Sam Altman: AI Is Entering the Era of Scientific Discovery — Math Breakthroughs Will Replicate in Biology and Energy

From Tool to Discoverer: AI's Role Is Shifting

In a recent interview, Sam Altman offered a key judgment: AI is entering an entirely new phase — helping humanity discover major new knowledge. He noted that while this trend has been discussed for a long time, it has only recently become concrete and tangible: "AI is doing very difficult things, figuring out very complex, important new knowledge."

The weight of this statement lies in how it repositions AI — from "a tool for boosting efficiency" to "an agent participating in knowledge production." In the past, we were used to AI writing copy, refactoring code, and producing summaries. What Altman describes is different: AI has already demonstrated the ability to solve highly difficult problems in mathematics, and that capability is now migrating to other foundational scientific fields.

You may not have noticed, but this shift didn't happen suddenly. Altman uses "maybe started earlier this year" to describe the starting point, suggesting that in his view, AI's discovery capability has gone through a period of accumulation and is now entering an acceleration phase.

Math Is the Starting Point; Biology and Materials Are Next

Altman laid out a clear migration path: the methods that work in mathematics will work equally well in biology, materials science, and energy.

Specifically, he mentioned that AI can help humanity develop new drugs, discover new materials, and find new energy sources. The underlying logic is that although these fields study different subjects, they all fundamentally involve high-dimensional search, pattern recognition, and hypothesis validation — precisely what today's AI excels at.

Observers note that mathematics became the first breakthrough point because it has clear validation criteria: a proof either holds or it doesn't. Feedback cycles in biology and materials are longer and noisier, but Altman believes the methodology itself can be reused.

If this judgment holds, AI's impact on science will no longer be at the auxiliary level of "accelerating literature review" but will directly intervene in hypothesis generation and validation. For industries like drug development — with long cycles and high costs — this is especially significant.

High-dimensional search and hypothesis validation: In the context of scientific discovery, this refers to finding solutions that satisfy constraints in problems with enormous parameter spaces — mathematical proof is searching through possible proof paths for a valid chain of reasoning; drug development is searching through vast molecular structures for candidates that meet target and safety requirements; materials discovery is searching through the combinatorial space of compositions and processes for specific properties. Traditional methods rely on expert intuition and trial-and-error, with costs spiraling out of control as dimensionality rises; AI's value lies in using learned priors to narrow the search space and in rapidly screening and iterating candidate hypotheses against automatically verifiable objectives (proof correctness, binding affinity, stability metrics).

The Most Exciting Direction, Yet the Least Perceptible

Altman candidly admits that scientific discovery has always been one of his "most exciting" directions. But he also offers a thought-provoking point: this impact won't make people directly feel like "I'm using AI."

His exact words: "It won't feel like directly using AI. It will just feel like — well, the world got better, society is functioning normally." This statement reveals an important pattern in how change spreads: truly foundational technological transformations often exist by "disappearing." Electricity and the internet both went through similar processes — once a technology is mature enough, it's no longer something that needs to be mentioned separately.

This also means that AI's value in science may not manifest as "blockbuster products" but will be distributed across cheaper drugs, more efficient energy, and more durable materials. For those focused on AI commercialization, this is a signal to adjust expectations: the biggest returns may not be in the AI application layer but in AI-driven physical industries.

Technology disappearing: This phrase is often used to describe how mature technologies recede into the background. Truly transformative technologies don't keep appearing as "new products" but become embedded in infrastructure and default conditions: people no longer ask "is this electricity?" — they only care about prices and reliability; they no longer ask "should I get online?" — they only care about whether services are available. Whether a technology is truly successful is often judged not by how much it's talked about but by whether it no longer needs to be talked about separately. For AI, this means its scientific value may be reflected more in industrial metrics like shortened drug development cycles and accelerated materials iteration than in visible products at the application layer.

Society Gets Better, But No One Will Applaud

Altman closes with "the world got better, society is doing what it should do." This phrasing seems plain but actually points to a deeper question: as AI-driven progress becomes routine, will the public and policymakers continue to pay attention to its risks and governance?

From a positive perspective, this is a sign of technological success — AI has truly integrated into humanity's knowledge production infrastructure. But from a governance perspective, when AI is no longer "conspicuous," related safety discussions, resource investment, and regulatory attention may also decline. Altman himself didn't elaborate on this implication in this segment, but the "imperceptible progress" he describes objectively places higher demands on science communication and policy-making.

Taken together, Altman's remarks, though brief, outline a clear trajectory: AI starts by solving mathematical problems, spreads to biology, materials, and energy, and ultimately changes society in ways that are hard to notice. Whether this trajectory materializes as expected will be a key benchmark for observing AI's scientific value in the coming years.

Imperceptible progress and governance attention: Public policy and risk regulation typically rely on visible focal events and sustained social discussion to allocate attention and resources. When technological benefits become routine and risks are distributed across many links, regulators face not a single object but foundational capabilities embedded in research, medicine, energy, and other systems. This creates two governance challenges: first, capability assessment requires cross-disciplinary expertise and is difficult to measure with unified standards; second, incentives to invest in safety research may be weaker than incentives to invest in capability improvements. Therefore, science communication needs to clearly articulate both the "imperceptible" benefits and potential risks, while the policy level needs to establish routine assessment and audit mechanisms rather than responding passively only when controversies erupt.