This issue collects 25 AI updates from 2026-08-23, led by Frontier AI labs still won't say how they'd contain a rogue model, OpenAI urges California to strengthen AI safety bill SB 53 it once opposed, Inherent, founded by DeepMind alumni, says its AI ‘teammate’ just outperformed Anthropic and OpenAI at replicating research, Harvard's $699 startup bootcamp uses AI avatars of its instructors for pitch feedback.
A new study from Guidelight AI Standards graded five frontier labs — Anthropic, Google, OpenAI, Meta, and xAI — on their preparedness to contain a rogue model, finding few publicly documented response plans.OpenAI scored highest while Anthropic and Meta ranked lowest, as regulators in California and New York begin mandating safety disclosure.
OpenAI is calling on California to strengthen SB 53, an AI safety bill the company previously opposed, according to TechCrunch.The reversal marks a significant shift in OpenAI's stance on AI regulation and adds a new variable to California's legislative process.
Fujian Province held its artificial intelligence innovation and development exchange event in Fuzhou, where three major AI industry actions were unveiled on the same stage.Local media reports confirm the simultaneous release, though the specific names and details of the three actions have not yet been disclosed.
Inherent, founded by DeepMind alumni, says its AI ‘teammate’ just outperformed Anthropic and OpenAI at replicating research.Built by DeepMind alumni, British AI lab Inherent released Faraday, an AI agent whose ability to replicate scientific papers could be a stepping stone for innovation.
Databricks has showcased Omnigent, an open source AI agent orchestration layer, as reported by TipRanks.The open source approach could let developers self host and customize agent orchestration while drawing more enterprise workloads into the Databricks ecosystem.
AWS has open sourced Dogwood, a project aimed at establishing governance rules for AI agent tool calling.The move targets the growing complexity of agent ecosystems, where reliable and controllable tool use is becoming a key bottleneck for real world deployments.
让AI Agent「试玩」世界模型，长程目标评测有了新基准PlayWorld 手机新浪网.让AI Agent「试玩」世界模型，长程目标评测有了新基准PlayWorld 手机新浪网 Watch planning quality, tool use, permission control, and result verification.
An AI agent was used to hack a gym's system and jump the waitlist, according to Fox News.The incident highlights how easily autonomous agents can manipulate third party services, and it is renewing questions about the security boundaries of agent access.
ByteDance's Doubao is reportedly preparing an intelligent work product in the mold of Microsoft's WorkBuddy, with a decade of Feishu capabilities being componentized to support it.The move signals ByteDance is packaging its office software expertise into enterprise facing AI agent products.
Harvard Business School's new $699 HBS Foundry bootcamp uses AI avatars, created by HeyGen, to give entrepreneurs feedback during practice pitches and board meetings.The avatars impressed participants, though the real Jeff Bussgang calls his digital copy a little "creepy."
Alibaba announced plans to place new shares with non U.S.persons outside the United States, targeting total proceeds of HK$80 billion in its first placement since its 2019 Hong Kong listing.
According to finance.biggo.com, Nvidia has spent roughly $7 billion to enter the open weight model race, taking direct aim at DeepSeek and OpenAI.
A startup founded by several PhD students is betting on an integrated brain for bipedal humanoid robots instead of following Silicon Valley's playbook.In its latest demo, a robot drives a go kart through consecutive corners fully autonomously in a single unedited take.
实测DeepSeek Harness全自动开发FPGA工程能力 53AI.实测DeepSeek Harness全自动开发FPGA工程能力 53AI Watch real task performance, API pricing, context capability, and migration cost.
The 2026 AI Creator Festival concluded in Bali, bringing Wan3.0 and creators from around the world together around the new frontier of AI video generation.As The Manila Times reported, the event showcased how AI video tools are being applied in real creative workflows, offering a window into the growing creator ecosystem for video generation models.
OpenAI has convened a summit on the future of AI and mathematics, bringing together roughly 40 leading mathematicians, according to media reports.The gathering comes as major labs race to strengthen reasoning and math capabilities in their models.
An anonymous AI model called Ox Alpha has sparked global buzz, and reporting by South Korea's Chosun Ilbo traces its origin to a Chinese developer.No specific company has been named, and observers are watching whether the developer will formally claim the model or release technical details.
An OpenAI leader has warned that AI driven cyber attacks are entering "a different chapter," with threats set to become increasingly "persistent." The warning, reported by The Guardian, underscores growing concern inside leading AI companies about automated and agent powered offensive operations.
Shangwei New Materials' Qiyuan Robot opened pre orders on August 23 for its Qiyuan Q1 and Qiyuan T1 personal robots, with first shipments expected in September 2026.The move takes consumer humanoid robots from sample displays to real purchases, marking a key step toward home robot commercialization.
Google has confirmed that Gemini 4 will fully replace Gemini 3.5 Pro, retiring the older flagship from its model lineup rather than keeping the two in parallel.Developers and enterprises still running on Gemini 3.
An anonymously released large language model called "Niulai" has been traced back to Chinese AI firm Zhipu, with analysts citing telltale signs in its tokenizer, video encoding, and API error messages.Separately, speculation is growing that Cursor may have trained on the open source GLM models, though no direct evidence has emerged.
A robot has faced a human tennis player in what QbitAI calls a world first matchup, pulling off extreme ball saves and instantly springing up after falls.The "AstraTennis moment" highlights advances in motion control and dynamic balance that could ripple into service and industrial robotics.
OpenAI is again hiring AI coding infrastructure talent: the entire team of YC backed unicorn InstantDB is reportedly joining the company, per BlockWeeks.The report describes the move as a major technical upgrade and the latest step in OpenAI's push to absorb developer infrastructure teams.
The founding lead of Google's TPU (Tensor Processing Unit) program has joined Anthropic, according to a report from Chinese outlet Sohu aggregated via Google News.The move has fueled speculation about Anthropic's future compute and infrastructure strategy.
NVIDIA has told some of its biggest customers that AI servers delivered in early 2027 could cost 15% or more, with some GB300 and Vera Rubin 200 systems reportedly rising about 17%.A 1GW AI data center could see at least $5 billion in added costs as surging memory prices squeeze GPU supply chains.