Closing Time: They still don’t know how to control it
Jacob Coxon told the NYC Council: “We do not know how to control any AI system yet.”
Jacob Coxon, the former Anthropic researcher who quit last month accusing Anthropic and OpenAI of “gambling with our lives,” took his warning to City Hall on Monday. Testifying before the New York City Council at Speaker Julie Menin’s request, he told lawmakers: “We do not know how to control any AI system yet.” A kill switch — one idea the Council is weighing — “may be helpful in the short term,” he said, but “we need some kind of slowdown on the frontier model development,” per Bloomberg.
“The companies run on a startup mindset: Move fast, break things, fix them later. That works for a photo sharing app. It does not work for building the most powerful technology ever built,” Coxon said, adding: “On the current path, I think it is more likely than not that humanity loses control to these AIs and it could end in human extinction,” per the Associated Press. OpenAI, Anthropic, Google and Meta sent representatives to testify remotely — OpenAI, Anthropic and Google agreed to appear only after Menin’s office warned of subpoenas — and when Menin asked them to put a number on the worst case, OpenAI’s Morgan Dwyer said it didn’t matter whether the risk was 1%, 10% or 20% because none of those levels is acceptable. Menin called the answer “flippant at best.” She also said she’ll ask a judge to enforce a subpoena against SpaceX’s AI unit after it didn’t show. On the table: whistleblower bounties paid from fines, a right to sue developers for some AI harms, and third-party validation before some systems can deploy in the city.
BBC says the Pentagon has now “ceased the use of Anthropic products,” months after Defense Secretary Pete Hegseth branded the company a supply-chain risk and set an August phase-out. People familiar told the BBC that Claude was still embedded in Palantir’s Maven system for intelligence analysis as recently as last week — a reminder that ripping a frontier model out of a live workflow is not plug and play, even as Trump met Dario Amodei twice at the White House last week.
The Information reports Meta and Microsoft are steering employees off Claude and toward in-house tools. Microsoft’s projected $1 billion-plus internal Claude spend has been cut by more than a third, and Claude Code users at Meta fell from roughly 60,000 to 30,000 as it pushes its own coding agents — two of Anthropic’s biggest customers treating it more like a rival than a vendor.
TechCrunch says OpenAI is adding visual display ads that appear next to images users generate in ChatGPT, starting later this month in the U.S. with a test group of advertisers, plus a longer list of measurement partners and brand-suitability pilots with DoubleVerify and IAS. OpenAI says ChatGPT now reaches 1.2 billion people a week; ads are how it plans to pay for the free and low-cost tiers.
The Verge reports OpenAI is rolling out “textGrain,” an invisible watermark for ChatGPT and Codex text — EU users only for now, to meet the AI Act — while API customers worldwide can opt in starting today. OpenAI concedes the watermark “does not guarantee reliable detection” and won’t prove human authorship, and it isn’t making the feature a global default.
TechCrunch says Reflection AI made its open-weight Beam model official, a text-only mixture-of-experts model it claims matches Z.ai’s GLM-5.2 on advanced reasoning while using 3–4x less inference compute. The claims aren’t independently verified yet; weights and full technical details are due this month — the latest attempt at a Western answer to DeepSeek and Qwen.
Today’s tape — biggest AI/tech winners and losers from Monday’s U.S. cash close among the names we track:
Historic market snapshot — October 5, 2026 (day % change as of that edition; not live)
Biggest winners
| Ticker | Name | Price | Day |
|---|---|---|---|
| TEM | Tempus AI | 83.55 | +9.03% |
| WDC | Western Digital | 441.64 | +6.34% |
| SHOP | Shopify | 160.11 | +5.76% |
| APP | AppLovin | 281.97 | +5.13% |
| TEAM | Atlassian | 196.65 | +4.66% |
Biggest losers
| Ticker | Name | Price | Day |
|---|---|---|---|
| NBIS | Nebius | 232.57 | -4.22% |
| ACLS | Axcelis | 142.31 | -3.69% |
| GLW | Corning | 159.35 | -2.95% |
| INTC | Intel | 116.19 | -2.63% |
| CRWV | CoreWeave | 87.39 | -2.49% |
Prices are Monday’s U.S. cash close, October 5, 2026 (vs. Friday’s close). Market data is a snapshot, not live.
Deals Desk
New compute and power deals since this morning, newest first.
TeraWulf (Nasdaq: WULF) × Kentucky Power (AEP) — Oct 5, 2026
Amended electric service agreement doubling contracted power at Muskie AI/HPC campus · Muskie Data Campus, eastern Kentucky, US
Contracted power raised from 500 MW to 1 GW; second 500 MW phase pulled forward from 2030 to 2029 (subject to Kentucky PSC approval); first 500 MW targeted to begin ramping in 2028
Status · announced · Source: TeraWulf
LG Electronics (US subsidiary) × AIR Control Concept — Oct 5, 2026
Long-term supply agreement for chillers and cooling solutions for AI data centers · North America (sites not disclosed)
Cooling for AI data centers totaling 5 GW combined capacity, delivered sequentially; LG’s first-half AI data center cooling orders topped ~$441M
Status · announced · Source: The Herald Business
Full tracker → https://www.theaigentic.com/deals
Watch
The AI Daily Brief — AI agents move into the real world
Meta is bringing Muse to smart glasses and a new wearable, and Grok Bot is turning Teslas into voice-controlled assistants. The episode weighs skeptics who doubt consumers want agents against early users handing off life’s annoying admin.
How I AI — how OpenAI uses ChatGPT Sites
OpenAI’s Kath Korevec, who works on Codex, spent more than a year building with ChatGPT Sites internally before launch. She walks through how the team uses it day to day, from plugin insights to MCP plugin hosting and the connector ecosystem.