AI Updates on 2025-11-16

Google's AlphaEvolve discovers solutions better than humans on certain math problems, including the Kissing problem, by repeatedly searching for solutions in parallel, verifying them, and performing natural selection to evolve ideas. Research by mathematician Terence Tao tested it on 67 problems and found that smarter AI base models converge to solutions quicker, parallelizing generally helps but adds compute cost, and reward hacking is common @deedydas
Future House team achieves breakthrough in AI-assisted scientific research, described as one of the most important impacts of AI @sama

Shopify was the first company outside of Microsoft to use GitHub Copilot, with their Head of Engineering sharing that being known for giving great feedback helped them get early access @GergelyOrosz
Some companies are finding that having developers use AI tools in interviews don't provide much signal, with at least one Silicon Valley startup eliminating "build something with AI" interviews @GergelyOrosz
Chinese models are already eating leading AI lab market share, with questions about whether this trend is more sticky within enterprises @natolambert
Microsoft's Fairwater datacenter in Atlanta has taken over 15 million labor hours to build, more than double the 7 million hours required for the Empire State Building @mustafasuleyman

Gmail introduces new smart scheduling feature that uses email context to find meeting times and automatically creates events when receiver selects a time, representing significant productivity improvement @deedydas
New version of llm-anthropic plugin adds support for structured outputs via official API and Anthropic's web search feature @simonw
Andrej Karpathy proposes that verifiability is the most predictive feature for AI automation in the new programming paradigm, where tasks that can be practiced, reset, and rewarded are most amenable to neural network optimization @karpathy
Experts at making AI are not necessarily experts at using AI, creating opportunities for domain specialists to figure out AI capabilities in their fields before others @emollick

Current AI benchmarking focuses too heavily on model ability through API calls rather than agentic work that combines tools and problem-solving ability, which matters more economically @emollick
Better benchmarking is needed to understand why agentic abilities break down, including vision weaknesses and "doom loops" where AI keeps trying the same failed approach @emollick
Windows faces criticism from developers for including ads in a paid OS and turning on OS-level AI features like Recall by default, which developers don't want @GergelyOrosz
Canadian medical system outside major cities has completely collapsed, with AI integration potentially mitigating staff shortages but still years away from implementation @AndrewCurran_