Blog
A complete list of notes, essays, and research from our team.
Physical AI: Quadruped market survey
Part 3 of the Physical AI series. The companies building quadrupeds, the work their robots do, and the emerging market in India.
Inference: Fewer, smaller forwards
Part 3 of the Inference Engineering series. How D2F and Funnel overlap canvases, commit tokens earlier, and shrink later denoising passes.
Physical AI: Quadruped mechanics
Part 2 of the Physical AI series. How ground contact, leg design, sensing, and control allow quadrupeds to move over uneven terrain.
Inference: Making every denoising pass fast
Part 2 of the Inference Engineering series. Replacing only the routed experts with Triton W4A16 under the unchanged official DiffusionGemma sampler.
Physical AI: Quadruped control architecture
Part 1 of the Physical AI series. Four control layers, their update rates, and three common controller architectures.
Inference: From noise to words
Part 1 of the Inference Engineering series. How discrete corruption, better training objectives, and block generation made text diffusion practical.
On systems of intelligence
Through systems of records, engagements, predictions, states and agents.
Looped language models
Trading unique parameters for recurrent depth, and attempting to make single-user decode compute-bound by keeping a looped transformer block resident in L2 cache.
Sarvam's illusion of safety
An independent mechanistic and adversarial audit of Sarvam-30B and Sarvam-105B across 14 Indian languages. The models are 6x more likely to comply with harmful requests in Indian languages than in English.
Releasing goedel-mHC-1B
Releasing the first open 1B+ language model with hyperconnections.