Amit Patnaik

Make the system legible

Thu, 23 Apr 2026 00:00:00 GMT

Why the through-line between explainers, MCP tools, and intelligence dashboards is legibility: building command surfaces for systems people need to understand and act on.

Week of 22 April – Jensen on Dwarkesh, Mythos zero-days, and the 7nm ≈ Hopper reframe

Wed, 22 Apr 2026 00:00:00 GMT

The inaugural briefing. One interview dominated the week's signal, and most of this maps back to it; future briefings range wider.

The configuration layer

Wed, 22 Apr 2026 00:00:00 GMT

Why the valuable Indian AI work over the next three years is not building frontier models but wrapping the ones that exist for the institutions and principals who will not build that wrapper themselves.

The AI stack

Wed, 22 Apr 2026 00:00:00 GMT

From electrons to tokens, one layer at a time. Tap any of the nine layers to unfold what it is, who operates there, and what breaks.

India's AI build

Wed, 22 Apr 2026 00:00:00 GMT

The functional map of who does what in Indian AI. Policy, compute, energy, models, applications, talent, data – area by area, with the gaps drawn in.

AI chip buyer's guide

Wed, 22 Apr 2026 00:00:00 GMT

Which chip for which workload. Eight chips (H100 / H200 / B200 / GB300 / TPU / Trainium / MI300X / Groq) compared on specs, ecosystem, availability, and fit – plus a three-question decision tree.

The RL post-training map

Wed, 22 Apr 2026 00:00:00 GMT

What happens to an LLM after pre-training. SFT, RLHF, DPO, PPO, GRPO, RLVR explained – with the pipeline, the terms, and why post-training is the growth slice of 2026 compute.

AI, term by term

Wed, 22 Apr 2026 00:00:00 GMT

What the words mean when you read about AI chips. Search across 100-plus terms, filter by category, click through to definitions and context.

Nine layers between electrons and tokens

Tue, 21 Apr 2026 00:00:00 GMT

A map of the AI compute stack, pulled apart one layer at a time, with failure modes drawn in. Interactive version.