Memos

Beyond Static Intelligence

Memo 01September 2026

Intelligence that stays effective through change

The real test is not completing a task once, but remaining effective as the world changes.

A system is not intelligent merely because it can reproduce an answer under a fixed distribution. In a dynamic world, intelligence means forming valid actions under uncertainty—and allowing the consequences of those actions to continually correct what the system believes and does next.

Intelligence is the ability to remain effective as the world changes.

No model can fully internalize an environment whose people, objects, constraints, and causal structure keep changing. A model may provide powerful priors, but a lifelong agent cannot operate from those priors alone. It must stay coupled to the world through a closed loop of grounding, execution, verification, and revision.

01 Grounding: locate intelligence in the present world

Grounding is more than recognizing what is visible. The system must estimate the current state of the world, identify what actions are actually afforded, and distinguish correlation from causal influence. It needs to know not only that an object exists, but whether it can be moved now, what else that movement will affect, and which assumptions remain uncertain. Grounding turns abstract capability into a belief about this particular situation—a belief that can be tested by action.

02 Experience: convert interaction into evidence

Long-horizon tasks make every action an intervention. Execution changes the state; verification compares the observed result with the intended effect; feedback exposes what the agent misunderstood. A useful lifelong agent does not discard this history after an episode. It turns outcomes into updated beliefs, memories, and reusable skills, so that a failed attempt can become supervision and a successful recovery can become a future strategy.

This is why closed-loop improvement matters more than isolated success. The agent should become better within the task as new evidence arrives, preserving continuity across many steps instead of repeatedly solving each moment from scratch.

03 Adaptation: make experience endure

Experience becomes valuable only when it continues to shape action. Adaptation must carry what was learned across tasks, across time, and across individuals without freezing it into a brittle rule. The goal is durable change: retaining causal knowledge and useful skills while remaining ready to revise them when the environment or the person changes.

A lifelong agent therefore improves through a continuing loop: ground itself in the present world, act and verify over long horizons, consolidate the result into experience, and adapt that experience to the next situation. Intelligence is not a static property stored inside a model. It is a relationship with a changing world that must be renewed through interaction.

Memo 02September 2026

Generation by Needs

Don’t generate when you can judge, estimate, or verify.

Embodied agents increasingly rely on LLMs and VLMs for components that do not inherently require open-ended generation: state recognition, action selection, success detection, safety checking, routing, ranking, and memory retrieval. Generation by Needs asks which of these components can be reformulated as efficient judgment, estimation, and verification problems.

Use generation only when the decision space cannot be represented efficiently.

The aim is not merely to make a generative model smaller. It is to determine which parts of intelligence fundamentally require generation and which can be reduced to fast, stable decision-making. Lightweight models should form the default control path; expensive generation should be invoked only when uncertainty or an out-of-distribution state makes it necessary.

01 From generation to decision

Many apparently generative modules have a constrained output space. Once that space is made explicit, the problem often becomes selection, prediction, scoring, or verification rather than decoding.

State understanding: classify or rank the current scene state and its transition.
Value estimation: replace reasoning-heavy scoring with a compact value or reward model.
Outcome verification: predict whether an action succeeded with a lightweight verifier.
Action selection: choose the next primitive from an available skill library instead of generating it.
Transition prediction: infer changes such as Door: Closed → Open directly from observations.
Memory selection: retrieve and score relevant experience instead of reasoning over the full history.

02 Why judgment, estimation, and verification?

A discriminative module can make one decision in a single forward pass. In a high-frequency control loop, lower latency is not only an efficiency gain: it allows the agent to observe, verify, and recover more often. Smaller decision models can also offer higher throughput, lower inference cost, more stable outputs, and easier deployment on edge devices and real robots.

The tradeoff is flexibility. This formulation requires a reasonably defined decision or action space, task-specific positive and negative supervision, and reliable detection of unseen states. Classification boundaries may fail under distribution shift, so open-ended planning and novel instruction following should remain available when the fast path is uncertain.

03 Embodied demonstrations

With privileged state and transition information, decision models can act quickly in simulation and generalize across tasks without repeatedly invoking open-ended generation. The unresolved challenge is the real world: useful information is partial, noisy, delayed, and rarely given in a form that a verifier or classifier can directly consume. How an embodied system can acquire reliable decision evidence from interaction remains an open problem.

One direction is to build hierarchical decision architectures that decompose a difficult judgment into progressively grounded decisions, improving accuracy without giving up speed. Parallel evaluation and compressed state representations may further expand the capability boundary of decision models by considering more evidence within the same latency and compute budget.

JEV for robotics

How much generation does an embodied agent actually need?

Generation by Needs treats generation as a scarce capability rather than a default interface. The resulting design principle is simple: decide, don’t decode—until the situation truly demands it.