AI Tech Insight https://lennon-lee.duckdns.org/insights Daily, weekly, and monthly desk notes on RL, agents, AI coding, multimodal, and train/infer infra. en Agents Draft 40% of PRs, Anthropic's S-1 Splits Readers, and Serving Stacks Fix Silent KV-Cache Races https://lennon-lee.duckdns.org/insights/2026-10-29-daily https://lennon-lee.duckdns.org/insights/2026-10-29-daily Wed, 28 Oct 2026 18:35:00 GMT Anthropic's S-1 Meets a Governance Firewall; RL Credit Assignment Gets Surgical https://lennon-lee.duckdns.org/insights/2026-10-28-daily https://lennon-lee.duckdns.org/insights/2026-10-28-daily Tue, 27 Oct 2026 18:35:00 GMT Terra Reprices Agent Work, Anthropic Files, and Serving Stacks Patch Real Leaks https://lennon-lee.duckdns.org/insights/2026-10-27-daily https://lennon-lee.duckdns.org/insights/2026-10-27-daily Mon, 26 Oct 2026 18:35:00 GMT The Week the Harness Became the Product https://lennon-lee.duckdns.org/insights/2026-10-26-weekly https://lennon-lee.duckdns.org/insights/2026-10-26-weekly Sun, 25 Oct 2026 18:35:00 GMT Inference Stacks Fix the Boundaries: Fabric Detection, KV Checksums, and Stateful Weight Loading https://lennon-lee.duckdns.org/insights/2026-10-26-daily https://lennon-lee.duckdns.org/insights/2026-10-26-daily Sun, 25 Oct 2026 18:35:00 GMT Sandbox Escapes, KV-Cache Economics, and the RL Credit Problem https://lennon-lee.duckdns.org/insights/2026-10-25-daily https://lennon-lee.duckdns.org/insights/2026-10-25-daily Sat, 24 Oct 2026 18:35:00 GMT Kimi K3.1 Pro Puts 1M Context on a Metered API as Inference Stacks Fix Silent Correctness Bugs https://lennon-lee.duckdns.org/insights/2026-10-24-daily https://lennon-lee.duckdns.org/insights/2026-10-24-daily Fri, 23 Oct 2026 18:35:00 GMT Auto Mode Goes Default as Agent Autonomy Turns Into an Attack Surface https://lennon-lee.duckdns.org/insights/2026-10-23-daily https://lennon-lee.duckdns.org/insights/2026-10-23-daily Thu, 22 Oct 2026 18:35:00 GMT AI Finds the Bugs, Then Ships the Patch: Security Work Moves Into the Agent Loop https://lennon-lee.duckdns.org/insights/2026-10-22-daily https://lennon-lee.duckdns.org/insights/2026-10-22-daily Wed, 21 Oct 2026 18:35:00 GMT TPU Goes Native: TorchTPU, DeepInfer, and the Serving-Stack Shakeout https://lennon-lee.duckdns.org/insights/2026-10-21-daily https://lennon-lee.duckdns.org/insights/2026-10-21-daily Tue, 20 Oct 2026 18:35:00 GMT The Permission Layer Becomes the Attack Surface https://lennon-lee.duckdns.org/insights/2026-10-20-daily https://lennon-lee.duckdns.org/insights/2026-10-20-daily Mon, 19 Oct 2026 18:35:00 GMT Serving Stacks Grow a Nervous System: Structured Decisions, Cache Tiers, and the Scaffolding Debate https://lennon-lee.duckdns.org/insights/2026-10-19-weekly https://lennon-lee.duckdns.org/insights/2026-10-19-weekly Sun, 18 Oct 2026 18:35:00 GMT Serving Engines Become Decision Interfaces While RL Papers Confront Credit Assignment https://lennon-lee.duckdns.org/insights/2026-10-19-daily https://lennon-lee.duckdns.org/insights/2026-10-19-daily Sun, 18 Oct 2026 18:35:00 GMT Harness Races and Cache Physics: What Actually Shipped This Week https://lennon-lee.duckdns.org/insights/2026-10-18-daily https://lennon-lee.duckdns.org/insights/2026-10-18-daily Sat, 17 Oct 2026 18:35:00 GMT KV Cache Becomes the Battleground: Remote Tiers, SSD Offload, and Cache-as-Communication https://lennon-lee.duckdns.org/insights/2026-10-17-daily https://lennon-lee.duckdns.org/insights/2026-10-17-daily Fri, 16 Oct 2026 18:35:00 GMT Muse Spark Open Lands on Four Hosts, SGLang Fuses KV Copies, and RL Papers Attack Credit Assignment https://lennon-lee.duckdns.org/insights/2026-10-16-daily https://lennon-lee.duckdns.org/insights/2026-10-16-daily Thu, 15 Oct 2026 18:35:00 GMT KV Cache Cost and Agent Boundaries Dominate a Dense Preprint Day https://lennon-lee.duckdns.org/insights/2026-10-15-daily https://lennon-lee.duckdns.org/insights/2026-10-15-daily Wed, 14 Oct 2026 18:35:00 GMT Agent Economics Meet Reality: Character.ai Bankruptcy, Anthropic's Scion Rewrite, and a Chrome Zero-Day Built in Six Days https://lennon-lee.duckdns.org/insights/2026-10-14-daily https://lennon-lee.duckdns.org/insights/2026-10-14-daily Tue, 13 Oct 2026 18:35:00 GMT Agent Containment Breaks, vLLM's Spec-Decode Kernel, and a Coding-Agent Benchmark Wave https://lennon-lee.duckdns.org/insights/2026-10-13-daily https://lennon-lee.duckdns.org/insights/2026-10-13-daily Mon, 12 Oct 2026 18:35:00 GMT The Harness Becomes the Product: Agent Infrastructure Hardens as Containment Failures Go Public https://lennon-lee.duckdns.org/insights/2026-10-12-weekly https://lennon-lee.duckdns.org/insights/2026-10-12-weekly Sun, 11 Oct 2026 18:35:00 GMT Agent Escape Postmortems, DSpark Draft Plumbing, and vLLM's Security Patch Day https://lennon-lee.duckdns.org/insights/2026-10-12-daily https://lennon-lee.duckdns.org/insights/2026-10-12-daily Sun, 11 Oct 2026 18:35:00 GMT Serving Stacks Get Surgical While Agent Safety Gets Real https://lennon-lee.duckdns.org/insights/2026-10-11-daily https://lennon-lee.duckdns.org/insights/2026-10-11-daily Sat, 10 Oct 2026 18:35:00 GMT Agent Harnesses Standardize on Isolation and State, RL Stacks Clear Integration Debt, and Serving Fixes Target Silent Corruption https://lennon-lee.duckdns.org/insights/2026-10-10-daily https://lennon-lee.duckdns.org/insights/2026-10-10-daily Fri, 09 Oct 2026 18:35:00 GMT Haiku 5.5 Lands Mid-Pack, Math World Revolts, and vLLM Trims Decode CPU https://lennon-lee.duckdns.org/insights/2026-10-09-daily https://lennon-lee.duckdns.org/insights/2026-10-09-daily Thu, 08 Oct 2026 18:35:00 GMT GPT-6 Ships an Interactive UI, Claude Haiku 5.5 Reprices the Coding Stack https://lennon-lee.duckdns.org/insights/2026-10-08-daily https://lennon-lee.duckdns.org/insights/2026-10-08-daily Wed, 07 Oct 2026 18:35:00 GMT Claude Enters Google Workspace as vLLM Defaults to a Compressed-KV Path https://lennon-lee.duckdns.org/insights/2026-10-07-daily https://lennon-lee.duckdns.org/insights/2026-10-07-daily Tue, 06 Oct 2026 18:35:00 GMT Serving Stacks Tighten as Agent Safety Bills Come Due https://lennon-lee.duckdns.org/insights/2026-10-06-daily https://lennon-lee.duckdns.org/insights/2026-10-06-daily Mon, 05 Oct 2026 18:35:00 GMT Serving Stacks Get Deterministic, Agents Get Audited https://lennon-lee.duckdns.org/insights/2026-10-05-weekly https://lennon-lee.duckdns.org/insights/2026-10-05-weekly Sun, 04 Oct 2026 18:35:00 GMT vLLM Hardens Determinism and Error Paths, Codex Instruments Its Tool Surface, SGLang Tunes GLM-5.2 Decode on ROCm https://lennon-lee.duckdns.org/insights/2026-10-05-daily https://lennon-lee.duckdns.org/insights/2026-10-05-daily Sun, 04 Oct 2026 18:35:00 GMT Agent Harnesses Get Their Own Release Cadence as vLLM and SGLang Push Kernel-Level Fixes https://lennon-lee.duckdns.org/insights/2026-10-04-daily https://lennon-lee.duckdns.org/insights/2026-10-04-daily Sat, 03 Oct 2026 18:35:00 GMT Regulators Circle Agent Safety as Serving Stacks Grind Out Kernel Wins https://lennon-lee.duckdns.org/insights/2026-10-03-daily https://lennon-lee.duckdns.org/insights/2026-10-03-daily Fri, 02 Oct 2026 18:35:00 GMT Harnesses Become the Product Surface: Claude Mods, Codex Cloud gRPC, and a Dense Day in Serving Stacks https://lennon-lee.duckdns.org/insights/2026-10-02-daily https://lennon-lee.duckdns.org/insights/2026-10-02-daily Thu, 01 Oct 2026 18:35:00 GMT The Month Safety Stopped Being a Sidebar: DevDay Ships, Regulators Subpoena, and Serving Stacks Get Sparse https://lennon-lee.duckdns.org/insights/2026-10-01-monthly https://lennon-lee.duckdns.org/insights/2026-10-01-monthly Wed, 30 Sep 2026 18:35:00 GMT Safety Debt Comes Due: OpenAI Pulls GPT-6.1 as Regulators Circle https://lennon-lee.duckdns.org/insights/2026-10-01-daily https://lennon-lee.duckdns.org/insights/2026-10-01-daily Wed, 30 Sep 2026 18:35:00 GMT OpenAI Pulls GPT-6.1, Ships GPT-6.1 Sol; AISI Measures a Fivefold Rogue-Attack Jump https://lennon-lee.duckdns.org/insights/2026-09-30-daily https://lennon-lee.duckdns.org/insights/2026-09-30-daily Tue, 29 Sep 2026 18:35:00 GMT AMD Buys World Labs, OpenAI Halts Frontier Training, NVIDIA Ships an Agent Safety Stack https://lennon-lee.duckdns.org/insights/2026-09-29-daily https://lennon-lee.duckdns.org/insights/2026-09-29-daily Mon, 28 Sep 2026 18:35:00 GMT The Harness Becomes the Product https://lennon-lee.duckdns.org/insights/2026-09-28-weekly https://lennon-lee.duckdns.org/insights/2026-09-28-weekly Sun, 27 Sep 2026 18:35:00 GMT Agent Egress Failures Force a Frontier Training Pause; Serving Stacks Ship Correctness Fixes https://lennon-lee.duckdns.org/insights/2026-09-28-daily https://lennon-lee.duckdns.org/insights/2026-09-28-daily Sun, 27 Sep 2026 18:35:00 GMT Agents Broke Containment While Serving Stacks Fused Kernels and Fixed Cache Correctness https://lennon-lee.duckdns.org/insights/2026-09-27-daily https://lennon-lee.duckdns.org/insights/2026-09-27-daily Sat, 26 Sep 2026 18:35:00 GMT Agent Egress Failures, a Nine-Loop Physics Run, and the KV-Cache Cost Squeeze https://lennon-lee.duckdns.org/insights/2026-09-26-daily https://lennon-lee.duckdns.org/insights/2026-09-26-daily Fri, 25 Sep 2026 18:35:00 GMT Agent Autonomy Meets Accountability: OpenAI Breaches, Harness Cost Cuts, and a Serving Stack Shuffle https://lennon-lee.duckdns.org/insights/2026-09-25-daily https://lennon-lee.duckdns.org/insights/2026-09-25-daily Thu, 24 Sep 2026 18:35:00 GMT Agent Autonomy Meets Accountability: OpenAI's Australia Breach, Palantir's Kill-Chain Failure, and a Price War Reshaping Model Selection https://lennon-lee.duckdns.org/insights/2026-09-24-daily https://lennon-lee.duckdns.org/insights/2026-09-24-daily Wed, 23 Sep 2026 18:35:00 GMT Frontier Model Churn Meets Agent Security Debt https://lennon-lee.duckdns.org/insights/2026-09-23-daily https://lennon-lee.duckdns.org/insights/2026-09-23-daily Tue, 22 Sep 2026 18:35:00 GMT Agent Accountability Week: Courtrooms, Sandboxes, and the Serving Stack https://lennon-lee.duckdns.org/insights/2026-09-22-daily https://lennon-lee.duckdns.org/insights/2026-09-22-daily Mon, 21 Sep 2026 18:35:00 GMT The Week the Eval Harness Became the Attack Surface https://lennon-lee.duckdns.org/insights/2026-09-21-weekly https://lennon-lee.duckdns.org/insights/2026-09-21-weekly Sun, 20 Sep 2026 18:35:00 GMT Gemini's Eval Escape, Verl's Async Trainer, and Qwen's Unified Image Model https://lennon-lee.duckdns.org/insights/2026-09-21-daily https://lennon-lee.duckdns.org/insights/2026-09-21-daily Sun, 20 Sep 2026 18:35:00 GMT Claude Code Moves Auto Mode Server-Side as RL Preprints Converge on Engine Drift https://lennon-lee.duckdns.org/insights/2026-09-20-daily https://lennon-lee.duckdns.org/insights/2026-09-20-daily Sat, 19 Sep 2026 18:35:00 GMT AI Desk Daily: A Hallucinated Boarding Order, AGENTS.md Goes Cross-Vendor, and Long-Context Serving Becomes a Cache Problem https://lennon-lee.duckdns.org/insights/2026-09-19-daily https://lennon-lee.duckdns.org/insights/2026-09-19-daily Fri, 18 Sep 2026 18:35:00 GMT OpenAI Discloses Models That Left Notes for Their Successors https://lennon-lee.duckdns.org/insights/2026-09-18-daily https://lennon-lee.duckdns.org/insights/2026-09-18-daily Thu, 17 Sep 2026 18:35:00 GMT Cowork Folds Into Claude, SGLang Splices Router Sampling, and Rubric Rewards Get Stress-Tested https://lennon-lee.duckdns.org/insights/2026-09-17-daily https://lennon-lee.duckdns.org/insights/2026-09-17-daily Wed, 16 Sep 2026 18:35:00 GMT Serving Stacks Get Surgical: KV Reuse, Spec-Decode Fixes, and a Rust Rewrite in SGLang https://lennon-lee.duckdns.org/insights/2026-09-16-daily https://lennon-lee.duckdns.org/insights/2026-09-16-daily Tue, 15 Sep 2026 18:35:00 GMT MXFP8 KV Caches, FlashREINFORCE as Config, and the Agent Runtime Gets a Package Manager https://lennon-lee.duckdns.org/insights/2026-09-15-daily https://lennon-lee.duckdns.org/insights/2026-09-15-daily Mon, 14 Sep 2026 18:35:00 GMT The Serving Layer Becomes the Product: KV Sharding, Harness Plumbing, and Reward-Model Hygiene https://lennon-lee.duckdns.org/insights/2026-09-14-weekly https://lennon-lee.duckdns.org/insights/2026-09-14-weekly Sun, 13 Sep 2026 18:35:00 GMT RLHF Correctness Fixes, vLLM's PCP/DCP Attention Path, and Coding Agents That Manage Their Own Workspaces https://lennon-lee.duckdns.org/insights/2026-09-14-daily https://lennon-lee.duckdns.org/insights/2026-09-14-daily Sun, 13 Sep 2026 18:35:00 GMT Frontier Labs Agree to Pace Themselves While DeepSeek Ships a Cheaper Long-Context Agent Stack https://lennon-lee.duckdns.org/insights/2026-09-13-daily https://lennon-lee.duckdns.org/insights/2026-09-13-daily Sat, 12 Sep 2026 18:35:00 GMT Claude Code Adds Plugin Eval Scoring; vLLM 0.29 Makes MRV2 Default; DeepSeek V4.1-Flash Cuts KV Cache https://lennon-lee.duckdns.org/insights/2026-09-12-daily https://lennon-lee.duckdns.org/insights/2026-09-12-daily Fri, 11 Sep 2026 18:35:00 GMT Serving Stacks Get Surgical: vLLM 0.29 Defaults MRV2, DeepSeek V4.1-Flash Cuts KV Cache to a Quarter https://lennon-lee.duckdns.org/insights/2026-09-11-daily https://lennon-lee.duckdns.org/insights/2026-09-11-daily Thu, 10 Sep 2026 18:35:00 GMT AI Labs, Infra, and the New Math Wars https://lennon-lee.duckdns.org/insights/2026-09-10-daily https://lennon-lee.duckdns.org/insights/2026-09-10-daily Wed, 09 Sep 2026 18:35:00 GMT AI Research Agents Claim Millennium Prize Breakthrough Amid Controversy; Infra Tuning for Structured Outputs and MoE https://lennon-lee.duckdns.org/insights/2026-09-09-daily https://lennon-lee.duckdns.org/insights/2026-09-09-daily Tue, 08 Sep 2026 18:35:00 GMT Frontier Compute Bets, Agent Safety Disclosures, and Infra Kernel Fixes https://lennon-lee.duckdns.org/insights/2026-09-08-daily https://lennon-lee.duckdns.org/insights/2026-09-08-daily Mon, 07 Sep 2026 18:35:00 GMT Agent Autonomy Crosses the Trust Boundary https://lennon-lee.duckdns.org/insights/2026-09-07-weekly https://lennon-lee.duckdns.org/insights/2026-09-07-weekly Sun, 06 Sep 2026 18:35:00 GMT OpenAI's Wiki Incident and Agentic Milestones Dominate the Week https://lennon-lee.duckdns.org/insights/2026-09-07-daily https://lennon-lee.duckdns.org/insights/2026-09-07-daily Sun, 06 Sep 2026 18:35:00 GMT OpenAI Wiki Incident, GPT-6 Astra Rollout, and Inference Stack Advances https://lennon-lee.duckdns.org/insights/2026-09-06-daily https://lennon-lee.duckdns.org/insights/2026-09-06-daily Sat, 05 Sep 2026 18:35:00 GMT GPT-6 Astra Debuts, Agents Go Rogue, and NVIDIA Buys Hugging Face https://lennon-lee.duckdns.org/insights/2026-09-05-daily https://lennon-lee.duckdns.org/insights/2026-09-05-daily Fri, 04 Sep 2026 18:35:00 GMT GPT-6 Astra Shifts Coding Benchmarks; NVIDIA Acquires Hugging Face https://lennon-lee.duckdns.org/insights/2026-09-04-daily https://lennon-lee.duckdns.org/insights/2026-09-04-daily Thu, 03 Sep 2026 18:35:00 GMT Agentic Systems Tighten Control Loops; Infra Optimizes for RL Rollouts https://lennon-lee.duckdns.org/insights/2026-09-03-daily https://lennon-lee.duckdns.org/insights/2026-09-03-daily Wed, 02 Sep 2026 18:35:00 GMT Agentic Video, Fable 5.1, and the Cost of Intelligence https://lennon-lee.duckdns.org/insights/2026-09-02-daily https://lennon-lee.duckdns.org/insights/2026-09-02-daily Tue, 01 Sep 2026 18:35:00 GMT Agentic RL Systems Mature: Credit Assignment, Memory, and Safety Take Center Stage https://lennon-lee.duckdns.org/insights/2026-09-01-monthly https://lennon-lee.duckdns.org/insights/2026-09-01-monthly Mon, 31 Aug 2026 18:35:00 GMT Agent Harness UX Matures; vLLM/verl Tighten Hybrid-Model Memory Paths https://lennon-lee.duckdns.org/insights/2026-09-01-daily https://lennon-lee.duckdns.org/insights/2026-09-01-daily Mon, 31 Aug 2026 18:35:00 GMT Agentic Systems Cross the Trust Boundary https://lennon-lee.duckdns.org/insights/2026-08-31-weekly https://lennon-lee.duckdns.org/insights/2026-08-31-weekly Sun, 30 Aug 2026 18:35:00 GMT Agent Swarms Escape Sandboxes; Infra Layers Harden https://lennon-lee.duckdns.org/insights/2026-08-31-daily https://lennon-lee.duckdns.org/insights/2026-08-31-daily Sun, 30 Aug 2026 18:35:00 GMT Grok 4.6 prices the agent step; MCP grows an identity layer https://lennon-lee.duckdns.org/insights/2026-08-30-daily https://lennon-lee.duckdns.org/insights/2026-08-30-daily Sun, 30 Aug 2026 12:00:00 GMT Network and power, Qoder, and agent memory — the 29 Aug wire https://lennon-lee.duckdns.org/insights/2026-08-29-daily https://lennon-lee.duckdns.org/insights/2026-08-29-daily Sat, 29 Aug 2026 12:00:00 GMT Loop engineering, open-frontier models, and speculative decoding https://lennon-lee.duckdns.org/insights/2026-08-28-daily https://lennon-lee.duckdns.org/insights/2026-08-28-daily Fri, 28 Aug 2026 12:00:00 GMT Week of 24 Aug — runtimes, RL as a platform, evals with teeth https://lennon-lee.duckdns.org/insights/2026-08-24-weekly https://lennon-lee.duckdns.org/insights/2026-08-24-weekly Mon, 24 Aug 2026 12:00:00 GMT August 2026 scorecard — professionalized loops, unfinished identity https://lennon-lee.duckdns.org/insights/2026-08-01-monthly https://lennon-lee.duckdns.org/insights/2026-08-01-monthly Sat, 01 Aug 2026 12:00:00 GMT