Episodios

  • ThursdAI - Jan 1 2026 - Will Brown Interview + Nvidia buys Groq, Meta buys Manus, Qwen Image 2412 & Alex New Year greetings

    ThursdAI - Jan 1 2026 - Will Brown Interview + Nvidia buys Groq, Meta buys Manus, Qwen Image 2412 & Alex New Year greetings
    Jan 1 2026

    Hey all,

    Happy new year! This is Alex, writing to you for the very fresh start of this year, it’s 2026 already, can you believe it?

    There was no live stream today, I figured the cohosts deserve a break and honestly it was a very slow week. Even the chinese labs who don’t really celebrate X-mas and new years didn’t come out with a banger AFAIK.

    ThursdAI - AI moves fast, we’re here to make sure you never miss a thing! Subscribe :)

    Tho I thought it was an incredible opportunity to finally post the Will Brow interview I recorded in November during the AI Engineer conference.

    Will is a researcher at Prime Intellect (big fans on WandB btw!) and is very known on X as a hot takes ML person, often going viral for tons of memes!

    Will is the creator and maintainer of the Verifiers library (Github) and his talk at AI Engineer was all about RL Environments (what they are, you can hear in the interview, I asked him!)

    TL;DR last week of 2025 in AI

    Besides this, my job here is to keep you up to date, and honestly this was very easy this week, as… almost nothing has happened, but here we go:

    Meta buys Manus

    The year ended with 2 huge acquisitions / aquihires. First we got the news from Alex Wang that Meta has bought Manus.ai which is an agentic AI startup we covered back in March for an undisclosed amount (folks claim $2-3B)

    The most interesting thing here is that Manus is a Chinese company, and this deal requires very specific severance from Chinese operations.

    Jensen goes on a new years spending spree, Nvidia buys Groq (not GROK) for $20B

    Groq which we covered often here, and are great friends, is going to NVIDIA, in a… very interesting acqui-hire, which is a “non binding license” + most of Groq top employees apparently are going to NVIDIA. Jonathan Ross the CEO of Groq, was the co-creator of the TPU chips at Google before founding Groq, so this seems like a very strategic aquihire for NVIDIA! Congrats to our friends from Groq on this amazing news for the new year!

    Tencent open-sources HY-MT1.5 translation models with 1.8B edge-deployable and 7B cloud variants supporting 33 languages (X, HF, HF, GitHub)

    It seems that everyone’s is trying to de-throne whisper and this latest attempt from Tencent is a interesting one. a 1.8B and 7B translation models with very interesting stats.

    Alibaba’s Qwen-Image-2512 drops on New Year’s Eve as strongest open-source text-to-image model, topping AI Arena with photorealistic humans and sharper textures (X, HF, Arxiv)

    Our friends in Tongyi decided to give is a new years present in the form of an updated Qwen-image, with much improved realism

    That’s it folks, this was a quick one, hopefully you all had an amazing new year celebration, and are gearing up to an eventful and crazy 2026.

    I wish you all happiness, excitement and energy to keep up with everything in the new year, and will make sure that we’re here to keep you up to date as always!

    P.S - I got a little news of my own this yesterday, not related to AI. She said yes 🎉



    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
    Más Menos
    30 m
  • 🔥 Someone Trained an LLM in Space This Year (And 50 Other Things You Missed)- ThursdAI yearly recap is here!

    🔥 Someone Trained an LLM in Space This Year (And 50 Other Things You Missed)- ThursdAI yearly recap is here!
    Dec 25 2025
    Ho Ho Ho, Alex here! (a real human writing these words, this needs to be said in 2025) Merry Christmas (to those who celebrate) and welcome to the very special yearly ThursdAI recap! This was an intense year in the world of AI, and after 51 weekly episodes (this is episode 52!) we have the ultimate record of all the major and most important AI releases of this year! So instead of bringing you a weekly update (it’s been a slow week so far, most AI labs are taking a well deserved break, the Cchinese AI labs haven’t yet surprised anyone), I’m dropping a comprehensive yearly AI review! Quarter by quarter, month by month, both in written form and as a pod/video! Why do this? Who even needs this? Isn’t most of it obsolete? I have asked myself this exact question while prepping for the show (it was quite a lot of prep, even with Opus’s help). I eventually landed on, hey, if nothing else, this will serve as a record of the insane week of AI progress we all witnessed. Can you imagine that the term Vibe Coding is less than 1 year old? That Claude Code was released at the start of THIS year? We get hedonicly adapt to new AI goodies so quick, and I figured this will serve as a point in time check, we can get back to and feel the acceleration! With that, let’s dive in - P.S. the content below is mostly authored by my co-author for this, Opus 4.5 high, which at the end of 2025 I find the best creative writer with the best long context coherence that can imitate my voice and tone (hey, I’m also on a break! 🎅) “Open source AI has never been as hot as this quarter. We’re accelerating as f*ck, and it’s only just beginning—hold on to your butts.” — Alex Volkov, ThursdAI Q1 2025🏆 The Big Picture — 2025 - The Year the AI Agents Became RealLooking back at 51 episodes and 12 months of relentless AI progress, several mega-themes emerged:1. 🧠 Reasoning Models Changed EverythingFrom DeepSeek R1 in January to GPT-5.2 in December, reasoning became the defining capability. Models now think for hours, call tools mid-thought, and score perfect on math olympiads.2. 🤖 2025 Was Actually the Year of AgentsWe said it in January, and it came true. Claude Code launched the CLI revolution, MCP became the universal protocol, and by December we had ChatGPT Apps, Atlas browser, and AgentKit.3. 🇨🇳 Chinese Labs Dominated Open SourceDeepSeek, Qwen, MiniMax, Kimi, ByteDance — despite chip restrictions, Chinese labs released the best open weights models all year. Qwen 3, Kimi K2, DeepSeek V3.2 were defining releases.4. 🎬 We Crossed the Uncanny ValleyVEO3’s native audio, Suno V5’s indistinguishable music, Sora 2’s social platform — 2025 was the year AI-generated media became indistinguishable from human-created content.5. 💰 The Investment Scale Became Absurd$500B Stargate, $1.4T compute obligations, $183B valuations, $100-300M researcher packages, LLMs training in space. The numbers stopped making sense.6. 🏆 Google Made a ComebackAfter years of “catching up,” Google delivered Gemini 3, Antigravity, Nano Banana Pro, VEO3, and took the #1 spot (briefly). Don’t bet against Google.By the NumbersQ1 2025 — The Quarter That Changed EverythingDeepSeek R1 crashed NVIDIA’s stock, reasoning models went mainstream, and Chinese labs took over open source. The quarter that proved AI isn’t slowing down—it’s just getting started.Key Themes:* 🧠 Reasoning models went mainstream (DeepSeek R1, o1, QwQ)* 🇨🇳 Chinese labs dominated open source (DeepSeek, Alibaba, MiniMax, ByteDance)* 🤖 2025 declared “The Year of Agents” (OpenAI Operator, MCP won)* 🖼️ Image generation revolution (GPT-4o native image gen, Ghibli-mania)* 💰 Massive infrastructure investment (Project Stargate $500B)January — DeepSeek Shakes the World(Jan 02 | Jan 10 | Jan 17 | Jan 24 | Jan 30)The earthquake that shattered the AI bubble. DeepSeek R1 dropped on January 23rd and became the most impactful open source release ever:* Crashed NVIDIA stock 17% — $560B loss, largest single-company monetary loss in history* Hit #1 on the iOS App Store* Cost allegedly only $5.5M to train (sparking massive debate)* Matched OpenAI’s o1 on reasoning benchmarks at 50x cheaper pricing* The 1.5B model beat GPT-4o and Claude 3.5 Sonnet on math benchmarks 🤯“My mom knows about DeepSeek—your grandma probably knows about it, too” — Alex VolkovAlso this month:* OpenAI Operator — First agentic ChatGPT (browser control, booking, ordering)* Project Stargate — $500B AI infrastructure (Manhattan Project for AI)* NVIDIA Project Digits — $3,000 desktop that runs 200B parameter models* Kokoro TTS — 82M param model hit #1 on TTS Arena, Apache 2, runs in browser* MiniMax-01 — 4M context window from Hailuo* Gemini Flash Thinking — 1M token context with thinking tracesFebruary — Reasoning Mania & The Birth of Vibe Coding(Feb 07 | Feb 13 | Feb 20 | Feb 28)The month that redefined how we work with AI.OpenAI Deep ...
    Más Menos
    1 h y 50 m
  • 📆 ThursdAI - Dec 18 - Gemini 3 Flash, Grok Voice, ChatGPT Appstore, Image 1.5 & GPT 5.2 Codex, Meta Sam Audio & more AI news

    📆 ThursdAI - Dec 18 - Gemini 3 Flash, Grok Voice, ChatGPT Appstore, Image 1.5 & GPT 5.2 Codex, Meta Sam Audio & more AI news
    Dec 19 2025
    Hey folks 👋 Alex here, dressed as 🎅 for our pre X-mas episode!We’re wrapping up 2025, and the AI labs decided they absolutely could NOT let the year end quietly. This week was an absolute banger—we had Gemini 3 Flash dropping with frontier intelligence at flash prices, OpenAI firing off GPT 5.2 Codex as breaking news DURING our show, ChatGPT Images 1.5, Nvidia going all-in on open source with Nemotron 3 Nano, and the voice AI space heating up with Grok Voice and Chatterbox Turbo. Oh, and Google dropped FunctionGemma for all your toaster-to-fridge communication needs (yes, really).Today’s show was over three and a half hours long because we tried to cover both this week AND the entire year of 2025 (that yearly recap is coming next week—it’s a banger, we went month by month and you’ll really feel the acceleration). For now, let’s dive into just the insanity that was THIS week.00:00 Introduction and Overview00:39 Weekly AI News Highlights01:40 Open Source AI Developments01:44 Nvidia's Nemotron Series09:09 Google's Gemini 3 Flash19:26 OpenAI's GPT Image 1.520:33 Infographic and GPT Image 1.5 Discussion20:53 Nano Banana vs GPT Image 1.521:23 Testing and Comparisons of Image Models23:39 Voice and Audio Innovations24:22 Grok Voice and Tesla Integration26:01 Open Source Robotics and Voice Agents29:44 Meta's SAM Audio Release32:14 Breaking News: Google Function Gemma33:23 Weights & Biases Announcement35:19 Breaking News: OpenAI Codex 5.2 MaxTo receive new posts and support my work, consider becoming a free or paid subscriber.Big Companies LLM updatesGoogle’s Gemini 3 Flash: The High-Speed Intelligence KingIf we had to title 2025, as Ryan Carson mentioned on the show, it might just be “The Year of Google’s Comeback.” Remember at the start of the year when we were asking “Where is Google?” Well, they are here. Everywhere.This week they launched Gemini 3 Flash, and it is rightfully turning heads. This is a frontier-class model—meaning it boasts Pro-level intelligence—but it runs at Flash-level speeds and, most importantly, Flash-level pricing. We are talking $0.50 per 1 million input tokens. That is not a typo. The price-to-intelligence ratio here is simply off the charts.I’ve been using Gemini 2.5 Flash in production for a while because it was good enough, but Gemini 3 Flash is a different beast. It scores 71 on the Artificial Analysis Intelligence Index (a 13-point jump from the previous Flash), and it achieves 78% on SWE-bench Verified. That actually beats the bigger Gemini 3 Pro on some agentic coding tasks!What impressed me most, and something Kwindla pointed out, is the tool calling. Previous Gemini models sometimes struggled with complex tool use compared to OpenAI, but Gemini 3 Flash can handle up to 100 simultaneous function calls. It’s fast, it’s smart, and it’s integrated immediately across the entire Google stack—Workspace, Android, Chrome. Google isn’t just releasing models anymore; they are deploying them instantly to billions of users.For anyone building agents, this combination of speed, low latency, and 1 million context window (at this price!) makes it the new default workhorse.Google’s FunctionGemma Open Source releaseWe also got a smaller, quirkier release from Google: FunctionGemma. This is a tiny 270M parameter model. Yes, millions, not billions.It’s purpose-built for function calling on edge devices. It requires only 500MB of RAM, meaning it can run on your phone, in your browser, or even on a Raspberry Pi. As Nisten joked on the show, this is finally the model that lets your toaster talk to your fridge.Is it going to write a novel? No. But after fine-tuning, it jumped from 58% to 85% accuracy on mobile action tasks. This represents a future where privacy-first agents live entirely on your device, handling your calendar and apps without ever pinging a cloud server.OpenAI Image 1.5, GPT 5.2 Codex and ChatGPT AppstoreOpenAI had a busy week, starting with the release of GPT Image 1.5. It’s available now in ChatGPT and the API. The headline here is speed and control—it’s 4x faster than the previous model and 20% cheaper. It also tops the LMSYS Image Arena leaderboards.However, I have to give a balanced take here. We’ve been spoiled recently by Google’s “Nano Banana Pro” image generation (which powers Gemini). When we looked at side-by-side comparisons, especially with typography and infographic generation, Gemini often looked sharper and more coherent. This is what we call “hedonistic adaptation”—GPT Image 1.5 is great, but the bar has moved so fast that it doesn’t feel like the quantum leap DALL-E 3 was back in the day. Still, for production workflows where you need to edit specific parts of an image without ruining the rest, this is a massive upgrade.🚨 BREAKING: GPT 5.2 CodexJust as we were nearing the end of the show, OpenAI decided to drop some breaking news: GPT 5.2 Codex.This is a specialized model optimized ...
    Más Menos
    39 m
  • 📆 ThursdAI - Dec 11 - GPT 5.2 is HERE! Plus, LLMs in Space, MCP donated, Devstral surprises and more AI news!

    📆 ThursdAI - Dec 11 - GPT 5.2 is HERE! Plus, LLMs in Space, MCP donated, Devstral surprises and more AI news!
    Dec 12 2025
    Hey everyone, December started strong and does NOT want to slow down!? OpenAI showed us their response to the Code Red and it’s GPT 5.2, which doesn’t feel like a .1 upgrade! We got it literally as breaking news at the end of the show, and oh boy! The new kind of LLMs is here. GPT, then Gemini, then Opus and now GPT again... Who else feels like we’re on a trippy AI rolercoaster? Just me? 🫨 I’m writing this newsletter from a fresh “traveling podcaster” setup in SF (huge shoutout to the Chroma team for the studio hospitality). P.S - Next week we’re doing a year recap episode (52st episode of the year, what is my life), but today is about the highest-signal stuff that happened this week.Alright. No more foreplay. Let’s dive in. Please subscribe. 🔥 The main event: OpenAI launches GPT‑5.2 (and it’s… a lot)We started the episode with “garlic in the air” rumors (OpenAI holiday launches always have that Christmas panic energy), and then… boom: GPT‑5.2 actually drops while we’re live.What makes this release feel significant isn’t “one benchmark went up.” It’s that OpenAI is clearly optimizing for the things that have become the frontier in 2025: long-horizon reasoning, agentic coding loops, long context reliability, and lower hallucination rates when browsing/tooling is involved.5.2 Instant, Thinking and Pro in ChatGPT and in the APIOpenAI shipped multiple variants, and even within those there are “levels” (medium/high/extra-high) that effectively change how much compute the model is allowed to burn. At the extreme end, you’re basically running parallel thoughts and selecting winners. That’s powerful, but also… very expensive.It’s very clearly aimed at the agentic world: coding agents that run in loops, tool-using research agents, and “do the whole task end-to-end” workflows where spending extra tokens is still cheaper than spending an engineer day.Benchmarks I’m not going to pretend benchmarks tell the full story (they never do), but the shape of improvements matters. GPT‑5.2 shows huge strength on reasoning + structured work.It hits 90.5% on ARC‑AGI‑1 in the Pro X‑High configuration, and 54%+ on ARC‑AGI‑2 depending on the setting. For context, ARC‑AGI‑2 is the one where everyone learns humility again.On math/science, this thing is flexing. We saw 100% on AIME 2025, and strong performance on FrontierMath tiers (with the usual “Tier 4 is where dreams go to die” vibe still intact). GPQA Diamond is up in the 90s too, which is basically “PhD trivia mode.”But honestly the most practically interesting one for me is GDPval (knowledge-work tasks: slides, spreadsheets, planning, analysis). GPT‑5.2 lands around 70%, which is a massive jump vs earlier generations. This is the category that translates directly into “is this model useful at my job.” - This is a bench that OpenAI launched only in September and back then, Opus 4.1 was a “measly” 47%! Talk about acceleration! Long context: MRCR is the sleeper highlightOn MRCR (multi-needle long-context retrieval), GPT‑5.2 holds up absurdly well even into 128k and beyond. The graph OpenAI shared shows GPT‑5.1 falling off a cliff as context grows, while GPT‑5.2 stays high much deeper into long contexts.If you’ve ever built a real system (RAG, agent memory, doc analysis) you know this pain: long context is easy to offer, hard to use well. If GPT‑5.2 actually delivers this in production, it’s a meaningful shift.Hallucinations: down (especially with browsing)One thing we called out on the show is that a bunch of user complaints in 2025 have basically collapsed into one phrase: “it hallucinates.” Even people who don’t know what a benchmark is can feel when a model confidently lies.OpenAI’s system card shows lower rates of major incorrect claims compared to GPT‑5.1, and lower “incorrect claims” overall when browsing is enabled. That’s exactly the direction they needed.Real-world vibes:We did the traditional “vibe tests” mid-show: generate a flashy landing page, do a weird engineering prompt, try some coding inside Cursor/Codex.Early testers broadly agree on the shape of the improvement. GPT‑5.2 is much stronger in reasoning, math, long‑context tasks, visual understanding, and multimodal workflows, with multiple reports of it successfully thinking for one to three hours on hard problems. Enterprise users like Box report faster execution and higher accuracy on real knowledge‑worker tasks, while researchers note that GPT‑5.2 Pro consistently outperforms the standard “Thinking” variant. The tradeoffs are also clear: creative writing still slightly favors Claude Opus, and the highest reasoning tiers can be slow and expensive. But as a general‑purpose reasoning model, GPT‑5.2 is now the strongest publicly available option.AI in space: Starcloud trains an LLM on an H100 in orbitThis story is peak 2025.Starcloud put an NVIDIA H100 on a satellite, ...
    Más Menos
    1 h y 37 m
  • 📆 ThursdAI - Dec 4, 2025 - DeepSeek V3.2 Goes Gold Medal, Mistral Returns to Apache 2.0, OpenAI Hits Code Red, and US-Trained MOEs Are Back!

    📆 ThursdAI - Dec 4, 2025 - DeepSeek V3.2 Goes Gold Medal, Mistral Returns to Apache 2.0, OpenAI Hits Code Red, and US-Trained MOEs Are Back!
    Dec 5 2025
    Hey yall, Alex here 🫡 Welcome to the first ThursdAI of December! Snow is falling in Colorado, and AI releases are falling even harder. This week was genuinely one of those “drink from the firehose” weeks where every time I refreshed my timeline, another massive release had dropped.We kicked off the show asking our co-hosts for their top AI pick of the week, and the answers were all over the map: Wolfram was excited about Mistral’s return to Apache 2.0, Yam couldn’t stop talking about Claude Opus 4.5 after a full week of using it, and Nisten came out of left field with an AWQ quantization of Prime Intellect’s model that apparently runs incredibly fast on a single GPU. As for me? I’m torn between Opus 4.5 (which literally fixed bugs that Gemini 3 created in my code) and DeepSeek’s gold-medal winning reasoning model.Speaking of which, let’s dive into what happened this week, starting with the open source stuff that’s been absolutely cooking. Open Source LLMsDeepSeek V3.2: The Whale Returns with Gold MedalsThe whale is back, folks! DeepSeek released two major updates this week: V3.2 and V3.2-Speciale. And these aren’t incremental improvements—we’re talking about an open reasoning-first model that’s rivaling GPT-5 and Gemini 3 Pro with actual gold medal Olympiad wins.Here’s what makes this release absolutely wild: DeepSeek V3.2-Speciale is achieving 96% on AIME versus 94% for GPT-5 High. It’s getting gold medals on IMO (35/42), CMO, ICPC (10/12), and IOI (492/600). This is a 685 billion parameter MOE model with MIT license, and it literally broke the benchmark graph on HMMT 2025—the score was so high it went outside the chart boundaries. That’s how you DeepSeek, basically.But it’s not just about reasoning. The regular V3.2 (not Speciale) is absolutely crushing it on agentic benchmarks: 73.1% on SWE-Bench Verified, first open model over 35% on Tool Decathlon, and 80.3% on τ²-bench. It’s now the second most intelligent open weights model and ranks ahead of Grok 4 and Claude Sonnet 4.5 on Artificial Analysis.The price is what really makes this insane: 28 cents per million tokens on OpenRouter. That’s absolutely ridiculous for this level of performance. They’ve also introduced DeepSeek Sparse Attention (DSA) which gives you 2-3x cheaper 128K inference without performance loss. LDJ pointed out on the show that he appreciates how transparent they’re being about not quite matching Gemini 3’s efficiency on reasoning tokens, but it’s open source and incredibly cheap.One thing to note: V3.2-Speciale doesn’t support tool calling. As Wolfram pointed out from the model card, it’s “designed exclusively for deep reasoning tasks.” So if you need agentic capabilities, stick with the regular V3.2.Check out the full release on Hugging Face or read the announcement.Mistral 3: Europe’s Favorite AI Lab Returns to Apache 2.0Mistral is back, and they’re back with fully open Apache 2.0 licenses across the board! This is huge news for the open source community. They released two major things this week: Mistral Large 3 and the Ministral 3 family of small models.Mistral Large 3 is a 675 billion parameter MOE with 41 billion active parameters and a quarter million (256K) context window, trained on 3,000 H200 GPUs. There’s been some debate about this model’s performance, and I want to address the elephant in the room: some folks saw a screenshot showing Mistral Large 3 very far down on Artificial Analysis and started dunking on it. But here’s the key context that Merve from Hugging Face pointed out—this is the only non-reasoning model on that chart besides GPT 5.1. When you compare it to other instruction-tuned (non-reasoning) models, it’s actually performing quite well, sitting at #6 among open models on LMSys Arena.Nisten checked LM Arena and confirmed that on coding specifically, Mistral Large 3 is scoring as one of the best open source coding models available. Yam made an important point that we should compare Mistral to other open source players like Qwen and DeepSeek rather than to closed models—and in that context, this is a solid release.But the real stars of this release are the Ministral 3 small models: 3B, 8B, and 14B, all with vision capabilities. These are edge-optimized, multimodal, and the 3B actually runs completely in the browser with WebGPU using transformers.js. The 14B reasoning variant achieves 85% on AIME 2025, which is state-of-the-art for its size class. Wolfram confirmed that the multilingual performance is excellent, particularly for German.There’s been some discussion about whether Mistral Large 3 is a DeepSeek finetune given the architectural similarities, but Mistral claims these are fully trained models. As Nisten noted, even if they used similar architecture (which is Apache 2.0 licensed), there’s nothing wrong with that—it’s an excellent architecture that works. Lucas Atkins later confirmed on the show that “Mistral ...
    Más Menos
    1 h y 34 m
  • ThursdAI Special: Google's New Anti-Gravity IDE, Gemini 3 & Nano Banana Pro Explained (ft. Kevin Hou, Ammaar Reshi & Kat Kampf)

    ThursdAI Special: Google's New Anti-Gravity IDE, Gemini 3 & Nano Banana Pro Explained (ft. Kevin Hou, Ammaar Reshi & Kat Kampf)
    Dec 2 2025

    Hey, Alex here,

    I recorded these conversations just in front of the AI Engineer auditorium, back to back, after these great folks gave their talks, and at the epitome of the most epic AI week we’ve seen since I started recording ThursdAI.

    This is less our traditional live recording, and more a real podcast-y conversation with great folks, inspired by Latent.Space. I hope you enjoy this format as much as I’ve enjoyed recording and editing it.

    AntiGravity with Kevin

    Kevin Hou and team just launched Antigravity, Google’s brand new Agentic IDE based on VSCode, and Kevin (second timer on ThursdAI) was awesome enough to hop on and talk about some of the product decisions they made, what makes Antigravity special and highlighted Artifacts as a completely new primitive.

    Gemini 3 in AI Studio

    If you aren’t using Google’s AI Studio (ai.dev) then you’re missing out! We talk about AI Studio all the time on the show, and I’m a daily user! I generate most of my images with Nano Banana Pro in there, most of my Gemini conversations are happening there as well!

    Ammaar and Kat were so fun to talk to, as they covered the newly shipped “build mode” which allows you to vibe code full apps and experiences inside AI Studio, and we also covered Gemini 3’s features, multimodality understanding, UI capabilities.

    These folks gave a LOT of Gemini 3 demo’s so they know everything there is to know about this model’s capabilities!

    Tried new things with this one, multi camera angels, conversation with great folks, if you found this content valuable, please subscribe :)

    Topics Covered:

    * Inside Google’s new “AntiGravity” IDE

    * How the “Agent Manager” changes coding workflows

    * Gemini 3’s new multimodal capabilities

    * The power of “Artifacts” and dynamic memory

    * Deep dive into AI Studio updates & Vibe Coding

    * Generating 4K assets with Nano Banana Pro

    Timestamps for your viewing convenience.

    00:00 - Introduction and Overview

    01:13 - Conversation with Kevin Hou: Anti-Gravity IDE

    01:58 - Gemini 3 and Nano Banana Pro Launch Insights

    03:06 - Innovations in Anti-Gravity IDE

    06:56 - Artifacts and Dynamic Memory

    09:48 - Agent Manager and Multimodal Capabilities

    11:32 - Chrome Integration and Future Prospects

    20:11 - Conversation with Ammar and Kat: AI Studio Team

    21:21 - Introduction to AI Studio

    21:51 - What is AI Studio?

    22:52 - Ease of Use and User Feedback

    24:06 - Live Demos and Launch Week

    26:00 - Design Innovations in AI Studio

    30:54 - Generative UIs and Vibe Coding

    33:53 - Nano Banana Pro and Image Generation

    39:45 - Voice Interaction and Future Roadmap

    44:41 - Conclusion and Final Thoughts

    Looking forward to seeing you on Thursday 🫡

    P.S - I’ve recorded one more conversation during AI Engineer, and will be posting that soon, same format, very interesting person, look out for that soon!



    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
    Más Menos
    46 m
  • 🦃 ThursdAI - Thanksgiving special 25’ - Claude 4.5, Flux 2 & Z-image vs 🍌, MCP gets Apps + New DeepSeek!?

    🦃 ThursdAI - Thanksgiving special 25’ - Claude 4.5, Flux 2 & Z-image vs 🍌, MCP gets Apps + New DeepSeek!?
    Nov 27 2025
    Hey ya’ll, Happy Thanskgiving to everyone who celebrates and thank you for being a subscriber, I truly appreciate each and every one of you!Just wrapped up the third (1, 2) Thanksgiving special Episode of ThursdAI, can you believe November is almost over? We had another banger week in AI, with a full feast of AI released, Anthropic dropped the long awaited Opus 4.5, which quickly became the top coding LLM, DeepSeek resurfaced with a math model, BFL and Tongyi both tried to take on Nano Banana, and Microsoft dropped a 7B computer use model in Open Source + Intellect 3 from Prime Intellect! With so much news to cover, we also had an interview with Ido Sal & Liad Yosef (their second time on the show!) about MCP-Apps, the new standard they are spearheading together with Anthropic, OpenAI & more! Exciting episode, let’s get into it! (P.S - I started generating infographics, so the show became much more visual, LMK if you like them) ThursdAI - I put a lot of work on a weekly basis to bring you the live show, podcast and a sourced newsletter! Please subscribe if you find this content valuable!Anthropic’s Opus 4.5: The “Premier Intelligence” Returns (Blog)Folks, Anthropic absolutely cooked. After Sonnet and Haiku had their time in the sun, the big brother is finally back. Opus 4.5 launched this week, and it is reclaiming the throne for coding and complex agentic tasks.First off, the specs are monstrous. It hits 80.9% on SWE-bench Verified, topping GPT-5.1 (77.9%) and Gemini 3 Pro (76.2%). But the real kicker? The price! It is now $5 per million input tokens and $25 per million output—literally one-third the cost of the previous Opus.Yam, our resident coding wizard, put it best during the show: “Opus knows a lot of tiny details about the stack that you didn’t even know you wanted... It feels like it can go forever.” Unlike Sonnet, which sometimes spirals or loses context on extremely long tasks, Opus 4.5 maintains coherence deep into the conversation.Anthropic also introduced a new “Effort” parameter, allowing you to control how hard the model thinks (similar to o1 reasoning tokens). Set it to high, and you get massive performance gains; set it to medium, and you get Sonnet-level performance at a fraction of the token cost. Plus, they’ve added Tool Search (cutting enormous token overhead for agents with many tools) and Programmatic Tool Calling, which effectively lets Opus write and execute code loops to manage data.If you are doing heavy software engineering or complex automations, Opus 4.5 is the new daily driver.📱 The Agentic Web: MCP Apps & MCP-UI StandardSpeaking of MCP updates, Can you believe it’s been exactly one year since the Model Context Protocol (MCP) launched? We’ve been “MCP-pilled” for a while, but this week, the ecosystem took a massive leap forward.We brought back our friends Ido and Liad, the creators of MCP-UI, to discuss huge news: MCP-UI has been officially standardized as MCP Apps. This is a joint effort adopted by both Anthropic and OpenAI.Why does this matter? Until now, when an LLM used a tool (like Spotify or Zillow), the output was just text. It lost the brand identity and the user experience. With MCP Apps, agents can now render full, interactive HTML interfaces directly inside the chat! Ido and Liad explained that they worked hard to avoid an “iOS vs. Android” fragmentation war. Instead of every lab building their own proprietary app format, we now have a unified standard for the “Agentic Web.” This is how AI stops being a chatbot and starts being an operating system.Check out the standard at mcpui.dev.🦃 The Open Source Thanksgiving FeastWhile the big labs were busy, the open-source community decided to drop enough papers and weights to feed us for a month.Prime Intellect unveils INTELLECT-3, a 106B MoE (X, HF, Blog, Try It)Prime Intellect releases INTELLECT-3, a 106B parameter Mixture-of-Experts model (12B active params) based on GLM-4.5-Air, achieving state-of-the-art performance for its size—including ~90% on AIME 2024/2025 math contests, 69% on LiveCodeBench v6 coding, 74% on GPQA-Diamond reasoning, and 74% on MMLU-Pro—outpacing larger models like DeepSeek-R1. Trained over two months on 512 H200 GPUs using their fully open-sourced end-to-end stack (PRIME-RL async trainer, Verifiers & Environments Hub, Prime Sandboxes), it’s now hosted on Hugging Face, OpenRouter, Parasail, and Nebius, empowering any team to scale frontier RL without big-lab resources. Especially notable is their very detailed release blog, covering how a lab that previously trained 32B, finetunes a monster 106B MoE model! Tencent’s HunyuanOCR: Small but Mighty (X, HF, Github, Blog)Tencent released HunyuanOCR, a 1 billion parameter model that is absolutely crushing benchmarks. It scored 860 on OCRBench, beating massive models like Qwen3-VL-72B. It’s an end-to-end model, meaning no separate detection and recognition steps. Great for parsing PDFs, docs, and ...
    Más Menos
    1 h y 21 m
  • 📆 ThursdAI - the week that changed the AI landscape forever - Gemini 3, GPT codex max, Grok 4.1 & fast, SAM3 and Nano Banana Pro

    📆 ThursdAI - the week that changed the AI landscape forever - Gemini 3, GPT codex max, Grok 4.1 & fast, SAM3 and Nano Banana Pro
    Nov 20 2025
    Hey everyone, Alex here 👋I’m writing this one from a noisy hallway at the AI Engineer conference in New York, still riding the high (and the sleep deprivation) from what might be the craziest week we’ve ever had in AI.In the span of a few days:Google dropped Gemini 3 Pro, a new Deep Think mode, generative UIs, and a free agent-first IDE called Antigravity.xAI shipped Grok 4.1, then followed it up with Grok 4.1 Fast plus an Agent Tools API.OpenAI answered with GPT‑5.1‑Codex‑Max, a long‑horizon coding monster that can work for more than a day, and quietly upgraded ChatGPT Pro to GPT‑5.1 Pro.Meta looked at all of that and said “cool, we’ll just segment literally everything and turn photos into 3D objects” with SAM 3 and SAM 3D.Robotics folks dropped a home robot trained with almost no robot data.And Google, just to flex, capped Thursday with Nano Banana Pro, a 4K image model and a provenance system while we were already live! For the first time in a while it doesn’t just feel like “new models came out.” It feels like the future actually clicked forward a notch.This is why ThursdAI exists. Weeks like this are basically impossible to follow if you have a day job, so my co‑hosts and I do the no‑sleep version so you don’t have to. Plus, being at AI Engineer makes it easy to get super high quality guests so this week we had 3 folks join us, Swyx from Cognition/Latent Space, Thor from DeepMind (on his 3rd day) and Dominik from OpenAI! Alright, deep breath. Let’s untangle the week.TL;DR If you only skim one section, make it this one (links in the end):* Google* Gemini 3 Pro: 1M‑token multimodal model, huge reasoning gains - new LLM king* ARC‑AGI‑2: 31.11% (Pro), 45.14% (Deep Think) – enormous jumps* Antigravity IDE: free, Gemini‑powered VS Code fork with agents, plans, walkthroughs, and browser control* Nano Banana Pro: 4K image generation with perfect text + SynthID provenance; dynamic “generative UIs” in Gemini* xAI* Grok 4.1: big post‑training upgrade – #1 on human‑preference leaderboards, much better EQ & creative writing, fewer hallucinations* Grok 4.1 Fast + Agent Tools API: 2M context, SOTA tool‑calling & agent benchmarks (Berkeley FC, T²‑Bench, research evals), aggressive pricing and tight X + web integration* OpenAI* GPT‑5.1‑Codex‑Max: “frontier agentic coding” model built for 24h+ software tasks with native compaction for million‑token sessions; big gains on SWE‑Bench, SWE‑Lancer, TerminalBench 2* GPT‑5.1 Pro: new “research‑grade” ChatGPT mode that will happily think for minutes on a single query* Meta* SAM 3: open‑vocabulary segmentation + tracking across images and video (with text & exemplar prompts)* SAM 3D: single‑image → 3D objects & human bodies; surprisingly high‑quality 3D from one photo* Robotics* Sunday Robotics – ACT‑1 & Memo: home robot foundation model trained from a $200 skill glove instead of $20K teleop rigs; long‑horizon household tasks with solid zero‑shot generalization* Developer Tools* Antigravity and Marimo’s VS Code / Cursor extension both push toward agentic, reactive dev workflowsLive from AI Engineer New York: Coding Agents Take Center StageWe recorded this week’s show on location at the AI Engineer Summit in New York, inside a beautiful podcast studio the team set up right on the expo floor. Huge shout out to Swyx, Ben, and the whole AI Engineer crew for that — last time I was balancing a mic on a hotel nightstand, this time I had broadcast‑grade audio while a robot dog tried to steal the show behind us.This year’s summit theme is very on‑the‑nose for this week: coding agents.Everywhere you look, there’s a company building an “agent lab” on top of foundation models. Amp, Cognition, Cursor, CodeRabbit, Jules, Google Labs, all the open‑source folks, and even the enterprise players like Capital One and Bloomberg are here, trying to figure out what it means to have real software engineers that are partly human and partly model.Swyx framed it nicely when he said that if you take “vertical AI” seriously enough, you eventually end up building an agent lab. Lawyers, healthcare, finance, developer tools — they all converge on “agents that can reason and code.”The big labs heard that theme loud and clear. Almost every major release this week is about agents, tools, and long‑horizon workflows, not just chat answers.Google Goes All In: Gemini 3 Pro, Antigravity, and the Agent RevolutionLet’s start with Google because, after years of everyone asking “where’s Google?” in the AI race, they showed up this week with multiple bombshells that had even the skeptics impressed.Gemini 3 Pro: Multimodal Intelligence That Actually DeliversGoogle finally released Gemini 3 Pro, and the numbers are genuinely impressive. We’re talking about a 1 million token context window, massive benchmark improvements, and a model that’s finally competing at the very top ...
    Más Menos
    1 h y 29 m