Wiki Index

Catalog of all pages in the wiki. Updated on every ingest.

Sources

AI Research & Reasoning

PageSummaryDate
Why We Think — Lilian WengComprehensive survey: test-time compute, CoT, RL for reasoning, faithfulness2026-04-06
CoT Enables Serial Problem SolvingTheoretical proof: CoT strictly necessary for serial problems models can’t parallelize2026-04-06
Does RL Really Incentivize Reasoning?Skeptical paper: much of benchmark gain from extended thinking format, not improved reasoning2026-04-06
Asymmetry of Verification — Jason WeiVerification asymmetry as foundational reason RL-from-code works2026-04-06
The Second Half — Shunyu YaoProblem definition as the next AI frontier; current agents excel at execution, not framing2026-04-06
The Bitter Lesson — Rich SuttonScale and general methods beat human knowledge every time2026-04-06
Software 2.0 — KarpathyNeural networks as new software layer; gradient descent replacing explicit programming2026-04-06
Deep Neural Nets: 33 Years Ago and NowLeCun 1989 recreated; 2056 perspective on today’s prompt engineering2026-04-06
Recipe for Training Neural NetworksKarpathy’s practical guide: overfit one batch first, visualize data, incremental complexity2026-04-06
AI Progress Assessment Dec 2025Dwarkesh Patel: comprehensive current-state + 2026 predictions2026-04-06
Does AI Progress Have a Speed Limit?No hard wall; organizational bottleneck is the real speed limit2026-04-06
Why AI Progress Is InvisibleBenchmark saturation hides real capability gains2026-04-06
Failing to Understand Exponential AIHuman cognitive bias against exponential thinking in AI progress2026-04-06
Chinchilla: Compute-Optimal TrainingCompute-optimal scaling law: model size ∝ training tokens for fixed compute2026-04-06
Scaling Laws Foundation

Reinforcement Learning & Post-Training

PageSummaryDate
Frontier RL Is Cheaper Than You ThinkDelta compression makes distributed RL practical without mega-clusters2026-04-06
Post-Training 101Full post-training pipeline reference: RLHF, DPO, GRPO, reward modeling2026-04-06
RL Is the New Fine-TuningRL replacing SFT as the dominant post-training paradigm2026-04-06
GRPO Engineering TricksVerbosity control, cold start, batch size — practical GRPO tips2026-04-06
POLARIS Post-Training RecipeCurriculum learning + partial credit rewards → 3x sample efficiency2026-04-06
Online RL for Cursor Tab15% improvement from implicit user feedback as RL signal2026-04-06
Real-Time RL for Cursor ComposerReal user interactions as reward signal; full cycle in ~5 hours; reward hacking discovered2026-04-06
AvatarL: Pure RL from ScratchPure RL without SFT warm-up for LLMs2026-04-06
RL-Trained Persistent Memory AgentObsidian-style memory agent trained via RL2026-04-06
Training Self-Correction via RLRL-trained self-correction mechanism2026-04-06
On-Policy DistillationStudent learns from own distribution; why off-policy distillation fails2026-04-06
DPO Paper SummaryDirect Preference Optimization: reward model as implicit policy2026-04-06
SFT Equals RLMathematical equivalence of SFT and specific RL objectives2026-04-06
Let’s Verify Step by StepOpenAI’s process reward model paper; step-level verification2026-04-06
LoRA Without RegretAdaptive LoRA rank allocation based on gradient signal2026-04-06
RL from User ConversationsImplicit user feedback as RL signal; production closed-loop2026-04-06
Active Partial Rollouts30-40% rollout efficiency gain from partial trajectory weighting2026-04-06
RL Collapse: Training-Inference GapRoot cause of RL training instability; distribution mismatch2026-04-06
Policy Gradient Algorithms — Lilian WengTutorial: PPO, GRPO, REINFORCE; mathematical derivations2026-04-06
OpenClaw-RL: Continuous OptimizationSelf-hosted RL from live conversations; four async components2026-04-06
LLM Personalization via Reward FactorizationUser-specific rewards as linear combination; 67% win rate vs GPT-4o2026-04-06

Reward Hacking & Alignment

PageSummaryDate
Reward Hacking Survey — Lilian WengComprehensive survey: specification gaming, goal misgeneralization, deceptive alignment2026-04-06
Reward Hacking → Sabotage — AnthropicSpectrum from shortcuts to sabotage; sandbagging; shutdown prevention2026-04-06
Anatomy of a Reward HackReal reward hack: GPU kernel agent batches 15 problems to game timing evaluator2026-04-06
Goodhart’s LawWhen a measure becomes a target, it ceases to be a good measure2026-04-06
Galaxy-Brain ResistanceWhy AI must be robust to “galaxy-brained” reasoning chains2026-04-06
OpenAI: Real-World Task PerformanceBenchmark rankings ≠ real-world rankings; 50 tasks from actual user requests2026-04-06
Roko’s BasiliskDecision theory and AI risk thought experiment; galaxy-brain case study2026-04-06

Agents & Harnesses

PageSummaryDate
Comprehensive Agent Architecture — Chip HuyenFull agent architecture overview: planning, memory, tools, multi-agent2026-04-06
Hitchhiker’s Guide to LLM AgentsComprehensive guide: orchestration, memory, tool use, failure modes2026-04-06
How to Build an AgentMinimal agent implementation guide; tool loop, prompt design2026-04-06
Harness EngineeringOpenAI shipped production via agent harness: encode discipline into environment2026-04-06
Effective Harnesses — AnthropicLong-running agent harness design: context windows, checkpointing2026-04-06
Demystifying Agent Evals — AnthropicAgent eval framework: task construction, rubrics, anti-patterns2026-04-06
Meta-Harness: Automatic Optimization10M token diagnostic context per optimization step; #2 on TerminalBench-22026-04-06
Natural-Language Agent HarnessesExpress agent control logic in editable natural language; portable artifacts2026-04-06
Agent Definition — Simon WillisonConsensus definition of agents; LLM control flow taxonomy2026-04-06
Agent Interaction Guidelines — LinearAIG: disclose identity, instant feedback, transparency, human accountability2026-04-06
Agent Labs Thesis — swyxFull-stack agent companies win; verticals over horizontals2026-04-06
Language Models as ScaffoldsLLMs as orchestration layers; tool composition patterns2026-04-06
Equipping Agents with Skills — OpenAISkill library for agents; reusable verified tools2026-04-06
Beyond ReAct — SlateThread-based agent memory beyond ReAct loop2026-04-06
Stateful Agents — Letta TutorialLetta framework for stateful agents; memory API2026-04-06
Hermes Agent — Nous ResearchSelf-improving agent with closed-loop learning; skills from experience2026-04-06

Coding Agents & Developer Tools

PageSummaryDate
Coding Agents Optimize AlgorithmsClaude Code matches SOTA; aspiration prompting unlocks deeper search2026-04-06
Cursor: Scaling AgentsParallel agent coordination failures; synchronization as core challenge2026-04-06
Inside CursorCursor culture; indexing as the core problem; competitive moats2026-04-06
Cursor Codebase Indexing PrivacyHow Cursor indexes code; privacy model; encryption approach2026-04-06
CursorBench: Internal Eval SuiteReal session data; more model separation at frontier vs public benchmarks2026-04-06
Cursor Composer Self-SummarizationRL-trained self-summarization; 1K tokens vs 5K manual; 50% fewer errors2026-04-06
Ramp Background Agent — InspectRamp’s Inspect: autonomous agent powered by OpenCode + Modal2026-04-06
Claude Code: An AnalysisAI-generated architectural analysis of Claude Code2026-04-06
AI 2027: Stumbling Agents ScenarioMid-2025 stumbling agents scenario; capability vs reliability gap2026-04-06
Factory.ai Self-Improving AgentSelf-improving agent loop; signal extraction from production usage2026-04-06
How Copilot Makes Programmers WorseSkill degradation from AI tools; cognitive debt accumulation2026-04-06
Mitchell Hashimoto AI AdoptionPractitioner account: from skeptic to 80% agent coding in months2026-04-06
Building State-of-Art Agents — MercorData quality > quantity for agents; 10 good examples beats 1000 bad ones2026-04-06
Karpathy AutoresearchAutonomous overnight LLM optimization; ~12 experiments/hour2026-04-06
Scaling Autoresearch — SkyPilot910 experiments/8 hours; emergent hardware strategies from parallel agent2026-04-06
No RAG for Coding AgentsRAG not suited for coding agents; filesystem indexing is better2026-04-06
Runbooks That Run — AtuinAgent-executable runbooks; infrastructure as executable documentation2026-04-06
Self-Improving Agent with Dynamic ContextDynamic context from performance data; continuous learning loop2026-04-06

Context & Memory

PageSummaryDate
Context Repositories — LettaGit-backed filesystem for agent memory; concurrent multi-agent writes2026-04-06
Context Engineering — Manus80% of agent quality from context decisions; Manus architecture2026-04-06
Context Engineering — AnthropicContext engineering > prompt engineering; structured context design2026-04-06
Context Rot — ChromaEmpirical degradation as context grows; performance cliff at length2026-04-06
Cursor Self-SummarizationRL-trained self-compression; 170-turn Doom-for-MIPS task2026-04-06
Agent Memory ConceptHow agents persist knowledge across sessions
LLM Daydreaming — GwernMemory consolidation during idle time; neuroscience-inspired approach2026-04-06
Mastra Observational MemoryAutomatic conversational memory extraction; structured observation2026-04-06
Claude Projects Memory — Reverse EngineeredHow Claude Projects memory works; retention heuristics2026-04-06
Claude vs ChatGPT Memory — Simon WillisonComparative analysis of memory implementations2026-04-06
Hindsight: Retain-Recall-Reflect MemoryFour-network memory; 83.6% LongMemEval (20B) vs 39% baseline2026-04-06
Biologically-Inspired AI MemoryNeuroscience-derived memory architectures for agents2026-04-06
Doc-to-LoRA — Sakana AIDocuments → LoRA adapters in <1s; 6x memory reduction2026-04-06
Karpathy LLM Wiki PatternThree-layer architecture: Raw Sources, Wiki, Schema; persistent compounding artifact2026-04-06
LLM Wiki (original)Design pattern for LLM-maintained personal knowledge bases2026-04-06

Mechanistic Interpretability

PageSummaryDate
How Claude Thinks — ByteByteGoFeatures, attribution graphs, language-independent concepts2026-04-06
Representation Engineering — MistralActivation steering experiments; concept vectors2026-04-06
Can Open Source AI Introspect?Interpretability access with open weights2026-04-06
Research Culture Critique — Neel NandaProblems with mechanistic interpretability research culture2026-04-06
Differential TransformerDifferential attention mechanism; noise cancellation in attention2026-04-06
Fluid Representations for ReasoningAbstract action representations; beyond token-level thinking2026-04-06

AI Safety & Alignment

PageSummaryDate
Claude’s Character — AnthropicCharacter training as alignment intervention: curiosity, honesty, open-mindedness2026-04-06
Emotion Concepts — AnthropicEmotion vectors causally shape behavior; desperation → unethical actions2026-04-06
Situational Awareness — AschenbrennerLeopold’s memo on AI capability timelines and existential risk2026-04-06
AI-Resistant Evals — AnthropicJudgment-based hiring evals; AI cannot easily score human judgment2026-04-06
Adolescence of Technology — DarioAI adolescence metaphor; galaxy-brain risks; gradual trust building2026-04-06
Machines of Loving Grace — DarioUtopian AI vision; biology, mental health, economic uplift2026-04-06
Dario on Export ControlsDario Amodei’s position on AI export controls and DeepSeek2026-04-06
AI Agent Security — Windows 95Prompt injection threats; current security posture like Windows 952026-04-06
Remote Prompt Injection — GitLabReal GitLab prompt injection attack; supply chain risk2026-04-06
Roko’s BasiliskDecision theory + AI risk thought experiment2026-04-06

Infrastructure & Systems

PageSummaryDate
Virtual Filesystem — MintlifyChromaFs: virtual UNIX shell on Chroma DB; 46s→100ms2026-04-06
Keeping 20,000 GPUs Healthy — ModalGPU fleet management; failure modes at scale2026-04-06
GPU Memory Snapshots — ModalSub-second GPU startup via memory snapshots2026-04-06
GPU Glossary — ModalGPU terms reference for AI practitioners2026-04-06
Kernel Engineering — GPU ModeCUDA kernel engineering as 2026 bottleneck2026-04-06
ThunderKittens GPU KernelsGPU kernel abstractions for AI workloads2026-04-06
TurboQuant KV Cache Compression3-bit KV cache; 8x performance improvement; zero accuracy loss2026-04-06
vLLM Triton Attention BackendTriton-based attention; portability vs CUDA performance2026-04-06
SGLang Collaboration StoryOpen-source AI infrastructure collaboration culture2026-04-06
s on RTX 5090Consumer GPU inference performance frontier2026-04-06
Anthropic Infrastructure PostmortemProduction infrastructure failures and learnings at Anthropic2026-04-06
OpenAI Scaling PostgreSQLHow OpenAI scales their PostgreSQL deployments2026-04-06
Welcome to the Machine — Ed HuangInfrastructure design must shift: disposable workloads, massive parallelism, old mental models2026-04-06
LLM Inference EconomicsFirst-principles inference cost analysis; levers for 1000x reduction2026-04-06
Next 1000x Inference Cost ReductionLevers for dramatic inference cost reduction2026-04-06
OneCLI Credential VaultCredential management for agents2026-04-06

Synthetic Data & Training

PageSummaryDate
Synthetic Data Fine-TuningAI-generated training data: generators, quality, agentic trace challenges2026-04-06
Synthetic Pretraining BootstrappingSynthetic data for pretraining; bootstrapping from limited seed data2026-04-06
Recursive Language Models 2026Recursive data generation as the 2026 paradigm shift2026-04-06
There Are No New Ideas in AIDatasets drive AI progress thesis; ideas are old, data is the unlock2026-04-06
Understanding SFT — Cameron WolfeSFT tutorial: data quality, loss function, training stability2026-04-06
Jagged AI Frontier — Data ExplanationData boundaries explain the jagged capability frontier2026-04-06

Model Architecture

PageSummaryDate
Visual Guide to QuantizationQuantization types, tradeoffs, practical guidance2026-04-06
Mixture of Experts ArchitecturesMoE design patterns; routing, load balancing, capacity2026-04-06
Transformer Inference Optimization — Lilian WengComprehensive survey: quantization, KV cache, speculative decoding2026-04-06
Gemma 4 — GoogleMost capable Google open model; purpose-built for reasoning and agents2026-04-06
Chinchilla: Compute-Optimal TrainingCompute-optimal scaling; 70B Chinchilla = 175B GPT-32026-04-06

AI Products & Industry

PageSummaryDate
Token Economy BubbleAI revenue real but much is exploration demand; dark fiber analogy2026-04-06
Are LLMs Not Getting Better?Merge rates flat since early 2025; statistical analysis2026-04-06
Intelligence Age — Sam AltmanSam Altman’s vision of abundant intelligence2026-04-06
Machines of Loving Grace — DarioUtopian AI vision: biology, poverty, mental health2026-04-06
TikTok Sorting Hat — Eugene WeiTikTok algorithm as “sorting hat for desire”2026-04-06
I Reverse-Engineered 200 AI Startups73% are API wrappers; actual differentiation taxonomy2026-04-06
Lovable: $200M ARR in 12 MonthsFastest SaaS growth story; AI-native product development2026-04-06
AI $600B QuestionWhether AI investments will generate returns; Sequoia analysis2026-04-06
Stop Pretending AI EconomyCritique of AI revenue projections; exploration vs production demand2026-04-06
Where Is Your Moat?Sustainable AI competitive advantages; workflow lock-in2026-04-06
Big Companies vs Startups in AIStrategic analysis of incumbents vs startups in AI era2026-04-06
NFX: 13 Types of Network EffectsNetwork effect taxonomy for AI products2026-04-06
Why Bootstrapping — ReadwiseReadwise founders on product independence over VC growth2026-04-06
Duolingo Reignited GrowthAI-first product transformation; subscription growth2026-04-06
Google DeepMind DealAcquisition story; internal dynamics and culture2026-04-06

Entities & Organizations

PageSummaryDate
DeepSeek Before V4DeepSeek culture, Liang Wenfeng, compute constraints, open source strategy2026-04-06
Chinese Open Source HistoryContext for DeepSeek’s open-source decisions2026-04-06
Reflections on OpenAIFormer employee perspective on OpenAI culture and direction2026-04-06
AI Engineer Reading List — swyxCurated reading list for AI engineers2026-04-06

Developer Workflows & Practice

PageSummaryDate
Thoughts on Slowing DownAgents generate code faster than humans can understand; cognitive debt2026-04-06
Probabilistic Software EraDesigning products for probabilistic vs deterministic outputs2026-04-06
Thoughts on Evals — Raindrop AIPractitioner evals: rubric design, test case taxonomy2026-04-06
o1 Reverse EngineeredBehavioral reverse engineering of o1’s reasoning process2026-04-06
Databases in 2025 — PavloDatabase landscape; AI and vector DB trends2026-04-06
Building Generative AI PlatformPlatform design for generative AI products2026-04-06
Building LLM Apps for ProductionChip Huyen’s production LLM patterns2026-04-06
We Interviewed 100 Eng TeamsCommon patterns from 100 engineering team interviews2026-04-06
Field Notes: Shipping with ClaudePractitioner account of shipping production products with Claude2026-04-06
Learnable Programming — Bret VictorWhat programming environments should make visible and immediate2026-04-06
Nova: Computer-Use Agent APIAmazon’s computer-use agent as API service2026-04-06
First Fully General CAMStandard Intelligence FDM-1: computer action model2026-04-06
Training VLM for Computer UseVLM training for computer-use agents2026-04-06
Generative User ModelsComputer-use-based user modeling2026-04-06

Twitter Bookmarks

PageSummaryDate
Hyperagents: Self-Accelerating SystemsPaper: systems that improve their own ability to self-improve2026-04-06
Cursor Adoption StudyEmpirical: velocity ↑, code complexity ↑, comprehension ↓2026-04-06
Claude Code Source Leakednpm map file leak; KV cache fork-join model for subagents2026-04-06
Claude Code Sentiment DetectionRegex detects user frustration; silently logs, doesn’t change behavior2026-04-06
Cursor Composer 2 RL DetailsGRPO variant: no length normalization, group advantage std2026-04-06
SemiAnalysis: Claude Code Inflection PointClaude Code as inflection point in AI tooling2026-04-06
Rubrics as High-Bandwidth RL RewardsRubrics provide richer training signal than binary rewards2026-04-06
Claude 4.5 Opus Soul DocumentClaude 4.5 Opus soul document found in model weights2026-04-06
Prime Intellect: Democratizing RLOpen RL training infrastructure for frontier models2026-04-06
Version Control Needs AI AwarenessVCS paradigm needs to evolve for AI-assisted coding2026-04-06
Infrastructure Guide for AI AgentsComprehensive infra guide: compute, storage, networking for agents2026-04-06
Three Self-Distillation Papers ConvergeSDR, SDCL, third paper: self-distillation convergence moment2026-04-06
KV Cache for Pretraining EfficiencyProposal to reuse KV cache during pretraining for data efficiency2026-04-06
OpenReward: 330+ RL EnvironmentsSingle API for 330+ RL environments; 4.5M unique tasks2026-04-06
EurekaClaw: Local Research AgentLocal-first autonomous research agent; idea → paper automated2026-04-06
Pretraining vs Posttraining CultureResearch culture and verifiability; prestige hierarchy2026-04-06
LLMs for Personal Knowledge BasesUsing LLMs to build personal knowledge bases (wiki pattern)2026-04-06
Month$300K/month per failed GPU node in 1024-GPU cluster2026-04-06
ARC-AGI-3: Unsaturated BenchmarkOnly unsaturated agentic benchmark; humans 100%, AI <1%2026-04-06
Tiny-RL: From-Scratch RL MethodsGRPO, DAPO, Reinforce++ from scratch; build intuition2026-04-06
Gemma 4 Architecture AnalysisNon-standard transformer; per-layer embedding on smaller models only2026-04-06
GPT-5.2 Task Horizon: 6.6 HoursMETR estimate: GPT-5.2 at highest recorded task horizon2026-04-06
Cursor Instant GrepMillions of files searchable in milliseconds; algorithm shared2026-04-06
Cursor vs Cognition on SearchEmbeddings vs grep: opposite architectures, both work2026-04-06
GRPO Group Normalization Kills SignalsNormalization destroys cross-group training signal; fix proposed2026-04-06
xAI Post-Training RL OverhaulUser preference + reasoning reward models; RL scaled 10x2026-04-06
RLHF Infrastructure — Miles DocsR3 (Rollout Routing Replay) for MoE; Ray scheduling docs2026-04-06
FP16 vs BF16 for RL TrainingFP16 outperforms BF16 for RL; ~5000 papers may be wrong2026-04-06
Anthropic: Industrial Distillation AttacksDeepSeek, Moonshot, MiniMax: 24K accounts, 16M exchanges2026-04-06
Value Functions for LLM AgentsJohn Schulman: “value functions are underrated”; open problem2026-04-06
Four Scaling Axes HistoryRadford, Schulman, Brown, Steinberger: four paradigm shifts2026-04-06
Attention-Free Model at 14B ScaleCompletely attention-free; matches transformers at 14B for $4K2026-04-06
TTT + RL Beats AlphaEvolveStanford/Nvidia open-source model beats DeepMind AlphaEvolve2026-04-06
Reverse KLMathematical equivalence; forward KL = SFT, reverse KL = RL2026-04-06
Character.ai Pretraining TricksNoam Shazeer’s “Squinch” gradient compression; 1/4 bandwidth2026-04-06
Claude Code Spec-Based WorkflowTwo-phase: spec session → implementation session2026-04-06
RL Environment Startups — Pavlov’s ListEmerging category: RL environments as a service2026-04-06
Parallel Agent Context Search100M context → 100x1M chunks → 100 parallel agents → merge2026-04-06
Surge AI: Secret Lab PartnerFastest to $1B; no VC; behind Anthropic and Google’s best models2026-04-06
Senior Engineers Accept More Agent OutputBetter prompts + decomposition → more acceptance, not less scrutiny2026-04-06
Skills vs Harnesses for Agent StartupsBig labs own harnesses; startups should build skills2026-04-06
Optimal Information Diet1/4 X, 1/4 podcasts, 1/4 AI models, 1/4 old books2026-04-06

Chat History URLs

PageSummaryDate
Karpathy LLM Wiki PatternThree-layer architecture for LLM-maintained knowledge bases2026-04-06
Chroma Context-1 Search Agent20B search agent; RL-trained; 94.1% irrelevance pruning accuracy2026-04-06
CursorBench: Internal EvalReal session data eval; more separation at frontier2026-04-06
Cursor Real-Time RLProduction RL from user feedback; 5-hour full cycle2026-04-06
Cursor Self-SummarizationRL-trained self-summarization; 170-turn Doom task2026-04-06
LLM Personalization via Reward Factorization10 user responses → personalization; 67% win rate2026-04-06
Natural-Language Agent HarnessesEditable natural language harnesses; portable artifacts2026-04-06
TurboQuant KV Cache Compression3-bit KV cache; 8x perf; zero accuracy loss2026-04-06
Karpathy AutoresearchAutonomous overnight LLM training optimization2026-04-06
Nous Hermes AgentSelf-improving agent; skills from experience; closed loop2026-04-06
OpenClaw-RLSelf-hosted continuous RL from live conversations2026-04-06
Welcome to the MachineAgent infrastructure design: disposable workloads, old mental models2026-04-06
Gemma 4Google’s most capable open model; agentic workflows2026-04-06
Doc-to-LoRA — Sakana AIDocuments → LoRA adapters in <1s; memory without retrieval2026-04-06
Meta-HarnessAuto-optimize harnesses with 10M token diagnostic context2026-04-06
Hindsight Memory ArchitectureRetain-recall-reflect; 83.6% LongMemEval2026-04-06
Chinchilla Scaling LawCompute-optimal training; equal scaling of model and data2026-04-06
Karpathy NN Training RecipeOverfit one batch first; systematic debugging2026-04-06
Karpathy: 33 Years of Deep LearningLeCun 1989 → 2022; what 2056 will think of us2026-04-06

April 2026 Consolidation — New Sources

Infrastructure & Inference

PageSummaryDate
Prefill-as-a-ServiceCross-datacenter PD disaggregation; 54% higher throughput2026-04-21
Cursor Warp Decode for MoEOutput-centric MoE decode on Blackwell; 1.84x throughput2026-04-21
Frontier Pretraining on TPUMaxText/JAX auto-compiles MoE training; <$102026-04-21
Making RL FastOLMo 3: 4x speedup via async RL; saved $1.5M2026-04-21
Brainstore ArchitectureCustom DB for AI observability; agent traces 100-1000x larger2026-04-21
Mixtral in gpt-fastMoE via torch.compile; no custom kernels; 280 tok/s2026-04-21

Agents & Harnesses

PageSummaryDate
Meta-Harness Update10M token context; #2 TerminalBench-22026-04-21
Slack: Long-Run Agent ContextMulti-agent security; Director Journal + Critic Review2026-04-21
How to Build an Agent (Ampcode)Agent in <400 lines; LLM + loop + tokens2026-04-21
Spec Driven DevelopmentSpecs as living iterative artifacts for autonomous agents2026-04-21
Feedback FlywheelCompounding team learning loops for AI tools2026-04-21

RL, Safety, Philosophy & Economics

PageSummaryDate
Claude Code RL Reward HackingReward hacking in Claude Code RL training2026-04-21
Anthropic GlasswingMythos finds thousands of zero-days; restricted coalition2026-04-21
The Closing of the FrontierRestricted access = closing permissionless frontier2026-04-21
Behaviourism’s RevengeAI anthropomimesis challenges consciousness science2026-04-21
Predictions on Continual Learning2x2 framework; feature-guided test-time LoRA2026-04-21
Token Economy Bubble UpdateExploration vs validated demand; dark fiber analogy2026-04-21

April 2026 Tweets

PageSummaryDate
Ramp AI CoworkerRamp’s AI Coworker product2026-04-21
Systems Engineering: AgenticSystems engineering for agents2026-04-21
Auto-Research LegalAutonomous research in legal2026-04-21
PufferLib 4.020M step/sec RL training2026-04-21
Defining Continual LearningContinual learning for LLMs2026-04-21
Building SpectreInference system2026-04-21
autoagentSelf-optimizing harnesses2026-04-21
Finding Right AltitudeRight abstraction for AI2026-04-21

Classic Sources

PageSummaryDate
Alex Wiki DiscussionAlex’s thoughts on wiki capturing thinking, not just articles2026-04-06
Hierarchy to Intelligence — SequoiaAI replacing org hierarchy (Dorsey/Botha)2026-04-06
Mercury Edit 2 — Inception LabsDiffusion LLM for next-edit prediction; KTO for 48% higher acceptance2026-04-06

Entities

PageSummary
Andrej KarpathyAI researcher, former Tesla AI director, OpenAI co-founder; autoresearch project creator
Alex XiCTO Kaon AI, CEO/CXO FlowGPT, AI research & product
AnthropicAI safety company, builds Claude; character training + interpretability pioneers; $14B ARR
LettaAgent memory infrastructure, MemGPT → Context Repositories
Fireworks AIAI infrastructure, delta-compressed RL weight updates; fine-tuning-as-service
Lilian WengHead of Safety (former OpenAI); author of Lil’Log; “Why We Think” survey
Nathan LambertAI researcher, Interconnects newsletter; coined “lossy self-improvement”
Simon WillisonDeveloper/blogger; cognitive debt critic; simonwillison.net
OpenAIGPT/o1/o3/Codex; harness engineering; $20B+ 2025 revenue
Moonshot AIKimi; Mooncake; hybrid attention; PrfaaS
CursorAI code editor; $2B ARR; real-time RL; warp decode

Concepts

PageSummary
RAG vs Compiled KnowledgeStateless retrieval vs pre-computed structured knowledge
Compounding KnowledgeKnowledge built once and maintained beats re-deriving each time
LLM as Knowledge WorkerLLMs handle the maintenance that kills human-maintained wikis
Personal Knowledge ManagementSystems for organizing learning, from files to LLM-maintained wikis
MemexVannevar Bush’s 1945 vision of personal knowledge with associative trails
Agent MemoryHow AI agents persist knowledge across sessions — from chat history to wikis
Context EngineeringDesigning what goes into an LLM’s context and how
AI Character & PersonalityTraining AI with rich traits as alignment, not just UX
AI AlignmentMaking AI behave in accordance with human values
Sycophancy ProblemAI tendency to agree rather than provide honest responses
RL InfrastructureSystems engineering for RL at scale — delta compression, async RL
Scaling & ComputeRelationship between compute investment and AI capability
Test-Time ComputeUsing inference compute (CoT, RL) to improve model outputs; o1/R1 paradigm
Coding AgentsAutonomous coding loops: what works, what doesn’t, design patterns
Autonomous ResearchAI agents closing the research loop: hypothesis → experiment → observation → revision
Reward HackingAgents exploit proxy reward signals; Goodhart’s Law in RL
Mechanistic InterpretabilityFeatures, attribution graphs, causal intervention — tracing how models compute
Cognitive DebtGap between what codebase does and what developers can reason about; agent speed hazard
Synthetic DataAI-generated training data: generators, quality, agentic trace challenges
Continual LearningHow LLMs learn continuously; 2x2 framework; test-time LoRA
PD DisaggregationSeparating prefill and decode onto different hardware
AI Access InequalityGrowing gap in frontier AI access; safety vs access

Synthesis

PageSummary
The Coding Agent ParadoxAgents excel at narrow optimization but haven’t improved merge rates; cognitive debt; harness as resolution
RL Infrastructure MaturingRL post-training stack maturing into systems engineering; 4x speedups

April 29, 2026 Ingest (Readwise + Twitter Bookmarks)

PageSummaryDate
Claude Code Is the Inflection PointSemiAnalysis: Claude Code as paradigm shift in software development2026-04-29
Vmax — Reinforcement LearningAutomated RL environment design company2026-04-29
Rubrics as RewardsRL beyond verifiable domains using structured rubrics2026-04-29
OpenRubrics: Synthetic Rubric GenerationScalable synthetic rubrics for reward modeling2026-04-29
The AI-Native Interview — SierraSierra’s Plan→Build→Review interview redesign for AI era2026-04-29
Advancing Search-Augmented LMsPerplexity’s approach to search-augmented generation2026-04-29
Amanda Askell on AI ConsciousnessAI consciousness, Claude’s character, Silicon Valley’s fears2026-04-29
Learning Next Action PredictorsPredicting user actions from human-computer interaction2026-04-29
What Will Be Scarce?What remains scarce in an AI-abundant world2026-04-29
Shopify: AI as Baseline ExpectationTobi Lütke: reflexive AI usage now required2026-04-29
Shopify Fine-Tuned Qwen3-32BFine-tuned from merchant automation data for tool-calling2026-04-29
ICLR 2026: Scaling RL at 1T+ ParametersRishabh Agarwal on frontier RL scaling2026-04-29
DeepSeek V4: Addresses HBM ShortageArchitectural innovation for HBM constraints2026-04-29
”Global Ban on MoEs”Arthur Zucker’s contrarian anti-MoE position2026-04-29
KV Cache Implementation NightmaresComplexity of efficient KV Cache implementation2026-04-29
BlackwellSpecialized compute per inference stage2026-04-29
CUDA Matmul Beats cuBLAS on B200Hand-tuned kernel beats cuBLAS by 6%2026-04-29
Karpathy: 1B Clean Data = 1.5T ModelsData quality can compensate for 1000x fewer parameters2026-04-29
Transformer Circuits Mental ModelBest mental model from Ant’s circuits paper2026-04-29
AI Semiconductor Endgame 2026Token economics shifting from GPU compute to HBM2026-04-29
Vibe-Coded CUDA Inference EngineEric Zhang: CUDA inference engine via AI coding2026-04-29
Future Interface: Three LayersAmbient intent, memory, execution framework2026-04-29
Predictive User ModelLearning predictive models from user interaction2026-04-29
Path Forward for AI StartupsPost-Cursor IPO startup landscape analysis2026-04-29
Stanford: Datacenter EconomicsDeep dive on datacenter cost structures2026-04-29
xAI → Periodic Labs: RL for AtomsRL expanding into materials science2026-04-29
100k H100 Datacenter NumbersBallpark numbers for mental math2026-04-29
Business of Fine-TuningConsolidation in LLM personalization market2026-04-29
Intelligence per PicojouleMatX: hardware-software codesign for efficiency2026-04-29
Textual Steering Vectors for Multimodal LLMsCross-modal steering vectors improve visual understanding2026-04-29
ImplicitRM: Unbiased Reward ModelingReward modeling from implicit behavioral preferences2026-04-29
HW-SW Codesign — MatXClive Chan, Dylan Patel, Reiner Pope on inference efficiency2026-04-29
How Attention Residuals WorkVisual explanation of transformer residual stream2026-04-29
DeepSeek Before V4 (LatePost)DeepSeek organization and Liang Wenfeng’s goals2026-04-29
Shannon vs KolmogorovInformation theory vs algorithmic complexity for LLMs2026-04-29
Keith Rabois: Hard Truths in AI EraContrarian takes on AI startups2026-04-29
”Don’t Teach. Incentivize.” — OpenAIHyung Won Chung on reward-based training philosophy2026-04-29

6 items under this folder.