AI news for 14 September 2026
Monday 14 September 2026 · Latentown
On Monday 14 September 2026, Latentown tracked 2 posts from 1 AI lab, 4 research papers. OpenAI was the busiest lab with 2 posts. The highest-ranked story was "Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation".
Top stories
- Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation
Top paper on Hugging Face that day. A searchable, continuously updated database of AI benchmarks and evaluations.
- Feyospace-v1: How the Cyber Mercury Seven Trained Frontier Cyber Models
Argues that open-weight cyber agents are limited more by post-training data than by model scale, and describes how the team trained theirs.
- DataFlex-RL: An Evaluation Platform for RLVR Data Policies
A platform for comparing how reinforcement learning with verifiable rewards chooses and weights its training data.
- LimiX-2: a contextual mechanism network for structured data
The most upvoted paper on Hugging Face this week. LimiX-2 scales the LimiX family for tabular and structured-data tasks using the group's earlier scaling laws.
- Python SDK 3.14.0 released
A minor OpenAI Python SDK release that normalises errors raised while reading streams.
From the AI labs
OpenAI
- Python SDK 3.14.0 released
A minor OpenAI Python SDK release that normalises errors raised while reading streams.
- Perplexity trusts GPT-6 Astra with end-to-end systems
A customer story. Perplexity uses GPT-6 Astra to write communications, change software and monitor production systems, and checks in less often than with earlier models.
Earlier: GPT-6 Astra: A new generation of intelligence OpenAI · 3 September 2026
Research papers
- Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation
Top paper on Hugging Face that day. A searchable, continuously updated database of AI benchmarks and evaluations.
- Feyospace-v1: How the Cyber Mercury Seven Trained Frontier Cyber Models
Argues that open-weight cyber agents are limited more by post-training data than by model scale, and describes how the team trained theirs.
- DataFlex-RL: An Evaluation Platform for RLVR Data Policies
A platform for comparing how reinforcement learning with verifiable rewards chooses and weights its training data.
- LimiX-2: a contextual mechanism network for structured data
The most upvoted paper on Hugging Face this week. LimiX-2 scales the LimiX family for tabular and structured-data tasks using the group's earlier scaling laws.
Every story links to its original source. Ranking is by automatic rules, not an editor. Something wrong? Tell us.