6:24 PM
Today in brief.
Gemini Decision 1 scores 1,000 questions in one pass
Google shipped a new model and API that scores yes/no, multiple-choice, and rubric questions in a single forward pass, with no output-token bill.
🎬 Media
Seedance 2.5 in 4K is live in Runway
Runway turned on Seedance 2.5 at 4K, a concrete creator upgrade you can use in the same app that just opened Praxis-1.
🤖 AI
Agents now dominate web traffic, and United just banned them
a16z says agents already make up most web traffic and most new production code, while United is blocking unauthorized AI agents from booking flights.
🔬 Science
Praxis-1 and PAN 2 open their world models
Two open-weight world models landed today: Runway's Praxis-1 for real robots, and MBZUAI's PAN 2 for simulating what happens next.
🔬 Science
AlphaGenome 2 predicts DNA effects in specific cells
DeepMind's new DNA model zooms from tissue averages to individual cell types, and it is live today for academic research through an API.
🔬 Science
Agents rediscover a Maya site in 72 hours
A non-archaeologist set agents loose on public lidar, reported a hilltop site to Mexico, then learned archaeologists had already mapped it last year.
🏛️ Policy
A backdoor in fewer than 10 neurons
Microsoft researchers showed they can hide a trigger in a handful of rarely used neurons, then make an agent emit an attacker's command without retraining the whole model.
🔧 Tech
An 0.8B on-device agent beats models 2.5x its size
Tether open-weighted BotPsy, a 0.8 billion parameter assistant that it says roughly doubles its base model's tool-use scores and beats a 2B sibling on most tests.
No stories in this category for this edition.