01
OpenAI ↗
Physicist Nathaniel Craig described how GPT-6 Astra helped develop a candidate lattice definition of the Standard Model. The November 15 post explains work presented in an October preprint, including assistance with the construction and detailed proof. It is a proposed mathematical approach under scientific scrutiny, rather than an established resolution of the problem.
02
Unite. AI ↗
Bharat Semi raised $24 million in seed funding to expand engineering and finish development of its AI-enabled image sensors. The New Delhi startup remains pre-revenue and is targeting initial commercial production in 2027. Info Edge Group, South Park Commons and Indus Bridge Ventures led the round; the sensors are still under development.
03
Harvey ↗
Wotton Kearney is expanding Harvey across 14 offices in Australia, New Zealand, Singapore and Thailand. Harvey says a six-month pilot involving 200 users achieved 94 percent adoption and more than 30,000 queries. The wider rollout supports research, drafting and document analysis, with client collaboration also being explored.
04
IBM Granite / Hugging Face ↗
IBM released experimental Granite 4.1 models with 3 billion and 8 billion parameters trained to use LLMON, a markup format for model context. It marks instructions and data with explicit structure to help models distinguish what to follow from what to process. IBM reports lower prompt-injection success on some benchmarks, but the models remain previews and the results do not establish universal protection.
05
MIT News ↗
MIT researchers built a system that asks engineers to review regions of a structure while an optimization algorithm designs it. In the reported study, human input produced structures that were easier to manufacture and tolerated higher buckling loads. The added interaction took about 15 seconds beyond the fully algorithmic process in the experiments.
06
Ai2 / Hugging Face ↗
Ai2 introduced XYBench, a benchmark built from 8,115 real-world questions where users may be pursuing a mistaken approach to their goal. The tested models often supplied a useful solution while continuing to focus on the flawed approach. The benchmark measures whether assistants can recognize those assumptions and redirect users toward a more effective answer.
07
arXiv ↗
WeightBench evaluates language models on 42 tasks involving reading, repairing or writing neural-network weights. The researchers report that frontier agents can understand small networks and produce working control policies in some cases. They still struggle to patch language models, compress them without data or distinguish impossible edits from merely difficult ones.
08
arXiv ↗
The FETTUCCINE research framework adapts speech-recognition models to turn recorded neural activity into text. Using public data from one intracortical speech-neuroprosthesis user, the authors report word error rates as low as 4.39 percent and operation on a commercially available mobile device. This is a research evaluation, not a consumer app or a demonstration of broad clinical effectiveness.
09
arXiv ↗
Researchers introduced Just Video Transformer, a video-generation design that works directly in pixel space without a pretrained autoencoder. Their experiments report quality competitive with latent-diffusion baselines in the tested settings. The system also generates aligned video and disparity information, improving camera following against the paper's RGB-only baselines.
10
WIRED ↗
Carnegie Mellon researcher Daehwa Kim developed Object Agent Twins: small E Ink displays that give everyday objects animated personalities. Sensors let a plant or dumbbell respond to care and use through an expressive avatar. WIRED describes a proof of concept with limited battery life; the creator has not decided whether to sell it as a product.