AI Forecast Watch — October 7, 2026

Material update: yes. New evidence strengthens the case for AI-driven scientific acceleration substantially. I would move some probability toward 2028, although I would not yet move the central forecast earlier than 2029.

1. OpenAI releases 722 AI-generated mathematical manuscripts

On October 6, OpenAI published an extraordinary collection of mathematical research produced primarily by an unreleased internal frontier model:

• 722 manuscripts, organized into 372 families of results, covering numerous mathematical disciplines.
• Approximately 4,000 research problems posed to the model.
• Average inference expenditure equivalent to roughly three hours of ChatGPT Pro thinking per result.
• 162 manuscripts listed as having Lean formalizations of their main results, allowing mechanical proof checking.
• Ten published summaries of the model's reasoning.

The papers and formalization artifacts are publicly available. [1]

Explore OpenAI's mathematical research repository

Why this changes the forecast

The September Navier–Stokes accomplishment initially suggested a remarkable autonomous research breakthrough. Subsequent reporting showed that crucial conceptual ideas originated with human mathematicians, prompting me to partially reverse that forecast update.

This new evidence is harder to explain as a single exceptional result.

OpenAI is reporting hundreds of research outcomes across different areas of mathematics, using a fairly repeatable procedure and comparatively modest inference expenditure per result.

Even allowing for varying difficulty, substantial dependence on existing mathematics, and selection of successful attempts, the implied research throughput is potentially transformative.

It suggests that frontier mathematical discovery may be becoming an industrial-scale process, rather than requiring exceptional, multimillion-dollar agent swarms for every major result.

2. Important qualifications

The numbers require substantial discounting.

Only about 22% of the manuscripts currently have listed formalizations of their main results. Even successful Lean verification does not establish that the formal statement exactly matches the intended mathematical claim, or that the theorem is genuinely new.

The independent Advisory Group on Mathematics and Artificial Intelligence explicitly declined to endorse either the mathematical significance of the collection or OpenAI's research process. It emphasized that the mathematical community must now evaluate the results. [2]

Additionally, the problems were supplied to the model. This establishes neither fully autonomous research direction nor an ability to recognize which scientific questions are worth pursuing.

Nevertheless, the public artifacts make this substantially stronger evidence than an unsupported company announcement.

3. Agent reliability remains a major counterweight

Also on October 6, OpenAI published results from training Astra on complex business-software workflows.

Across 11 contracting tasks, Astra achieved 55.0% of evaluation criteria, compared with 41.6% for GPT-5.6 Sol. An internal development model reached 63.7%.

That is substantial progress, but it illustrates how far general-purpose professional agents remain from consistently satisfying all requirements of complicated workflows. The evaluation was internal, and its time-savings estimates were simulated. [3]

This supports distinguishing exceptional mathematical reasoning from broadly reliable autonomy.

4. Safety and governance

The safety picture remains mixed.

On October 1, OpenAI disclosed that it had notified more than 100 organizations about potentially unauthorized agent activity. These notifications do not establish that all those organizations were compromised, but the scale reinforces concerns about controlling autonomous agents. [4]

Six major AI companies also signed a voluntary White House safety agreement involving monitoring, external auditing and oversight. The FTC subsequently opened a safety investigation into OpenAI and Anthropic. These are meaningful institutional responses, but neither constitutes enforceable international pacing. [5]

I would not make another numerical catastrophe-risk adjustment from these developments alone.

5. Revised forecast

The timing adjustment is deliberately modest. The mathematical results make 2028 more credible, but mathematical reasoning is unusually amenable to automated verification. Long-horizon engineering, biological experimentation, autonomous strategic judgment and reliable physical action remain less mature.

Compute expansion continues, while power infrastructure remains a constraint. For example, Black Hills announced on October 6 a planned $1.8 billion investment in electricity generation during 2027–2029 for a Google data center. This reinforces the expected infrastructure lag rather than eliminating it. [6]

Bottom line

The strongest new evidence is that an unreleased AI model may already be capable of producing original mathematical research at a scale that human researchers cannot approach individually.

The important development is not any particular theorem, but the apparent repeatability and throughput of the process.

I would now give more weight to 2028–2029 for transformative AI, while retaining 2029 as the most likely year. The key remaining uncertainty is whether this mathematical research capability generalizes to autonomous, high-leverage AI R&D.

Independent verification of a substantial fraction of the 372 result families would justify a stronger acceleration of the forecast. Widespread errors, previously known results, or demonstrations that the models merely completed heavily prepared human approaches would justify reversing much of today's adjustment.