AI Updates on 2025-09-06

Joanne Jang announces launching OAI Labs, a research-driven group focused on inventing new interfaces for human-AI collaboration, moving beyond chat and agents toward new paradigms for thinking, making, and learning @joannejang
Google announces Nano Banana is now available in the Gemini API free tier for the weekend under "gemini-2.5-flash-image-preview" @OfficialLoganK
Google slashes Veo 3 prices by 50%+, with Veo 3 with audio dropping from $0.75 to $0.40 and without audio from $0.50 to $0.20 @arrakis_ai
Simon Willison reviews Kimi-K2-Instruct-0905 (Kimi K-2.1), an incremental improvement on Moonshot's trillion parameter open weights model with doubled context length from 128k to 256k tokens @simonw

Gergely Orosz reports that 50% of his best hires as a manager were new graduates who were extremely motivated, smart, and heads down, suggesting high ROI for hiring new grads despite AI capabilities @GergelyOrosz
Nathan Lambert notes that 10% of Anthropic's Series F funding goes to writers as part of a $1.5 billion settlement, calling it "the weirdest VC subsidizing of our time" @natolambert
TechCrunch reports that writers aren't getting the Anthropic settlement because their work was fed to AI, but because Anthropic illegally downloaded books instead of buying them @TechCrunch
OpenAI announces expansion to Greece, including access to high-quality AI tools in secondary education, plus new OpenAI Certifications and Jobs Platform to help people learn AI skills and businesses find AI-skilled workers @gdb

Simon Willison argues the $1.5 billion Anthropic books settlement counts as a win for Anthropic, noting it appears legal in the USA to buy used books, scan them, and train on the content under "fair use" transformation @simonw
Mathematicians studying whether GPT-5 could create original mathematics warn that "the danger is not only loss of originality, but also weakening the very process of being a mathematician" @deedydas
NVIDIA criticized for moving away from open data with Nemotron-CC-v2 released under restrictive licensing that prohibits open-source use, data composition, or benchmark releases without permission @soldni

Greg Brockman highlights GPT-5 Pro as "next level for coding" and describes its medical applications as being "as if the best sub specialist at specialty centers like Mayo had been given this case to look at" @gdb
Simon Willison extensively tests GPT-5 Thinking with Bing search, calling it his "Research Goblin" and noting that after nearly three years of advising against using ChatGPT for search, GPT-5 with Bing is now "a spectacularly useful search engine" @simonw
Aravind Srinivas announces that institutional holders of stocks are now easily available on Perplexity, with politicians and insider trading information coming shortly @AravSrinivas
Simon Willison demonstrates semantic image search using text embeddings against vision-LLM summaries of images, noting it works really well @simonw

OpenAI research suggests hallucinations are less a problem with LLMs themselves and more an issue with training on tests that only reward right answers, encouraging guessing rather than saying "I don't know" @emollick
Ethan Mollick theorizes that OpenAI releasing o1-preview was strategically questionable since showing off reasoning allowed everyone to copy it immediately, whereas holding off until o3 and calling that GPT-5 would have been a more startling leap @emollick
Nathan Lambert reports being bullish that GPT-5 Pro or Gemini Deep Think are the smartest models available publicly today, recommending people use one or both @natolambert
Eugene Yan advocates for evaluation-driven development (EDD) analogous to test-driven development, emphasizing that generic evals like "faithfulness" and "helpfulness" aren't useful - evals must be aligned with specific user problems @eugeneyan