ExpertsFMEpisode 10 - How AI Is Accelerating Scientific Discovery Today and What's Ahead
OpenAI Podcast · November 20, 2025
5 Selected Lessons:
reasoning_pattern · 3 case · 1 framework · 1
AI-accelerated drug discovery and regulatory pipeline, step by step
Source
OpenAI Podcast — Episode 10 - How AI Is Accelerating Scientific Discovery Today and What's Ahead (November 20, 2025) · at 29:53 ▸
When this applies
when trying to accelerate a multi-stage pipeline that includes a large combinatorial search phase and a slow bureaucratic/validation phase
The takeaway
To accelerate a long multi-stage pipeline, identify and apply AI acceleration separately at each distinct bottleneck stage (search/filtering, physical validation, and documentation/approval), not just the most visible one.
Why it works
AI reduces the huge candidate search space by pruning unlikely options before costly physical experiments, and separately speeds the downstream paperwork/synthesis-of-findings stage required for regulatory approval, compounding acceleration across the whole pipeline rather than at just one point.
One level up
End-to-end acceleration of a pipeline requires attacking every sequential bottleneck, not optimizing a single stage in isolation.
Where it stops holding
This stops working when physical experiments remain the dominant bottleneck regardless of how well the search space is pruned, or when regulatory bodies cannot keep pace with AI-accelerated submissions.
Detail
procedure:
- Use AI to intelligently prune a huge combinatorial search space of candidate drugs/molecules.
- Run remaining physical real-world experiments on the narrowed candidate set.
- Use AI to help write and assemble the large regulatory submission documents that synthesize findings.
- Apply AI support at each step from candidate generation through to real-world deployment.
benchmarks: {}
rationale: Even though physical experiments at the end are unavoidable, pruning the search space earlier converges on likely winners faster, and AI assistance with regulatory documentation shortens the final bottleneck to real-world impact.
In their words
The more that models can learn how to prune that search space, even if you're going to end up with a bunch of physical real-world experiments to run at the end of the day, if you can intelligently prune the search space, then you can more rapidly converge on the drugs that are likely to work in particular scenarios.
Persist across model generations before declaring a tool incapable
Source
OpenAI Podcast — Episode 10 - How AI Is Accelerating Scientific Discovery Today and What's Ahead (November 20, 2025) · at 36:43 ▸
When this applies
when someone forms a lasting judgment about a fast-improving tool based on an old or limited trial
The takeaway
Periodically re-test a rapidly improving tool rather than freezing your judgment of it at the time of first failure, and account for which tier/version you tested.
Why it works
Because the underlying models improve rapidly (new versions, longer thinking time, paid tiers), a negative result from months ago, or from a free/limited version, no longer reflects current capability.
One level up
In fast-moving technology, beliefs about capability decay quickly and must be actively refreshed, not assumed stable.
Where it stops holding
This stops working when the tool genuinely has plateaued or when retesting costs more than the expected value of a changed outcome.
Detail
approach: When someone says a tool 'doesn't work' for their problem, check how long ago they tried it and whether they used the free/limited tier, then suggest retrying periodically with the best available version.
signals:
- Colleague says 'I tried it' with no time reference
- Colleague used free tier with limited thinking time
- Several months have passed since last trial
rationale: People form durable impressions from one-time trials even though the tool itself is evolving extremely fast, leading to outdated judgments.
In their words
I've encountered people who I really respect and the scientists are like, oh, I tried it, and I'm like, I tried it 18 months ago. They're not used to a tool evolving that quickly.
Raise benchmark difficulty as models saturate existing tests
Source
OpenAI Podcast — Episode 10 - How AI Is Accelerating Scientific Discovery Today and What's Ahead (November 20, 2025) · at 39:00 ▸
When this applies
when any measurement/evaluation tool becomes too easy for the thing it's measuring and stops differentiating performance
The takeaway
When a performance metric approaches its ceiling, replace it with a harder one rather than concluding progress has stopped.
Why it works
As the measured entity's capability rises, a fixed-difficulty test converges toward ceiling performance and loses discriminative power, so evaluators must continuously raise difficulty to keep the measurement informative and to surface remaining capability gaps.
One level up
Measurement systems must evolve at the same pace as the capability they measure, or they will falsely signal a plateau.
Where it stops holding
This stops working when no harder, well-defined test exists yet to replace the saturated one, leaving a measurement gap.
Detail
approach: Track benchmark saturation (model scores approaching/exceeding human baseline) as the trigger to design new, harder evaluations (e.g., frontier science/math questions, GDPVal for economically valuable tasks).
signals:
- Model score approaching or exceeding human baseline on an existing benchmark
- Public perception that models 'aren't impressive anymore' despite continued improvement
rationale: Every remaining gap where the model fails is valuable feedback for further training; harder benchmarks keep surfacing useful signal as capability grows.
In their words
I think it came out in 2023, and GPT-4 originally was at 39 percent on this benchmark. Humans, by the way, are at about 70 percent, but now you fast forward two years and our latest models are nearly at 90 percent.
Watch for the gap between frontier model capability and actual community usage
Source
OpenAI Podcast — Episode 10 - How AI Is Accelerating Scientific Discovery Today and What's Ahead (November 20, 2025) · at 33:46 ▸
When this applies
when a powerful new capability exists but a target user community hasn't caught up to using it effectively
The takeaway
Treat the gap between a technology's actual capability and its users' perceived/realized capability as a primary lever for value creation — closing that gap (via education, awareness, access) can be as impactful as improving the technology itself.
Why it works
Because models improve faster than most users' mental model of them updates, a persistent and growing gap opens between what the tool can actually do and what the community believes/uses it for, representing untapped value.
One level up
In fast-moving technology adoption, 'awareness and usage lag' is often a bigger bottleneck to realized value than raw capability improvement.
Where it stops holding
This stops working when the community has already fully adopted and is using the tool at its frontier, in which case the gap-closing strategy adds little marginal value.
Detail
approach: Identify where community usage of a tool lags its demonstrated capability (e.g. via internal stronger models not yet public, or examples like Alex's), then invest in outreach/education/access to close that gap rather than only pushing raw capability.
signals:
- Internal/advanced models already demonstrably stronger than what's publicly used
- Experts' mental models of the tool are outdated (see 'persistence across model generations' lesson)
rationale: Kevin states there's 'a big gap between what the models can do and what the science community uses them for,' and bridging it is explicitly one of OpenAI for Science's goals.
In their words
I think there's a big gap between what the models can do and what the science community uses them for. One of our goals here at OpenAI for Science is to start bridging that gap
Quantum gravity symmetry problem: warm-up priming unlocks a hard result
Source
OpenAI Podcast — Episode 10 - How AI Is Accelerating Scientific Discovery Today and What's Ahead (November 20, 2025) · at 14:29 ▸ · at 18:12 ▸ · at 21:03 ▸
When this applies
when a highly capable tool fails a hard problem on first try but might succeed if first given an easier, related warm-up version
The takeaway
If a capable system fails a hard problem outright, don't give up — give it an easier related sub-problem first to 'warm up' its reasoning context, then retry the hard version.
Why it works
Solving an easier, structurally related version of the problem first let the model build the relevant conceptual scaffolding (e.g., recognizing conformal symmetry) in context, which it could then reuse and extend to crack the harder, previously-failed version.
One level up
Difficulty at the frontier is often not a hard capability ceiling but a matter of priming and scaffolding; the right sequencing of problems can unlock what direct attack cannot.
Where it stops holding
This stops working when there's no simpler related sub-problem to warm up on, or when the harder problem doesn't actually share structure with the easier one.
Detail
situation: Alex asked GPT-5 Pro directly for the symmetries of a hard black-hole spacetime equation from his own cutting-edge research; the model said 'no symmetries,' which was wrong.
decision: At Mark Chen's suggestion, Alex instead gave the model an easier flat-space warm-up version of the same type of problem, then re-asked the original hard question in the same context.
outcome: The model solved the warm-up correctly (found conformal symmetry), and then, primed by that success, fully and correctly solved the original hard problem that it had failed minutes before.
lesson: Sequencing an easier related problem before a hard one can convert a model failure into a success on a problem at the edge of expert human ability.
In their words
And then Mark said, okay, well, but now that it's been primed on the warm-up example, try again in this instance of chat the harder problem.