AI So Whats
Edition

Saturday, August 22, 2026

3 issues 23h window ~3 min read

The money is rotating from AI software to AI's physical layer

2 sources

Two issues point at the same trade from different ends. The AI Opportunity backs Apollo Atomics — a $26M seed for a light-water reactor 40x smaller, targeting 3c/kWh, already sitting on 20 GW of hyperscaler LOIs because utility interconnection queues now run a decade. Superhuman covers the other end: Unitree's Shanghai debut closed up over 460% at roughly a $50B valuation, and SoftBank put $200M into Gravis Robotics. Power and actuators, not model weights, are where the new capital is landing.

Humanoids got faster and cheaper in the same week

1 source

Unitree's "Superman" robot is claimed at 12.7 m/s against Usain Bolt's 12.4, while Nor Robotics put a home humanoid, the Nori A3, at $1,688 — two orders of magnitude under the six-figure norm. Generalist AI's GEN-1.5 meanwhile learns a new physical skill from a single 3–12 second demo, hitting 59% on the first try and 83% after brief practice. Capability and cost are moving in opposite directions at once, which is usually the point a category stops being a demo.

Agent memory is converging on plain markdown files

1 source

Ben Tossell rebuilt his personal agent around four hand-edited files — AGENTS.md, code.md, todos.md and a memory.md pointer — and uses git commits instead of a log, explicitly rejecting vendor auto-memory because stale saved preferences were over-steering his answers. The same file set runs unchanged across Claude, Codex and ChatGPT. For anyone standardizing internal agent tooling, the portable-file pattern is looking like the safer bet against per-vendor memory features.

Why I Invested in Apollo Atomics

So what

Firm nuclear power, not chips or capital, is now the binding constraint on AI data center buildout through 2030.

The AI Opportunity's Guillermo explains why he invested in Apollo Atomics, a stealth nuclear startup that shrank the reactor's steam generator 20x, enabling a reactor 40x smaller and aimed at beating natural gas on cost.

  • Uses proven light-water reactor tech and existing supply chains instead of exotic fuels, easing NRC approval
  • Targets 3c/kWh power and a 300MW plant built in under 24 months, 4-5x cheaper than conventional designs
  • Raised $26M seed led by FCVC, with YC, Paul Graham and others; its 40kW MIT demonstrator already went critical
  • Reports 20 GW of signed LOIs from hyperscalers facing decade-long utility interconnection queues
  • Roadmap: 1MW demonstrator in 2027, commercial deployment in 2028

⚡ Unitree's robot breaks Usain Bolt's speed record

So what

Unitree's $50B IPO pop signals China's humanoid robotics race is now attracting serious public capital, not just VC money.

Unitree's shares surged over 460% in its Shanghai IPO debut, valuing the humanoid robot maker at ~$50B, just as it unveiled a "Superman" robot claiming to beat Usain Bolt's speed record.

  • Unitree stock closed at 845 yuan, ~$50B valuation; founder Wang Xingxing's stake now worth $11B+; Superman robot reportedly hits 12.7 m/s vs Bolt's 12.4 m/s
  • Generalist AI's GEN-1.5 learns new physical skills from one 3-12s demo, succeeding 59% on first try, 83% after brief practice
  • Nor Robotics unveiled a $1,688 home humanoid (Nori A3), undercutting six-figure industry pricing
  • SoftBank put $200M into Gravis Robotics; Uber and Zipline partnered on drone delivery targeting 1M daily deliveries by 2029
  • Waymo won approval to expand across the Bay Area and LA

How I built this

So what

A minimal AGENTS.md/skills file pattern, not vendor memory features, is emerging as the reusable architecture for personal and team AI agents.

Ben Tossell rebuilt his personal AI agent from scratch, ditching automatic memory-saving for a minimal, manually-edited file structure that works the same across Claude, Codex, and ChatGPT.

  • Core files: AGENTS.md (instructions), code.md (build preferences), todos.md, and a memory.md pointer file, using git commits instead of a log.md for history
  • Rejected auto-saved memory because it over-steered responses based on stale preferences instead of letting him brainstorm freely
  • Chat history is already searchable via local session files, so no separate memory store is needed for "what did we discuss last week"
  • Distinguishes instructions (agent identity/rules) from skills (repeatable task processes like email triage), each just another .md file
  • Assembled 06:30 PT from newsletters received in the previous 23 hours. Issues covered here have been archived out of your inbox.