Agents that do real science: faster disc ...

Agents that do real science: faster discovery from the lab to the cloud

Oct 29, 2025
155

Agent talk is cheap until it produces results in hard domains. Scientific discovery is a good test, and 2025 brought credible steps. MIT researchers introduced SCIGEN, a tool that lets scientists encode design rules into generative models so outputs respect physics and domain constraints. Instead of hoping the model coughs up useful candidates, SCIGEN bakes in what “valid” means. The payoff is a higher hit rate for materials with exotic properties relevant to quantum computing and energy.

A companion thread is orchestration across many data types and steps. MIT’s CRESt platform learns from heterogeneous scientific information, then plans and runs experiments to discover new materials. This is agentic in the practical sense. It retrieves, reasons, chooses, and acts. The value is not a single clever output, but a closed loop that moves a hypothesis through execution toward evidence.

The same mindset is reshaping infrastructure. Google Research described LAVA, a scheduling algorithm that continually re-predicts virtual machine lifetimes and adapts placements in large cloud data centers. It is an agentic controller for resources. The result is better efficiency and lower cost, which matters when model training and serving chew through compute and energy budgets. Science and cloud ops are different arenas, but the pattern is shared: plan, act, measure, adapt.

If you run R&D, do not start with a monolithic “AI scientist.” Start by codifying constraints, automating the planning around them, and instrumenting every decision so it can be audited. If you run platforms, apply the same loop to compute. Let agents plan and re-plan the boring but expensive stuff so your people focus on ideas.

The headline is simple. Agentic AI is not just writing emails a bit faster. It is starting to speed up discovery and make the compute that powers discovery less wasteful. That is a real win.

Sources
https://news.mit.edu/2025/new-tool-makes-generative-ai-models-likely-create-breakthrough-materials-0922
https://news.mit.edu/2025/ai-system-learns-many-types-scientific-information-and-runs-experiments-discovering-new-materials-0925
https://research.google/blog/solving-virtual-machine-puzzles-how-ai-is-optimizing-cloud-computing/

Ti piace questo post?

Offri un caffè a Don Teague

Altro da Don Teague

PrivacyTermini