Model Releases

LFM2.5-2.6B: Liquid AI’s Edge-Optimized Foundation Model — Liquid AI drops a 2.6B parameter liquid foundation model built for edge deployment with 16-language support. The liquid architecture (continuous-time, adaptive computation) means it should handle variable-length sequences more efficiently than standard transformers. Base and fine-tuned variants on HF if you want to kick the tires.

Open Source Releases

anthropics/claude-code v2.1.221: Focus view and sandbox credential masking — New Focus view in VSCode (Ctrl+Alt+F) collapses tool activity into expandable per-turn summaries with a live running indicator. Also adds "mode": "mask" for sandbox credential files on Linux/WSL — sandboxed commands read a sentinel copy instead of the real thing. Small but meaningful quality-of-life improvements.

sourcecode 4.10.3 - Persistent structural context for AI coding agents — Builds persistent ASTs, call graphs, and type indexes so AI agents can query codebases semantically without re-parsing everything on each request. Incremental updates mean the index stays fresh as you edit. If you’re building tooling around AI code navigation, this is the infrastructure layer you’ve been waiting for.

Research Worth Reading

Output-Aware Rotation for INT2 KV-Cache Quantization — Rotation-based INT2 quantization that optimizes for the actual attention readout instead of proxy cache metrics. Directly minimizes output distortion in the attention computation, which is where the memory/bandwidth pain lives for long-context inference. Practical angle on a real bottleneck.

Beyond the Hivemind: Escaping LLM Homogeneity via Meta-Persona Anchoring and Sequential Temperature Scaling — Documents an “Artificial Hivemind” effect: LLMs converge to 0.80-0.90 inter-response similarity even at high temperatures. Their fix combines meta-persona anchoring with sequential temperature scaling to restore diversity without tanking quality. If you’ve noticed every model sounding the same lately, this explains why.

HyperAgent: Planning and Acting over Tool-Schema Hypergraphs for Tool-Use LLM Agents — Represents tool schemas as hypergraphs to enable explicit planning over tool compositions. Moves beyond “let the LLM figure it out” by structuring tool relationships and execution constraints in a queryable format. Sensible approach for complex tool-use scenarios where implicit reasoning falls short.

BAP-SQL: Budget-Aware Observation Planning for Agentic Text-to-SQL — Treats observation formation in agentic text-to-SQL as a budget-control problem. Plans observations within token/compute budgets upfront rather than relying on post-hoc compression that can’t recover omitted rows. Pragmatic framing for a real cost problem.

When Compression Scores Cannot Decide: Information Boundaries for Group-Robust LLM Pruning — High-reliability compression stats (0.906 split-half) still pick suboptimal pruning candidates — 6-7.7% worse than controls. Models the gap via “information interfaces” that delimit which distinctions each statistic actually supports. Important if you’re doing structured pruning and trusting your metrics.

LLMs Can Annotate Attribution Graphs — Automates the manual clustering step in circuit tracing by prompting LLMs to group features/neurons into supernodes from their descriptions. Practical pipeline for scaling mechanistic interpretability beyond toy models. The “use LLMs to interpret LLMs” loop continues.

AI Dev Tools

Uber ADR: Enterprise AI agent security and observability platform — Production-deployed framework at Uber for agent governance: observability, security benchmarking, threat detection. Built for enterprise requirements around runtime protection and auditability. Worth studying if you’re figuring out what “agent security” actually means in practice.

anthropics/claude-code v2.1.222: Worktree isolation fixes and hook security — Fixes worktree-isolated sessions to actually prevent destructive git commands against the main checkout (isolation now covers file edits and Bash in all session types). Also patches PreToolUse auto-allow hooks bypassing restrictions in background agent tasks. Security hardening that should’ve been there from day one.

cline/cline SDK v0.0.70: Plan mode hard-blocks file-editing commands — Plan mode now hard-blocks mutating shell commands (sed -i, perl -i, redirection, mutating git, package installs, nested commands) instead of just prompting. Read-only investigation commands still work. Finally, a plan mode that actually prevents accidents.

Claude Code How-To: Visual Guide with Copy-Paste Templates — Example-driven visual guide from basics to advanced agents, with immediately usable copy-paste templates for common workflows. Good reference if you’re onboarding team members or just want a cheat sheet for patterns you keep forgetting.

Today’s Synthesis

The AI coding agent stack is quietly hardening. cline/cline SDK v0.0.70 finally makes plan mode actually prevent mutations — hard-blocking sed -i, mutating git, package installs — instead of just prompting. anthropics/claude-code v2.1.222 plugs the worktree isolation gap so background agents can’t touch your main checkout. Meanwhile sourcecode 4.10.3 gives agents persistent ASTs, call graphs, and type indexes that update incrementally, so semantic navigation doesn’t require re-parsing the world on every request. And if you’re running this at scale, Uber ADR shows what production agent governance looks like: observability, threat detection, audit trails. The pattern is clear — we’re moving from “give the LLM a shell and pray” to proper isolation boundaries, semantic infrastructure, and runtime governance. If you’re building internal tooling around AI coding, these four pieces (isolation, plan-mode guardrails, semantic indexing, observability) are the minimal viable platform. Skip any of them and you’re still in the YOLO phase.