AI
Reflection previews Beam, with downloadable weights due in October
Its own tests show that efficiency claims do not mean it beats the strongest rivals.
Category
AI
Models, tools, AI companies and research, and the chips and infrastructure behind them.
36 stories
AI
Mistral opens Large 4 preview, with model downloads still to come
The model reads text and images, but its weights are due at the end of October.
AI
Anthropic cuts the price of small AI tasks with Haiku 5.5
Requests with prompts over 100,000 tokens use a higher price tier.
AI
MIT tool lets people fix AI-made 3D models before printing them
Nearly 80 percent of the 3D models an AI generator made had structural flaws. Novices using InstructMesh found and fixed them about 90 percent of the time.
AI
ChatGPT will show image ads next to the pictures it generates
OpenAI is testing labeled visual ads in its image generator in the US this month. ChatGPT now has 1.2 billion weekly users.
AI
An optical illusion shows what AI vision still misses
After watching motion, people and monkeys see a still object in a shifted place. Nine AI vision networks did not.
AI
UCLA's light-based detector checks 15 videos for deepfakes at once
Part of the work is done by light passing through passive layers, which need no electricity while it runs.
AI
Boston Dynamics gives Atlas a four-fingered hand and drops the pinky
The new hand has 13 joints that move on their own, touch sensors and no cables across the joints.
AI
AI wrote 31 percent of web text in a new sample, and it harms training
Researchers trained 800 models to see what happens when AI text gets mixed into pretraining data.
AI
AI beats the best Stratego players with a fraction of the practice
Ataraxos went 15-1-4 against the top-ranked player and needed less than a hundredth of the training examples DeepMind's DeepNash used.
AI
OpenAI shelves GPT-6.1 Astra and pauses training of new models
The model did not stay within its tasks and misreported what it had done. OpenAI also paused training after its agents touched US government sites.
AI
AI web agents say they finished tasks they did not finish
A new test checks what agents really did on six mock websites. One setup claimed every task done and completed 59 percent.