AIThis post was created with the assistance of artificial intelligence (AI).

By Thorsten Meyer | November 15, 2026 | ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

  • Business-only prices and quantity discounts
  • Tax-exempt purchasing
  • Multiple users, one account, clear invoices
As an affiliate, we earn on qualifying purchases.

Key takeaways

  • The core idea: the tool being improved can also contribute to the work of improving it. That is why I take the possibility of rapid progress toward superintelligence seriously.
  • The evidence is getting concrete. On November 5, 2026, the independent evaluation organization METR published its assessment of AI research acceleration at Anthropic. It judged that Anthropic is likely improving its models’ capabilities on well-specified tasks at around twice the earlier rate, with high uncertainty (it “would not be surprised” by 1.3x or 3x), and that the work it described was primarily AI assistance to researchers. Anthropic redacted much of the evidence METR found most informative.
  • A second example: Google DeepMind reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, which cut Gemini’s overall training time by 1%.
  • Small gains can compound. In a deliberately simple illustration, three independent 10% savings leave a job needing about 73% of the original compute, so the same budget runs roughly 37% more of those jobs.
  • The loop has brakes: experiment time and hardware, the gap between small experiments and large systems, choosing the research direction, continuity, and data quality.
  • The transition will be uneven, and plans should allow for a pace of improvement that changes. Who can use these systems, who owns the infrastructure and how the benefits are shared will depend on deployment and institutions as well as technical capability.
Video: When AI Starts Improving AI. Narration: AI voice clone of Thorsten Meyer.

The research loop · schematic

When the tool helps build its successor

The sequence at the heart of this essay: a model contributes to research, tests filter out what does not hold up, and what survives goes into a stronger model that joins the work.

Model Model Model helps with research Tests decide what holds up Survivors are kept Stronger model joins in
  1. The model helps with researchIt proposes changes, writes code, runs experiments and checks results.
  2. Tests are the gateMany ideas fail or bring no real gain. Only what holds up under fair tests gets through.
  3. What survives is keptThe improvements that pass are used to build a stronger model.
  4. The stronger model joins the researchAnd the loop starts again, now with a better helper.

Sources: METR, independent assessment of AI R&D acceleration at Anthropic (published Nov 5, 2026); Google DeepMind, AlphaEvolve (May 14, 2025). The loop is a schematic of the essay’s argument, not a measured process at any lab.

Graphic: Powered by Thorsten Meyer AI

Products Worth Considering

Bose QuietComfort Headphones (2nd Gen), Wireless Over-Ear Noise-Cancelling Headphones, Black

Bose QuietComfort Headphones (2nd Gen), Wireless Over-Ear Noise-Cancelling Headphones, Black

Amazon Kindle Scribe (32GB) with Premium Pen, Tungsten (Like-New)

Amazon Kindle Scribe (32GB) with Premium Pen, Tungsten (Like-New)

As an affiliate, we earn on qualifying purchases.

The next model helps build the one after it

Imagine a research team finishing a new AI model. They give it access to their code, their experiment results, and the problems they have been trying to solve.

The model finds a way to run one part of the training process more efficiently. It helps test a different approach to preparing data. It spots a mistake that was making an experiment look better than it really was.

Some of those contributions survive testing. The team uses them to build a stronger model.

Then that stronger model joins the research process.

That sequence is why I take the possibility of rapid progress toward superintelligence seriously. The tool being improved can also contribute to the work of improving it.

I build with AI, so my interest starts with what these systems let people do. Can you turn an idea into working software? Can you investigate something properly? Can you maintain a service, find a failure, and fix it?

Now apply those same questions to an AI laboratory. Can the model help design the next experiment? Can it tell whether the result means anything? Can it help produce a successor that does those things better?

As more of that work becomes reliable, the pace of development itself can change. And that would affect every business and every person trying to plan around AI.

I think this deserves a practical explanation. What exactly is being automated? How could the gains compound? Where could the process slow down? And how would this connect to a world where far more work can be done with far less human labor?

Start with an ordinary research day

The phrase automating AI research can make the whole subject sound impossibly abstract. So imagine one fairly ordinary assignment.

A team wants to reduce the amount of computing needed to reach a particular level of model performance.

Someone has to read the relevant research. Someone has to inspect the existing implementation. An experiment needs code, data, configuration, and a way to measure the result. Failed runs need investigating. Promising runs need repeating. Eventually, someone has to decide whether an apparent improvement is worth using.

There are many opportunities for assistance along the way. There are also many opportunities to make a convincing mistake.

Illustration: suppose an agent changes the training code and reports a 5% improvement. That sounds encouraging. But perhaps it accidentally made the test easier. Perhaps the two runs used different amounts of compute. Perhaps the gain disappears when you use a different random seed.

Useful research includes finding those problems.

An automated researcher would need to make a proposal, implement it, test it fairly, examine what happened, and revise its approach. It would also need to recognize when the evidence is too weak to justify a conclusion.

That gives us a concrete way to think about progress. We can ask how much of that sequence a system can handle, how much supervision it needs, and whether its contributions remain valuable after independent checks.

Even partial automation can matter. If a human researcher can investigate more ideas and spend less time diagnosing routine failures, the same team can accomplish more. Full autonomy would extend that process further, including choosing what to investigate next.

Evidence that the process is already changing

We have a recent assessment that makes this discussion much more concrete.

On November 5, 2026, the independent evaluation organization METR published its assessment of AI research acceleration at Anthropic. Its team had worked with internal information from August through October.

METR judged that Anthropic’s capabilities on well-specified tasks were improving at roughly twice the earlier rate. It considered AI-assisted research likely responsible for most of the current acceleration, with substantial uncertainty. The work it described was primarily assistance to researchers.

Graphic titled Around twice the earlier rate, with a wide range. METR, November 5, 2026: pace of capability gains at Anthropic on well-specified tasks compared with the historical rate of March 2024 to October 2025, as of October 2026. Highlighted: METR's overall judgment, about 2x, and it would not be surprised by 1.3x to 3x. Benchmark trend AECI since October 2025: 1.9x, 95 percent confidence interval 1.5x to 2.8x. Notes: R&D is accelerated primarily via AI assistance rather than fully autonomous contributions; Anthropic redacted a majority of the evidence METR found most informative for these conclusions
METR's finding and its range. Source: METR, "Independent Assessment of AI R&D Acceleration at Anthropic in October 2026", November 5, 2026. Graphic: Thorsten Meyer AI.

The public evidence also has a limitation. Anthropic redacted much of the information METR considered most informative. That leaves outsiders relying heavily on the evaluators’ judgment.

We covered the METR report in detail when it came out: Is AI Now Speeding Up Its Own Development? What METR Found Inside Anthropic, and What Stayed Redacted.

A different example comes from Google DeepMind’s AlphaEvolve. It combines language models that propose programs with automated evaluation and an evolutionary search process.

DeepMind reported that AlphaEvolve helped optimize a matrix-multiplication component used in Gemini. That component became 23% faster, translating into a 1% reduction in overall training time.

Graphic titled A faster component, a shorter training run. How AlphaEvolve works: Gemini models propose new programs, automated evaluators verify the answers, an evolutionary framework improves on the most promising ideas. A matrix-multiplication kernel in Gemini's architecture became 23 percent faster. Highlighted: 1 percent less overall Gemini training time, shown as a schematic bar that is not to scale
AlphaEvolve: a component 23% faster, overall training time 1% shorter. Source: Google DeepMind, May 14, 2025. Graphic: Thorsten Meyer AI.

One percent can sound modest. Across a costly training process, it can be valuable. More importantly for this discussion, it illustrates a specific route through which AI can contribute to the machinery used to develop AI.

These examples give us something tangible to build an argument around. We can examine which parts of the process improved and then ask how much further the same approach could go.

Why measurable tasks are an attractive starting point

Consider three possible assignments for an AI research agent.

The first is to train a small model to the same target accuracy using less compute. The second is to improve how reliably a model solves a set of unfamiliar programming tasks. The third is to make an existing model respond faster while preserving the required quality.

Each assignment has something we can measure. We can record compute consumption, evaluate the outputs, and compare the result with a baseline.

That creates an environment where an agent can try an idea and receive feedback. It also makes it possible to reject many poor ideas before a human researcher has to examine them in detail.

Illustration: imagine that an agent proposes 10 changes. Seven fail immediately. Two offer no meaningful gain. One looks promising. It repeats that experiment and discovers that the improvement only works with a particular kind of data.

Now it has a more specific question to investigate. What property of that data makes the difference? Can the benefit survive a change in scale? Does it help on a fresh evaluation?

That is already a substantial research process. It includes failure, comparison, interpretation, and another experiment.

The difficult part is choosing measurements that remain connected to the real objective. A model can become excellent at a familiar test while still failing on the work we actually need it to perform.

So a convincing improvement should survive different tasks, new data, repeated trials, and comparison under the same resource constraints. Some ideas will also need a large and expensive experiment before we know whether they matter.

The opportunity comes from making this whole process more productive. Better proposals help. Better tests help. Faster debugging helps. So does abandoning a misleading result before it absorbs another month of work.

How improvement could compound

Once an AI contributes to developing its successor, we have a feedback loop. A better research system may help produce a better research system again. That is the central possibility behind recursive self-improvement.

The loop can run through several different mechanisms.

A model might help make training more efficient. It might improve the software used to run experiments. It might help create better training tasks. It might make the research agents themselves cheaper to operate.

Those changes have different consequences. A reduction in inference cost can allow more agent work within the same budget. A reduction in training cost can allow more experiments. An improvement in research quality can make each experiment more informative.

Here is a deliberately simple numerical example.

Illustration: suppose three independent improvements each reduce the compute needed for a particular job by 10%, and suppose they combine cleanly. The job now needs about 73% of the original compute (0.9 × 0.9 × 0.9 = 0.729). The same budget could support roughly 37% more of those jobs (1 ÷ 0.729 ≈ 1.37).

Graphic titled How three small savings compound, labelled illustration, not measured data. Compute needed for one job with three independent improvements each saving 10 percent: start 100 percent, after the first 90, after the second 81, after the third 72.9 percent. 0.9 times 0.9 times 0.9 equals 0.729, so the job needs about 73 percent of the original compute. Highlighted: 1 divided by 0.729 is about 1.37, so the same budget runs roughly 37 percent more of those jobs. Assumption: the gains combine cleanly
The essay's compounding example, as an illustration that assumes the gains combine cleanly. Arithmetic: Thorsten Meyer AI.

That is an illustration of compounding. In an actual system, the improvements might overlap, conflict, or apply to different parts of the workload. The arithmetic only works under those assumptions.

But it explains why a sequence of relatively small gains can matter. An expensive experiment becomes affordable. A slow test becomes fast enough to run routinely. A research assistant becomes cheap enough to use more extensively.

There is also an advantage in copying a successful process. If an effective agent can work on several independent problems, a lab can deploy additional instances within its compute budget. Useful procedures and results can be shared among them.

The value of those copies depends on what they actually contribute. Ten agents proposing the same failed idea add very little. Ten agents exploring genuinely different approaches and recording reliable results could add a great deal.

So I would judge an automated lab by the quality of its accepted discoveries and the resources they consumed. That tells us more than a count of agents or experiments running in the background.

What could slow the loop down?

One immediate constraint is the time required for experiments.

An agent may prepare an experiment quickly, but the training run still has to happen. If the next decision depends on its result, that part of the work has to wait. More agents also compete for the same hardware unless the lab expands its capacity.

Automation could help use that capacity more effectively. An agent might identify a broken run earlier, suggest a smaller preliminary test, or examine results that would otherwise receive little attention. How much that helps is an empirical question.

Another constraint is the relationship between small experiments and large systems.

An idea that looks excellent in a small model may behave differently when the model, dataset, or training duration changes. Researchers still need a way to establish which results will transfer. Making that judgment more reliable would itself be a valuable contribution.

Then there is the problem of choosing the research direction.

You can give an agent a clearly defined objective and measure its performance. Deciding whether that objective is the right thing to pursue can be harder. A lab needs to allocate resources among incremental improvements, uncertain new approaches, and work that may only pay off much later.

An automated research organization would need to make those decisions too. Producing more experiments gives it more evidence, but interpreting that evidence remains part of the job.

Continuity is another challenge. A research project needs a reliable record of what was tried, why it failed, and what questions remain open.

There are several ways to support that continuity. Agents can use experiment logs, retrieve earlier findings, maintain project records, and carry information through a workflow. Updating the model’s learned parameters is another mechanism, with its own difficulties.

For the research project, the practical test is whether earlier experience improves later decisions. Can the system avoid repeating an expensive mistake? Can it recover the assumptions behind an old result? Can it recognize that two apparently different failures have the same cause?

Finally, there is data quality. Generating more examples is useful when those examples teach something relevant and reliable. If they contain systematic errors, repetitions, or shortcuts, greater volume can amplify the problem.

Graphic titled What could slow the loop down. Highlighted: experiment time and hardware, the training run still has to happen and more agents compete for the same hardware. Small to large: an idea that works in a small model may behave differently at scale. Choosing the direction: measuring a goal is easier than knowing it is the right goal. Continuity: a reliable record of what was tried, why it failed and what is still open. Data quality: more examples help only if they are relevant and reliable
Five brakes on the loop, from the essay's own analysis; not measured data. Graphic: Thorsten Meyer AI.

These constraints are reasons to measure the whole process carefully. They also identify work that increasingly capable research agents could help address. The pace will depend on how quickly those improvements arrive and which constraints remain after they do.

How far does this take us toward superintelligence?

For this discussion, I use superintelligence to mean a system that substantially exceeds the best human performance across a broad range of important cognitive work.

AI research is one route toward that possibility because improvements in learning methods can affect many downstream capabilities. A system that learns more efficiently, reasons more reliably, or uses tools more effectively could become useful across many domains.

The crucial issue is transfer. How much of an improvement in a controlled research setting carries over to an unfamiliar real-world problem?

Illustrations: suppose an AI has become exceptionally good at optimizing software. Now give it responsibility for improving a construction project. It will need to understand incomplete records, physical conditions, scheduling constraints, and disagreements between people with different information.

Or ask it to help develop a medical treatment. It may contribute to literature analysis and experiment design, while the evidence needed to establish safety and effectiveness takes time to collect.

Or ask it to run a new business. Writing the software and building a financial model are only part of the task. Customers have to respond. Distribution has to work. Some consequences become visible months later.

Greater intelligence could improve decisions in all of these settings. It could also make better use of each piece of feedback. But we would have to observe that performance in the relevant environment.

This is why I expect the transition to be uneven. Digital work with rapid feedback is well suited to repeated testing. Work involving physical deployment, human relationships, or delayed outcomes will develop on different schedules.

Even within a single organization, you could see extraordinary automation in one department while another remains limited by practical conditions outside the model.

For anyone building with AI, that unevenness creates specific opportunities. Look at the work that can already be performed and checked reliably. Then ask what additional information, tools, or feedback would make the next part possible.

What I would watch over the next few years

I take substantial acceleration seriously. A fixed date for superintelligence would require confidence in several unresolved questions at once.

How quickly will agents become dependable across an entire research project? How much will their contributions improve their successors? How strongly will those gains transfer to work outside the laboratory?

I would watch those questions through concrete milestones:

  • Can an agent complete a valuable research project with much less intervention?
  • Do its results hold up when another team reproduces them?
  • Does a successor model become measurably better at research because of those contributions?
  • Do the gains remain worthwhile after accounting for the compute spent finding them?

Repeated success on those measures would move my expectations much more than a spectacular isolated demonstration.

Several paths are plausible. These are scenarios, not forecasts:

  • Progress could remain sustained but uneven. Agents absorb more research work, yet difficult experiments and integration keep the development cycle relatively long.
  • There could be a period of much faster progress as several bottlenecks ease together. That acceleration could then slow while hardware, data, or evaluation methods catch up.
  • Or the feedback could become strong enough that successive improvements keep shortening the next cycle. That would create a much more compressed transition, especially if research quality and compute efficiency improved together.

Which path develops depends on the balance between the gains from automation and the increasing difficulty of further progress.

My working view is that planning should allow for a pace of improvement that changes. A business strategy built around the capabilities of one model generation can become outdated if the process producing the next generation is also getting faster.

Why this matters beyond the AI labs

For me, the most interesting consequence is what this could make possible for people with limited resources.

Many ideas stop because carrying them out requires too much time, too much specialist work, or a team that someone cannot afford to hire. More capable and affordable AI could change those constraints.

A small company could investigate a technical problem more thoroughly. A researcher could explore an approach that would otherwise wait for funding. A person with deep knowledge of an industry could build and maintain software for it with much less help.

Those are the possibilities that connect this subject to my interest in a post-labor economy.

As more economically valuable work becomes automatable, we have to think about how people gain access to the results. Who can use the systems? Who owns the infrastructure? How are the benefits shared? What happens to income when the market needs fewer human hours for the same output?

The answers will depend on deployment and institutions as well as technical capability.

If increasingly powerful tools remain affordable and widely accessible, individuals and small organizations could take on projects that currently require much larger institutions. If access becomes highly concentrated, the distribution of those benefits could look very different.

I am optimistic about the amount of useful work that AI could make possible. I also care about whether people gain more agency as that happens. The ability to create, investigate, solve problems, and pursue meaningful work is a significant part of the opportunity.

Research acceleration matters because it could bring those capabilities forward. It also gives us less time to adapt if we only begin thinking about the consequences after they arrive.

What is not known

  • How strong the acceleration really is. METR’s own range runs from about 1.3x to 3x the historical rate, and it has additional uncertainty about how the acceleration translates to broader and less well-specified tasks.
  • What the redacted evidence shows. Anthropic redacted a majority of the evidence METR found most informative, so the public has to rely on the evaluators’ judgment.
  • How dependable agents become across an entire research project, and how much less human direction they will need.
  • How much their contributions improve their successors, and whether the gains remain worthwhile after accounting for the compute spent finding them.
  • How strongly the gains transfer to work outside the laboratory: physical projects, medicine, business.
  • Which path develops. The three paths above are scenarios, not forecasts. I give no date for superintelligence.

The next thing to measure

When the next major model is released, I will be interested in what it can do. I will also want to know how it was built.

What research did the previous generation contribute? Which improvements survived testing? How much human direction did the process require? And does the new model make that same process more effective?

Those questions tell us whether AI development is becoming increasingly able to accelerate itself.

A research team using AI effectively can already change what it accomplishes. A succession of models that each help build a better successor could change how quickly the rest of us encounter new capabilities.

I want to understand that process closely, because I see enormous potential in what people could build with it.

If you are using AI in your own work, tell me where it has changed what you can actually accomplish. Which part of a project has become easier, and where do you still have to step in?

I will keep writing about AI, superintelligence, and what increasingly capable systems could mean for work and everyday life.

Sources

  • METR: “Independent Assessment of AI R&D Acceleration at Anthropic in October 2026”, November 5, 2026. metr.org
  • Google DeepMind: “AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms”, May 14, 2025. deepmind.google

This is an opinion essay; the views are my own. Hypothetical examples are labelled as illustrations, and the future paths are scenarios, not forecasts. Researched and drafted with AI assistance, including Anthropic’s Claude. Anthropic is one of the companies this essay discusses; neither Anthropic, Google DeepMind nor METR reviewed or sponsored it. Our own text and graphics contain no affiliate links; offer and product boxes added by the site may contain Amazon affiliate links. Sources checked on November 15, 2026. Images: AI-generated illustration (Runware), animation and graphics by Thorsten Meyer AI; no real people or logos are shown in our images. Boxes added by the site may show product photos or people.

You May Also Like

The Model They BuiltBut Won’t Release

AIThis post was created with the assistance of artificial intelligence (AI). The…

Doge Uses AI to Draft Federal Regulation Cuts

Doge reportedly using AI tool to create ‘delete list’ of federal regulations, streamlining policies for efficiency.

Claude Did Not Invent Scorpion Antivenom – The Verge

A headline says Claude did not invent scorpion antivenom, but the available material provides no article details to verify the claim or explain its basis.

The Channel Move: Anthropic, Wall Street, and the Acquisition of the Real Economy

AIThis post was created with the assistance of artificial intelligence (AI).By Thorsten…