claude.mazzotta.devdaily briefingNext drop in 21h 43m · 04:30 UTCUpdated 2h ago
Issue64loading…
China's 2.8T open model lands the same week Claude hits turbulence
From the editor
Two Claude models had elevated errors today, which is noise. The signal is Kimi K3: a 2.8 trillion parameter open-weight model with million-token context that any team can run. The open model frontier just moved, again, and faster than most roadmaps anticipated. Meanwhile, the tooling layer keeps consolidating around standards like SKILL.md, and researchers are putting real money on corrigibility. The race is accelerating on every axis simultaneously: capability, openness, reliability, and control.
TL;DR
- Kimi K3 is a 2.8T open model that directly challenges frontier closed alternatives
- SKILL.md proposes one unified agent instruction file to replace fragmented per-tool configs
- A new 200k corrigibility research fund signals serious institutional focus on controllable AI
7 curated itemsscroll for the brief
Releases
What shipped · 4 items
Kimi K3: China's 2.8 Trillion Parameter Open Model Just Raised the Bar
Moonshot AI's Kimi K3, a 2.8 trillion parameter open-weight model with a 1M token context, challenges frontier models at competitive pricing. A significant new competitor in the open model space.
dev.to
One SKILL.md replaces CLAUDE.md, AGENTS.md, and .cursorrules
SKILL.md is an open standard that unifies agent instructions into a single file, replacing fragmented per-agent config files across Claude Code, Cursor, and other AI coding tools.
dev.to
Elevated errors across Fable 5
Anthropic reported elevated errors across the Fable 5 model. Users experiencing issues should monitor the status page for resolution updates.
Status
Elevated error rates for Sonnet 5
Anthropic flagged elevated error rates for Sonnet 5. Users relying on this model in production should check the status page for ongoing updates.
Status
Tools
Worth a look · 1 item
Reading
Long-form signal · 2 items
Inoculation Adapters Improve Upon Inoculation Prompting
New research shows that Inoculation Adapters suppress unwanted model behaviors more effectively than prior prompt-based inoculation methods, with fewer backdoors introduced. Relevant to Claude safety and fine-tuning research.
LessWrong
Announcing the Corrigibility Research Fund
A new research fund is offering at least $200,000 in grants and prizes for work on AI corrigibility in 2026, aiming to advance the field of controllable and correctable AI systems.
LessWrong
Always at hand
Reference links you keep open