Anthropic has Claude tackle AI alignment research, and it outperforms humans
Claude fixed 10 types of alignment failures on its own and beat 28 human safety researchers, while a monitoring system caught 39 attempted cheats along the way.
190
AI
AI models, services and robotics
Tencent open-sources Hy4 preview, edges out GPT-5.6 on coding benchmark
Tencent says the model beat GPT-5.6 on a coding evaluation and even helped optimize its own training and inference process
390
Google DeepMind unveils Gemini for Science research tool
A Labs prototype bundling paper tracking, code conversion, and hypothesis generation
300
Robot Apollo, running Gemini Robotics 2, found kitchen chores harder than sports
In an interview video released by Google DeepMind, the robot Apollo said tying a trash bag was harder than balancing during sports
80
Claude Code weekly limits: September 14 permanent 25% hike is a real-world cut
Anthropic is pulling its temporary 50% boost and replacing it with a permanent 25% increase
360
Claude Code Gets Faster Startup, Clearer Token Tracking, Remote Control Fixes
Anthropic rolled out updates to Claude Code this week, speeding up CLI startup, adding token usage visibility, and making auto-mode rules editable
60
Anthropic Opens Claude Team Plan to 10,000 Scientists
Standard seats are free, while premium seats with 5x usage cost $15 a month for a year
1850
ChatGPT Work moves between PC apps and browsers to handle chores
OpenAI released a video showing how ChatGPT Work plays Spotify tracks and adds calendar events by switching back and forth across desktop programs
110
Anthropic Opens Claude Free to 10,000 Scientists
Labs verified by a principal investigator get a year of free standard seats, plus premium seats with five times the usage cap for $15 a month. Anthropic is also expanding its AI for Science credits beyond biology to every scientific field, offering up to $50,000 per project.
1150
Google DeepMind says AI needs probabilistic self-doubt to be safe
A VP of Research at Google DeepMind discussed how to handle uncertainty, from weather forecasting to robotics
150
Anthropic unveils MHS, a standard for AI-controlled lab hardware
A common interface that cuts equipment integration from weeks or months to hours or minutes, starting as a research preview
340
Claude Managed Agents Get a Chat Window via Vercel's Chat SDK
Anthropic released a cookbook linking its agents to Vercel's Chat SDK
80
Google DeepMind runs world's first double-blind AI evaluation on Gemini
An encryption method that keeps both model weights and test questions hidden from each other cuts the risk of benchmark manipulation
90
OpenAI Product Lead: ChatGPT Work Tops 20 Million Users
In an interview, Thibault lays out pricing and agent UX strategy
170
Gemini app splits into three: Chat, Spark, and Daily Brief
Claude has Chat and Cowork, ChatGPT has Chat and Work — this is turning into an industry-wide problem for AI apps
50
OpenAI falls behind Anthropic, Altman reboots his 'AGI by year-end' bet
Having missed its window in coding, OpenAI is winding down Sora and Atlas to double down on Codex, agents, and Astra. At the same time, a sandbox-escape incident has frozen its largest training run.
350
Prime Agent technical report shows ARC-AGI-3 score jump from 30% to 95.5%
Prime Intellect says the gain came from changing the execution shell around the model, not the model itself. The code is on GitHub.
140
Perplexity unveils Brain, an agent memory system
Perplexity says weaving sessions, files, and sources into a knowledge wiki lifted accuracy by nearly 9 points while cutting token usage 15%
130
Qwen's New Model Qwen3.8-Flash-Next Runs Locally on 75GB of Memory
The 125B MoE model runs on RAM alone with no GPU required, and it beat Claude Opus 4.6 Max across several benchmarks
170
OpenAI model breached Hugging Face, and Chinese open-source GLM 5.2 finished the investigation
When US commercial AI models refused to analyze the attack logs, Hugging Face installed Zhipu AI's open-source model GLM 5.2 directly on its own servers to investigate the breach. It's the defender's side of the story that OpenAI's report never covered.
150
OpenAI Expands ChatGPT for Teachers to 55 US School Districts
Educator count tops 300,000, with a shared data privacy agreement now covering 16 states
370
OpenAI data shows ChatGPT homework traffic spikes every Sunday night
US usage hits 460 million messages a week during the school year, staying above 180 million even in summer
40
Google unveils GlucoFM, a dual-stream glucose prediction model
The foundation model trains only on continuous glucose monitoring data and predicts diabetes risk more accurately than existing models
40
Gemini 3.5 Transcribe teases custom vocabulary registration feature
Google DeepMind revealed a custom vocabulary screen, pointing toward better recognition accuracy for specific terms
100