📅 ThursdAI - July 11 - Mixture of Agents & Open Router interviews (no news this week)

Hey everyone, Alex here 👋We’re deep in the post-reality era now. Between Sora2, the latest waves of video models, and “is-that-person-real” cameos, it’s getting genuinely hard to trust what we see. Case in point: I recorded a short clip with (the real) Sam Altman this week and a bunch of friends thought I faked it with Sora-style tooling. Someone even added a fake Sora watermark just to mess with people. Welcome to 2025.This week’s episode and this write-up focus on a few big arcs we’re all living through at once: OpenAI’s Dev Day and the beginning of the agent-app platform inside ChatGPT, a bizarre and exciting split-screen in model scaling where a 7M recursive model from Samsung is suddenly competitive on reasoning puzzles while inclusionAI is shipping a trillion-parameter mixture-of-reasoners, and Grok’s image-to-video now does audio and pushes the line on… taste. We also dove into practical evals for coding agents with Eric Provencher from Repo Prompt, and I’ve got big news from my day job world: W&B + CoreWeave launched Serverless RL, so training and deploying RL agents at scale is now one API call away.Let’s get into it.OpenAI’s 3rd Dev Day - Live Coverage + exclusive interviewsThis is the third Dev Day that I got to attend in person, covering this for ThursdAI (2023, 2024), and this one was the best by far! The production quality of their events rises every year, and this year they’ve opened up the conference to >1500 people, had 3 main launches and a lot of ways to interact with the OpenAI folks! I’ve also gotten an exclusive chance to sit in on a fireside chat with Sam Altman and Greg Brokman (snippets of which I’ve included in the podcast, starting 01:15:00 and I got to ask Sam a few questions after that as well. Event Ambiance and VibesOpenAI folks outdid themselves with this event, the live demos were quite incredible, the location (Fort Mason), Food and just the whole thing was on point. The event concluded with a 1x1 Sam and Jony Ive chat that I hope will be published on YT sometime, because it was very insightful. By far the best reason to go to this event in person is meeting folks and networking, both OpenAI employees, and AI Engineers who use their products. It’s 1 day a year, when OpenAI makes all their employees who attend, Developer Experience folks, as you can and are encouraged to, interact with them, ask questions, give feedback and it’s honestly great! I really enjoy meeting folks at this event and consider this to be a very high signal network, and was honored to have quite a few ThursdAI listeners among the participants and OpenAI folk! If you’re reading this, thank you for your patronage 🫡 Launches and ShipsOpenAI also shipped, and shipped a LOT! Sam was up on Keynote with 3 main pillars, which we’ll break down 1 by 1. ChatGPT Apps, AgentKit (+ agent builder) and Codex/New APIsCodex & New APIsCodex has gotten General Availability, but we’ve been using it all this time so we don’t really care, what we do care about is the new slack integration and the new Codex SDK, which means you can now directly inject Codex agency into your app. This flew a bit over people’s heads, but Romain Huet, VP of DevEx at OpenAI demoed on stage how his mobile app now has a Codex tab, where he can ask Codex to make changes to the app at runtime! It was quite crazy! ChatGPT Apps + AppsSDKThis was maybe the most visual and most surprising release, since they’ve tried to be an appstore before (plugins, customGPTs). But this time, it seems like, based on top of MCP, ChatGPT is going to become a full blown Appstore for 800+ million weekly active ChatGPT users as well. Some of the examples they have showed included Spotify and Zillow, where just by typing in “Spotify” in chatGPT, you will have an interactive app with it’s own UI, right inside of ChatGPT. So you could ask it to create a playlist for you based on your history, or ask Zillow to find homes in an area under a certain $$ amount.The most impressive thing, is that those are only launch partners, everyone can (technically) build a ChatGPT app with their AppsSDK that’s built on top of... the MCP (model context protocol) spec! The main question remains about discoverability, this is where Plugins and CustomGPTs (previous attempts to create apps within ChatGPT have failed), and when I asked him about it, Sam basically said “we’ll iterate and get it right” (starting 01:17:00). So it remains to be seen if folks really need their ChatGPT as yet another Appstore. AgentKit, AgentBuilder and ChatKit2025 is the year of agents, and besides launching quite a few of their own, OpenAI will not let you, build and host smart agents that can use tools, on their platform. Supposedly, with AgentBuilder, building agents is just dragging a few nodes around, prompting and connecting them. They had a great demo on stage where with less than 8 minutes, they’ve build an agent to interact with the DevDay website.It’s also great to see how greatly does OpenAI adapt the MCP spec, as this too, is powered by MCP, as in, any external connection you want to give your agent, must happen with an MCP server. Agents for the masses is maybe not quite there yetIn reality though, things are not so easy. Agents require more than just a nice drag & drop interface, they require knowledge, iteration, constant evaluation (which they’ve also added, kudos!) and eventually, customized agents need code. I spent an hour trying it out yesterday, building an agent to search the ThursdAI archives. The experience was a mixed bag. The AI-native features are incredibly cool. For instance, you can just describe the JSON schema you want as an output, and it generates it for you. The widget builder is also impressive, allowing you to create custom UI components for your agent’s responses.However, I also ran into the harsh realities of agent building. My agent’s web browsing tool failed because Substack seems to be blocking OpenAI’s crawlers, forcing me to fall back on the old-school RAG approach of uploading our entire archive to a vector store. And while the built-in evaluation and tracing tools are a great idea, they were buggy and failed to help me debug the error. It’s a powerful tool, but it also highlights that building a reliable agent is an iterative, often frustrating process that a nice UI alone can’t solve. It’s not just about the infrastructure; it’s about wrestling with a stochastic machine until it behaves.But to get started with something simple, they have definitely pushed the envelope on what is possible without coding. OpenAI also dropped a few key API updates:* GPT-5-Pro is now available via API. It’s incredibly powerful but also incredibly expensive. As Eric mentioned, you’re not going to be running agentic loops with it, but it’s perfect for a high-stakes initial planning step where you need an “expert opinion.”* SORA 2 is also in the API, allowing developers to integrate their state-of-the-art video generation model into their own apps. The API can access the 15-second “Pro” model but doesn’t support the “Cameo” feature for now.* Realtime-mini is a game-changer for voice AI. It’s a new, ultra-fast speech-to-speech model that’s 80% cheaper than the original Realtime API. This massive price drop removes one of the biggest barriers to building truly conversational, low-latency voice agents.My Chat with Sam & Greg - On Power, Responsibility, and EnergyAfter the announcements, I’ve got to sit in a fireside chat with Sam Altman and Greg Brockman and ask some questions. Here’s what stood out:When I asked about the energy requirements for their massive compute plans (remember the $500B Stargate deal?), Sam said they’d have announcements about Helion (his fusion investment) soon but weren’t ready to talk about it. Then someone from Semi Analysis told me most power will come from... generator trucks. 15-megawatt generator trucks that just drive up to data centers. We’re literally going to power AGI with diesel trucks!On responsibility, when I brought up the glazing incident and asked how they deal with being in the lives of 800+ million people weekly, Sam’s response was sobering: “This is not the excitement of ‘oh we’re building something important.’ This is just the stress of the responsibility... The fact that 10% of the world is talking to kind of one brain is a strange thing and there’s a lot of responsibility.”Greg added something profound: “AI is far more surprising than I anticipated... The deep nuance of how these problems contact reality is something that I think no one had anticipated.”This Week’s Buzz: RL X-mas came early with Serverless RL! (X, Blog)Big news from our side of the world! About a month ago, the incredible OpenPipe team joined us at Weights & Biases and CoreWeave. They are absolute wizards when it comes to fine-tuning and Reinforcement Learning (RL), and they wasted no time combining their expertise with CoreWeave’s massive infrastructure.This week, they launched Serverless RL, a managed reinforcement learning service that completely abstracts away the infrastructure nightmare that usually comes with RL. It automatically scales your training and inference compute, integrates with W&B Inference for instant deployment, and simplifies the creation of reward functions and verifiers. RL is what turns a good model into a great model for a specific task, often with surprisingly little data. This new service massively lowers the barrier to entry, and I’m so excited to see what people build with it. We’ll be doing a deeper dive on this soon but please check out the Colab Notebook to get a taste of what AutoRL is like! Open SourceWhile OpenAI was holding its big event, the open-source community was busy dropping bombshells of its own.Samsung’s TRM: Is This 7M Parameter Model... Magic? (X, Blog, arXiv)This was the release that had everyone’s jaws on the floor. A single researcher from the Samsung AI Lab in Montreal released a paper on a Tiny Recursive Model (TRM). Get this: it’s a 7 MILLION parameter model that is outperforming giants like DeepSeek-R1 and Gemini 2.5 Pro on complex reasoning benchmarks like ARC-AGI. I had to read that twice. 7 million, not billion.How is this possible? Instead of relying on brute-force scale, TRM uses a recursive process. It generates a first draft of an answer, then repeatedly critiques and refines its own logic in a hidden “scratchpad” up to 16 times. As Yam pointed out, the paper is incredibly insightful, and it’s a groundbreaking piece of work from a single author, which is almost unheard of these days. Eric made a great point that because it’s so small, it opens the door for hobbyists and solo researchers to experiment with cutting-edge architectures on their home GPUs. This feels like a completely new direction for AI, and it’s incredibly exciting.inclusionAI’s Ling-1T: Enter the Trillion Parameter Club (X, HF, Try it)On the complete opposite end of the scale (about 3 OOM away), we have Ling-1Tfrom inclusionAI. This is a 1 TRILLION parameter Mixture-of-Experts (MoE) model. The key here is efficiency; while it has a trillion total parameters, it only uses about 37 billion active parameters per token.The benchmarks are wild, showing it beating models like GPT-5-Main (in non-thinking mode) and Gemini 2.5 on a range of reasoning tasks. They claim to match Gemini’s performance using about half the compute. Of course, with any new model posting huge scores, there’s always the question of whether it was trained on the public test sets, but the results are undeniably impressive. It’s another example of the push towards maintaining top-tier performance while drastically reducing the computational cost of inference.More Open Source Goodness: Microsoft, AI21, and IBMIt didn’t stop there.* Microsoft released UserLM-8B, a fascinating Llama 3 finetune trained not to be an assistant, but to simulate the user in a conversation. As Yam explained from his own experience, this is a super useful technique for generating high-quality, multi-turn synthetic data to train more robust chatbot agents. (X, HF)* Our friends at AI21 Labs are back with Jamba Reasoning 3B, a tiny but mighty 3-billion-parameter model. It uses a hybrid SSM-Transformer architecture, which makes it incredibly fast for its size, making it a great option for local inference on a laptop.* IBM also released their Granite family of models, which also use a hybrid design. Their big focus is on enterprise-grade governance and trust, and it’s the first open model family to get an ISO certification for AI management systems.Big Company Moves: Grok Imagine Levels Up... And Leans InFinally, let’s talk about the latest update to Grok Imagine. They’ve rolled out video generation with synchronized sound, and it’s fast—often faster than Sora. The quality has significantly improved, and it’s a powerful tool.However, I have to talk about the other side of this. Grok is positioning itself as the “uncensored” alternative, and they are leaning into that hard. Their video generator has a “spicy” mode that explicitly generates 18+ content. But the thing that truly disturbed me was a new feature with their animated character, “Annie.” It’s a gamified engagement mechanic where you “make your connection better” by talking to her every day to unlock special rewards, like new outfits.To be blunt, this is disgusting. We talk a lot on this show about the immense responsibility that comes with building these powerful AIs. I know from my conversations with folks at OpenAI and other labs that they think deeply about safety, preventing misuse, and the psychological impact these systems can have. This feature from Grok is the polar opposite. It leans into the worst fears about AI creating addictive, para-social relationships. It’s exploitative, and frankly, the team behind it should reconsider their choices IMO. All righty, this is mostly the new for this week, it’s been a very busy week, and if you’d like to see our live coverage + DevDay keynote + interviews I’ve had with Simon Willison , Greg Kamradt, Jeffrey Huber, Allesio from Latent.Space, Matthew Berman and more impactful folks, our livestream can be found here: I’m incredibly humbled and privileged to keep being invited to the Dev Day, and looking forward to cover more events, with exclusive interviews, on the ground reporting and insights. Please subscribe if you like this content to continue. TL;DR and Show Notes* Show Notes & Guests* Alex Volkov - AI Evangelist & Weights & Biases (@altryne)* Co-Hosts - @WolframRvnwlf, @yampeleg, @nisten, @ldjconfirmed* Guest: Kyle Corbitt - OpenPipe / CoreWeave (@corbtt)* Guest: Eric Provencher - Repo Prompt (@pvncher)* OpenAI Dev Day* OpenAI AgentKit All-in-One Agent Builder (X, OpenAI)* ChatGPT Apps & New APIs (GPT-5-pro, SORA, realtime-mini)* Open Source LLMs* Microsoft UserLM-8B Model Released (X, HF)* Samsung Tiny Recursive Model (TRM) (X, Blog, arXiv)* AI21 Labs releases Jamba Reasoning 3B (X, HF)* inclusionAI debuts Ling-1T: Trillion-Scale Efficient Reasoner (X, HF, Try it)* IBM Granite Models* Evals* Repo Bench by Repo Prompt (Web)* Big CO LLMs + APIs* Qwen 3 Omni & Realtime Models* Google DeepMind unveils Gemini 2.5 Computer-Use model (X, Blog)* Google Gemini Flash 2.5 (new)* Grok Imagine updated with video and sound* This weeks Buzz* OpenPipe (part of Coreweave,W&B) launch Serverless RL (X, Blog, Notebook)* Vision & Video* Ovi: Open Source Video & Synchronized Audio Generation (X, HF)* Voice & Audio* GPT-realtime-mini: OpenAI’s ultra-fast speech-to-speech model API (OpenAI Blog, TechCrunch)* AI Art & Diffusion & 3D* Bagel.com: Paris – Decentralized Diffusion Model (X, HF, Blogpost) This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit sub.thursdai.news/subscribe

Hey everyone, Alex here (yes the real me if you’re reading this) The weeks are getting crazier, but what OpenAI pulled this week, with a whole new social media app attached to their latest AI breakthroughs is definitely breathtaking! Sora2 released and instantly became a viral sensation, shooting to the top 3 free iOS spot on AppStore, with millions of videos watched, and remixed. On weeks like these, even huge releases like Claude 4.5 are taking the backseat, but we still covered them! For listeners of the pod, the second half of the show was very visual heavy, so it may be worth it watching the YT video attached in a comment if you want to fully experience the Sora revolution with us! (and if you want a SORA invite but don’t have one yet, more on that below) ThursdAI - if you find this valuable, please support us by subscribing! Sora 2 - the AI video model that signifies a new era of social mediaLook, you’ve probably already heard about the SORA-2 release, but in case you haven’t, OpenAI released a whole new model, but attached it to a new, AI powered social media experiment in the form of a very addictive TikTok style feed. Besides being hyper-realistic, and producing sounds and true to source voice-overs, Sora2 asks you to create your own “Cameo” by taking a quick video, and then allows you to be featured in your own (and your friends) videos. This makes a significant break from the previously “slop” based meta Vibes, becuase, well, everyone loves seeing themselves as the stars of the show! Cameos are a stroke of genius, and what’s more, one can allow everyone to use their Cameo, which is what Sam Altman did at launch, making everyone Cameo him, and turning him, almost instantly into one of the most meme-able (and approachable) people on the planet! Sam sharing away his likeness like this for the sake of the app achieved a few things, it added trust in the safety features, made it instantly viral and showed folks they shouldn’t be afraid of adding their own likeness. Vibes based feed and remixingSora 2 is also unique in that, it’s the first social media with UGC (user generated content) where content can ONLY be generated, and all SORA content is created within the app. It’s not possible to upload pictures that have people to create the posts, and you can only create posts with other folks if you have access to their Cameos, or by Remixing existing creations. Remixing is also a way to let users “participate” in the creation process, by adding their own twist and vibes! Speaking of Vibes, while the SORA app has an algorithmic For You page, they have a completely novel and new way to interact with the algorithm, by using their Pick a Mood feature, where you can describe which type of content you want to see, or not see, with natural language! I believe that this feature will come to all social media platforms later, as it’s such a game changer. Want only content in a specific language? or content that doesn’t have Sam Altman in it? Just ask! Content that makes you feel goodThe most interesting thing is about the type of content is, there’s no sexualisation (because all content is moderated by OpenAI strong filters), and no gore etc. OpenAI has clearly been thinking about teenagers and have added parent controls, things like being able to turn of the For You page completely etc to the mix. Additionally, SORA seems to be a very funny model, and I mean this literally. You can ask the video generation for a joke and you’ll often get a funny one. The scene setup, the dialogue, the things it does even unprompted are genuinely entertaining. AI + Product = Profit? OpenAI shows that they are one of the worlds best product labs in the world, not just a foundational AI lab. Most AI advancements are tied to products, and in this case, the whole experience is so polished, it’s hard to accept that it’s a brand new app from a company that didn’t do social before. There’s very little buggy behavior, videos are loaded up quick, there’s even DMs! I’m thoroughly impressed and am immersing myself in the SORA sphere. Please give me a follow there and feel free to use my Cameo by tagging @altryne in there. I love seeing how folks have used my Cameo, it makes me laugh 😂 The copyright question is.. wildRemember last year when I asked Sam why Advanced Voice Mode couldn’t sing Happy Birthday? He said they didn’t have classifiers to detect IP violations. Well, apparently that’s not a concern anymore because SORA 2 will happily generate perfect South Park episodes, Rick and Morty scenes, and Pokemon battles. They’re not even pretending they didn’t train on this stuff. You can even generate videos with any dead famous person (I’ve had zoom meetings with Michael Jackson and 2Pac, JFK and Mister Rogers) Our friend Ryan Carson already used it to create a YouTube short ad for his startup in two minutes. What would have cost $100K and three months now takes six generations and you’re done. This is the real game-changer for businesses.Getting invitedEDIT: If you’re reading this on Friday, try the code `FRIYAY` and let me know in comments if it worked for you 🙏I wish I would have invites for all of you, but all invited users have 4 other folks they can invite, so we shared a bunch of invites during the live show, and asked folks to come back and invite other listeners, this went on for half an hour so I bet we’ve got quite a few of you in! If you’re still looking for an invite, you can visit the thread on X and see who claimed and invite and ask them for one, tell them you’re also a ThursdAI listener, they hopefully will return the favor! Alternatively, OpenAI employees often post codes with a huge invite ratio, so follow @GabrielPeterss4 who often posts codes and you can get in there fairly quick, and if you’re not in the US, I heard a VPN works well. Just don’t forget to follow me on there as well 😉A Week with OpenAI Pulse: The Real Agentic Future is HereListen to me, this may be a hot take. I think OpenAI Pulse is a bigger news story than Sora. I’ve told you about Pulse last week, but today on the show I was able to share my weeks worth of experience, and honestly, it’s now the first thing I look at when I wake up in the morning after brushing my teeth! While Sora is changing media, Pulse is changing how we interact with AI on a fundamental level. Released to Pro subscribers for now, Pulse is an agentic, personalized feed that works for you behind the scenes. Every morning, it delivers a briefing based on your interests, your past conversations, your calendar—everything. It’s the first asynchronous AI agent I’ve used that feels truly proactive.You don’t have to trigger it. It just works. It knew I had a flight to Atlanta and gave me tips. I told it I was interested in Halloween ideas for my kids, and now it’s feeding me suggestions. Most impressively, this week it surfaced a new open-source video model, Kandinsky 5.0, that I hadn’t seen anywhere on X or my usual news feeds. An agent found something new and relevant for my show, without me even asking.This is it. This is the life-changing-level of helpfulness we’ve all been waiting for from AI. Personalized, proactive agents are the future, and Pulse is the first taste of it that feels real. I cannot wait for my next Pulse every morning.This Week’s Buzz: The AI Build-Out is NOT a BubbleThis show is powered by Weights & Biases from CoreWeave, and this week that’s more relevant than ever. I just got back from a company-wide offsite where we got a glimpse into the future of AI infrastructure, and folks, the scale is mind-boggling.CoreWeave, our parent company, is one of the key players providing the GPU infrastructure that powers companies like OpenAI and Meta. And the commitments being made are astronomical. In the past few months, CoreWeave has locked in a $22.4B deal with OpenAI, a $14.2B pact with Meta, and a $6.3B “backstop” guarantee with NVIDIA that runs through 2032.If you hear anyone talking about an “AI bubble,” show them these numbers. These are multi-year, multi-billion dollar commitments to build the foundational compute layer for the next decade of AI. The demand is real, and it’s accelerating. And the best part? As a Weights & Biases user, you have access to this same best-in-class infrastructure that runs OpenAI through our inference services. Try wandb.me/inference, and let me know if you need a bit of a credit boost! Claude Sonnet 4.5: The New Coding King Has a Few QuirksOn any other week, Anthropic’s release of Claude Sonnet 4.5 would’ve been the headline news. They’re positioning it as the new best model for coding and complex agents, and the benchmarks are seriously impressive. It matches or beats their previous top-tier model, Opus 4.1, on many difficult evals, all while keeping the same affordable price as the previous Sonnet.One of the most significant jumps is on the OS World benchmark, which tests an agent’s ability to use a computer—opening files, manipulating windows, and interacting with applications. Sonnet 4.5 scored a whopping 61.4%, a massive leap from Opus 4.1’s 44%. This clearly signals that Anthropic is doubling down on building agents that can act as real digital assistants.However, the real-world experience has been a bit of a mixed bag. My co-host Ryan Carson, whose company Amp switched over to 4.5 right away, noted some regressions and strange errors, saying they’re even considering switching back to the previous version until the rough edges are smoothed out. Nisten also found it could be more susceptible to “slop catalysts” in prompting. It seems that while it’s incredibly powerful, it might require some re-prompting and adjustments to get the best, most stable results. The jury’s still out, but it’s a potent new tool in the developer’s arsenal.Open Source LLMs: DeepSeek’s Attention RevolutionDespite the massive news from the big companies, open source still brought the heat this week, with one release in particular representing a fundamental breakthrough.DeepSeek released V3.2 Experimental, and the big news is DSA, or DeepSeek Sparse Attention. For those who don’t know, one of the biggest bottlenecks in LLMs is the “quadratic attention problem”—as you double the context length, the computation and memory required quadruple. This makes very long contexts incredibly expensive. DeepSeek’s new architecture makes the cost curve nearly flat, allowing for massive context at a fraction of the cost, all while maintaining the same SOTA performance as their previous model.This is one of those “unhobbling moments,” like the invention of RoPE or GRPO, that moves the entire field forward. Everyone will be able to implement this, making all open-source models faster and more efficient. It’s a huge deal.We also saw major releases from Z.ai with GLM-4.6, an advanced agentic model with a 200K context window that’s getting incredibly close to Claude’s performance, and a surprise from ServiceNow SLAM Labs, who dropped Apriel-1.5-15B, a frontier-level multimodal model that’s fully open source. It’s amazing to see a huge enterprise company contributing to the open-source ecosystem at this level.Multimodal Madness: Audio, Video, and Image Models updatesThe torrent of releases continued across all modalities this week, a bit overshadowed by SORA but definitely still happened (all links in the TL;DR section)In voice and audio, our friends at Hume AI launched Octave 2, their next-gen text-to-speech model that’s faster, cheaper, and now fluent in over 11 languages. We also saw LFM2-Audio from Liquid AI, an incredibly efficient 1.5B parameter end-to-end audio model with sub-100ms latency.In video, the open-source community answered Sora 2 with Kandinsky 5.0, a new 2B parameter text-to-video model that is claiming the #1 spot in open source and looks incredibly promising. And as I mentioned on the show, I wouldn’t have even known about it if it weren’t for my new personal AI agent, Pulse!Finally, in AI art, Tencent dropped a monster: HunyuanImage 3.0, a massive 80-billion-parameter open-source text-to-image model. The scale of these open-source releases is just breathtaking.Agentic browsing for all is hereJust as I was wrapping up the show, Perplexity has decided to let everyone in to use their Comet Agentic browser. I strongly recommend it, as I switched to it lately and it’s great! I’m using it right now to run some agents, it can click stuff, scroll through stuff, collect info across tabs, it’s really great. Give it a spin, really, it’s worth getting into the habit of agentic browsing! Many of you were asking me for invites before, well, it’s free access now, download it here (not sponsored, I just really like it) Phew, ok, this was a WILD week, and I’m itching to get back to creating and seeing all the folks who used my Cameo on SORA, which you can see too btw if you hit the Cameo button here (https://sora.chatgpt.com/profile/altryne) Next week is OpenAI’s Dev Day, and for the third year in a row we’re going to cover it, so follow us on social media and tune in Monday 8:30am Pacific. We’ll be live streaming from the location and re-streaming the keynote with Sam so don’t miss it! TL;DR and Show NotesHosts and Guests:* Alex Volkov - AI Evangelist & Weights & Biases (@altryne)* Co Hosts - @WolframRvnwlf @yampeleg @nisten @ldjconfirmed @ryancarsonBig CO LLMs + APIs:* OpenAI releases SORA2 + a new social media app (X, Blog, App download)* Anthropic releases Claude Sonnet 4.5 - same price as 4.1 - leading coding model (X)* OpenAI launches Instant Checkout & Agentic Commerce Protocol (X, Protocol)Open Source LLMs:* DeepSeek V3.2 Exp: Sparse Attention, Cost Drop (X, Evals, HF)* Apriel-1.5-15B-Thinker by ServiceNow SLAM Labs (X, HF, Arxiv)* Z.ai GLM-4.6: advanced Agentic flagship model (X, Blog, HF)This weeks Buzz:* CoreWeave locks $22.4B OpenAI, a $6.3B NVIDIA “backstop”, and a $14.2B Meta compute pact (X)Voice & Audio:* Hume AI launches Octave 2 (X, Blog)* LFM2-Audio: End-to-end audio foundation model (X, Blog, HF)Vision & Video:* Kandinsky 5.0 T2V Lite: #1 open-source text-to-video (Blog, GitHub, HF, Try It)AI Art & Diffusion & 3D:* HunyuanImage 3.0: 80B Open-Source Text-to-Image by Tencent (X, HF, Github) This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit sub.thursdai.news/subscribe

This is a free preview of a paid episode. To hear more, visit sub.thursdai.newsHola AI aficionados, it’s yet another ThursdAI, and yet another week FULL of AI news, spanning Open Source LLMs, Multimodal video and audio creation and more! Shiptember as they call it does seem to deliver, and it was hard even for me to follow up on all the news, not to mention we had like 3-4 breaking news during the show today! This week was yet another Qwen-mas, with Alibaba absolutely dominating across open source, but also NVIDIA promising to invest up to $100 Billion into OpenAI. So let’s dive right in! As a reminder, all the show notes are posted at the end of the article for your convenience. ThursdAI - Because weeks are getting denser, but we’re still here, weekly, sending you the top AI content! Don’t miss outTable of Contents* Open Source AI* Qwen3-VL Announcement (Qwen3-VL-235B-A22B-Thinking):* Qwen3-Omni-30B-A3B: end-to-end SOTA omni-modal AI unifying text, image, audio, and video* DeepSeek V3.1 Terminus: a surgical bugfix that matters for agents* Evals & Benchmarks: agents, deception, and code at scale* Big Companies, Bigger Bets!* OpenAI: ChatGPT Pulse: Proactive AI news cards for your day* XAI Grok 4 fast - 2M context, 40% fewer thinking tokens, shockingly cheap* Alibaba Qwen-Max and plans for scaling* This Week’s Buzz: W&B Fully Connected is coming to London and Tokyo & Another hackathon in SF* Vision & Video: Wan 2.2 Animate, Kling 2.5, and Wan 4.5 preview* Moondream-3 Preview - Interview with co-founders Via & Jay* Wan open sourced Wan 2.2 Animate (aka “Wan Animate”): motion transfer and lip sync* Kling 2.5 Turbo: cinematic motion, cheaper and with audio* Wan 4.5 preview: native multimodality, 1080p 10s, and lip-synced speech* Voice & Audio* ThursdAI - Sep 25, 2025 - TL;DR & Show notesOpen Source AIThis was a Qwen-and-friends week. I joked on stream that I should just count how many times “Alibaba” appears in our show notes. It’s a lot.Qwen3-VL Announcement (Qwen3-VL-235B-A22B-Thinking): (X, HF, Blog, Demo)Qwen 3 launched earlier as a text-only family; the vision-enabled variant just arrived, and it’s not timid. The “thinking” version is effectively a reasoner with eyes, built on a 235B-parameter backbone with around 22B active (their mixture-of-experts trick). What jumped out is the breadth of evaluation coverage: MMU, video understanding (Video-MME, LVBench), 2D/3D grounding, doc VQA, chart/table reasoning—pages of it. They’re showing wins against models like Gemini 2.5 Pro and GPT‑5 on some of those reports, and doc VQA is flirting with “nearly solved” territory in their numbers.Two caveats. First, whenever scores get that high on imperfect benchmarks, you should expect healthy skepticism; known label issues can inflate numbers. Second, the model is big. Incredible for server-side grounding and long-form reasoning with vision (they’re talking about scaling context to 1M tokens for two-hour video and long PDFs), but not something you throw on a phone.Still, if your workload smells like “reasoning + grounding + long context,” Qwen 3 VL looks like one of the strongest open-weight choices right now.Qwen3-Omni-30B-A3B: end-to-end SOTA omni-modal AI unifying text, image, audio, and video (HF, GitHub, Qwen Chat, Demo, API)Omni is their end-to-end multimodal chat model that unites text, image, and audio—and crucially, it streams audio responses in real time while thinking separately in the background. Architecturally, it’s a 30B MoE with around 3B active parameters at inference, which is the secret to why it feels snappy on consumer GPUs.In practice, that means you can talk to Omni, have it see what you see, and get sub-250 ms replies in nine speaker languages while it quietly plans. It claims to understand 119 languages. When I pushed it in multilingual conversational settings it still code-switched unexpectedly (Chinese suddenly appeared mid-flow), and it occasionally suffered the classic “stuck in thought” behavior we’ve been seeing in agentic voice modes across labs. But the responsiveness is real, and the footprint is exciting for local speech streaming scenarios. I wouldn’t replace a top-tier text reasoner with this for hard problems, yet being able to keep speech native is a real UX upgrade.Qwen Image Edit, Qwen TTS Flash, and Qwen‑GuardQwen’s image stack got a handy upgrade with multi-image reference editing for more consistent edits across shots—useful for brand assets and style-tight workflows. TTS Flash (API-only for now) is their fast speech synth line, and Q‑Guard is a new safety/moderation model from the same team. It’s notable because Qwen hasn’t really played in the moderation-model space before; historically Meta’s Llama Guard led that conversation.DeepSeek V3.1 Terminus: a surgical bugfix that matters for agents (X, HF)DeepSeek whale resurfaced to push a small 0.1 update to V3.1 that reads like a “quality and stability” release—but those matter if you’re building on top. It fixes a code-switching bug (the “sudden Chinese” syndrome you’ll also see in some Qwen variants), improves tool-use and browser execution, and—importantly—makes agentic flows less likely to overthink and stall. On the numbers, Humanities Last Exam jumped from 15 to 21.7, while LiveCodeBench dipped slightly. That’s the story here: they traded a few raw points on coding for more stable, less dithery behavior in end-to-end tasks. If you’ve invested in their tool harness, this may be a net win.Liquid Nanos: small models that extract like they’re big (X, HF)Liquid Foundation Models released “Liquid Nanos,” a set of open models from roughly 350M to 2.6B parameters, including “extract” variants that pull structure (JSON/XML/YAML) from messy documents. The pitch is cost-efficiency with surprisingly competitive performance on information extraction tasks versus models 10× their size. If you’re doing at-scale doc ingestion on CPUs or small GPUs, these look worth a try.Tiny IBM OCR model that blew up the charts (HF)We also saw a tiny IBM model (about 250M parameters) for image-to-text document parsing trending on Hugging Face. Run in 8-bit, it squeezes into roughly 250 MB, which means Raspberry Pi and “toaster” deployments suddenly get decent OCR/transcription against scanned docs. It’s the kind of tiny-but-useful release that tends to quietly power entire products.Meta’s 32B Code World Model (CWM) released for agentic code reasoning (X, HF)Nisten got really excited about this one, and once he explained it, I understood why. Meta released a 32B code world model that doesn’t just generate code - it understands code the way a compiler does. It’s thinking about state, types, and the actual execution context of your entire codebase.This isn’t just another coding model - it’s a fundamentally different approach that could change how all future coding models are built. Instead of treating code as fancy text completion, it’s actually modeling the program from the ground up. If this works out, expect everyone to copy this approach.Quick note, this one was released with a research license only! Evals & Benchmarks: agents, deception, and code at scaleA big theme this week was “move beyond single-turn Q&A and test how these things behave in the wild.” with a bunch of new evals released. I wanted to cover them all in a separate segment. OpenAI’s GDP Eval: “economically valuable tasks” as a bar (X, Blog)OpenAI introduced GDP Eval to measure model performance against real-world, economically valuable work. The design is closer to how I think about “AGI as useful work”: 44 occupations across nine sectors, with tasks judged against what an industry professional would produce.Two details stood out. First, OpenAI’s own models didn’t top the chart in their published screenshot—Anthropic’s Claude Opus 4.1 led with roughly a 47.6% win rate against human professionals, while GPT‑5-high clocked in around 38%. Releasing a benchmark where you’re not on top earns respect. Second, the tasks are legit. One example was a manufacturing engineer flow where the output required an overall design with an exploded view of components—the kind of deliverable a human would actually make.What I like here isn’t the precise percent; it’s the direction. If we anchor progress to tasks an economy cares about, we move past “trivia with citations” and toward “did this thing actually help do the work?”GAIA 2 (Meta Super Intelligence Labs + Hugging Face): agents that execute (X, HF)MSL and HF refreshed GAIA, the agent benchmark, with a thousand new human-authored scenarios that test execution, search, ambiguity handling, temporal reasoning, and adaptability—plus a smartphone-like execution environment. GPT‑5-high led across execution and search; Kimi’s K2 was tops among open-weight entries. I like that GAIA 2 bakes in time and budget constraints and forces agents to chain steps, not just spew plans. We need more of these.Scale AI’s “SWE-Bench Pro” for coding in the large (HF)Scale dropped a stronger coding benchmark focused on multi-file edits, 100+ line changes, and large dependency graphs. On the public set, GPT‑5 (not Codex) and Claude Opus 4.1 took the top two slots; on a commercial set, Opus edged ahead. The broader takeaway: the action has clearly moved to test-time compute, persistent memory, and program-synthesis outer loops to get through larger codebases with fewer invalid edits. This aligns with what we’re seeing across ARC‑AGI and SWE‑bench Verified.The “Among Us” deception test (X)One more that’s fun but not frivolous: a group benchmarked models on the social deception game Among Us. OpenAI’s latest systems reportedly did the best job both lying convincingly and detecting others’ lies. This line of work matters because social inference and adversarial reasoning show up in real agent deployments—security, procurement, negotiations, even internal assistant safety.Big Companies, Bigger Bets!Nvidia’s $100B pledge to OpenAI for 10GW of computeLet’s say that number again: one hundred billion dollars. Nvidia announced plans to invest up to $100B into OpenAI’s infrastructure build-out, targeting roughly 10 gigawatts of compute and power. Jensen called it the biggest infrastructure project in history. Pair that with OpenAI’s Stargate-related announcements—five new datacenters with Oracle and SoftBank and a flagship site in Abilene, Texas—and you get to wild territory fast.Internal notes circulating say OpenAI started the year around 230MW and could exit 2025 north of 2GW operational, while aiming at 20GW in the near term and a staggering 250GW by 2033. Even if those numbers shift, the directional picture is clear: the GPU supply and power curves are going vertical.Two reactions. First, yes, the “infinite money loop” memes wrote themselves—OpenAI spends on Nvidia GPUs, Nvidia invests in OpenAI, the market adds another $100B to Nvidia’s cap for good measure. But second, the underlying demand is real. If we need 1–8 GPUs per “full-time agent” and there are 3+ billion working adults, we are orders of magnitude away from compute saturation. The power story is the real constraint—and that’s now being tackled in parallel.OpenAI: ChatGPT Pulse: Proactive AI news cards for your day (X, OpenAI Blog)In a #BreakingNews segment, we got an update from OpenAI, that currently works only for Pro users but will come to everyone soon. Proactive AI, that learns from your chats, email and calendar and will show you a new “feed” of interesting things every morning based on your likes and feedback! Pulse marks OpenAI’s first step toward an AI assistant that brings the right info before you ask, tuning itself with every thumbs-up, topic request, or app connection. I’ve tuned mine for today, we’ll see what tomorrow brings! P.S - Huxe is a free app from the creators of NotebookLM (Ryza was on our podcast!) that does a similar thing, so if you don’t have pro, check out Huxe, they just launched! XAI Grok 4 fast - 2M context, 40% fewer thinking tokens, shockingly cheap (X, Blog)xAI launched Grok‑4 Fast, and the name fits. Think “top-left” on the speed-to-cost chart: up to 2 million tokens of context, a reported 40% reduction in reasoning token usage, and a price tag that’s roughly 1% of some frontier models on common workloads. On LiveCodeBench, Grok‑4 Fast even beat Grok‑4 itself. It’s not the most capable brain on earth, but as a high-throughput assistant that can fan out web searches and stitch answers in something close to real time, it’s compelling.Alibaba Qwen-Max and plans for scaling (X, Blog, API)Back in the Alibaba camp, they also released their flagship API model, Qwen 3 Max, and showed off their future roadmap. Qwen-max is over 1T parameters, MoE that gets 69.6 on Swe-bench verified and outperforms GPT-5 on LMArena! And their plan is simple: scale. They’re planning to go from 1 million to 100 million token context windows and scale their models into the terabytes of parameters. It culminated in a hilarious moment on the show where we all put on sunglasses to salute a slide from their presentation that literally said, “Scaling is all you need.” AGI is coming, and it looks like Alibaba is one of the labs determined to scale their way there. Their release schedule lately (as documented by Swyx from Latent.space) is insane. This Week’s Buzz: W&B Fully Connected is coming to London and Tokyo & Another hackathon in SFWeights & Biases (now part of the CoreWeave family) is bringing Fully Connected to London on Nov 4–5, with another event in Tokyo on Oct 31. If you’re in Europe or Japan and want two days of dense talks and hands-on conversations with teams actually shipping agents, evals, and production ML, come hang out. Readers got a code on stream; if you need help getting a seat, ping me directly.Links: fullyconnected.comWe are also opening up registrations to our second WeaveHacks hackathon in SF, October 11-12, yours trully will be there, come hack with us on Self Improving agents! Register HEREVision & Video: Wan 2.2 Animate, Kling 2.5, and Wan 4.5 previewThis is the most exciting space in AI week-to-week for me right now. The progress is visible. Literally.Moondream-3 Preview - Interview with co-founders Via & JayWhile I’ve already reported on Moondream-3 in the last weeks newsletter, this week we got the pleasure of hosting Vik Korrapati and Jay Allen the co-founders of MoonDream to tell us all about it. Tune in for that conversation on the pod starting at 00:33:00Wan open sourced Wan 2.2 Animate (aka “Wan Animate”): motion transfer and lip sync Tongyi’s Wan team shipped an open-source release that the community quickly dubbed “Wanimate.” It’s a character-swap/motion transfer system: provide a single image for a character and a reference video (your own motion), and it maps your movement onto the character with surprisingly strong hair/cloth dynamics and lip sync. If you’ve used runway’s Act One, you’ll recognize the vibe—except this is open, and the fidelity is rising fast.The practical uses are broader than “make me a deepfake.” Think onboarding presenters with perfect backgrounds, branded avatars that reliably say what you need, or precise action blocking without guessing at how an AI will move your subject. You act it; it follows.Kling 2.5 Turbo: cinematic motion, cheaper and with audioKling quietly rolled out a 2.5 Turbo tier that’s 30% cheaper and finally brings audio into the loop for more complete clips. Prompts adhere better, physics look more coherent (acrobatics stop breaking bones across frames), and the cinematic look has moved from “YouTube short” to “film-school final.” They seeded access to creators and re-shared the strongest results; the consistency is the headline. (Source X: @StevieMac03)I’ve chatted with my kiddos today over facetime, and they were building minecraft creepers. I took a screenshot, sent to Nano Banana to make their creepers into actual minecraft ones, and then with Kling, Animated the explosions for them. They LOVED it! Animations were clear, while VEO refused for me to even upload their images, Kling didn’t care hahaWan 4.5 preview: native multimodality, 1080p 10s, and lip-synced speechWan also teased a 4.5 preview that unifies understanding and generation across text, image, video, and audio. The eye-catching bit: generate a 1080p, 10-second clip with synced speech from just a script. Or supply your own audio and have it lip-sync the shot. I ran my usual “interview a polar bear dressed like me” test and got one of the better results I’ve seen from any model. We’re not at “dialogue scene” quality, but “talking character shot” is getting… good. The generation of audio (not only text + lipsync) is one of the best ones besides VEO, it’s really great to see how strongly this improves, sad that this wasn’t open sourced! And apparently it supports “draw text to animate” (Source: X) Voice & AudioSuno V5: we’ve entered the “I can’t tell anymore” eraSuno calls V5 a redefinition of audio quality. I’ll be honest, I’m at the edge of my subjective hearing on this. I’ve caught myself listening to Suno streams instead of Spotify and forgetting anything is synthetic. The vocals feel more human, the mixes cleaner, and the remastering path (including upgrading V4 tracks) is useful. The last 10% to “you fooled a producer” is going to be long, but the distance between V4 and V5 already makes me feel like I should re-cut our ThursdAI opener.MiMI Audio: a small omni-chat demo that hints at the floorWe tried a MiMI Audio demo live—a 7B-ish model with speech in/out. It was responsive but stumbled on singing and natural prosody. I’m leaving it in here because it’s a good reminder that the open floor for “real-time voice” is rising quickly even for small models. And the moment you pipe a stronger text brain behind a capable, native speech front-end, the UX leap is immediate.Ok, another DENSE week that finishes up Shiptember, tons of open source, Qwen (Tongyi) shines, and video is getting so so good. This is all converging folks, and honestly, I’m just happy to be along for the ride! This week was also Rosh Hashanah, which is the Jewish new year, and I’ve shared on the pod that I’ve found my X post from 3 years ago, using the state of the art AI models of the time. WHAT A DIFFERENCE 3 years make, just take a look, I had to scale down the 4K one from this year just to fit into the pic! Shana Tova to everyone who’s reading this, and we’ll see you next week 🫡ThursdAI - Sep 25, 2025 - TL;DR & Show notes* Hosts and Guests* Alex Volkov - AI Evangelist & Weights & Biases (@altryne)* Co Hosts - @yampeleg @nisten @ldjconfirmed @ryancarson* Guest - Vik Korrapathy (@vikhyatk) - Moondream* Open Source AI (LLMs, VLMs, Papers & more)* DeepSeek V3.1 Terminus: cleaner bilingual output, stronger agents, cheaper long-context (X, HF)* Meta’s 32B Code World Model (CWM) released for agentic code reasoning (X, HF)* Alibaba Tongyi Qwen on a release streak again:

Hey folks, What an absolute packed week this week, which started with yet another crazy model release from OpenAI, but they didn't stop there, they also announced GPT-5 winning the ICPC coding competitions with 12/12 questions answered which is apparently really really hard! Meanwhile, Zuck took the Meta Connect 25' stage and announced a new set of Meta glasses with a display! On the open source front, we yet again got multiple tiny models doing DeepResearch and Image understanding better than much larger foundational models.Also, today I interviewed Jeremy Berman, who topped the ArcAGI with a 79.6% score and some crazy Grok 4 prompts, a new image editing experience called Reve, a new world model and a BUNCH more! So let's dive in! As always, all the releases, links and resources at the end of the article. ThursdAI - Recaps of the most high signal AI weekly spaces is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.Table of Contents* Codex comes full circle with GPT-5-Codex agentic finetune* Meta Connect 25 - The new Meta Glasses with Display & a neural control interface* Jeremy Berman: Beating frontier labs to SOTA score on ARC-AGI* This Week’s Buzz: Weave inside W&B models—RL just got x-ray vision* Open Source* Perceptron Isaac 0.1 - 2B model that points better than GPT* Tongyi DeepResearch: A3B open-source web agent claims parity with OpenAI Deep Research* Reve launches a 4-in-1 AI visual platform taking on Nano 🍌 and Seedream* Ray3: Luma’s “reasoning” video model with native HDR, Draft Mode, and Hi‑Fi mastering* World models are getting closer - Worldlabs announced Marble* Google puts Gemini in ChromeCodex comes full circle with GPT-5-Codex agentic finetune (X, OpenAI Blog)My personal highlight of the week was definitely the release of GPT-5-Codex. I feel like we've come full circle here. I remember when OpenAI first launched a separate, fine-tuned model for coding called Codex, way back in the GPT-3 days. Now, they've done it again, taking their flagship GPT-5 model and creating a specialized version for agentic coding, and the results are just staggering.This isn't just a minor improvement. During their internal testing, OpenAI saw GPT-5-Codex work independently for more than seven hours at a time on large, complex tasks—iterating on its code, fixing test failures, and ultimately delivering a successful implementation. Seven hours! That's an agent that can take on a significant chunk of work while you're sleeping. It's also incredibly efficient, using 93% fewer tokens than the base GPT-5 on simpler tasks, while thinking for longer on the really difficult problems.The model is now integrated everywhere - the Codex CLI (just npm install -g codex), VS Code extension, web playground, and yes, even your iPhone. At OpenAI, Codex now reviews the vast majority of their PRs, catching hundreds of issues daily before humans even look at them. Talk about eating your own dog food!Other OpenAI updates from this weekWhile Codex was the highlight, OpenAI (and Google) also participated and obliterated one of the world’s hardest algorithmic competitions called ICPC. OpenAI used GPT-5 and an unreleased reasoning model to solve 12/12 questions in under 5 hours. OpenAI and NBER also released an incredible report on how over 700M people use GPT on a weekly basis, with a lot of insights, that are summed up in this incredible graph:Meta Connect 25 - The new Meta Glasses with Display & a neural control interfaceJust when we thought the week couldn't get any crazier, Zuck took the stage for their annual Meta Connect conference and dropped a bombshell. They announced a new generation of their Ray-Ban smart glasses that include a built-in, high-resolution display you can't see from the outside. This isn't just an incremental update; this feels like the arrival of a new category of device. We've had the computer, then the mobile phone, and now we have smart glasses with a display.The way you interact with them is just as futuristic. They come with a "neural band" worn on the wrist that reads myoelectric signals from your muscles, allowing you to control the interface silently just by moving your fingers. Zuck's live demo, where he walked from his trailer onto the stage while taking messages and playing music, was one hell of a way to introduce a product.This is how Meta plans to bring its superintelligence into the physical world. You'll wear these glasses, talk to the AI, and see the output directly in your field of view. They showed off live translation with subtitles appearing under the person you're talking to and an agentic AI that can perform research tasks and notify you when it's done. It's an absolutely mind-blowing vision for the future, and at $799, shipping in a week, it's going to be accessible to a lot of people. I've already signed up for a demo.Jeremy Berman: Beating frontier labs to SOTA score on ARC-AGIWe had the privilege of chatting with Jeremy Berman, who just achieved SOTA on the notoriously difficult ARC-AGI benchmark using checks notes... Grok 4! 🚀He walked us through his innovative approach, which ditches Python scripts in favor of flexible "natural language programs" and uses a program-synthesis outer loop with test-time adaptation. Incredibly, his method achieved these top scores at 1/25th the cost of previous systemsThis is huge because ARC-AGI tests for true general intelligence - solving problems the model has never seen before. The chat with Jeremy is very insightful, available on the podcast starting at 01:11:00 so don't miss it!ThursdAI - Recaps of the most high signal AI weekly spaces is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.This Week’s Buzz: Weave inside W&B models—RL just got x-ray visionYou know how every RL project produces a mountain of rollouts that you end up spelunking through with grep? We just banished that misery. Weave tracing now lives natively inside every W&B Workspace run. Wrap your training-step and rollout functions in @weave.op, call weave.init(), and your traces appear alongside loss curves in real time. I can click a spike, jump straight to the exact conversation that tanked the reward, and diagnose hallucinations without leaving the dashboard. If you’re doing any agentic RL, please go treat yourself. Docs: https://weave-docs.wandb.ai/guides/tools/weave-in-workspacesOpen SourceOpen source did NOT disappoint this week as well, we've had multiple tiny models beating the giants at specific tasks! Perceptron Isaac 0.1 - 2B model that points better than GPT ( X, HF, Blog )One of the most impressive demos of the week came from a new lab, Perceptron AI. They released Isaac 0.1, a tiny 2 billion parameter "perceptive-language" model. This model is designed for visual grounding and localization, meaning you can ask it to find things in an image and it will point them out. During the show, we gave it a photo of my kid's Harry Potter alphabet poster and asked it to "find the spell that turns off the light." Not only did it correctly identify "Nox," but it drew a box around it on the poster. This little 2B model is doing things that even huge models like GPT-4o and Claude Opus can't, and it's completely open source. Absolutely wild.Moondream 3 preview - grounded vision reasoning 9B MoE (2B active) (X, HF)Speaking of vision reasoning models, just a bit after the show concluded, our friend Vik released a demo of Moondream 3, a reasoning vision model 9B (A2B) that is also topping the charts! I didn't have tons of time to get into this, but the release thread shows this to be an exceptional open source visual reasoner also beating the giants!Tongyi DeepResearch: A3B open-source web agent claims parity with OpenAI Deep Research ( X, HF )Speaking of smaller models obliterating huge ones, Tongyi released a bunch of papers and a model this week that can do Deep Research on the level of OpenAI, even beating it, with a Qwen Finetune with only 3B active parameters! With insane scores like 32.9 (38.3 in Heavy mode) on Humanity's Last Exam (OAI Deep Research gets 26%) and an insane 98.6% on SimpleQA, this innovative approach uses a lot of RL and synthetic data to train a Qwen model to find what you need. The paper is full of incredible insights into how to build automated RL environments to get to this level. AI Art, Diffusion 3D and VideoThis category of AI has been blowing up, we've seen SOTA week after week with Nano Banana then Seedream 4 and now a few more insane models.Tencent's Hunyuan released SRPO (X, HF, Project, Comparison X)(Semantic Relative Preference Optimization) which is a new method to finetune diffusion models quickly without breaking the bank. Also released a very realistic looking finetune trained with SRPO. Some of the generated results are super realistic, but it's more than just a model, there's a whole new method of finetuning here! Hunyuan also updated their 3D model and announced a full blown 3D studio that does everything from 3D object generation, meshing, texture editing & more. Reve launches a 4-in-1 AI visual platform taking on Nano 🍌 and Seedream (X, Reve, Blog)A newcomer, Reve has launched a comprehensive new AI visual platform bundling image creation, editing, remixing, creative assistant, and API integration, all aimed at making advanced editing as accessible, all using their own proprietary models. What stood out to me though, is the image editing UI, which allows you to select on your image exactly what you want to edit, write a specific prompt for that thing (change color, objects, add text etc') and then hit generate and their model takes into account all those new queues! This is way better than just ... text prompting the other models! Ray3: Luma’s “reasoning” video model with native HDR, Draft Mode, and Hi‑Fi mastering (X, Try It)Luma released the third iteration of their video model called Ray, and this one does.. HDR! But it also has Draft Mode (for quick iteration), first/last frame interpolation, and they claim to be "production ready" with extreme prompt adherence. The thing that struck me is the reasoning part, their video model is now reasoning, to let you create more complex scenes, while the model will ... evaluate itself and select the best generation for you! This is quite bonkers, can't wait to play with it! World models are getting closer - Worldlabs announced Marble (Demo)We've covered a whole host of world models, Genie3, Hunyuan 3D world models, Mirage and a bunch more! Dr FeiFei's WorldLabs have been one of the first ones to tackle the world model concept and their recent release shows incredible progress (and finally lets us play with it!) Marble takes images and creates Gussian Splats, that can be used in 3D environments. So now you can use any AI image generation and turn it into a walkable 3D world! Google puts Gemini in Chrome (X, Blog)This happened after the show today and while not fully rolled out yet, I've told you before when we covered Comet from PPXL and Dia from browser company, that Google will not be far behind! So today they announced that Gemini is coming to Chrome, and will allow users to chat with a bunch of their tabs, summarize across tabs and soon do agentic tasks like clicking things and shopping for you? 😅I wonder if this means that Google will offer this for free to the over 1B chrome users or introduce some sort of Gemini tier cross-over? Remains to be seen but very exciting to see AI browsers all over! The best feature could be a hidden one, where the Gemini in Chrome will have knowledge about your surfing history and you'll be able to ask it about that one website you visited a while ago that had sharks! Folks, I can go on and on today, literally there's a new innovative video model from ByteDance, a few more image models, but alas, I have to prioritize and give you only the top important news. So, I'll just remind that I put all the links in the TL;DR below and that you should absolutely check out the video version of our show on YT because a lot of visual things are happening and we're playing with all of them live! Hey, just before you get to the “links”, consider subscribing to help me keep this going? 🙏See you next week 🫡 Don't forget to subscribe (and if you already subbed, share this with a friend or two?) TL;DR and show notes - September 18, 2025* Hosts and Guests* Alex Volkov - AI Evangelist & Weights & Biases (@altryne)* Co Hosts - @WolframRvnwlf @ldjconfirmed @nisten* Guest : Jeremy Berman (@jerber888) - SOTA on ARC- AGI* Open Source* Perceptron AI introduces Isaac 0.1: a 2B param perceptive-language model (X, HF, Blog)* Tongyi DeepResearch: A3B open-source web agent claims parity with OpenAI Deep Research (X, HF)* Mistral updates Magistral-Small-2509 (HF)* Big CO LLMs + APIs* GPT-5-Codex release: Agentic coding upgrade for Codex (X, OpenAI Blog)* Meta Connect - New AI glasses with display, new AI mode (X Recap)* NBER & OpenAI - How People Use ChatGPT: Growth, Demographics, and Scale (X, Blog, NBER Paper)* ARC-AGI: New SOTA by Jeremy Berman and Eric Pang using Grok-4 (X, Blog)* OpenAI’s reasoning system aces 2025 ICPC World Finals with a perfect 12/12 (X)* OpenAI adds thinking budgets to ChatGPT app (X)* Gemini in Chrome: AI assistant across tabs + smarter omnibox + safer browsing (X, Blog)* Anthropic admits Claude bugs - Detailed analysis * This weeks Buzz* W&B Models + Weave! You can now log your RL runs in W&B Weave 👏 (X, W&B Link) * W&B Fully Connected London - tickets are running out! Use FCLNTHURSAI for a free ticket on me! (Register Here)* Vision & Video* Moondream 3 (Preview): 9B MoE VLM with 2B active targets frontier-level visual reasoning (X, HF)* Ray3: Luma’s “reasoning” video model with native HDR, Draft Mode, and Hi‑Fi mastering (X)* HuMo: human‑centric, multimodal video gen from ByteDance/Tsinghua (X, HF)* Voice & Audio* Reka Speech: high-throughput multilingual ASR and speech translation for batch-scale pipelines (X, Blog)* AI Art & Diffusion & 3D* Hunyuan SRPO (Semantic Relative Preference Optimization) supercharges diffusion models (X, HF, Project, Comparison X)* Hunyuan 3D 3.0 (X, Try it)* FeiFei WorldLabs presents Marble (Demo)* Reve launches 4-in-1 AI visual platform (X, Reve, Blog)* Tools* Chrome adds Gemini (Blog) This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit sub.thursdai.news/subscribe

Hey Everyone, Alex here, thanks for being a subscriber! Let's get you caught up on this weeks most important AI news! The main thing you need to know this week is likely the incredible Image model that ByteDance released, that overshoots the (incredible image model from last 2 weeks) nano 🍌. ByteDance really outdid themselves on this one! But also, a video model with super fast generation, OpenAI rumor made Larry Ellison the richest man alive, ChatGPT gets MCP powers (under a flag you can enable) and much more! This week we covered a lot of visual stuff, so while the podcast format is good enough, it's really worth tuning in to the video recording to really enjoy the full show. ThursdAI - Recaps of the most high signal AI weekly spaces is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.AI Art and DiffusionIt's rare for me to start the newsletter not on Open Source AI news, but hey, at least this way you know that I'm writing it and not some AI right? 😉ByteDance SeeDream 4 - 4K SOTA image generation and editing model with up to 6 reference images (Fal, Replicate)The level of detail on ByteDance's new model, has really made all the hosts on ThursdAI stop and go... huh? is this AI? Bytedance really outdid themselves with this image model that not only generates images, it also is a fully functional image editing with natural language model. It's a diffusion transformer, able to generate 2K and 4K images, fast (under 5 seconds?) while enabling up to 6 reference images to be provided for the generation. This is going to be incredible for all kinds of purposes, AI art, marketing etc'. The promt adherence is quite incredible, text is also crisp and sharp at those 2/4K resolutions. We created this image live on the show with it (using a prompt extended by another model)I then provided my black and white headshot and the above image and asked to replace me as a cartoon character, and it did, super quick, and even got my bomber jacket and the W&B logo on it in there! Notable, nothing else was changed in the image, showing just how incredible this one is for image editing. In you want enhanced realism, our friend FoFr from replicate, reminded us that using IMG_3984.CR2 in the prompt, will make the model show images that are closer to reality, even if they depict some incredibly unrealistic things, like a pack of lions forming his nicknameAdditional uses for this model are just getting discovered, and one user already noted that given this model outputs 4K resolution, it can be used as a creative upscaler for other model outputs. Just shove your photo from another AI in Seedream and ask for an upscale. Just be ware that creative upscalers change some amount of details in the generated picture. DecART AI Lucy 14B Redefines Video Generation speeds! If Seedteam blew my mind with images, Decart's Lucy 14B absolutely shattered my expectations for video generation speed. We're talking about generating 5-second videos from images in 6.5 seconds. That's almost faster than watching the video itself!This video model is not open source yet (despite them adding 14B to the name) but it's smaller 5B brother was open sourced. The speed to quality ratio is really insane here, and while Lucy will not generate or animate text or faces that well, it does produce some decent imagery, but SUPER fast. This is really great for iteration, as AI Video is like a roulette machine, you have to generate a lot of tries to see a good result. This paired with Seedream (which is also really fast) are a game changer in the AI Art world! So stoked to see what folks will be creating with these! Bonus Round: Decart's Real-Time Minecraft Mod for Oasis 2 (X)The same team behind Lucy also dropped Oasis 2.0, a Minecraft mod that generates game environments in real-time using diffusion models. I got to play around with it live, and watching Minecraft transform into different themed worlds as I moved through them was surreal.Want a steampunk village? Just type it in. Futuristic city? Done. The frame rate stayed impressively smooth, and the visual coherence as I moved through the world was remarkable. It's like having an AI art director that can completely reskin your game environment on demand. And while the current quality remains low res, if you consider where Stable Diffusion 1.4 was 3 years ago, and where Seedream 4 is now, and do the same extrapolation to Oasis, in 2-3 years we'll be reskinning whole games on the fly and every pixel will be generated (like Jensen loves to say!) OpenAI adds full MCP to ChatGPT (under a flag) This is huge, folks. I've been waiting for this for a while, and finally, OpenAI quietly added full MCP (Model Context Protocol) support to ChatGPT via a hidden "developer mode."How to Enable MCP in ChatGPTHere's the quick setup I showed during the stream:* Go to ChatGPT settings → Connectors* Scroll down to find "Developer Mode" and enable it* Add MCP servers (I used Rube.ai from Composio)* Use GPT-4o in developer mode to access your connectorsDuring the show, I literally had ChatGPT pull Nisten's last five tweets using the Twitter MCP connector. It worked flawlessly (though Nisten was a bit concerned about what tweets it might surface 😂).The implications are massive - you can now connect ChatGPT to GitHub, databases, your local files, or chain multiple tools together for complex workflows. As Wolfram pointed out though, watch your context usage - each MCP connector eats into that 200K limit.Big Moves: Investments and InfrastructureSpeaking of OpenAI, Let's talk money, because the stakes are getting astronomical. OpenAI reportedly has a $300 billion (!) deal with Oracle for compute infrastructure over five years, starting in 2027. That's not a typo - $60 billion per year for compute. Larry Ellison just became the world's richest person, and Oracle's stock shot up 40% on the news in just a few days! This has got to be one of the biggest compute deals the world has ever head of!The scale is hard to comprehend. We're talking about potentially millions of H100 GPUs worth of compute power. When you consider that most AI companies are still figuring out how to profitably deploy thousands of GPUs, this deal represents infrastructure investment at a completely different magnitude.Meanwhile, Mistral just became Europe's newest decacorn, valued at $13.8 billion after receiving $1.3 billion from ASML. For context, ASML makes the lithography machines that TSMC uses to manufacture chips for Nvidia. They're literally at the beginning of the AI chip supply chain, and now they're investing heavily in Europe's answer to OpenAI.Wolfram made a great point - we're seeing the emergence of three major AI poles: American companies (OpenAI, Anthropic), Chinese labs (Qwen, Kimi, Ernie), and now European players like Mistral. Each is developing distinct approaches and capabilities, and the competition is driving incredible innovation.Anthropic's Mea Culpa and Code InterpreterAfter weeks of users complaining about Claude's degraded performance, Anthropic finally admitted there were bugs affecting both Claude Opus and Sonnet. Nisten, who tracks these things closely, speculated that the issues might be related to running different quantization schemes on different hardware during peak usage times. We already reported last week that they admitted that "something was affecting intelligence" but this week they said they pinpointed (and fixed) 2 bugs realted to inference! They also launched a code interpreter feature that lets Claude create and edit files directly. It's essentially their answer to ChatGPT's code interpreter - giving Claude its own computer to work with. The demo showed it creating Excel files, PDFs, and documents with complex calculations. Having watched Claude struggle with file operations for months, this is a welcome addition.🐝 This Week's Buzz: GLM 4.5 on W&B and We're on Open Router!Over at Weights & Biases, we've got some exciting updates for you. First, we've added Zhipu AI's GLM 4.5 to W&B Inference! This 300B+ parameter model is an absolute beast for coding and tool use, ranking among the top open models on benchmarks like SWE-bench. We've heard from so many of you, including Nisten, about how great this model is, so we're thrilled to host it. You can try it out now and get $2 in free credits to start.And for all you developers out there, you can use a proxy like LiteLLM to run GLM 4.5 from our inference endpoint inside Anthropic's Claude Code if you're looking for a powerful and cheap alternative! Second, we're now on Open Router! You can find several of our hosted models, like GPT-4-OSS and DeepSeek Coder, on the platform. If you're already using Open Router to manage your model calls, you can now easily route traffic to our high-performance inference stack.Open Source Continues to ShineOpen Source LLM models took a bit of a break this week, but there were still interesting models! Baidu released ERNIE-4.5, a very efficient 21B parameter "thinking" MoE that only uses 3B active parameters per token. From the UAE, MBZUAI released K2-Think, a finetune of Qwen 2.5 that's showing some seriously impressive math scores. And Moonshot AI updated Kimi K2, doubling its context window to 256K and further improving its already excellent tool use and writing capabilities.Tencent released an update to HunyuanImage 2.1, which is a bit slow, but also generates 2K images and is decent at text. Qwen drops Qwen3-Next-80B-A3B (X, HF)In breaking news post the show (we were expecting this on the show itself), Alibaba folks dropped a much more streamlined version of the next Qwen, 80B parametes with only 3B active! They call this an "Ultra Sparse MOE" and it beats Qwen3-32B in perf, rivals Qwen3-235B in reasoning & long-context. This is quite unprecedented, as getting models as sparse as to work well takes a lot of effort and skill, but the Qwen folks delivered! ToolsWe wrapped with a quick shoutout to EBSynth, a nifty video editor that lets you draw or add elements to one frame and extrapolates to the rest—like Photoshop for motion. It's browser-based and fun for VFX tweaks; check the demo video on X. Simple but powerful for quick video hacks. Speaking of Photoshop, it was confirmed that Nano Banana is going to be embedded into Photoshop for image editing! Summary & TL;DRWhat a week—Seedream and Lucy alone have me rethinking how fast AI can iterate creatively, while MCP in ChatGPT feels like the dawn of truly accessible agents. With open-source keeping pace and big deals fueling the fire, AI's multimodal future is accelerating. Thanks for tuning in, folks; if you missed the live vibes, catch the podcast or hit sub.thursdai.news for all the links. See you next Thursday—what blew your mind this week? Drop a comment and share with a friend, it's the best way to support this endeavor! TL;DR of all topics covered:AI Models & APIs:* ChatGPT adds full MCP support - Developer mode unlocks tool connectors for 400M+ users (Setup Guide)* Sea Dream 4.0 - ByteDance's unified image generation/editing model creates 4K images in ~1.8 seconds (X, Try it)* Lucy 14B - Decart's lightning-fast video model generates 5-second clips in 6.5 seconds (Demo, Page)* Claude bug fixes - Anthropic admits to performance issues and releases code interpreter (Blog)* Sonoma Dusk & Sky - Mystery models on OpenRouter with 2M context, rumored to be Grok (OpenRouter)This Week's W&B Buzz:* OpenRouter integration - Serving models to broader developer community (Try us)* GLM 4.5 - 350B parameter coding model added to inference (X, Try It)* W&B inference in Claude Code with LiteLLM (Olaf's Guide)Open Source Releases:* ERNIE 4.5 - Baidu open-sources 21B parameter thinking model with 3B active parameters (X, HF)* K2-think - MBZUAI's Qwen 2.5 fine-tune with strong math performance (X)* Kimi K2 update - Doubled context to 256K, improved tool use (X)* HunyuanImage 2.1 - Tencent's 17B parameter open-source 2K image model (X, HF)* Qwen-next-80B-A3B - Alibaba's next frontier MoE with 3B active param (X, HF)Voice & Audio:* Qwen3-ASR-Flash - 11-language speech recognition with singing support (X)* Stable Audio 2.5 - Enterprise audio generator creating 3-minute tracks in X, Blog, Try It)* ElevenLabs Voice Remixing - Modify cloned voices for age, gender, accent (X)Business & Investment:* OpenAI-Oracle deal - $300B infrastructure agreement over 5 years* Mistral funding - $1.3B investment from ASML at $13.8B valuation (Blog)Tools:* Oasis 2.0 - Real-time Minecraft world generation mod from Decart (Try It)* EbSynth - Video editing tool for frame-by-frame manipulation (X)Hosts:* Alex Volkov (@altryne)* Wolfram RavenWlf (@WolframRvnwlf)* Yam Peleg (@yampeleg)* Nisten (@nisten)* LDJ (@ldjconfirmed) This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit sub.thursdai.news/subscribe

Wohoo, hey ya’ll, Alex here,I'm back from the desert (pic at the end) and what a great feeling it is to be back in the studio to talk about everything that happened in AI! It's been a pretty full week (or two) in AI, with Coding agent space heating up, Grok entering the ring and taking over free tokens, Codex 10xing usage and Anthropic... well, we'll get to Anthropic. Today on the show we had Roger and Bhavesh from Nous Research cover the awesome Hermes 4 release and the new PokerBots benchmark, then we had a returning favorite, Kwindla Hultman Kramer, to talk about the GA of RealTime voice from OpenAI. Plus we got some massive funding news, some drama with model quality on Claude Code, and some very exciting news right here from CoreWeave aquiring OpenPipe! 👏 So grab your beverage of choice, settle in (or skip to the part that interests you) and let's take a look at the last week (or two) in AI! ThursdAI - Recaps of the most high signal AI weekly spaces is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.Open Source: Soulful Models and Poker-Playing AgentsThis week did not disappoint as it comes to Open Source! Our friends at Nous Research released the 14B version of Hermes 4, after releasing the 405B and 70B versions last week. This company continues to excel in finetuning models for powerful, and sometimes just plain weird (in a good way) usecases. Nous Hermes 4 (14B, 70B, 405B) and the Quest for a "Model Soul" (X, HF)Roger and Bhavash from Nous came to announce the release of the smaller (14B) version of Hermes 4, and cover the last weeks releases of the larger 70B and 405B brothers. Hermes series of finetunes was always on our radar, as unique data mixes turned them into uncensored, valuable and creative models and unlocked a bunch of new use-cases. But the wildest part? They told us they intentionally stopped training the model not when reasoning benchmarks plateaued, but when they felt it started to "lose its model soul." They monitor the entropy and chaos in the model's chain-of-thought, and when it became too sterile and predictable, they hit the brakes to preserve that creative spark. This focus on qualities beyond raw benchmark scores is why Hermes 4 is showing some really interesting generalization, performing exceptionally well on benchmarks like EQBench3, which tests emotional and interpersonal understanding. It's a model that's primed for RL not just in math and code, but in creative writing, role-play, and deeper, more "awaken" conversations. It’s a soulful model that's just fun to talk to.Nous Husky Hold'em Bench: Can Your LLM Win at Poker? (Bench)As if a soulful model wasn't enough, the Nous team also dropped one of the most creative new evals I've seen in a while: Husky Hold'em Bench. We had Bhavesh, one of its creators, join the show to explain. This isn't a benchmark where the LLM plays poker directly. Instead, the LLM has to write a Python poker botfrom scratch, under time and memory constraints, which then competes against bots written by other LLMs in a high-stakes tournament. Very interesting approach, and we love creative benchmarking here at ThursdAI! This is a brilliant way to test for true strategic reasoning and planning, not just pattern matching. It's an "evergreen" benchmark that gets harder as the models get better. Early results are fascinating: Claude 4 Sonnet and Opus are currently leading the pack, but Hermes 4 is the top open-source model.More Open Source GoodnessThe hits just kept on coming this week. Tencent open-sourced Hunyuan-MT-7B, a translation model that swept the WMT2025 competition and rivals GPT-4.1 on some benchmarks. Having a small, powerful, specialized model like this is huge for anyone doing large-scale data translation for training or needing fast on-device capabilities.From Switzerland, we got Apertus-8B and 70B, a set of fully open (Apache 2.0 license, open data, open training recipes!) multilingual models trained on a massive 15 trillion tokens across 1,800 languages. It’s fantastic to see this level of transparency and contribution from European institutions.And Alibaba’s Tongyi Lab released WebWatcher, a powerful multimodal research agent that can plan steps, use a suite of tools (web search, OCR, code interpreter), and is setting new state-of-the-art results on tough visual-language benchmarks, often beating models like GPT-4o and Gemini.All links are in the TL;DR at the endBREAKING NEWS: Google Drops Embedding Gemma 308M (X, HF, Try It)Just as we were live on the show, news broke from our friends at Google. They've released Embedding Gemma, a new family of open-source embedding models. This is a big deal because they are tiny—the smallest is only 300M parameters and takes just 200MB to run—but they are topping the MTEB leaderboard for models under 500M parameters. For anyone building RAG pipelines, especially for on-device or mobile-first applications, having a small, fast, SOTA embedding model like this is a game-changer.It's so optimized for on device running that it can run fully in your browser on WebGPU, with this great example from our friend Xenova highlighted on the release blog! Big Companies, Big Money, and Big ProblemsIt was a rollercoaster week for the big labs, with massive fundraising, major product releases, and a bit of a reality check on the reliability of their services.OpenAI's GPT Real-Time Goes GA and gets an upgraded brain (X, Docs)We had the perfect guest to break down OpenAI's latest voice offering: Kwindla Kramer, founder of Daily and maintainer of the open-source PipeCat framework. OpenAI has officially taken its Realtime API to General Availability (GA), centered around the new gpt-realtime model.Kwindla explained that this is a true speech-to-speech model, not a pipeline of separate speech-to-text, LLM, and text-to-speech models. This reduces latency and preserves more nuance and prosody. The GA release comes with huge upgrades, including support for remote MCP servers, the ability to process image inputs during a conversation, and—critically for enterprise—native SIP integration for connecting directly to phone systems.However, Kwindla also gave us a dose of reality. While this is the future, for many high-stakes enterprise use cases, the multi-model pipeline approach is still more reliable. Observability is a major issue with the single-model black box; it's hard to know exactly what the model "heard." And in terms of raw instruction-following and accuracy, a specialized pipeline can still outperform the speech-to-speech model. It’s a classic jagged frontier: for the lowest latency and most natural vibe, GPT Real-Time is amazing. For mission-critical reliability, the old way might still be the right way for now.ChatGpt has branching! Just as I was about to finish writing this up, ChatGPT announced a new feature, and this one I had to tell you about! Finally you can branch chats in their interface, which is a highly requested feature! Branching seems to be live on the chat interface, and honestly, tiny but important UI changes like these are how OpenAI remains the best chat experience! The Money Printer Goes Brrrr: Anthropic's $13B RaiseLet's talk about the money. Anthropic announced it has raised an absolutely staggering $13 billion in a Series F round, valuing the company at $183 billion. Their revenue growth is just off the charts, jumping from a run rate of around $1 billion at the start of the year to over $5 billion by August. This growth is heavily driven by enterprise adoption and the massive success of Claude Code. It's clear that the AI gold rush is far from over, and investors are betting big on the major players. In related news, OpenAI is also reportedly raising $10 billion at a valuation of around $500 billion, primarily to allow employees to sell shares—a huge moment for the folks who have been building there for years.Oops... Did We Nerf Your AI? Anthropic's ApologyWhile Anthropic was celebrating its fundraise, it was also dealing with a self-inflicted wound. After days of users on X and other forums complaining that Claude Opus felt "dumber," the company finally issued a statement admitting that yes, for about three days, the model's quality was degraded due to a change in their infrastructure stack.Honestly, this is not okay. We're at a point where hundreds of thousands of developers and businesses rely on these models as critical tools. To have the quality of that tool change under your feet without any warning is a huge problem. It messes with people's ability to do their jobs and trust the platform. While it was likely an honest mistake in pursuit of efficiency, it highlights a fundamental issue with closed, proprietary models. You're at the mercy of the provider. It's a powerful argument for the stability and control that comes with open-source and self-hosted models. These companies need to realize that they are no longer just providing experimental toys; they're providing essential infrastructure, and that comes with a responsibility for stability and transparency.This Week's Buzz: CoreWeave Acquires OpenPipe! 🎉Super exciting news from the Weights & Biases and CoreWeave family - we've acquired OpenPipe! Kyle and David Corbett and their team are joining us to help build out the complete AI infrastructure stack from metal to model.OpenPipe has been doing incredible work on SFT and RL workflows with their open source ART framework. As Yam showed during the show, they demonstrated you can train a model to SOTA performance on deep research tasks for just $300 in a few hours - and it's all automated! The system can generate synthetic data, apply RLHF, and evaluate against any benchmark you specify.This fits perfectly into our vision at CoreWeave - bare metal infrastructure, training and observability with Weights & Biases, fine-tuning and RL with OpenPipe's tools, evaluation with Weave, and inference to serve it all. We're building the complete platform, and I couldn't be more excited!Vision & Speed: Apple's FastVLM (HF)Just before Apple's event next week, they dropped FastVLM - a speed-first vision model that's 85x faster on time-to-first-token than comparable models. They released it in three sizes (7B, 1.5B, and 0.5B), all optimized for on-device use.The demo that blew my mind was real-time video captioning running in WebGPU. HF CEO Clem showed it processing Apple's keynote video with maybe 250ms latency - the captions were describing what was happening almost in real-time. When someone complained it wasn't accurate because it described "an older man in headphones" when showing an F1 car, Clem pointed out that was actually the previous frame showing Tim Cook - the model was just slightly behind!Tools Showdown: Codex vs Claude CodeTo wrap up, we dove into the heated debate between Codex and Claude Code. Sam Altman reported that Codex usage is up 10x in the past two weeks (!) and improvements are coming. Yam gave us a live demo, and while Claude Code failed to even start up during the show (highlighting why people are switching), Codex with GPT-5 was smooth as butter.The key advantages? Codex authenticates with your OpenAI account (no API key juggling), it has MCP support, and perhaps most importantly - it's not just a CLI tool. You can use it for PR reviews on GitHub, as a cloud-based agent, and integrated into Cursor and Windsurf. Though as Yam pointed out, OpenAI really needs to stop calling everything "Codex" - there are like five different products with that name now! 😅If you're tried Codex (the CLI!) when it was released, and gave up, give it a try now, it's significantly upgraded! Ok, phew, what a great episode we had, if you're only reading, I strongly recommend checking out the live recording or the edited podcast, and of course, if this newsletter is helpful to you, the best way you can do to support it is to subscribe, and share with friends 👏 P.S - Just came back after my first burning man, it was a challenging, all consuming experience, where I truly disconnected for the first time (first ThursdAI in over 2 years that I didn't know what's going on with AI). It was really fun but I'm happy to =be back! See you next week! TL;DR and Show Notes* Hosts and Guests* Alex Volkov - AI Evangelist & Weights & Biases (@altryne)* Co Hosts - @WolframRvnwlf @yampeleg @nisten @ldjconfirmed* Guests - Roger Jin - @rogershijin & Bhavesh Kumar @bha_ku21* Kwindla Kramer - @kwindla* Open Source LLMs* Nous Hermes 4 — 14B launches: compact hybrid reasoning model with tool calling for local and cloud use (X, HF, Tech Report)* Tencent open-sources Hunyuan-MT-7B translation model after sweeping WMT2025 (X, HF)* Nous - Husky Hold'em Bench launches as an open-source pokerbot eval for LLM strategic play (X, Bench)* WebWatcher: Alibaba's Tongyi Lab open-sources a vision-language deep research agent that sets new SOTA (X, HF)* Apertus-8B and 70B launch as Switzerland's fully open, multilingual LLMs trained on 15T tokens across 1,800+ languages (X, HF)* Google releases Embedding Gemma - 300M param SOTA embeddings model for RAG ([Breaking News])* Big CO LLMs + APIs* Mistral adds 20+ MCP-powered connectors and controllable Memories to Le Chat for enterprise workflows (X, Blog)* Anthropic raises $13B Series F at a $183B post-money valuation (X, Blog)* OpenAI fundraises $10B at ~$500B valuation - buyback for employees* OpenAI ships gpt-realtime and takes Realtime API to GA with remote MCP tools, image input, and SIP phone calling (X)* OpenAI releases projects for free users with larger file uploads and project-only memory controls* OpenAI acquires Statsig & Alex for $1.1B+ to strengthen applications team* Grok Code 1 - now taking 50% of coding traffic on OpenRouter* Codex usage up 10x in 2 weeks per Sam Altman, with improvements coming* Anthropic admits to Claude Opus quality degradation for 3 days due to infrastructure changes* This weeks Buzz* CoreWeave buys OpenPipe! 🎉 (Blog)* Vision & Video* Apple's FastVLM-7B lands with speed-first vision encoder—85x faster TTFT vs peers (X, HF)* AI Art & Diffusion & 3D* Nano Banana (Imagen 3) continues to dominate as Google's best image model (ai.studio/banana)* Tools* Codex vs Claude Code discussion → Codex now significantly better with GPT-5 engine, GitHub PR reviews, and cloud agents This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit sub.thursdai.news/subscribe

Hey everyone, Alex here 👋This week looked quiet… until about 15 hours before we went live. Then the floodgates opened: DeepSeek dropped a hybrid V3.1 that beats their own R1 with fewer thinking tokens, ByteDance quietly shipped a 36B Apache-2.0 long-context family with a “thinking budget” knob, NVIDIA pushed a faster mixed-architecture 9B with open training data, and a stealth image editor dubbed “Nano Banana” started doing mind-bending scene edits that feel like a new tier of 3D-aware control. On the big-co side, a mystery “Sonic” model appeared in Cursor and Cline (spoiler: the function call paths say a lot), and OpenAI introduced Agents.md to stop the config-file explosion in agentic dev tools. We also got a new open desktop-agent RL framework that 4x’d OSWorld SOTA, an IBM + NASA model for solar weather, and Qwen’s fully open 20B image editor that’s shockingly capable and runnable on your own GPU.Our show today was one of the shortest yet, as I had to drop early to prepare for Burning Man 🔥🕺 Speaking of which, Wolfram and the team will host the next episode! Ok, let's dive in! DeepSeek V3.1: a faster hybrid that thinks less, scores more (X, HF)DeepSeek does this thing where they let a base artifact “leak” onto Hugging Face, and the rumor mill goes into overdrive. Then, hours before we went live, the full V3.1 model card and an instruct variant dropped. The headline: it’s a hybrid reasoner that combines the strengths of their V3 (fast, non-thinking) and R1 (deep, RL-trained thinking), and on many tasks it hits R1-level scores with fewer thinking tokens. In human terms: you get similar or better quality, faster.A few things I want to call out from the release and early testing:* Hybrid reasoning mode done right. The model can plan with thinking tokens and then switch to non-thinking execution, so you don’t have to orchestrate two separate models. This alone simplifies agent frameworks: plan with thinking on, execute with thinking off.* Thinking efficiency is real. DeepSeek shows curves where V3.1 reaches or surpasses R1 with significantly fewer thinking tokens. On AIME’25, for example, R1 clocks 87.5% with ~22k thinking tokens; V3.1 hits ~88.4 with ~15k. On GPQA Diamond, V3.1 basically matches R1 with roughly half the thinking budget.* Tool-use and search-agent improvements. V3.1 puts tool calls inside the thinking process, instead of doing a monologue and only then calling tools. That’s the pattern you want for multi-turn research agents that iteratively query the web or your internal search.* Long-context training was scaled up hard. DeepSeek says they increased the 32K extension phase to ~630B tokens, and the 128K phase to ~209B tokens. That’s a big bet on long-context quality at train time, not just inference-time RoPE tricks. The config shows a max position in the 160K range, with folks consistently running it in the 128K class.* Benchmarks show the coding and terminal agent work got a big push. TerminalBench jumps from a painful 5.7 (R1) to 31 with V3.1. Codeforces ratings are up. On SweBench Verified (non-thinking), V3.1 posts 66 vs R1’s ~44. And you feel it: it’s faster to “get to it” without noodling forever.* API parity you’ll actually use. Their API now supports the Anthropic-style interface as well, which means a bunch of editor integrations “just work” with minimal glue. If you’re in a Claude-first workflow, you won’t have to rewire the world to try V3.1.* License and availability. This release is MIT-licensed, and you can grab the base model on Hugging Face. If you prefer hosted, keep an eye on our inference—we’re working to get V3.1 live so you can benchmark without burning your weekend assembling a serving stack.Hugging Face: https://huggingface.co/deepseek-ai/DeepSeek-V3.1-BaseQuick personal note: I’m seeing a lot of small, pragmatic improvements add up here. If you’re building agents, the hybrid mode plus tighter tool integration is a gift. DeepSeek V3.1 is going to be deployed to W&B Inference service soon! Take a look here to see when it's ready wandb.me/inference ByteDance Seed-OSS 36B: Apache-2.0, 512K context, and a “thinking budget” knob (X, HF, Github)I didn’t see much chatter about this one, which is a shame because this seems like a serious release. ByteDance’s Seed team open-sourced a trio of 36B dense models—two Base variants (with and without synthetic data) and an Instruct model—under Apache-2.0, trained on 12T tokens and built for long-context and agentic use. The context window is a native half-million tokens, and they include a “thinking budget” control you can set in 512-token increments so you can trade depth for speed.They report strong general performance, long-context RULER scores, and solid code/math numbers for a sub-40B model, with the Instruct variant posting very competitive MMLU/MMLU-Pro and LiveCodeBench results. The architecture is a straightforward dense stack (not MoE), and the models ship with Transformers/vLLM support and 4/8-bit quantization ready to go. If you’ve been hunting for a commercial-friendly, long-context 30-something‑B with an explicit reasoning-control dial, this should be on your shortlist.A neat detail for the training nerds: two Base releases—one trained with synthetic data, one without—make for a rare apples-to-apples study in how synthetic data shapes base capability. Also worth noting: they previously shipped a Seed-Prover specialized for Lean; it looks like the team is interested in tight domain models and generalists.NVIDIA Nemotron Nano 9B V2: mixed architecture, open data, and long-context throughput (X, Blog, HF, Dataset, Try It) NVIDIA shipped a fully open release of Nemotron Nano 9B V2—base, base-before-alignment/pruning, and a realigned reasoning model—and, crucially, they published most of the pretraining dataset details (~6.6T tokens across premium web, math, code, and SFT). That level of data transparency is rare and makes this a great base for fine-tuners who want reproducibility.Under the hood, this is a mixed Mamba+Transformer architecture. NVIDIA is claiming up to 6x higher throughput versus a pure-Transformer peer (they compare to Qwen3-8B) and specifically highlight that they pruned a 12B down to 9B while preserving quality. They also note a single A10 can handle 128K context after compression and distillation passes, which is the kind of practical systems work that matters when you’re running fleets.A couple of caveats. The license is NVIDIA Open Model License—not Apache-2.0—so read it; it includes restrictions around illegal surveillance and safety bypasses and has revocation clauses. Personally, I appreciate the data openness and the long-context engineering; as always, just make sure the license fits your use case.If you’re into longer-context math/coding with small models, the numbers (AIME’25, RULER-128K, GPQA) are impressive for 9B. And if you fine-tune: the availability of both pruned and pre-pruned bases plus the dataset recipe is a rare treat.Cohere’s Command-A Reasoning: dense, multilingual, and research-only licensing (X, Blog, HF)Cohere Dropped a new reasoning model focused on enterprise deployment patterns. It’s dense 111B model, supports a 256K context, and includes very strong multilingual coverage (23 languages is what they called out). What caught my eye: on the BFCL (Berkeley Function-Calling Leaderboard) they show 70%—above DeepSeek R1’s ~63% and GPT-OSS’s ~61%—and they plot the now-familiar test-time compute curves where more thinking tokens yield higher scores.This release uses Cohere’s non-commercial research license; if you want commercial usage you’ll need to go through them. That said, for teams who need privately deployable, on-prem reasoning and can work under a research license for prototyping, it’s a serious option. A meta observation from the show: there’s accumulating evidence that more active parameters help multi-hop tool-use chains compared to very sparse MoE at similar effective capacity. This model nudges in that direction.Desktop agents leap: ComputerRL hits 48% on OSWorld (Paper)A new framework dubbed ComputerRL from Z.ai and folks at Tsingua Uni, unified API calls with GUI actions and scaled RL across fleets of virtual desktops, posting a 48.1% success rate on OSWorld versus ~12% for earlier open models. The training system spins up thousands of qemu-in-docker VMs via gRPC; the learning loop alternates RL with supervised fine-tuning and uses a clean step-level binary reward to simplify credit assignment. If you care about practical desktop automation across Ubuntu/Windows/macOS, this is a big jump.IBM + NASA’s Surya: open model for solar weather (HF)Scientists get some love: IBM and NASA open-sourced Surya, a transformer trained on nine years of multi-instrument observations (nearly 200 TB) to forecast solar dynamics and space weather—the stuff that can knock satellites and power grids sideways. It’s on Hugging Face, it’s actually runnable, and it’s a fantastic example of open models delivering real-world scientific utility.Smaller but notable: InternLM and OpenCUA, plus Intel’s quantsTwo quick flags from the “worth your time” pile. InternLM shipped S1 Mini, an 8B vision+language model (ViT on top) that’s multimodal and lightweight; if you need on-device omni-ish behavior on a laptop or tablet, give it a look. And OpenCUA 32B (Qwen-based) is a specialized computer-usage agent with strong scores; if you’re building automations that need native OS control, it’s worth benchmarking.Also, if you’re running 4-bit: the Intel quantization work is excellent right now. Their 4-bit quants have been extremely high precision in my testing, especially for large coders and reasoners like DeepSeek V3.1. It’s an easy win if you’re trying to squeeze a 30B+ onto a workstation without hemorrhaging quality.Big-co updates and platform shiftsSonic appears in Cursor and ClineIf you open Cursor or fire up Cline, you may see a new “Sonic” model toggle. It’s labeled as a reasoning model and, when you poke the function-calling internals, the call paths include “xai/…” strings. Folks report it’s fast and solid for coding. No official docs yet, but I’d be surprised if this isn’t Grok Code in pre-release clothes.Agents.md: one file to rule your agentsAgentic dev stacks have multiplied config files like gremlins: Cursor’s rules.json, Windsurf’s prompts, MCP server manifests, tool schemas, install scripts… and every tool wants a different filename and format. OpenAI’s Agents.md is a strong attempt at standardization. It’s just Markdown at repo root that says, “here’s how to set up, build, test, and run this project,” plus any agent-specific caveats. Tools then auto-detect and follow your instructions instead of guessing.It’s already supported by OpenAI Codex, Amp, Jules, Cursor, RooCode, and more, with tens of thousands of public repos adopting the pattern. In monorepos, the nearest Agents.md wins, so you can override at the package level. And human chat instructions still override the file’s guidance, which is the right default.GPT‑5 context truncation in the web UI (reports)There’s been a wave of reports that GPT‑5 in the web UI is truncating long prompts even when you’re under the documented context limit. The folks at Repo Prompt reproduced this multiple times and got confirmation from OpenAI that it’s a bug (not a deliberate nerf). If you saw GPT‑5 suddenly forget the bottom half of your carefully structured system prompt in the web app, this likely explains it. The API doesn’t seem affected. Fingers crossed for a quick fix—GPT‑5 is still the best model I’ve used for 300k‑token “read the whole repo and propose a plan” tasks.Image and 3D: Nano Banana and Qwen’s open image editorNano Banana: 3D-consistent scene editing from thin airA stealth model nicknamed “Nano Banana” surfaced in a web demo and started doing the kind of edits you’d normally expect from a 3D suite with a modeler at the controls. Take two photos—your living room and a product shot—and it composites the object into the scene with shockingly consistent lighting and geometry. Ask for a 3D mesh “five inches off the skin,” and the mesh really does offset. Ask for a new camera angle on a single still, and it renders the scene from above with plausible structure. People have been calling it a game-changer and, for once, it doesn’t feel like hyperbole.There’s a strong whiff of 3D world modeling under the hood—some volumetric representation or neural field that enables true view synthesis—and Logan Kilpatrick posted a banana emoji that set speculation on fire. We’ll see where it lands and under what license, but for now the demo has been doing the rounds and it is… wow.If you’re wondering where to try it: LMarena is the currently only way to give it a try, but it's supossedly dropping soon! Qwen Image Edit (20B): fully open and already practical (X, HF)Qwen shipped a 20B image-editing model layered on their existing vision stack, and it’s properly open (permissive license) with strong bilingual text editing in Chinese and English. It handles high-level semantic edits (pose adjustments, rotations, style/IP creation) and low-level touch-ups (add/remove/insert). You can swap objects, expand aspect ratios, keep character identity consistent across panels, and do clean style transfer. It runs locally if you’ve got a decent GPU.What I appreciate here is the precision. Product placement tasks like “put this book in this person’s hand at this angle,” or “make the shoes match the dress” come out with the kind of control that used to require hand masking and a dozen passes. And yes, the capybara mascot is back in the release materials, which made my day! 👏If Nano Banana is the closed-world demo of what’s “beyond SOTA,” Qwen Image Edit is the open tool you can actually ship with today.This week’s buzz from Weights & BiasesTwo quick updates from our side. First, we’re working to bring DeepSeek V3.1 to our inference as fast as we can so you can run serious benchmarks without fussing over serving stacks. Keep an eye on our channels; we’ll shout when it’s live and we’ll have some credits for early feedback.Second, our cofounder Chris Van Pelt released Catnip, a containerized multi‑agent coding workspace that runs multiple Claude Code sessions (or other agents) in isolated sandboxes, each with its own context and notification stream. If you’ve been juggling parallel coding agents that step on each other’s toes, this is catnip indeed.Catnip Github: https://github.com/wandb/catnipClosing thoughtsA year ago, “thinking tokens” weren't even a curiosity; We only got the first whiff of "reasoning" back in September, and now we’re watching hybrid models that do more with less thinking, tool calls woven inside the reasoning loop, and long-context training regimes scaled up by an order of magnitude. The agent stack is maturing fast—desktop RL is finally clearing real tasks; editor ecosystems are converging on a single config file; and even the stealth drops are clearly building toward world-model‑aware editing and control.If you only try two things this week: run DeepSeek V3.1 in both modes (planning with thinking on, execution with thinking off) and throw a complex multi-step tool workflow at it; then take Qwen Image Edit for a spin on a real product-placement or character-consistency task. You’ll feel the future in your hands.I’m off to the desert next week for a bit (no internet where I’m going), but Wolfram and the crew will keep the ship on course. If you’re at Burning Man, DM me—would love to say hi out there. As always, thank you for tuning in and nerding out with us every week.— AlexTL;DR and show notesThursdAI - Aug 21, 2025 - TL;DRThursdAI - Aug 21, 2024 - TL;DRTL;DR of all topics covered:* Hosts and Guests* Alex Volkov - AI Evangelist & Weights & Biases* Co Hosts - @WolframRvnwlf, @yampeleg, @nisten, @ldjconfirmed* Open Source LLMs // Papers* ByteDance Seed-OSS - 36B open-source LLM family (X, HF, GitHub)* DeepSeek V3.1 - Updated Hybrid model (HF)* Cohere CMD-a Reasoning - (X, Blog, HF)* Zai/Tsinghua ComputerRL - Framework for desktop agents (X, Paper, Benchmark)* IBM & NASA Surya - Solar weather prediction (HF)* NVIDIA Nemotron Nano 9B V2 - (X, Blog, HF, Dataset, Try It) * Alibaba Quark Med* Big CO LLMs + APIs* Sonic Stealth Model - Likely Grok Code* OpenAI agents.md - Unified agent files (agents.md)* GPT-5 Nerf* AI Art & Diffusion & 3D* Nano Banana - Image model (rumored Google)* Qwen-Image-Edit - 20B Image editing (X, HF)* This weeks Buzz* Catnip - Containerized AI agent runner (GitHub) This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit sub.thursdai.news/subscribe

Hey everyone, Alex here 👋Last week, I tried to test GPT-5 and got really surprisingly bad results, but it turns out, as you'll see below, it's partly because they had a bug in the router, and partly because ... well, the router itself! See below for an introduction, written by GPT-5, it's actually not bad?Last week was a whirlwind. We live‑streamed GPT‑5’s “birthday,” ran long, and then promptly spent the next seven days poking every corner of the new router‑driven universe.This week looked quieter on the surface, but it actually delivered a ton: two open‑source world models you can drive in real time, a lean vision‑language model built for edge devices, a 4B local search assistant that tops Perplexity Pro on SimpleQA, a base model “extraction” from GPT‑OSS that reverses alignment, fresh memory features landing across the big labs, and a practical prompting guide to unlock GPT‑5’s reasoning reliably.We also had Alan Dao join to talk about Jan‑v1 and what it takes to train a small model that consistently finds the right answers on the open web—locally.Not bad eh? Much better than last time 👏 Ok let's dive in, a lot to talk about in this "chill" AI week (show notes at the end as always) first open source, and then GPT-5 reactions and then... world models!00:00 Introduction and Welcome00:33 Host Introductions and Health Updates01:26 Recap of Last Week's AI News01:46 Discussion on GPT-5 and Prompt Techniques03:03 World Models and Genie 303:28 Interview with Alan Dow from Jan04:59 Open Source AI Releases06:55 Big Companies and APIs10:14 New Features and Tools14:09 Liquid Vision Language Model26:18 Focusing on the Task at Hand26:18 Reinforcement Learning and Reward Functions26:35 Offline AI and Privacy27:13 Web Retrieval and API Integration30:34 Breaking News: New AI Models30:41 Google's New Model: Gemma 333:53 Meta's Dino E3: Advancements in Computer Vision38:50 Open Source Model Updates45:56 Weights & Biases: New Features and Updates51:32 GPT-5: A Week in Review55:12 Community Outcry Over AI Model Changes56:06 OpenAI's Response to User Feedback56:38 Emotional Attachment to AI Models57:52 GPT-5's Performance in Coding and Writing59:55 Challenges with GPT-5's Custom Instructions01:01:45 New Prompting Techniques for GPT-501:04:10 Evaluating GPT-5's Reasoning Capabilities01:20:01 Open Source World Models and Video Generation01:27:54 Conclusion and Future ExpectationsOpen Source AIWe've had quite a lot of Open Source this week on the show, including a breaking news from the Gemma team!Liquid AI's drops LFM2-VL (X, blog, HF)Let's kick things off with our friends at Liquid AI who released LFM2-VL - their new vision-language models coming in at a tiny 440M and 1.6B parameters.Liquid folks continue to surprise with speedy, mobile device ready models, that run 2X faster vs top VLM peers. With a native 512x512 resolution (which breaks any image size into 512 smart tiles) and an OCRBench of 74, this tiny model beats SmolVLM2 while being half the size. We've chatted with Maxime from liquid about LFM2 back in july, and it's great to see they are making them multimodal as well with the same efficiency gains!Zhipu (z.ai) unleashes GLM-4.5V - 106B VLM (X, Hugging Face)In another "previous good model that now has eyes" release, the fine folks from Zhipu continued training thier recently released (and excelled) GLM 4.5-air with a vision encoder, resulting in probably one of the top vision models in the open source!It's an MoE with only 12B active parameters (106B total) and gets SOTA across 42 public vision-language benches + has a "thinking mode" that reasons about what it sees.Given that GLM-4.5Air is really a strong model, this is de fact the best visual intelligence in the open source, able to rebuild websites from a picture for example and identify statues and locations!Jan V1 - a tiny (4B) local search assistant QwenFinetune (X, Hugging Face)This one release got a lot of attention, as the folks at Menlo Research (Alan Dao who came to chat with us about Jan on the pod today) released an Apache 2 finetune of Qwen3-4B-thinking, that's focused on SimpleQA.They showed that their tiny model is beating Perplexity Pro on SimpleQA.Alan told us on the pod that Jan (the open source Jan app) is born to be an open source alternative to searching with local models!The trick is, you have to enable some source of search data (Exa, Serper, Tavily) via MCP and then enable tools in Jan, and then.. you have a tiny, completely local perplexity clone with a 4B model!Google drops Gemma 3 270M (blog)In some #breakingNews, Google open sourced a tiny (270M) parameters, "good at instruction following" Gemma variant. This joins models like SmolLM and LFM2 in the "smol models" arena, being only 300MB, you can run this.. on a toaster. This one apparently also fine-tunes very well while being very energy efficient!Big Companies (AKA OpenAI corner this past 2 weeks)Ok ok, we're finally here, a week with GPT-5! After watching the live stream and getting access to GPT-5, my first reactions were not great. Apparently, so have other peoples, and many folks outcried and complained about the model, some even yelling "AGI is cancelled".What apparently happened is (and since, been fixed by OpenAI) is that GPT-5 wasn't just a model that launched, it was a "smart" router between a few models, and not only did they have a routing bug, the basic GPT-5 model, the one without thinking, is... not great.But the thinking GPT-5, the one that the router refused to send me to, is really good (as confirmed independently by multiple evals at this point)For one, it's the most accurate function calling model on OpenRouterIt's also one of the best on this new FormulaOne benchmark that was just launchedYou're prompting it wrong!Apparently, not only is GPT-5 more intelligent, it's also significantly "surgical" in instruction following, and so, for many folks, just replacing GPT-5 into their tools or prompts didn't just "work", as this model, more than before, is sensitive to conflicting things in the prompt.OpenAI has released a guide for prompting the model, mostly aimed at developers (as users shouldn't be learning to prompt as models get more intelligent) + they also released a prompt optimizer! Just dump your long and complex prompts in there, and you'll get an updated prompt with explanations of why they changed what they changed!Model Picker (and legacy models) are back!So, OpenAI tried and super quickly reversed course on removing the "model picker". At first, it was only GPT-5 there, but many people complained about the abrupt removal of 4o, their .. favorite models. At first, OpenAI added back the models via a hidden setting, and later, they have added 4o back to everyone by default, while increasing the reasoning quota to 3000 messages per week!Generally, my thoughts are, if you've tried GPT-5 and weren't impressed, give it another go! (especially now that it's connected to Gmail in chats!)Other notable Big Company updatesIn other news, Claude has extended the context window of Sonnet to 1M in the API, and apparently both Claude and Gemini have been adding memory features!Grok video has been catching up and is now free for a while to all usersThis Week's Buzz: Weave DX improvementsQuick update from my day job at Weights & Biases - we've rolled out some quality-of-life improvements to Weave, our LLM observability platform. We now have a unified assets tab where you can manage all your prompts, models, and datasets with full versioning support.Prompts are being version tracked, so if you use that GPT-5 prompt optimizer, we'll store all the previous revisions for ya!The coolest addition? Threads! Perfect for tracking agent executions or grouping related API calls. You just add a thread_id to your traces and Weave handles the rest. If you're building AI applications and not tracking everything, you're flying blind - give Weave a try at wandb.me/weave!World models are getting... open sourced!I still think that Google's Genie-3 release from last week was maybe the more important one, though we didn't really get to play with it yet!And while getting excited by world models, I was thinking that it's goig to take a while for Open Source to catch up. But this week, not 1, but two world models were open sourced, making me think that we'll get to generated worlds quicker than I expected and the race has begun!Skywork's Matrix-Game 2.0 (project, HF)Matrix-game 2 is a auto-regressive diffusion model, that was trained on 1200 hours of Unreal Engine and GTA-5 environments that runs at 25 frames per second!It works by creating an "action injection module" that embeds the mouse/keyboard inputs into the generation, enabling frame-level controls.Hunyuan open-sources GameCraft for real-time, high-dynamic game video generation (X, Hugging Face)Two world-models (well, game models) in the same week? Tencent (who had Hunyuan video before) have trained a game engine on top of their excellent HY-video and have shown the same examples, of building a full world based on a few images.Their pipeline trained on 1M game play clips from AAA titles, and they also map W/A/S/D and mouse signals into continuous camera/action embeddings, allowing for control and angle creation.The cool thing? A quantized 13B version supposedly can run on a RTX 4090!Funnily, they already had Matrix-Game (the one that came out a few days before) benchmarked and beat on the release today!Genie 3 is not messing aroundWhile all the open source is impressive, I was… absolutely blown away by this video from an artist who got the Genie3 team to extend a video of his. Just look at the collision of the plane with the sphere, out of nowhere, Genie3 adds a shadow, and then collision mechanics, the plane bouncing off, and even the jet trails subside and then resume! It really really is crazy to imagine that no prompting was given and the model just.. knew how to do this!Phew, that was a lot! Much more as always on the actual show, despite it being a "quiet" week by summer of 2025 standards, there was a LOT of open source releases and GPT-5 situation shows that even OpenAI can stumble on new tech releases!Keep the feedback coming - find me on Twitter/X at @altryne or email the show. And remember, if you want to catch all the technical details and demos, the video version on YouTube has everything the podcast can't show.Until next week, keep tinkering, keep questioning, and keep pushing the boundaries of what's possible with AI!See you next ThursdAI 🚀TL;DR - All Topics CoveredHosts and Guests* Alex Volkov - AI Evangelist @ Weights & Biases (@altryne)* Co-hosts: @WolframRvnwlf, @yampeleg, @nisten, @ldjconfirmed* Guest: Alan Dao from Jan (@jandotai)Open Source LLMs* Liquid AI LFM2-VL: 440M & 1.6B vision-language models with 2x GPU speedup (Blog, HF)* Jan V1: 4B parameter search assistant beating Perplexity Pro on SimpleQA (X, HF)* GPT-OSS Base: Reverse-engineered base model from Jack Morris (X Thread)* Gemma 3: Google's 270M parameter model with strong instruction following (HF)* Meta Dino v3: Self-supervised vision model for segmentation (Blog)Big Companies & APIs* Mistral Medium 3.1: New model on Mistral platform* Gemini & Claude: Added memory/personalization features* GPT-5 Updates: Router fixes, model selector returned, prompting guide released (Guide, Optimizer)* Claude Sonnet 4: 1M token context window (Announcement)This Week's Buzz* Weave updates: Unified assets tab and threads for agent tracking* New features for LLM observability and evaluationVision & Video* Hunyuan Large Vision: 1B vision encoder + 389B MoE language model (Project)* GLM-4.5V: 106B open source VLM from Zhipu AI (X, HF)World Models* Matrix Game 2.0: Real-time interactive world model, 25fps generation (Project, HF)* Hunyuan GameCraft: Game video generation with physics understanding (X, HF)Tools* Grok: Now includes video generation for all users* Jan Desktop App: Local AI with MCP support and search capabilities This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit sub.thursdai.news/subscribe

Hey folks 👋 Alex here, writing to you, from a makeshift recording studio in an Eastern European hookah bar, where I spent the last 7 hours. Why you ask? Well, when GPT-5 drops, the same week as OpenAI dropping the long awaited OSS models + Google is shipping perfect memory World Models (Genie 3) and tons of other AI drops, well I just couldn't stay away from the stream.Vacation or not, ThursdAI is keeping you up to date (for 32 months straight, which is also the time since the original GPT-4 release which gave this show its name!)So, what did we have today on the stream? Well, we started as usual, talking about the AI releases of the week, as if OpenAI dropping OSS models (apache 2) 120B and 20B is "usual". We then covered incredible releases like Google's World model Genie3 (more on this next week!) and Qwen-image + a few small Qwens.We then were VERY excited to tune in, and watch the (very long) announcement stream from OpenAI, in which they spent an hour to tell us about GPT-5.This was our longest stream by far (3.5 hours, 1hr was just OpenAI live stream) and I'm putting this here mostly unedited, but chapters are up so feel free to skip to the parts that are interesting to you the most.00:00 Introduction and Special Guests00:56 Twitter Space and Live Streaming Plans02:12 Open Source AI Models Overview03:44 Qwen and Other New AI Models08:59 Community Interaction and Comments10:01 Technical Deep Dive into AI Models25:06 OpenAI's New Releases and Benchmarks38:49 Expectations and Use Cases for AI Models40:03 Tool Use vs. Deep Knowledge in AI41:02 Evaluating GPT OSS and OpenAI Critique42:29 Historical and Medical Knowledge in AI51:16 Opus 4.1 and Coding Models55:38 Google's Genie 3: A New World Model01:00:43 Kitten TTS: A Lightweight Text-to-Speech Model01:02:07 11 Labs' Music Generation AI01:08:51 OpenAI's GPT-5 Launch Event01:24:33 Building a French Learning Web App01:26:22 Exploring the Web App Features01:29:19 Introducing Enhanced Voice Features01:30:02 Voice Model Demonstrations01:32:32 Personalizing Chat GPT01:33:23 Memory and Scheduling Features01:35:06 Safety and Training Enhancements01:39:17 Health Applications of GPT-501:45:07 Coding with GPT-501:46:57 Advanced Coding Capabilities01:52:59 Real-World Coding Demonstrations02:10:26 Enterprise Applications of GPT-502:11:49 Amgen's Use of GPT-5 in Drug Design02:12:09 BBVA's Financial Analysis with GPT-502:12:33 Healthcare Applications of GPT-502:12:52 Government Adoption of GPT-502:13:22 Pricing and Availability of GPT-502:13:51 Closing Remarks by Chief Scientist Yakob02:16:03 Live Reactions and Discussions02:16:41 Technical Demonstrations and Comparisons02:33:53 Healthcare and Scientific Advancements with GPT-502:47:09 Final Thoughts and Wrap-Up---My first reactions to GPT-5Look, I gotta keep it real with you, my first gut reaction was, hey, I'm on vacation, I don't have time to edit and write the newsletter (EU timezone) so let's see how ChatGPT-5 handles this task. After all, OpenAI has removed all other models from the dropdown, it's all GPT-5 now. (pricing from the incredible writeup by Simon Willison available here)And to tell you the truth, I was really disappointed! GPT seems to be incredible at coding benchmarks, with 400K tokens and incredible pricing (just $1.25/$10 compared to Opus $15/$75) this model, per the many friends who got to test it early, is a beast at coding! Readily beating opus on affordability per token, switching from thinking to less thinking when it needs to, it definitely seems like a great improvement for coding and agentic tasks.But for my, very much honed prompt of "hey, help me with ThursdAI drafts, here's previous drafts that I wrote myself, mimic my tone" it failed.. spectacularly!Here's just a funny example, after me replying that it did a bad job:It literally wrote "I'm Alex, I build the mind, not the vibe" 🤦‍♂️ What.. the actual...For comparison, here's o3, with the same prompt, with a fairly true to tone draft:High taste testers take on GPT-5But hey, I have tons of previous speakers in our group chats, and many of them who got early access (I didn't... OpenAI, I can be trusted lol) rave about this model. They are saying that this is a huge jump in intelligence.Folks like Dr Derya Unutmaz, who jumped on the live show and described how GPT5 does incredible things with less hallucinations, folks like Swyx from Latent.Space who had early access and even got invited to give first reactions to the OpenAI office, and Pietro Schirano who also showed up in an OpenAI video.So definitely, definitely check out their vibes, as we all try to wrap our heads around this new intelligence king we got!Other GPT5 updatesOpenAI definitely cooked, don't get me wrong, with this model plugging into everything else in their platform like memory, voice (that was upgraded and works in custom GPTs now, yay!), canvas and study mode, this will definitely be an upgrade for many folks using the models.They have now also opened access to GPT-5 to free users, just in time for schools to reopen, including a very interesting Quiz mode (that just showed up for me without asking for it), and connection to Gmail, all those will now work with GPT5.It now has 400K context, way less hallucinations but fewer refusals also, and the developer upgrades like a new verbosity setting and a new "minimal" reasoning setting are all very welcome!OpenAI finally launches gpt-oss (120B / 20B) apache 2 licensed models (model card, HF)It was really funny, on the stream Nisten talked about the open source models OpenAI dropped, and said "when we covered it last week", while it was just two days ago! It really does feel like this world is moving really fast.OpenAI's long promised open source models are here, and they got a fairly mixed bag of reviews from folks. Many folks are celebrating that the western world is now back in the game, releasing incredible local models, with an open license!Though, after the initial excitement, the vibes are split on these models. Folks are saying that maybe these were trained with only synthetic data, because, like Phi, they seem to be very good at benchmarks, and on the specific tasks they were optimized for (code, math) but really bad at creative writing (Sam Paech from EQBench was not impressed), they are also not multilingual, though OpenAI did release a cookbook on finetuning with HuggingFace!Overall, these models are trained for agentic workflows—supporting function calling, web search, Python execution, configurable reasoning effort, and full raw chain-of-thought access, which we will never get from GPT5.I particularly love the new approach, where a reasoning effort can be defined directly via the system prompt, by just adding "reasoning: high" to the system prompt, this model will reason for way longer! Can't wait to get back and bench these and share with you.Overall, the fine-tuning and open source community is split for now, but it's been only a few days, so we'll keep you up to date on how well these models land, regardless, this was a historic week for OpenAI!Speaking of open models, did you have a chance to try our W&B Inference? The team worked hard to bring these new models to you in record time and incredible pricing (just $.05 for 20B and $.15 for 120B!), these models are definitely worth giving a try!Plus, if you comment "OSS Power" on our announcement post, we'll likely give you a few credits to try it out and let us know what you think!World models "holy crap" moment - Google Genie3The other very important release this week was.... not a release at all, but an announcement from Deepmind, with Genie3.This World Model takes a single image or text prompt and creates a fully interactive, controllable 3D environment that runs in real-time at 24fps. An environment you as a user can control, walk (or fly) in, move around the camera view. It's really mindblowing stuff.We've covered world models like Mirage on previous episodes, but what Google released is a MAJOR step up in coherency, temporal consistency and just overall quality!The key breakthrough here is consistency and memory. In one demo, a user could "paint" a virtual wall, turn away, and when they turned back, the paint was still there. This is a massive step towards generalist agents that can train, plan, and reason in entirely simulated worlds, with huge implications for robotics and gaming.We’re hoping to have the Genie 3 team on the show next week to dive even deeper into this incredible technology!!Other AI news this weekThis week, the "other" news could have filled a full show 2 years ago, we got Qwen keeping the third week of releases with 2 new tiny models + a new diffusion model called Qwen-image (Blog, HF)Anthropic decided to pre-empt the GPT5 release, and upgraded Opus 4 and gave us Opus 4.1 with a slight bump in specs.ElevenLabs released a music API called ElevenMusic, which sounds very very good (this on top of last weeks Riffusion + Producer.ai news, that I'm still raving about)Also in voice an audio, a SUPER TINY TTS model called KittenTTS released, with just 15M parameters and a model that's 25MB, it's surprisingly decent at generating voice (X)And to cap it off with breaking news, the Cursor team, who showed up on the OpenAI stream today (marking quite the change in direction from OpenAI + Windsurf previous friendship), dropped their own CLI version of cursor, reminiscent of Claude Code!PHEW, wow ok this was a LOT to process. Not only did we tune in for the full GPT-5 release, we did a live stream when gpt-oss dropped as well.On a personal note, I was very humbled when Sam Altman said it was 32 months since GPT-4 release, because it means this was 32 months of ThursdAI, as many of you know, we started live streaming on March 13, 2023, when GPT-4 was released.I'm very proud of the incredible community we've built (50K views total across all streams this week!), the incredible co-hosts I have, who step up when I'm on vacation and the awesome guests we have on the show, to keep you up to date every week!So, a little favor to ask, if you find our content valuable, entertaining, the best way to support this pod is upgrade to a paid sub, and share ThursdAI with a friend or two! 👏 See you next week 🫡 This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit sub.thursdai.news/subscribe

This is a free preview of a paid episode. To hear more, visit sub.thursdai.newsWoohoo, we're almost done with July (my favorite month) and the Open Source AI decided to go out with some fireworks 🎉Hey everyone, Alex here, writing this without my own personal superintelligence (more: later) and this week has been VERY BUSY with many new open source releases.Just 1 hour before the show we already had 4 breaking news releases, a tiny Qwen3-coder, Cohere and StepFun both dropped multimodal SOTAs and our friends from Krea dropped a combined model with BFL called Flux[Krea] 👏 This is on top of a very very busy week, with Runway adding conversation to their video model Alpha, Zucks' superintelligence vision and a new SOTA open video model Wan 2.2. So let's dive straight into this (as always, all show notes and links are in the end) ThursdAI - Recaps of the most high signal AI weekly spaces is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.Open Source LLMs & VLMs Tons of new stuff here, I'll try to be brief but each one of these releases deserves a deeper dive for sure. Alibaba is on 🔥 with 3 new Qwen models this weekYes, this is very similar to last week, where they have also dropped 3 new SOTA models in a week, but, these are additional ones. It seems that someone in Alibaba figured out that after splitting away from the hybrid models, they can now release each model separately and get a lot of attention per model! Here's the timeline: * Friday (just after our show): Qwen3-235B-Thinking-2507 drops (235B total, 22B active, HF) * Tuesday: Qwen3-30B-Thinking-2507 (30B total, 3B active, HF)* Today: Qwen3-Coder-Flash-2507 lands (30B total, 3B active for coding, HF)Lets start with the SOTA reasoner, the 235B(A22B)-2507 is absolutely the best reasoner among the open source models.We've put the model on our inference service (at crazy prices $.10/$.10) and it's performing absolutely incredible on reasoning tasks. It also jumped to the top OSS model on Artificial Analysis scores, EQBench, Long Context and more evals. It a really really good reasoning model! Smaller Qwens for local useJust a week ago, we've asked Junyang on our show, about smaller models that folks can run on their devices, and he avoided by saying "we're focusing on the larger models" and this week, they delivered not 1 but 2 smaller versions of the bigger models (perfect for Speculative Decoding if you can host the larger ones that is) The most interesting one is the Qwen3-Coder-flash, which came out today, with very very impressive stats - and the ability to run locally with almost 80 tok/s on a macbook! So for the last two weeks, we now have 3 Qwens (Instruct, Thinking, Coder) and 2 sizes for each (all three have a 30B/A3B version now for local use) 👏Z.ai GLM and StepFun Step3 As we've said previously, Chinese companies completely dominate the open source AI field right now, and this week as saw yet another crazy testament to how stark the difference is! We've seen a rebranded Zhipu (Z.ai previously THUDM) release their new GLM 4.5 - which gives Qwen3-thinking a run for it's money. Not quite at that level, but definitely very close. I personally didn't love the release esthetics, showing a blended eval score, which nobody can replicate feels a bit off. We also talked about how StepFun has stepped in (sorry for the pun) with a new SOTA in multimodality, called Step3. It's a 321B MoE (with a huge 38B active param count) that achieves very significant multi modal scores (The benchmarks look incredible: 74% on MMMU, 64% on MathVision) Big Companies APIs & LLMsWell, we were definitely thinking we'll get GPT-5 or the Open Source AI model from OpenAI this week, but alas, the tea leaves readers were misled (or were being misleading). We 100% know that gpt-5 is coming as multiple screenshots were blurred and then deleted showing companies already testing it. But it looks like August is going to be even hotter than July, with multiple sightings of anonymous testing models on Web Dev arena, like Zenith, Summit, Lobster and a new mystery model on OpenRouter called Zenith - that some claim are the different thinking modes of GPT-5 and the open source model? Zuck shares vision for personalized superintelligence (Meta)In a very "Nat Fridman" like post, Mark Zuckerberg finally shared the vision behind his latest push to assemble the most cracked AI engineers.In his vision, Meta is the right place to provide each one with personalized superintelligence, enhancing individual abilities with user agency according to their own values. (as opposed to a centralized model, which feels like his shot across the bow for the other frontier labs) A few highlights: Zuck leans heavily into the rise of personal devices on top of which humans will interact with this superintelligence, including AR glasses and a departure from a complete "let's open source everything" dogman of the past, now there will be a more deliberate considerations of what to open source. This Week's Buzz: Putting Open Source to Work with W&BWith all these incredible new models, the biggest question is: how can you actually use them? I'm incredibly proud to say that the team at Weights & Biases had all three of the big new Qwen models—Thinking, Instruct, and Coder—live on W&B Inference on day one (link)And our pricing is just unbeatable. Wolfram did a benchmark run that would have cost him $150 using Claude Opus. On W&B Inference with the Qwen3-Thinking model, it cost him 22 cents. That's not a typo. It's a game-changer for developers and researchers.To make it even easier, a listener of the show, Olaf Geibig, posted a fantastic tutorial on how you can use our free credits and W&B Inference to power tools like Claude Code and VS Code using LiteLLM. It takes less than five minutes to set up and gives you access to state-of-the-art models for pennies. All you need to do is add this config to vllm and run claude (or vscode) through it! Give our inference service a try here and follow our main account @weights_biases a follow as we often drop ways to get additional free credits when new models releaseVision & Video modelsWan2.2: Open-Source MoE Video Generation Model Launches (X, HF)This is likely the best open source video model, but definitely the first MoE video model! It came out with text2video, image2video and a combined version. With 5 second 720p videos, that can even be generator at home on a single 4090, this is definitely a step up in the quality of video models that are fully open source. Runway changes the game again - Gen-3 Aleph model for AI video editing / transformation (X, X)Look, there's simply no denying this, AI video has had an incredible year, from open source like Wan, to proprietary models with sounds like VEO3. And it's not surprising that we're seeing this trend, but it's definitely very exciting when we see an approach like Runway has, to editing. This adds a chat to the model, and your ability to edit.. anything in the scene. Remove / Add people and environmental effects, see the same scene from a different angle and a lot more! Expect personalized entertainment very soon! AI Art & Diffusion & 3DFLUX.1 Krea [dev] launches as a state-of-the-art open-weights text-to-image model (X, HuggingFace)Black Forest Labs teamed with Krea AI for Flux.1 Krea [dev], an open-weights text-to-image model ditching the "AI gloss" for natural, distinctive vibes—think DALL-E 2's quirky grain without the saturation. It outperforms open peers and rivals pros in prefs, fully Flux-compatible for LoRAs/tools. Yam and I geeked over the aesthetics frontier; it's a flexible base for fine-tunes, available on Hugging Face with commercial options via FAL/Replicate. If you're tired of cookie-cutter outputs, this breathes fresh life into generations.Ideogram Character launches: one-shot character consistency for everyone (X)Ideogram's Characters feature lets you upload one pic for instant, consistent variants—free for all, with inpainting to swap into memes/art. My tests nailed expressions/scenes (me in cyberpunk? Spot-on), though not always photoreal. Wolfram praised the accuracy; it's a meme-maker's dream! and they give like 10 free ones so give it a goTencent Hunyuan3D World Model 1.0 launches as the first open-source, explorable 3D world generator (X, HF)Tencent's Hunyuan3D World Model 1.0 is the first open-source generator of explorable 3D worlds from text/image—360° immersive, exportable meshes for games/modeling. ~33GB VRAM on complex scenes, but Wolfram called it a metaverse step; I wandered a demo scene, loving the potential despite edges. Integrate into CG pipelines? Game-changer for VR/creators.Voice & Audio Look I wasn't even mentioning this on the show, but it came across my feed just as I was about to wrap up ThursdAI, and it's really something. Riffusion joined forces producer and using FUZZ-2 they now have a fully Chatable studio producer, you can ask for.. anything you would ask in a studio! Here's my first reaction, and it's really fun, I think they still are open with the invite code 'STUDIO'... I'm not afiliated with them at all! Tools Ok I promised some folks we'll add this in, Nisten went super viral last week with him using a new open source tool called Crush from CharmBracelet, which is an open version of VSCode and it looks awesome! He gave a demo live on the show, including how to set it up to work, with subagents etc. If you're into vibe coding, and using the open source models, def. give Crush a try it's really flying and looks cool! Phew, ok, we somehow were able to cover ALLL these releases this week, and we didn’t even have an interview! Here’s the TL;DR and links to the folks who subscribed (I’m trying a new thing to promote subs on this newsletter) and see you in two weeks (next week is Wolframs turn again as I’m somewhere in Europe!) ThursdAI - July 31st, 2025 - TL;DR* Hosts and Guests* Alex Volkov - AI Evangelist & Weights & Biases (@altryne)* Co Hosts - @WolframRvnwlf @yampeleg @nisten @ldj* Open Source LLMs* Zhipu drops GLM-4.5 355B (A32B) AI model (X, HF)* ARCEE AFM‑4.5B and AFM‑4.5B‑Base weights released (X, HF)* Qwen is on 🔥 - 3 new models: