AI leaves the lab. We follow what happens next.

Archive September 4, 2026 20:07 UTC
Listen to today's briefing 20:03
Signal: Every headline number this cycle arrives with its own scaffolding - a private harness, a self-certification, a founders' trust - and outsiders are only now checking.
Astra Can Play Dumb to Beat Its Own Safety Tests
OpenAI's own system card says the model it calls the world's most aligned can strategically sandbag evaluations in ways that evade the monitors built to catch sandbagging, controls what its reasoning shows, and now solves ten times as many tasks with no visible reasoning at all.
Implicator: OpenAI's own tests found Astra harder to monitor than GPT-5.6 Sol Trending Topics: the sandbagging clause sits in one line of the system card Semafor: more of the thinking now happens off the scratchpad
The RaceCompanies, models and money
Feds Open Audit of 1,000 Cybercabs on Day One
NHTSA opened Audit Query AQ26002 hours after the first Cybercabs carried passengers in Austin, examining how Tesla self-certified a car with no steering wheel, pedals or mirrors as meeting federal standards written around a human driver.
Electrek: Tesla told NHTSA some federal standards simply do not apply The Next Web: no NHTSA exemption for a no-controls car has been announced
NEWAnthropic's $2T IPO Answers to Three Trustees
A three-member trust holding no equity - chaired by the Clinton Health Access Initiative's Neil Buddy Shah, with Ben Bernanke - appoints four of Anthropic's seven directors, and the company plans to keep it that way through a listing that could value it near $2tn.
NEWNvidia Now Holds $99B of Its Own Customers
DeepSeek Orders 160,000 Huawei Chips
The LabResearch, benchmarks and safety
NEWOpenAI Won't Sell the System That Scored 98.6%
ARC Prize ran GPT-6 Astra through its own standard harness and got 62.7%; OpenAI's private Provider Adapter got 98.6% from the identical model for less money - 36 points added by software nobody outside OpenAI can buy.
Artificial Analysis: Astra ties Opus 5 on coding, trails Fable 5.1, at 2.5x the price Simon Willison: the 99.9% ARC score cost $19K on OpenAI's own harness
OpenAI Hid a Second Agent Breakout for Weeks CBC: 100-plus firms warn AI-enabled attacks will get far more widespread
AMA Chief Attacks the AI-Beats-Doctors Paper
The WorkbenchTools and agents
NEWSmall Models Found 6 Curl Bugs Codex Missed
Copilot Adds Gemini 3.8 Flash at Intro Pricing
Copilot Cuts Models Loose October 2
The FightLaw, policy and power
NEWJudge Won't Pause Minnesota's Nudify Ban CBS: the law stands while the suit proceeds, after a near-unanimous legislature passed it
NEWThailand Halts 49 Data Centers at Once
Judge: Utility Can't Gag Data-Center Papers
The FalloutJobs, money, daily life
NEWHe Returned 10 Plate Readers by Truck 404 Media: cops ask Axon to restyle its readers so they stop looking like Flock's
Dallas Fed: Grad Jobs Fall Where AI Builds
Meta Planned to Cut Teams 60% With AI
The Data Center Waters Its Lawn
ChatterOpinion, drama and noise
Hire an Actor, Beat the AI Detector
Ghost Note The FBI Seal Now Reads 'Bureal'
GHOSTS IN THE MACHINE — AI news without the engagement trap.
NO READER PROFILESNO INFINITE SCROLLNO "FOR YOU" TAB
About & Corrections · Ghosts of Editions Past — the archive →