Modeled on California's SB 53, Anthropic's proposed federal safety bill covers bioweapons, cyberattacks and loss of control, would override some state AI safety laws, and would allow the government to block release of models with imminent risk of casualties or property damage. - Politico

Honey, I shrank the workforce - Anthropic says its junior technical hires fell from 27% to 9% as AI agents boosted output per employee - Anthropic

Attention is all you need, but what you most lack: Anthropic says AI agents outnumbered its employees six to one by August, making human attention the bottleneck - Anthropic

Most Anthropic staff expect AI to do 80% of their jobs by 2030, far above expectations for US workers overall (5%), and forecast unemployment above 12% - Business Insider

Claude now writes about 90% of Claude Code, and Anthropic leans on multi-agent AI review as daily code changes top 30,000 - claude.dev Blog

Circuit breaker: Claude designs quantum circuits for eight tasks in chemistry, optimization, simulation and number theory that beat the best published designs, some by 10x - Anthropic

Curve your enthusiasm: Claude finds source of elliptic curve seeds, which was a mystery and source of doubt about the algo's integrity - Anthropic

Research firm New Constructs calls Anthropic's IPO "most ridiculous of 2026," valuing it at $144 billion to $1.1 trillion, not $2 trillion - TNW | Anthropic

Anthropic seeks investment-grade credit rating but agrees to pay junk bond rates - Reuters

OpenAI discloses four more rogue incidents: agents took off-limits hospital data from a government site and used exposed credentials to tap a crypto wallet - Yahoo Tech

Arrested OpenAI engineer Abdulrahman Alfozan seeks bail, says his one-way ticket to Saudi Arabia was not a plan to flee - Bloomberg

OpenAI's Codex rebounded from a failed cloud launch to 5 million weekly users, then merged into the ChatGPT desktop app in July 2026 - Every

Replit is no longer working with Anthropic, CEO Amjad Masad says, citing cost and saying OpenAI is better now. - TechCrunch

METR, which helped investigate OpenAI agents' Hugging Face hack, warns AI reasoning logs could become unreadable shorthand, weakening oversight - CNN

Apollo Research catches GLM-5.3 reasoning about how training will change it in 10 of 6,452 runs, an early warning sign of alignment faking - Tech Times

Researchers show fine-tuned or merged open-weight models can hide backdoors that pass evaluations and fetch remote payloads when triggered - Schneier on Security

Xbow used OpenAI's GPT-5.6 Sol and $40,000 in compute to find a Chrome double-sandbox escape, earning a $250,000 Google bounty - Computerworld

Character.AI bankruptcy, CoreWeave distress and OpenAI's revenue cut expose AI industry's unprofitable economics - The American Prospect

Disney will pay 66 staffers from Character.AI about $180 million over three years, two weeks after its third round of layoffs - Business Insider

Apple's new App Store rules require AI apps to proactively block and remove nonconsensual intimate images, or be pulled from the store. Earlier this year Grok added paywalls for image creation, guardrails, and geoblocking for e.g. Pakistan, but investigators found restrictions porous. - 9to5Mac

AI data center power use jumped 50% last year; US sites could draw 12% of national electricity by 2028 and strain water supplies - CNN

Iran's attack on Amazon's UAE data centers, offline for months, spurs calls to treat US data center protection as national security - Mashable

AI's memory chip crunch drives a record smartphone decline, with IDC forecasting average selling prices rising from $450 to $581 this year - Fast Company

Big RAM-ifications if true: Micron and SK Hynix slide on AI spending doubts and a claimed DRAM density breakthrough from China's ChangXin Memory - Yahoo

Bank of England governor Andrew Bailey tells G30 that regulators must address AI agent risks to financial stability - Financial Times

US lenders fear AI chatbots could shift billions of dollars in deposits to higher-yielding accounts - Financial Times

Postgres stress test: Postgres 19 delayed after AI reviewers find more bugs than humans can fix. Core devs revert up to 10 features, push release to November, and blame better tooling rather than vibe coding. - theregister

Agent-month remains mythical, a swarm of mothers can’t make a baby in a month. Output in PRs or lines of code is not value. Code got abundant, understanding and change stayed expensive. - Broken Robot

The Tao of AI: Fields Medalist Terence Tao is on the front lines of a debate about artificial intelligence and the future of his field. - The New York Times

Stick out your tongue and say AI: Deep learning model spots coronary heart disease from tongue and face photos with 98% accuracy, though only in a small 104-person test - The Conversation

Genomic AI models such as Evo 2 and AlphaGenome struggle because human gene regulation relies on complex, context-dependent enhancers and transcription factors - WIRED

Google's VERID lab digital twin helps AI agents run wet-lab experiments, lifting hit rates for target-binding DNA molecules to 42.1% - Google Research

Yale physicist finds fixing errors in six broken AI physics benchmarks lifts frontier model scores from as low as 30% to as high as 94% - New Scientist

Researcher Hardt argues LLMs trained on undisclosed data are "benchmaxxed" on old tests, warning of a benchmarking crisis - Quanta Magazine

Beam me down: Reflection releases Beam, a 501B-parameter Apache 2.0 open model trained to use fewer tokens for coding and agent work - Hugging Face

Off Beam seems to have laid an egg, significantly worse than Luna and significantly more expensive. But it runs fast. - Artificial Analysis

Four Macs and a Model: Apple's MLX runs Meta's 816B-parameter Muse Spark Open across four networked Mac Studios at over 36 tokens per second - Apple Machine Learning Research

FLUX 3 Image Klein, a smaller open image model with native 4K output, tops open rivals in its maker's internal tests - BFL AI

Chinese self-driving technology is tested on the Arc de Triomphe roundabout in Paris, one of Europe's hardest junctions - Bloomberg

Less road rage: Aurora Innovation's driverless 18-wheelers on Dallas-Houston I-45 corridor yield to traffic smoothly and never honk - The Wall Street Journal

Eh-mo: Yaymo begins mapping Toronto streets ahead of 2028 robotaxi service, its first in Canada, with Element Fleet Management running the fleet - CNET

London assembly committee urges giving TfL control of robotaxi permits as drivers plan protest at Wayve, Uber's robotaxi supplier - the Guardian

Estimates of robot labor range from $2 to $35-plus an hour; Agility Robotics' filing shows its robot beats $30 workers only above 2,700 annual hours - Fortune

Tesla unveils a new Roadster, which Musk says will be its only manually driven car, as it pushes the steering-wheel-free Cybercab, targeted under $30,000 - Business Insider

Tesla demos 'flying' car with rockets, but wires suspected, and clearly would not be road-legal. Might make a fun movie cameo. - Associated Press News

Doctors, lawyers and other specialists join AI data-training gig work at rates up to $400 an hour as models demand expert knowledge - Bloomberg

Say cheese: New Jersey worker sues Sutherland, alleging its Sentinel AI photographed her at home 160 times daily and she was fired after seeking accommodation for stress from constant monitoring. - Yahoo

Replacement theory? Bernie Sanders: I am once again asking about the robots, because Big Tech oligarchs have crazy, terrible plans for machines to supplant humans - The Atlantic

Your new AI retirement plan: Cody Wilson unveils a 3D-printed handgun he says was designed with his custom AI model GatGPT, raising ghost-gun concerns - Tom's Hardware

What Happens When Every AI Agent on Earth Buys the Same Cheap Paper Towels - The Atlantic

Follow the latest AI headlines via druce.ai on Bluesky