Hugging Face Post Introduces Consistency Guidelines to Close AI Agent Reliability Gap
Hugging Face post introduces tools to halve a 24.4-point consistency gap for a GPT-4.1 ReAct agent on AppWorld.
Latest news, developments and reporting on Large Language Models.
Hugging Face post introduces tools to halve a 24.4-point consistency gap for a GPT-4.1 ReAct agent on AppWorld.
iOS 27 is now available, according to 9to5Mac Daily, which also covered Apple data sharing, a T-Mobile iPhone Handoff fee, and Craig Federighi comments.
Meta launched Meta One bundles globally, adding AI usage and premium social features for $7.99 to $499 a month.
Elon Musk called for leading AI companies, including top Chinese firms, to let rivals test their models before release, as AI leaders warn of risks and the Trump administration rejects slowdown calls.
Reuters: DeepSeek plans to name GL Ventures partner Yan Wentao, 35, as its first CFO, ahead of a STAR Market IPO.
Pioneers of modern AI, including Yoshua Bengio, Geoffrey Hinton and Aidan Gomez, are adding their warnings to a wave of safety alarms from leading labs as Washington debates regulation and President Trump opposes new oversight.
Mozilla's State of Open Source AI report, published September 15, says U.S. frontier models lead the best Chinese open-weights models by 4.4 months, and that most organizations should make open models their default.
The OpenAI Foundation is funding efforts to create biological datasets for AI, including a $500,000 grant to acquire failed biotech companies’ drug-development files.
Warnings from leading AI executives that model development should slow have unsettled data-center stocks and forced investors to reassess the industry's growth narrative. Data-center executives and analysts say cloud and inference demand remains strong, while some job seekers are rejecting AI-led interviews, according to CNBC and MIT Technology Review.
Leiphone reported that StepFun launched five StepAudio 3 voice models for real-time interaction, speech recognition, TTS, audio and music generation; several ranked first on Artificial Analysis.
Salesforce unveiled Koa, an Nvidia Nemotron-based reasoning model for CRM tasks, and deepened AWS and Google Cloud integrations at Dreamforce 2026.
Nvidia detailed its IMO gold math pipeline; other AI firms reported advances in edge, video, robotics and biotech.
NPR reports the Supreme Court blocked Trump’s mail-in voting limits before the midterms, while CNBC reports Trump rejected AI leaders’ calls for slower model development and new guardrails.
Recursive self-improvement is the AI industry’s new battleground as frontier labs report automated research gains and critics warn of control risks.
Two TechRadar pieces urge control over AI models, data and workflows to avoid lock-in.
NYU professor Tristan Buckmaster accuses OpenAI of copying his work and says AI makes publishing races pointless.
Meta FAIR and University of Washington researchers propose distilling language models at the byte level instead of the token level, predicting a downstream accuracy ceiling about four points above conventional token distillation.
Liam Fedus, OpenAI's former post-training VP, answered a statement by 25 Fields Medalists who warned that AI's race for headline results is misaligned with mathematics. He agreed that understanding matters, but said today's limits should not define what AI will do next.
ByteDance CEO Liang Rubo said Doubao, Feishu and Volcano Engine are integrated and more will go to enterprise AI.
BYD introduced its Didi Xia super intelligent agent and the Denza N8L pure electric vehicle on Sept. 14, expanding an AI cockpit partnership with Alibaba Cloud that uses Qwen models and ecosystem agents.
New Mexico court fines lawyer $5,000 and holds him in contempt after ChatGPT-fabricated testimony appeared in a murder appeal.
Agent-net released Webagent, an Apache 2.0 Go harness that turns a website into a guarded AI agent via a JSON spec and nine pluggable slots.
QuantumBit reports that seven doctoral students at Beijing Zhongguancun Academy trained ZGCM-1, a 7B model, from scratch in three months using hundreds of AI agents for data, experiments and evaluation. They released weights, code, data recipes, checkpoints and logs.
Nvidia open-sourced SoL-Pi, an agent workflow harness that cuts token use by up to 64% and API costs by 50–54%.