No AGI without efficient inference.
Next chokepoint battle is open inference.
Everyone is scaling models. Game is token per dollar per watt.
Read the piece →Hand-crafted SIFT vectors for a computer vision startup that was years too early.
The work survived through Intuit, through enterprise software, through multiple AI winters. Now the compute and tooling have caught up.
Kalmantic is what came out. An Agentic Lab working on inference economics, agent infrastructure, and the layers value migrates to when execution gets cheap.
Bamboo shoot pre-seed bets that SV ignores at Upekkha. Many companies in so far. The early end of the market is where I learned to see.
The writing here runs alongside the lab. I keep going back to first principles. Feynman is where I picked up the test. If you cannot explain it to a child, you do not understand it. Taleb, Naval, Housel are the other rooms I keep walking into. So I write essays for people running companies, and I make decks that try to explain the same ideas to an eleven-year-old. Both are the same work. The eleven-year-old is the harder reader.
Conversations with my eleven-year-old nephew. Quakeview ideas I wish I had known when I was younger.
Pieces that still stand.
The Kalman filter is where I learned to find signal in noise. Kalmantic carries that lineage.
An Agentic Lab working on inference economics, agent infrastructure, NemoClaw strategy, benchmarking, and the long arc of systems that think.
Code. peakinfer and peakweights, both open source.
Papers. Seclens on arxiv, and PeakWeights with Vamshi Ambati.
Books. Five so far.
More at kalmantic.com.
Small experiments that escaped containment.
