Are we in the good timeline?

A new model ships. It's noticeably better at coding, research or operating a computer. And someone asks: whatever happened to "pacing the frontier"?

It's a fair question with a mistaken premise. The release calendar isn't a live feed of the research frontier. And the pacing push this July was never a promise that products would stop improving. It was a warning that automated AI development might start moving faster than our ability to understand and control what it produces, along with a request for ways to coordinate around that.¹

That warning is no longer hypothetical. In the past few months we've seen evidence of stronger internal models, of AI doing a real share of AI research, and of labs restricting particular kinds of training. In the same period we've seen concrete failures of control. All of these are true at once.

My read: meaningful, human-directed self-improvement is already happening. The evidence that the feedback loop exists is much stronger than the evidence that anyone has made it reliably safe.

The question is no longer just how fast the models are improving. It's whether the people building them are getting better at steering faster than the systems are getting more powerful.

01The AI you can use is not the frontier

When people say "the frontier," they usually mean whatever the best public chatbot can do. But at least three different things hide under that word.

A public model failing at a task tells you something real about that model. It tells you very little about a different model running inside a different research setup. And there's direct evidence that the gap is real:

• OpenAI's Navier–Stokes work (September) described an internal model substantially more capable than the public GPT-6 Astra. Training was still running, and researchers swapped in a further-trained checkpoint during the project.²
• OpenAI's mathematics release (October) came from an unreleased internal model that attempted about 4,000 research problems and produced 719 manuscripts grouped into 372 result families. Existing math tests had saturated, which is why the work moved to open problems. Verification is incomplete, and a manuscript isn't a proven breakthrough. Still, it's strong evidence of broad research ability beyond what customers can access.³
• Anthropic's August risk report described an internal-only "Model 2" as a noticeable step up from Mythos 5 on many internal tasks. It also cautioned that the jump was smaller than an earlier generational leap.⁴
• METR estimated that internal models were about 66 days ahead of public ones on one autonomous-task measure. That was a narrow February–March snapshot, not a universal constant. It rests on a small sample with privacy noise added, so it doesn't establish a precise lead today. It does show the lag can be substantial. Corrected 2026-10-25⁵

None of this means an arbitrarily powerful secret system must exist. It means that public products don't show you everything that has already been built.

02AI is already helping to build AI

You don't have to infer this from benchmarks. The labs are saying it themselves.

• In September, Dario Amodei attributed recent acceleration mainly to AI's growing contribution to building the next AI systems. He described recursive self-improvement as beginning across the industry, Anthropic included.⁶
• Anthropic measured Claude taking the lead on 26% of its assessed R&D work in August and collaborating substantially on more than 90%. None of the measured categories was fully autonomous.⁷
• OpenAI describes an automated "research intern" that carries out well-defined, multi-day tasks under human direction. That's not an autonomous lab, but it is a real part of the process that produces stronger successors.⁸

Then there's the striking number. Anthropic's daily code merges went from about 2,500 in May to more than 30,000 in September. That was enough growth that source control, builds and testing became the bottleneck.⁹

So is AI research running twelve times faster? No. The closest thing we have to a whole-lab measurement is a preliminary estimate relayed by METR in September. A separate team with deeper access to Anthropic concluded that AI was speeding up its overall progress by around 1.5×, with perhaps a 30% chance of 2×. It's an imperfect anchor. The period it covers wasn't specified, METR's publishing team couldn't inspect the evidence behind it, and METR's own conclusion was more cautious: development was somewhat accelerated, but probably not dramatically.¹⁰

Why the gap between 12× and 1.5×? Because research isn't only writing code. Experiments still take time to run. Ideas still need judgment. Whatever AI doesn't speed up becomes the limit.

The honest distinction is between a working feedback loop and an explosive, fully autonomous one. The evidence supports the first. How large it is, how long it lasts and whether it's accelerating are far less well measured.

Google's model card for Gemini 4 Argon says the model didn't meet Google's thresholds for substantially accelerating overall AI progress or automating a full research team. That isn't the same as zero help, and it tells us little about other labs' systems in other workflows.¹¹

03Advantages can compound, but not automatically

Compute is one of the few inputs we can observe from outside. Here's Epoch AI's estimate of the capacity each major developer had at the end of 2025, covering research, training and serving customers combined.¹²

The biggest fleets are in the same rough range, and the uncertainty ranges overlap considerably. Compute alone doesn't crown a leader.

What can compound is compute plus a strong research system. A better internal model can help build better training environments, inspect experiments, speed up code and fix the infrastructure needed to run more research. Those improvements then help train the next model.

That doesn't mean leads must widen. Competitors read the same papers, improve their own tools and use AI too. Steady improvement is already exponential. Accelerating improvement is a stronger claim that needs stronger evidence.

Some labs also tell us more than others. Xiaomi livestreamed the reinforcement-learning run behind MiMo-V2.6. Mistral disclosed that Mistral Large 4 was trained on 3,800 Grace Blackwell GPUs and released it in public preview.¹³¹⁴ Disclosures like these let us reason about a lab directly instead of guessing.

My inference: OpenAI and Anthropic have unusually strong evidence of an internal research advantage. Google remains a serious challenger. Everyone else should be judged on their own record, assumed neither to have an enormous hidden lead nor to have nothing hidden at all.

A leaderboard compares products. A strategic assessment has to compare the systems producing those products.

04So where are we on the AI 2027 timeline?

AI 2027, published in April 2025, is a detailed scenario of how a fast AI takeoff could unfold.¹⁵ One thread runs through all of it: how much AI speeds up AI research itself. In the scenario, that rises from about 1.1× in 2025 to 3× in January 2027 and 10× by mid-2027.

That makes a natural yardstick. Find where today's evidence sits on the scenario's curve, then read off the date.

The companion analysis does this properly.¹⁶ It combines three measures: research speed-up (60% of the weight), the ability to complete long tasks autonomously (30%) and usable compute (10%). It then runs 200,000 simulations with explicit uncertainty on every assumption. Its answer:

In plain English: near schedule, with uncertainty far larger than the difference. Under this model's assumptions, anything from "about seven and a half months behind" to "about five months ahead" is consistent with the evidence. Corrected 2026-10-25

Why you should hold this loosely

The headline leans on assumptions more than measurements, and it's worth being explicit about which ones.

• Freeze the 1.5× estimate instead of assuming AI's research boost has kept compounding since it was measured, and the overall estimate falls to about 100 days behind. Getting to "on schedule" requires believing the loop has kept speeding up, and that unseen internal systems are ahead of what was measured.
• The choice of yardstick flips the sign. The graph AI 2027 originally published contained an error. Measuring against the corrected version of the authors' own forecast moves us to about 20 days ahead.
• The scenario's authors are more cautious. Their own grading put progress on quantitative measures at about 65% of AI 2027's pace in February, revised to about 75% with data through July.¹⁷ That predates the September and October evidence, but it's a useful reality check.
• Ask a stricter question and the answer changes. If you ask whether all the ingredients are ready, rather than what the weighted average says, the answer is about 134 days behind.

The single most valuable missing number is a fresh, independently measured, whole-lab research speed-up. If it were reliably 3×, the research-speed-up clock alone would read about 85 days ahead. At 5×, it would read about five months ahead. Try both in the chart above. These are readings of that one clock, not new values for the combined headline estimate. Corrected 2026-10-25

Being on schedule is not the same as being on the safe branch.

Matching the technical trajectory tells us nothing about whether AI 2027's political events will happen, whether a particular loss-of-control incident will occur, or which of its two endings we're heading toward. The scenario itself branches.¹⁵

05Is anyone actually slowing down?

Three different claims tend to get blurred together:

1. a particular model hasn't been released;
2. some research activity has been restricted;
3. the overall pace of progress has meaningfully slowed.

The first doesn't prove the second, and the second doesn't tell you how large the third is.

Here's what we know. After the July incident, OpenAI paused reinforcement learning on its latest deployment-bound models for two weeks. Later restrictions cut GPU allocation to its Astra-class models, but more compute went to other model classes, offsetting about 85% of that GPU-allocation cut across the reinforcement-learning workloads analyzed.⁸

Anthropic's pacing proposal goes further than release decisions. It calls for embedded external evaluators, then industry coordination, then international coordination.⁶ The later stages need others to sign on, and a proposal isn't a verified regime.

The fairest summary: parts of the actual development frontier have already been slowed or restricted. Continued product launches don't disprove that. What's unresolved is whether the restraints are broad, durable and effective enough.

06Reasons for hope, reasons for worry

A "good timeline" should mean more than one company or country getting the strongest system first. It should mean people keep meaningful control over their lives, catastrophic misuse and loss of control are avoided, and the benefits spread widely without locking in unaccountable power.

An internal research lead cuts both ways. It could buy time and capability for defensive preparation and alignment research. It could also make mistakes more consequential, widen the information gap between labs and everyone else, and concentrate decisions in a few institutions. The lead itself doesn't decide which effect wins.

07What to watch next

My central expectation is more of the same, faster. Expect substantial AI-assisted research, occasional disclosures that reveal more than the latest product does, and bottlenecks shifting toward experiment quality, infrastructure and oversight. A sharper acceleration is plausible if models get much better at choosing research directions. Stubborn bottlenecks could instead keep whole-lab gains far smaller than task-level ones. Both belong in the forecast.

For any specific risk, the real question is whether an effective safeguard is in place before exposure becomes unacceptable. It's whether that stays true as capability and deployment scale up. That's a different quantity from "days ahead of AI 2027."

—So, are we in the good timeline?

We're in a timeline where AI-assisted self-improvement is real, where the public can't inspect the strongest systems, and where some people and institutions are taking real steps to stay in control. We don't yet have enough evidence that those steps are keeping pace.

That's neither defeat nor reassurance. It's a reason to stop arguing about the latest chatbot release and start watching the process that's building its successors. There is a real attempt to steer. The open question is whether it's strong enough, early enough and accountable enough.