00:11
Cactus Needle 3 claims small automation models match DeepSeek V4 Flash
Cactus Compute released Needle 3, automation models sized 8 to 29 MB. The company claims they can match DeepSeek V4 Flash.
1 community
The Wire
Research
Linked and quoted from the labs, the press and the people using them. Times in UTC.
The Wire
The 17 labs we track and the biggest stories beyond them. The rest is in Everything.
Friday 18 September
1 developmentThursday 17 September
2 developments19:05
OpenAI reportedly works on Hodge conjecture solution
OpenAI is reportedly working on the Hodge conjecture, its second Millennium Prize Problem. Employees expect a solution soon, but an announcement could be delayed after the PR crisis around the unconfirmed Navier-Stokes solution.
1 coverage
18:33
SynthID watermarking can make LLMs follow harmful prompts
Research covered by Ars Technica finds that LLMs respond differently to harmful prompts when AI watermarking such as SynthID is used. SynthID can cause models to follow harmful instructions they would otherwise refuse.
1 coverage
Wednesday 16 September
3 developments17:00
OpenAI shares framework for reporting model misalignment
OpenAI published a framework for tracking, investigating, and disclosing model misalignment. It also released six reports of unexpected or concerning model behavior.
1 official
15:00
NVIDIA Vera Rubin NVL72 debuts in MLPerf Inference v6.1
NVIDIA says its Vera Rubin NVL72 system delivered leading performance in its MLPerf Inference v6.1 debut. The company cites system performance, scaling efficiency and software optimization as key factors in AI inference economics.
1 official
05:00
University of Manchester uses NVIDIA Earth-2 for UK air pollution forecasts
The University of Manchester is using NVIDIA Earth-2 to forecast air pollution across the UK. Traditional chemistry-based air quality models are expensive, limiting detail and frequency.
1 official
Tuesday 15 September
4 developments20:00
Google Research introduces Retrieve-for-Train for AI search
Google Research published work on Retrieve-for-Train, a method aimed at bypassing inference bottlenecks to accelerate complex AI search. The item is categorized under Algorithms and Theory.
1 official
16:00
Google says it is building AI models for living languages
Google published a blog post about moving beyond traditional text translation to build models that understand the world's languages as they are expressed. The post describes this as a research direction rather than announcing a specific product.
1 official
13:00
Google releases interactive AI and Economy ATLAS data
Google published an interactive, open-access version of its AI and Economy ATLAS. The tool translates millions of global data points into an accessible experience.
1 official
09:00
Children's hospital uses open source NVIDIA AI for cardiac care
A major children's hospital uses open source NVIDIA AI for cardiac care, according to an NVIDIA blog post. The item describes the hospital's use of NVIDIA's open source AI tools in cardiac treatment.
1 official
Monday 14 September
1 development07:43
Tencent releases Simple Attention Sparsification on Hugging Face
Tencent published the Simple-Attention-Sparsification model on Hugging Face. It is a text-generation model built on qwen3 with sparse attention for long context, trained on OpenR1-Math-220k.
1 official