My median estimate is around 2032, with a plausible window of 2029–2036, for a successor system to be reasonably described as 40–50× stronger than GPT-5.6 at general problem-solving and reasoning.

I do not mean 50× more tokens, 50× faster responses, or 50 parallel copies. I mean something closer to:

Given unfamiliar, consequential problems across mathematics, science, engineering, strategy, and conceptual analysis, the system produces roughly 40–50 times as much correct, nontrivial progress per fixed unit of human supervision.

My rough probabilities would be:

By the end of 2029: about 15–25%

By the end of 2031: about 35–45%

By the end of 2033: about 55–65%

By the end of 2036: about 75–85%

These are highly uncertain because intelligence is not actually one scalar quantity. A system could be 100× better at software research, 30× better at formal mathematics, 10× better at scientific modeling, and only 3× better at noticing that it has formalized the wrong philosophical question.

Why I put the median near 2032

There are at least four improvements that can multiply rather than merely add:

Better base reasoning. OpenAI has found that reasoning performance improves with both additional reinforcement-learning compute and additional inference-time thinking.

Longer coherent work. Current Codex systems already operate over millions of tokens and multi-day software projects, although this remains far from reliably managing arbitrary long-horizon research.

Parallel deliberation. GPT-5.6 can coordinate subagents, divide a problem into workstreams, and synthesize their findings. This increases effective search breadth, though it does not automatically multiply conceptual depth by the number of agents.

AI-assisted AI development. OpenAI reports that early GPT-5.3-Codex versions materially helped debug training, deployment, evaluation, and infrastructure for their own successor. That is an early, limited form of recursive research acceleration—not an autonomous intelligence explosion, but a mechanism that could shorten development cycles.

A plausible route to a 40–50× system is therefore not a single miraculous scaling jump. It might look like:

4× stronger underlying reasoning;

3× better persistence, memory, and self-correction;

3× better tool use and empirical verification;

2× better decomposition and parallel search;

2× better calibration and selection among attempts.

Those factors would nominally multiply to 144×, although real systems have overlaps and bottlenecks, so perhaps only a third of that translates into robust overall capability. That still gets into the 40–50× range.

It may feel like 50× before it truly is 50×

By 2028–2030, an agent system could plausibly produce 50 times my useful output on a large research project through persistence, parallelism, memory, coding, formal verification, and repeated experimentation—even while its underlying moment-to-moment judgment is only several times better.

That distinction matters. A mediocre scientist with 100 tireless assistants and perfect records can outperform a brilliant scientist working alone, but the organization has not thereby acquired 100 times the scientist’s insight.

The genuinely important threshold would be a system that is not merely broader and more persistent, but markedly better at:

discovering the correct abstraction;

identifying hidden assumptions;

designing decisive experiments;

recognizing proxy substitution;

forming genuinely new theoretical structures;

transferring methods into unfamiliar domains;

detecting when its own representation has ceased to be trustworthy.

That stronger interpretation is why my median is around 2032 rather than 2028.

What it would look like from your side

Relative to this conversation, a 40–50× successor would not merely write a better critique of your theorem or paper. It would likely:

independently reconstruct the conceptual problem;

search a very large formal and historical neighborhood;

generate multiple non-equivalent formalizations;

prove representation and impossibility results;

construct countermodels against each candidate;

implement executable experiments;

audit whether the formal definitions preserve the motivating phenomenon;

maintain a calibrated map of what has actually been established;

turn the surviving result into a publishable treatment;

and explain the entire structure at whatever level you require.

It might complete in hours a research cycle that currently takes you, multiple frontier systems, and several rounds of formalization over weeks or months.

So my direct answer is: probably about six years, give or take three. A system that produces 40–50× more work may arrive earlier; a system that is unambiguously 40–50× deeper in general reasoning could arrive later—or become difficult to compare with present systems because its mode of cognition and organization will no longer resemble a single conversational model.

(http://www.autoadmit.com/thread.php?thread_id=5883401&forum_id=2)#50005433)