20 topics across 42 posts. The 19 with real depth come first.
Everything written about llm.
Anthropic called the settlement proof that training on books is fair use. The settlement covers something narrower: how the books were obtained. The court also did something unexpected, which is why the "largest copyright recovery in history" quote ran backwards. Two months of coverage missed the actual holding.
Epoch AI reviewed 15 major AI benchmarks and flagged 9 as flawed, including Humanity's Last Exam, where a hand-audited sample of 48 questions found 22 defective. I took their findings and ran them: filtering the broken items out of a live leaderboard, and the scores barely move. Both findings are true and they are not in tension.
I ran the first head-to-head of the two open "System One" decision models, Laya and Jev, on 751 byte-identical inputs: same states, same questions, same seed, question hashes verified before comparing. Jev wins most suites for $0.008. Laya wins where it counts for a free model. A fourth column with Qwen 1.5B under constrained decoding completed the picture.
IFM released six foundation models from 0.9B to 375B, plus roughly 21.5TB of the actual training datasets, all Apache 2.0. I verified the dataset sizes and benchmarked the three small models myself: the 0.9B does real reasoning at 19 tokens per second in 2.2GB of VRAM. Here is what shipped, what is still missing, and where the catches are.
I stress-tested Ternary Bonsai 2 27B on two Tesla T4s: 15 tokens per second on a single weak GPU, the full 262k context loading in 26GB of VRAM, and a complete todo webapp built and self-tested inside a local coding agent. The claims mostly check out, with a few catches.
Cloudflare Workers AI serves 50+ models from its edge network with 10,000 free Neurons a day. Since July 2026 the frontier coding models need the $5/mo Paid plan, but the free tier still covers real daily coding through OpenCode. What is free, what moved to Paid, and how I wire it up.
Two stories keep getting merged into one: the Bartz v. Anthropic copyright case ($1.5B settlement) and Anthropic's claim that Moonshot AI routed 300,000 customer requests through Claude to harvest responses. The public record supports a narrower conclusion than either side's loudest claims.
Real Qwen3.8-27B numbers on the DGX Spark GB10: the memory wall caps single-stream decode at 11-12 tok/s, and speculative decoding is the only way out. SGLang with DSpark hits 34 tok/s, veloGB10 reaches about 40 tok/s, and NVFP4 keeps quality within noise of FP8.
I ranked nine free GPU notebook platforms against Google Colab in 2026. Kaggle (Tesla P100, 30 hrs/week, 29GB RAM) and AWS SageMaker Studio Lab (T4, 4 GPU hrs/day, persistent storage) take S tier, Paperspace Gradient is A tier with persistent storage, and Colab itself lands B tier for vague limits and 2-3 hour disconnects.
VRAM is everything for local AI agents: the 4-bit cheat sheet runs 7B at about 5GB, 14B at 10GB, 30B at 20GB, and 70B at 40GB. Three builds cover it, from a $1,200-1,500 RTX 4060 Ti 16GB starter to the used 24GB-card sweet spot that runs 32B models at near-cloud quality for daily work.
How Heretic removes refusal behavior from open-weight LLMs: my evaluation philosophy, Optuna TPE search learnings, the ARA next-generation ablation, why single constant vectors fail on MoE models, and what publishing 34 abliterated models taught me about attention-versus-MLP placement across Qwen and Gemma families.
Claude Code ranked the best local AI model for each VRAM tier from 4GB to 384GB - Nanbeige 3B at 4GB to Motif-3 314B at 384GB. Fact-checked July 2026: all five picks are real, available, and the 5.9GB Bonsai 27B even runs on an iPhone.
Why about 1B parameters is the sweet spot for agent swarms: MiniCPM5-1B and six rival small models profiled from roughly 200 t/s on a $0.01/hr T4 down to 5 t/s on a Raspberry Pi 5, with the swarm economics at 1K, 100K, and 10M agents per day.
The memory wall ends here: NVIDIA's DGX Station pairs 252GB of HBM3e at 7.1TB/s with 496GB of LPDDR5X into one 748GB coherent pool on GB300 Grace Blackwell Ultra, adds NVFP4 quantization and 800Gbps ConnectX-8 networking (two boxes pool about 1.5TB), at roughly $94,930.
OKF stores organizational knowledge as plain markdown files with YAML frontmatter. Instead of RAG's retrieve-every-time loop, it uses compile once, read always - and AI agents can read it directly, no vector database or SDK needed.
Every frontier model gets dumber as context grows — long before the window is full. Recursive Language Models (RLMs) fix it by treating the prompt as a variable in a Python REPL and recursing over it: GPT-5-mini beats GPT-5 by 34 points, and 10M+ token inputs stop collapsing.
Every AI system runs on one of five chip types — CPU, GPU, TPU, LPU, or NPU. Each makes a different tradeoff between flexibility and speed. This guide explains all five in plain English, with enough depth to actually help you decide which one matters for your work.
No cluster, no budget, no PhD — just a MacBook Air with 8GB of RAM, a gaming PC, and every free GPU I could find. How I abliterated 38 models (97k+ downloads), rebranded the entire collection, and ended up decensoring models I trained myself.
