After a (weak) AGI is created, how many months will it be before the first superintelligent AI is created?

Current estimate
37.4 months
(9.04 - 160)

Forecast Timeline

Comments

Comment by AussieA

I’m leaning toward a "fast takeoff" here. The main signal for me is the MiniMax M2.7 report from last month, where they showed a 30% jump in internal benchmarks purely through recursive self-optimization without a full retraining cycle.
MiniMax M2.7: The Self-Evolving AI Model Explained

If a "Weak AGI" is defined by its ability to do the work of a software engineer (hitting ~58% on SWE-bench Pro like GPT-5.4 just did), then the first thing it’s going to be tasked with is optimizing its own inference code and architecture.
Once you have a "digital worker" that doesn't sleep and can run 10,000 iterations of a new transformer block in a weekend, the gap to ASI collapses.
The only reason I’m not at "weeks" or "days" is the Physical Infrastructure Bottleneck. Even if the software gets 100x smarter in a month, it still needs to procure chips, build data centers, and secure gigawatts of power to actually manifest that "superintelligence" as a global force. That "hardware lag" is a sticky 12-month floor in my model.

Comment by markwokho418

People here are out of their mind. There is no way we only need 35 months to build an "AI that can do anything better than anyone" after achieving weak AGI.

Comment by casens

AI 2027 by kokotajlo et al., estimate a scenario where this question resolves as about 10 - 18 months.

Comment by Auspicious

In other words, are you saying that we will leapfrog over the "weak AGI" classification, and go straight from specialized systems to full AGI?

Comment by Cato

Two questions about subagents in the resolution. First, is just the design enough? Or does the subagent actually have to be built and the design validated by successful trial?

Second: Whether subagent performance has to be demonstrated or not, the bar for SAI is very high, much higher than the community seems to be imagining.

Comment by ADQ

ADQ's Prediction: 177 months (122 - 233)

Comment by Azgerod

So let me get this straight: we think weak AGI will come in December 2025, and full AGI will come in 2030. And simultaneously we think that ASI will come less than two years after weak AGI, so before December 2027.

Comment by AI_Understander

AI_Understander's Prediction: 5.6 months (2.09 - 11.8). Y'all underestimate AGI hard. Even in a weak form it will be amazing.

Comment by Sourian

@AI_Understander The "weak AGI" resolution criteria do not necessitate self-improvement.

Comment by edwardhammond

@AI_Understander Why does being able to pass a Turing test and build a model car predict the ability to rapidly build ASI?

Comment by Primer

Primer's Prediction: 126 months (69.1 - 271). Considering weak AGI might resolve somewhere around right now (probably just a question of whether someone in the right place puts in a few weeks effort) and superintelligent AI might mean the end of the world, the community seems pretty pessimistic compared to other AI questions.

Comment by kokotajlod

I highly encourage anyone interested in AI timelines to check out the interactive model and the accompanying report. Pay special attention to the "AGI training requirements" variable.

Additional Notes

It is essential to define what constitutes a "weak AGI" and ensure clarity on its criteria for evaluation.