Hacker NewsKarpathy MicroGPT Full GPT Training in 200 Lines 994 PointsKarpathy·high signalXBlueskyLinkedInCopy linkComplete GPT training and inference in 200 lines Python. 4,192 parameters. Landmark educational resource at nearly 1000 HN points.SourceSource pageKarpathy↳ Follow the threadPolicy dependency / Stack layerEnvs-FORGE Solves a Per-Seed MILP to Synthesize Agent-RL Environments, Hitting 77.1% on SWE-bench Verified vs 73.4% BasearXiv 2608.14312Stack layer / ContrastMOSS-VL Ships an 11.3B Open-Weight VLM That Perceives While It Speaks — 66.0 vs 37.5 on OmniMMI Proactive Alerting and 5.1x Faster TTFT Than Qwen3-VL-8BarXiv (via HuggingFace Daily Papers)Stack layer / Contrast'How Do Agents Fail on AutoResearch': 800 Trajectories Across 8 Harness-Model Combos Find One Failure Every Model Shares — They Never Check Output Against EvidencearXiv (via HuggingFace Daily Papers)Stack layer / ContrastQwen3.8 27B Hits 50.44 tok/s at Full 256K Context on a Single 24GB Blackwell Card — With 515 MiB of VRAM Headroom Leftpiszczek.pl / Hacker NewsStack layer / ContrastShow HN: MoEspresso Squeezes DeepSeek V4 Flash Coder From 84GB to 56.8GB by Deleting ~80B Parameters, Then Writes a C Compiler on a MacHugging Face / Hacker NewsStack layer / Update threadTrain the harness, not just the model: 6K examples of harnessed agentic RL lifts Qwen3.5-9B on SWE-bench Verified from 41.8% to 56.4%arXiv 2608.17528Stack layer / Threat patternHow you lay out your repo changes prompt-injection success: highly modular workspaces measurably lower attack success ratearXiv 2608.14876Stack layer / Follow-up threadGruber Calls Claude's SynthID Watermark a 'Perversion of Writing' — 348 HN Comments, and the Unrebutted Objection Is the Detection Oracle, Not the ProseDaring Fireball / Anthropic / Hacker News