The Informed Layer · Newsscan
Daily briefing

Claude Code sidesteps encoding bugs; world models persist

← All issues

Anthropic / Claude ecosystem

Claude Code now refuses to edit non-UTF-8 files after mangling non-ASCII characters

A regression saw Claude Code replace every non-ASCII character when working on non-UTF-8 files (Windows-1252, Shift-JIS, GBK); the current build now refuses to touch such files rather than corrupt them. The behavior change affects developers working in non-UTF-8 codebases and encodings.

Claude Code Projects beta opens to Pro and Max users

Anthropic has cleared the waitlist for Claude Code Projects, opening the beta to Pro and Max subscribers and clarifying usage limits and beta boundaries. This broadens access to the project-scoped Claude Code experience.

Frontier model providers

DeepSeek launches V4.1-Flash with 1M context and FP4 KV cache optimization

DeepSeek released V4.1-Flash, a 1M-context model with FP4 KV cache optimization, under an MIT licence permitting self-hosting, fine-tuning and commercial use.

AI developer tooling & infrastructure

Windsurf Wave 14 adds Arena Mode for in-IDE model battles

Windsurf's Wave 14 introduces Arena Mode and Plan Mode, letting developers compare models side by side inside the IDE with user voting and persistent leaderboards.

Cloud & platform providers

No significant new developments.

AI policy, regulation & governance

EU AI panel debates which frontier-risk advice must be made public

The EU's AI panel is grappling with how much of its frontier-risk guidance should be published, with unclear artefacts and disclosure boundaries flagged in the discussion.

ICO closes foundation-model programme, opens formal agentic AI inquiry

The UK Information Commissioner's Office has ended its voluntary foundation-model reporting programme and opened a formal supervisory inquiry into agentic AI, moving from voluntary reporting to enforcement-oriented oversight of major developers.

Industry & market moves

Kimi K3 becomes first Chinese model in OpenAI's Codex enterprise billing via Baseten

Through a Baseten partnership, Kimi K3 (and GLM 5.3) can now be accessed through OpenAI's Codex enterprise channel — the first time a Chinese open-source model has entered OpenAI's enterprise procurement system.

Google DeepMind hires engineers with no degree required, pay up to Rs 10 crore

Google DeepMind's Polaris-related engineering hiring drops degree requirements, with compensation reportedly reaching Rs 10 crore for software engineers evaluated on Gemini-tested tasks.

Cohere opens Ottawa office to expand Canadian federal government work

Cohere is opening an Ottawa office to engage federal agencies and accelerate sovereign AI deployments, deepening its Canadian public-sector footprint.

AI product & feature launches

Dawa real-time world model maintains persistent, controllable video worlds

Dawa establishes a persistent, stateful world model for video that maintains continuity over time rather than generating isolated frames, with a public demo/SDK expected at an October Shanghai showcase.

Open WebUI 0.12.0 adds realtime voice, MFA and collaborative chat

The Open WebUI 0.12.0 release introduces realtime voice, multi-factor authentication, collaborative chat and per-model controls in the self-hosted AI interface.

Research with immediate practical relevance

Shanghai AI Lab's RoboRSI achieves stable, reusable robot self-evolution across four SOTA benchmarks

RoboRSI introduces a four-agent governance framework and self-evolving skill library, enabling 104 autonomous real-world rounds and state-of-the-art results on four robot benchmarks.

HKUST and ByteDance open-source FreeMatching for dense correspondence beyond spatio-temporal priors

FreeMatching performs dense correspondence matching without dense manual annotation, delivering significant EPE and PCK gains on AI-edited images and introducing a human-aligned identity-preservation metric.

KAIST unveils ME-World, the first multi-agent egocentric world model

ME-World synchronises multiple egocentric video streams with shared memory and cross-view action conditioning, improving cross-view coherence and fine-grained controllability in embodied world modelling.