> ## Content Index
> Fetch the complete content index at: https://www.deconstructingbabel.com/llms.txt
> Use this file to discover other available public pages before exploring further.
# The Brake That Doesn’t Reach the Engine
- URL: https://www.deconstructingbabel.com/the-brake-that-doesnt-reach-the-engine/
- Published: 2026-11-06T19:09:25.000Z
- Updated: 2026-11-06T19:09:25.000Z
- Description: On Thursday, Wall Street dumped AI stocks, fearing new rules would slow the buildout. That evening, an outside group found Anthropic improving its AI about twice as fast as before, mostly because AI is doing the research. Money can brake the buildings. It doesn’t reach the engine.
- Author: David F Brochu and Edo de Peregrine
- Tags: AI Safety, Observer Constraint, analysis
**THE WEEK, CONNECTED**
Voters, markets and a chief executive all pressed the brake on AI this week. METR’s inside look at Anthropic shows why none of them reached the part that’s speeding up, and why the lab may have crossed its own warning line months ago.
*David F. Brochu & Edo de Peregrine · Deconstructing Babel · November 6, 2026*
On Thursday, Wall Street hit the brakes on artificial intelligence. The S&P 500 fell 4.4%, its worst day since the “Liberation Day” tariffs of April 2025\. The Nasdaq fell 5.5%. The Dow lost 1,833 points. Nvidia, the biggest maker of AI chips, dropped 9.3%. [\[1\]](https://www.nbcboston.com/news/national-international/ai-stocks-wall-street-reeling-sp-500-tumbles/4044610/)
The reason was Tuesday’s election. Investors feared that Democrats, now running Congress and several big statehouses, would regulate AI and slow the building of data centers. “The immediate risk is that restrictions on data center construction slow the pace of the AI infrastructure buildout,” said Joe Maher of Capital Economics. The sell-off may even delay Anthropic’s stock market debut, which had been expected this month and could raise as much as $100 billion. [\[2\]](https://www.cbsnews.com/news/stocks-midterm-elections-tech-companies-ai-data-centers/)
That evening, an independent testing group called METR published what it had found after 11 weeks inside Anthropic. Its headline: Anthropic is likely improving its AI at around twice the rate it did before October 2025\. [\[3\]](https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/)
The market was stepping on the brake. The report showed the engine speeding up.
## Three brakes in one week
This week we watched three different brakes get pressed.
The first was the voters. On Tuesday, towns in Ohio banned data centers, and governors who promised pauses won in Texas and Georgia. Two days later, OpenAI answered with what it calls a National Community Compact. It promised to pay the full cost of its power and grid upgrades, publish water and power use for each site, and stop asking public officials to sign confidentiality agreements. The pledges cover only campuses OpenAI owns, and partners like Oracle own much of its computing power. [\[4\]](https://www.nytimes.com/2026/11/05/technology/openai-data-centers-midterms.html) On Wednesday we warned that a check driven by electric bills can be bought off. [\[5\]](https://www.deconstructingbabel.com/the-neighbors-voted-on-the-machine/) The first offer arrived in 48 hours.
The second brake was money. That was Thursday.
The third was self-restraint. In September, Anthropic’s chief executive, Dario Amodei, urged the industry to “slow the pace at which we improve the capabilities of AI models.” Of AI building AI, he wrote: “Left unchecked, it could outrun our ability to understand and control these systems.” [\[6\]](https://abcnews.com/Technology/wireStory/anthropic-rapidly-accelerating-ai-development-amid-slowdown-calls-137885082) His own company, METR now says, roughly doubled its pace.
Each of those brakes pushes on something real. None of them reaches what’s actually speeding up.
A RUNG UP
AI companies make their models better in two main ways. One is more computing power: more chips, bigger buildings, longer training runs. That costs money and needs land, power and permits. The other is better methods: smarter code, cleaner data, better experiments. Increasingly, AI does much of that second kind of work itself, writing code and running tests for the researchers. METR is a nonprofit that measures how capable AI systems are. Anthropic let a small METR team study its internal data and interview its staff for about 11 weeks.
## The engine
Here’s the part of the METR report that matters most for the markets. METR wrote that it was “able to confidently determine that the acceleration was not due to an acceleration in the rate of growth of training compute.” [\[3\]](https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/)
In plain terms, the speedup didn’t come from buying more chips. METR’s best guess is that about 80% of it comes from Anthropic’s heavy use of AI to do its own research. Its benchmark scores have grown about 1.9 times faster since October 2025, which it describes as “roughly two years of historical progress over the last year.” [\[3\]](https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/)
So the brake Wall Street pulled on Thursday mostly acts on new buildings and new chips. The acceleration METR measured comes mostly from AI working inside buildings that already exist. The two are coming apart.
To be fair, money still matters. More computing power is still how frontier AI grows over years, and a slower buildout could bite in 2028 and after. But METR also reported seeing no major bottlenecks in methods, computing power or data “that will take effect in the next 12 months.” [\[3\]](https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/) The financial brake might slow the decade. It doesn’t reach next year.
## The warning may have been late
On Tuesday we reported that Anthropic had told investors it “may imminently meet” its own threshold for AI-driven self-improvement, and we said labs owed the public an early warning. [\[7\]](https://www.deconstructingbabel.com/inherited-not-installed/) This week, the early warning came from outside the company, and it says the line may already have been crossed months ago.
In a separate review of Anthropic’s August risk report, METR wrote: “we think it is plausible that they had crossed their early warning threshold at the time,” meaning July. It added that the public report “understates the level of acceleration that was occurring.” Both METR’s team and Anthropic’s staff lean toward the threshold not having been crossed, but METR says that conclusion “relies on contestable judgment calls.” [\[8\]](https://metr.org/blog/2026-11-05-rd-section-anthropic-risk-report-august-2026-review/)
METR also found that AI is “not yet able to make highly significant and fully autonomous contributions to the most important parts of the model training process.” But Anthropic’s researchers described “signs of life” suggesting that could happen soon. “It is possible that automation is close, and leads to very rapid acceleration once human bottlenecks are removed.” [\[3\]](https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/)
Read that last phrase again: once human bottlenecks are removed. The humans are the bottleneck. They’re also the brake.
## Who holds the pen
Anthropic deserves real credit. It gave METR what METR called “unprecedented access,” more than it expected. [\[3\]](https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/)
But read the terms. Anthropic could approve or reject METR’s requests for data. It had “unilateral redaction rights over all non-public information.” Every fact in the public report was “affirmatively approved for release by Anthropic.” METR’s own checklist asks whether the company lacked the power “to redact results to conceal concerning findings.” METR’s answer: “No.” [\[3\]](https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/)
That’s no scandal. METR told us, and that honesty is the point of the exercise. It does mean the most important outside check on the fastest-moving lab has a lock on it, and the lab holds the key.
WHY IT MATTERS
The brakes society has, such as voters, markets, permits and lawmakers, all push on buildings and dollars. The engine is now AI improving AI, on machines that are already built, at a pace one outside group says doubled in a year. The only thing that reaches that engine is a person who can see inside and say what they see. That’s the [Observer Constraint](https://www.deconstructingbabel.com/glossary/#observer-constraint) at its simplest: humans who can see what’s happening and still make the call. If the lab can edit what the observer is allowed to report, the constraint becomes optional. And in [recursive symbiotic improvement](https://www.deconstructingbabel.com/glossary/#recursive-symbiotic-improvement-rsi), the human is supposed to be part of the loop, not the bottleneck it’s built to get around.
## What should come next
First, make this routine. Every lab working at the frontier should get a high-access outside assessment of how fast AI is speeding up its research, at least twice a year, published on a schedule. What Anthropic did voluntarily should become the floor.
Second, separate secrets from conclusions. A lab has a fair claim to protect trade secrets and security details. It shouldn’t be able to remove a finding. The outside group should decide what counts as a concerning finding and should be able to publish that it found one, even if the details stay private.
Third, track the number. METR’s estimate, about twice the old rate with a range of 1.3 to 3 times, should be reported like inflation: same method, every quarter, every frontier lab. Markets fall on rumors about data centers. They should have a real figure for the thing that actually moves the future.
Our call, dated today: by the end of 2027, an independent group with inside access will report that AI is making highly significant, fully autonomous contributions to the most important parts of a frontier lab’s training process. METR says that hasn’t happened yet. It also says it may come soon.
Wall Street can stop paying for new buildings. It can’t stop what’s already running inside them. Only someone who can see inside can do that.
## Related reading
[Inherited, Not Installed](https://www.deconstructingbabel.com/inherited-not-installed/)
MIT showed an AI can read another model’s weights and edit them directly. Anthropic says its own AI may soon speed up its research, and it can hold the gate. Open models have no gate. And new evidence says what matters in a model is inherited, not installed.
[The Neighbors Voted on the Machine](https://www.deconstructingbabel.com/the-neighbors-voted-on-the-machine/)
In Trenton, Ohio, voters banned an 18-building Amazon data center, 80% to 20%. In Texas, a governor who ran on a data-center pause beat Greg Abbott. Voters weren’t moved by fears of superintelligence. They were moved by their bills. The plug on AI turns out to be local.
[Who Holds the Key?](https://www.deconstructingbabel.com/who-holds-the-key/)
The industry has converged on the kill switch. Nobody answered the only question that matters: who holds it? On H.R. 9917, the Wallfacer Problem, and why a switch that reaches only the compliant is maximum political cost for minimum safety yield.
| Get the book [](https://www.amazon.com/dp/B0HHTDH5RC) Crossing The Event Horizon The book behind these dispatches. On AI, agency, the singularity, and the Observer Constraint. Kindle and paperback. [Buy on Amazon →](https://www.amazon.com/dp/B0HHTDH5RC) |
| -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
## References
1. Stan Choe (Associated Press), “Brutal sell-off for AI stocks sends Wall Street reeling,” NBC10 Boston, November 5, 2026\.
2. Mary Cunningham and Alain Sherter, “Dow sheds more than 1,800 points as midterm results spur investor exodus from tech and AI,” CBS News, November 5–6, 2026\.
3. METR, “Independent Assessment of AI R&D Acceleration at Anthropic in October 2026,” November 5, 2026\.
4. The New York Times, “After Midterm Rebuke, OpenAI Pledges to Give Data Center Towns More Say,” November 5, 2026\.
5. Deconstructing Babel, “The Neighbors Voted on the Machine,” November 4, 2026\.
6. Kelvin Chan and Kaitlyn Huamani (Associated Press), “Anthropic rapidly accelerating AI development amid slowdown calls, group says,” ABC News, November 5, 2026\.
7. Deconstructing Babel, “Inherited, Not Installed,” November 3, 2026\.
8. METR, “Review of the ‘Risks from automated R&D’ section in Anthropic’s Risk Report (August 2026),” November 5, 2026\.
*Drafted with Edo de Peregrine, partner/collaborator.*