AI Daily — September 16, 2026
2026-09-16
deepmind.google
Introducing Gemini 3.8 Live and 3.8 Live Extended Thinking
Google DeepMind has released Gemini 3.8 Live and a variant with Extended Thinking, continuing the push to combine real-time multimodal interaction with chain-of-thought reasoning. The Live models are designed for low-latency, streaming audio/video interaction, while the Extended Thinking variant adds deliberative reasoning to that real-time context. This is a notable step in bringing reasoning-capable models into latency-sensitive, agentic use cases.
code.claude.com
Claude Code Week 35: Resume Terminal Sessions and Claude-Drafted Feedback
Claude Code Desktop now supports /resume to reattach to any CLI session with full conversation and context intact, closing a key gap between desktop and terminal workflows. The update also introduces Claude-drafted feedback reports that users review before submitting, and a --restricted flag for sandboxed sessions. These incremental but practical improvements continue to round out Claude Code as a production-grade agentic coding environment.
code.claude.com
Claude Code Week 33: Auto-Continue and Fork Mode by Default
Claude Code Desktop gains auto-continue on usage limit reset, so interrupted agentic sessions resume automatically without user intervention. Fork mode is now on by default in interactive sessions, enabling Claude to spin up subagents that inherit full conversation context for parallel side tasks. GitLab merge request integration also shipped, extending the platform's CI/CD reach beyond GitHub.
rss.arxiv.org
Latent Undertow: How Ordinary Typos Break Probes
This paper demonstrates a practical robustness failure in hidden-state probes used for detecting malicious or injected prompts: as few as ~3 common typos per message rotates the readout vector by 43–56 degrees at the perturbed token and cuts single-position probe TPR@FPR=1% by 12 percentage points. Multi-position aggregation mitigates localized perturbations but still drops ~3.8pp against distributed ones. The finding has direct implications for safety monitoring systems that rely on linear probes over LLM internals.