Scaling Intelligence in the Physical Arena

Frontier models currently struggle with design, manufacturing and assembly. Our thesis is that this is because they lack the engineering priors needed to reason reliably about physical systems. Blacksmith builds training data, environments and evals to develop those priors.

BS-0217_bearing-housing_asm.step STEP AP214 · 32 parts · mm
Loading assembly
A bearing housing assembly going together one part at a time: base plate, housing, flange bolts, a 6206 bearing built up from its rings, balls, cage and seals, then pressed into its seat, followed by the shaft, retaining ring, end cap and cap screws.

Coding agents demonstrate the potential of training at scale with verifiable feedback. Models can now make substantial contributions to software development, including the software used to build their successors. This creates a path toward recursive self-improvement.

Our bet is that physical engineering can follow a similar trajectory. Engineering projects contain decades of accumulated expertise, but much of it remains inaccessible to model training. We source projects from leading manufacturers and design teams around the world and turn them into training data spanning mechanical design, machining, simulation and electronics. We’re also collecting long-horizon engineering trajectories that connect design decisions to performance in production and real-world use.

We build environments where models use engineering software to complete tasks drawn from these projects. Our evals measure generated solutions against hard requirements such as load capacity, manufacturing constraints and assembly tolerances. This provides a reward signal for models and a measure of engineering competence for researchers.

We also want to understand how much of that experience transfers to unfamiliar problems. The research question is whether training across diverse projects builds engineering priors that generalize, reducing the need for task-specific training. If it does, a wider range of engineering work becomes economical to automate.

Take building AI, for example. Engineering-capable models could help develop hardware around the computational requirements of future models, while informing model design with physical constraints. We think this could extend recursive self-improvement into hardware. The payoff is just as large outside AI. Jet engines have grown more efficient mostly by running hotter, and their turbine blades already sit in gas above their own melting point; every degree they can take is repaid on each flight over decades of service. In data centers, power and heat are often the binding limits, so each watt saved in a voltage regulator, or carried off by a better cold plate, is a watt returned to compute.

We want a small team to build what currently takes an entire company. Plenty of good hardware never gets made because the engineering costs more than the idea is worth. Think of a lab that needs an instrument built around one experiment, or a factory that wants a robot for one awkward job on its line. We want projects like these to be worth trying on a small budget.

We’re hiring builders, engineers and researchers. Get in touch.