The blue tsunami won't sink AI, when one model edits another, and agents double Anthropic's research pace

America's midterm elections have been read as a vote against artificial intelligence. Democrats took both houses of Congress and a run of governorships on Tuesday. Voters turned against data centres in 44 of 45 local ballots. By Thursday the Nasdaq was in a correction, Nvidia had given up its gains for the year and the chip index had fallen 12% in a day, on the fear that the people who had just won would slow the AI build-out. However, we've gone through the delivery schedules of 77 of America's largest AI sites, and it turns out the fear is misplaced. Political changes as a result of the election will delay a few gigawatts and remove almost none, most of that delay was already in train, and some of it was initiated by a Republican governor in Texas. As we've covered before, the weak point of the AI boom is debt and, in the US, electrical grid capacity. This week we explore why the record-breaking election result won't change the direction of the AI build-out at least until the next election.

When commentators say AI was on the ballot, they usually mean it loosely, as a theme that shaped how people voted for governors and senators. In Trenton, Ohio, a city of about 13,000 people north of Cincinnati, as in many other communities, it was on the ballot in the literal sense. Below the candidates' names, voters found Issue 17, a charter amendment that began "The construction of a data center is hereby prohibited" and defined a data centre as any facility drawing more than 25 megawatts.

In reality the larger decisions were made through the candidates chosen. Twenty-six gubernatorial candidates across 21 states backed some form of pause on data centre development ahead of the election, and the winners include Gina Hinojosa in Texas, Keisha Lance Bottoms in Georgia, Amy Acton in Ohio and Deb Haaland in New Mexico. Those four states hold the largest share of capacity scheduled for 2027 and 2028. But most of what has been written since Tuesday confuses three different quantities: capacity that is exposed to a possible policy, capacity that is likely to be delayed, and capacity that will not be built. Cleanview's widely shared chart showing 41% of planned capacity in states whose governor backs a pause carried the caveat "this is exposure, not a forecast of cancellations" in its footnote.

To try to understand the true implications for the AI build-out, we built a forecast from the delivery schedules of the 77 large American AI campuses that Epoch AI tracks. For each block of capacity we asked which state approval it still needs and how long any pause is likely to last, then ran a simulation. Half of the capacity due by the end of 2028 is in a state with a pause in force or pledged. Weighted by what is likely to happen, the sites run about 6% behind schedule in early 2028 and 9% behind at the end of that year. Half of that shortfall is made up by the middle of 2029. Capacity still more than doubles, from about 13 gigawatts today to about 30. The local bans change even less. Trenton's ban leaves the Amazon campus under construction next door, outside the city limits, untouched.

Of the 3 gigawatts we expect to be late at the end of 2028, about 1.8 was already in train. Greg Abbott, the Republican governor of Texas, halted new grid connections in August and in September ordered the state's environmental agency to stop issuing permits for data centre infrastructure. The agency complied within a month and, when permits kept slipping through, shut its online filing system. A pending air permit in New Mexico and a transmission case in Wisconsin make up the rest. The election adds the remaining 1.2 gigawatts, mainly by extending the Texas pause. Thirteen of the 20 communities that voted on bans were in Republican-held districts. Nearly every moratorium on record exempts sites already approved, because the construction unions are not to be trifled with.

The Democratic majorities will hold hearings and issue subpoenas, and some members will introduce bills to pause frontier AI development or ban superintelligence outright. Those bills will not become law. The president holds a veto until January 2029 and has built his domestic agenda around the AI build-out, most Senate legislation needs 60 votes, and the new caucus is not what the headlines suggest: Innovation for Good, the pro-AI Democratic PAC, saw its endorsed candidates win 39 of 40 called races. Chuck Schumer's speech on Wednesday setting out the Senate's priorities named the cost of living, corruption and ending the Iran war, and did not mention AI. Democrats know they will be judged in 2028 on electricity and grocery bills, and a federal data centre moratorium would deliver no visible benefit to any voter before then.

Trump's own policy framework in March asked Congress for a ratepayer protection pledge to shield household electricity rates from data centres. Schumer's energy plan includes a promise to make data centres pay their fair share of grid costs. A cost-sharing bill that requires large users to fund their own interconnection and upgrades fits a cost-of-living agenda and has been proposed by both sides. It would regulate who pays for the grid rather than the technology, and it is the only AI-related law we expect to see signed in the next two years. On Thursday OpenAI published a community compact and policy blueprint for what it now calls compute campuses. At the sites it owns and operates from here on, it says it will pay the full cost of its electricity and of the grid upgrades it needs, use closed-loop cooling, publish each site's energy, water and noise figures, and stop asking public officials to sign non-disclosure agreements.

The five largest spenders laid out $230 billion of capital last quarter against $199 billion of operating cash flow and raised the difference from bonds and share sales, a pattern we have written about for months. This week's market volatility sorted the industry by that exposure rather than by politics. The equipment suppliers and the hosts that build with borrowed money fell hardest; Microsoft, Amazon and Alphabet fell least.

As things stand, the spending will continue. Hyperscalers would lose more by pausing while their rivals build than by overspending alongside them. No American government, of either party, will slow its own labs while China's advance.

Takeaways: Nothing decided this week will slow the AI build-out in any meaningful way, and the capacity that matters for 2027 is already under construction. A data centre built to the standards now being demanded, with closed-loop cooling, careful screening, its own power and a guarantee that neighbours' bills will not rise, occupies a few hundred acres and asks very little of the people around it in return for what it adds to economic growth, science, medicine and education. The opposition to it is understandable all the same, because the building is the physical manifestation of AI that can be seen and voted on, while the things people actually fear, about jobs, about who controls the technology and about whether we can maintain control over it, have no ballot box of their own. Fighting the buildings does not answer any of those questions. The harder work of deciding how machine intelligence should be governed, and who benefits from it, still has to be done.

Four researchers at MIT's Computer Science and Artificial Intelligence Laboratory published a blog post this week called "Sparks of Weight-Space Programming", and it has become one of the most discussed pieces of AI research in recent times. They gave OpenAI's GPT-6 Astra the raw weights of Alibaba's small Qwen3.5-0.8B model and asked it to improve the model directly, without any training. Astra performed surgery on over 130,000 of its weights (its internal numerical neurons), and the smaller model's score on the MATH-500 benchmark rose from 8% to 48%.

Within hours, commentators were comparing the work to gene editing for neural networks, and threads on Reddit and Hacker News filled with speculation about models that repair other models and, eventually, themselves. Some readers were more cautious. They noted that the work is a blog post rather than a peer-reviewed paper, that the strongest results without tools came from very small networks, and that the Qwen edit fixed one known fault, a tendency to fall into repetitive reasoning loops. When the team tried to improve a Gemma model that did not have this problem, the gains did not appear.

The method is easiest to place against the established ways of changing what a model does, listed in order of how far each reaches inside the model:

Four ways to change what a model does

Ordered by how far each method reaches inside the model, from the request to the weights themselves. The last is the method in MIT's report.

• Prompting

  Weights unchanged

  Instructions, examples or reference documents go in with each request. The weights stay as they are, so the change costs nothing to apply and lasts only for that request.

• Fine-tuning

  All weights, or an adapter

  The model is trained further on a set of examples, and its weights move a little in whatever direction the data pushes. It needs a dataset, training compute and tests that nothing else has got worse. Adapter methods such as LoRA do the same through a small add-on set of weights.

• Activation steering

  Run time only

  A chosen direction is added to the model's internal activity while it runs, or one feature is clamped at a set level. Nothing is saved to the weights, so the change can be adjusted or switched off, and it needs access to the model's internals.

• Weight-space programming

  Written by a model

  A model is handed another model's weights, works out what they do and writes new values straight into them. There is no training data and no training run, and the editing model rather than a human-written procedure decides what to change.

We wanted to know how practical this would be for a small team, so we ran our own experiment on a small open 135M parameter model. We started from a familiar problem, AI gender bias. Given a sentence about various occupations, the model favoured "he" over "she" more strongly for professions like coach, doctor and engineer. With Astra designing the experiment, we recorded the model's internal calculations, tested which components carried the association and converted the most promising changes into fixed weight edits.

Our first attempts reduced the association by less than 2%. A broader search across 2,010 single edits and 207 combinations, which took about ten minutes on a local Blackwell GPU, produced a four-matrix patch that reduced our measure by 36%. It introduced no new errors on prompts that stated a person's gender, it could be reversed exactly, and it left general prediction quality almost unchanged. The improvement was uneven, however, because the six occupations used during the search improved by 49%, eight new occupations improved by only 15%, and chef and paramedic became worse.

Takeaways: The MIT research shows that a frontier model can read and change another model's weights. Our own experiment shows that it's possible to find a useful, reversible edit for a narrowly defined problem on modest hardware. The difficult work lies in defining the problem, searching widely and measuring what else the edit changes. Editing weights directly could well become a normal way to maintain and improve AI models, and the organisations that benefit most will be the ones with the strongest evaluation practices.

This week's chart comes from METR, the independent evaluation group given access to Anthropic's internal research metrics between August and October. The orange line is the internal frontier, the models Anthropic uses inside the company. The green line is what the public sees, months behind. The dashed lines after today show the two things nobody can yet measure: how much of the acceleration comes from AI itself, and whether anything will slow it down.

Source: METR, Independent assessment of AI R&D acceleration at Anthropic, 5 November 2026. The chart is illustrative and not plotted from data.

METR's answer to the first question is cautious but clear. Anthropic is likely improving its models on well-specified tasks at about twice the rate it managed between March 2024 and October 2025. The 95% range runs from 1.5x to 2.8x. METR believes most of that gain comes from AI assisting researchers rather than from more compute, though it admits the attribution is uncertain and that much of the supporting evidence has been redacted. Staff polled on Slack said their most capable internal model made them about 12x more productive. The system as a whole moved at 2x.

The timing matters. Three days earlier, Anthropic amended its IPO prospectus to say it may imminently meet its own threshold for dramatically accelerated progress attributable to AI. The filing concedes that past accelerations were identified only after they had happened, because the indicators lag. In the same week, investors reportedly pushed the target valuation from $1.5 trillion towards $1 trillion, and the listing may slip.

Several researchers who had expected slower progress now talk about recursive self-improvement within months. METR's own report does not support that reading. It says AI cannot yet make fully autonomous contributions to the most important training work. But whilst Anthropic's financial fortunes are being shaped by politics and market forces, its technical process looks less constrained.