Bringing you the best information on AI so you can stay updated. Clear, practical breakdowns of models, agents and tools. Stay informed. Stay ahead.
United States
Joined November 2022
- Tweets2.8K
- Following479
- Followers665
- Likes923
Morph just launched Reflex, a 27B model for making decisions from text or images.
One useful detail is cached pricing: teams asking several questions about the same support ticket can reuse the input rather than paying the full input rate each time.
The practical test will be whether those decisions stay accurate enough for the workflow.
Mistral AI showed its Large 4 model investigating a Django vulnerability, with 133 agents working through the code and testing attack scenarios before a patch.
The useful question for teams is whether the same approach catches issues in their own code, and what it costs to run and review.
How would you test that against your existing security tools?
Anthropic just shared FlashProve, research that lets developers verify whether a response came from a particular AI model.
That could help teams check whether the model they tested is the same one serving users. The useful next question is how much time and computing power those proofs need in a real deployment.
Ten AI launches from this week in one list.
Google had the busiest week with Genie 4, Gemini 4 Carbon, document editing in the Gemini app and PlanAssist. Mistral and NVIDIA both put out open weights.
Which one would you try first?
Y Combinator Labs put 50 AI agents in Minecraft with three hours to build a working civilization.
The first night took 22 of the run's 34 deaths. After that, the agents mostly hid. Their informal leader, Ada, told everyone to idle in lit shelters until they got short tasks.
That's a very human reaction to getting burned.
Would you call that caution or a failure to adapt?
What would you have changed in the setup?
Put a few Microducks near each other and they find one another over Bluetooth, sync up, and start singing. Nobody is conducting. They make the song up on the spot, in harmony.
Microduck comes from Pollen Robotics, the French team that is now part of Hugging Face. It started as a side project by one of their engineers, Antoine Pirron, who wanted his own little walking droid inspired by Disney's BDX robots.
It stands 25 cm tall and weighs about 780 g. Inside are 15 motors, a front camera, a small depth sensor, two IMUs, mics and a speaker. It has no hands, so it folds in half and grabs things with its beak. Clip wheels on its feet and it roller skates, which meant learning a whole new way to balance.
The skills are trained with reinforcement learning in the MuJoCo simulator and then copied onto the real robot. Sim and reality never match perfectly, so they test on the real duck, adjust and train again. The software is open source under Apache 2.0, so anyone can train a new trick at home.
Pre-orders are $399 with first deliveries targeted before Christmas. Jensen Huang got down on one knee with four of them at GTC Berlin.
What skill would you teach one first?
Full video: youtube.com/watch?v=j25BJ32_…
GitHub Copilot can now send some requests to AI models running on your own computer. It's an opt-in preview in the CLI, desktop app and VS Code.
One detail surprised me while making our video on it. Running the model locally doesn't make the session offline. The choice of where each request goes still happens in GitHub's cloud.
Would you turn this on if it still needs a connection?
Or is fully offline the whole point for you?
Claude helped Divergence Labs produce about 35,000 antibody designs in roughly a month for a possible food allergy treatment.
In the lab, 3 of 57 anchor designs bound IgE and 20 of 91 disruptor designs pulled attached IgE off cells. Three disruptors appeared to beat Xolair at that one job.
It's still a lab assay, so nothing yet on safety in people.
Which number matters more to you here?
Would you trust AI-picked leads like these?