06-reference

innermost loop kimi k3 frontier moat

2026-07-17·reference·source: Innermost Loop·by Alex Wissner-Gross

"Welcome to July 17, 2026" — @theinnermostloop

Why this is in the vault

This issue lands the week Kimi K3 — a 2.8T-parameter open model from a 300-person lab — posted benchmark wins against Claude Fable 5 and GPT-5.6 Sol on code, browsing, and chip design tasks, prompting a serious analyst debate about whether the compute moat that underwrites $650B in closed-lab valuations is cracking. Simultaneously: Tata begins India's first wafer fab (90nm), Valar Atomics raises $1B at $6B for small reactors to feed data centers, and the Schema harness hits ~99% on ARC-AGI-3 via scaffolding alone. Each of these is a direct signal for the Markov capital-cycle phase tracker and RDCO's agentic toolchain thesis.

Issue contents

Open weights crack the compute-moat thesis. Moonshot AI published detailed evals for Kimi K3: trails Fable 5 and GPT-5.6 Sol on aggregate, but takes outright wins on Program Bench, SWE Marathon, SpreadsheetBench 2, and BrowseComp; matches Fable 5 on GPU kernel optimization, Triton-compiler builds, and autonomous chip design. A DeepMind researcher concluded "the frontier is no longer something money can buy," pointing to a 300-person lab compressing frontier training out of compute scarcity. An ex-Meta PM piled on: if open weights rival closed models, why pay Fable prices, why ban chip exports, why is Anthropic worth 50x Moonshot? Counterpoints: K3 still loses to a five-month-old Mythos Preview; Anthropic reportedly sits on a 10T-parameter model; recursive self-improvement needs research taste plus compute, not just weights. AAII scorecard: K3 at $0.94/task (half of Opus, triple its predecessor) with a 13-point intelligence jump.

Scaffolding as free intelligence. The Schema harness — models write each game's mechanics as an executable program — hit ~99% on ARC-AGI-3 with zero weight changes. AI forecasters became statistically indistinguishable from superforecasters, outranking them on market questions. Fable 5.1 rumored next week; GPT-6 within 1.5 months. Google reportedly months behind on Gemini 3.5 Pro.

Where models meet markets. Netflix: ~300 titles used GenAI, including 17 minutes of enhanced documentary footage produced 2× faster at half cost. Satya Nadella told Copilot engineers Fable's limits "don't make sense" — "editorially controlled." Anthropic arranging billions in bank credit ahead of planned IPO. Z.AI pacing toward $1B ARR; K3 lopped 20% off its shares.

Geopolitics and governance. Xi Jinping pitched China as AI partner to the Global South with 5,000 training slots. Demis Hassabis proposed an international watchdog to vet frontier models pre-release, citing Mythos's cyber capabilities as the warning shot. MLB banned AI from dugout iPads after a third of the league fed live games into decision engines.

Substrate and energy. Tata will make India's first wafers on 90nm technology — decades old, but a new national rung on the fab learning curve. A top AWS executive defecting to Meta to build data centers. Valar Atomics raising ~$1B at $6B to supply small reactors to those centers.

Robotics and embodiment. Chinese MMA robot lost its head mid-fight and kept swinging. New York district hired humanoid teaching assistant "Sally" for classroom support and 24/7 homework help.

Biotech and cognition. Regulators approved Merck's Lipfendra, an oral cholesterol pill beyond statins. Kalshi opened prediction markets on clinical trials. EEG research: switching attention between speakers creates a brief window where both streams are encoded simultaneously.

Space. Musk says Starlink V3 boosts space bandwidth ~100×; Starship 13 scrubbed on engine starts, next attempt early next week. Astronomers detected a helium atmosphere around LHS 1140 b — a rocky, Earth-like world 48 light-years out in a habitable zone — first direct atmosphere detection for any rocky exoplanet.

Mapping against Ray Data Co

Tata's 90nm wafer fab and Valar Atomics' $1B small-reactor raise are the two most actionable signals here for the Markov phase-tracker. The Markov spec places chip-fab/memory in Phase 2 (capacity announcements); a new national entrant (India, 90nm) and a purpose-built energy supply raise aimed at data centers are Phase 2 confirmation signals, not noise. Tata 90nm feeds the "new-entrant capacity" category the tracker watches; Valar is an upstream energy-infrastructure bet on sustained data-center capex.

The more complex overlay: Kimi K3's compute-moat crack is a counter-risk to the Phase 2 bull case. If open weights commoditize frontier capability, closed-lab revenue compression follows — which could pull Phase 3 (capacity-online/glut) forward or flatten the demand curve for HBM. The "skeptic" framing in the issue (K3 still trails on 5-month-old Mythos Preview; true frontier stays legally sandbagged) preserves the bull case for now, but this issue is the clearest articulation yet of the counter-thesis. Worth flagging in the Markov tracker's risk-register.

The Schema harness result (scaffolding = ~99% ARC-AGI-3, zero weight changes) is an independent signal relevant to RDCO's agentic COO approach: harness design is still high-alpha even at today's model capabilities — free intelligence via scaffolding is exactly the mode RDCO operates in.

Related