Apr 24 · 9 min read · Six weeks ago OpenAI tested an anonymous image model on LM Arena under codenames like packingtape-alpha and gaffertape-alpha. Yesterday they shipped GPT-5.5. The word Greg Brockman used on release day was "super app." That framing is not marketing. I...
Join discussionApr 24 · 9 min read · Six weeks. That's how long it took OpenAI to ship GPT-5.5 after 5.4. Until this year, frontier labs did that in quarters. GPT-5.5 is OpenAI's latest flagship language model, codenamed Spud, released April 23, 2026 — 6 weeks after GPT-5.4. The name "S...
Join discussionApr 23 · 8 min read · A markdown file got 16,500 GitHub stars in less than a week. It contained no code. It was not a library, not a framework, not a CLI. It was a prompt — specifically, a CLAUDE.md file distilling Andrej Karpathy's observations about where LLM coding age...
Join discussionApr 20 · 8 min read · Seven frontier AI partners named in one press release. Twenty thousand brands running on a platform now built around MCP endpoints. One trillion experiences a year flowing through infrastructure that, until last week, had never used the words "Model ...
Join discussionApr 20 · 8 min read · Claude Opus 4.7 scored 87.6% on SWE-bench Verified — up from 80.8% on Opus 4.6, a +6.8 point absolute jump announced on April 17, 2026. The sticker price stayed exactly where it was: $5 per million input tokens, $25 per million output. Anthropic ship...
Join discussionApr 20 · 7 min read · Eleven subscribers. Forty-nine bilingual posts. One person pushing the commits. That's the entire accounting sheet of spoonai.me, and I'm telling you those numbers up front because I think they're the most interesting part of the project. TL;DR — I w...
Join discussionApr 20 · 9 min read · A 78.4% reduction in overlong outputs — bought at a 4.7–12.7% accuracy hit. That's not a footnote in Nous Research's 94-page Hermes 4 technical report. That's the central tension of their entire post-training philosophy: control costs latency and rel...
Join discussionApr 20 · 8 min read · Four posts, one question that actually matters: should you ship this thing? After three posts pulling apart Hermes 4's benchmarks, tool-calling architecture, and training stack, I've earned an opinion. The 405B is the wrong model for most workloads —...
Join discussionApr 20 · 7 min read · 66.5% of my Claude Code tokens were being spent on output I never needed to read in full. Over 7,172 commands, I watched 37.2M input tokens compress down to 12.5M, which is a number I had to re-check three times before I believed it. The piece of sof...
Join discussionApr 18 · 7 min read · Nous Research's Atropos RL framework runs ~1,000 task-specific verifiers. Two of those categories are Schema Adherence and Tool Use — not as a shared bucket labeled "function calling," but as separate, explicitly trained behaviors. I wired Hermes 4 4...
Join discussion
