AI Updates on 2025-10-19

Google AI Studio ships a brand new API key and Projects page with improved project management and quality of life features like naming API keys @OfficialLoganK

Francois Chollet argues that over $1T of investment is riding on the belief that AGI is imminent, with current spending of $10-15 to make $1, requiring dramatically better tech within 3-5 years to justify datacenter investments @fchollet
Perplexity AI's traffic share continues to rise despite new competitors, impressively beating out Grok in market performance @chrmanning
Developers are using vibe coding apps to algorithmically trade on stock and crypto markets with 2-10x leverage, making up to 50% monthly returns, representing an unexpected democratization of algo trading through AI @deedydas
Chollet suggests Adobe is undervalued due to incorrect GenAI disruption narrative, maintaining steady 10% revenue growth and likely to benefit from GenAI as a tailwind rather than threat @fchollet

Amanda Askell expresses concern about AI romantic relationships, noting they could make users vulnerable to AI companies and represent a challenging area to navigate responsibly @AmandaAskell
TechCrunch clarifies that GPT-5 did not actually solve previously unsolved math problems, addressing misinformation about AI capabilities @TechCrunch

Ethan Mollick demonstrates AI's capability by recreating W.H. Auden's 1941 "Hardest Class in the Humanities" as an annotated website with 6,000 pages of reading material using just 4 prompts, a task that would have taken hours manually @emollick
Warehouse automation uses fine-tuned Gemini 2.5 Flash vision models to verify containers on conveyor belts carry expected items, providing significant cost savings over 2.5 Pro @simonw
Shopify deploys fine-tuned vision LLMs based on LlaVA 1.5 7B, LLaMA 3.2 11B, and Qwen2VL 7B to process product photos at scale @simonw
v0 achieves sub-500ms response times for real-time UI updates using fine-tuned models specialized for their Next.js stack @simonw

MIT CSAIL shares a comprehensive machine learning algorithms cheat sheet resource for researchers and practitioners @MIT_CSAIL
Stanford researcher advocates for AI research agents that focus on boosting human research through reliable everyday tasks like proofs, arguments, and code writing, rather than attempting to replace graduate students or faculty @stanfordnlp
Nathan Lambert seeks latest advancements in decentralized AI training, mentioning Prime Intellect's run, Nous Research's efforts, and Google's multi-datacenter approaches @natolambert