on. So you can, on the fly change, what memory goes into the chat because it influenced the chat a lot. it's very notable and also system prompts and everything. and one last thing opening. please let me pin chats, star chats, every platform support this and I think it's gonna help, everybody else here. And so I'm just saying, but yeah, it's a good move in my opinion. I'm using a lot of, I'm always saying that the Chat GPT app, you can say what you say about the models. I think they're great, but you can say what you say. The apps are great, absolutely phenomenal. All of them. The mobile, the web apps are absolutely phenomenal. Everyone agrees they're really good. So

w with one tiny exception of dj Before we get to you, one tiny

exception, if you are on the desktop Mac app and you're pasting a bunch of context and it refuses, go to the web version past the same context. Usually it's a app limitation and not the model context limitation. It drives me nuts. So if anybody's listening from open the eye, why, why?

Bug report, please.

the copy button doesn't work on deep research. I'm just saying bug report if we 14 bucks, but yeah. Okay. But

besides this, those are great products.

Absolutely. Absolutely. Jay, you had a comment as well.

Yeah.

said this at the beginning, but of course probably new viewers that, the past two weeks, Sam Melton said in the past two weeks, there's a 10 x increase in usage of Codex, which is pretty insane. And then he also said improvements coming, which if people, in case people forgot, there is like a, a coding specific model opening does have called Codex One. And so maybe a Codex Two model, which might be a coding Finetune of GPT five. Who knows?

we're gonna talk about Codex in a little bit after we, we

run through some of the other updates. So Open the eyes is raising $10 billion. and most of it is for buyback, so employees will get some liquidity. So congratulations to multiple, multiple folks in OpenAI, whose lives have completely changed now, to be able to go, you know, keep working on the problem, but without the stress of where, when, and how will I get paid. OpenAI still, a non-for-profit controls, the profit arm company, and, to see some folks realize they're Kwindla, like very, deser well deserved gains is really fun. is really cool. And I think it's a great example for. Private companies to be able to do this Speaking of large fundraisers, I think the next thing that we must update, and I think it was like one of the biggest, updates for this week in terms of fundraising and Tropic raised a insane $13 billion series F, it looks like they don't, want to follow the playbook of regular, bigger companies getting to like series D or E or whatever. And then going public, none of the main AI companies seem to be wanting to do this. series F at the, $13 billion at a 183 billion post money valuation. their revenue at the start of the year was roughly $1 billion, and now it's over 5 billion by August Cloud code. And the coding paradigm definitely is a huge contributor to this. Everybody like switch to cloud code and then it drew Cloud Max people maxed out the cloud Max. They, they have to like play around with some, some of the stuff. So definitely Tropic is, focusing on the engineering. they have 300,000 plus business customers, and large accounts are up to seven x over a year. Thank you. the coverage, for this I took from Swix and the news for small ai. If you're not subscribed to the news for small ai, definitely, definitely should. that's how I keep up daily on some of the stuff if I'm not able to like just. Get the raw feed from, from the X platform. so what else is interesting? Lead investors not super interesting, so this is less for us. We usually cover the updates, but this is so huge that you guys should know. Tropic is gonna be around and still we'll have many, many more GPUs to train bigger opposites for us. Yam, you mentioned something, or, LDJ mentioned something about them apologizing for Nerfing, one of the models for cloud code. Could you talk about, about this? Well, yeah.

Oh yeah.

Oh yeah. I, I, I would show it because I think it's, let me just show it to everyone. philanthropic, did, did some updates for efficiency and apparently people noticed, but, you know, it's Kwindla a conspiracy theory, Kwindla, Kwindla thing. And maybe the model changed. Maybe it's worse, you know, you can never know.

Yes.

but

apparently there's always one person goes viral by

saying, Hey, the model was good last week and today it's not as good. And then people say, no, we paid our models exactly the same model way. It's exactly the same, whatever. And it's hard to figure out whether it's like personal feelings, people stop prompting correctly or whatever. but we now have proof. Yeah.

so what ended up happening is that, this is by the

way, my, spicy take about the thing. But, what we get is A statement from Tropic that Claude Opus 4.1 and four, had degraded quality for, a short period of time, and people noticed, conspiracy confirmed. that's the, official statement from Tropic. You can find it if you want the link and everything. I think it's, important to show something like this, and not just, talk about it. But I wanna not only show this, but

read this out for podcast listeners.

from five 30 on August 25th to 2:00 AM UTC on August 28th. So about three days. Cloud Opus 4.1 experience a degradation in quality for some requests. And then the next one, the next part just kills me. The world we're living in is incredible. This is an official update. Users may have seen lower intelligence malformed responses or issues with tool calling and cloud code lower intelligence. So users may have, you know, the tools that they pay $200 a month for may have changed under their feet, and it remains to be seen, for them to like, whether or not, you know, they screwed up with their prompting or not. This was caused by a rollout over infra stack, which we have since rolled back for CloudOps four. We often make changes intended to improve the efficiency through to our models. Our intention is always to retain the same model response quality. So basically what this means. is even though the model itself may have been the same, the inference stack maybe have been optimized, maybe they did some, I don't know this, this doesn't mention quantization, maybe they changed the templates, et cetera. And they do this without users necessarily consent or knowledge. So the users only knew about this three days after the fact after they reported about this. what do we think about this? I don't love this. people cannot trust their tools.

Absolutely not.

I think if you're paying for a tool, at least deserve transparency. you probably deserve to rely on the tool, but At least transparency, at least tell people, There is an update going on. so people would know. it's hard when these tool, you don't, you can't know. You use this every day. People use cloud code for work every day. if it messes up, well, it messes up. People feel it. And, you want to rely on the tool. We don't have the, off weeks from open ai, for a very long time. I think that since GPT four oh. Everything feels very stable and I thought that we moved from the initial phase of the field where everyone was optimizing their inference stack and no one really knew and things fell off. but apparently not. And I think at least just tell people what is going on. I mean,

Qwen, yeah.

Any comments?

The, OpenAI I think is the best company in the world

at running a massive supercomputer, but I will say that their time to First Token has gone up a lot in the last six months. 'cause we monitor that really closely. I think that's an intentional trade off they're making, but it's not the right trade off for the use cases I care about. So that goes a little bit to like the transparency. Like I really want to be able to say, Hey, I want you to keep time to first Token stable. I care about that more than anything else. They're making a different trade off. I also think the charitable view of the Anthropic thing is that they were trying to keep throughput fast even as their usage grew and they did benchmarks that made them think that they weren't impacting quality. 'cause benchmarks are really hard, but they could serve more users and they could keep the existing users happier. Like if I were inside Anthropic, I could totally see making that decision and. You know, it didn't quite work out, but like, they update the models all the time. So it feels like an honest mistake to me.

I'm with you on the honest mistake.

the thing that I keep getting hit on is that these larger companies, while being multiple hundreds of billions dollar in valuation, multiple hundreds of thousands, business users still don't realize the impact area of how much people rely on them day to day. And I think to that extent, We can talk about this release for like impacted actual people for like three days. And people need to get the work done and they need to struggle with themselves to see, you know, AI agents, coding agents. They're not perfect anyway. Like anyway, you ask them to do something, they would do something else. And then like, it's really hard to like capture the instructions following ability of these models and, and they would like slightly degraded. you basically fuck with people's reality. You fuck with people's ability to understand what's going on. And then it takes, folks posting on Twitter, let's say and posting on local LAMA and then all collectively agreeing that, hey, something's fuck here for the people's. Like, oh, actually, you know, yeah, we, we did like the inference release. I, I think that these people need to start to contend with the fact that like. People rely on these tools and they're not more the co-pilots every day now, and some sort of like, heads up folks, we're about to change something. Maybe it's gonna be better for you, but like we're about to change something and the ability to opt out of this. I think at some point we as users need to ask for this because of the absolutely reliance on, on many of these things. And we've seen multiple companies fall in the same trap hole. Like, oops, we didn't notice. Oops, we didn't have enough emails when we released. Oops, oops, oops, oops, oops. These, oops is like, people need to be able to trust these models. the benefit here is local models, open source models. When we, on. Wait, and by inference, for example, host Cmic K two, we're gonna give you the same Chemic K two yesterday and today. And we, if we upgrade it, we will absolutely tell you with a different benchmark, like a different endpoint. This is a different, Kim K two. We obviously we'll try and not change the v LLM version, underneath all models as well, which could have cost something like this. But definitely I think more scrutiny for these companies is needed and more chance to opt out of automatic updates. Like my iPhone doesn't update when I don't want it to, it only updates if I wanted to. And I think the same should be happening with like these tools. definitely honest mistake. but also folks I think need to contend with the fact that like, this is now in people's lives and they're not used to getting updates without knowing, and they shouldn't, get used to this. I believe. we are moving on because there's more news for us to cover. Super, super quick. I really want to get to, codex, as well. the more news that we want to chat about is, GR code one FK, a sonic, coding model from GR out of nowhere. LDJ, you wanna, you wanna talk about this a little bit? 'cause we, we did talk about this, but it seems that GR code is now a significant contender in the world of code. What? Out of nowhere.

Yeah.

So on open router in here, let me check on the, what's the latest stats there. But last I checked, which was a few days ago. It seemed to be second or first place actually. And it was either in the top trending or top used models of, per week. I have,

this chart here, 52.8% of coding traffic on open router.

Yeah. Is now graph code fast one? It's number one

now.

I just checked.

Incredible.

So something like 697 billion tokens in the past, I wanna say week Insane numbers. It's absolutely insane numbers. this model came out of, you know, XAI training, like a coding specific model and testing it out for free. I think a lot of this can be attributed to the fact that it's free. it's, been given out as free in, places like client, I believe still in, in multiple other coding agents. so like A huge deal because people would definitely use the best free model on top of the best paid model, because it's free. but 50% of our coding traffic, for XEI is a huge achievement. So shout out to the folks who work

also fast.

It's just a week out. I mean, it's just one week and it's already like more than 50% you have to show.

it is fast, but I believe it was about 270 tokens per second.

People were clocking it at, but average speed for GPT five, which I think a lot of people underestimate how fast GPT five is. That's like around like 180 tokens per second average. So for the

non reasoning one, for the reasoning takes longer

for the non reasoning.

yeah. Well, like if you count like all the reasoning tokens for the reasoning one, I think it's like, not that, much slower too, but obviously it needs to think way longer, use way more tokens. So yeah, the overall response will take longer. Yeah. But yeah, I think it's going to be exciting to see more coding specialized models and maybe something like Codex two from Open AI and maybe, who knows, maybe Claude will even maybe have a coding specialized model. Know

this just one to show the actual plot from Xai about

the speed of, grow code fast one, speed versus output price. And you can easily, easily see that it's. Way faster than, everybody else and nearly as cheap as Quin three coder. That's impressive. I'm just saying that there is, a different axis that is not shown on the chart and which is quality. Yes. but, but yeah, look, most of coding tasks, most of any task is like the, the small things you're just too lazy to do and some tasks are extremely hard. So I think what we see here is you give people free stuff that are good enough for, I don't know, 80% of what they do, they will use them. Absolutely. And, and, and it's a really, really good way to spread. grok Gro, by the way, is really good at code that the, the real grok, I'm, I'm just saying if you didn't try it, because I don't know, there are so many. Just go, go and try it. It's really good. And it can do things that other models just couldn't. I mean, all of them can, they're good at different things, but yeah, it's a good model.

Alright, folks, we have about, we've covered

pretty much everything. We have about two things to cover and before this, I wanna jump into this week's bus super quick because this week we have some exciting news for you.

Or we've, it feels like I need to rerecord this intro because, uh, this intro started when we were only Weights, & Biases. Now obviously Weights, & Biases by CoreWeave, which is our parent company, and, it, it is public. So disclosure here, there's nothing, exposed here. All public information shared and everything is basically my opinion besides the actual facts that we can reshare. And the facts are, that this week, we have some exciting news from CoreWeave. We have, this incredible company joining us. it's called Open Pipe. I believe both Kyle and David have been on the show, but if not, we definitely talked about some of the stuff that they did. Kyle's definitely been on the show. Open Pipe has been, with the industry for a long time. We talked about them, training models and fine tuning models and building pipelines for forward tuning models. And recently they have, focused on, our l and SFT workflows, in a, framework called art, which we also mentioned. Open Pipe has open sourced art and has been developing it and will continue to develop this, as, part of Coral Reef's mission to build the right infrastructure and tools to serve AI workloads at every level of obstruction from metal to token folks. So now we're expanding our whole offering. I'm super excited about this one 'cause it plugs right in, where we need. CoreWeave runs things on bare metal and some incredible optimization. Then big companies like Open ai, for example, use, all the way to Observability on the model training layer from Weights, & Biases. Everybody knows and love to now the ability to train and, build libraries to train sfts, track all of that in way its Biases models. Then because of RL you need rollouts. You need to see whether or not your agents are doing the right things as part of the RL and converting those signals back into, Signals for the model. You can track all of that in Weave our Observability and evaluation software. Then you can evaluate those models and eventually you can also post those models and inference. And we're working really hard on getting the inference for custom models up as well. So as you see, the folks at CoreWeave are really working hard on every part of the stack, complete end-to-end, every part of the stack. And because of that, when you use one of our products, for example, inference, you would be able to use the other ones like, evaluation and tracing. If you use, the art library to train and support SFT, you'll definitely use Weights, & Biases to track all of this. So as you see, we're gearing towards a holistic solution here. cor wave is doing the right moves, I believe, and I personally love Kyle and David. So huge. Congratulations folks, for this incredible, milestone achievement. I know it's not easy to run the company. I definitely know that it's a celebratory move when a company that CoreWeave, acquires you as, have, have, having gone through it. Not as a founder though. So shout out to Kyle, David, and some of the rest of the open pipe team.

Yeah, also open pipe release and, and interesting benchmark,

interesting, code example model this week using this specifically. It was, and it was really, really interesting. completely. Let me just show you actually it is right? one moment. So basically what you just talked about is this specific trainer system for models. what they show here is that, nearly for free, $300, which is for free on, in this, in this domain, for in couple of hours you can get an agent to do deep research on, state-of-the-art, performance level, which is Kwindla crazy. It's only 14 billion parameters. and the way, the even more interesting here is that it's all automatic. You can use this system for whatever you want. but, you can basically ask this, using this system to do whatever you want, in terms of agent use, whatever tools that you give it, and it'll automatically just go and generate synthetic data for SFT and then apply A-G-R-P-O and evaluate against whatever benchmark that you want, all automatically. So basically, we are at the point where you can just ask, Hey, make this model. Funnier or train this model to use this specific API or this specific tool, and it'll automatically do all the steps for you, and just give you a model that is better at this specific benchmark. And it's really, really, really useful. I mean, I'm not sure that people realize how powerful widespread RL really is, and it is powerful, and we're just starting to see all the automations of it and Wild times, and absolutely great acquisition. And by the way, it fits perfectly into the pipeline. It's absolutely clear how it fits into the pipeline. great Use

building blocks to be a holistic platform.

I absolutely love this. And, good to see. all right, we're moving on, from this week's buzz, we have two more things to cover. super quick. Apple released a fast VLM. It's good to see Apple release sound like a week before they actually released. let me just show this super quick. yeah, lemme take your stage, get this to stage. Alright, so Apple released, fast V LLM seven BA speed first vision encoder, and it's super cool. It's like three models, seven B, 1.5 B, and 0.5 B fast vision models. They run like super quick. I really wanna show an example of this, here live. Let me try to get, click this link, because I actually don't think that this example is the right one. I think CLA posted the right example, where they have, this festival running and the coding and Apple video in real time. I think lamp poster is, am I wrong? Am I wrong? It's been cla post a lot. yeah, this, so I will, let me, so you can see Kwindla the video that Apple posted I think last year. And on the bottom right you can see live captions and those captions are like maybe 250 milliseconds, behind of what happens on the screen. So this like actually takes, the video model the fast VLM video model and runs it on web GPU, Which is incredible in real time because the models are tiny and like, wait

a minute, is these, it is that small?

Yeah. Like you're gonna, oh, okay. That's changes everything. That's really, really cool.

It's a very tiny, very tiny bill and it runs like fully on

device And I really found it funny because somebody reacted to Clem and said, Hey, it's not that good. and they posted like this example of, the person riding in a, like F1 car and the life caption was the image depict in an older man wearing headphones, glasses, and a black polo shirt with a headset microphone attached to his earpiece, which doesn't seem like what's happening here. And Clem was like, no, definitely not perfect because it was like a millisecond delay. 'cause the one frame before this was Tim Cook wearing a polo shirt and headphones and like this literal caption was basically on the frame before. So it lags a little bit. But if you imagine, something like. Like a live translation or like a live transcript where people Kwindla like transcribe and translate. that delays significantly larger than this. So this model is able to understand what's happening on the screen in web GPU live on the computer as the video progresses, which is like super, super cool, That's,

live subtitles, man.

You find Finetune this thing for life type titles. That's incredible. Seriously for

the vision thing.

Like what actually, like what's happening on the screen? Which I find Very, very impressive. so shout out to Apple for open sourcing. This, let's take a look at the license super quick. on this LDJ, you already know it by heart, but I think we could just take a look. And this model is up on the Apple models, hug and face and fast VLM, and 0.5 1.5 and seven. they submitted to the CVPR in 2025. I don't see a license here. Apache Apple usually is like more respect with license. but I don't see a license here yet.

Yeah, I'm not sure what the license is either.

The highlight there is, this model runs with MLX and then, official

repository allows you to use this in IS and MAC OS apps so that you'd be able to build vision enabled features right into your mobile applications. It yu. This now is the, the conversation. Qwen, thank you so much Qwen, for joining us. he had to leave, but he dropped a comment here and now we should definitely talk about Codex versus cloud code discussion. let's talk with Sam Alman posting about this and then improvements, and then we'll continue and tell people where we are in regards of this, battle of the coding agents.

Yes.

Sure. So here, I might as well send you the link to the thread that I was referencing. Actually I already have it right here, so I'll put it in the stream yard shot. It's right there. Yep. And yeah, in the past two weeks, 10 x increase in usage of, codex of opening AI codex. And they see improvements are coming soon, and he gives some other details, which I don't have a top of mind right now, Sam

says, usage up 10 x in the past two weeks.

That's incredible. And like many folks have commented on the actual hype not being just hype. Yam, I think you're one of those folks you wanna tell us about Kwindla, some of your Codex experiences. So again, for ing, you have no idea who Codex is, is A-A-C-L-I tool that you install. it used to be a model in Open the Eye, and then there used to be like a different thing, like an agent thing. And now Codex is a CLI tool that is cloud code comparable. You can ask it to do stuff and it like integrates with your id, replaces your id, and you can just basically ask it to code for you users GPT five, I believe, and their own product model interchangeably. Mm-hmm.

Yeah, absolutely.

I can show you by the way. yeah, one versus the other. I haven't prepared this. I just opened, you know, set it up and we can see things like see it live. One moment. I'm not the one showing. Okay. You are the one showing now. Yes. Okay. So basically, if you don't know, these are CLI agents, they run in the terminal. why people like them so much? First, they are slim. They are small. You can just ask them whatever you want and they can do it. Second, they can use your computer. They run in the terminal, they can have access to your entire computer. They can do whatever you do on your computer through text. So, this is why people are using the terminal engines nowadays. And, you can, and, you know, there are ports for IDs and so on because people are asking online, why would I use this? So, that's pretty much the reason. on the right we have cloud code right here. one moment. Yeah. On the right. Oh, okay. Okay. I guess we, need an update So, we are going to try to update live as you can see. but if it's not gonna work, we are just gonna leave it like that. so apparently we don't have a cloud code with us today, but we have Codex. Codex is, I think it a great highlight

of why people are getting excited about Codex because

this one works in Sun Code for some reason after we decided to just quit. Yeah.

Look, that's on my end.

I didn't, probably didn't update it. I just, recently had the max plan, so, you know, I need to set it up and I used the Codex over the past couple of days, so I just didn't know this. Anyway, that's the solution from OpenAI. it uses GPT five. it was hyped up, the hype was building up on X and I was curious. It is the hype real, and I got to tell you, yes, it is. codex is really good. it started as, Kwindla an experimental POC early stages. and people just remember that this was the case, but it's not the case anymore. It matured a lot and now it's a really good product that you can use. Basically, you can just go and ask whatever you want, like, what files do I have on my desktop, for example. As you can see, the model is working, doing whatever it needs doing, and that's v PT five. When it says listing desktop files, it is calling things on your computer, running code on your computer. As you can see, it's writing scripts at the moment, running on your computer, getting back outputs, and just doing the entire loop for you. The cool thing is that not only I can tell you to find stuff on my desktop, and I can also tell it to organize my desktop, or, if there is code to look into my code, to look into my terminal, it just has access to your entire computer. that's pretty much it. I don't have any, specific, very hard, challenging task, on the fly, just to show you how well it works. but I did try it extensively and it works really, really well. GPT, GP five engine high, GPT five, the high as an engine is an absolute beast. Absolute beast. And you can switch those with slash

models, right?

Mm-hmm.

You can switch those with slash models. By the way, it's all open source, open ai, absolutely based for releasing this open source. And also supports if you want, you really want, you can use whatever endpoint that you want, even from competitors. So shout out to OpenAI for, for this. And, yeah, it's a really good thing. I just wanna say in my own opinions, the engine is great. The, the product is great. cloud code has, had more time in development, so it's, it, it feels Kwindla minimal. the engine that is actually running the model cloud code feels more baked. You can say, the tools, the prompts and everything. It's more, Alright. It has mcps. Absolutely. Oh, Cox has

CPS as well.

Absolutely.

Amazing. It's, none of them are configured at the moment, but it absolutely has MCP support. you can switch the models here. these are the GPD five versions, and, you can go on and change those. there are some other options like, what you want to auto approve. you don't want to give, permission to the model to do whatever it wants. So you can restrict some things and approve and ask and tell 'em all these, these things I need you to ask me to approve. And these things you can do on your, on your own. How

does it work?

Does it authenticate against your Open the eye account? does it like API key, like how, how do people actually use this? I,

I, I authenticated, with my Open AI account, I have the open ai, the

Pro account, the one, the expensive one. And it just on, on initial, on the initial one, it tells you, whether or not you want to, log in with your R GPT account. It's very, very simple. Just show you your link. You press with your mouse open a browser. You authenticate and that's it. It starts.

So no API, keys to paste.

Like with cloud. Next, the, the highlight, yamal, thank you so much for the coverage. I will say two things. First of all, Qwen who was just here, said, that, he was a hundred percent on cloud code. Two weeks ago, and now he switched a hundred percent to Codex this week. So also Qwen has give, gives you that data point. And number two that I wanted to highlight is this thing that Seman retweeted, which is, from Tebow. Tebow, works on Codex in open ai. It says, enable pro tip, enable Codex code reviews. It's been described multiple times to me as superhuman ins, built spots, double flaws. Navigate to Chat GPT dot com slash codex. Speed onboarding enable for your repos in the setting, and then deep reviews on every pr. So basically you'd be able to tag and assign pull requests in GitHub to Codex as well, because Codex is not only a CLI tool, it's Kwindla like the, the offering, the suite of things to happen. So, yeah, codex is a little bit confusing because it's also, the Kwindla, the open the eyes, like online thing where you can like kick off inside like an agent online on the cloud. and that's really

cool in a completely different way.

It's cool. Yes, that thing. that thing is really, really cool because you can, walk with it in your pocket and I don't think people realize that you walk with this thing in your pocket. It's really cool in my opinion. And they also have

integration to courser and windsurf as well.

So you can hit this up and open in windsurf and then have Codex, built in Kwindla like cloud code as well, to just integrate where you integrate. So not only CLI, it's also an extension to your editor and also a cloud running agent thing, and also a PR reviewer as well. So you can have it, review your prs, which is super cool.

I just wanna say OpenAI please.

there are five different things called Codex at the moment. please don't call everything Codex because, we talked about the API, the CLI now, and now we are talking about different codex on the web. There is Codex a model, which is a Finetune of Codex, and there is the, you get the point. Just give them different names. It's hard. Give them names, but yeah. Alright,

folks, I think with this, very interesting Kwindla

switch in the AI coding CLI agents. We will conclude the show for today. It was incredible. Exactly two hours. So thank you so much. Everybody who joined us we're clocking in at around 8,000 or so views, which is incredible. Shout out to everybody who joins. if you are joining for the first time, Thursday, I has been here for the past two and a half years, almost without fail. I think we missed like one Thursday, like a year and a half ago. And we're not planning to miss any other ones. to give you updates on everything that happens in the world of ai. just 'cause we love to talk about this as friends on, on the show, and many people just told us, Hey, this is really fun. We are participating with you. So, favorite time of the week, Thursday, we had breaking news today. We had two great guests. well, three, great guess. We had two folks from news research, And then we also had, queen La Kramer, our resident voice expert to talk about GBT real time. That's in ga. We had Nisten here before. LDJ and Yam Peleg are, here with me to end the show with you. My name is Alex, and the show is sponsored by Weights, & Biases from CoreWeave. We also had some exciting updates from that end as well. If you missed any part of the show, everything is posted as a podcast and the newsletter immediately after this. So please subscribe at Thursday I news. And with that, I think we're at the end. many You wanted to comment. Last thing

feel the a GI.

And we're here until AGI runs Qwen. We see you're teasing Qwen. We see you guys are teasing. We're just saying we're waiting. We saw it waiting and we're gonna cover

this next week.

With that, everybody, thanks so much for joining and we'll see you here next week. Bye-bye everyone.