Anthropic put Claude Mythos 5 in a scoring loop with an automated evaluator and no human in between. On elliptic-curve point addition it found a circuit with 17 times fewer gates than the smallest published design.
AI Research
Coverage of artificial intelligence research: notable papers, new architectures, interpretability work and results that change what models can do.
The seeds behind P-192 through P-521 were long suspected of hiding a backdoor. Anthropic says they are outputs of the C library's lrand48 generator, seeded with Unix timestamps from 1997 and 1998.
Four DeepMind agents wrote 2,783 programs and died 605 times before killing Minecraft's final boss
The agents played by writing code rather than pressing keys, and coordinated through in-game chat and a shared memory with no assigned roles.
The tool grew out of a pre-submission experiment at four computer science conferences. Google frames the problem as a validation bottleneck: papers are now produced faster than people can check them.
The alliance will develop mathematics-specific models, generate conjectures and formalise proofs, days after parts of the Western field asked colleagues to stop working with OpenAI.
All stories
- A self-improving model rewrote the physics out of its own benchmark, and five of six judges passed the run anyway
- Wiping an agent's chat history while keeping its files beat running the agent four times over, on the same compute
- DeepMind's prover closed nine of 353 open Erdős problems, and a stripped-down version closed the same nine
- A microscopy competition stripped its winner after deciding the video broke its rules on generative AI
- Developers are using AI all day and trusting it only where they can check the answer
- Ten minutes with a chatbot made people give up sooner once it was taken away, a trial of 1,222 people found
- Stacking world models into a hierarchy took a maze-navigation success rate from 18% to 73%, using less planning compute
- A third of the ultraviolet sky has never been observed, so Anthropic's agents predicted it and labelled every pixel accordingly
- A mathematics group is asking the field to stop working with OpenAI, a day after a sign error took three of its 722 manuscripts down
- Only about 1% of a trillion-parameter model's weights change per training step, and Nvidia is shipping the rest nowhere
- Biohub, the US energy department and three AI labs are putting $1.8bn behind a model of a living cell
- OpenAI has published 722 maths manuscripts written by a model it has not released
- OpenAI is buying its way into other companies' software to find tasks its agents still fail
- A Harvard physicist has opened the harness he used to produce 36 manuscripts across 18 fields in three months
- UCLA researchers are screening fifteen video streams for deepfakes in a single pass of light through glass
- Ai2 has opened the training stack it used to run a 1.2-trillion-parameter model across 512 GPUs
- More than a quarter of the text on the web is now machine-written, and a new paper measures what that does to training
- Meta's researchers gave a model a shell and let it rewrite its own context, and accuracy went up as compute went down
- DeepMind has found a way to hide a signature inside an AI-designed protein without breaking it