Medium-Term Directions

AI for scientific discovery

Why it is credible. AI has already contributed to real scientific results, protein structure prediction is the standout demonstrated case, and models are being applied to hypothesis generation, materials search, and experiment design. The pattern of "AI narrows a vast search space for human scientists" is proven in at least one domain.

What stands in the way. Verification and grounding. A model can propose a molecule or a proof sketch, but nature and mathematics are the arbiters, not the model's confidence. Discovery requires closing the loop with real experiments or formal checking, which is slow and expensive. Generalizing from the domains where this works (well-structured, verifiable) to open-ended science is the open challenge.

World models

Why it is credible. Systems that learn a predictive model of an environment, "if I take this action, this is what happens next", enable planning and data-efficient learning, and have shown strong results in games and simulated control. A model that can imagine consequences is a natural route to agents that plan.

What stands in the way. Fidelity and generalization. Learned world models drift from reality over long rollouts, and transferring from simulation to the messy physical world is a long-standing, unsolved gap. Whether rich, reliable world models scale to open, real-world environments is genuinely uncertain.

Continual and lifelong learning

Why it is credible. The value is obvious, models that keep learning after deployment, adapting to new information without a full retrain, and there is active, measurable progress on the core obstacle.

What stands in the way. Catastrophic forgetting. When a neural network learns new tasks, it tends to overwrite what it knew before. Techniques exist to mitigate this, but a system that learns continuously, indefinitely, without degrading is not yet a solved capability. Until it is, "learning" in production usually means periodic retraining, not true online adaptation.

Much cheaper inference

Why it is credible. This is the most solid of the five, it is really the KNOWN efficiency trend, projected forward. Quantization, distillation, better hardware, and improved serving keep lowering cost per unit of capability.

What stands in the way. Not feasibility but magnitude and timing. That inference gets cheaper is reliable; how much cheaper, how fast, and whether it unlocks qualitatively new applications (pervasive on-device AI, for instance) is the uncertain part.

1. 1Find the mechanismIs there a concrete reason this could work, a technique, an early result, or just a wish?
2. 2Name the obstacleWhat specifically has to be solved? Compounding error, verification, forgetting, sim-to-real?
3. 3Look for the proof-of-conceptHas it worked in at least one narrow domain? Protein folding for discovery; games for world models.
4. 4Hold it as EMERGINGCredible mechanism plus real obstacle equals 'watch closely', not 'coming next year'.

Further reading

• Highly accurate protein structure prediction (AlphaFold), the clearest demonstrated case of AI advancing science.
• Overcoming catastrophic forgetting in neural networks, the core obstacle to continual learning, and one line of attack.
• World Models, learning a predictive model of an environment to plan and imagine.
• Stanford HAI AI Index Report, tracks which of these directions are producing measurable results versus talk.