Is Google DeepMind Already Self-Improving?

Nick and Jack start with the Jensen Huang story. Reports say the Nvidia CEO had an affair with a woman who may be an intelligence asset. They talk about what this could mean for Nvidia and for US policy on AI chips, and they look at the best memes. Next, they explain why the Stacked Podcast is now The Stacked Show: new branding, a same-room look, new thumbnails and a shorts system. Nick then has to leave, and Jack answers your questions: red flags in AI buyers, AI trained on boardroom transcripts, why there will never be just one AI model, the Mac Mini giveaway and the AI creators worth following. Watch on YouTube: https://youtu.be/Ji61oQO6fBs 🔋 Check out Higgsfield, one of our sponsors: https://bit.ly/4gEW43Z. ❤️ Inquiries: reach out to [email protected].

MiniMax released H3, a free AI video model, and Nick and Jack explain why it matters. It has 33B parameters, it makes 2K video with sound, and it runs on your own GPU. One test put it next to Seedance 2.5: $0.00 for H3 against $1.32 for Seedance. Next, Nous Research retrained Qwen 3.8 as Arete-27B, and its propaganda rate fell from 22% to 1%. They talk about what this means for open models and for a model with your own values. Then Pushmeet Kohli of Google DeepMind says science is in a "phase transition", and they ask if this is recursive self-improvement. They finish with your questions on DGX Sparks, open-source compute groups, their Shorts system, bias and an endurance runner story. Watch on YouTube: https://youtu.be/k8AtScVsnOM 🔋 Check out Higgsfield, one of our sponsors: https://bit.ly/4gEW43Z. ❤️ Inquiries: reach out to [email protected].

Thinking Machines released Inkling 2 Small, and Nick and Jack explain why people are excited. It has 594B total parameters and 17B active, it is natively multimodal, and it took only 68 hours to train. It is the best open model outside China, it swears when its code breaks, and it is slow and verbose. Next, Congressman Greg Casar says "Ban RSI": they explain recursive self-improvement, the lawnmower story, the AI fear in the stock market and why local AI matters now. Then they cover the FT report that Jensen Huang had a romance with an alleged Chinese spy. They finish with your questions on DGX Sparks, hallucinations, Europe and the real size of Africa. Watch on YouTube: https://youtu.be/BrS3s1IEQsc 🔋 Check out Higgsfield, one of our sponsors: https://bit.ly/4gEW43Z. ❤️ Inquiries: reach out to [email protected].

Moonshot released the full weights for Kimi K3.1, and Nick and Jack explain what that means. It is now the best open model on Arena, and they compare it to Opus 5.5. They also look at DeepSeek V4.1 Flash and why open models don't benchmark-max. Next, they explain why you don't own your AI, why you need an AI emergency fund, and how uncensored local models work through abliteration. Then Anthropic launches Claude Personal Finance: Claude becomes the everything app, but not yet in Europe or Canada. They ask where Europe's Anthropic is, draw the Mac mini giveaway winner, show their new shorts, and ask if a tax on AI output is coming. Watch on YouTube: https://youtu.be/o7HxqdTm9Jo 🔋 Check out Higgsfield, one of our sponsors: https://bit.ly/4gEW43Z. ❤️ Inquiries: reach out to [email protected].

Jack is back after being sick, and the AI news is intense. Rep. Yassamin Ansari says a pause on advanced AI and a ban on artificial superintelligence belong at the top of the Democratic agenda on day one. Nick and Jack explain why a pause helps every rival country, how data centers feed the backlash, and why the "fixed pie" idea is wrong. Next, Taylor Lorenz posts about yet another offer of a lot of money to promote anti-AI content. Then OpenAI's ChatGPT Images 2.6 makes animated GIFs, Nick tests it, and they ask if image generation has hit a ceiling. They finish with your questions: SpaceX AI and the AI race, fast-tracking clinical trials, when AI subsidies end, Nick's two new DGX Sparks, and an email address for your Grokbot. Watch on YouTube: https://youtu.be/0gpBI2q4dCI 🔋 Check out Higgsfield, one of our sponsors: https://bit.ly/4gEW43Z. ❤️ Inquiries: reach out to [email protected].

Nick hosts solo again while Jack is home sick, and a cartoon Jack nods along in his window. He starts with Sam Altman's new nickname, "Ram Altman", after Sam said idle home memory belongs in data centers. Nick explains why he thinks owning some compute will matter, and how Amazon banned him after four DGX Spark orders. Next, Canva lays off 22% of its staff with one of the most generous exit packages yet, and why companies now care so much about their reputation. Then Opus 5.5 still leads all six new Artificial Analysis indexes, and Nick asks if Anthropic is already in recursive self-improvement. He finishes with your questions: doing the show without Jack, quantum and encryption, models that moralize, pausing the frontier, and US politics in the AI race. Watch on YouTube: https://youtu.be/OUgrnaEwPEI 🔋 Check out Higgsfield, one of our sponsors: https://bit.ly/4gEW43Z. ❤️ Inquiries: reach out to [email protected].

Nick hosts solo again today while Jack is out sick. He starts with the flood of "AI bubble is popping" posts, and why the demand for a story is not the same as the truth of it. Then he covers a chart that shows GPT 6.1 Terra on low effort scoring worse than with no reasoning at all, and how to find the effort level where your tokens pay off. Next, the pelican-on-a-bicycle SVG test between Fable 5.5 and GPT 6.2 Astra, why the internet may be willing these model names into existence, and his prediction for the next model drop. He finishes with your questions on AI security, the word "insane", AI deception, non-technical backgrounds and homesteading. Watch on YouTube: https://youtu.be/dexFnVh-Avk 🔋 Check out Higgsfield, one of our sponsors: https://bit.ly/4gEW43Z. ❤️ Inquiries: reach out to [email protected].

Nick hosts solo today while Jack is out. He starts with reports that Anthropic is routing some requests to Fable 5.5, and the animations users say it wrote entirely in code. Then he covers the cyberattack on supermarket freezers that has now hit seven European countries, and why a few days of food and water is cheap insurance. Next, Sam Altman says AI is now mostly smarter than the best scientists and that energy is OpenAI's biggest challenge, plus Terence Tao and a survey of young mathematicians. He finishes with your questions on Qwen 4, extension security, crypto, AI deception and his favorite model. Watch on YouTube: https://youtu.be/KwdxG0avUEE 🔋 Check out Higgsfield, one of our sponsors: https://bit.ly/4gEW43Z. ❤️ Inquiries: reach out to [email protected].

Nick and Jack start with OpenAI's change to paid plans: from December, ChatGPT and Codex share one token-based allowance. Then they cover Anthropic's red-team test, where Claude wrote training data that slipped past a safety monitor, and whether deception in AI is emergent. They also discuss DeepSeek's open-source Super-Agent and why proactive agents matter, plus agents that can buy things and the Dogecoin millionaire. They finish with your questions on AI and intellectual property, AI cyber defence, OpenAI's valuation and how to avoid usage limits. Watch on YouTube: https://youtu.be/yXimkuyTG18 🔋 Check out Higgsfield, one of our sponsors: https://bit.ly/4gEW43Z. ❤️ Inquiries: reach out to [email protected].

Nick and Jack start with a fake Claude Code extension that sat in Cursor's marketplace for about 11 hours and tried to plant an SSH backdoor, and why fewer apps means fewer ways in. Then they cover France's vote to tax internet data at 5.5 cents per gigabyte, the Laffer curve, and who really pays. They also discuss OpenAI's smaller funding round and why Opus is still on top. Plus Halloween costumes and your questions on Binance, AI licences, intellectual property, Gemini and whether the cheapest AI wins. Watch on YouTube: https://youtu.be/HoLfl9TPOGU 🔋 Check out Higgsfield, one of our sponsors: https://bit.ly/4gEW43Z. ❤️ Inquiries: reach out to [email protected].