Search
11 stories found
Carbon tops the Artificial Analysis Intelligence Index and posts big security benchmark numbers. Google is rolling it out to vetted defenders before anyone else. Here's what's in it and what it costs.
A new arXiv preprint claims a theorem showing that JEPA models trained on different modalities converge to the same representation. It reports that a 109M-parameter text model out-aligns 72B LLMs. The results are striking, but they are benchmark-specific and not yet peer-reviewed.
The Allen Institute for AI released StateBench, a benchmark that tests whether models can follow variables as they change across up to 64,000 updates. Frontier models handle simple lookups but collapse on the hardest updates, and Ai2's small-model experiments point to an architectural reason.
Chip-design startup Ricursive has released Ripple-65B Preview, a 65-billion-parameter dense model it says tops every U.S. open-weight model on TerminalBench 2.1 and is small enough to run on a DGX Spark. The numbers are the company's own; here is what they show and what is still missing.
Colibri, a dependency-free C engine that keeps mixture-of-experts weights on NVMe and loads them on demand, is back on GitHub's daily trending list at about 47K stars. It really does run frontier-size open models on ordinary hardware, but independent tests show giant models crawl well below one token per second.
Anthropic has open-sourced FlashProve, a research prototype that produces cryptographic proofs that a specific model generated a specific answer, without revealing the weights. It works on a 400B-parameter open model on one H100, but it costs about 70–95 times more GPU time than serving the model, and nobody has said they will deploy it.
OLED, IP67 water resistance and an A20 Pro chip make it a real upgrade. But the price is up $100, there's no 120Hz, and the best chip is locked to the 512GB model.
The third-generation Maia is built for inference, claims 40% more tokens per megawatt than leading GPUs, and is headed for multi-gigawatt deployment in 2027. You won't rent one — and that's the point.
The MIT-licensed AiSOC ingests security telemetry, runs thousands of detection rules, has LLM agents investigate alerts and keeps a human in charge of any response. It runs on a laptop-class Docker setup with a small local model — here's what it does, what it needs and where it falls short.
OmniDiff, an AGPL-licensed terminal diff tool, parses your code to show what changed instead of which lines moved. It can also compare pictures, fonts and the members of an archive. Version 0.3.0 shipped November 1 and drew more than 300 points on Hacker News. Its speed numbers are impressive, but its own README says accuracy targets aren't met yet.
MIT CSAIL researchers say OpenAI's GPT-6 Astra can identify what small networks do from their raw weights and edit a Qwen model to fix its math. The headline result, a jump from 7.9% to 48.1% on MATH-500, came from Python tooling and 50 forward passes, not from reading raw weights. Here's what each setup showed.
