AI Updates on 2025-09-07

Elon Musk announces big update to Imagine arriving in a few weeks, with 'compelling half-hour episodes' of generative video by next year, targeting coherent 15-minute video generations from a single prompt by end of this year @AndrewCurran_
Tencent Hunyuan achieves top two spots on Hugging Face trending charts with Hunyuan-MT-7B and HunyuanWorld-Voyager models @huggingface

ASML expected to get a seat on Mistral's board after committing $1.5 billion to their raise and becoming the top shareholder, forming a Euro AI alliance @AndrewCurran_
Perplexity hiring data scientists to work on evals for Assistant, requiring work experience improving complex AI systems at scale @alexgraveley
Nathan Lambert describes paying for better AIs as a way to "pay to win" in your career, comparing it to video game dynamics @natolambert
Paul Graham retweets observation about AI agents enabling decoupling of output (value) from human input (time) in knowledge work for the first time @paulg

Logan Kilpatrick demonstrates using NanoBanana in Google AI Studio for experimentation @OfficialLoganK
Simon Willison provides follow-up on Google's new "AI mode" being very good and massively different from "AI overviews" which he considers terrible @simonw
Greg Brockman shares example of codex CLI with web search integration @gdb

Ethan Mollick discusses nuanced findings about GPT-5 Pro being able to do novel mathematics but only when guided by a math professor, highlighting the speed of advance since GPT-4 @emollick
Hugging Face releases FinePDFs, the largest PDF dataset spanning over half a billion documents with 3T tokens from high-demand domains like legal and science, showing 2x longer context than web text @huggingface
Alex Graveley implements token level reranker idea as referenced research @alexgraveley
Ethan Mollick notes that multimodal LLMs have been weak at seeing fine visual details, making visual benchmarks important to watch for progress tracking @emollick
François Chollet explains that deep learning models can only generalize via interpolation on parametric curves, leading to hallucinations, and suggests causal symbolic graphs as the fix for exact truthiness propagation @fchollet