We publish what the lab learns. No corporate blog, no SEO sludge — engineering memos, research summaries, and the occasional polemic. Roughly monthly. Always written by someone whose hands were on the keyboard.
A field survey of where the bottleneck has actually moved — from algorithms to electrons, from FLOPs to fab capacity, from VCs to hyperscaler treasuries. Twelve calls and twelve we got wrong.
┌─────────────────────────┐ │ ATLAS · Q1 2026 │ │ ─────────────────── │ │ > 47 production stacks │ │ > 12 power-grid maps │ │ > 9 case studies │ │ > 3 contrarian takes │ │ │ │ [ READ ALL ▾ ] │ └─────────────────────────┘
Marine fouling, sensor degradation, and seasonal biology mean a perception model trained in spring is wrong by autumn. Numbers from the Poseidon fleet on what to retrain when.
Sara Mendoza
When the GPU bill exceeded the salaries of the team running the workload, something had to change. Notes on moving inference to on-prem at scale.
Marta Chen
A field-tested calibration protocol developed across 800 grow-bench installations. Why the cheap-and-noisy sensor with a Bayesian wrapper beats the expensive-and-precise one almost every time.
Mira Patel
A frank account of the studio process — the working sessions that mattered, the term-sheet edits that didn't, and the day we found out salmon farming had a perception problem.
Ahmet Alpat
Most reported numbers from fine-tuning are leakage in disguise. A working protocol for the eval harness we now ship across every venture in the cohort.
Marta Chen & Daniel Park
Headline yields, eval methodology, and a frank discussion of the failure modes we found across nine crops in three countries.
Mira Patel
Operational lessons from the Mutfak.ai pilot rollout. Mostly about installer training, not firmware.
Tom Bauer
A retrosynthesis tool isn't a model — it's a pipeline. Notes on why ChemSynth's edge has turned out to be data infrastructure, not the T5 fine-tune.
Aisha Khan
One email a month. Real engineering content. We don't sell the list and we don't pad to fill the schedule.