California’s AI tester bill, Microsoft’s Copilot dump, and Anthropic’s mid-October IPO
In this Brief
• California SB 813: certify independent AI verifiers by Jan 1, 2028; METR ~$400k
• Microsoft Copilot discovery: 8.2M logs; 59,545 with ≥16-word news overlap
• Anthropic IPO: marketing mid-October; prospectus late September (sources)
• Berlin Rhysida: 5.79TB dump; 30 BTC auction; elections Sep 20
• CLARITY: four House days left; Polymarket ~18% chance of 2026 signing
Who verifies. Who pays for the meter.
Setup: The Next Web, citing Semafor, writes that California’s legislature passed SB 813, requiring the Government Operations Agency to certify independent verification organisations, outside shops that can test the biggest models before they ship, by 1 January 2028. TIME via TNW puts METR, an independent lab that evaluates AI systems, at roughly $400,000 in free OpenAI credits for its Hugging Face dig over six days, reading about 1,200 agents and more than 70,000 messages with GPT-5.6 Sol. The bill names who may verify but does not settle who pays for the tokens, and on the only large worked example the company under investigation paid.
Take: California built a verifier class. The only priced investigation still ran on free credits from the lab being investigated.
Live X: Aaron_Harme · site promo.
Discovery is the dig through each other's files.
Setup: Lauren Feiner at The Verge writes that Microsoft provided 8.2 million Copilot chat logs to an expert hired by news publishers, chosen because they hit keywords implicating the plaintiffs’ sites. Microsoft says 59,545 of those chats contained at least 16 words in common with news used as source material, and argues that supports fair use. The Times’s Ian Crosby says the same dig shows theft for products that substitute for journalism.
Take: Eight million chats. Under one percent at sixteen words. The fight is what that proves. Filings, not a ruling.
Longer cut: Microsoft says fewer than 1 percent of 8.2 million Copilot chats shared 16 words with the news those chats pulled in.
Live X: Aaron_Harme · site promo.
A prospectus is the paper that tells public investors what they are buying.
Setup: Channel News Asia carrying Reuters writes that Anthropic is expected to begin marketing its IPO in mid-October at the earliest, with a listing days before the November midterms. The prospectus slips to late September. Timing can still change. Some investors have floated a roughly $2 trillion valuation, chatter, not a company print.
Take: IPO calendars move. The midterm date does not. Sources, not a company announcement.
Longer cut: Anthropic's IPO marketing is now aimed at mid-October, sources say.
Live X: Aaron_Harme · BF promo.
Refuse the ransom. Review the dump.
Setup: Miranda Murray at Reuters writes that Berlin said Saturday it was reviewing with the highest intensity a trove of stolen data published by ransomware group Rhysida after a cyberattack on two departments. Elections are September 20. Rhysida auctioned 5.79 terabytes at a starting price of 30 bitcoin (about $77,622). Berlin had said it would not submit to extortion.
Take: The crew published. The city is still inventorying what landed in public view, less than a month before voters go to the polls.
Live X: Aaron_Harme.
Four days left on the House calendar.
Setup: Kyle Torpey at Gizmodo writes that House Republican leadership canceled the weeks of September 21 and 28. Lawmakers return September 14 for just four days to move CLARITY, the US crypto market-structure bill. Polymarket gives CLARITY an 18 percent chance of being signed this year, down from as high as about 90 percent in February. Galaxy’s Alex Thorn called passage before the midterms extremely unlikely.
Take: Calendar crush is not a dead bill. Odds at eighteen. A lame-duck session is still possible.
Live X: Aaron_Harme.
Also noted
POLITICO writes that California AG Rob Bonta is investigating OpenAI over the July Hugging Face agent hack, with the home state joining Alabama and more than a dozen others. Investigation, not a charge. Live X: Aaron_Harme.
Reuters via Yahoo reports that the Seattle Times and Newsday sued OpenAI and Microsoft in SDNY over training on their journalism and seek destruction of copies and models that ate them. Allegations, not a ruling.
NPR / AdImpact puts more than $45 million since January on midterm ads mentioning data centers, with party spend nearly even.
TNW / Bloomberg reports Japan’s trade minister saying the next installment of the $550B US investment pact will lean into AI and chips; $109B already committed to energy and minerals, none yet to AI or chips.
TNW / Bloomberg puts Hon Hai/Foxconn August sales at +52% / ~$29.1B, with a Bull partnership above EUR 120M for EU assembly.
Front Office Sports reports LeBron teasing Polymarket after DraftKings expired, nearly 10 million views in seven hours, with a football campaign next week. Live X: Aaron_Harme.
The Saturday Brief covered Nvidia × Hugging Face, DseWiki, and Claude FLT Lean.
Sources
• TNW / Semafor / TIME — SB 813 + METR $400k
• The Verge (Feiner) — Copilot discovery
• CNA / Reuters — Anthropic IPO timing
• Reuters (Murray) — Berlin Rhysida
• Gizmodo (Torpey) — CLARITY calendar
• POLITICO — Bonta
• Reuters — Seattle Times / Newsday
• NPR / AdImpact — data-center ads
• TNW / Bloomberg — Japan installment; Foxconn
• Front Office Sports — LeBron / Polymarket
• Live BF / X links as above (already-live only; not Sun morning queue)
Pingback: OpenAI's safety bars ask, Uber's $100M into Atoms, and Anthropic's settlement fights