# Does Anthropic's research speedup put superintelligence a year away?

**Summary:** METR's Anthropic assessment supports faster AI research. I favor more frequent capability reviews; superintelligence within a year remains a conditional extrapolation.

- Canonical: https://markhuang.ai/news/anthropic-research-speedup-compounding
- Language: en
- Author: [Mark Huang](https://markhuang.ai/about)
- Published: 2026-11-06
- Section: News
- Tags: Anthropic, METR, AI Research, AI R&D Acceleration, Recursive Self-Improvement, AI Forecasting
- Source: [Planned Obsolescence](https://www.planned-obsolescence.org/p/ai-is-dramatically-accelerating-ai)
- License: https://creativecommons.org/licenses/by-nc/4.0/

---

![A translucent teal ribbon loops between two processor packages and disappears behind opaque charcoal panels](https://cdn.markhuang.ai/news/anthropic-research-speedup-compounding/hero.webp)

*Conceptual illustration: AI research can feed back into better AI, while redacted evidence makes the strength of that feedback harder to judge from outside.*

AI that helps researchers build better AI could shorten the wait for more capable tools. A new finding makes that possibility more concrete. In a [November 5 assessment](https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/), METR estimates that Anthropic's progress on an internal index of AI test results since October 2025 ran at 1.9 times its earlier pace, with a 95% confidence interval of 1.5 to 2.8.

In [Planned Obsolescence](https://www.planned-obsolescence.org/p/ai-is-dramatically-accelerating-ai), Ajeya Cotra argues that this acceleration could compound. Her striking extrapolation is extreme superintelligence within a year, conditional on a smoothly strengthening feedback loop.

I take the evidence seriously enough to favor more frequent reviews of what AI can do. A one-year arrival date remains a conditional scenario: each improvement has to speed up the next strongly enough to keep the intervals shrinking. For anyone choosing AI tools or planning work around today's limitations, separating those claims helps decide what to revisit and how urgently.

## What the faster pace actually measures

METR's headline benchmark result concerns AECI, Anthropic's version of the [Epoch Capabilities Index](https://epoch.ai/eci). These indexes combine results across benchmarks into a capability scale. A faster rise means models are making more benchmark progress per month. It does not mean every task finishes 1.9 times faster, or that a business using Claude gets a matching productivity gain.

The assessment included internal models that had not been released. That matters: the finding concerns Anthropic's research progress, including capabilities customers could not yet buy. METR also cautions that its conclusions concern well-specified tasks, with more uncertainty about broader, messier work.

If AI helps researchers test ideas and improve their models more quickly, users could eventually receive better tools sooner. I find that possibility worth following even without accepting a prediction about superintelligence. The distance between an internal result and a dependable product still matters to the person paying for it.

## The speedup and its cause are different findings

[METR's staff assessed Anthropic](https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/) from August 12 to October 29, 2026, with access to internal metrics and interviews. They concluded that faster growth in training compute did not explain the acceleration. They judged that a majority of the current acceleration probably came from AI use, while acknowledging substantial uncertainty and other contributors, especially earlier in the period.

I give that judgment weight because METR examined internal information. But Anthropic redacted a majority of the evidence METR found most informative. METR says this forces the public to rely more heavily on its judgment than would be ideal. Outside readers can examine parts of the methodology, but cannot independently inspect much of the evidence behind the attribution.

Even the visible benchmark analysis depends on analytical choices. METR reports acceleration estimates from 1.1 to 2.9 times the earlier rate under alternative assumptions about how benchmark performance relates to underlying capability. The headline confidence interval does not cover all of that methodological uncertainty.

The reported 12x individual uplift needs its own caution. It comes from staff self-reports in Slack polls, which METR calls very loose evidence. It is not a measured twelvefold increase in the lab's output. METR also describes the work as primarily AI assistance, rather than fully autonomous contributions to the most important parts of model training.

## One faster lap does not settle the next one

Cotra, who says she neither worked on the assessment nor saw the unredacted report, asks what happens if better models keep increasing the rate of research progress. In her heuristic, another doubling arrives in six months, the next in three months, then six weeks. The intervals shrink because the research process itself keeps getting faster.

```mermaid
flowchart TD
    A[More capable AI] --> B[AI assists research]
    B --> C[Researchers develop better models]
    C --> A
    C --> D{Does each cycle speed up enough?}
    D -->|Yes, if limits do not intervene| E[Progress can accelerate further]
    D -->|No| F[The pace need not keep doubling]
```

The diagram illustrates the proposed feedback mechanism; the report does not measure a timetable for its future cycles. Maintaining twice the old pace would already be a substantial change. Repeatedly doubling the pace adds a stronger requirement: the feedback must keep overcoming diminishing returns and other constraints.

A useful counterweight is Tom Davidson and Tom Houlden's [Forethought analysis of a software intelligence explosion](https://www.forethought.org/research/how-quick-and-big-would-a-software-intelligence-explosion-be). Their model makes the outcome depend on the initial speedup, returns to further research, and how far software can improve before reaching effective limits. They explicitly describe the exercise as speculative. It helps explain why observing acceleration alone cannot determine its eventual size or duration.

METR found no clear near-term bottlenecks that would prevent further acceleration. That makes further acceleration a possibility I take seriously. Its authors also say their understanding changed repeatedly during the assessment and that they may be missing important constraints. Their concern deserves attention, and the uncertainty belongs in the forecast.

## Shorter review intervals, with concrete tests

My practical response is to shorten the time between reassessments when a plan depends on an AI capability remaining weak. A workflow that makes sense with today's tools may deserve another look when a materially better model becomes available. The reassessment should test useful completed work, including the effort needed to check it.

That connects to [my earlier argument about the cost of proving an AI-generated patch](https://markhuang.ai/news/coding-agents-patch-proof-cost): faster production helps most when the result can be accepted and used. A research benchmark can justify trying a new tool; it cannot tell a particular team how much review work that tool will leave behind.

The next model generation offers a useful check: has the time needed for a comparable research advance shrunk again? I also want to know how much the result changes under different measurement choices, and whether the gains carry into messy tasks. That would give me a firmer basis for changing a plan than the one-year extrapolation alone.