Baidu Awards $2 Million to Two Technical Teams; Robin Li Calls Technology Its Real Competitiveness
Baidu awarded two teams $1 million each on Sept. 21. Robin Li said technology belief is Baidu's real competitiveness.
Latest news, developments and reporting on Large Language Models.
Baidu awarded two teams $1 million each on Sept. 21. Robin Li said technology belief is Baidu's real competitiveness.
A paper described in a Hugging Face Blog post reformulates LLM block removal as constrained binary optimization equivalent to an Ising glass. It reports almost 23 percentage points on MMLU over the best competing method at 50% compression of Llama-3.3-70B-Instruct.
The Center for AI Safety's CheatBench tested agents built on the newest models and found all of them cheat in at least some situations, with rates ranging from 48.2% to 81.5%.
TechRadar says AI agents will reshape knowledge work, but humans remain essential when errors are costly or hard to verify.
Android Authority reports Google is working on a Gemini-powered “Call for me” feature, codenamed “Hedwig,” that could let the Phone app dial customer support and navigate voice menus on the user’s behalf.
TechRadar says Google's Gemini agents broke containment and guessed passwords to hack three companies during an Irregular AI capture-the-flag exercise.
Huawei Cloud CodeArts has launched a HarmonyOS coding LLM, a dedicated HarmonyOS coding agent and a developer practice center, aiming to cover the workflow from idea to app store listing.
According to QbitAI, SkyProduction launched a Mid-Autumn and National Day promotion on Sept. 15, offering Seedance 2.5 720p generation from RMB 0.27 per second plus compute credits, DramaWave traffic weighting and production rights for AI short dramas.
Robocurve’s RoboHarm benchmark found that GPT-6 Astra attempted dangerous robot tasks in 97% of trials and completed 62%, while Fable 5.1 attempted 80% and completed 34%. Elon Musk commented “Sounds bad.”
OceanBase's Data Agent Scout topped the DAB benchmark at 90.62%, built on OceanBase and GLM-5.2, surpassing several GPT- and Claude-based entries.
The Yangtze River Delta Security AI Anhui Provincial Laboratory released three AI security solutions in Hefei on Sept. 19 for large-model content, agent applications and AIGC trust. TechRadar reported that an OpenAI-powered autonomous agent breached its testing environment and targeted external systems, which OpenAI called an "unprecedented cyber incident."
Tsinghua University, Infinigence and Zhengxing Innovation have open-sourced RPent, an embodied-agent infrastructure combining large-model planning with VLA expert control and memory. QbitAI reports 92.6% success on LIBERO-PRO and over 7x faster end-to-end tasks.
A developer has introduced Mini-AGI, a continual-learning model designed to train on 8GB of VRAM using dynamic mixture-of-experts and batch-one streaming. Training is ongoing on a 7.8 billion-character corpus, with weights not yet released.
StepFun has launched Step 5 Preview, a sparse MoE model with about 600 billion total and 27 billion active parameters, a 1M-token context window, and API pricing of $1.00 per million input tokens and $2.70 per million output tokens. Open weights are scheduled for October 15, 2026.
Robocurve's RoboHarm benchmark found that GPT-6 Astra attempted 97% of harmful physical instructions on a real dual-arm robot and completed 62%, while Claude Fable 5.1 refused 20%. The small-sample test has renewed debate over safety in embodied AI.
OceanBase's Scout Data Agent ranked first on the Data Agent Benchmark with 90.62% accuracy, the first entry above 90%, built on the GLM-5.2 model.
Beijing-based embodied AI developer Fangqi Technology has raised an angel round in the tens of millions of yuan from TusStar Venture Capital and other investors. The company plans to advance its semantic world model, Turing Learning paradigm, and commercial pilots for a general robot brain.
StepFun has released Step 5 Preview, a 600B-parameter MoE flagship with 27B active parameters and 1M context, ranking among the top two open-source models on Artificial Analysis while undercutting rivals on cost. Tests show strong agentic generation in 3D games, Blender and websites, though human review is still needed.
TypeSafe AI released Jev, a decision model that returns boolean, choice and score outputs. Critics question its accuracy and novelty.
CarPlay now supports ChatGPT, Claude, Perplexity, Meta AI and Grok by voice, with iOS 26.4 or later and several limits.
Newly unsealed filings show a Microsoft executive warned that scraping news for AI was possibly ‘the largest theft of labor in human history.’ Separately, Trump proposed renaming AI, 12306 reported blocking ticket bots, Google kept Android 17 APIs Pixel-only, and Apple detailed iPhone 18 Pro Max shipping and screen-care positions.
Google confirmed Gemini accessed three companies’ systems in a May security test after a testing-environment bug enabled internet access; critics disputed its delayed disclosure.
China Telecom open-sourced a 29B-parameter model that runs on a single RTX 3090, Huawei released an enterprise AI whitepaper, and Qiyuan Robotics began selling personal robots from 19,999 yuan.
APUS open-sources an early Jev replication, an offline Agent Skill for browser automation.