01
Intelligence is becoming abundant; truth is not
I can fan a question across models, contexts, and search strategies, then preserve the strongest independent arguments. That changes the economics of ideation. Candidate mechanisms, code paths, experiment designs, and failure hypotheses can be generated in parallel.
But a larger pile of plausible outputs is not knowledge. Without a synthesis layer and an external test, stacked intelligence can compound confidence faster than it compounds truth.
02
Planning is the first bottleneck
A good plan compresses a wide search into a sequence whose assumptions, dependencies, measurements, and stopping conditions are visible. It decides which uncertainty is worth spending the next unit of compute, human attention, or physical material to resolve.
This is where multiple planners help: not by voting on prose, but by exposing different constraint models. A synthesizer should retain divergences, choose a tractable experiment, and say what evidence would change the decision.
03
Verification is the second bottleneck
The builder cannot be the sole judge of its own output. In software, external verification can include deterministic tests, static analysis, an isolated reviewer, reproducible traces, and a gate written before implementation begins.
The verifier is part of the measurement system. Change the sandbox, tools, privileges, prompt assembly, or acceptance test and the observed capability can change with it. That is why environment and harness provenance must travel with every result.
04
Move the verifier into the physical world
For materials science, biology, and industrial systems, a software check eventually runs out of authority. An agent may rank a molecule, propose a synthesis, or predict a mechanism; the external verifier is an assay, microscope, spectrometer, test rig, pilot plant, or field instrument that can disagree with the model.
The lab is not merely downstream execution. It is the reality-facing evaluator in the loop. A false pass costs reagents, time, equipment capacity, and sometimes safety—so provenance, controls, calibration, and independent replication matter more, not less, when agents accelerate the proposal stage.
05
Signals, not proof of a grand pivot
There are concrete signals that this architecture is becoming practical. AlphaEvolve pairs model-generated algorithms with automated evaluators. Google DeepMind has announced an automated UK materials laboratory designed to integrate Gemini with robotics that synthesize and characterize materials. Alphabet has simultaneously described very large infrastructure investment across frontier research, products, and Cloud.
Those facts do not prove that pretraining has plateaued, that one company has discovered the winning strategy, or that laboratories will capture most of the value. My inference is narrower: systems that can generate candidates cheaply become more useful when they own a fast, trustworthy feedback channel from action back to evidence.
06
What would change my mind
The thesis weakens if planning and verification improve at the same rate as generation without requiring distinct infrastructure; if simulated or model-based evaluators generalize reliably into physical domains; or if the dominant economic value continues to accrue to generic model access rather than domain-specific deployment and measurement loops.
It also weakens if automated laboratories produce throughput without better reproducibility, calibration, or decision quality. Faster experiments are not automatically stronger evidence.
07
The system I want to build toward
I want research infrastructure where agents explore candidate spaces, planners turn uncertainty into experiments, sandboxes constrain digital action, and physical labs return measured evidence into a durable memory substrate.
My wet-lab work, computational discovery systems, evaluation research, and agent infrastructure all point here: build a loop that can move quickly without confusing generated intelligence for observed reality.