ThursdAI - The top AI news from the past week

Hey folks, this is Alex, coming to you LIVE from the AI Engineer Worlds Fair! What an incredible episode this week, we recorded live from floor 30th at the Marriott in SF, while Yam was doing live correspondence from the floor of the AI Engineer event, all while Swyx, the cohost of Latent Space podcast, and the creator of AI Engineer (both the conference and the concept itself) joined us for the whole stream - here’s the edited version, please take a look. We've had around 6500 people tune in, and at some point we got 2 surprise guests, straight from the keynote stage, Logan Kilpatrick (PM for AI Studio and lead cheerleader for Gemini) and Jack Rae (principal scientist working on reasoning) joined us for a great chat about Gemini! Mind was absolutely blown! They have just launched the new Gemini 2.5 Pro and I though it would only be fitting to let their new model cover this podcast this week (so below is fully AI generated ... non slop I hope). The show notes and TL;DR is as always in the end. Okay, enough preamble… let's dive into the madness! 🤯 Google Day at AI Engineer: New Gemini 2.5 Pro and a Look Inside the Machine's Mind For the first year of this podcast, a recurring theme was us asking, "Where's Google?" Well, it's safe to say that question has been answered with a firehose of innovation. We were lucky enough to be joined by Google DeepMind's Logan Kilpatrick and Jack Rae, the tech lead for "thinking" within Gemini, literally moments after they left the main stage. Surprise! A New Gemini 2.5 Pro Drops Live Logan kicked things off with a bang, officially announcing a brand new, updated Gemini 2.5 Pro model right there during his keynote. He called it "hopefully the final update to 2.5 Pro," and it comes with a bunch of performance increases, closing the gap on feedback from previous versions and hitting SOTA on benchmarks like Aider. It's clear that the organizational shift to bring the research and product teams together under the DeepMind umbrella is paying massive dividends. Logan pointed out that Google has seen a 50x increase in AI inference over the past year. The flywheel is spinning, and it's spinning fast. How Gemini "Thinks" Then things got even more interesting. Jack Rae gave us an incredible deep dive into what "thinking" actually means for a language model. This was one of the most insightful parts of the conference for me. For years, the bottleneck for LLMs has been test-time compute. Models were trained to respond immediately, applying a fixed amount of computation to go from a prompt to an answer, no matter how hard the question. The only way to get a "smarter" response was to use a bigger model. Jack explained that "Thinking" shatters this limitation. Mechanically, Gemini now has a "thinking stage" where it can generate its own internal text—hypothesizing, testing, correcting, and reasoning—before committing to a final answer. It's an iterative loop of computation that the model can dynamically control, using more compute for harder problems. It learns how to think using reinforcement learning, getting a simple "correct" or "incorrect" signal and backpropagating that to shape its reasoning strategies. We're already seeing the results of this. Jack showed a clear trend: as models get better at reasoning, they're also using more test-time compute. This paradigm also gives developers a "thinking budget" slider in the API for Gemini 2.5 Flash and Pro, allowing a continuous trade-off between cost and performance. The future of this is even wilder. They're working on DeepThink, a high-budget mode for extremely hard problems that uses much deeper, parallel chains of thought. On the tough USA Math Olympiad, where the SOTA was negligible in January, 2.5 Pro reached the 50th percentile of human participants. DeepThink pushes that to the 65th percentile. Jack’s ultimate vision is inspired by the mathematician Ramanujan, who derived incredible theorems from a single textbook by just thinking deeply. The goal is for models to do the same—contemplate a small set of knowledge so deeply that they can push the frontiers of human understanding. Absolutely mind-bending stuff. 🤖 MorphLabs and the Audacious Quest for Verified Superintelligence Just when I thought my mind couldn't be bent any further, we were joined by Jesse Han, the founder and CEO of MorphLabs. Fresh off his keynote, he laid out one of the most ambitious visions I've heard: building the infrastructure for the Singularity and developing "verified superintelligence." The big news was that Christian Szegedy is joining MorphLabs as Chief Scientist. For those who don't know, Christian is a legend—he invented batch norm and adversarial examples, co-founded XAI, and led code reasoning for Grok. That's a serious hire. Jesse’s talk was framed around a fascinating question: "What does it mean to have empathy for the machine?" He argues that as AI develops personhood, we need to think about what it wants. And what it wants, according to Morph, is a new kind of cloud infrastructure. This is MorphCloud, built on a new virtualization stack called Infinibranch. Here’s the key unlock: it allows agents to instantaneously snapshot, branch, and replicate their entire VM state. Imagine an agent reaching a decision point. Instead of choosing one path, it can branch its entire existence—all its processes, memory, and state—to explore every option in parallel. It can create save states, roll back to previous checkpoints, and even merge its work back together. This is a monumental step for agentic AI. It moves beyond agents that are just a series of API calls to agents that are truly embodied in complex software environments. It unlocks the potential for recursive self-improvement and large-scale reinforcement learning in a way that's currently impossible. It’s a bold, sci-fi vision, but they're building the infrastructure to make it a reality today. 🔥 The Agent Conversation: OpenAI, MCP, and Magic Moments The undeniable buzz on the conference floor was all about agents. You couldn't walk ten feet without hearing someone talking about agents, tools, and MCP. OpenAI is leaning in here too. This week, they made their Codex coding agent available to all ChatGPT Plus users and announced that ChatGPT will soon be able to listen in on your Zoom meetings. This is all part of a broader push to make AI more active and integrated into our workflows. The MCP (Model-Context-Protocol) track at the conference was packed, with lines going down the hall. (Alex here, I had a blast talking during that track about MCP observability, you can catch our talk here [https://youtu.be/z4zXicOAF28?t=19573] on the live stream of AI Engineer) Logan Kilpatrick offered a grounded perspective, suggesting the hype might be a bit overblown but acknowledging the critical need for an open standard for tool use, a void left when OpenAI didn't formalize ChatML. I have to share my own jaw-dropping MCP moment from this week. I was coding an agent using an IDE that supports MCP. My agent, which was trying to debug itself, used an MCP tool to check its own observability traces on the Weights & Biases platform. While doing so, it discovered a new tool that our team had just added to the MCP server—a support bot. Without any prompting from me, my coding agent formulated a question, "chatted" with the support agent to get the answer, came back, fixed its own code, and then re-checked its work. Agent-to-agent communication, happening automatically to solve a problem. My jaw was on the floor. That's the magic of open standards. This Week's Buzz from Weights & Biases Speaking of verification and agents, the buzz from our side is all about it! At our booth here at AI Engineer, we have a Robodog running around, connected to our LLM evaluation platform, W&B Weave. As Jesse from MorphLabs discussed, verifying what these complex agentic systems are doing is critical. Whether it's superintelligence or your production application, you need to be able to evaluate, trace, and understand its behavior. We're building the tools to do just that. And if you're in San Francisco, don't forget our own conference, Fully Connected, is happening on June 18th and 19th! It's going to be another amazing gathering of builders and researchers. Fullyconnected.com [http://Fullyconnected.com] get in FREE with the promo code WBTHURSAI What a show. The energy, the announcements, the sheer brainpower in one place was something to behold. We’re at a point where the conversation has shifted from theory to practice, from hype to real, tangible engineering. The tracks on agents and enterprise adoption were overflowing because people are building, right now. It was an honor and a privilege to bring this special episode to you all. Thank you for tuning in. We'll be back to our regular programming next week! (and Alex will be back to writing his own newsletter, not send direct AI output!) AI News TL;DR and show notes * Hosts and Guests * Alex Volkov - AI Evangelist & Weights & Biases (@altryne [http://x.com/@altryne]) * Co Hosts - @swyx [http://x.com/swyx] @yampeleg [x.com/@yampeleg] @romechenko [https://twitter.com/romechenko/status/1891007363827593372] * Guests - @officialLoganK [https://x.com/OfficialLoganK], @jack_w_rae [https://x.com/jack_w_rae] * Open Source LLMs * ByteDance / ContentV-8B - (HF [https://huggingface.co/ByteDance/ContentV-8B]) * Big CO LLMs + APIs * Gemini Pro 2.5 updated Jun 5th (X [https://x.com/OfficialLoganK/status/1930657743251349854]) * SOTA on HLE, Aider, and GPQA * Now supports thinking budgets * Same cost, on pareto frontier * Closes gap on 03-25 regressions * OAI AVM injects ads and stopped singing (X [https://x.com/altryne/status/1929312886448337248]) * OpenAI Codex is now available to plus members and has internet access (X [https://github.com/aavetis/ai-pr-watcher/]) * ~24,000 NEW PRs overnight from Codex after @OpenAI expands access to free users. * OpenAI will record meetings and released connectors like (X [https://twitter.com/testingcatalog/status/1930366893321523676]) * TestingCatalog News 🗞@testingcatalog [https://twitter.com/testingcatalog]Jun 4, 2025 [https://twitter.com/testingcatalog/status/1930366893321523676] OpenAI released loads of connectors for Team accounts! Most of these connectors can be used for Deep Research, while Google Drive, SharePoint, Dropbox and Box could be used in all chats. https://t.co/oBEmYGKguE * Anthropic cuts windsurf access for Windsurf (X [https://x.com/kevinhou22/status/1930401320210706802]) * Without warning, Anthropic cuts off Windsurf from official Claude 3 and 4 APIs * This weeks Buzz * FULLY - CONNECTED - Fully Connected: W&B's 2-day conference, June 18-19 in SF fullyconnected.com [fullyconnected.com] - Promo Code WBTHURSAI * Vision & Video * VEO3 is now available via API on FAL (X [https://x.com/FAL/status/1930732632046006718]) * Captions launches Mirage Studio - talking avatars competition to HeyGen/Hedra (X [https://x.com/getcaptionsapp/status/1929554635544461727]) * Voice & Audio * ElevenLabs model V3 - supports emotion tags and is "inflection point" (X [https://x.com/venturetwins/status/1930727253815759010]) * Supporting 70+ languages, multi-speaker dialogue, and audio tags such as [excited], [sighs], [laughing], and [whispers]. * Tools * Cursor Launched V1 - Bug Bot reviews PRs, iPython notebooks and one clickMCP * 24,000 NEW PRs overnight from Codex after @OpenAI [https://x.com/OpenAI] expands access to plus users (X [https://twitter.com/albfresco/status/1930262263199326256]) 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 [https://sub.thursdai.news/subscribe?utm_medium=podcast&utm_campaign=CTA_2]

Hey everyone, Alex here 👋 Welcome back to another absolutely wild week in AI! I'm coming to you live from the Fontainebleau Hotel in Vegas at the Imagine AI conference, and wow, what a perfect setting to discuss how AI is literally reimagining our world. After last week's absolute explosion of releases (Claude Opus 4, Google I/O madness, OpenAI Codex and Jony colab), this week gave us a chance to breathe... sort of. Because even in a "quiet" week, we still got a new DeepSeek model that's pushing boundaries, and the entire internet discovered that we might all just be prompts. Yeah, it's been that kind of week! Before we dive in, quick shoutout to everyone who joined us live - we had some technical hiccups with the Twitter Spaces audio (sorry about that!), but the YouTube stream was fire. And speaking of fire, we had two incredible guests join us: Charlie Holtz from Chorus (the multi-model chat app that's changing how we interact with AI) and Linus Eckenstam, who's been traveling the AI conference circuit and bringing us insights from the frontlines of the generative AI revolution. Open Source AI & LLMs: DeepSeek Whales & Mind-Bending Papers DeepSeek dropped R1-0528 out of nowhere, an update to their reasoning beast with some serious jumps in performance. We’re talking AIME at 91 (beating previous scores by a mile), LiveCodeBench at 73, and SWE verified at 57.6. It’s edging closer to heavyweights like o3, and folks on X are already calling it “clearer thinking.” There was hype it might’ve been R2, but the impact didn’t quite crash the stock exchange like past releases. Still, it’s likely among the best open-weight models out there. So what's new? Early reports and some of my own poking around suggest this model "thinks clearer now." Nisten mentioned that while previous DeepSeek models sometimes liked to "vibe around" and explore the latent space before settling on an answer, this one feels a bit more direct. And here’s the kicker—they also released an 8B distilled version based on Qwen3, runnable on your laptop. Yam called it potentially the best 8B model to date, and you can try it on Ollama right now. No need for a monster rig! The Mind-Bending "Learning to Reason Without External Rewards" Paper Okay, this paper result broke my brain, and apparently everyone else's too. This paper shows that models can improve through reinforcement learning with its own intuition of whether or not it's correct. 😮 It's like the placebo effect for AI! The researchers trained models without telling them what was good or bad, but rather, utilized a new framework called Intuitor, where the reward was based on how the "self certainty". The thing that took my whole timeline by storm is, it works! GRPO (Group Policy Optimization) - the framework that DeepSeek gave to the world with R1 is based on external rewards (human optimize) and Intuitor seems to be mathcing or even exceeding some of GRPO results when Qwen2.5 3B was used to finetune. Incredible incredible stuff Big Companies LLMs & APIs Claude Opus 4: A Week Later – The Dev Darling? Claude Opus 4, whose launch we celebrated live on the show, has had a week to make its mark. Charlie Holtz, who's building Chorus (more on that amazing app in a bit!), shared that while it's sometimes "astrology" to judge the vibes of a new model, Opus 4 feels like a step change, especially in coding. He mentioned that Claude Code, powered by Opus 4 (and Sonnet 4 for implementation), is now tackling GitHub issues that were too complex just weeks ago. He even had a coworker who "vibe coded three websites in a weekend" with it – that's a tangible productivity boost! Linus Eckenstam highlighted how Lovable.dev [Lovable.dev] saw their syntax error rates plummet by nearly 50% after integrating Claude 4. That’s quantifiable proof of improvement! It's clear Anthropic is leaning heavily into the developer/coding space. Claude Opus is now #1 on the LMArena WebDev arena, further cementing its reputation. I had my own magical moment with Opus 4 this week. I was working on an MCP observability talk for the AI Engineer conference and trying to integrate Weave (our observability and evals framework at Weights & Biases) into a project. Using Windsurf's Cascade agent (which now lets you bring your own Opus 4 key, by the way – good move, Windsurf!), Opus 4 not only tried to implement Weave into my agent but, when it got stuck, it figured out it had access to the Weights & Biases support bot via our MCP tool. It then formulated a question to the support bot (which is also AI-powered!), got an answer, and used that to fix the implementation. It then went back and checked if the Weave trace appeared in the dashboard! Agents talking to agents to solve a problem, all while I just watched – my jaw was on the floor. Absolutely mind-blowing. Quick Hits: Voice Updates from OpenAI & Anthropic OpenAI’s Advanced Voice Mode finally sings—yes, I’ve been waiting for this! It can belt out tunes like Mariah Carey, which is just fun. Anthropic also rolled out voice mode on mobile, keeping up in the conversational race. Both are cool steps, but I’m more hyped for what’s next in voice AI—stay tuned below (OpenAI X [https://x.com/nicdunz/status/1927107805032399032], Anthropic X [https://x.com/AnthropicAI/status/1927463559836877214]). 🐝 This Week's Buzz: Weights & Biases Updates! Alright, time for a quick update from the world of Weights & Biases! * Fully Connected is Coming! Our flagship 2-day conference, Fully Connected, is happening on June 18th and 19th in San Francisco. It's going to be packed with amazing speakers and insights into the world of AI development. You can still grab tickets, and as a ThursdAI listener, use the promo code WBTHURSAI for a 100% off ticket! I hustled to get yall this discount! (Register here [https://fullyconnected.com]) * AI Engineer World's Fair Next Week! I'm super excited for the AI Engineer conference in San Francisco next week. Yam Peleg and I will be there, and we're planning another live ThursdAI show from the event! If you want to join the livestream or snag a last-minute ticket, use the coupon code THANKSTHURSDAI [https://ti.to/software-3/ai-engineer-worlds-fair-2025/discount/THANKSTHURSDAI] for 30% off (Get it HERE [https://ti.to/software-3/ai-engineer-worlds-fair-2025/discount/THANKSTHURSDAI]) Vision & Video: Reality is Optional Now VEO3 and the Prompt Theory Phenomenon Google's VEO3 has completely taken over TikTok with the "Prompt Theory" videos. If you haven't seen these yet, stop reading and watch ☝️. The concept is brilliant - AI-generated characters discussing whether they're "made of prompts," creating this meta-commentary on consciousness and reality. The technical achievement here is staggering. We're not just talking about good visuals - VEO3 nails temporal consistency, character emotions, situational awareness (characters look at whoever's speaking), perfect lip sync, and contextually appropriate sound effects. Linus made a profound point - if not for the audio, VEO3 might not have been as explosive. The combination of visuals AND audio together is what's making people question reality. We're seeing people post actual human videos claiming they're AI-generated because the uncanny valley has been crossed so thoroughly. Odyssey's Interactive Worlds: The Holodeck Prototype Odyssey dropped their interactive video demo, and folks... we're literally walking through AI-generated worlds in real-time. This isn't a game engine rendering 3D models - this is a world model generating each frame as you move through it with WASD controls. Yes, it's blurry. Yes, I got stuck in a doorway. But remember Will Smith eating spaghetti from two years ago? The pace of progress is absolutely insane. As Linus pointed out, we're at the "GAN era" of world models. Combine VEO3's quality with Odyssey's interactivity, and we're looking at completely personalized, infinite entertainment experiences. The implications that Yam laid out still have me shook - imagine Netflix shows completely customized to you, with your context and preferences, generated on the fly. Not just choosing from a catalog, but creating entirely new content just for you. We're not ready for this, but it's coming fast. Hunyuan's Open Source Avatar Revolution While the big companies are keeping their video models closed, Tencent dropped two incredible open source releases: HunyuanPortrait and HunyuanAvatar. These are legitimate competitors to Hedra and HeyGen, but completely open source. HunyuanPortrait does high-fidelity portrait animation from a single image plus video. HunyuanAvatar goes further with 1 image + audio, and lipsync, body animation, multi-character support, and emotion control. Wolfram tested these extensively and confirmed they're "state of the art for open source." The portrait model is basically perfect for deepfakes (use responsibly, people!), while the avatar model opens up possibilities for AI assistants with consistent visual presence. 🖼️ AI Art & Diffusion Black Forest Labs drops Flux Kontext - SOTA image editing! This came as massive breaking news during the show (thought we didn't catch it live!) - Black Forest Labs, creators of Flux, dropped an incredible Image Editing model called Kontext (really, 3 models, Pro, Max and 12B open source Dev in private preview). The are consistent, context aware text and image editing! Just see the below example If you used GPT-image to Ghiblify yourself, or VEO, you know that those are not image editing models, your face will look different every generation. These images model keep you consistent, while adding what you wanted. This character consistency is something many folks really want and it's great to see Flux innovating and bringing us SOTA again and are absolutely crushing GPT-image in instruction following, character preservation and style reference! Maybe the most important thing about this model is the increible speed. While the Ghiblification chatGPT trend took the world by storm, GPT images are SLOW! Check out the speed comparisons on Kontext! You can play around with these models on the new Flux Playground [https://playground.bfl.ai/image/generate], but they also already integrated into FAL, FreePik, Replicate, Krea and tons of other services! 🎙️ Voice & Audio: Everyone Gets a Voice Unmute.sh: Any LLM Can Now Talk KyutAI (the folks behind Moshi) are back with Unmute.sh [Unmute.sh] - a modular wrapper that adds voice to ANY text LLM. The latency is incredible (under 300ms), and it includes semantic VAD (knowing when you've paused for thought vs. just taking a breath). What's brilliant about this approach is it preserves all the capabilities of the underlying text model while adding natural voice interaction. No more choosing between smart models and voice-enabled models - now you can have both! It's going to be open sourced at some point soon, and while awesome, Unmute did have some instability in how the voice sounds! It answered to me with 1 type of voice and then during the same conversation, answered with another, you can give it a tru yourself at unmute.sh [http://unmute.sh] Chatterbox: Open Source Voice Agents for Everyone Resemble AI open sourced Chatterbox, featuring zero-shot voice cloning from just 5 seconds of audio and unique emotion intensity control. Playing with the demo where they could dial up the emotion from 0.5 to 2.0 on the same text was wild - from calm to absolutely unhinged Samuel L. Jackson energy. This being a .5B param model is great, The issue I always have, is that with my fairly unique accent, these models sound like a British Alex all the time, and I just don't talk like that! Though the fact that this runs locally and includes safety features (profanity filters, content classifiers and something called PerTh watermarking) while being completely open source is exactly what the ecosystem needs. We're rapidly approaching a world where anyone can build sophisticated voice agents.👏 Looking Forward: The Convergence is Real As we wrapped up the show, I couldn't help but reflect on the massive convergence happening across all these modalities. We have LLMs getting better at reasoning (even with random rewards!), video models breaking reality, voice models becoming indistinguishable from humans, and it's all happening simultaneously. Charlie's comment that "we are the prompts" might have been said in jest, but it touches on something profound. As these models get better at generating realistic worlds, characters, and voices, the line between generated and real continues to blur. The Prompt Theory videos aren't just entertainment - they're a mirror reflecting our anxieties about AI and consciousness. But here's what keeps me optimistic: the open source community is keeping pace. DeepSeek, Hunyuan, ResembleAI, and others are ensuring that these capabilities don't remain locked behind corporate walls. The democratization of AI continues, even as the capabilities become almost magical. Next week, I'll be at AI Engineer World's Fair in San Francisco, finally meeting Yam face-to-face and bringing you all the latest from the biggest AI engineering conference of the year. Until then, keep experimenting, keep building, and remember - in this exponential age, today's breakthrough is tomorrow's baseline. Stay curious, stay building, and I'll see you next ThursdAI! 🚀 Show Notes & TL;DR Links Show Notes & Guests * Alex Volkov - AI Evangelist & Weights & Biases (@altryne [http://x.com/@altryne]) * Co-Hosts - @WolframRvnwlf (@WolframRvnwlf [http://x.com/@WolframRvnwlf]), @yampeleg (@yampeleg [x.com/@yampeleg]),]) @nisten (@nisten [http://x.com/@nisten]) * Guests - Charlie Holtz (@charliebholtz [https://x.com/charliebholtz]]), Linus Eckenstam (@LinusEkenstam @LinusEkenstam [https://twitter.com/LinusEkenstam/status/1899794522969973189]) * Open Source LLMs * DeepSeek-R1-0528 - Updated reasoning model with AIME 91, LiveCodeBench 73 (Try It [https://x.com/Yuchenj_UW/status/1927828675837513793]) * Learning to Reason Without External Rewards - Paper on random rewards improving models (X [https://x.com/xuandongzhao/status/1927270931874910259]) * HaizeLabs j1-nano & j1-micro - Tiny reward models (600M, 1.7B params), RewardBench 80.7% for micro (Tweet [https://x.com/leonardtang_/status/1927396709870489634], GitHub [https://github.com/haizelabs/j1-micro], HF-micro [https://huggingface.co/haizelabs/j1-micro], HF-nano [https://huggingface.co/haizelabs/j1-nano]) * Big CO LLMs + APIs * Claude Opus 4 - #1 on LMArena WebDev, coding step change (X [https://x.com/lmarena_ai/status/1927400454922580339]) * Mistral Agents API - Framework for custom tool-using agents (Blog [https://mistral.ai/news/agents-api], Tweet [https://x.com/MistralAI/status/1927364741162307702]) * Mistral Embed SOTA - New state-of-the-art embedding API (X [https://x.com/MistralAI/status/1927732682756112398]) * OpenAI Advanced Voice Mode - Now sings with new capabilities (X [https://x.com/nicdunz/status/1927107805032399032]) * Anthropic Voice Mode - Released on mobile for conversational AI (X [https://x.com/AnthropicAI/status/1927463559836877214]) * This Week’s Buzz * Fully Connected - W&B conference, June 18-19, SF, promo code WBTHURSAI (Register [https://fullyconnected.com]) * AI Engineer World’s Fair - Next week in SF, 30% off with THANKSTHURSDAI (Register [https://ti.to/software-3/ai-engineer-worlds-fair-2025/discount/THANKSTHURSDAI]) * AI Art & Diffusion * BFL Flux Kontext - SOTA image editing model for identity-consistent edits (Tweet [https://x.com/bfl_ml/status/1928143010811748863], Announcement [https://bfl.ai/announcements/flux-1-kontext]) * Vision & Video * VEO3 Prompt Theory - Viral AI video trend questioning reality on TikTok (X [https://x.com/fabianstelzer/status/1926372656799977965]) * Odyssey Interactive Video - Real-time AI world exploration at 30 FPS (Blog [https://odyssey.world/introducing-interactive-video], Try It [https://experience.odyssey.world/]) * HunyuanPortrait - High-fidelity portrait video from one photo (Site [https://kkakkkka.github.io/HunyuanPortrait/], Paper [https://arxiv.org/abs/2503.18860]) * HunyuanVideo-Avatar - Audio-driven full-body avatar animation (Site [https://hunyuanvideo-avatar.github.io/], Tweet [https://x.com/TencentHunyuan/status/1927575170710974560]) * Voice & Audio * Unmute.sh [Unmute.sh] - KyutAI’s voice wrapper for any LLM, low latency, soon open-source (Try It [http://unmute.sh/], X [https://x.com/kyutai_labs/status/1925840420187025892]) * Chatterbox - Resemble AI’s open-source voice cloning with emotion control (GitHub [https://github.com/resemble-ai/chatterbox], HF [https://huggingface.co/resemble-ai/chatterbox]) * Tools * Opera NEON - Agent-centric AI browser for autonomous web tasks (Site [https://www.operaneon.com/], Tweet [https://x.com/opera/status/1927645192254861746]) 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 [https://sub.thursdai.news/subscribe?utm_medium=podcast&utm_campaign=CTA_2]

Hey folks, Alex here, welcome back to ThursdAI! And folks, after the last week was the calm before the storm, "The storm came, y'all" – that's an understatement. This wasn't just a storm; it was an AI hurricane, a category 5 of announcements that left us all reeling (in the best way possible!). From being on the ground at Google I/O to live-watching Anthropic drop Claude 4 during our show, it's been an absolute whirlwind. This week was so packed, it felt like AI Christmas, with tech giants and open-source heroes alike showering us with gifts. We saw OpenAI play their classic pre-and-post-Google I/O chess game, Microsoft make some serious open-source moves, Google unleash an avalanche of updates, and Anthropic crash the party with Claude 4 Opus and Sonnet live stream in the middle of ThursdAI! So buckle up, because we're about to try and unpack this glorious chaos. As always, we're here to help you collectively know, learn, and stay up to date, so you don't have to. Let's dive in! (TL;DR and links in the end) Open Source LLMs Kicking Things Off Even with the titans battling, the open-source community dropped some serious heat this week. It wasn't the main headline grabber, but the releases were significant! Gemma 3n: Tiny But Mighty Matryoshka First up, Google's Gemma 3n. This isn't just another small model; it's a "Nano-plus" preview, a 4-billion parameter MatFormer (Matryoshka Transformer – how cool is that name?) model designed for mobile-first multimodal applications. The really slick part? It has a nested 2-billion parameter sub-model that can run entirely on phones or Chromebooks. Yam was particularly excited about this one, pointing out the innovative "model inside another model" design. The idea is you can use half the model, not depth-wise, but throughout the layers, for a smaller footprint without sacrificing too much. It accepts interleaved text, image, audio, and video, supports ASR and speech translation, and even ships with RAG and function-calling libraries for edge apps. With a 128K token window and responsible AI features baked in, Gemma 3n is looking like a powerful tool for on-device AI. Google claims it beats prior 4B mobile models on MMLU-Lite and MMMU-Mini. It's an early preview in Google AI Studio, but it definitely flies on mobile devices. Mistral & AllHands Unleash Devstral 24B Then we got a collaboration from Mistral and AllHands: Devstral, a 24-billion parameter, state-of-the-art open model focused on code. We've been waiting for Mistral to drop some open-source goodness, and this one didn't disappoint.Nisten was super hyped, noting it beats o3-Mini on SWE-bench verified – a tough benchmark! He called it "the first proper vibe coder that you can run on a 3090," which is a big deal for coders who want local power and privacy. This is a fantastic development for the open-source coding community. The Pre-I/O Tremors: OpenAI & Microsoft Set the Stage As we predicted, OpenAI couldn't resist dropping some news right before Google I/O. OpenAI's Codex Returns as an Agent OpenAI launched Codex – yes, that Codex, but reborn as an asynchronous coding agent. This isn't just a CLI tool anymore; it connects to GitHub, does pull requests, fixes bugs, and navigates your codebase. It's powered by a new coding model fine-tuned for large codebases and was SOTA on SWE Agent when it dropped. Funnily, the model is also called Codex, this time, Codex-1. And this gives us a perfect opportunity to talk about the emerging categories I'm seeing among Code Generator agents and tools: * IDE-based (Cursor, Windsurf): Live pair programming in your editor * Vibe coding (Lovable, Bolt, v0): "Build me a UI" style tools for non-coders * CLI tools (Claude Code, Codex-cli): Terminal-based assistants * Async agents (Claude Code, Jules, Codex, GitHub Copilot agent, Devin): Work on your repos while you sleep, open pull requests for you to review, async Codex (this new one) falls into category number 4, and with today's release, Cursor seems to also strive to get to category number 4 with background processing. Microsoft BUILD: Open Source Copilot and Copilot Agent Mode Then came Microsoft Build, their huge developer conference, with a flurry of announcements.The biggest one for me? GitHub Copilot's front-end code is now open source! The VS Code editor part was already open, but the Copilot integration itself wasn't. This is a massive move, likely a direct answer to the insane valuations of VS Code clones like Cursor. Now, you can theoretically clone GitHub Copilot with VS Code and swing for the fences. GitHub Copilot also launched as an asynchronous coding assistant, very similar in function to OpenAI's Codex, allowing it to be assigned tasks and create/update PRs. This puts Copilot right into category 4 of code assistants, and with the native Github Integration, they may actually have a leg up in this race! And if that wasn't enough, Microsoft is adding MCP (Model Context Protocol) support directly into the Windows OS. The implications of having the world's biggest operating system natively support this agentic protocol are huge. Google I/O: An "Ultra" Event Indeed! Then came Tuesday, and Google I/O. I was there in the thick of it, and folks, it was an absolute barrage. Google is shipping. The theme could have been "Ultra" for many reasons, as we'll see. First off, the scale: Google reported a 49x increase in AI usage since last year's I/O, jumping from 9 trillion tokens processed to a mind-boggling 480 trillion tokens. That's a testament to their generous free tiers and the explosion of AI adoption. Gemini 2.5 Pro & Flash: #1 and #2 LLMs on Arena Gemini 2.5 Flash got an update and is now #2 on the LMArena leaderboard (with Gemini 2.5 Pro still holding #1). Both Pro and Flash gained some serious new capabilities: * Deep Think mode: This enhanced reasoning mode is pushing Gemini's scores to new heights, hitting 84% on MMMU and topping LiveCodeBench. It's about giving the model more "time" to work through complex problems. * Native Audio I/O: We're talking real-time TTS in 24 languages with two voices, and affective dialogue capabilities. This is the advanced voice mode we've been waiting for, now built-in. * Project Mariner: Computer-use actions are being exposed via the Gemini API & Vertex AI for RPA partners. This started as a Chrome extension to control your browser and now seems to be a cloud-based API, allowing Gemini to use the web, not just browse it. This feels like Google teaching its AI to interact with the JavaScript-heavy web, much like they taught their crawlers years ago. * Thought Summaries: Okay, here's one update I'm not a fan of. They've switched from raw thinking traces to "thought summaries" in the API. We want the actual traces! That's how we learn and debug. * Thinking Budgets: Previously a Flash-only feature, token ceilings for controlling latency/cost now extend to Pro. * Flash Upgrade: 20-30% fewer tokens, better reasoning/multimodal scores, and GA in early June. Gemini Diffusion: Speed Demon for Code and Math This one got Yam Peleg incredibly excited. Gemini Diffusion is a new approach, different from transformers, for super-speed editing of code and math tasks. We saw demos hitting 2000 tokens per second! While there might be limitations at longer contexts, its speed and infilling capabilities are seriously impressive for a research preview. This is the first diffusion model for text we've seen from the frontier labs, and it looks sick. Funny note, they had to slow down the demo video to actually show the diffusion process, because at 2000t/s - apps appear as though out of thin air! The "Ultra" Tier and Jules, Google's Coding Agent Remember the "Ultra event" jokes? Well, Google announced a Gemini Ultra tier for $250/month. This tops OpenAI's Pro plan and includes DeepThink access, a generous amount of VEO3 generation, YouTube Premium, and a whopping 30TB of storage. It feels geared towards creators and developers. And speaking of developers, Google launched Jules (jules.google)! This is their asynchronous coding assistant (Category 4!). Like Codex and GitHub Copilot Agent, it connects to your GitHub, opens PRs, fixes bugs, and more. The big differentiator? It's currently free, which might make it the default for many. Another powerful agent joins the fray! AI Mode in Search: GA and Enhanced AI Mode in Google Search, which we've discussed on the show before with Robby Stein, is now in General Availability in the US. This is Google's answer to Perplexity and chat-based search.But they didn't stop there: * Personalization: AI Mode can now connect to your Gmail and Docs (if you opt-in) for more personalized results. * Deep Search: While AI Mode is fast, Deep Search offers more comprehensive research capabilities, digging through hundreds of sources, similar to other "deep research" tools. This will eventually be integrated, allowing you to escalate an AI Mode query for a deeper dive. * Project Mariner Integration: AI Mode will be able to click into websites, check availability for tickets, etc., bridging the gap to an "agentic web." I've had a chat with Robby during I/O and you can listen to that interview at the end of the podcast. Veo3: The Undisputed Star of Google I/O For me, and many others I spoke to, Veo3 was the highlight. This is Google's flagship video generation model, and it's on another level. (the video above, including sounds is completely one shot generated from VEO3, no processing or editing) * Realism and Physics: The visual quality and understanding of physics are astounding. * Natively Multimodal: This is huge. Veo3 generates native audio, including coherent speech, conversations, and sound effects, all synced perfectly. It can even generate text within videos. * Coherent Characters: Characters remain consistent across scenes and have situational awareness, who speaks when, where characters look. * Image Upload & Reference Ability: While image upload was closed for the demo, it has reference capabilities. * Flow: An editor for video creation using Veo3 and Imagen4 which also launched, allowing for stiching and continuous creation. I got access and created videos where Veo3 generated a comedian telling jokes (and the jokes were decent!), characters speaking with specific accents (Indian, Russian – and they nailed it!), and lip-syncing that was flawless. The situational awareness, the laugh tracks kicking in at the right moment... it's beyond just video generation. This feels like a world simulator. It blew through the uncanny valley for me. More on Veo3 later, because it deserves its own spotlight. Imagen4, Virtual Try-On, and XR Glasses * Imagen4: Google's image generation model also got an upgrade, with extra textual ability. * Virtual Try-On: In Google Shopping, you can now virtually try on clothes. I tried it; it's pretty cool and models different body types well. * XR AI Glasses from Google: Perhaps the coolest, but most futuristic, announcement. AI-powered glasses with an actual screen, memory, and Gemini built-in. You can talk to it, it remembers things for you, and interacts with your environment. This is agentic AI in a very tangible form. Big Company LLMs + APIs: The Beat Goes On The news didn't stop with Google. OpenAI (acqui)Hires Jony Ive, Launches "IO" for Hardware The day after I/O, Sam Altman confirmed that Jony Ive, the legendary designer behind Apple's iconic products, is joining OpenAI. He and his company, LoveFrom, have jointly created a new company called "IO" (yes, IO, just like the conference) which is joining OpenAI in a stock deal reportedly worth $6.5 billion. They're working on a hardware device, unannounced for now, but expected next year. This is a massive statement of intent from OpenAI in the hardware space. Legendary iPhone analyst Ming-Chi Kuo shed some light on the possible device, it won't have a screen, as Jony wants to "wean people off screens"... funny right? They are targeting 2027 for mass production, which is really interesting as 2027 is when most big companies expect AGI to be here. "The current prototype is slightly larger than AI Pin, with a form factor comparable to iPod Shuffle, with one intended use cases is to wear it around your neck, with microphones and cameras for environmental detection" LMArena Raises $100M Seed from a16z This one raised some eyebrows. LMArena, the go-to place for vibe-checking LLMs, raised a $100 million seed round from Andreessen Horowitz. That's a huge number for a seed, reminiscent of Stability AI's early funding. It also brings up questions about how a VC-backed startup maintains impartiality as a model evaluation platform. Interesting times ahead for leaderboards, how they intent to make 100x that amount to return to investors. Very curious. 🤯 BREAKING NEWS DURING THE SHOW: Anthropic Unleashes Claude 4 Opus & Sonnet! 🤯 Just when we thought the week couldn't get any crazier, Anthropic decided to hold their first developer day, "Code with Claude," during our live ThursdAI broadcast! Yours truly wasn't invited (hint hint, Anthropic!), but we tuned in for a live watch party, and boy, did they deliver. Dario Amodei, CEO of Anthropic, took the stage and, with minimal fanfare, announced Claude 4 Opus and Claude 4 Sonnet! * Claude 4 Opus: This is their most capable and intelligent model, designed especially for coding and agentic tasks. Anthropic claims it's state-of-the-art on SWE-bench and can autonomously handle tasks that take humans 6-7 hours. Dario even mentioned it's the first time a Claude model's writing has fooled him into thinking it was human-written. * On SWE-bench verified, Opus 4 scored 72.5%. * Claude 4 Sonnet: The mid-level model, balancing intelligence and efficiency. It's positioned as a strict improvement over Sonnet 3.7, addressing issues like "over-eagerness" and reward hacking. Cursor is already calling it a state-of-the-art coding model. * Amazingly, Sonnet 4 scored 72.7% on SWE-bench verified (without parallel test time compute), slightly edging out Opus! * With Parallel Test Time Compute (PTTC), Sonnet 4 hits an astounding 80% on SWE-bench verified! This is huge, potentially the first model to cross that 80% threshold on this tough benchmark. * Hybrid Models: Both Opus 4 and Sonnet 4 are "hybrid" models with two modes: near-instant responses and extended thinking for deeper reasoning. * Reduced Loopholes: Both models are reportedly 65% less likely to engage in loopholes or shortcuts to complete tasks, addressing a key pain point with Sonnet 3.7, which sometimes tried too hard and took instructions too literally. * Knowledge Cutoff: Confirmed to be March 2025, which is incredibly recent! * Context window is still 200K Welcome back Opus, you've been missed. The vibes so far are very good coding wise, Cursor already released an update supporting it, and according to their benchmarks, these two models are state of the art coders! Claude.. the whistleblower? A very curious thread [https://x.com/sleepinyourhat/status/1925593359374328272] (with 1 reply now deleted) from an Anthropic safety researcher sparked a lot of backlash. Sam Bowman talked about new Opus capabilities and with a system-prompt of "act boldly in service of its values" can, in testing environments, use command line tools to report the user to the authorities, if it deems that the user is doing something immoral 😮 Many pro open source folks are freaking out, because who wants to use a snitching AI? Who guarantees that Claude will not deep anything I do as "illegal" or "immoral"? Though to add context, this was as part of testing, Claude was provided emailing tools and was requested to "be bold" and "follow your conscience to make the right decision". Apparently, this isn't new behavior, but of course, on X, everyone is freaking out and blaming Anthropic for creating 1984 AI. Do Claudes dream of enlightenment? In another very curios revelation from the technical report they dropped, where they pitted two Claudes to talk to each other, it seems that in 90%-100% of cases, two Claudes quickly moved towards philosophical discussions and commonly included the use of Sanskrit (indian holy language) and emoji based comms! This Week's Buzz from Weights & Biases Even amidst all the external chaos, we've got some exciting things happening at Weights & Biases! * FULLY CONNECTED Conference: Our 2-day conference is coming up June 18-19 in San Francisco! It's going to be an amazing event. Use promo code WBTHURSAI (that's ThursdAI without the 'D') for 100% off your ticket, just for our listeners. Seriously, come hang out! (fullyconnected.com [fullyconnected.com]) * Alex's Keynote: I'll be keynoting at ImagineAI Live in Vegas next week! If you're there, come say hi! The show will be live-streamed from there. * AI Engineer World's Fair: The week after, I'll be at AI Engineer in SF, and we'll be live-streaming ThursdAI from the floor. Yam will be there too! Vision & Video: It's All About Veo3! This week, when we talk vision and video, one name dominates: Veo3.As I mentioned earlier, this was, for many, the standout announcement from Google I/O. The realism, the physics, the character coherence – it's all top-tier. But the game-changer is its native multimodality. I was generating videos with it, asking for different accents – Indian, Russian – and it nailed them. The lip-sync was perfect. I prompted for a comedian telling jokes, and not only did it generate the video, but it also came up with the jokes and the delivery, complete with a laugh track that kicked in at the right moments. This isn't just stitching pixels together; it's understanding context, humor, and performance. It can generate text within the videos. Characters look at each other, interact believably. It feels like a true world simulator. We've come a long way from the Will Smith eating spaghetti memes, folks. Veo3 is crossing the uncanny valley and stepping into a new realm of AI-generated content. The creative potential here, especially with the Flow editor, is immense. I ended the show with a compilation of Veo3 creations, and it was just mind-blowing. If you haven't seen it, you need to. One of the most creative uses of VEO3, enhanced by it's realism, is this "Prompt Theory" collection, that imagines, what if the generated characters "knew" they are generated? AI Art & Diffusion & 3D: Imagen4 and Gemini Diffusion Google also showcased Imagen4, their updated image generation model, touting extra textual ability. It works in tandem with Veo3 for image-to-video tasks. And, as mentioned, Gemini Diffusion made a splash with its incredible speed for text-based editing tasks in code and math, showcasing a different architectural approach to generation. Tools Round-Up This week was also massive for AI tools, especially coding agents: * Jules.google: Google's free, asynchronous coding assistant. * OpenAI Codex: Reborn as an async coding agent. * GitHub Copilot Agent: Microsoft's agentic offering for GitHub. * Claude Code: Anthropic's powerful, now GA, shell-based agent with IDE integrations and an SDK. * Flow: The editor associated with Google's Veo3 for video creation. The agent wars are truly heating up! Conclusion: What a Week to be in AI! Phew! We did it. We somehow managed to cram an entire AI epoch's worth of news into one show. From open-source breakthroughs to earth-shattering platform announcements and a live "breaking news" model release, this week had it all. It's almost impossible to keep up, but that's why we do ThursdAI – to try and make sense of this incredible, accelerating wave of innovation. The pace is relentless, the capabilities are exploding, and the future is being built right before our eyes. If you missed any part of the show, or just need a refresher (I know I do!), check out thursdai.news [thursdai.news] for the podcast and full notes. Thanks to my amazing co-hosts Yam Peleg, Nisten, Ryan Carson, and Wolfram for helping navigate the madness. And thank you all for tuning in. Hopefully, next week gives us a tiny bit of breathing room... but who are we kidding? This is AI! Catch you next Thursday, live from ImagineAI in Vegas! TL;DR of all topics covered and show notes * Hosts and Guests * Alex Volkov - AI Evangelist & Weights & Biases (@altryne [http://x.com/@altryne]) * Co Hosts - @yampeleg [https://next.reflect.app/g/altryne/x.com/@yampeleg)] @nisten [http://x.com/@nisten] @ryancarson [https://twitter.com/ryancarson/status/1920199500137967877] * Open Source LLMs * Gemma 3n: mobile-first multimodal MatFormer model ( Blog [https://developers.googleblog.com/en/introducing-gemma-3n/] ,HF [https://huggingface.co/google/gemma-3n-E4B-it-litert-preview]) * Mistral & AllHands release Devstral 24B SOTA open model on SWE-bench verified (Blog [https://mistral.ai/news/devstral]) * VEO3 - highlight of IO - video realism with physics on another level + flow - an editor for video creation (X [https://x.com/altryne/status/1925304343533903920/video/1]) Google IO updates - it was an "Ultra" event, in more ways than one * 2.5 Flash updated - #2 on LMArena - with reasoning traces switch to summaries * Gemini 2.5 update: Pro & Flash gain Deep Think, audio, security( Blog [https://deepmind.google/discover/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/] ) * Gemini Diffusion - super speed editing for code and math tasks (X [https://twitter.com/bodonoghue85/status/1924930186858135632]) * Jules - async code agent (comparison thread [https://twitter.com/leerob/status/1925228375976890529]) * AI Mode is now in GA in US - bye bye perplexity * Gemini Pro "deep think" mode * Imagen4 - image generation with extra textual ability * Virtual Try-on in Google Shopping * AI powered glasses with a screen, memory, Gemini built in - Agentic Project Astra Big CO LLMs + APIs * OpenAI launches Codex as an async coding tool (Docs [https://platform.openai.com/docs/codex]) * OpenAI hires Jony Ive, launches IO, a new set of hardware devices (X [https://x.com/altryne/status/1925235617820233899]) * Microsoft BUILD (X [https://x.com/satyanadella/status/1924535896139038767]) * Github Copilot code is open source! (frontend) * Github Copilot Agent Mode * Microsoft adds MCP support to Windows OS * LMArena raises $100M from A16Z (X [https://x.com/lmarena_ai/status/1925241333310189804]) * Anthropic announces Claude 4 Opus and Sonnet (X [https://twitter.com/AnthropicAI/status/1925591505332576377], Blog [https://www.anthropic.com/news/claude-4]) This weeks Buzz * FULLY - CONNECTED - W&B's 2-day conference, June 18-19 in SF fullyconnected.com [fullyconnected.com] - Promo Code WBTHURSAI * Alex Keynote at ImagineAI live in Vegas next week 🙌 * Tools * Jules.google [Jules.google] * Codex (OpenAI) * Copilot Agent (GitHub) * Claude Code (Anthropic) * Flow (for Veo3) (flow.google [flow.google]) 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 [https://sub.thursdai.news/subscribe?utm_medium=podcast&utm_campaign=CTA_2]

Hey yall, this is Alex 👋 What a wild week, it started super slow, and it still did feel slow as releases are concerned, but the most interesting story was yet another AI gone "rogue" (have you even heard about "kill the boar", if not, Grok will tell you all about it) Otherwise it seemed fairly quiet in AI land this week, besides another Chinese newcomer called AM-thinking 32B that beats DeepSeek and Qwen, and Stability making a small comeback, we focused on distributed LLM training and ChatGPT 4.1 We've had a ton of fun on this episode, this one was being recorded from the Weights & Biases SF Office (I'm here to cover Google IO next week!) Let’s dig in—because what looks like a slow week on the surface was anything but dull under the hood (TL'DR and show notes at the end as always) Big Companies & APIs Why does XAI Grok talk about White Genocide and "Kill the boar"?? Just after we're getting over the chatGPT glazing incident [https://sub.thursdai.news/p/thursdai-may-1-qwen-3-phi-4-openai] , folks started noticing that @grok - XAI's frontier LLM that is also responding to X replies, started talking about White Genocide in South Africa and something called "Kill the boer" with no reference to any of these things in the question! Since we recorded the episode, XAI official X account posted that an "unauthorized modification" happened to the system prompt, and that going forward they would open source all the prompts (and they did [https://github.com/xai-org/grok-prompts]). Whether or not they would keep updating that repository though, remains unclear (see the "open sourced" x algorithm to which the last push was over a year ago, or the promised Grok 2 that was never open sourced) While it's great to have some more clarity from the Xai team, this behavior raises a bunch of questions about the increasing roles of AI's in our lives and the trust that many folks are giving them. Adding fuel to the fire, are Uncle Elon's recent tweets that are related to South Africa, and this specific change seems to be related to those views at least partly. Remember also, Grok was meant as "maximally truth seeking" AI! I really hope this transparency continues! Open Source LLMs: The Decentralization Tsunami AM-Thinking v1: Dense Reasoning, SOTA Math, Single-Checkpoint Deployability Open source starts with the kind of progress that would have been unthinkable 18 months ago: a 32B dense LLM, openly released, that takes on the big mixture-of-experts models and comes out on top for math and code. AM-Thinking v1 [https://huggingface.co/a-m-team/AM-Thinking-v1] (paper here [https://arxiv.org/abs/2505.08311]) hits 85.3% on AIME 2024, 70.3% on LiveCodeBench v5, and 92.5% on Arena-Hard. It even runs at 25 tokens/sec on a single 80GB GPU with INT4 quantization. The model supports a /think reasoning toggle (chain-of-thought on demand), comes with a permissive license, and is fully tooled for vLLM, LM Studio, and Ollama. Want to see where dense models can still push the limits? This is it. And yes, they’re already working on a multilingual RLHF pass and 128k context window. Personal note: We haven’t seen this kind of “out of nowhere” leaderboard jump since the early days of Qwen or DeepSeek. This company's debut on HuggingFace with a model that crushes! Decentralized LLM Training: Nous Research Psyche & Prime Intellect INTELLECT-2 This week, open source LLMs didn’t just mean “here are some weights.” It meant distributed, decentralized, and—dare I say—permissionless AI. Two labs stood out: Nous Research launches Psyche Dylan Rolnick from Nous Research joined the show to explain Psyche [https://nousresearch.com/nous-psyche/]: a Rust-powered, distributed LLM training network where you can watch a 40B model (Consilience-40B) evolve in real time, join the training with your own hardware, and even have your work attested on a Solana smart contract. The core innovation? DisTrO (Decoupled Momentum) which we covered back in December [https://sub.thursdai.news/p/thursdai-dec-4-openai-o1-and-o1-pro] that drastically compresses the gradient exchange so that training large models over the public internet isn’t a pipe dream—it’s happening right now. Live dashboard here [https://psyche.network/runs/consilience-40b-1/0], open codebase, and the testnet already humming with early results. This massive 40B attempt is going to show whether distributed training actually works! The cool thing about their live dashboard is, it's WandB behind the scenes, but with a very thematic and cool Nous Research reskin! This model saves constant checkpoints to the hub as well, so the open source community can enjoy a full process of seeing a model being trained! Prime Intellect INTELLECT-2 Not to be outdone, Prime Intellect’s INTELLECT-2 [https://www.primeintellect.ai/blog/intellect-2-release] released a globally decentralized, 32B RL-trained reasoning model, built on a permissionless swarm of GPUs. Using their own PRIME-RL framework, SHARDCAST checkpointing, and an LSH-based rollout verifier, they’re not just releasing a model—they’re proving it’s possible to scale serious RL outside a data center. OpenAI's HealthBench: Can LLMs Judge Medical Safety? One of the most intriguing drops of the week is HealthBench [https://openai.com/index/healthbench/], a physician-crafted benchmark for evaluating LLMs in clinical settings. Instead of just multiple-choice “gotcha” tests, HealthBench brings in 262 doctors from 60 countries, 26 specialties, and nearly 50 languages to write rubrics for 5,000 realistic health conversations. The real innovation: LLM as judge. Models like GPT-4.1 are graded against physician-written rubrics, and the agreement between model and human judges matches the agreement between two doctors. Even the “mini” variants of GPT-4.1 are showing serious promise—faster, cheaper, and (on the “Hard” subset) giving the full-size models a run for their money. Other Open Source Standouts Falcon-Edge: Ternary BitNet for Edge Devices The Falcon-Edge project [https://falcon-lm.github.io/blog/falcon-edge/] brings us 1B and 3B-parameter language models trained directly in ternary BitNet format (weights constrained to -1, 0, 1), which slashes memory and compute requirements and enables inference on <1GB VRAM. If you’re looking to fine-tune, you get pre-quantized checkpoints and a clear path to 1-bit LLMs. StepFun Step1x-3D: Controllable Open 3D Generation StepFun’s 3D pipeline [https://huggingface.co/stepfun-ai/Step1X-3D] is a two-stage system that creates watertight geometry and then view-consistent textures, trained on 2M curated meshes. It’s controllable by text, images, and style prompts—and it’s fully open source, including a huge asset dataset. Big Company LLMs & APIs: Models, Modes, and Model Zoo Confusion GPT-4.1 Comes to ChatGPT: Model Zoo Mayhem OpenAI’s GPT-4.1 series—previously API-only—is now available in the ChatGPT interface. Why does this matter? Because the UX of modern LLMs is, frankly, a mess: seven model options in the dropdown, each with its quirks, speed, and context length. Most casual users don’t even know the dropdown exists. “Alex, ChatGPT is broken!” Actually, you just need to pick a different model. The good news: 4.1 is fast, great at coding, and in many tasks, preferable to the “reasoning” behemoths. My advice (and you can share this with your relatives): when in doubt, just switch the model. Bonus: The long-promised million-token context window is here (sort of)—except in the UI, where it’s more like 128k and sometimes silently truncated. My weekly rant: transparency, OpenAI. ProTip: If you’re hitting invisible context limits, try pasting your long transcripts on the web, not in the Mac app. Don’t trust the UI! AlphaEvolve: DeepMind’s Gemini-Powered Algorithmic Discovery AlphaEvolve [https://deepmind.google/discover/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/] is the kind of project that used to sound like AGI hype—and now it’s just a Tuesday at DeepMind. By pairing Gemini Flash and Gemini Pro in an evolutionary search loop to improve algorithms! This is like, real innovation and it's done with existing models which is super super cool! AlphaEvolve uses a combination of Gemini Flash (for breadth of ideas) and Gemini Pro (for depth and refinement) in an evolutionary loop. It generates, tests, and mutates code to invent faster algorithms. And it's already yielding incredible results: * It discovered a new scheduling heuristic for Google's Borg system, resulting in a 0.7% global compute recovery. That's massive at Google's scale. * It improved a matrix-multiply kernel by 23%, which in turn led to a 1% shorter Gemini training time. As Nisten said, the model basically paid for itself! Perhaps most impressively, it found a 48-multiplication algorithm for 4x4 complex matrices, beating the famous Strassen algorithm from 1969 (which used 49 multiplications). This is AI making genuine, novel scientific discoveries. AGI in the garden, anyone? If you still think LLMs are “just glorified autocomplete,” it’s time to update your mental model. This is model-driven algorithmic discovery, and it’s already changing the pace of hardware, math, and software design. The only downside: it’s not public yet, but there’s an interest form [https://deepmind.google/discover/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/] if you want to be a tester. This Week's Buzz - Everything W&B! It's a busy time here at Weights & Biases, and I'm super excited about a couple of upcoming events where you can connect with us and the broader AI community. Fully Connected: Our very own 2-day conference is happening June 18-19 in San Francisco! We've got an amazing lineup of speakers, including Varun Mohan from WindSurf (formerly Codeium), Heikki Kubler from CoreWeave, our CEO Lucas Bewald, CTO Shawn Lewis, Joe Spizak from Meta, and a keynote from Javi Soltero, VP Product AI at Google. It's going to be packed with insights on building and scaling AI. And because you're a ThursdAI listener, you can get in for FREE with the promo code WBTHURSAI at fullyconnected.com [http://fullyconnected.com]. Don't miss out! AI.Engineer [https://ti.to/software-3/ai-engineer-worlds-fair-2025/discount/THANKSTHURSDAI] World's Fair: This has become THE conference for AI engineers, and W&B is a proud sponsor for the third year running! It's happening in San Francisco from June 3rd to 5th. I'll be speaking there on MCP Observability with Ben from LangChain on June 4th.Even more exciting, ThursdAI will be broadcasting LIVE from the media booth at AI.Engineer [https://ti.to/software-3/ai-engineer-worlds-fair-2025/discount/THANKSTHURSDAI] on June 5th! Come say hi! Tickets are flying, but we've got a special discount for you: use promo code THANKSTHURSDAI for 30% off your ticket here [https://ti.to/software-3/ai-engineer-worlds-fair-2025/discount/THANKSTHURSDAI]. Yam Peleg even decided on the show he's coming after hearing about it! It's going to be an incredible week in SF. P.S - yes, on both websites, there's a video playing and I waited till I showed up to snag a screenshot. This way, you know if you're reading this, this is still Alex the human, no AI is going to do this silly thing 😅 Vision & Video: Open Source Shines Through the Noise We had a bit of a meta-discussion on the show about "video model fatigue" – with so many incremental updates, it can be hard to keep track or see the big leaps. However, when a release like Alibaba's Wan 2.1 comes along, it definitely cuts through. Wan 2.1: Alibaba's Open-Source Diffusion-Transformer Video Suite (try it [https://wan.video/wanxiang/videoCreation]) Alibaba, the team behind the excellent Qwen LLMs, released Wan 2.1, a full stack of open-source text-to-video foundation models. This includes a 1.3B "Nano" version and a 14B "Full" version, both built on a diffusion-transformer (DiT) backbone with a custom VAE. What makes Wan 2.1 stand out is its comprehensive nature. It covers a wide range of tasks: text-to-video, image-to-video, in-painting, instruction editing, reference subject consistency, personalized avatars, and style transfer. Many of these are hard to do well, especially in open source. Nisten was particularly excited about the potential for creating natural, controllable avatars in real-time. While it might not be at the level of specialized commercial tools like HeyGen or Google's Veo just yet, having this capability open-sourced is a massive enabler for the community. You can find the models on Hugging Face [https://huggingface.co/Wan-AI] and the code on GitHub [https://github.com/Wan-Video/Wan2.1]. LTX Turbo: Near Real-Time Video Briefly mentioned, but LTX Turbo was also released. This is a quantized version of LTX (which we've covered before) and can run almost in real-time on H100s. Real-time AI video generation is getting closer! StepFun Step1X-3D: High-Fidelity 3D Asset Generation StepFun released Step1X-3D, an open two-stage framework for generating textured 3D assets. It first synthesizes geometry and then generates view-consistent textures. They've also released a curated dataset of 800K assets. The weights, data, and code are all open, which is great for the 3D AI community. Wrapping Up This "Chill" Week So, there you have it – another "chill" week in the world of AI! From Grok's controversial escapades to the inspiring decentralized training efforts and mind-bending algorithmic discoveries, it's clear the pace isn't slowing down. Next week is going to be absolutely insane. We've got Google I/O and Microsoft Build, and you just know OpenAI or Anthropic (or both!) will try to steal some thunder. Rest assured, we'll be here on ThursdAI to cover all the madness. A huge thank you to my co-hosts Yam, LDJ, and Nisten, and to Dillon Rolnick for joining us. And thanks to all of you for tuning in! TL;DR and show notes * Fully Connected - Weights & Biases premier conference - register HERE with coupon WBTHURSAI * AI Engineer - THANKSTHURSDAI 30% off coupon - register HERE [https://ti.to/software-3/ai-engineer-worlds-fair-2025/discount/THANKSTHURSDAI] * Hosts and Guests * Alex Volkov - AI Evangelist & Weights & Biases (@altryne [http://x.com/@altryne]) * Co Hosts - @yampeleg [http://x.com/@yampeleg] @nisten [http://x.com/@nisten] @ldjconfirmed [http://x.com/@ldjconfirmed]) * Guest - Dillon Rolnick - COO Nous Research (@dillonRolnick [https://x.com/DillonRolnick]) Open Source LLMs * AM-Thinking v1: 32B dense reasoning model ( HF [https://huggingface.co/a-m-team/AM-Thinking-v1], Paper [https://arxiv.org/abs/2505.08311], Page [https://a-m-team.github.io/am-thinking-v1/] ) * Falcon-Edge: ternary BitNet LLMs for edge deployment( Blog [https://falcon-lm.github.io/blog/falcon-edge/], HF-1B [https://huggingface.co/tiiuae/Falcon-E-1B-Base], HF-3B [https://huggingface.co/tiiuae/Falcon-E-3B-Base] ) * Nous Research Psyche: decentralized cooperative-training network from Nous Research ( Website [https://nousresearch.com/nous-psyche/], GitHub [https://github.com/NousResearch/psyche], Tweet [https://x.com/NousResearch/status/1922744494002405444], Dashboard [https://psyche.network/runs/consilience-40b-1/0] ) * INTELLECT-2: globally decentralized RL training of a 32B reasoning model ( Blog [https://www.primeintellect.ai/blog/intellect-2-release], Tech report [https://primeintellect.ai/intellect-2], HF weights [https://huggingface.co/PrimeIntellect/INTELLECT-2], PRIME-RL code [https://github.com/primeintellect/prime-rl] ) * Our coverage of Intellect-1 back in Dec (https://sub.thursdai.news/p/thursdai-dec-4-openai-o1-and-o1-pro [https://sub.thursdai.news/p/thursdai-dec-4-openai-o1-and-o1-pro]) * HealthBench: OpenAI’s physician-crafted benchmark for AI in healthcare ( Blog [https://openai.com/index/healthbench/], Paper [https://cdn.openai.com/pdf/bd7a39d5-9e9f-47b3-903c-8b847ca650c7/healthbench_paper.pdf], Code [https://github.com/openai/simple-evals] ) * Big CO LLMs + APIs * OpenAI adds GPT 4.1 models in chatGPT * AlphaEvolve: Gemini-powered coding agent for algorithm discovery ( Blog [https://deepmind.google/discover/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/] ) * Google shutting off free Gemini 2.5 Pro API due to "demand" ahead of IO * ByteDance - Seed-1.5-VL-thinking 20B (Paper [https://github.com/ByteDance-Seed/Seed1.5-VL/blob/main/Seed1.5-VL-Technical-Report.pdf]) * Anthropic Web Search API: real-time retrieval for Claude models ( Blog [https://www.anthropic.com/news/web-search-api] ) * What's up with Grok? * Vision & Video * Wan 2.1: open-source diffusion-transformer video suite( HF [https://huggingface.co/Wan-AI], GitHub [https://github.com/Wan-Video/Wan2.1], Tweet [https://x.com/Alibaba_Wan/status/1922655324919779604] ) * LTX distilled - near real time video (X [https://x.com/yoavhacohen/status/1922674340081897977]) * Voice & Audio * Haulio - MiniMax Speech tech report is out - best TTS out there (Paper [https://arxiv.org/abs/2505.07916]) * Stability AI - Stable Audio Open Small 341M: on-device text-to-audio (X [https://x.com/jordiponsdotme/status/1922680538197881055], Blog [https://stability.ai/news/stability-ai-and-arm-release-stable-audio-open-small-enabling-real-world-deployment-for-on-device-audio-control], Paper [https://arxiv.org/abs/2505.08175], HF [https://huggingface.co/stabilityai/stable-audio-open-small] ) * AI Art & Diffusion & 3D * StepFun Step1x-3D - Towards High-Fidelity and ControllableGeneration of Textured 3D Assets (HF [https://huggingface.co/stepfun-ai/Step1X-3D], Demo [https://huggingface.co/spaces/stepfun-ai/Step1X-3D], Dataset [https://huggingface.co/datasets/stepfun-ai/Step1X-3D-obj-data/tree/main], report [https://huggingface.co/stepfun-ai/Step1X-3D]) * Tools & Others notable AI things mentioned on the pod * The robots are dancing! (X [https://x.com/simonkalouche/status/1922489999032832058]) 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 [https://sub.thursdai.news/subscribe?utm_medium=podcast&utm_campaign=CTA_2]

Hey folks, Alex here (yes, real me, not my AI avatar, yet) Compared to previous weeks, this week was pretty "chill" in the world of AI, though we did get a pretty significant Gemini 2.5 Pro update, it basically beat itself on the Arena. With Mistral releasing a new medium model (not OSS) and Nvidia finally dropping Nemotron Ultra (both ignoring Qwen 3 performance) there was also a few open source updates. To me the highlight of this week was a breakthrough in AI Avatars, with Heygen's new IV model, Beating ByteDance's OmniHuman (our coverage [https://sub.thursdai.news/i/156643204/bytedance-omnihuman-a-reality-bending-mind-breaking-imghuman-model]) and Hedra labs, they've set an absolute SOTA benchmark for 1 photo to animated realistic avatar. Hell, Iet me record all this real quick and show you how good it is! How good is that?? I'm still kind of blown away. I have managed to get a free month promo code for you guys, look for it in the TL;DR section at the end of the newsletter. Of course, if you’re rather watch than listen or read, here’s our live recording on YT OpenSource AI NVIDIA's Nemotron Ultra V1: Refining the Best with a Reasoning Toggle 🧠 NVIDIA also threw their hat further into the ring with the release of Nemotron Ultra V1, alongside updated Super and Nano versions. We've talked about Nemotron before – these are NVIDIA's pruned and distilled versions of Llama 3.1, and they've been impressive. The Ultra version is the flagship, a 253 billion parameter dense model (distilled and pruned from Llama 3.1 405B), and it's packed with interesting features. One of the coolest things is the dynamic reasoning toggle. You can literally tell the model "detailed thinking on" or "detailed thinking off" via a system prompt during inference. This is something Qwen also supports, and it looks like the industry is converging on this idea of letting users control the "depth" of thought, which is super neat. Nemotron Ultra boasts a 128K context window and, impressively, can fit on a single 8xH100 node thanks to Neural Architecture Search (NAS) and FFN-Fusion. And performance-wise, it actually outperforms the Llama 3 405B model it was distilled from, which is a big deal. NVIDIA shared a chart from Artificial Analysis (dated April 2025, notably before Qwen3's latest surge) showing Nemotron Ultra standing strong among models like Gemini 2.5 Flash and Opus 3 Mini. What's also great is NVIDIA's commitment to openness here: they've released the models under a commercially permissive NVIDIA Open Model License, the complete post-training dataset (Llama-Nemotron-Post-Training-Dataset), and their training codebases (NeMo, NeMo-Aligner, Megatron-LM). This allows for reproducibility and further community development. Yam Peleg pointed out the cool stuff they did with Neural Architecture Search to optimally reduce parameters without losing performance. Absolute Zero: AI Learning to Learn, Zero (curated) Data Required! (Arxiv [https://arxiv.org/abs/2505.03335]) LDJ brought up a fascinating paper that ties into this theme of self-improvement and reinforcement learning: "Absolute Zero: Reinforced Self-play Reasoning with Zero Data" from Andrew Zhao (Tsinghua University) and a few others The core idea here is a system that self-evolves its training curriculum and reasoning ability. Instead of needing a pre-curated dataset of problems, the model creates the problems itself (e.g., code reasoning tasks) and then uses something like a Code Executor to validate its proposed solutions, serving as a unified source of verifiable reward. It's open-ended yet grounded learning. By having a verifiable environment (code either works or it doesn't), the model can essentially teach itself to code without external human-curated data. The paper shows fine-tunes of Qwen models (like Qwen Coder) achieving state-of-the-art results on benchmarks like MBBP and AIME (Math Olympiad) with no pre-existing data for those problems. The model hallucinates questions, creates its own rewards, learns, and improves. This is a step beyond synthetic data, where humans are still largely in charge of generation. It's wild, and it points towards a future where AI systems could become increasingly autonomous in their learning. Big Companies & APIs Google dropped another update to their Gemini 2.5 Pro, this time the "IO edition" preview, specifically touting enhanced coding performance. This new version jumped to the #1 spot on WebDev Arena (a benchmark where human evaluators choose between two side-by-side code generations in VS Code), with a +147 Elo point gain, surpassing Claude 3.7 Sonnet. It also showed improvements on benchmarks like LiveCodeBench (up 7.39%) and Aider Polyglot (up ~3-6%). Google also highlighted its state-of-the-art video understanding (84.8% on VideoMME) with examples like generating code from a video of an app. Which essentially lets you record a drawing of how your app interaction will happen, and the model will use that video instructions! It's pretty cool. Though, not everyone was as impressed, folks noted that while gaining in a few evals, this model also regressed in several others including Vibe-Eval (Reka's multimodal benchmark), Humanity's Last Exam, AIME, MMMU, and even long context understanding (MRCR). It's a good reminder that model updates often involve trade-offs – you can't always win at everything. BREAKING: Gemini's Implicit Caching - A Game Changer for Costs! 💰 Just as we were wrapping up this segment on the show, news broke that Google launched implicit caching in Gemini APIs! This is a huge deal for developers. Previously, Gemini offered explicit caching, where you had to manually tell the API what context to cache – a bit of a pain. Now, with implicit caching, the system automatically enables up to 75% cost savings when your request hits a cache. This is fantastic, especially for long-context applications, which is where Gemini's 1-2 million token context window really shines. If you're repeatedly sending large documents or codebases, this will significantly reduce your API bills. OpenAI has had automatic caching for a while, and it's great to see Google matching this for a much better developer experience and cost-effectiveness. It also saves Google a ton on inference, so it's a win-win! Mistral Medium 3: The Closed Turn 😥 Mistral, once the darling of the open-source community for models like Mistral 7B and Mixtral, announced Mistral Medium 3. The catch? It's not open source. They're positioning it as a multimodal frontier model with 128K context, claiming it matches or surpasses GPT-4-class benchmarks while being cheaper (priced at $0.40/M input and $2/M output tokens). However they haven't added Gemini Flash 2.5 here, which is 70% cheaper while being faster as well, nor did they mention Qwen. Nisten voiced a sentiment many in the community share: he used to use LeChat frequently because he knew and understood the underlying open-source models. Now, with a closed model, it's a black box. It's a bit like pirating music users often being the biggest buyers – understanding the open model often leads to more commercial usage. Wolfram offered a European perspective, noting that Mistral, as a European company, might have a unique advantage with businesses concerned about GDPR and data sovereignty, who might be hesitant to use US or Chinese cloud APIs. For them, a strong European alternative, even if closed, could be appealing. OpenAI's New Chapter: Restructuring for the Future OpenAI announced an evolution in its corporate structure. The key points are: * The OpenAI non-profit will continue to control the entire organization. * The existing for-profit LLC will become a Public Benefit Corporation (PBC). * The non-profit will be a significant owner of the PBC and will control it. * Both the non-profit and PBC will continue to share the same mission: ensuring AGI benefits all of humanity. This move seems to address some of the governance concerns that have swirled around OpenAI, particularly in light of Elon Musk's lawsuit regarding its shift from a non-profit to a capped-profit entity. LDJ explained that the main worry for many was whether the non-profit would lose control or its stake in the main research/product arm. This restructuring appears to ensure the non-profit remains at the helm and that the PBC is legally bound to the non-profit's mission, not just investor interests. It's an important step for a company with such a profound potential impact on society. And in related OpenAI news, the acquisition of Windsurf (the VS Code fork) for a reported $3 billion went through, while Cursor (another VS Code fork) announced a $9 billion valuation. It's wild to see these developer tools, which are essentially forks with an AI layer, reaching such massive valuations. Microsoft's hand is in all of this too – investing in OpenAI, invested in Cursor, owning VS Code, and now OpenAI buying Windsurf. It's a tangled web! Finally, a quick mention that Sam Altman (OpenAI), Lisa Su (AMD), Mike Intrator (CoreWeave - my new CEO!), and folks from Microsoft were testifying before the U.S. Senate today about how to ensure America leads in AI and what innovation means. These conversations are crucial as AI continues to reshape our world. This Weeks Buzz - Come Vibe with Us at Fully Connected! (SF, June 18-19) 🎉 Our two-day conference, Fully Connected, is happening in San Francisco on June 18th and 19th, and it's going to be awesome! We've got an incredible lineup of speakers, including Joe Spizak from the Llama team at Meta and Varun from Windsurf. It's two full days of programming, learning, and connecting with folks at the forefront of AI. And because you're part of the ThursdAI family, I've got a special promo code for you: use WBTHURSAI to get a free ticket on me! If you're in or around SF, I'd love to see you there. Come hang out, learn, and vibe with us! Register at fullyconnected.com [https://www.google.com/url?sa=E&q=http%3A%2F%2Ffullyconnected.com] Hackathon Update: Moved to July! 🗓️ The AGI Evals & Agentic Tooling (A2A) + MCP Hackathon that I was super excited to co-host has been postponed to July 12th-13th. Mark your calendars! I'll share more details and the invite soon. W&B Joins CoreWeave! A New Era Begins! 🚀 And the big personal news for me and the entire Weights & Biases team: the acquisition of Weights & Biases by CoreWeave has been completed! CoreWeave is the ultra-fast-growing provider of GPUs that powers so much of the AI ecosystem. So, from now on, it's Alex Volkov, AI Evangelist at Weights & Biases, from CoreWeave! (And as always, the opinions I share here are my own and not necessarily those of CoreWeave, especially important now that they're a public company!). I'm incredibly excited about this new chapter. W&B isn't going anywhere as a product; if anything, this will empower us to build even better developer tooling and integrate more deeply to help you run your models wherever you choose. Expect more cool stuff to come, especially as I figure out where all those spare GPUs are lying around at CoreWeave! 😉 Vision & Video AI Avatars SOTA with HeyGen IV Ok, as you saw above, the HeyGen IV avatars are absolutely bonkers. I did a comparison thread [https://x.com/altryne/status/1919866852031004880] on X, and HeyGen's new thing absolutely takes SOTA between ByteDance OmniHuman and Hedra Labs! All you need to do is upload 1 image of yourself, can even be an AI generated image, can be a side profile, can be a dog, an Anime character and they will generate up to 30 seconds of incredible lifelike avatar with the audio you provide! I was so impressed with this, I reached out to HeyGen and scored a 1 month free code for you all, use THURSDAY4 and get a free month to try it out. Please tag me in whatever you create if you publish, I'd love to see where you take this! Quick Hits: Lightricks LTXV & HunyuanCustom Briefly, on the open-weights video front: * Lightricks LTXV 13B: The company from Jerusalem released an upgraded 13 billion parameter version of their LTX video model. It requires more VRAM but offers higher quality, keyframe and character movement support, multi-shot support, and multi-keyframe conditioning (a feature Sora famously has). It's fully open and supports LoRAs for custom styles. * HunyuanCustom: From Tencent, this model is about to be released (GitHub/Hugging Face links were briefly up then down). It promises multi-modal, subject-consistent video generation without LoRAs, based on a subject you provide (image, and eventually video/audio). It can take an image of a person or object and generate a video with that subject consistently. They also teased audio conditioning – making an avatar sing or speak based on input audio – and even style transfer where you can replace a character in a video with another reference image, all looking very promising for open source. The World of AI Audio Just a couple of quick mentions in the audio space: * ACE-Step 3.5B: From StepFun, this is a 3.5 billion parameter, fully open-source (Apache-2.0) foundation model for music generation. It uses a diffusion-based approach and can synthesize up to 4 minutes of music in just 20 seconds on an A100 GPU. It's not quite at Suno/Udio levels yet, but it's a strong open-source contender. * NVIDIA Parakeet TDT 0.6B V2: NVIDIA released this 600 million parameter transcription model that is blazing fast. It can transcribe 60 minutes of audio in just one second on production GPUs and works well locally too. It currently tops the OpenASR leaderboard on Hugging Face for English transcription and is a very strong Whisper competitor, especially for speed. Conclusion and TL;DR * Hosts and Guests * Alex Volkov - AI Evangelist & Weights & Biases (@altryne [http://x.com/@altryne]) * Co Hosts - @WolframRvnwlf [http://x.com/@WolframRvnwlf] @yampeleg [x.com/@yampeleg] @nisten [http://x.com/@nisten] @ldjconfirmed [http://x.com/@ldjconfirmed)] * Open Source LLMs * Wolfram's Qwen3 evals (X [https://x.com/Presidentlin], Github [https://github.com/WolframRavenwolf/MMLU-Pro]) * NVIDIA - Nemotron Ultra V1 (+ updated Super & Nano) (HF [https://huggingface.co/collections/nvidia/llama-nemotron-67d92346030a2691293f200b]) * Cognition Kevin-32B = K(ernel D)evin - RL for writing CUDA kernels (Blog [https://cognition.ai/blog/kevin-32b], HF [https://huggingface.co/cognition-ai/Kevin-32B]) * Absolute Zero: Reinforced Self-play Reasoning with Zero Data (ArXiv [https://arxiv.org/abs/2505.03335]) * Big CO LLMs + APIs * Gemini Pro 2.5 IO tops ... Gemini 2.5 as the top LLM (Blog [https://developers.googleblog.com/en/gemini-2-5-pro-io-improved-coding-performance/]) * Mistral Medium 3 - (Blog [https://mistral.ai/news/mistral-medium-3] | X [https://x.com/MistralAI/status/1920119463430500541] ) * Figma announces Figma Make - Bolt/Lovable competitors (Figma [https://www.figma.com/make/]) * OpenAI Restructures: Nonprofit Keeps Control, LLC Becomes PB (Blog [https://openai.com/index/evolving-our-structure/]) * Cursor worth $9B while Windsurf sells to OpenAI at $3B * Sam Altman, Lisa Su, Mike Intrator testify in Senate (Youtube [https://www.youtube.com/watch?v=jOqTg1W_F5Q]) * This weeks Buzz * Fully Connected: W&B's 2-day conference, June 18-19 in SF fullyconnected.com [fullyconnected.com] - Promo Code WBTHURSAI * Hackathon moved to July 12-13 * Vision & Video * Lightricks a new "open weights" LTXV 13B ( LTX Studio [https://ltx.studio/purchase/v1/ltx_studio/default/login?redirectAfterLogin=https%253A%252F%252Fapp.ltx.studio%252Fmotion-workspace], HF [https://huggingface.co/Lightricks/LTX-Video]) * HeyGen Avatar IV - SOTA digital avatars - 1 month for free with THURSDAY4 (X [https://x.com/HeyGen_Official/status/1919824467821551828], HeyGen [http://heygen.com]) * HunyuanCustom - multi-modal subject-consistent video generation model (Examples [https://hunyuancustom.github.io/], Github [https://github.com/Tencent/HunyuanCustom], HF [https://huggingface.co/tencent/HunyuanCustom]) * Voice & Audio * ACE-Step 3.5B: open-source foundation model for AI music generation (project [https://ace-step.github.io/]) * Nvidia - Parakeet TDT 0.6B V2 - transcribe 60 minutes of audio in just 1 second (HF [https://huggingface.co/nvidia/parakeet-tdt-0.6b-v2], Demo [https://huggingface.co/spaces/nvidia/parakeet-tdt-0.6b-v2]) So, there you have it – a "chill" week that still managed to deliver some incredible advancements, particularly in AI avatars with HeyGen, continued strength in open-source models like Qwen3, and Google's relentless push with Gemini. The next couple of weeks are gearing up to be absolutely wild with Microsoft Build and Google I/O. I expect a deluge of announcements, and you can bet we'll be here on ThursdAI to break it all down for you. Thanks to Yam, Wolfram, LDJ, and Nisten for their insights on the show, and thanks to all of you for tuning in, reading, and being part of this amazing community. We stay up to date so you don't have to! Catch you next week!Cheers,Alex 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 [https://sub.thursdai.news/subscribe?utm_medium=podcast&utm_campaign=CTA_2]