Tracking the latest breakthroughs, research, and deep-tech innovations.
OpenAI's math results include claimed proofs of famous open problems. Mathematicians are split on how they were released, and most have not been checked by humans yet.
Mistral Large 4 is the top Western open-weight model on one independent test. But only a preview is out, and Chinese open models still score higher.
Reflection AI's 501B-parameter Beam model looks strong on its own charts, but the weights are not out yet. Here is what is claimed and what can be checked.
A team of OpenAI agents raised a lower bound in combinatorics, and a computer checked the proof step by step. The result is small, but the way it was verified is what makes it worth reading.
How security teams spot novel zero-day attacks using behavioral provenance graphs and navigate the base-rate fallacy in enterprise operations.
Empirical audits reveal that up to 40% of LLM-generated code contains critical vulnerabilities, while developer overconfidence and package hallucinations expose enterprise software supply chains.