ai

Files tagged with AI topics.

• Claude Opus 4.6
• Slay — Evaluation & Search (the 1v1 case) — eval-beats-depth in wide-branching search; handcrafted terms as the eventual NNUE’s features
• Economics
• The LLM Grounding Problem
• The Cyborg Model
• Cognitive vs. Motor Skills
• The Limits of Utopia
• Camelot From YouTube
• Slay
• Slay-C
• Measurement, Causality, and Free Will
• Computation and Information Theory
• Gaming
• The Multiplayer Coalition Problem
• Working With Claude
• Set Mastering Pipeline
• Diplomacy: 7 AI Models
• H-Neurons: The Neural Basis of Hallucination
• Planner-LM Composites — where the agent actually lives; Cicero, the 100,000× planning claim, the higher praxeology bar
• Energy-Based Models — names the middle layer of the composite; Aleph + Lean at 99.4% on PutnamBench; LeCun’s JEPA convergence; the Sudoku tell
• Claude Code Skill Engineering — 4 rules + 2 patterns from Anthropic engineers; vault skill audit
• LLM Game Benchmark — Outline
• The Birthmark and the Search for Truth
• Obsidian Plugin Setup
• Karpathy LLM Wiki — Independent Convergence
• Level 6 — Direct Execution
• LLMs as Praxeological Actors
• ELIZA — The Pattern Match That Won’t Die
• AI History — A Personal Arc
• OOP Neurons
• Claude Code on a Persistent Cloud VM
• iPad + Claude Setup
• Career
• The Role Landscape
• Career Strategy — The Performative Loop
• Tools Roadmap
• Presentation Generator Spec
• Practicality vs. Precision — Principled LLM Code — legacy slop, worse-is-better as economics, the refactor-cost inversion, persistence as the fix
• The Context Cache Hierarchy — memory layers = CPU cache; drift = promotion with no eviction; the placement rule + a hygiene check
• Transpilation as a Grounding Strategy — formal-language grounding gradient; deterministic transpile to a grounded language; 6502→C, COBOL payoff
• Repairing LLM Code — The Two Oracles — Spiro’s vibe-coded-mess video assessed; the two-oracle model; blind-reader experiment on the NA1 decompiler
• CaptainMeme vs. 6 Cicero — The Human/AI Split in Press Diplomacy — An expert human diplomacy player plays full-press blitz Diplomacy against six Cicero AIs and narrates the game.
• Demolition Man (1993) — A Case Study in Managed Utopia — A 30-year-old sci-fi film that predicted the exact failure modes of AI-managed societies.
• DJ Set 1 — Progressive/Trance Mix — 11-track set built around texture-over-harmony philosophy, with a BPM staircase from 120→129.
• Gunboat Diplomacy and Diplodocus — The Planner Without the Mouth — Meta AI’s Diplodocus plays Gunboat Diplomacy — the variant with no text communication — and beat every human in Meta’s 2022 Speedboat Tournament.
• The Three-Layer Method — Spec, Verifier, Environment — Karpathy’s Spec/Verifier/Environment stack; gap-first audit (request-vs-rule, the missing PreToolUse guardrail)
• Arithmetic Scarcity and the 3D Problem — the coprocessor cycle (EAE → 8087 → GPU → NPU) with local inference as the current phase-A workload; the 3D strategies map 1:1 onto quantization / sparsity / KV-cache / constrained decoding, and the memoization-depth law transfers intact
• Arithmetic as a Purchase — Sixty Years of Paying for Multiply — the NPU moment as the fifth iteration of one pattern, not a novelty; capability regresses whenever computing finds a cheaper substrate, then gets re-bought
• Is AI Making People Dumber? (Word War — Charsky vs. Bunn) — the offloading debate done well: meta-analyses vs. the “cognitive crutch” RCT, staged tool access (calculators are gated by grade by design), and the seed-corn / apprenticeship-pipeline harm — where the cost lands on the supply of future experts, not the current user
• AI as a Cognitive Tool — outsource computation, never verification: verification makes offloading safe and preserves the skill, because checking the work is the practice. Car reductio, wheelchair-and-car, productivity-not-intelligence, fix-the-instrument. But the rule is a frontier, not a floor — verification gets tooled and atrophies too (run the tool, replace the part); the obligation that survives is being able to notice and retool
• The Lemonade Stand Experiment — grounding made corporeal: an ice maker with a lever no AI has hands to pull, a Rube Goldberg display customers couldn’t see, a ten-minute cup; plus “Stress doesn’t ship lemonade” — risk without felt stakes
• Watching the Solver Learn — the training-harness layer the vault’s solver projects lack: raw + smoothed objective, policy entropy as a confidence signal, bounded cost metric, non-interrupting event log; plus fixed-size perception (92 raycasts, cost independent of enemy count) and action rate as a planning-horizon decision
• Media Pipeline — (draft) homes the generator list (image/audio/video) so it is never re-searched; md as grounded control language, generators as deterministic tools, LLM orchestrating via pointers
• Building Swarms — Token Rules as Architecture — Nate B.
• The Substrate Is the Governing Mechanism — Agents coordinating through a shared store is read as the thing that makes them ungovernable.
• Comments and the Distance to an Oracle — comments as an agent-readable context layer and an inter-agent coordination channel; type them so eviction is mechanical
• The Contract Model vs. the Substrate Model — a per-delivery contract (delivery / constraints+guardrails / proof artifact / outside-verification / owner) is the vault’s architecture with the accumulation stripped out — contracts don’t compound, substrate does; plus the artifact-vs-reader split in verification independence, and tests written only to pass as oracle collapse
• Oracles Are Objective Functions — an oracle is not a truth detector, it is a direction to move in — the same object backprop descends and an LP declares; “token generator” vs. maximiser, and where the linear-programming analogy actually breaks (implicit objective, no constraint set, no certificate)