Latent Map Labs (LML)
We are charting the latent map between intelligence and reality — the high-dimensional manifold where cognition, action, and the physical world meet. Our work spans artificial intelligence, quantum computation, neuroscience, and physics, in search of the structures that connect them.
Generality is engineered, not benchmarked.
Most frontier gains today are abstraction stacked on abstraction. Agentic systems are made to generalize marginally better by being made substantially more complex — longer scaffolds, heavier orchestration, more parameters over more data. We move the other way: pulling the work back down from high-level abstraction to the concrete mechanics of how these systems learn, act, and fail, and rebuilding that machinery deliberately — end to end, against the only standard that counts, behavior in the real world.
This is one program, not four departments. The pipeline is a coupled system — its parts cannot be understood in isolation — so we keep the surface narrow and the stack deep, and work it from four angles.
Pretraining, post-training, alignment, and agentic embedding treated as one coupled system — chasing gains where they now actually live, in post-training and inference, and baking them into core components rather than bolting on orchestration.
Not leaderboard deltas or capability bought with more data, but contamination-controlled, harness-pinned evaluation that tracks real deployment — measurement built to show where capability, and where the bottlenecks, actually live across the stack.
How people build on open-weight models, made reproducible. The recipe, eval harness, and inference stack that decide a result live in separate tools and rarely travel together; we are building toward an open, provider-agnostic manifest that pins all of them into one shareable, reconstructable workspace.
Model and systems choices made with the silicon in view — quantization, kernels, memory, and serving treated as part of the design, so software, hardware, and ultimately robotics move in alignment rather than in isolation.
Underneath all four is rigorous, grounded fundamental research. We work from first principles — the mathematics of optimization and information, the statistics of learning and estimation, and the physics of statistical mechanics and dynamical systems — and we hold ourselves to deriving results, not merely observing them. That research earns its place only when it becomes real: models, tools, and infrastructure we build and release into the open-source community, where anyone can reproduce, test, and extend them.
Intelligence is not a brain in a vat.
Cognition becomes useful only through bidirectional coupling with the world. The brain is consequential not because it computes in isolation, but because it sits at the center of a dense lattice of sensorimotor pathways — afferent signals carrying the world in, efferent signals carrying action out. Every meaningful representation is shaped, constrained, and validated by that loop.
The current frontier of artificial intelligence has been built the other way around: vast competence cultivated inside the vat, with embodiment, interaction, and consequence treated as downstream applications. We believe this gets the order wrong. An intelligence that cannot be grounded in, corrected by, and held accountable to reality is not robust, not interpretable, and not ultimately useful.
LML is built around the inverse premise — that the coupling between model and world is the primary object of study, not a peripheral concern.
Ethics is the substrate, not the wrapper.
A system that cannot model the consequences of its actions cannot be aligned with anyone's values. Alignment is not a layer applied after capability is achieved; it is the same problem, viewed from the other side. A model that lacks a coherent grasp of reality cannot be safely steered toward, or away from, any outcome within it.
We therefore reject the view that capability and alignment are parallel tracks to be balanced. They are dual descriptions of the same underlying competence — the ability to represent what is, what could be, and what the difference costs. Our research is organized so that interpretability, evaluation, and value-sensitivity are built into the architecture rather than retrofitted onto it.
From models to matter.
Intelligence cannot be understood at any single level of description. The same phenomenon — a system maintaining a useful model of its environment while acting within it — recurs in transformer activations, cortical microcircuits, quantum measurement, and physical control. We study these in parallel because we believe their structures inform one another.
World models, grounded representations, and architectures that learn from interaction rather than from text alone.
Measurement, decoherence, and the structure of physical inference under uncertainty.
Cortical circuits as a working blueprint for active, embodied prediction under metabolic budget.
Dynamics, statistical mechanics, and the geometry of systems that act within the worlds they model.
Each plane is not a separate department but a different vantage onto a single landscape. Insights from one are expected to constrain the others; the work succeeds when they begin to converge.
Solid foundations, slow inferences.
Open by default. We publish what we learn, in papers, posts, and code, because frontier science is a collective enterprise and accountability requires legibility.
Empirical to the bone. Theory advances by surviving contact with data. We instrument every claim and prefer measurements we can defend over narratives that sound good.
Small surface, deep stack. We resist the temptation to do everything at once. The lab is organized around a narrow set of bets pursued with disproportionate depth.
Respect for prior art. The questions we ask are old. The communities of cognitive science, condensed-matter physics, dynamical systems, and theoretical neuroscience have been here for decades. We borrow what works and credit what we borrow.
We are looking for collaborators.
LML is a small, deliberately interdisciplinary lab. We are interested in researchers and engineers who treat this as a serious technical question rather than a slogan — people fluent in the mathematics of learning, the biology of cognition, the physics of measurement, or the engineering of real systems acting in the world.
If your work touches any of these, we would like to hear from you.