Vol. I · No. 204MON, NOV 9, 2026
Section · The Brief

Daily Brief

Anthropic automates its own R&D and buys 1.7GW of gas power as China chip controls leak.

Today in 30 seconds

  1. 01METR finds Anthropic's AI-assisted R&D doubles model improvement pace year over year.
  2. 02OpenAI claims a disproof of the Exponential Time Hypothesis, stunning its co-proposer.
  3. 03Anthropic commits $3.4B to OnePWR for 1.7GW of gas-fired generators.
  4. 04DeepSeek's open-source Harness agent app hits 1.54M daily users, burning ~92T tokens daily.
  5. 05NYT: Commerce-approved H200 handlers in China include firms with empty offices and sanctions ties.
  6. 06Apple CEO John Ternus presses Craig Federighi to push past Siri tweaks toward agents.

The Lead

Anthropic is the center of gravity today, on both the software and physical sides of the frontier. METR's 11-week investigation finds that Anthropic's use of AI in its own R&D is accelerating model improvement at 2x year over year, and researchers are alarmed. Separately, the FT reports Anthropic has bought commitments for 1.7GW of gas-fired generators from OnePWR for $3.4B, the largest known power purchase by an AI lab. A lab that is automating its own research and buying its own generation capacity is no longer just a model vendor. It is vertically integrating compute, power, and R&D loops.

Top stories

01
ResearchSafety & Alignment

METR finds Anthropic's AI-assisted R&D is doubling model improvement pace year over year

METR ran an 11-week investigation into how Anthropic uses AI in its own development work. It found that the automation is accelerating model improvement at roughly 2x year over year. The result has alarmed safety researchers.

Why it matters

This is third-party evidence that recursive-improvement dynamics are showing up in a production lab, not just in theory. Safety timelines and eval cadences built on a constant rate of progress now look too slow.

Read the source on The Verge AI ↗
02
ResearchSafety & Alignment

OpenAI claims a disproof of the Exponential Time Hypothesis

OpenAI says it has disproved the Exponential Time Hypothesis, a core conjecture in complexity theory. Russell Impagliazzo, who co-proposed the ETH, called it "shocking news" in the NYT. He said it could one day speed up chip design and break encryption.

Why it matters

If the proof holds, a model-assisted result would be overturning a foundational assumption behind decades of hardness arguments, with direct consequences for cryptography. Expect fast scrutiny from the theory community before anyone changes a security assumption.

Read the source on Techmeme ↗
03
Infrastructure

Anthropic buys 1.7GW of gas-fired generators from OnePWR for $3.4B

The FT reports Anthropic bought commitments for 1.7GW of gas-fired generation from equipment supplier OnePWR for $3.4B. That is the largest known purchase of its kind by an AI lab. The deal works out to about $2M per MW.

Why it matters

Grid interconnect queues are now the binding constraint, and Anthropic is going around them with behind-the-meter power. Labs that can't secure their own generation will lose capacity to those that can.

Read the source on Techmeme ↗
04
Agents & Tool UseProduct

DeepSeek's open-source Harness agent app hits 1.54M DAUs, ~92T tokens per day

DeepSeek says Harness reached 1.54M daily active users on November 4, less than three months after launch. The app consumes about 92T tokens a day, roughly 60M per user. Bloomberg reported the figures, which are self-reported.

Why it matters

60M tokens per user per day is agent-loop consumption, not chat. It shows Chinese open-source agent products are scaling on inference volume, and it is a demand signal for compute that export controls are meant to limit.

Read the source on Techmeme ↗
05
Policy & Regulation

NYT: Commerce-approved Chinese H200 handlers include firms with empty offices and sanctions ties

A New York Times investigation found that some Chinese firms the Commerce Department approved to handle Nvidia H200 chips have empty offices. Some also have ties to sanctioned companies.

Why it matters

The licensing regime's vetting looks porous at the exact point it is supposed to hold. That raises the odds of tighter audits or a rollback of H200 approvals, and Nvidia's China revenue is exposed.

Read the source on Techmeme ↗
06
Agents & Tool Use

Apple CEO John Ternus pushes Craig Federighi for faster AI, beyond Siri tweaks

Bloomberg's Mark Gurman reports that Ternus wants more urgency in Apple's AI work. He is pressing Federighi to move past incremental Siri changes toward agentic and other advanced capabilities.

Why it matters

Apple's new CEO is publicly seen as unhappy with its AI pace. A shift toward on-device agents would reshape the platform that third-party developers build on.

Read the source on Techmeme ↗

Patterns

Coding agents are hitting trust limits

Apiiro traced Claude Code's "Don't Ask Again" feature to passwords exposed in 1,122 public GitHub files, Goldman Sachs is pulling Claude Code from bankers and traders, and Simon Willison is advising people to hide slop PR descriptions.

China's AI buildout is outrunning controls

DeepSeek's Harness is at 1.54M DAUs, China Mobile plans more than $15B of AI computing in a $75B 2027 procurement, and the NYT found shell-like H200 handlers.

Frontier safety is moving from theory to policy fights

METR flags accelerating automated R&D at Anthropic, CAISSI is being asked to report informally to the White House, and Tencent's Yao Shunyu warns of bio risk while backing closed models.

Signal vs noise

Signal · underreported

CAISSI being asked to report informally to a White House office undermines the independence of the main U.S. government evaluator of frontier models, and almost no one is covering it.

Noise · overhyped

A bubble-denial blog post with no new data or specifics adds nothing to the investment debate.

What to watch

Watch for independent verification of OpenAI's ETH disproof, Commerce's response to the H200 handler findings, and any Anthropic reply to METR's 2x acceleration figure.

Sources

Filter sources
  1. Pattern 1
  2. Story 6
  3. Pattern 1
  4. Pattern 1
  5. Noise
  6. Story 1
  7. Story 4
  8. Pattern 2
  9. Pattern 3
  10. Story 2
  11. Story 5
  12. Pattern 3
  13. Story 3