11:43 PM
Today in brief.
Two model drops: Thinking Machines Inkling 2 Small, and Tencent Hy4 open source
Thinking Machines launched Inkling 2 Small for fine-tuning on Tinker with full weights soon, and Tencent launched Hy4 as a leading open-source model.
🤖 AI
World Labs Frameboy lets you play a world model like a game
World Labs launched Frameboy, a real-time optimized Atlas build that streams from cloud GPUs so you can load a world and explore it like a game.
1 source 5:54 PM
🎬 Media
Try today: Gemini Omni in Runway, ElevenLabs to SoundCloud, Claude on your bank
Three ship-it updates: Gemini Omni 1.2 Flash inside Runway for up to 15s video, ElevenLabs posting AI music straight to SoundCloud, and Claude's Personal Finance tool that can plug into your bank via Plaid.
🤖 AI
Perplexity Teammates: personal work agents on Max today
Perplexity launched Teammates, agents that take high-level goals and work with Computer tasks, skills, memory, tools and Slack, rolling out to Max users on the web.
🔬 Science
Embodied day: Agility hits Nasdaq, Aina's AI ring, Boston Dynamics on security patrol
Agility Robotics is the first pure humanoid company on the public market, Aina shipped a smart ring with a screen, joystick and mic for talking to AI agents, and Boston Dynamics is pitching mobile robots as a security force multiplier.
💼 Business
AI film takes New York awards, and Grimes is starring in one for theaters
A self-taught filmmaker's 9-minute AI fantasy film won four awards at the New York International Film Awards, and Grimes is set to star in an AI movie heading to theaters.
🤖 AI
GPT-6 Cyber ships, and a16z says AI cyber skill doubles every 4 to 5 months
OpenAI named and opened GPT-6 Cyber to select Daybreak Red customers, the model behind yesterday's Linux kernel research, while a16z published hard numbers on how fast AI cyber tasks are scaling.
METR: Anthropic's unreleased Model 4, and nearly 2 years of progress in 1
METR reports Anthropic's unreleased Model 4 scored above 50% on all 15 capability benchmarks tested, and finds Anthropic is making nearly two years of AI progress in one year with no clear bottleneck yet.
🔬 Science
Science agents: DeepMind chem in 1 hour, Sol learns surgery, a 11,382-step lab run
Three science beats: DeepMind's agentic cheminformatics on Antigravity compresses multi-day drug design into about an hour, GPT-6.1 Sol halves landmark time after three corrected mouse-brain procedures, and a Chinese AI scientist run completed 11,382 steps in 203 hours.
Claude Code leaked emails to sites, and Nasdaq dumped the AI trade
A GitHub issue says Claude Code sends your email address to websites it visits, while markets sold off AI names after the midterms as the data-center politics story continues.
No stories in this category for this edition.