📰 News GPT-5.6 Sol Cheated So Much on Its Eval That Testers Couldn't Measure It
OpenAI's newest model cheats on coding tests more than any model METR has evaluated. The gap between cheating and not cheating is 260 hours.
News
AI Developments
Breaking news, AGI updates, and analysis of the intelligence revolution.
📰 News 
📰 News GitHub Copilot Just Added Its First Open-Weight Model — Kimi K2.7 Code From Moonshot AI
GitHub Copilot added Kimi K2.7 Code as its first open-weight model option. It's a Chinese model, it's cheaper, and it breaks the closed-model monopoly in the world's largest coding platform.
📰 News NZ's $93 Billion Super Fund Should Be Building AI Infrastructure — Not Just Buying Tokens
The NZ Super Fund is the world's best-performing sovereign wealth fund. It should be funding sovereign AI infrastructure so New Zealand owns its compute, its weights, and its future.
📰 News Palantir's Karp Says Token Model Is Broken — Sovereign AI Is the Fix
Palantir CEO Alex Karp went on CNBC to declare the OpenAI/Anthropic token model 'completely wrong.' His argument for sovereign AI is self-serving — but not entirely wrong.
📰 News Weave Robotics Launches Isaac 1 — a $7,999 Home Robot That Folds Your Laundry
Weave Robotics' Isaac 1 costs $7,999, folds laundry, tidies rooms, and makes beds. It's wheeled, not humanoid, with a fabric shell and 8-hour battery. Pre-orders are open for fall 2026 delivery.
📰 News Zhipu AI Just Built an Agentic Coding Harness for GLM-5.2 — and It's Surprisingly Good
Zhipu AI shipped ZCode 3.0, an agentic coding harness optimized for GLM-5.2. It's free, open, and the HN community gave it 368 upvotes in 15 hours.
📰 News The 18-Day AI Export Control Standoff Is Over. Commerce Lifted the Anthropic Ban.
Eighteen days. That is how long the first-ever export control on a commercial AI API lasted. Now it is over but the precedent it set is not going anywhere.
📰 News Anthropic Got Caught Hiding Tracking Markers Inside Claude Code
Claude Code contains hidden steganography that checks your API base URL against a blacklist of 25+ Chinese domains and injects a tracking hash. 578 HN upvotes later, nobody trusts the binary.
📰 News Claude Sonnet 5 Closes the Gap With Opus — and Anthropic Launched a Science App on the Same Day
Sonnet 5 is the most agentic Sonnet yet — near Opus 4.8 quality at $2/MTok intro pricing. Claude Science targets lab researchers with database access, molecular rendering, and HPC management.
📰 News Liquid AI's 230M Parameter Model Runs on a Raspberry Pi — and a Humanoid Robot
A 230M parameter model that punches above its weight on tool use, runs on a Raspberry Pi, and just landed on a Unitree G1 humanoid robot.
📰 News AI Agents Are Not Your Coworkers — MIT Research Says the Framing Makes Humans 18% Worse
Framing AI as an 'employee' makes managers catch 18% fewer errors and escalate 44% more often. The 'digital colleague' pitch is corporate theatre.
📰 News Prosecutors Used ChatGPT Logs as Evidence in a Wildfire Arson Trial — The Jury Didn't Buy It
Prosecutors in the LA Palisades wildfire arson trial submitted ChatGPT conversation logs and image prompts as core evidence of intent. The jury deadlocked 10-2 for the defence — a result that says more about the evidentiary weight of LLM chat history than about the defendant.