NESTFRONTIER / ARCHIVE
Technical analysis, model releases, research, and deployment notes.
Bridgewater fine-tuned Qwen3-235B on expert-labeled financial data and beat every frontier model at 1/14th the cost. The era of renting the biggest generalist may be ending.
A GitHub issue reveals GPT-5.5 Codex responses cluster at exactly 516 reasoning tokens, correlating with task failures. 390K records analyzed, 82% of exact-516 events come from GPT-5.5.
GitHub Models dies July 30. The free playground, API, and model catalog are all gone. Here's what happened, where to migrate, and why this keeps happening.
Microsoft gave 5,000 engineers Claude Code. The annual AI budget lasted four months. Here's the token math that's breaking enterprise budgets.
Rendering code as images before sending it to LLMs can cut token costs by 60%. But the savings come with accuracy tradeoffs and a pricing model that may not last.
claude-real-video uses scene-change detection to extract the right frames from any video, giving LLMs meaningful visual data instead of random snapshots.
Claude Sonnet 5's $2/M intro price hides a 30% tokenizer inflation that makes it cost more per task than Opus 4.8.
Moonshot AI says K2.7 Code cuts reasoning tokens by 30%. Independent testers aren't convinced by the benchmarks. The pricing math still works.
ZAI shipped a desktop app that runs Claude Code with GLM-5.2, an open-weight model that matches Opus on benchmarks at a fraction of the cost.
Lilian Weng traces the full history of scaling law research, showing where Kaplan and Chinchilla disagree and why the exponents matter more than anyone admits.