Keep on to blur preview images; turn off to show them clearly
Okay and then final step is to take the start and end frame for each section and put into Veo 3.1. I ended up muting the audio for each clip and overlaying game music. And that's the walk-through video for @NeurIPSConf 2025! Prompt: "pov playing a pixel style video game. you as a character walk forward and then turn right to see a pixel style character presenting an ai conference poster."
Partner @a16z AI 🤖 and twin to @omooretweets | Investor in @elevenlabsio, @krea_ai, @bfl_ml, @hedra_labs, @wabi, @WaveFormsAI, @ViggleAI, & more
I think, if Europeans want to enlist China in the pro-Ukraine coalition, they have just one good card. ASML EUV JV. Supply chain in the US gets burned, but China will bail ASML out. In 5 years it'll be working again. They'll pay tens-hundreds of billions on top. Rest is noise.
We're in a race. It's not USA vs China but humans and AGIs vs ape power centralization. @deepseek_ai stan #1, 2023–Deep Time «C’est la guerre.» ®1
I don't understand this bullshit. Russia is not all-important but you've got to offer SOMETHING OF VALUE to get China to burn the bridges. Probably there's nothing of value enough in France. What did Macron even offer? Win-win Airbus deals? Privilege of investment? …Izzat?
I think, if Europeans want to enlist China in the pro-Ukraine coalition, they have just one good card. ASML EUV JV. Supply chain in the US gets burned, but China will bail ASML out. In 5 years it'll be working again. They'll pay tens-hundreds of billions on top. Rest is noise.
Experimenting with some new workflows today. Saw a post about a new way of considering memory for agents and realized that my new cass tool would be the perfect bedrock to build such a system on top of, since it's super performant, already integrates with all coding agents, and is built for use by coding agents via the CLI. I asked the hit squad: GPT 5.1 Pro, Gemini3 Ultra, Grok 4.1, and Opus 4.5 separately to come up with detailed planning proposals that combine cass with those memory ideas from the blog post and academic paper. I had to clarify that I wanted something that worked for all coding agents, not just Claude, but each conversation just had two short prompts from me (shown in the screenshots in this post). Then I took the resulting conversations (mostly using my handy new tool for this purpose, see recent tweets for that) and put them as markdown files in the folder. Then I asked Claude Code with Opus 4.5 to take the best parts of all the plans and make a hybrid superior plan, which it just finished doing. Then I'll turn that plan into a ton of beads, then use bv to figure out how the optimal order to build them all, and have a swarm of agents make it all today for me. We live in the future.
Just realized I posted two of the same screenshot above… here’s the other one I missed:
I'm at the neurips workshops today! DM me if you wanna chat about life at Thinking Machines, Pytorch, or ML systems in general!
@thinkymachines Formerly @PyTorch "My learning style is Horace twitter threads" - @typedfemale
This 👇happens at 11:30 in 11AB.. 🤞 #NeurIPS2025 The time has come, the walrus said To talk of many things Of Think Tokens, Semantics, Complexity and whether R1 RL even has any RL.. https://t.co/7L1QzsuUzy
AI researcher & teacher @SCAI_ASU. Former President of @RealAAAI; Chair of @AAAS Sec T. Here to tweach #AI. YouTube Ch: https://t.co/4beUPOmf6y Bsky: rao2z