MCP gets interactive apps, OpenAI's Responses shape becomes an open spec, Kimi K2.5 brings trillion-scale open weights, and Claude Code starts the 2.1 line.
The first issue of this series. Once a month I collect what actually moved in AI: model releases, agent tooling, research, and whatever else affects builders. Every link goes to the primary source. January belonged to the agent stack standardizing: interactive UIs for MCP, an open Responses spec, and open weights closing in on the frontier.
A 1T mixture-of-experts model with 32B active, vision, 256k context, and a modified MIT license, plus native decomposition of tasks into parallel sub-agents (Agent Swarm). One of the strongest open options for agentic workloads, self-hosted or routed cheaply.
The Codex variant of GPT-5.2 became directly callable through the Responses API, so you can build it into your own harnesses instead of only using the Codex product. Tuned for agentic coding, with context compaction for long tasks.
The technical report for Mistral's small-model line. Relevant if you run local or edge models and want documented training and eval detail rather than a bare weights drop.
MCP tools can return interactive HTML UIs, rendered in sandboxed iframes with JSON-RPC over postMessage and an official SDK. Claude, Goose, VS Code Insiders, and ChatGPT shipped client support the same week, so one UI runs across hosts.
A big release for skill and agent authors: skill hot-reload, context: fork to run a skill in a forked sub-agent, hooks in frontmatter, and wildcard Bash permission patterns like Bash(npm *). Rapid 2.1.x point releases followed through January.
An open-source spec for interoperable LLM interfaces built on the Responses API shape. It matters if you maintain provider-abstraction layers: a candidate successor to "everyone clones the Chat Completions schema".
The paper argues embedding scaling outperforms adding MoE experts under matched budgets, challenging the default recipe. Useful context for reading between the lines of current MoE release notes.
A full report on the fast MiMo-V2 variant, part of a January wave of unusually transparent reports from Chinese labs. Practical reading if you care about speed-versus-quality trade-offs in open models.
The Blackwell successor is in production, claiming around five times the inference performance and tokens at a fraction of today's cost, with systems due in the second half of the year. That curve decides which agent workloads are affordable next year.
A long essay laying out five risk categories for powerful AI and a concrete mitigation plan, drawing on Anthropic's own red-team findings. It sets the frame for the safety debates that will reach builders as policy.
