ThursdAI - Apr 3rd - OpenAI Goes Open?! Gemini Crushes Math, AI Actors Go Hollywood & MCP, Now with Observability?

Alright, fine.

Here we go. Welcome to Thursday ai. I'm your Oso enthusiastic host, Alex here to bring you the latest in AI today. We've got some thrilling or not news for you from open AI's announcements to whatever else is on the agenda. So grab your coffee or don't, and let's get this over with.

Woo.

Welcome to ThursdAI everyone welcome, welcome. ThursdAI April 3rd, not April 1st. Nobody to be fooled. April 3rd. Welcome folks. My name is Alex Volkov as always, I'm an evangelist with weights and biases. I'm your host for today and, there's so much to cover today that hopefully we'll get through all of it. But there's a bunch of open source, a very, very exciting week, one of these weeks where I started by being very busy at work. I'm gonna tell you some of the stuff that I did at work that relate to Thursday as well, and I was playing with the native image mode of GPT, like all of us, probably like the 130 million people that used it in the past week and I was playing with LGBT. I was doing some stuff. I was learning. I wasn't like really, obviously I was paying attention to everything that happens. And then yesterday when I prepared my notes for the show, I looked at a mountain of updates that we need to go through, just like an absolute mountain, a bunch of open source stuff, a bunch of big companies stuff, open the eye is about to be open again, folks, we're gonna cover this as well, but apparently open the eye is about to open source, something soon and you can even sign up for a thing and they talk about this being like, basically we've been waiting for it. finally, we got some updates for them as well. And a lot to cover this week. but this is why we're here to keep you up to date. For those of you who are new to the show, a brief housekeeping. I don't know what to call this. ThursdAIce is brought to you by weights and biases. We don't have other sponsors. Everything we talk about here, unless we specify in advance or during the thing, nothing is sponsored. we just love the folks we bring on and the people we talk to. And, the motto is, we stay up to date. So you don't have to. And that's been the deal with Thursday. I, for the past two years. Today is show number 99, and so our next show is gonna be a hundred episodes, at least according to Substack, which we have. And if you go on Substack, you find ThursdAI please subscribe and share with your friends. at least according to Substack, today's show, 99, which is crazy to me, that we did a hundred shows, because there were a lot of streams that weren't planned. we did some streams on Friday that didn't count as an episode, but officially on Substack, since we started posting, we're up to a hundred shows. We also crossed, and here I will use some of my audio effects. I just noticed that we've crossed a hundred thousand downloads. At least on Substack. So we crossed more. this is crazy to me that a hundred thousand downloads of our episodes has been around for the last two years. And many people write to me and write to my friends and write to my colleagues and say, Hey, ThursdAIs something we listen to every week. this fills me with a lot of joy as well, but also a lot of commitment to keep this professional for you guys and make sure that this is high signal and not fluff. And this is also a point for me to invite feedback. If you have any type of feedback about how to improve the show as listeners, as community members of ThursdAI I would welcome it. And, thank you for participating in the community with this. How about we start TLDR.

All right folks, we are at the TLDR. So this is everything we're going to talk about outside of breaking news. Okay? So we'll also have breaking news, but this is everything that we're going to talk about on Thursday. I, so I want to run through this just in case you won't be here for the next two hours or so. so you don't wanna miss any part of this. Everything also will be in the show notes as well. All these links I'm sending them every week in the show notes on ThursdAI do news. So if you're missing any part of that, you shouldn't be worried about, oh, I don't have a link. All the links and everything you'll receive in your inbox if you subscribe. So I recommend you subscribe to this. we'll start the show with open source. There's a bunch of stuff in open source lms. the first big one that we're gonna chat about is open. The AI is about to be open ai. get it. they're about to open source a reasoning model. Apparently Sam Alman posted something like ta-da. We're about to open source something. we assume it's a reasoner. I don't know if they a hundred percent confirm it, but I've seen like multiple folks talk about this, so we'll talk about that. They didn't open source yet, but we'll cover it. they did open source some other stuff though. So OpenAI, also open sourced paper bench. A new benchmark. A new eval that actually goes to show how awesome they are in doing, because they released an eval where they are not the leader. Claude is the leader there. So we're gonna talk about paper bench. Paper bench is super cool. we'll also cover Dream seven B. It's a diffusion model. Do you guys remember the, we talked about diffusion language models. So we're gonna talk about this diffusion model, soon. it's Dream seven B is like the claim state of the art diffusion model. yam was very particularly excited about this, so I'm looking forward to hear from Yam, about this model as well. We also have open Hands, theory two b, a fine tune from Quinn, and for this we'll have a guest. I see I'm hoping I pronounce this correctly. Shia Wang from Open Hands. You guys remember Open Hands? We've talked with. With open hands about, their coding agent. and, now they also released a model and there's like also a cloud offering. So we're gonna chat about this reaching out very soon. then we will also talk about, it's not in my notes, but I have that here. we're gonna talk about, Nomics multimodal embed, which, I have Zach nbo, who's been on the show before. He's gonna join and, chat to us about how, they've trained a state-of-the-art multimodal model for embedding PDFs, which is super cool. Then in the big companies, LLMs and APIs, we obviously have to mention, although these prices, like these tickers don't make sense anymore. OpenAI raised an additional $40 billion at a $300 billion valuation, which is more than Coca-Cola and other big companies, more than Disney, more like a bunch of other stuff. Open the eyes on track to become like a trillion dollar valuation company. and also to cross 1 billion users. So that's insane. they also have this new Monday voice, which we'll talk with Monday. Super cool. I don't know if you guys chatted with the EMO GBT yet, or it's pretty cool. So basically Amazon announced this NOVA Act, which is a computer use API. We're gonna chat about this as well. there's a new benchmark from Math Arena called U-S-A-M-O, so Math Olympiad. And, up until Gemini 2.5, none of the models get more than 5%. Gemini crushed it. And we're gonna chat about what this means with 24%. So Gemini 2.5 on this one benchmark compared to every other big model out there gets 24%. Everybody else gets less than five. we have to try to figure out what's here. and then, also there's a new update from Core Weave and NVIDIA, where they announced like 800 tokens per second for LAMA 400, B on their new ml perf, like five benchmark or something. So we're gonna chat about this as well. That to me is crazy. Just absolutely crazy. 800 tokens per second for LAMA 4 0 5 B in this week's buzz. I now call this week's Buzz MCP coverage as well, because to me, they're combined. We're leaning so hard into MCP and all. I like just keep telling you about this. I have an update, I have a request for you as well. I have a whole story to tell you about MCP in this week's buzz. Something that I'm leading personally, inmates and biases and I'm very excited about and really want to share with you. and also, yeah, so this week's buzz in this week is worth, waiting for. We're probably gonna do it an hour in, and I will talk to you about MCP observability, which is a thing that I think that I needs to exist. And, I hope to convince you that it's a thing that needs to exist as well. Vision and video category has been crazy this week because runway started this week with announcing Gen four and Runway has been in the video generation game since Gen one, gen two. The Gen three, gen four looks absolutely ridiculous and will show some example videos. it just looks so, so good and I've seen a lot of, creators already play with this model. Do you guys remember Omni Human from Ance? on the show, you may remember so only human is this like avatar creating. Model from Biden's that they announced but didn't release. Finally, it's available. It's available via the dream in of website and we're gonna chat about that. But not only that, there's another contender that we haven't talked about called Hetero Labs that I've used and, recently many people use as well because they want to animate their image generation from Chad g pt. So I'm gonna show you both Omni Human and hetero labs and show you kinda the comparison that I think, is worth looking at. Now, basically, both of them receive, an image and some audio, and then they generate the avatar. So it's not like a lip sync where you take a video. And then, we also want to, in this. Era as well. We wanna chat about Mocha from Meta, which is a movie, great talking character synthesis. And this absolutely looks just insane, just absolutely some of them cross down Candy Valley for me very easily. So we're definitely gonna show this, folks who are only listening, you'll get from the context when you need to tune into the video. So don't worry, you can go about your day and just listen. once we get to the video though, this specific segment, vision and video is gonna be very visual and very obvious that it's worth watching. So definitely worth, watching as well. If you are listening to the podcast, by the way, I have now recently started embedding the live show that we have. So the thing that people are watching right now, hey folks, at the top of the newsletter. So if you are listening to the show and you actually wanna see some parts of it, the YouTube show is going to be, which chapter. So you can skip to the exact chapter that you want, and that link is on the top of the substack for you, to find super quick. Voice and audio. I have two updates here. Gladia, our friends from Gladia French, company that does transcription models that always try to beat Whisper. They just, decided, that they announced that they released solaria, which is, they call it best in class AI transcription model and translation as well. it's super, super fast. we're gonna chat about this as well, and Helio Minimax audio speech too. That's a mouthful. Helio AI is a company that has a video model and honestly, they have a bunch of other models. Their audio, though I've talked to you about their audio me is the best by far. People still think it's 11 labs because 11 Labs has the better brand, but you guys know that I bring you the nitty gritty of ai and Hallelu is absolutely much better as far as I'm concerned. In audio, they have emotional speech, they have voice cloning that is crazy good. And they have long context as well. So long text, you can send up to 30,000 characters in there. you can generate like almost a whole podcast in there. The voices are crazy. The cloning is crazy. The emotional control they have is unparalleled. not open source at all, but, speech O2 from, minimax is definitely great. AI art and diffusion. We should just mention that, the whole world became AI artists because more than, 130 million people have used and generated over 700 million images with Chet since last week. we're coming up on a billion images generated with Chet since last week, and especially in India as well. it looks like India is onboarding super hard on TGBT. I don't have any other news in the AI art. we'll see if something comes up, but folks who are listening from the AI art community would love to know from you if there's any news that I missed, because I know that this field is very hot and oh, yeah, I see already. So potentially we'll see like a new Midjourney, today. Thank you, Colleen. there's been talk about, about new mid journey as well. Then in the tool section, in the last section that we're gonna talk about, I wanna shout out our friends from Windsurf. If you guys remember, we had Kevin here from Windsurf talk about, cascade and why Windsurf agents et cetera is better. I use Windsurf since then. I really like it. they, added a new bunch of releases including deployment. So now from Windsor you can actually deploy your stuff, which is super, super cool and they're gunning for the bolt and the lovable type of products that we also covered. And then also notebook, LM adds discoverability of sources. So in notebook, LM the thing that generates the podcast from Google. Yeah, you guys remember, you no longer have to bring your own sources, you bring some sources and it discovers other sources for you, which is pretty cool. So folks, this is our TLDR. With this, I wanna welcome LDJ to the stage. LDJ. welcome man. It is been the minute since we've seen you. How are you? How are you doing? Let's jump in into the open source stuff. Yep, let's do it. How you been, man? I've been pretty good. How about yourself? very good. All right, I think with this we're gonna start with open source. We're gonna invite a few folks here to the stage. but before, we'll just, do the open source opening. Folks, let's get it.

Open source ai. Let's get it started, folks, and now we are welcoming multiple folks. So we, you guys know we have L-D-J-L-D-J, cohost, thank you LJ for joining. And now we have two more folks to welcome. we're welcoming Zach Nusbaum, not your first time on the show. You've been here before. Welcome, I think. Not in video form though, so welcome. Great to see you, Zach. Zach, you work at noc, correct? yep, that's correct. Thanks again for having me. Absolutely. It's always a pleasure to have you on. And we also have, Shing Wang from Open all hands. So all hands, the company Open hands, the AI coding agent. You guys also really something super cool this week. So we'd love to chat and hear from you what that's about and what you trained. but I think, just by order of who's been here first and who I invited first, we'll start with the Nomic stuff. And, Zach, I think that you guys have an announcement this week, not the first time that you guys are announcing things on Thursday. I, I would love to hear from you directly what we're talking about and like talk, talk through the details.

Yeah.

Thanks again for having me. always fun to chat about the cool research that we've been up to and, do it Cool. show like that, that you guys have. but yeah, we released, NOC embed multimodal, state of the art, embedding model that allows you to, embed image text, interleaved. previously most multimodal embedding models required, two, encoders. So like clip style. You'd have a separate text encoder, and a separate vision encoder. But a lot of the problems would be that your modalities would get embedded separately. require you to have, two separate, vector stores, one for your image vectors and one for your, text vectors. we took a bunch of, our learnings from, training, high performing text embedding models and applied this to, the multimodal work that, the great folks that led the co poly team and the DSC work. I think Manuel and, I'm forgetting, the DSC leads, Team name, but yeah. we basically, built, on top of great open source work, and we've released, two, really great Apache two license, seven B embedding models that are state of the art. K noc, is the top of the leaderboard for, I'm gonna butcher the name of the leaderboard, but, uh, video, video re or video retrieval or visual retrieval. this one, VV two. Yes, exactly.

Yeah.

so yeah, we're top of the leaderboard there with our

multi-vector or later interaction models. So this leverages a lot of the same approaches that, called bear models use. and then we also, have a dense model, just noic embed multimodal seven B, which, outperforms all o other, single vector, multimodal and embedding models.

That's so cool.

So first of all, thanks for coming up and talking to us about this and, huge congrats on the state of the art embedding models. I wanna talk to you about sizes for a little bit. Okay. 3 billion and 7 billion parameters for both these omic and Omic embed multimodel. what is the difference between the two? Like how do folks select which ones they want?

Yeah, so the seven B, both these models are built

off top of the, Quinn 2.5 vl. the seven B comes with an Apache two license, and you'll get slightly, better performance, the three Bs. I believe they are under a Quinn license. so I think, for research purposes, it's all good, but I think for any commercial purposes, it falls under, a non-commercial license. with the single vector models as well, the, embedding dimension is slightly larger for the seven B, but you also get, a decent performance boost. so it all depends on your compute setup. I think right now most people on, apple computers or CPUs, can run, seven B parameter models, fairly well. nice thing about this is you're not generating any tokens auto aggressively. You're just outputting a single, vector. should be faster than, if you were running the Quin vl, through generation. we also have some, demos, later in the blog post. I can also drop them in the chat as well, showing how to use both the multi-vector, late interaction model and SIE L Vector model.

That's awesome.

one more question I have for you, Zach. Embedding models. we talked about embedding models with the backbone of, like mistrial, for example. And this is the first time that I remember we talked about like huge embedding models. And by huge in the small LLM category, but huge compared to embedding models. embedding models, they're tiny. in the hundreds, millions of parameters. how should folks choose, like why go all the way to A seven B when there's is the performance that much better so that folks need to actually run like a 7 billion parameter model? Or is it, more useful for folks to run like a very small, very quick model? Because sometimes for embeddings, maybe when you create the vector store. that speed doesn't necessarily matter because that's kinda offline. When you create this, maybe update this, but at query time, right? When the user drops something you wanna do, like a search, you want to embed that like super quick. So maybe for that, you want like a much, much faster model. how are you guys thinking about what are the best practices in using like smaller or bigger models?

yeah.

this is a really good question. I think it gets to the crux of a lot of the things that we've been thinking about, when we're training our models, you bring up a good point about query time, right? if you have a larger model, and this is in your rag pipeline, you don't want your retriever to, be the bottleneck in your pipeline. which is why I think, we tend to aim for smaller models, and, allows you to also run these models, on device, on, consumer hardware. Another, practical consideration is the, dimension of your embedding. So with larger models, your output dimension gets larger and you pay a linear increase in memory. compared to A one B and, 150 million, parameter model, like you're paying, a two or three x, increase in memory, which, adds up when you have a lot of, data, right? And you have, millions to hundreds of millions of documents. one challenge I think right now that exists in the open source ecosystem is I think there's, potentially I. I don't think that there's, many high performing VMs out there that are under 3 billion. in general, right now, I think that choosing the best model, and for multimodal data, I think requires you to have a very, high performing and strong VLM base.

which is, Kuen 2.5.

Vl. Yeah. And, shout out to our friends from Alibaba Quinn. By the way, shout out to Junior Yang, Han and like the whole team for this just incredible workhorse of the industry because like many folks, this is what they're now building on top of. And Zach, thank you so much for coming up because it connected to this. The next thing that we're gonna talk about is also a Quent thing. Zach, feel free to, to drop or stick around with us if you want. we are gonna move to talk about the next part, which is, also, quent fine tune, I believe. Xia, would you like to present yourself? this is your first time on the show. We've had, Graham before. And, but this is your first time on the show. Would you like to present yourself first, and then we're gonna talk about what you guys released.

Yeah, sure, definitely.

hey everyone, I'm Elle, I'm a co-founder of All Hands and also like early contributor of Open Hands as the open source project we are building.

yeah.

So Open Hands once you guys, launched it, I believe it was also, junior young, related to this. I remember like Junior young posted let's do some open hand stuff and then the community started from this and then, yeah. and then you guys actually took it into a whole thing. So we've covered, all hands before we've covered Open Hands, which is the, which started as Open Devvin, I believe, if I'm remembering correctly, right? So folks, if you guys remember Devvin De Coder, where's Devvin now? I dunno. but not even in the ranks. Fine, but, that kinda excitement about like a coding agent that, I don't know, they got like a crazy valuation, got folks to think, Hey, we can build this in the open. And then folks started contributing as, you did as well. And then afterwards, all Hands came around as a company as well. which you are a co-founder of, correct?

Yeah.

Yeah.

Awesome.

three of us met through the open source.

right after the day of demo release, we have open demo and then we just jump into the community simultaneously.

That is so cool.

And I remember we talked about open dev, with Graham before, and then we covered like a month later. the reason why you are here is because this week you guys also have some news, would love to hear from you what you guys announced discussed this week. I think you have two updates.

Definitely.

the first update we have is we have this open hands LI release, which is it's more about a research preview, but like we got some encouraging preliminary result. We got about 37.2%. Also binge verify, which kind of the performance is a comparable to model with 20 times four parameter, for example, like Deep Sea. So one complaint I heard from community members, like we lack like a decent open source model that people can just self post it and run by run for themselves. So this is actually like our first attempt here.

so let me just, recap this for folks because

I find it super incredible. you guys started as a rectangle, folks on GitHub doing some open source stuff to try to replicate them, and then you became a company and then you released one of the best performing agents for software engineering. That's now like one of the best performing ones. And we're gonna talk about open, live Suite bench in a second that you guys are also topping. and then we've chatted about like open Hands as an idea, and then now you are releasing a model, like a full language model of fine tune on top of Quinn 2.5 as well, right? Which is yeah, to provide code there. Yeah. 32 billion parameter coding model specifically that gets a very high performance on multiple things. so let's talk about, some of the parameters. Folks. Love to hear the numbers.

yeah, definitely.

so primary number we are chasing here is specifically on the venture verify, which we are getting about 37.2. for example, the latest Deep C got about 38% within the open hand harness. So this is more about customizing the coin model towards open hand specific use cases rather than trying to improve its underlying capability, from all dimensions.

Yeah.

so you're fine tuning, you have a specific benchmark that you're dropping. are you being careful at not overfeeding for this one benchmark? Could you talk about 'cause folks always say when there's one, higher for on some eval than others. there's the scare that, some data is leaking. how are you guys being like, careful about not having the sweep bench stuff actually in the training data?

Oh yeah, that's definitely a very big concern for people.

So actually we have a research paper called switching. So this is in collaboration with some folks at uc, Berkeley and stuff. So we actually built like 2,400 tasks. within this switch chain repository, we make sure that there's no overlap between switch chain and swb bench. So there's like definitely no, data contamination. We are not training on the test sets. So yeah, and then we actually set up a environment for all of these repository. Basically we make sure that those units can run within the docker images. So in training time we simply run open hands on a lot of this, task. And then basically open hands, we generate a bunch of agent trajectories and to solve the given task. by the end we just run the unit test to see, if some of the trajectory already solves the task, then we simply train on those successful trajectories. and then the open hands LM release is simply a scale up version of the paper. in the paper we find that, if we are only training with 500 trajectory or 32 billion model with 2K context window, we can already get like a 13%, result rate improvements on the 32 billion model. And then the interesting thing we found is like switching 30

13% improvement on top of the wan.

That's already good at coding,

Yeah.

I think the caveat there is like the good at coding doesn't necessarily means they are good at a coding agent style. Because coding agent style is more about multiterm tool use I think it's not a hundred percent related to the coding capability of the agent. Rather about whether the agent is able to, for example, self-correct itself When it runs into some errors or how to properly use tools to basically edit files. So those are the actual things we are, targeting in improving, at least in this version of Openhand sale I

That's awesome.

Yeah. I think because we

already believe Quinn is like already pretty strong

at the generating code, so Yeah.

I think the distinction is very important.

LDJ like we should definitely like, pause here and talk about this. coding ability does not necessarily mean real world ability. That's why SW Bench exists. That's why SW Lancer exists, which I believe SWR translates, tasks from Upwork and actually gives them a specific price and then SW will answer. Yeah. From also open the eye, adjust, how much money would this agent be able to get from the pool of tasks within Upwork. and so I think that like you're absolutely right. There's also like fill in the middle, there's tools, there's like bunch of stuff. Xing I wanna ask you, you mentioned you guys are testing for agent's ability or like the model's ability to like self-Correct. We are now in the third like expansion laws, whatever, with test time compute, which is basically the ability of the model, to get the aha moment and realize it's wrong and try to correct itself. You didn't release a reasoning model. ha. Have you guys played with reasoning traces at all of trying to like, fine tune, et cetera for this effort? Like how do you think about that aspect of self correctness from perspective of there's a whole crop of models now that do this.

Yeah, that's a great question.

We actually have one Preprint earlier where we study a bunch of like recent models in like agent context. So one thing we found, at least for now, most recent model today aren't really good at agent stuff. So they're just like living in their own dreams. For example, they would just say that inside their reasoning context, they'll say, okay, I want to open this file. And then they will hallucinate the context of that file and then just execute a bunch of command and it successfully solve the task. So that's still one thing, but I think we've definitely see things getting much better with open most recent, O three mini, where they actually train the calling along with the reasoning task. So I think going forward, I think the critical thing is you need to balance the capability from both reasoning task as well as calling. Because if you're only training on problem of competitive coding and math problem is like the problem description is self-contained. Basically, all you need to do is think really hard to find a solution. But like a coding agent, you are given a GitHub issue and the agent actually needs to go out there to the GitHub repository and execute a bunch of commands to gather information before they need to start thinking. Yeah. So I think that there is like a gap there. I think we need to bridge them before we can see like the model become like exceptionally for calling.

So to speak to this, I want to talk about, obviously there's this we

bench verified, which we've talked about. weight and biases used to be the number one agent, generally on the top there. and now, recently this week we've been beaten by a little bit, but, those like agents do like multiple traces. You guys have been on the top there as well. I remember Graham, you're also still like runners, but there's a new suite bench like benchmark that I wanted to talk about as well called Live Suite Bench from the folks from Live Bench. we covered, supported by Abaco AI and live Suite Bench just launched and it's kinda like a different attempt at these things. And Live Suite Bench has you guys on the second, like almost on top and all these models, they don't use the model that we just talked about. the all hands, 32 B, all these models they use Claude. I believe 2.7 is VLM, and I think this is new like literally from today. I want to congratulate you guys on taking the second step here. And the reason why is basically what you just talked about, right? Is because just coding ability is not enough. This is like the instrumentation, the tool calling, the kind of the reasoning, the stuff that you built in maybe with prompts. could you maybe add. More reasons of why you guys at the top and not like windsurf, for example. Windsurf is also like doing great work in like energetic stuff. could you speculate as why you guys getting average of 46% and let's say, co cloud code, which everybody loves, which op like anthropic trained and it's their own thing gets like 39 on this task and you guys getting like 46. So almost more than seven points difference. Could you speculate as to why you guys are so good at this benchmark?

yeah.

Def definitely. So I think one primary difference between like suite agent and open hands versus like other competitive here is like we, we have a dedicated run time, which is like a containerized like symbol environment where the agent can actually run command and execute command. So I think like a cur, a cursor or wings serve, like that might not be easy because you are actually running command on user's machine. You certain don't want to destroy users' machine, but like that restriction is not there within open hand system. So I would probably suspect, but like one thing is because open hand is free to edit file as well as try to reproduce the bug and running command just to try to figure out what's actually going wrong there and then. Running a bunch of unit tests. Yeah, and I guess another thing is I think because in our development flow, we are actually trying to build open hands into a shape where it can resolve GitHub issue for us. So open hands actually improving Open hands itself, and we use it constantly. So it's like unintentionally we optimize the agent towards that direction where it solves task as GitHub issues. So that might also have some unintended effect.

I guess that makes perfect sense to me.

That explanation specifically because we live, we bench and SW bench verify, those are GitHub tasks for folks who don't know who they're, those are GitHub tasks that people collected. They have a very specific solution and basically the models need to get to the solution. and then if you optimize open hands for that work specifically, unlike, for example, windsurf, which is like generally built for open coding things like you just code everything with windsurf and so it's like maybe optimized for that. so Winf is still getting a very high score here. copilot is getting a high score cursor as well. very interesting that AER is getting a lower score. Maybe Aiden is optimized for like in, in, in place in film. one last thing, X you also had a cloud announcement, I believe. I want give you guys the chance to also announce that and then we'll move on to other things. But feel free to kinda shout out the friends who worked on this and then kinda share your thing. That's very welcome here for Friends of the pod. Which you guys are.

yeah, definitely.

So we, so now everybody can just log into App dot All Hands app, and then you can just use all the hands from your browser. We give everybody like 50 bucks in free credits and there're like a higher rate limits on Aaro credits. And we have, you are able to kick up like, Openhand conversation with your phone or inside GitHub, basically you can just tag it, any sort of GitHub issues and you just go out and fix it for you. Or you can just log into the web and then select a GitHub repo to get started. And then just ask it to like, basically, fix, put like fix merch conflicts and doing like short stuff. You don't want to do it yourself. So open hands, like what's also designed as a generalist agent, but like we just use it to resolve GitHub issues

That's awesome.

folks go to app, do all hands dev, and then you'll get 50 bucks rates. we're not sponsored, which is like level hands and they're absolutely one of The best coding agents right now for solving GitHub issues. Definitely. go and give this a try. And you also train your own model, which, what is the license? Just remind me.

license.

Everything's mit,

everything's MIT, which is great.

Shingo, thank you so much for coming up and talking to us about. Yeah, thanks for having me. We'll need to move on because we have a bunch of stuff to cover, but folks, definitely give all hands, a try. Open hands, like super cool project. Cheers, sh out. Thank you so much for coming up and we wanna welcome ya as well. Yam. Welcome. How are you man? How are you doing?

I'm fine.

I wanted to talk to the guy. he did some bunch of stuff, but no time.

yeah.

we'll catch him next time. They're friends of the pod. Definitely. LDJ. How are you doing, man? What do you think about their effort? I'm fine tuning.

Yeah, I've been good.

And, yeah, I think it's really important to have these different frameworks and not only just continue advancing and using newer, coding benchmarks, but also just like in general. Yeah. for open hands and all hands. but what you have on the screen now too. This is really interesting and I saw this yesterday. Yeah, go on Alex.

I'm just saying there's folks who are only listening.

So let's describe this. Oh, sure. Okay. I excited to, to talk to both of you about this specific thing.

Yeah.

so we wanna move on to the dream diffusion model. Yeah. Let's talk about dream. Yep. Yep. Okay. Yeah. this model here, it's called, I think it stands for, diffusion Reasoning Model. I think they just take the first few letters of each of those words and they end up with dream. but they called this a diffusion reasoning model. And apparently, based on these benchmarks that Alex has up on the screen right now, this is getting like on, like generally popular benchmarks, but they're starting to be less used nowadays 'cause they're so easy for the frontier models. But ARC E and Arc C, it's getting 10% or more compared to models like. Higher than Quin 2.57 B higher than LAMA three, eight B, higher than menstrual seven B. And even just like in, in big bench hard, even in benchmarks like MMLU, it's really impressive.

Yeah.

Human eval is also up there, 60 ish percent, almost 55. But what is interesting about this is that this diffusion, right? So could you like clarify for folks who are not following this closely, what is the difference between this and like 2.57 B?

Yeah, so there's different ways to do language diffusion, but in general, people

mostly remember diffusion from the main way that image models are doing things. And so the difference between auto aggression, which is what language models usually do versus diffusion, is auto aggression is essentially you have the typical, you're generating one token after another, one word after another. Whereas diffusion is Much more often doing everything in parallel all at once. And that might look something like in images starting out with a bunch of noise and then doing diffusion steps where for each step you do, it's making the whole image clearer and better and closer to the goal. And here what Alex has up on the screen as well, for language, it's like a bunch of random words or tokens are showing up from different parts of the output. So like things from the last paragraph and the first paragraph all showing up around the same time and it's constantly filling in the gaps just for each step that happens.

I found this like crazy to me.

We are humans. We're sequential. We think about the story, and we write it like from the beginning of the sense to the end. That's how we learn. And this is how auto aggressive models also they predict like next tokens. This thing just doesn't, like I see how this could be good for I don't know, an image model, but for language and the fact that it works and for code, the fact that it works just breaks my mind because the model can generate the last word in the sentence or in the paragraph before the second word. it don't make sense. what's exciting about this to you? Y why'd you guys so excited about like the digital

Look for, I can talk about this for hours if you really want to.

What I'm saying is that, many times you need to use the language models in real life, like for coding and so on. And you need to use them like somehow let them, okay, edit this part of the code. The thing is that they're not built for that. They're predicting the next token. So the tricks everyone like is doing is like. Putting, I don't know, some marker somewhere. And then, they predict after the text, and then you like, move it by yourself or tell the model to, what string would you like to replace? And then they output a string to replace and replace by what? And then you go, this is how coor and wind, all sorts of tricks. The thing that it's not native, if you think about it, it's not really native. It's like too many steps. Diffusion in, in two, in two words. The thing is that, I, images have, there are many, combinations of colors you can say, more than, way more colors on your screen than there are tokens, the types of tokens. So at the end of the day, diffusion models, work, work well mostly on data, which is continuous like images, like pixels and so on. when it comes to text, they haven't shown this level of success yet until this paper. And what is interesting about this is that if you look at the results. it doesn't really matter how the results are obtained. Like what exactly do you diffuse? Getting to these results is incredible. it's on power with models of this size. I cannot stress this enough. There are some projects that are waiting, just waiting for someone to back them up. And they're so important. that if they work, they're gonna be massive. And this is one of them. Because if you tax this to 70 B and you maintain the same level of performance, everyone can use this on their courser at home, it, all of a sudden starts to make sense. And so it's very important. I just wanna say it's a completely different, concept and first time where we see. some method that starts to, shine and might even, be competitive. Amazing. Amazing. Seriously amazing. Absolutely.

So this is so awesome.

I just wanna highlight a few things here. they're beating other models in this same category. So LAMA three, eight B, Quinn 2.57 B, Misra seven B is here, deep six, seven B is also here. They're fairly competitive on benchmarks with all of them. one thing we talk about is human eval. Human here is at 57%. So state of the art across these small models as well. Only Quinn 2.5 comes as close, but they have planning tasks here, and they have a countdown, Sudoku and tree planning. I just wanna highlight the huge difference in Sudoku. for Sudoku, this makes sense to me. Sudoku is a three on three grid of three on three numbers or something, right? Something like this. people have played Sudoku for a while, diffusing Sudoku. Makes perfect sense that you would need to look at the whole kind of thing and plan itself like that. They are solving 81% of Sudoku, and the closest one between the ones that they've tested is lada, a b diffusion, also a diffusion model. So 46 and Quin 2.5. So they're solving 81%. I don't know if it's percent, but like 81 points. And 2.5 solves only 20%. So that's very interesting because Sudoku is numbers, but kind of text, but not. so that's very interesting. Trip planning. They have like also bid a score. So very interesting. As Yam just said, if you guys have GPUs that you would like to contribute to some of these efforts, these folks, reach out to Chen and we'll add the links to the notes and then just sponsor this effort. I don't think this is a commercial effort at all. I think it's like some folks, correct. I'm so like, definitely very interesting. And we want to see the scale because we know that we scale stuff come and sometimes we scale stuff appear that they haven't been trained on before and then you can like then use this learning and then pour it into small models like we saw with the aha moment from deep seek when they got to this aha moment and they talked about this in a paper and they spent a lot of compute together and then suddenly people who have replicated that aha moment would like under 50 bucks or something. So definitely if you're listening to this and if you're from a big lab that has GPUs and wanna contribute, you have two choices. You can go and give them to Sam Alman that asks for more GPUs. I don't know if you guys saw this, or you can sponsor some open source efforts like this, like Dream seven B. the thing is I added it under open source, but I haven't seen the actual model. I don't know. The JM if you guys seen the model, I scoured their blog and I don't think that they released the model. So while we did put this under open source, because conceptually makes sense, I haven't seen the model yet, so it's not on hugging face. you cannot play with Dream as far as I saw, yet, but this is a shout out to the team from Dream. If you're listening to ThursdAIf you guys know them, tell them that we want these models in the open source, I think it's gonna benefit the whole community. I think we've covered most of it. Ah, yeah. Now that you guys are here, we haven't covered the biggest thing in open source in announcement. Open the eye is about to be well open. Let's talk about this. This is so exciting, folks. we knew about this because Sam Alman posted like a poll if you remember, and asked if you want a model size model, mobile size model, or like the best performing, general size, which I think we want both is the answer. And now, this week Sam Alman actually posted about Hey, in the next two months, I think this is the quote, legit, correct me if I'm wrong, in the next two months we're going to release a state of the art reasoning. I think I don't know if he said reasoning, but I think people conferred the reasoning model as well. so open air is about to give us some open source folks. and they claim that they want to make this right, and so they're seeking feedback from the community, to which I tagged all of you guys, everybody who's on Thursday. I just tagged in the reply to Kevin. and so they're gonna host feedback sessions, encouraging developers to sign up. So there's a form if you guys are interested. I'm gonna add this form to the show notes. LDJ, what do you think about this? Open, open. Open the eye.

Yeah, I think it's really cool.

And I just wanted to, correct, I think you said two months, but the tweet he put out, I think it just said few months. Ah, yeah. Okay. Few months. Yeah. but yeah, that's still maybe could be in, the next couple months or something.

from Kevin will coming months.

Yeah. So you're not super committed. It's,

yeah.

So it's exciting though. I remember they put out that poll. It was, here I was two February 17th, they put out that poll. so like roughly a couple months ago. and I was wondering like after the poll results, it ended up saying. That people wanted O three mini, basically the options were, would it be more useful to do an O three mini level model that is pretty small, but still needs to run on GPUs or the best phone size model we can do? And that's me reading directly from this tweet. And people voted 54% for the O three, like the poll result titled O three mini people voted 46% for the phone size model. And we didn't hear anything about it for a few weeks. And I was wondering are they just gonna act like that poll never happened? But luckily they're acknowledging it now. And yeah, it's

cool.

I'm pretty sure that release is gonna be tied to some huge Google breakthrough, like very close, like a day before maybe Google io. Google io is coming at May, 2021. Yeah, we're gonna cover Google io, directly from Google io. By the way. maybe a day before, if you guys remember last year, a day before Google announced Project Astra, which is like a multimodal, which by the way feels so. Common. Now we all have this, like we can talk to ai, we can interrupt. A year ago. This is crazy. A year ago, this is like a huge announcement. Like folks, we have a multimodal they can watch, et cetera. And the day before, open AI dropped advanced voice mode. and just showed us what's possible in this realm. So like a lot has happened this year. That's one. But two, open The Eyes is very known at undercutting Google's releases, including the last Gemini 2.5, which is incredible. We still haven't talked about how incredible Gemini 2.5 is. We'll cover this in a second. A day before that, or the same, I think it's the same day. Open The Eyes is like ta da native generation for images that we've talked about a year and now we're like finally releasing this. And they broke the internet basically with Ghibli. So OpenAI loves undercutting Google. So we can assume that maybe when Google's about to release something super cool, OpenAI will have an open source release, which is like absolutely take over the airwaves. yeah. What do you think about this, this thing and also the limitation that Sam said that we're not gonna do some silly stuff like limit this to under 700 million users attacking Zach directly, I think.

I think it's fair to say that everyone understand that it

doesn't really matter at the end of the day, if they release an open source model or not, it's not gonna hurt, their coffees or anything like this. People are still gonna use, their services because the services are good when I use the products, because the products are good. the API is managed and it's really good. So even if in future they're going to go with even larger model or more sophisticated model, I don't think it's gonna impact anything profits and so on. Just like Google, releasing demo and meta releasing Lama and so on and so forth. that's it. I think it's absolutely great, all for it. open ai, backing open source again, very positive. Very positive news. Absolutely.

We're very excited about this here specifically.

'cause our focus is like strongly on open source. We start with open source every week. we barely cover open AI and open source. Sometimes they release some stuff like paper bench, which we're gonna cover in a second. but generally, we haven't seen any models from them and we're very excited about this. So we're definitely gonna cover everything. They're probably gonna go livestream folks. We will let you know once you can use these models for sure. let's cover the next, and the last maybe in open source stuff, from Open the Eye as well, which is, you guys seen paper bench, a new eval that they released. and I would love to chat about this because it's very interesting. paper Bench is a new benchmark evaluation benchmark from Open ai, which I will show in a second on the screen that tries to replicate evaluating AI's ability to replicate AI research. Yam. As an AI researcher, how does this make you feel?

look, it's not easy to replicate ai.

I think even if you're an a GI, it's not that easy to replicate AI research. I must say that I haven't had the time to read about everything that's going on every single day. I just didn't get to this one yet, but it sounds exciting. That's why we're here. That's why we're

paper bench is a new benchmark from them.

agents must replicate 20 ICML, 2024 Spotlight and papers from scratch. So they actually, from the paper,

from like you read the paper, you replicate the

code and how do you measure this? Like unit test or something like verify unit test.

They,

oh, this is a hard benchmark.

Really?

they replicate the papers and they have LLM based judges to grade

the replication attempts against rubrics. And they also have my favorite thing, which they assess judge performance by creating a separate benchmark for judges. So this is like a super cool effort from multiple angles. First of all, there's 8,316 gradable tasks on several things within the paper. They collaborated with the researchers from my CML with the papers to kinda try to figure out what would it mean to be better at this task, right? but also, they have LLM as a judge to trained on the outputs of these people to actually scale this up a little bit because it's impossible. There's 8,000 tasks there. You're not gonna set a human and try to like, solve all these tasks as well. the thing that struck me the most is that, first of all, it's open source. Not only that, they also released an eval framework called Nano Eval. so a bunch of open source from open AI here, and then they evaluated a bunch of frontier models on paper bench. and the best performing tested agent is 3.5 sonnet new. So not 3.7, not 3.5, the old one. we call this new sonnet, 3.5 sonnet new, which was released somewhere in like December or November of last year. with open source scaffolding achieves an average replication score of 21%. So 21% replication score of an ICML 2024 paper. That's done with an ai. and they recruit top ML PhDs to attempt a subset of paper bench finding the models do not yet outperform the human baseline, which is, which Sounds good. So this is open source. folks are welcome to try this at, paper bench and on GitHub, the links will be in the notes as well. and, this, yeah, this is super cool. It's a hard benchmark, and I applaud open the eye for releasing a benchmark in which they're not the top. I think that's like super cool. That's every, applauded everywhere. LDJ, what are your thoughts on this?

yeah, I think just all around Cool and good progress that we're making here.

So absolutely, great effort from open ai and we

welcome the open source as well. I will just say, we'll, welcome the other open source much better than another eval. Another eval is always great, especially in the, PhD ml PhDs. So we'll definitely appreciate this, but I think we'll welcome the other open source, much better. I think that now we're talking about Open the eye, we should move towards big companies and APIs because there's like a bunch of updates there. Hi folks. I think it's time for me to reset the show. We're exactly at one hour. We just covered open source. I love when we get to actually cover like a full hour open source. And we chatted with Xiao from all Hands and Open Hands about their new release model. We also chatted with Zack Nesbaum from NOIC about the NOC state-of-the-art multimodal embed. So if you're interested in that conversation, that will be posted on, the newsletter, the podcast, and actually on YouTube live as well, if you're interested in that conversation. With that said, with an hour in, we want to tell you folks, thank you for joining. And, we usually do, multiple other categories. So we also need to talk about big companies. We have vision and video coming up. We may actually talk with some of the researcher from the meta paper. Actually, I'm trying to bring 'em on. Until that happens though, in the middle of ThursdAI usually, talk about this week's buzz. This week is the category where I actually cover, and stuff. Usually in this week's buzz, I let my co-host go and have a smoke break or whatever. Not this week's buzz is now attached with MCP stuff. And ya, I know you've been getting into MCP, getting excited and I know that our listeners as well, this is like a big thing. So I will absolutely, love for you to just like tune in and listen to this. for folks who are listening in, this week's buzz is something that I specifically spearhead at WIS and Bias. I would love for you to hear what that's about. So let's get it started. Let's do what this week does.

All right folks, welcome to this week's buzz, a category of ThursdAI where I talk about the things that happen, weights and biases. And this week the things that are happening are things that I kind push forward for because if you guys remember, we all got Mc peeled, March 6th on the show. We had Jason, NEN and we had Dina Koslow from CloudFlare. And we basically just said, Hey folks, what does MCP tell us about us? And get excited. And I basically wanted to translate this excitement to you and also tell you about why I think now, that CP is a very big deal. After getting mc peeled and realizing how potentially big this could be. And then we saw multiple things happen. We saw open the eye, announced support for MCP in the agent SDK and upcoming support in the chair, GPT app and the completions, API and the other API that I keep forgetting the new one that they have. we also saws, PPI talk about MCP and Sundar just posted to MCP or not, et cetera, something like this. We see an incredible adoption for this protocol. And basically we saw from Swix, our friend of the pod from latent space, a post on latent space, outlining why MCP won. And after open the AI jumping on and supporting this and like fully supporting this, by the way, in their eval dashboard, like everywhere MCP supported, they're going like full on into this. And then also Sundar is hinting support usually when a, when the CEO asks, should we support something, the answer, they already been working on this for a while, like Sam Alman with the open source. So with all that said, that excitement translates into action. like this thing is real. This thing is here. for folks who have no idea what I'm talking about, MCP stands for Model context Protocol, which is a. Standardized way of letting your agents or your chat apps get access to tools. And by standardized, I mean you have to write the standardization once and all the tools that support the standard will be supported. Think about this as USBC, that's the anthropic way of saying this. the metaphor USBC for apps and tools, basically every tool that's built with MCP will be supported by every client that supports MCP. That's the concept. I like to think about MCP as HTTP for agents, because for folks who, who are developers, who, which is the audience that I want to listen in right now. you could technically have built the web on top of TCP sockets, but not until H TT P came and browsers came along with the text transfer protocol. Did we get the web? So I think that, this is how MCP is gonna be because I think it's that big and agents are gonna be everywhere with that kind of caveat. And like with that announcement, the thing that I've noticed super quick is that we, at weights and biases build, weave is a AI, LLM Observability and Evaluation toolkit. As I've talked to you multiple times before, observability is the bigger part, of that as it concerns to MCP specifically because, there's this thing with MCP where right now. Before you use MCP, if you want to add tools to your code, to your agent, to your chat bot, et cetera, you just write them in whatever language you are and you execute them together with your agent, which means you can observe them. You can observe what happens. The more folks will start externalizing their tool use, the more agents will start being using MCP, the less observable this whole thing will become, because a lot of the execution will now happen somewhere else in an MCP server, for example. This is not great. CPS are great, the protocol is great, centralization is awesome. But the more CP is getting adopted, the bigger of a black boxes collection of execution you will have as a developer. This is not great. This is the part that I'm talking about. We're not great and not only not great for us as an observability platform because we will be able to just give you less of a product that is useful to you. It's not great for you as a developer because less and less you'll have the chance to debug something that you would be able to debug had you not used cps, for example. And for that reason, I started thinking about, okay, this is a problem that can be solved because MCP is an open protocol. MCP is hosted by Tropic and the folks are on tropic are supporting this, but we saw OpenAI support this with whether had it not been an open protocol. Google is about like thinking about joining as well. And so I was like, okay. We should talk about MCP observability and how this should look like and what this means and et cetera. And for that, I wanna tell you that I'm starting a new initiative. I called it Observable Tools. This is an actual domain observable tools. If you go there right now, there's not gonna be nothing super interesting because we haven't announced this yet. you guys on Thursday are getting, a little bit of a sneak peek. So right now, observable tools, all it is a redirect. And this redirect is going directly to, the model context protocol GitHub, where a discussion about how this should work, a technical discussion about how this should work, exists, If you go there, it will redirect you to the GitHub discussion of MCP. the proposal that I wrote together with folks at Points and Biases of how we think MCP should adopt observability. For many of you, this makes, no sense. Many folks are listening, not developers, maybe developers, they don't care about MCP. Maybe they care about MCP, don't care about observability, that's fine. I just wanna tell you that for me and for us, it's very important that this exists. And for those of you for whom this is interesting, I would love for you to join the conversation on GitHub and to show by support, by giving a thumbs up, a star Show the tropic team that this is something that you care about because, this is how open protocols work. People suggested ideas all the time and they say, Hey, we need this. We think this should be that. We went a little bit further than this and actually suggested actual implementation with diagrams, with how we think the protocol should change with traces. I'm not gonna go super deep into this, but I'm gonna highlight a few points of how we think this should happen. Specifically, there's another protocol that's called Open Telemetry. Open Telemetry is a protocol that's been, created with, multiple folks also collaborating so that. A company that does observability, for example, like us and some other company, once they support Open Telemetry, a user can just switch to where they send the telemetry, the traces, the spans, the kind of the details about how the application performs, the observability details. They can just switch where they send them. Basically just opening up this field, many big companies, Datadog all these companies, like Open Telemetry is supported by multiple people right now, because nobody wants to get sherlocked into one company solution and then get really stuck with them. So this is why Open Telemetry makes sense. Open Telemetry is also an open protocol, which means it's also a vendor neutral protocol. And so basically we saw that these two protocols have a beautiful, conceptual connection together, vendor neutral. Supports LLLM, specifically the very helpful for user experience and developer experience. And so we suggest an open telemetry support for MCP. This is basically what this, proposal is all about. with that, I already see there's like participants in here. I would encourage you to go to observable tools. we'll post the manifesto next week. Look out for that. I don't wanna show you a sneak peek, but I really do wanna show you a sneak peek. We work on something super cool for the manifesto and, I, this will go live next Monday, so I'm not gonna spoil the surprise, but please follow for, for this to to join. I encourage you to join the discussion. I encourage you to also participate, give thumbs up, and generally help us let Tropic know. And generally folks who work on MCP know that this is something that we think is needed. This is something that we are very excited about and we think that the developer experience will improve. There's also a thing about MCP. Security recently. Yeah, I don't know if you saw this, but there's like a one, a few posts about if you install Mc P tools, you're basically giving them access to your computer. If you're running them like locally. if you don't know what they're doing and they could do a bunch of stuff on your computer, they, without you even knowing about leaking your secrets, like there's a bunch of stuff that like, in every new field is like scary. So we believe that observability is also good for security part as well. I'm very excited to, I'm leading this effort within weights and biases. I really encourage you as a community to just like join and participate and give comments and also discuss this and let us know, let on topic know how you think this should be, happening. We haven't implemented anything. So this is like open, discussion. Yeah. so this is, this week's buzz. Another thing that I want to just briefly mention that, for some very interesting reason, also MCP related reason I, I asked the community, I asked like a bunch of folks. About, how to, like I just said, I'm a simple man. All I want in life is an MCP client that just like uses, all the possible ais and then, lets me add a bunch of tools and that post of my went like stupid viral because apparently I'm not the only one who wants it. and so I just wanna show you that the amount of folks that came in there and talked about this is just incredible. I haven't seen like that excitement about anything. Ya, you probably saw this as well. I got so many pings from people, Hey, if you're not gonna build this, whatever, I'm gonna build this. So I just wanna show you that yeah, almost 1,400 likes. and like bookmarks. So people are bookmarking this more than they're liking this because there's a bunch of solutions. Like 200 folks chime in there, like with solutions, with ideas. I just wanna tell you that I haven't yet seen a very good solution for this, but I will keep you up to date. I got dms from multiple folks saying that they're working on something like this. So once we get something like this, something like an MCP local client that you can chat with local models, but also online models that supports MCP tools or toolboxes, which is a concept, I don't wanna use cloud desktop as what I said. I will let you know about some of these solutions once they come. all right folks. I think that this was this week's buzz and let's move on to big companies.

So we are back from this week's buzz, and now we need to talk about big companies and APIs and different things that they, have announced. And we will start with OpenAI, as well, because OpenAI just had this great week. OpenAI announced a new funding round that we should talk about. oh. just as this is a tiny funding ground from SoftBank, $40 billion. Additional funding for OpenAI. That's a B, not an m. We're used to funding rounds like this for, for millions of dollars. And now this is like a $40 billion validating OpenAI at $300 billion valuation. Which is insane because that's bigger than Coca-Cola and that's bigger than McDonald's. It's bigger than like multiple, like decade old companies. And this just kinda, for us, that's not that crazy, right? we gonna see it. If a GI is coming, this is gonna take over a SI is coming after that's gonna like absolutely take over. They're building the frontier. Like to, to me it's not that crazy. What do you guys think about this? valuation run and like investment, additional investment?

Yeah, I think it's pretty impressive and I think surprising to at least

some people of the rate at which they're like starting these new raises. I think it was just like, less than six months ago that they announced like a $5 billion raise or something around there, right?

Yep.

what you mean $500 billion for Stargate?

No, even before that, like just a couple months before that.

like specifically for open ai,

so they raise, they keep raising these funds.

This one is from SoftBank. and I think that the excitement about the last updates with the image generation, is highlighting how much interest is there. I've seen a few, diagrams of how many people search for open AI and the growth stagnated. And then now it jumped significantly up as well. Probably because of the image generation. They are growing like crazy in India as well enough so that some ultimate posts images of him playing cricket. I don't know if you guys saw this.

Yeah, I just put the link to, I found it.

So it was in October that OpenAI last raised 6.6 billion and then now they have this, 40 billion round and Stargate to be clear, the Stargate initiative, it seems to be set up as its own company technically. And I think they stated OpenAI has equity and Stargate, but it's technically not funding of OpenAI, even though it is basically open AI's compute.

Yeah.

so in this announcement they mentioned that 500 million people, Useche, GPT, every week, 500 million people. I remember last week we talked about million people that keeps growing. Sam Altman mentioned, something crazy because of the image generation. He said, when CG PT launched, they added 1 million people in five days. The first five days, they just added, onboarded, new 1 million people in an hour because of this image generation, which is like just absolutely insane. And I love the fact that like when we started, I. I'll just remind this like for a second. When we started changing, Thursday ai two years ago, I, my name myself, AI evangelist because like you needed to evangelize ai. Now we're getting to a point where one eighth of the world is about to use ai, if not already, like every week, which is absolutely insane. So shout out to AI and the launch. I wanted to highlight this as well. I thought it was funny. It's not super crazy important, but this is the whole blog. Two paragraphs. They raised $40 billion. The whole blog that talks about this is two paragraphs with basically no details when they launch a research blog is multiple pages long with a paper. I just found it funny that the number of words on the blog is proportionately opposite to how much money was raised. because $40 billion is absolutely insane. We saw another valuation of an AI company, by the way, this week we should mention AI has purchased x. Which I don't know what it means, but it, because it's same daddy for both companies, in all stock. But basically that, transaction valued X at some like crazy amount, and also valued at XAI. So XAI is now, I think around, do you remember, I don't remember exactly. I haven't written it down, but like X AI was also evaluated very highly, which is very interesting because, I think I saw a post that means what does this mean for tropic and why IST tropic not as validated as these two like other companies. so very interesting. We're getting like very high valuations for AI companies, AI foundation companies.

Yeah.

of the information here pulled up. it was, in December XAI raised 6 billion.

The raise 6 billion and now post the, the acquisition of XXAI as

valued at, I think at 80 or something. I don't remember exactly. yeah. Alright, folks, let's move on to other news that I wanna talk to you about. we have Amazon NOVA Act. Do you guys see this? This is crazy. Amazon announced NOVA Act, which is their computer use agent, only via API and they're flexing on top of Amazon. Amazon is flexing on open the eye on traffic with this chart that they say that on multiple things like screen spot web, text eval, and web icon eval, and then also ground UI web. They probably cherry picked like these three because we know there's like a web agent evals as well, but on these ones they're beating both cloth 3.5 sonnet and open the eyes QA model, which is computer use agent model, which we know runs operator and we've talked about that being run on the open, state of the art. they're beating both the foundation labs with this model, with this NOVA Act model. We talked about NOVA as well. NOVA is their kind of series of foundational models for Amazon. NOVA Act specifically is, computer use actor. And so very interesting. there's an SDK that you are able to sign up to get. I try to sign up and I get seven different. Captures in the middle. very Amazon, but fine. so it's not like a open model, but I found it very interesting that Amazon is like playing in this field. And then somebody told me like, Hey, Amazon bought Acqui, hired Adept. You guys remember Adep, the company that was started by some of the co writers of the Transformers paper and then recently went through some terminals, didn't release a thing, promised us like a web page and didn't release it, blah, blah, blah, blah. apparently Amazon Echo hired, so this could be the result of the dev folks. so I don't think, I haven't used it. I haven't, seen anything. I don't know if there's any more commentary, but folks in the audience, if you used, or try this one, please. please let us know how this works. I wanna highlight a few more things super quick before we move on to the vision stuff. And we actually have maybe a guest, to, to join us afterwards on the vision stuff. Gemini 2.5, let's talk about how good this is for a second. General 2.5 was released two weeks ago. Doshi on the show from the DeepMind team. and she's the, sorry, the Google DeepMind and she's leading the product for Gemini 2.5. Gemini 2.5 is incredible. Folks. I kinda I'm very into this model. I've loved Gemini models. Generally. Gemini 2.5 released is a thinking model as a unified thinking model. and it became my number one for multiple things. And I can tell you the things that I've used Gemini 2.5 for. I recently am very happy to announce that I finished filing my personal taxes. I still have corporate taxes to do for, but my personal taxes, which is a, like for folks who are listening and not in the States and like the com, the normal government takes care of your taxes for you. And you don't have to stress about this. You don't understand how stressful that period is. And you can hire CPA, but it's still stressful. You ask you a bunch of questions like, oh shit, if I answer incorrectly, I'm gonna go to jail. There's like a whole thing. And, I've used for my taxes this time, and I did them myself this time, like a dumb ass. I used both open the eyes, oh, one pro. supposedly the leading math, supporting model, and this time I use Gemini. And Gemini is amazing because the context window is insane. And I don't know what it did in this model. I posted this and immediately deleted this video because I didn't want to expose my personal text to the world. But if you post your 10 40, which is like the standard four for taxes in the US into Gemini and say, Hey, review this, and then you open the thinking process of that. I imagine that this is how a CPA's mind works because literally you can see the model just thinks about every little paragraph, every little box in the tax code. it, it is crazy. It tries to compare. There's like a bunch of math that you do. There's nevermind. I don't wanna talk about taxes. It's horrible. Sorry for bringing this up. But the fact is just seeing how this model thinks about taxes and Google does expose the thinking traces, unlike open the eye, let's open them up folks. Everybody exposed thinking traces, come on. There's no, no mother. Google gives you the actual thinking traces of Gemini 2.5 and they're incredible. I sometimes ask a very difficult question and I go to just the thinking traces just to see what it thinks. I don't even care about the answer, And so an additional highlight of how awesome this model is and you can, you guys can call us Gemini s Shells. You can call us open eye shills. We are here to promote AI in a positive way. When it's good. We're gonna tell you when it's bad. Like the refrigerator stuff from lg. We're gonna also tell you, this is good folks. And to highlight how good this is on math, I wanted to show this thing where, meth Olympiad problems were released late March. So they are pretty held out. So there's not a lot of chances that releasing late March, that model was already trained by this. We know the model trained it, it takes a while. they're held out. They're so difficult that any score better than zero is better than what 99% of the people is capable of. That's how one person on X described them. Okay? Any score better than zero is better than 99.9 Percent of the people who read like general Twitter are capable So I wanna show this chart, and for folks who are watching my screen, this chart is so high that I need to scroll through the chart to actually see the whole chart. this is an accuracy, and date on the x excess are dates from when this kind of U-S-A-M-O, was released. And then, models that are getting some percentage, most models on this very difficult benchmark. they get barely up to 5%. I think deep CR one is getting the closest to, four point 76%, on this USAM mo. And then somebody measured, Gemini 2.5 Pro. Folks and the models are here are like, oh, one pro on high O one Pro with high thinking. There's a lot of, sequential tokens in there. There a lot of like test time, compute, all three meaning with high, a lot of sequential test time, compute Cloud 3.7 with Think enabled, Gemini 2.4, 2.0 flash, though previous thinking, deep Seek R one. All of them get less than 5% on this benchmark. And here comes Gemini 2.5 Pro and we can see how much Google cooked with this and we can see why, bringing back no for $2.5 billion, whatever was worth it. Noam Shair, the lead author on Transformers, Gini 2.5, gets 24.4% on this super, super hard math benchmark. just just the graph is crazy. I just wanna highlight this. if you still think that AI cannot do math, just look at this chart, look at the same old thing. all right, so Gemini it generally great use it. oh. And by the way, it's free now in the Gemini app as well, so you can always use it for free in the AI Studio, Gemini app itself, there's an app, there's a mobile app as well. Gemini 2.5, is now also on that app as well. Last thing that I wanna chat about is this thing, super quick and big company's APIs without too much fanfare, without mentioning any type of other news that I'm not allowed to mention. COR Wave and Nvidia announced 800 tokens per second for LAMA 4 0 5 B. On the new ML PERF 5.0, which is like an evaluation for GPU runs, with the new Blackwell Super Chipps GB 200. There's a mess of like acronyms here. and I, I'm waiting for this to help me like wade through them. But basically, core Weave the hyperscaler for ai, and that's backed by media like a bunch of other folks, announced that 800 tokens per second for LAMA 4 0 5 B, the bigger one, the chunky, it's not an MOE, it's a, it is a dense model. and then, there's new results for H 200 GPU achieving 33,000, not 3300, 30 3000 tokens per second on LAMA two 70 B, which is a 40% improvement on top of H 100. Now, some of it is, is due to core waves, getting the chips early. Some of this is due to the core waves, infrastructure as well, and some optimizations. I just wanted to highlight this insane number, 33,000 tokens per second. Like what? Hell, and I'm very excited to see how, how LAMA four, when it comes out will perform as well. so I wanted to highlight these like insane speed ups for the new generation of, Nvidia chips. Ya, how many comments on this?

I feel the A GII.

Seriously, there's nothing to, it's just, I wanna use it pretty much.

All right.

And you, if you do feel the a GI, I wanna show you, the open ai, gave us a way to feel the emo ai. But basically there's a black circle now in advanced voice mode in Open the Eye, and that's called Monday. And when you talk to Monday answers back and it's emo and it just like hates answering to you. And it's just so fun. I don't know if folks use this. I really wanna try, to talk to Monday Monday. How are you?

Alright, fine.

Here we go. Welcome to Thursday ai. I'm your Oso enthusiastic host, Alex here to bring you the latest in AI today. We've got some thrilling or not news for you from open AI's announcements to whatever else is on the agenda. So grab your coffee or don't, and let's get this over with. Ugh.

For some reason, this is so fun.

For some reason, having an AI answers like an email and inside folks have leaked System message for it. it is just so fun to try and cheer this model up when it's just gloomy. There's no crazy stuff. It's just really fun. So shout out to open the eye for this, supposedly open, April 1st full trick. But essentially voice replaced Santa 'cause remember Santa This Way. I think we've covered the big companies. LMS and API in a good way, folks. good job to us. And then I think we're moving on to vision and video and I would love to invite, Kong to come up to the stage with us as well. Kong that was, working with Meta on Mocha, once Kong joined. We're gonna talk about this. Oh, Kong is right here. I also want to highlight the two things that happened in the world of vision and video. So for folks who are not watching, I promised this, to let you know when now is when. So we'll start with, gen four and I'll show this on my screen. So if you are listening only and you want to look, now is the good time to look. let's start with runway. Runway announced gen four, and some of the videos are just like, just keep getting like incredibly better. Flapping of wings and birds, like people are building, like all kinds of things. I honestly, folks, I'm like running out of ways to get excited about video models because they just keep getting better and better. In this case, this is like animated stuff, that the runway shows. It's just absolutely incredible. So we're seeing like all kinds of cute felt animals and I think they're highlighting the consistency there. just absolutely great folks who are like using runway. I'm like running out of ways to get excited about video models because they just keep being insane, to the level of open AI that just released the native image generation. We're getting there with video models, very soon. But the consistency in runway is, I think the highlight. I think people are, used to judging whether or not something is an AI video by. By these three things that I at least judge. The AI video ones. One, it's all shortcuts. Very shortcuts. These models usually generate like very short form, maybe five to 10 seconds. And so the people cut two. Consistency between frames is inconsistent. So the same character doesn't look the same in like the next frame, et cetera. They look a little bit different. That's how we judge AI videos right now. Runway release, something that looks incredible inconsistency. They're rebuilding worlds. so shout out to runway, for this. Now the next thing that I wanna chat about is omni human. Do you guys remember Omni Human? from Ance? We talked about omni human, and we get super excited. Omni Human is taking one picture So you basically take one image and Omni Human from Bidens has trained probably on all of TikTok, and it just generates incredible like lifelike portraits. So Omni Human didn't release, they announced like a paper, et cetera. They didn't release. But now Omni Human is available. I wanted to tell you guys, but I'm like super excited. there's a website called dreama dot cap co.com. So this is like their, AI playground and omni human lip-syncing is absolutely available here now. And so you go to Lip-Sync. And here's an example that I wanna show you of me from last week. You guys remember we got excited about Image Generation from last week? Was that last week? Jesus. It was last week, right? It's crazy. Image generation came out a week ago. Gemini 2.5 came out a week ago. Like it feels like a whole month has passed. here's me, animated and lip synced from one image. This is not video, this is one image. so I don't know if you guys, if you folks can hear this, let me see if I can actually show you the thing with the sound, because I think it's very important to see where the sound, window, no, sorry for sound. I need this. Yeah. Okay. Mina. And let's see if you guys can see this. Yeah, you can see this. So here's a, oh, no, this is a, this is the thing that I don't wanna show you yet. Okay. going back. this is further for next week. okay. So I'll share it to the, in the show notes, but also share it on x, the, no, actually share it to next Hold on. Technical issues, folks, but it's fine. We will get over them. here, let's go here. And then let's find, I wanna just play this for you so you'll hear, on train. So omni human animates from one image. And here's an example of it that I posted on my X somewhere here. Yes. We'll say it was an insane week, and we have a few guests coming to talk about the how insane. Let's jump into the DLDR because time is of the essence today. every Thursday, but specifically this week. So you can see the lips move, you can see the, everything. they're not actually trained on, cartoon specifically. here's Tom, like robots that build like with an actual hero. Hold second, look at the lips moment. It's crazy.

I was just very excited to let you guys know that this is now available and you can try it. It is not an open source model, but you can do animated characters. You can do like a bunch of like very crazy stuff. in addition to this, another effort came out this week that I wanna talk to, to the author of. So I wanna welcome to the Stage Kong Way. Hey Kong, how are you? Hey. Hey, Alex. Thank you for hopping up. very, impromptu. You're, listening to the show. You actually tagged me in some of the releases as well. you guys announced Mocha this week, right? Yes. could you, first of all, welcome to the show. This is your first time, so feel free to introduce yourself, who you are. and then afterwards, let's talk about Mocha. I'm like very excited to hear from you.

Thank you.

Thank you so much for that kind of invitation. I really appreciate it. Hi everyone. I'm from the University of Waterloo and yeah, the Mo the Mocha project, I don't know how to pronounce it. It's either moca or Mocha

let's say moca.

Yeah,

mocha.

Okay. Was done like during my internship at, meta AI and in collaboration with my supervisor professor. Mocha is like the first model that can generate talking characters animation directly from speech and text. talking characters. So we define it as digital characters that don't just speak, they show emotions and natural body languages and act realistic really, in video, just like human actors movie. And I think the most interesting part about Mocha is generating characters directly from speech and So it doesn't rely on any accelerated signals such as traditional talking head models that only animate human faces.

I wanna highlight this point for just one second.

What I just showed with Omni Human and with hetero labs, they require an image and some audio. Let's say this is not that you don't require an image, right? Yeah. You generate these characters from where does these characters come from?

No, from the noise space, right?

From the laing space.

You basically just bring them up to life, which is, I found incredible. let's watch. I think we now can have a sound. Let's watch like some of the example of the launch trail that you guys have.

I said it before and I'll say it again.

Life moves pretty fast.

Are you behind on your credit card Bills?

Good. Pick up the phone and start dialing. Boo. Tired of 'em being gone.

For folks who are listening, some of these characters, like this

character we're watching right now, holds a cup of coffee and there's like a bunch of things happening. What I presume is New York in the background, people are walking, it looks realistic. This character that we're watching now is a woman in a sweater. She holds a cup of coffee with two hands and you can see the cup of coffee. First of all, Kudos on the hands, man. What the hell? The hands look realistic. Like all the fingers are here. All the right amount of fingers. She holds the cup with the two fingers. but also like she, she talks and then she drinks from the cup. What is going on here? So help us understand like what you guys cooked here, because this looks to me like crazy.

I don't know.

It's just I would try this prompt and just saying yeah, this woman is drinking while talking. And then we send the model speech. She's going to talk and now we found this, the model generated this

directly.

It's so good. It really, some of these that I walked through, like this specific guy, they look so realistic. It's unbelievable. this looks like a scene from a movie.

Life moves pretty fast.

You don't stop

folks, who are listening.

This guy is holding a cigar and the life moves, the smoke trail follows the ci perfectly pretty fast. this looks, you don't, I don't think that if I showed this to my mom today, my mom will be able to even understand that this is not real life. This is not was taken of a person. This is crazy. what was hard about this? tell us about some of the work. Tell us about like how did you get this done and what is the kind of the implications of some of this work? implications. Yeah. Tell us a little bit about challenges that you found. how was this trained as much as you can? I know that meta is, not always open fully, but like you guys also published a paper. so give us some highlights for folks who are following. Yeah. how is this good?

Yeah, I can walk you guys through the promotional tweet basically

summarize the architectures and stuff.

yeah, absolutely.

Oh, can I share my screen?

Yeah, you

Sure.

Okay. One sec. I'll just give you a heads up. we love deep dives, but we're, we don't have like too much time for a super deep dive, but we would love to hear from you kind of some insights and some incredible things from training this model. And folks will absolutely check it out. we'll add this to the show notes as well, because I, yeah. Can you submit a little bit, please?

Yeah.

Basically just for, basically, yeah. for tweets, could you just, zoom in, please?

this?

No, just go back into the hit control. Yeah. Like this is great.

Basically this is the war architecture.

It's we project the speech and the text and also the video to tokens. And then the whole thing is like the DIT architecture, DIT is,

diffusion

transformers.

Diffusion transformers. Yes. And then we use self attention for the video tokens. And cause attention for the text tokens and also our window cross attention. Basically, I'm gonna describe it later in the next one. And the model directly generating the videos. And I think one thing to notice that the captions, the text prompt is pretty long. It basically highlights the camera shots. And then

for folks who are listening, we're watching like a text prompt.

And the text prompt is a closeup shot of a woman standing in the modern house speaking to the camera while facing straight ahead. She has shoulder length, dark hair, the background feature, sleek modern furniture. She holds a single rose in her hands, gently running her fingers along its pedals. The static camera captures her face and upper body. So this, just describe the scene. There's no like text here, text comes from somewhere else, right? It comes from the sound.

yes.

Come it is independent from the sound. Yeah. yeah. But it's pretty detailed. It captures the actions and the appearance. And the next one is about the cross window attention. basically we design this architecture to align the audio tokens with the, video tokens. So basically we found that like the speech associated with like short term, phenoms, right? And then. the text associated with the Intel video. So basically we introduced a strong inductive bias here. Like for, let's say for one, second video token. It only associated to the one second audio token. At that time, it should not look at other audio tokens. Yeah. that's a key here to

when you say the speech and the text, you mean the prompt

that describes the scene text? Yeah. The prompt should describe the entire scene. And then the speech is like literally what its about to say. And then the lips.

Yes.

exactly. that's a key here for the moca to generating very accurate lipsync characters.

I wanna highlight the lip sync is incredible.

I've been following field for a while, and I've looked at wav to vec and the lip sync and deep, lat sync from Biden's recently released. very good lip sync as well. you guys are spot on. Perfect. I haven't seen any issues with lip sync that I could detect, which is very hard. I'm assuming this is because of data, because of something else. Could you speak about the lip sync stuff specifically? what,

data could helps a lot.

these architectures, yeah. Helps, a lot and also. I think the data we use is higher quality. We do a lot of, filtering And also, we have training, starting from data with facial regions. so the model can learn the lip sync quickly instead of training all the data, labeled with audios together. most of the data, not close up shot. It's basically a medium shot. A medium shot, and the feature region is too small in the frame it harder for the model to learn. starting from just model, a square image of the face

Oh, so you're giving just like the square video of the face talking.

yeah, exactly. Starting training from that. Yes. I wanna show more of this if you allow me, I wanna show folks, because this to me is like absolutely incredible. I just, let me find a video. Yeah. You wanna continue with one more, one more kinda, yeah. Insight.

There's this one's talking about oh yeah, we forgot to mention that.

Moan generating like multi character interactively conversation

yeah.

Sorry about this. it's not only one character, this is like multi character as well, which is like absolutely crazy.

Yeah.

Basically the speech signal can. Implicitly tell the model to Switch the shot. switch the clip and the prompt. In the prompt that we design like characters. Basically we're using some tagging mechanism, like person one, person two, so basically like person one, associate for person with certain appearance, and then also person two. Then we described the first clip and second clip. In this standard format. The reason we're doing this is because I'll show an example.

go ahead.

Continue. But I wanna show an example because the example is like absolutely crazy. the multi character speak. let's say which one is your favorite? I like this one. Yeah, this one's it's just the people who no one imagines anything of who do the thing. Oh, there's only one. Yeah, there's only one. It's a bottom. Oh, it's multi character. Oh, BU's dialogue. okay, this one. Let's see.

You don't even know me.

I have the rest of my life to find out.

This is crazy.

So you took, I'm assuming, sounds from movies like these sounds, the like segments.

for the, demonstration purpose, Bruce, not using them.

Yeah, Absolutely.

But I'm saying like, these are like very high produced, like quality.

ent,

image.

Bruce Wayne, Rachel's told me everything about you.

So we just heard like a little segment of the dark night

probably, but folks are sitting in the spaceship, which is like insane. And they're talking the both. Turn base conversations. con I wanna ask you about like, why is it only five seconds? What's going on here?

the base model is only support five seconds.

And as an intern, I don't have too much compute switching huge models, supporting like 10 seconds or something.

I see, so is this just a function of compute and

generally models can generally for the method, should works for

10 seconds or whatever lens?

Yeah.

Wow, okay. what about consistency from like, when I wanna generate multiple things for a longer video? Is that possible

think it's possible.

'cause like it works for the two person talk, right? Yeah. I think there's like another demo, like I think in the website. I don't know

Let's show this and then we'll move on to we

have one last thing to cover. but show me which one you, it's

the third one in the first row of the

This one?

Yeah.

This one is actually, you can see the characters shown like these.

follow my own plane.

The character show on the right.

The next in the second place.

Oh, the first character is shown from the back with the same coat.

On the same hair you mean? yeah. Absolutely. This one actually is the one that I okay. I see an AI picture here. Like I see her face is moving weird. Some of them are like so realistic that I cannot like. cannot discern. but this one, I could like discern a little bit. This was like an ai. anyway, con thank you so much for coming up and talking to us about your, your effort here. Absolutely ridiculous. Absolutely ridiculous. That I can now fool my mom. last maybe question here is obviously this a research preview. Any, I don't any way we can test this, try this, at some point in the future. are there any plans for productionizing this? what, folks that are watching this will, will they be able to do generations like this going forward?

I think, yeah, that's, gonna require the improvement from meta,

but personally I can try to implement on some like open source models.

Yeah.

If it's possible in open source and you need support in compute and we have folks who are listening who have compute to support, follow Kong to, I think you said, right? Yeah. To, yeah. Tong. and follow to, and then, please, also, let's have this in open source as well. This is crazy folks. This is crazy. I encourage you to go to, to, to mocha and check out some of the videos. Some of them just are absolutely ridiculous song. Thank you so much for coming up. folks, we need to cover a few more things in the audio and You're good. Okay. so folks, so Mocha, definitely check it out. The next thing, and I think the last thing that we're gonna cover folks for today, unless I miss some multiple things, is, I wanna show you or have you listened to the two things. Okay. So Mojo is unreleased, omni Human that I told you about, is now possible in, dream Meina website and you can try this. But also now with I image generation from JGPT, I saw Olivia Moore from a 16 Z Talk about hydro labs, and I wanna show you hedge as well. Hedge is also doing Hydro Labs. One second please. Hedge labs. Yeah, like this. What am I doing? What am I doing? I wanna show you, I wanna make it, I want to make it so you can hear. Chrome only allows here to, sorry, one second please. So Hedge Labs also does, animation for characters. And, in order to show you what this means, I need to maybe go here. Okay. We'll skip Head Labs. I wanna show you now. hi Helio. Ilio audio number two. One second, please. we've talked about, minimax, the model and the video models from Hi Lio ai. and they have released an update to their voice models and they're just absolutely incredible. I will try to play something for you, because just absolutely remarkable, but for some reason the technical reasons, don't let me do it. we'll generate something. Okay. Here's an example. this one Lifelike speech. Okay? this is an example. they added authentic multi-language coverage, 99 vocal similarity is what they claim. Zero glitch and rhythm and studio, great quality. and then they also have like history management. So let's take a look. They have, and we've talked about these models before and they general like multiple things. but number like speech two HD is just like absolutely crazy. you can say, Hey, folks. Welcome to Thursday, ai. Let's see if this works. Come on. Generate, of course. Lemme log in. one second. I'll log in and I'll tell you guys. Yeah, now it works. Alright.

Hey folks.

Welcome to Thursday ai. First of all, right off the

bat, this is one of the first text generations that said

ThursdAI correctly, without me having to spell it differently, right? But the thing is that they have here, they have voice modifier and they have emotions. You can have 'em say this with a happy thing. thank you for tuning in every week.

now made them happy.

Hey folks, welcome to Thursday. I thank you for tuning in.

Every week.

I'm like, they have a voice emotion thread thing and you can try like happy, you can try. sad. Let's try Surprise with the same voice.

Hey folks, welcome to Thursday.

I thank you for tuning in.

Anyway.

They have emotion control. The thing that they have though that I love is voice cloning. That is crazy. Good. it takes a while to do and, I'll just add it to the show notes so you guys will see, but basically I clone some voices with this. I've never, like 11 labs or nobody comes very close to this. So I just wanted to highlight that this release was there, all of the tests that we just talked about together come to this point where like everybody can create multiple things right now. You can go to chat, g pt, native image generation, generate whatever you want with any resemblance that you want. Then you can take this into multiple video models and now character models as well. So like Hera and Omni Human from Dreama. You can lip sync those. You can generate any voice you want with any person clone you want 99% similarity to a person. And then you can take all these and generate whatever you want basically. And so I think this is a great place to stop our show at exactly two hours by saying that the creative space of what folks can do now with AI is incredible and hopefully some of the tools that I mentioned to you are useful to you to go and generate and do whatever you want. Start with image generation, put it in Hera, and then generate the voice Minimax, which is absolutely to me, state of the art voice generation folks with this. We're at exactly two hours for Thursday. I wanna thank, my friends here, LDJ, and Yam for joining me as co-hosts. I wanna thank multiple guests that we had. You guys know that one of the funnest things that we get to do is talk to the researchers and developers of the things that they released this week. And this week we talk with song just now about Mocha, which is an effort from Meta to generate, movie style quality characters that talk with lip sync, chat with, Shingo Wang from All Hands. about all hands releasing their own model that they've Fine tune on top of Quin 2.5 B, 32 B, and that performs well and also open hands. Their open competitor to Devon now is very close to the state of the art of agent generation coding generation models. We also chatted with Zach Nusbaum from NOIC about no multimodal embeddings as well. So always great to have, expert folks that did the work to come and share the work on Thursday.

have you seen that in two?

Yes.

Breaking news, I dunno if it's is called Worth Calling Back News. Okay. One, one for fun, one for fun. AI breaking news coming at you only on Thursday. Ai.

All right, super quick folks. Thanks to our community. Thanks, yam. Devin was announced Devon two. Devon is the AI coding editor that, folks, got super excited about and then open. Devon tried to replicate, they call this a new agent, native ID experience, are reacting, oh, breaking news. It generally available for starting at $20. Devon previously cost $500 a month, so this is now $20 for Devon. new experience let's you spin parallel Devons, to take multiple tech at once. Devons is, familiar via code environment. What is this interesting there? collaborative planning partner. Start with their ideas instead of detailed r and d. PRDs. Devon search actually 20 bucks a month. I'll try it. I haven't signed up to Dev for 500. It didn't do the work as we know from our friends from, an ai, but 20 bucks a month, it's actually breaking news. last but not least, each. A CU now goes almost two Xs four. We measure a CU efficiency on internal benchmark of junior dev tasks. so folks, here you have it, breaking news. Thanks Yam. at the end of the show. Agent Native, IDE from Devon. and they cut the price from Teams Price for 500 bucks a month. Two pay as you go, but 20 bucks minimum. So minimum it could go over. Okay? So it's just wa watch your billing folks Yam. Thank you. folks with this I think is a great point to stop the show as, always. If you missed any part of our show, please go to Thursday. I news all the links to YouTube to Spotify folks. We're almost at a thousand followers on YouTube. Please give us a follow on YouTube. We're almost there. It's very exciting. YouTube, Spotify. Apple Podcasts, wherever you get your podcasts. If you want any links of the show, the newsletter is gonna get posted with all the links in the show notes. Probably today, a little bit less of a full newsletter, but still, with all the links of everything that we've talked about, basically all my notes converted into a blog post form. if you missed any part of the show and you wanna support this podcast, the best way to support us is to share it with a friend and give us five star reviews on the podcast platforms. It really helps if you wanna support financially. I never talk about this, but like the substack, you can, Substack is free if you don't want to, but if you wanna support, Substack is possible to support financially as well. This helps run the show smoothly. but other than that, folks next week I'm gonna talk about, observable tools. It's also a great way to support as well. Thank you everybody for tuning in from week to week. It's really a great pleasure of ours to host these, incredible guests, to talk about the work they do to host all of you in the community, to give us like updates about everything that happened and keep us up to date as well. We here live on stage are getting up to date from one another as well, and hopefully this translates to you as well. With that, happy Thursday. Thank you so much for joining, and we'll see you here next week. bye bye everyone.