Sam Altman’s 2025 AI Coding Forecast and Three Observations

In February 2025, I argued that AI coding agents could soon outperform individual human coders and be copied at enormous scale. I also expect AI to advance scientific discovery and change society, although those outcomes and their timing remain forecasts, not demonstrated results.

Sam Altman spoke in Japan about rapid model progress, reasoning and lower costs. He said successive GPT generations used roughly 100 times as much pretraining compute and acquired new capabilities. That is a claim about training resources, not a measurement that intelligence increased 100-fold.

Altman said reinforcement learning had improved performance on some benchmarks enough that smaller reasoning models could achieve results previously expected from much larger models. He also emphasized that the new approach does not improve every capability equally.

At inference time, a reasoning model can spend longer working on a prompt. Training with reinforcement learning and allocating more computation at inference are distinct techniques, although they can complement one another in models such as the reasoning systems discussed here.

I compare the pace of AI progress with Moore's law, the historical observation about transistor density. Improvements in AI capability or service prices are not directly comparable to a doubling of transistor density: they measure different things.

Verifiable rewards help train models on tasks where a proposed solution can be checked, such as some mathematics and programming problems. A verifier can use an answer, tests or other criteria; it need not know every solution in advance. Subjective tasks such as poetry do not have the same straightforward check.

Altman expects that combining larger pretraining runs with these methods could bring the first signs of genuinely new scientific knowledge. This is an expectation about future models, not evidence that such a breakthrough has already occurred.

I return to Leopold Aschenbrenner's illustrative intelligence-explosion scenario: if AI research produces systems that improve AI research in turn, a feedback loop might accelerate progress. The chart is a scenario, and both the loop and its speed are uncertain.

I think coding could be an early area of exceptional AI performance. I draw an analogy with AlphaGo, whose self-play uncovered strategies such as its famous move 37; the analogy is suggestive, but software engineering involves more than a closed game with an agreed win condition.

Imagine copying a capable coding agent thousands or millions of times and having those agents work continuously. That scenario depends on actual model capability and on available compute, energy and practical deployment, rather than copying being free or unlimited.

Altman says the models he is discussing can program well when people have already done the relevant work, but are less capable of inventing new algorithms or science. He predicts that further increases in scale could change that; I regard the prospect as close, though the timing is speculative.

Altman describes a progression in competitive-programming estimates: an early reasoning model around the top million, another around the top 10,000, o3 around 175th and an internal result around 50th. He suggests the top rank might be reached by the end of 2025.

Those numbers concern competitive programming, not a direct ranking of every professional developer. I see coding, like chess and Go, as a domain where solutions can often be verified, but benchmark scores alone do not establish that a model is the world's best software engineer.

Altman hopes that by the end of 2025 a Pro-tier user could give a model a difficult question and let it work for hours, using tools if necessary. He excludes problems requiring new scientific discovery; this is an aspiration about future capability, not a promise that AI can do virtually anything better than people.

I turn to Altman's essay 'Three Observations,' which says systems pointing toward AGI are coming into view. Its footnote explicitly says the essay's use of AGI is not intended to interpret or change the definitions governing OpenAI's relationship with Microsoft.

Altman presents AGI both as another tool in humanity's accumulation of technology and as a potentially different kind of change. I speculate that it might be the last major tool humans need to invent, while acknowledging that the consequences could be positive or negative.

The essay imagines curing disease, giving families more time and expanding creativity. I contrast that optimism with Altman's earlier warning against expecting an immediate AGI deployment. These are aspirations and predictions, not evidence of imminent cures or a settled deployment date.

The first observation links model capability roughly to the logarithm of resources used in training and inference, including compute and data. Altman describes predictable gains across a range of scales; extrapolating those trends indefinitely, or to a particular discovery, remains uncertain.

The second observation is that the cost of using a given level of AI has fallen dramatically, with a claimed roughly tenfold annual reduction. The essay compares GPT-4 and GPT-4o token pricing; prices and model equivalence depend on the task, provider and input or output rates, and are not a universal physical law.

The third observation proposes that the socioeconomic value of increasing intelligence grows super-exponentially. I interpret this as a claim that value may increase faster than capability itself, but the size and distribution of that value are forecasts rather than established measurements.

Altman asks readers to imagine an engineering agent able to do many tasks that take a skilled engineer up to a couple of days, then to imagine thousands or a million copies. I extend the thought experiment to other knowledge work; it should not be mistaken for a demonstrated workforce at that scale.

The essay compares AI with the transistor: an invention that spread through the economy until its effects became ordinary. I recommend Chip War for context on the history and geopolitics of semiconductors, while expecting AI's economic effects to become hard to ignore.

I expect disruption and friction as jobs change, but I am optimistic that people will find new activities. Mechanized agriculture reduced the share of workers needed on farms and helped make other occupations possible; that historical transition does not guarantee an easy or equally shared AI transition.

Faster scientific progress and lower prices for many goods are among the benefits I hope for. Altman also notes that scarce assets such as land and luxury goods might become more expensive, so improvements would not necessarily be distributed evenly.

Altman warns that the balance of power between labor and capital could go wrong and may require early intervention. I worry that if owners can buy highly capable automated labor without the bargaining constraints of human workers, workers' negotiating power could weaken. Much remains unresolved even as I remain optimistic about the possibilities.