The lab for
AI-assisted invention.
Neural-guided program synthesis models that invent from perception.
SOTA AI explores only what it has already seen.
Current state-of-the-art models perform poorly in R&D because of their inability to explore outside of their distribution.
Generalization
Transformer-based architectures require massive scale in compute and data to generalize. Even then, asked to perform outside their distribution, they usually fall short.
Black Box
LLMs and VLAs are purely deep-learning based, and therefore black boxes — difficult to predict, struggling to follow logical rules, and not interpretable.
Restricted Search
SOTA architectures search only within their distribution, and therefore only within what they have seen — limiting them in scientific discovery and invention.
Neural networks combined with program synthesis.
Together they give artificial intelligence the foundation it needs to learn and invent efficiently.
Neural-guided program synthesis learns from just 10–100 examples — orders of magnitude less data than the large-scale networks that dominate AI today. These systems follow logical rules and use Helmholtz dreaming, an unsupervised method for "imagining" new ways to use their current understanding of a program space.
The Perception Layer
A model must perceive its environment — through language or vision — to make sense of its world. Observations let it learn the rules that change the state of its environment.
The Brain Layer
Converting perceptions into neural representations that guide program search lets the model learn state transitions from few examples, driving causal understanding and finding the primitive rules of its world.
The Invention Layer
With an internal model of the world and the programs that represent its state transitions, the model extrapolates to find new structure in that space — leading to inventions.
The current R&D market is sized at ~$3 trillion.
Estimated value projections (USD) across the global research & development landscape.
Leif
A neural-guided program synthesis model that perceives an environment, learns its primitive rules from a handful of examples, and extrapolates that structure into invention.
Perceive
Leif observes its environment through language or vision, grounding raw signal into the entities and states it can reason about.
Reason
Neural representations guide a program search that recovers the fundamental, causal rules governing state transitions — from just 10–100 examples.
Invent
With an internal world model in hand, Leif extrapolates to find new structure in the program space — proposing designs no training set contained.
Helmholtz Dreaming
An unsupervised learning loop where Leif "imagines" new ways to compose the programs it already knows — expanding its library of reusable primitives without new labelled data.
Interpretable by Construction
Because invention is expressed as explicit programs, every output is inspectable, verifiable, and follows logical rules — not a black box.
1000× Less Data
Learning from 10–100 examples rather than internet-scale corpora makes Leif viable in data-scarce R&D domains like aerospace.
Beyond Distribution
Program search lets Leif explore structure outside of what it has seen — the prerequisite for genuine scientific discovery.
MicroWorlds
As far as we know, the first benchmark built to measure an AI's ability to invent from primitives — not to recall, but to discover new structure in a program space.
Invention, not recall
Each world hides a small set of primitive rules. A model must induce them from a few observations, then compose them into solutions it was never shown.
Out-of-distribution by design
Tasks are constructed so that pattern-matching fails — success requires genuine extrapolation beyond the training distribution.
Few-shot primitives
Worlds expose 10–100 examples. Scoring rewards both correctness and the compositional reuse of discovered primitives.
Three founders.
AI research, computer vision, and hands-on R&D engineering — the disciplines it takes to build machines that perceive, reason, and invent.
Kade Carlson
2+ years researching zeroth-order optimization for AI-driven design, with a publication in AISTATS. 3+ years of industry RL at Sandia National Labs, Reservoir Computing research at the AFRL, and leading AI research projects at P1 AI.
Noah Muthler
4+ years developing computer vision algorithms for docking, refueling, and debris-removal satellite missions. 2+ years researching data-driven non-cooperative satellite pose estimation and 3DGS vSLAM for defunct satellite repair.
Micah Delattre
3+ years of academic experience researching vehicle path-following algorithms and an autonomous depth control system for a bioinspired underwater robot platform. 1+ years of professional experience developing vehicle control and management software for defense applications.
Contact us.
Building the lab for AI-assisted invention. We'd love to hear from collaborators, partners, and the curious.