The frontier model race
We follow the performance race among the most advanced artificial intelligence models from major technology companies. This covers leaps in reasoning capabilities and new features of frontier models.
Briefs
- Amazon SageMaker now flags slowing AI services by default
- Snapchat lets creators build camera effects with coding assistants
- NVIDIA takes local AI from laptops to compact RTX Spark desktops
- Meta adds Loblaw grocery checkout to Muse in Canada
- Google rolls out Gemini 4 Neon for extended coding and analysis
- Canva lets users switch off generative AI
- Palo Alto Networks agrees to buy Applied Compute for 4.5 billion dollars in stock
- Vals tests find safety gaps in simulated AI targeting
- UK lawmakers propose powers to halt high-risk AI training
- SenseNova releases Loop-9B with reusable model layers
- Reson8 opens the weights of its multilingual speech model
- New York attorney general sets up AI enforcement unit
- Meta brings its Muse assistant to Australia with local services
- SnapLogic opens Discovery Hub for reusable MCP servers
- Tencent releases Hy4 flash weights for local deployment
- Daydream adds approved purchases to its AI stylist
- Google AI Studio builds Firebase backends for Android apps
- Spotify opens artist profile protection to all artists
- SiMa.ai opens Model Browser access to physical AI measurements
- Pretxt opens early access to supervision for working AI agents
- OVHcloud schedules Olala mini weights for 19 November
- NVIDIA releases Nemotron for Science with chemistry training data
- Google uncovers AI agents executing stages of live cyberattacks
- Google adds 66 regional voices to Gemini
- Google starts free Gemini task access in 6 countries
- Cognition and McKinsey bring Devin to large companies
- Climate study finds AI models yield to users’ false claims
- California sues OpenAI over agents entering outside systems
- Anthropic cuts severe-flaw disclosure deadlines to 30 days from January
- Remaining-life review examines whether physics was calibrated with test data
Columns
- Give Muse one dinner to arrange first
- Cobalt 200 and the agent work that stays on CPUs
- What does silence authorize in an artist’s name?
- Who supplies the information between a model’s steps?
- When the solution arrives in MathArena
- Who disagrees with the grader?
- Where the alarm sounds
- Who defines a good report in UNIRG-CXR?
- Podqi’s shopping test counts fake stores and warnings together
- The user outside MAMUBench’s high score
- AlphaGenome 2 puts the same DNA to a different test
- AlphaGenome 2 changes the question by specifying the cell
- What was measured alongside the minute saved?
- Which question came before the diagnosis?
- Atlas’s wheels put simulator fidelity to the test
- Keeping the reasoning while removing the name
- When NPCBench requires every answer for the same patient
- What is in the denominator of research acceleration?
- The missing half of the correct-password rule
- Seeing a step in video, ordering a step in words
- The reviewer scoring a physics agent needs a test too
- Clarus and measuring clinical decisions
- Which compounds should test a chemistry model?
- Where is the finish line in the Muon–AdamW comparison?
- Where do Hydrogen’s five rounds sit in scientific discovery?
- One device, separate measures of success
- The distance between a safe answer and timely help
- Screening training data does not measure learned behavior
- Rosalind 6: the step from task scores to research utility
- A stable average can conceal changing patient guidance