This issue collects 41 AI updates from 2026-09-03, led by India's richest man now wants to turn aging computers into AI ready PCs, empirik.ai emerges from stealth with $21M to build the 'AI Agent for Infrastructure Change', OpenAI's Astra becomes its first AI model with 'Critical' hacking abilities, DOJ urges judge to rule for OpenAI and Microsoft in New York Times copyright lawsuit.
TechCrunch reports that Jio — the company controlled by India's richest man — is betting it can turn aging computers into AI ready PCs for as little as about $11 for two months.The plan is aimed at price sensitive users and the vast installed base of older machines as a cheap on ramp to AI.
Broadcom beat market expectations in its latest quarter, with revenue up 86% year over year to $29.59 billion on the strength of AI semiconductors.The company now expects AI chip sales to roughly double to about $115 billion in fiscal 2027 and reach $230 billion in fiscal 2028, with Anthropic and OpenAI custom chip deals at the center of that growth.
TechCrunch Disrupt 2026’s new Real World AI Stage features Nvidia, robots, and extinct animals.On our new Real World AI stage, we’ll be focusing on the intersection between the digital and physical, and all the ways we’ll continue to see a blending of the two.
Sohu reports that large models from Alibaba Cloud, Zhipu and other providers have opened up on Tmall, where users can buy tokens as easily as topping up a mobile phone.The move takes AI API credits onto mainstream e commerce shelves and could lower the barrier for individuals trying large models.
empirik.ai has emerged from stealth with $21 million in funding to build what it calls the AI Agent for Infrastructure Change, as reported by Yahoo Finance.The funding targets a high stakes corner of enterprise IT, where AI agents could take over the planning and execution of infrastructure changes that have long been manual and risky.
PCMag reports that a group of laid off developers has created an AI model designed to take over CEO and other executive roles.The story highlights how AI agents are moving from task automation toward management decisions, while raising fresh questions about accountability when an AI executive gets something wrong.
A report from capacityglobal.com, filed under the Datacloud USA banner, says a $500 billion pact involving Nvidia spotlights a structural gap in financing data centres across a 20 year horizon.The piece argues that AI compute demand is turning data centres into multi decade capital projects where financing tools built for shorter cycles no longer fit.
Eddie AI, an AI video editing company, has launched an agent marketplace whose editing agents are built on the experience of great filmmakers, per No Film School.Packaging editing craft into selectable agents pushes video editing from a software chore toward a style driven agent service, and brings the "agents plus marketplace" model into creative tools.
TechCrunch reports, citing sources, that Palo Alto Networks has paid roughly $500 million to acquire Console, a Thrive backed startup in AI IT service automation.If confirmed, the deal would leave Sequoia backed Serval as the de facto startup leader in the category, a sign of rapid consolidation in the young market.
ByteDance's AI assistant Doubao is heading back into the phone market in the form of AI agent devices, according to a Pandaily report published on September 3.The move signals that ByteDance wants its assistant to go beyond apps and compete at the level of dedicated agent hardware.
Listed banks' half year reports are exposing a new competitive front: AI, with token consumption surging and AI agents going live in batches, according to a Red Net report.The disclosures suggest large model usage inside Chinese banks is shifting from pilots toward measurable, production scale workloads.
A report on September 3 says Meta is imposing its new AI agent on employees, rolling the tool out across the company.The report does not name the agent or detail its capabilities, but a mandate style rollout shows Meta is hardening its agent technology inside real workflows.
ByteDance is set to receive a US$29.6 billion syndicated loan — the second largest deal of its kind in Asia this year — according to a report from Jiemian News.The financing gives the AI and content giant deep reserves for its push across Doubao, the Seed model family, and Volcano Engine at a time when the large model race is burning through capital.
NVIDIA announced on September 3 that it has agreed to acquire Hugging Face for approximately $12.9 billion, bringing the platform that hosts more than 3 million models under the AI chip giant's roof.The deal, confirmed by TechCrunch, raises immediate questions about how much independence the open source hub will keep and how it will reshape developer access to AI.
OpenAI's Astra has become the first of the company's AI models to demonstrate hacking abilities rated Critical, according to a Decrypt report.The rating marks a step change in what OpenAI's own models are measured to do in security sensitive tasks, sharpening the question of how such capabilities should be governed and deployed.
Google has rolled out Gemini 3.8 Flash Cyber, a security focused version of its Gemini 3.8 Flash model aimed squarely at vulnerability patching.
Google has added Gemini 3.8 Flash to AI Mode, according to Search Engine Journal, bringing the newest Flash series model into its AI powered search experience.Flash tier models typically mean faster responses and lower per query costs, making the update the latest step in Google's ongoing refresh of AI Mode's underlying model.
TechCrunch reports that OpenAI's upcoming Astra model will use a technique called 'recurrent depth,' allowing it to reason outside the sequential chain most reasoning models rely on.AI safety experts are concerned that once reasoning can no longer be traced step by step, misalignment and misuse will become far harder to catch.
Meta has released Muse Spark 1.3, which it calls its most powerful AI model yet, with Chief AI Officer Alexandr Wang saying its coding beats OpenAI's GPT 5.6 Sol and is on par with Anthropic's Claude Fable 5.
Chinese media report that Google's Gemini 3.8 has made its debut, with the news spreading quickly after publication.In an early take, Yao Shunyu says the model delivers a huge leap on the RSI metric, though concrete capability details are still thin.
Industrial Bank and China UnionPay have jointly launched an AI credit card, marking the latest push to bring AI into China's card business.The collaboration between a major lender and the national card network signals that AI credit cards could spread from a single bank's innovation to an industry trend.
Chinese robotics company Zibianliang (自变量) unveiled TwinDex, a dexterous operating system whose robot ran a continuous chemistry experiment — 24 sub actions across three tools — after post training on a few hundred bodyless demonstrations and zero real robot teleoperation data.The three finger hand pairs with an isomorphic wearable capture rig that produces roughly 5.3x the usable trajectories per unit time, and experiments show bodyless data can substitute for nearly 100% of teleop data in training.
Hong Kong stocks gapped higher at the open on September 3, with JPMorgan raising its target price for Chinese AI company Zhipu to HK$2,000.The move signals institutional confidence in one of China's leading foundation model developers and offers a fresh read on sentiment toward Chinese AI assets.
Chen Danian, the Shanda co founder behind WiFi Master Key, has returned to the AI race with a new startup, StartLux, which just unveiled its first local model, StartLux V1.0 27B Preview.The 27B parameter model scored 39.
Google is now offering university students in the United Arab Emirates one year of free access to its Gemini AI, according to a report from Gulf Business on September 3.The student focused promotion extends Google's push to seed Gemini usage among younger users in the Middle East education market.
NVIDIA says 26 new games are coming to GeForce NOW this month, led by NBA 2K27 with the DLSS 5 3D Guided Neural Rendering feature.Built with Visual Concepts and 2K, DLSS 5 brings a new level of lifelike lighting and material detail to the court, marking the newest AI rendering technology's arrival in cloud gaming.
Fine tuning a 350M Model for Better Structured Outputs in 100 GRPO Steps.According to huggingface.co, Fine tuning a 350M Model for Better Structured Outputs in 100 GRPO Steps.
Training a coding model to paint watercolours with TRL and OpenEnv.According to huggingface.co, Training a coding model to paint watercolours with TRL and OpenEnv.
On September 3, Google DeepMind and Google Research released WeatherNext 3, a new AI weather forecasting model that predicts key variables at 5 kilometer resolution, is 60% better at rain evaluations than WeatherNext 2, and produces hourly forecasts.Google says the model will feed weather information shown in Search, Google Maps and Gemini and will be available to users and researchers on its cloud platforms.
The U.S.Department of Justice has urged the judge in the New York Times' copyright lawsuit against OpenAI and Microsoft to rule in favor of the two companies, according to The Washington Post.
The United States is urging G 20 countries to permit AI training on creators' work, a permissive copyright stance reported by The Straits Times.The push pits Washington against creator groups and stricter regimes, and its outcome could shape cross border rules for generative AI training data.
Top U.S.tech CEOs and Trump administration officials pressed the G20 Innovation Ministerial in Chapel Hill on September 2 to embrace lighter AI regulation, arguing that excessive restrictions would slow innovation.
U.S.lawmakers have unveiled a new bill to secure AI agents, Axios reports, in direct response to the breach involving OpenAI on Hugging Face.
Security vendor ONEKEY has launched an AI agent purpose built for automated firmware security analysis, per Security Today.The release targets a discipline that has long been slow and heavily dependent on human experts, marking another step in the spread of agents into specialized security workflows.
Zhaogang Group has teamed up with Tencent WorkBuddy to launch NextB2B, billed as China's first AI agent for the B2B trading industry.The release pairs an industrial trading platform with an enterprise agent product, signaling that AI agents are moving into core transaction workflows.
South Korean security vendor Capsule Security has released an AI agent behavior protection system built on NVIDIA's Nemotron models.The launch targets the behavioral risks of autonomous agents, adding a defense layer for enterprises as agents gain tools and account permissions.
Precise Biometrics has won a design win with PamirAI, which will integrate its FPC AllKey Pro biometric system into a consumer Linux based Agent Computer that hosts always on AI agents and is slated for fall 2026.The device pairs long running agent tasks with biometric approval for sensitive actions, with Seeed Studio handling production.
Hugging Face has open sourced funes, a durable memory layer that gives coding agents such as Claude Code, Codex, pi, and Hermes a searchable record built from the sessions already on your machine.It runs fully local by default, and can be bound to a private Hugging Face dataset you own so the memory follows you across devices and agents.
Amber International Holding Limited (Nasdaq: AMBR) reported second quarter revenue of US$13.9 million, up 38.8% quarter over quarter, with a 79.
Attacks targeting Langflow, an open source AI agent building tool, are surging, according to a new BankInfoSecurity report.The uptick underscores how agent development tooling has become an attractive target as enterprises rush to deploy AI agents.
On September 3, Hcompany introduced NeoMME, a family of 260M and 800M parameter multilingual multimodal encoders that process text and raw image patches in a single bidirectional Transformer, with Apache 2.0 weights and day zero Hugging Face Transformers support.The NeoMME Retriever variant tops every model under 800M parameters on the ViDoRe v3 document retrieval benchmark and compresses late interaction index storage from about 1.
