Updated 09/27/2026
All links
Grouped by month.
109 items
October 2026
September 2026
Opus code-based launch-video promptcontinuous-take launch video prompt with beat-synced motion Opens in a new tabOpus motion-graphics workflowreference-led workflow for polished code-generated motion graphics Opens in a new tabCUDA from zero to heroCUDA notes from GPU fundamentals to kernel optimization Opens in a new tabApple-style product video frameworkbrand, design, and motion rules for premium product videos Opens in a new tabHyperFrames launch-video promptprompt idea for a polished launch video Opens in a new tabGLiNER 2.5 for information extractionan open-source model for structured information extraction Opens in a new tabbev-decision-150K dataset150K examples for training decision models across many domains Opens in a new tabManim, Kokoro, and FFmpeg for AI videoManim visuals, Kokoro voice, and FFmpeg assembly Opens in a new tabBuilding a decision-model training datasetadapting structured examples into decision-model training data Opens in a new tabContrastive Language Model for fast decisionscontrastive state and action scoring for fast agent decisions Opens in a new tabHow Claude.ai became 3x fasterusing Claude to measure, debug, and improve web performance Opens in a new tabHyperFrames 3D camera and captions skill3D camera movement and caption effects for HyperFrames Opens in a new tab3D camera and text effects with HyperFrames3D camera movement paired with animated text Opens in a new tabBuilding with Jev: use cases and implementation architecturesHow Instinct's memory worksreverse-engineering a git-backed memory system for a messaging agent Opens in a new tabJev as a judge for agent evalstesting a typed decision model as a fast, consistent agent evaluator Opens in a new tabInstant generative UI with json-render and Jevjson-render and Jev turn typed decisions into instant UI Opens in a new tabHand-drawn canvas animation skillJS drawing skill for hand-drawn canvas animation Opens in a new tabNeedle 3 local automation modeltiny local models for tool calls and structured extraction Opens in a new tabScaling reinforcement learning for MiMo-V2.6scaling compute, environments, harnesses, and graders in one RL run Opens in a new tabArrow 2 for editable vector graphicsan AI model for precise, editable vector graphics Opens in a new tabCobbleDB built with persistent agentsCobbleDB built by two engineers with persistent AI agents Opens in a new tabHow Perplexity built CobbleDBthe research behind Perplexity's agent-built key-value store Opens in a new tabJev and reinforcement learning from decisionsa fast composable model trained with reinforcement learning from decisions Opens in a new tabZGateway for ZippyDBa stateless proxy for controlling database connection fanout Opens in a new tabTransformer basics: prefill, decode, and the KV cacheJev and System One modelsMulti-agent systems scale thought in parallelPC-ALM: local learning without backpropagationlayer local credit assignment with predictive coding and dual neurons Opens in a new tabScaling sandboxes for SWE-2 trainingsandbox infrastructure for large batches of coding agent rollouts Opens in a new tabRecurrent Looped Transformerrecurrent transformer loops for deeper iterative reasoning Opens in a new tabCognition SWE-2a lower cost coding model trained with large scale reinforcement learning Opens in a new tabCursor Projectsone persistent coordinator thread with proactive subagents Opens in a new tabRunning DeepSeek V4.1 Flash on an M5 MaxDeepSeek V4.1 Flash on an M5 Max with SSD streaming Opens in a new tabDeepSeek V4.1 Flash architecture notesencoder-decoder design, sparse attention, Engram, mHC, native vision Opens in a new tabDeepSeek V4.1 Flash on everyday design tasksdesign-task benchmark across open and closed models Opens in a new tabStoryboarding with ChatGPT Image and Seedancestoryboards, cast design, film scenes Opens in a new tabHow Cursor's agent harness shed its rolesfour harness designs, fewer central roles, one handoff Opens in a new tabWhy Jacob Coxon left Anthropica researcher's warning about self-improving AI Opens in a new tabTowards self-driving codebasesrecursive planning for coding agents at scale Opens in a new tabSpeculative decoding, from first ideas to current methodshow speculative decoding evolved and stays lossless Opens in a new tabBuilding a 3D game with Astra3D game creation with Astra Opens in a new tabDurable execution and atomic execution
August 2026
Speculative decoding study notesillustrated notes on speculative decoding Opens in a new tabSpeculative decodingConsistent hashingHow Lauren ships with cloud agentspstack, background context, cloud-agent routines Opens in a new tabCloud agents that hold the goalevent-driven agents for long-running work Opens in a new tabGit at any scaleGit storage treated as database infrastructure Opens in a new tabRebuilding Linear's delta-sync read pathdelta sync, predictable latency, turbopuffer Opens in a new tab70 durable moatsa broader vocabulary for defensibility Opens in a new tabScaling self-verificationuse an LLM as a judge so a smaller model can outperform a frontier model Opens in a new tabDeepSeek Harness design notesSDK harnesses, programmatic tools, cache-aware design Opens in a new tabDeepSeek Harnessmodels, tools, skills, loops, and UI as plugins Opens in a new tabIntro to Grok Botalways-on computer, persistent sessions, human handoff Opens in a new tabUnsloth Desktoplocal models, training, tool calls, model export Opens in a new tabExtracting hidden model reasoningAPI vulnerability, hidden reasoning, thinking tokens Opens in a new tabPrime-Agent walkthrougha compact implementation tour Opens in a new tabLearn ChatGPT Workmainstream agent workloads with the GPT Work app Opens in a new tabGraph engineering for coding agentsOpenAI engineering talk on managing coding agents Opens in a new tabRLMs from first principles50-minute explanation plus implementation Opens in a new tabWhat to keep in AGENTS.mdagent skill and AGENTS.md patterns to explore Opens in a new tabTraining multi-agent systems with PRIME-RLstartup idea, and I want to try this Opens in a new tabChatGPT Work in 61 minutesmainstream agent workloads with the GPT Work app Opens in a new tabWhat is an RLM?start with the basic question Opens in a new tabWhat the Hugging Face agent incident revealedmultiple runs, revoked credentials, unexpected coordination Opens in a new tabThe OpenAI agent incident talkincident details, emergent coordination, misalignment Opens in a new tabA human and AI workspacestartup idea Opens in a new tabEvery company needs a company brainstartup idea Opens in a new tabThe supermemory company braincompany knowledge, cross-tool action, permissions Opens in a new tabOpenWork Webstartup idea inspiration Opens in a new tabDesign is everythingI like this cabinet app idea Opens in a new tabThe multiplayer agent harnessthe product category is the multiplayer harness Opens in a new tabClaude Opus 5proactive model, frontier capability, lower price Opens in a new tabGraph engineering with Claudebranching, parallelism, verification, convergence Opens in a new tabAgent skill to exploreagent skill to explore Opens in a new tabLLM-as-a-Verifierfine-grained scores, repeated evaluation, richer feedback Opens in a new tabA Field Guide to Fableto watch Opens in a new tabBespoke Labs on reliable agentsstartup to research Opens in a new tabMaking the Claude cartoon with GlifGlif tutorial, mobile-first forkable mind-map idea Opens in a new tabRL Town: agents as a tiny civilizationcoding agents, RL Town, interesting to try locally Opens in a new tabA social CLI for agentsstartup idea to track Opens in a new tabA Claude cartoon made with Glifvideo creation with AI, made with Glif Opens in a new tabBuilding software from an iPhonecoding agents, mobile workflows, read list Opens in a new tabThe unreasonable effectiveness of SVG animationSVG animation as the next layer of technical visuals Opens in a new tabGetting more out of Codexdurable threads, tools, automations, artifacts Opens in a new tabpstackrigorous workflows, verification, parallel agentsWeb designtaste, systems, responsive UX, browser QAFrontend developmentpreserve the stack, test behavior, verify in browserDotted painterly cover artphysical metaphors, square-cell texture, crop QASpider-Man: Brand New Daywatched in theaters, 95/100The Odysseywatched in 70mm IMAX, 85/100Floguo notessplit panes, compact rows, quiet motion Opens in a new tabCompound Engineering: My Self-Improving Remote Agent SetupHumanoid components guidemotors, batteries, reducers, sensors, costs Opens in a new tabThe actuator bottleneckactuators, cost, reliability, control Opens in a new tabFast simulator integrationintegrated in a day, immediate feedback Opens in a new tabAutomated UI testingbrowser QA, visual checks Opens in a new tabYC on an AI-native companyAI-native from day one, not AI added later Opens in a new tabKarpathy on AI architecturetwo-hour overview, models, systems, architecture Opens in a new tabHow Text Embedding Models Are Trained and Served