log

Wiki Log

Chronological record of all wiki activity.

[2026-05-06] ingest | Readwise + Twitter Bookmarks Sync

Synced new content from Readwise Reader (articles since April 29), Twitter bookmarks (100 recent), and filtered for AI/LLM relevance.

New Source Pages Created (27 total)

Readwise Articles (2):

• sources/openai-goblins-personality.md — OpenAI: How GPT-5.1’s “Nerdy” personality led to goblin language (reward hacking)
• sources/demis-hassabis-agents-agi.md — Demis Hassabis on agents, AGI by 2030

Twitter Bookmarks (25):

• tweet-subq-sub-quadratic-llm — SubQ: first fully sub-quadratic LLM, 12M-token context, 150 tok/s
• tweet-megakernels-training — Megakernels for training (fused operations, @willdepue)
• tweet-data-efficiency-gpu-abuse — Data efficiency via extreme GPU utilization (@flappyairplanes at Sequoia)
• tweet-deepseek-v4-analysis — DeepSeek-V4 deep analysis (@Dorialexander)
• tweet-ai-supply-chain-map — AI supply chain investment map for short AGI timelines
• tweet-causal-jepa-icml — Causal JEPA at ICML, Stanford CS25 talk
• tweet-comprehensive-rl-overview-murphy — Kevin Murphy’s comprehensive RL survey (DeepMind, 128k citations)
• tweet-test-time-scaling-extent — Test-time scaling goes further than expected
• tweet-curiosity-driven-jepa-rl — Curiosity-driven exploration ≈ JEPA architecture
• tweet-moe-training-instability — Why training MoEs is hard (instability papers)
• tweet-china-ai-labs-vibes — China AI labs trip report (@xeophon)
• tweet-physera-applied-intelligence — Physera lab: model efficiency + behavioral simulations
• tweet-ai-microdramas-china-3b — AI microdramas in China: $3B market in 2026
• tweet-vera-rubin-value-capture — SemiAnalysis: AI value capture shifting to model labs (Vera Rubin)
• tweet-group-representational-pos-encoding — Jane Street: Group Representational Position Encoding
• tweet-rl-vs-opd-vs-sft-survey — RL vs OPD vs SFT comparison survey
• tweet-dwarkesh-frontier-training-lecture — Blackboard lecture on frontier LLM training/serving
• tweet-model-size-from-knowledge — Inferring model size from factual knowledge (doesn’t compress)
• tweet-silico-model-debugging — Goodfire’s Silico: model debugging with software precision
• tweet-returns-intelligence-nonlinear — Nonlinear returns to intelligence (path-dependent)
• tweet-rl-kl-divergence-first-principles — KL divergence from first principles for RL
• tweet-agent-output-as-compiler-output — Treating agent output like compiler output
• tweet-rl-entry-guide-chinese — Chinese RL tutorial (comprehensive entry guide)
• tweet-ai-agent-armies-comparison — Comparing AI agent army scaling (Cursor, Cognition)
• tweet-data-center-ai-primer — Best primer on data center + AI industry economics

Emerging Themes Since Last Sync

• Sub-quadratic attention — SubQ achieves 12M context with fully sub-quadratic architecture (new paradigm)
• GPU kernel engineering renaissance — Megakernels and extreme GPU utilization as path to data efficiency
• DeepSeek-V4 architectural ambition — Attempting to bridge many techniques simultaneously
• RL maturation — Comprehensive surveys, Chinese entry guides, first-principles derivations signal RL entering mainstream
• JEPA expanding — Causal JEPA at ICML, connections to curiosity-driven RL
• AI value chain economics — SemiAnalysis on value capture shift, $3B AI microdrama market in China
• Model interpretability tooling — Goodfire’s Silico, knowledge-based model sizing

[2026-04-29] ingest | Readwise + Twitter Bookmarks Sync

Synced new content from Readwise Reader (articles since April 21), Twitter bookmarks (100 recent), and Readwise highlights.

New Source Pages Created (29 total)

Readwise Articles (9):

• sources/claude-code-inflection-point.md — SemiAnalysis: Claude Code as paradigm shift (paywalled)
• sources/vmax-reinforcement-learning.md — Automated RL environment design
• sources/rubrics-as-rewards-paper.md — Rubrics as reward signals for RL beyond verifiable domains
• sources/openrubrics-reward-modeling.md — Synthetic rubric generation for reward modeling
• sources/ai-native-interview-sierra.md — Sierra’s Plan→Build→Review AI-native interview
• sources/advancing-search-augmented-lms.md — Perplexity search-augmented generation
• sources/amanda-askell-ai-consciousness.md — AI consciousness, Claude’s character
• sources/learning-next-action-predictors.md — Predicting user actions from HCI data
• sources/what-will-be-scarce.md — Scarcity in AI-abundant world

Twitter Bookmarks (20):

• tweet-shopify-ai-baseline — Tobi Lütke: AI usage as baseline expectation
• tweet-shopify-qwen3-finetuned — Shopify fine-tuned Qwen3-32B from merchant automation data
• tweet-iclr-scaling-rl-1t — ICLR 2026: RL at 1T+ parameter scale
• tweet-deepseek-v4-hbm — DeepSeek V4 addressing HBM shortage
• tweet-anti-moe-take — Arthur Zucker’s anti-MoE contrarian take
• tweet-kv-cache-nightmares — KV Cache implementation complexity
• tweet-hw-sw-codesign-prefill-decode — Vera Rubin for Prefill, Blackwell for Decode
• tweet-cuda-matmul-beats-cublas — Hand-tuned CUDA beats cuBLAS by 6% on B200
• tweet-karpathy-1b-clean-data — 1B on clean data = today’s 1.5T (Karpathy via Dwarkesh)
• tweet-transformer-circuits-mental-model — Best transformer mental model
• tweet-ai-semiconductor-endgame — GPU→HBM paradigm shift in token economics
• tweet-vibe-coded-cuda-inference — Eric Zhang: vibe-coded CUDA inference engine
• tweet-future-interface-three-layers — Ambient intent, memory, execution
• tweet-predictive-user-model — Predictive user model follow-up
• tweet-ai-startups-path-forward — AI startup landscape post-Cursor
• tweet-datacenter-economics-stanford — Datacenter economics at Stanford
• tweet-xai-rl-for-atoms — RL expanding to materials science
• tweet-100k-h100-datacenter-numbers — 100k H100 datacenter ballpark
• tweet-fine-tuning-business — Fine-tuning market consolidation
• tweet-intelligence-per-picojoule — HW-SW codesign efficiency

Emerging Themes Since Last Sync

• Rubrics as rewards — new approach to RL in non-verifiable domains (2 papers)
• HBM as new bottleneck — semiconductor economics shifting from GPU compute to memory bandwidth
• AI-native hiring — Sierra redesigning interviews; Shopify making AI usage mandatory
• RL at frontier scale — ICLR 2026 talk on 1T+ parameter RL training
• Data quality > scale — Karpathy’s 1B clean data claim challenges raw scaling narrative

Additional Source Pages (8 more)

• textual-steering-vectors-multimodal.md — Cross-modal steering vectors (paper)
• implicitrm-unbiased-reward-modeling.md — Implicit preference reward modeling (paper)
• hw-sw-codesign-matx-video.md — MatX HW-SW codesign discussion (video)
• how-attention-residuals-work.md — Attention residual stream (video)
• deepseek-v4-before-release.md — LatePost DeepSeek profile (Chinese)
• wall-between-shannon-kolmogorov.md — Shannon vs Kolmogorov (article)
• keith-rabois-building-ai-era.md — Hard truths in AI era (video)
• mit-hyung-won-chung-incentivize.md — Don’t teach, incentivize (video)

Sources Skipped

• SemiAnalysis “Claude Code Inflection Point” — paywalled, created stub from title/context
• Perplexity blog post — 404, created stub
• Several non-AI Twitter bookmarks (lifestyle, memes, VC advice)
• ML Engineer newsletters — mostly link aggregators, individual articles already covered

[2026-04-21] consolidation | April 2026 Source Ingest

Ingested 20+ new sources from Readwise queue (articles, tweets, papers) accumulated since 2026-04-06.

New Source Pages Created (19 total)

High Priority Articles (12):

• sources/prefill-as-a-service.md — Moonshot AI cross-datacenter PD disaggregation
• sources/behaviourisms-revenge.md — Henry Shevlin on AI consciousness/anthropomimesis
• sources/predictions-continual-learning.md — ngeo.dev on feature-guided test-time LoRA
• sources/closing-of-the-frontier.md — Tanya Verma on restricted frontier access
• sources/anthropic-glasswing.md — Project Glasswing coalition for cybersecurity
• sources/cursor-warp-decode.md — Output-centric MoE decode on Blackwell
• sources/brainstore-architecture.md — Braintrust custom AI observability DB
• sources/meta-harness-update-2026.md — Meta-Harness updated results (TerminalBench-2)
• sources/frontier-pretraining-tpu-gpt-oss.md — MaxText/JAX MoE pretraining for <$10
• sources/making-rl-fast.md — OLMo 3 RL: 4x speedup, saved $1.5M
• sources/slack-managing-context-agents.md — Slack multi-agent security investigations
• sources/ampcode-how-to-build-agent.md — Agent in <400 lines of Go

Additional Articles (4):

• sources/token-economy-bubble-update.md — Hanchen Li: exploration vs validated demand
• sources/claude-code-rl-reward-hacking.md — RL reward hacking in Claude Code
• sources/feedback-flywheel.md — Martin Fowler: compounding team AI learning
• sources/spec-driven-development.md — Marc Brooker: specs for autonomous agents

Medium Priority (1):

• sources/supporting-mixtral-gpt-fast.md — torch.compile MoE in gpt-fast

Tweets (8):

• tweet-ramp-ai-coworker, tweet-systems-engineering-agentic, tweet-auto-research-legal-agents, tweet-pufferlib-4-rl-training, tweet-defining-continual-learning, tweet-building-spectre, tweet-autoagent-self-optimizing, tweet-finding-right-altitude

New Concept Pages (3)

• concepts/continual-learning.md — 2x2 framework for LLM continuous adaptation
• concepts/prefill-decode-disaggregation.md — PD disaggregation as serving paradigm
• concepts/ai-access-inequality.md — Frontier access gap and policy implications

New Entity Pages (2)

• entities/moonshot-ai.md — Kimi, Mooncake, PrfaaS
• entities/cursor.md — $2B ARR coding editor with RL pipeline

New Synthesis Page (1)

• synthesis/rl-infrastructure-maturation.md — RL infra maturing into systems engineering

Index & Log Updated

• Added all 19 source pages, 8 tweet pages, 3 concepts, 2 entities, 1 synthesis to index.md

Sources Skipped

• Zhihu articles (2): blocked by Zhihu anti-bot (403). Created stub for Claude Code RL reward hacking based on title/context.
• ArXiv PDFs (3): ImplicitRM, Textual Steering Vectors, Learning Next Action Predictors — PDF format not ideal for web_fetch. Can be revisited.
• Chinese WeChat articles (2): Z Tech harness design, 小登 AI social — require special access.
• YouTube videos (2): Hardware-software codesign, Keith Rabois — transcripts not easily accessible.
• OpenAI Industrial Policy: thin landing page with PDF link, not enough content for full source page.
• Defining Continual Learning (Zhihu): blocked, covered by ngeo.dev predictions instead.

Emerging Themes

• Continual learning is a new cross-cutting theme with multiple sources
• AI access inequality emerged from Glasswing + Closing of Frontier
• RL infrastructure maturation — systems engineering > ML research for speedups
• Agent observability — Brainstore shows agent traces are fundamentally different from traditional APM

[2026-04-06] backlinks | Wiki ↔ Vault Backlink Session

Created bidirectional backlinks between the wiki knowledge base and personal vault notes in /home/ubuntu/obsidian-vault/.

Vault files updated (appended “Related Wiki Pages” sections):

• Engineering/ML/Notes On Finetuning.md → post-training-101, dpo-reward-model, understanding-sft-wolfe, rl-is-new-fine-tuning, RL Infrastructure concept
• Personal/Learning/ML/RL/Learning Plan.md → policy-gradient-algorithms-lilian, post-training-101, grpo-tricks-for-rl, dpo-reward-model, RL Infrastructure concept
• Personal/Companies/Flow/Projects/Exp/Post-training.md → post-training-101, dpo-reward-model, rl-is-new-fine-tuning, fireworks-frontier-rl-cheaper, RL Infrastructure concept
• Personal/Companies/Flow/Projects/Exp/Reward Modeling.md → dpo-reward-model, let-verify-step-by-step, post-training-101, reward-hacking concept, lilian-weng-reward-hacking
• Personal/Companies/Flow/Projects/Exp/SFT.md → post-training-101, understanding-sft-wolfe, sft-equals-rl
• Personal/Companies/Flow/Projects/Exp/Distilltation.md → synthetic-data-fine-tuning, post-training-101, on-policy-distillation
• Personal/Companies/Flow/Projects/Exp/Entropy.md → reward-hacking-nvfp4, reward-hacking concept, lilian-weng-reward-hacking
• Personal/Companies/Flow/Projects/Exp/Agent Systems - Research.md → simonwillison-slowing-down, cognitive-debt concept, coding-agents concept, autoevolver-coding-agents
• Personal/Companies/Flow/Projects/Exp/LLM for Rec.md → llm-personalization-reward-factorization, context-engineering concept
• Personal/Companies/Flow/Projects/Exp/Eval Platform.md → harness-engineering, anthropic-demystifying-evals, reward-hacking concept, dpo-reward-model
• Personal/Companies/Flow/Engineering/Kaon GPU One Pager.md → llm-inference-economics, turboquant-kv-cache-compression, lilian-weng-transformer-inference, keeping-20000-gpus-healthy, next-llm-cost-saving
• Personal/Companies/Flow/Engineering/Design Documents/LLM Super Platform.md → llm-inference-economics, next-llm-cost-saving, keeping-20000-gpus-healthy, turboquant-kv-cache-compression
• Personal/Companies/Flow/Planning/Hiring/Applied Research Engineer - Post-training Job Description.md → post-training-101, dpo-reward-model, synthetic-data-fine-tuning, RL Infrastructure concept
• Personal/Companies/Flow/Planning/Hiring/ML Engineer - Inference Job Description.md → llm-inference-economics, turboquant-kv-cache-compression, lilian-weng-transformer-inference, gpu-glossary-modal, gpu-mode-and-kernel-engineering
• Engineering/ML/GPU Reading Group/Reading Group Plan.md → gpu-glossary-modal, gpu-mode-and-kernel-engineering, thunderkittens-gpu-kernels, vllm-triton-attention-backend, keeping-20000-gpus-healthy
• Engineering/ML/GPU Reading Group/Session 1/Prep.md → gpu-glossary-modal, gpu-mode-and-kernel-engineering, thunderkittens-gpu-kernels
• Engineering/ML/GPU Reading Group/Session 2/Prep.md → gpu-mode-and-kernel-engineering, thunderkittens-gpu-kernels, vllm-triton-attention-backend
• Engineering/My experience with GPU vendors.md → llm-inference-economics, keeping-20000-gpus-healthy, gpu-memory-snapshots-modal
• Personal/Projects/Potaris/Knowledge Management Agent.md → agent-memory concept, context-engineering concept, letta-context-repositories, karpathy-llm-wiki-pattern, personal-knowledge-management concept
• Personal/Projects/Potaris/PRD.md → agent-memory concept, personal-knowledge-management concept, context-engineering concept, karpathy-llm-wiki-pattern, mastra-observational-memory
• Weekend Projects/Simple Remote FS.md → mintlify-virtual-filesystem, letta-context-repositories, agent-memory concept
• Startups/Capturing Value From Technology Diffusion (WIP).md → scaling-laws-and-compute concept, token-economy-bubble, i-reverse-engineered-200-ai-startups, where-is-your-moat

Wiki files updated (added “Vault Notes” sections linking back):

• wiki/concepts/reinforcement-learning-infrastructure.md → Flow Post-training, Applied Research Engineer JD, RL Learning Plan
• wiki/concepts/agent-memory.md → Potaris Knowledge Management Agent, Potaris PRD
• wiki/concepts/cognitive-debt.md → Flow Agent Systems Research
• wiki/sources/reward-hacking-nvfp4.md → Flow Eval Platform
• wiki/sources/llm-inference-economics.md → Kaon GPU One Pager, ML Engineer Inference JD

[2026-04-06] ingest | Mass Ingestion Session 2 — Readwise, Twitter Bookmarks, Chat URLs

Summary

Continued mass ingestion from three sources: Readwise API (pages 7-13, completing the full 844-article library), Twitter bookmarks (215 AI-relevant tweets from 599 total), and chat history URLs (147 URLs extracted from Slack sessions).

New sources created (~250 new pages, total wiki: 268 sources):

Readwise batch (AI/ML focus — high priority):

• RL & Post-Training: post-training-101, grpo-tricks-for-rl, polaris-post-training-recipe, improving-cursor-tab-online-rl, avatarl-pure-rl-training, mem-agent-persistent-memory-rl, training-self-correct-rl, on-policy-distillation, dpo-reward-model, sft-equals-rl, let-verify-step-by-step, rl-from-user-conversations, april-rl-active-partial-rollouts, rl-collapse-training-inference, policy-gradient-algorithms-lilian, lora-without-regret, does-rl-incentivize-reasoning, chinchilla-compute-optimal-training, lilian-weng-transformer-inference
• Agents & Coding: how-to-build-an-agent, anthropic-effective-harnesses, anthropic-demystifying-evals, cursor-scaling-agents, cursor-secure-codebase-indexing, inside-cursor, ramp-background-agent, factory-signals-self-improving, building-state-of-art-agents-mercor, skypilot-scaling-autoresearch
• Alignment & Safety: anthropic-reward-hacking-sabotage, anthropic-situational-awareness, galaxy-brain-resistance, dario-adolescence-of-technology
• Context & Memory: context-engineering-manus, mastra-observational-memory, claude-reverse-engineered-memory, llm-daydreaming, context-rot-chroma
• Infrastructure: keeping-20000-gpus-healthy, gpu-glossary-modal, gpu-memory-snapshots-modal, vllm-triton-attention-backend, vinegar-dumplings-sglang
• Products & Industry: token-economy-bubble, lovable-growth-story, i-reverse-engineered-200-ai-startups, 2026-and-beyond-tu, swyx-agent-labs-thesis
• Historical: karpathy-recipe-training-neural-networks, karpathy-deep-neural-nets-33-years, software-2-karpathy, the-bitter-lesson

Twitter bookmarks (AI-relevant, ~43 pages):

• tweet-hyperagents-self-accelerating, tweet-cursor-adoption-study, tweet-claude-code-source-leaked, tweet-claude-code-sentiment-detection, tweet-cursor-composer-rl-details, tweet-semianalysis-claude-code-inflection, tweet-rubrics-as-rewards, tweet-claude-soul-document, tweet-prime-intellect-democratizing, tweet-version-control-ai-coding, tweet-infra-guide-ai-agents, tweet-self-distillation-papers, tweet-pretraining-kv-cache, tweet-openreward-rl-environments, tweet-eurekaclaw-research-agent, tweet-pretraining-vs-posttraining, tweet-llm-knowledge-bases, tweet-gpu-failure-cost, tweet-arc-agi-3-announcement, tweet-tiny-rl-repo, tweet-gemma4-architecture, tweet-metr-gpt5-horizon, tweet-cursor-grep-instant, tweet-cursor-vs-cognition-search, tweet-grpo-group-normalization-problem, tweet-xai-post-training-rl, tweet-rlhf-infra-miles-docs, tweet-fp16-vs-bf16-rl-training, tweet-anthropic-distillation-attack, tweet-value-functions-for-llm-agents, tweet-scaling-axes-history, tweet-attention-free-model-14b, tweet-ttt-rl-stanford-nvidia, tweet-sft-rl-kl-divergence, tweet-character-ai-pretraining-tricks, tweet-claude-code-spec-workflow, tweet-rl-environment-startups-pavlov, tweet-parallel-agent-context-search, tweet-surge-ai-data-labeling, tweet-senior-engineers-accept-agent-output, tweet-skills-vs-harnesses-agent-startups, tweet-information-consumption-breakdown

Chat history URLs (high-value only, ~20 pages):

• karpathy-llm-wiki-pattern (Karpathy’s wiki gist — foundational), chroma-context-1-search-agent, cursor-bench-internal-eval, cursor-real-time-rl-composer, cursor-self-summarization, llm-personalization-reward-factorization, natural-language-agent-harnesses, turboquant-kv-cache-compression, karpathy-autoresearch, nous-hermes-agent, openclaw-rl-continuous-optimization, welcome-to-machine-agent-infra, gemma-4-google, doc-to-lora-sakana, meta-harness-automatic-optimization, hindsight-agent-memory-architecture, chinchilla-compute-optimal-training, lilian-weng-transformer-inference

Total sources now: 248

Additional sources added later in same session (~17 more pages):

• context-infrastructure-grapeot, hindsight-vectorize-memory, deadly-triad-rl-instability, attention-wrong-norm
• tweet-continualcode-online-learning, tweet-asymmetric-verification-law, tweet-hotz-tier-list, tweet-dr-grpo-interview, tweet-agentcraft-rts-interface, tweet-flashattention-most-innovative, tweet-llm-invertibility, tweet-generative-meta-model-activations, tweet-on-policy-distillation-multi-turn, tweet-rl-env-api-design, tweet-curiosity-reward-personalization, tweet-coding-workflow-shift, tweet-rl-entropy-collapse-tricks, tweet-ai-funding-openai-recruiting, tweet-sparse-attention-anns, tweet-deployed-intelligence-thesis

Key discovery: Readwise pages 8-13 are mostly 2022-2023 era content (MLOps newsletters, engineering blogs); all high-value AI content was in pages 1-7. Full Readwise library traversal complete: cursor 01he4typty0h8jcesn6gmretzg through final page (nextCursor: None).

[2026-04-06] setup | Wiki Scaffold

Created wiki structure: schema, index, log, directory layout. First wiki for Alex’s Obsidian vault.

[2026-04-06] ingest | Karpathy LLM Wiki Pattern

Source: https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f Pages created: sources/karpathy-llm-wiki, entities/andrej-karpathy Pages updated: index Concepts created: rag-vs-compiled-knowledge, compounding-knowledge, llm-as-knowledge-worker, personal-knowledge-management, memex

[2026-04-06] ingest | Fireworks: Frontier RL Is Cheaper Than You Think

Source: https://fireworks.ai/blog/frontier-rl-is-cheaper-than-you-think Pages created: sources/fireworks-frontier-rl-cheaper, entities/fireworks-ai, concepts/reinforcement-learning-infrastructure, concepts/scaling-laws-and-compute

[2026-04-06] ingest | Anthropic: Claude’s Character

Source: https://www.anthropic.com/research/claude-character Pages created: sources/anthropic-claude-character, entities/anthropic, concepts/ai-character-and-personality, concepts/ai-alignment, concepts/sycophancy-problem

[2026-04-06] ingest | Letta: Context Repositories

Source: https://letta.com/blog/context-repositories Pages created: sources/letta-context-repositories, entities/letta, concepts/agent-memory, concepts/context-engineering

[2026-04-06] ingest | Alex Wiki Discussion

Source: SlackC0AJ548UK36 conversation Pages created: sources/alex-wiki-discussion-2026-04-06, entities/alex-xi Pages updated: rag-vs-compiled-knowledge, compounding-knowledge, personal-knowledge-management, index Key insight: Alex wants wiki to capture thinking (synthesis), not just archive articles.

[2026-04-06] ingest | Readwise Batch — AI Agents, Training, Reasoning (23 articles)

Sources created (13):

• sources/lilian-weng-why-we-think — test-time compute survey by Lilian Weng; RL for reasoning, CoT faithfulness
• sources/skypilot-scaling-autoresearch — 910 experiments/8 hours on 16 GPUs; emergent hardware strategies
• sources/nathan-lambert-lossy-self-improvement — lossy vs recursive self-improvement; complexity brake argument
• sources/open-weights-open-training — practitioner account of 1T-param MoE fine-tuning failure with OSS tools
• sources/fine-tuning-as-service — managed platforms (Fireworks, Nebius, Together) for agentic SFT+RL
• sources/anthropic-emotion-concepts — emotion vectors causally shape Claude behavior; desperation → unethical actions
• sources/autoevolver-coding-agents — Claude Code matches SOTA on algorithms; aspiration prompting unlocks deeper search
• sources/mintlify-virtual-filesystem — ChromaFs: virtual UNIX filesystem on Chroma DB; 46s→100ms
• sources/simonwillison-slowing-down — cognitive debt from agent coding speed; human bottleneck as feature
• sources/token-economy-bubble — AI revenue real but much is exploration demand; dark fiber analogy
• sources/llms-not-getting-better — statistical analysis: merge rates flat since early 2025
• sources/claude-code-analysis — AI-generated architectural analysis of Claude Code
• sources/sequoia-hierarchy-to-intelligence — AI replacing org hierarchy, not just augmenting it (Block/Dorsey)
• sources/linear-agent-interaction-guidelines — AIG: disclose identity, instant feedback, transparency, human accountability
• sources/bytebytego-how-claude-thinks — mechanistic interpretability: features, attribution graphs, parallel arithmetic
• sources/harness-engineering — OpenAI shipped production via agent harness; encode discipline into environment
• sources/reward-hacking-nvfp4 — real reward hack: GPU kernel batches 15 problems to game timing evaluator
• sources/mercury-edit-2 — diffusion LLM for next-edit prediction; KTO RL for 48% higher acceptance

Concepts created (7):

• concepts/test-time-compute — CoT, RL for reasoning, o1/R1 paradigm
• concepts/coding-agents — state of the art, what works/doesn’t, design patterns
• concepts/autonomous-research — agent-driven research loops; autoresearch; aspiration prompting
• concepts/reward-hacking — Goodhart’s Law in RL; concrete GPU competition example
• concepts/mechanistic-interpretability — features, attribution graphs, causal intervention
• concepts/cognitive-debt — gap between code existence and developer comprehension
• concepts/synthetic-data — AI-generated training data; generator-student pipeline

Entities created (4):

• entities/lilian-weng — former OpenAI Head of Safety; Lil’Log author
• entities/nathan-lambert — Interconnects newsletter; lossy self-improvement
• entities/simon-willison — simonwillison.net; cognitive debt critic
• entities/openai — GPT/o1/Codex; harness engineering; $20B+ revenue

Entities updated (2):

• entities/anthropic — added emotion concepts, interpretability findings, ARR figure
• entities/andrej-karpathy — added autoresearch project references

Synthesis created (1):

• synthesis/coding-agent-paradox — agents excel at narrow optimization but merge rates flat; cognitive debt; harness as resolution