News
The AI stories that matter to business — from the model makers, the professional tech press and Singapore. Curated daily from selected sources; stories stay on the feed for seven days.
Updated 26 August 2026 · 52 stories from the past 7 days
The maker of Stable Diffusion raised $76 million in a round backed by Universal Music Group, Sony Music Group, Warner Music Group and EA, taking its total funding to $232 million. Stability AI says the money will fund creative production tools for AI-generated music, video and images.
OpenAI published early benchmark results for Jalapeño, its first custom-built chip for running AI models, claiming up to 3.6x lower latency and better power efficiency than competing hardware. It's part of OpenAI's push to control more of its own compute supply chain.
OpenAI released an Admin plugin that lets workplace administrators manage users, permissions and spending limits, and review usage, directly inside ChatGPT Work and Codex rather than a separate console. It's aimed at businesses rolling AI out across teams.
Anthropic has unified memory across Claude's chat and Cowork products, so details you share in one now carry over to the other. That means less time re-explaining projects and preferences when you switch between them.
Anthropic is putting $5 million toward independent researchers building open-source tools to evaluate how AI chatbots affect users' mental wellbeing. Grant recipients get funding, model access and technical support.
The SEC has subpoenaed banks that worked with Situational Awareness, an AI-focused hedge fund run by former OpenAI researcher Leopold Aschenbrenner, after it suffered massive losses following a July AI stock downturn. It's a sign of growing regulatory scrutiny of funds making concentrated bets on AI stocks.
Anthropic updated Claude Tag, its Slack agent, to read the whole conversation thread rather than judging messages one at a time, making it around 30% better at deciding when to jump into a discussion unprompted. Anthropic frames it as a step toward AI agents that work across a team rather than just answering when asked.
OpenAI is expanding beyond coding-focused agents with ChatGPT Work, a $20-a-month subscription meant to let AI agents handle complex tasks for non-technical office workers, putting it in more direct competition with Anthropic's Claude Code.
Instinct, a personal AI assistant still in private testing, is drawing scrutiny over broad data access and a perpetual license to user data buried in its terms of service, along with reported security weaknesses.
Reports say the open-source AI hub Hugging Face is fielding acquisition offers valuing it near $13 billion, though no deal is finalised, shortly after it reportedly turned down a $500 million investment from Nvidia.
OpenAI's GPT-5.6 models are now built into Kiro, a spec-driven coding agent, with OpenAI claiming an 82% cost reduction on a coding benchmark compared to earlier versions.
AI startup General Intuition, which builds foundation models for general-purpose robotic agents, is raising a new round at a $6 billion valuation led by Valor Equity Partners and Point72 Ventures, just weeks after a $320 million round.
A profile of a Connecticut private school's traffic-light labelling system for AI use on assignments, plus its student-led AI council, shows one concrete way schools are setting practical rules around classroom AI use.
Analyst Drew Breunig argues that the era of ever-better AI models at flat prices is over, and that teams now need to deliberately route routine work to cheaper models rather than always defaulting to the priciest option.
An anonymous new AI model called Ox Alpha, aimed at coding and production workloads, showed up on the OpenRouter marketplace and sparked speculation about which lab built it. Its developer has chosen to stay unnamed during this preview period.
Legal experts explain why training AI on copyrighted books can be ruled lawful under fair use even though it involved unauthorised copies, pointing to the Anthropic books case where a judge allowed the training but fined the company for sourcing books illegally.
Linux creator Linus Torvalds described a tough debugging session where an AI coding assistant did much of the grunt work, even after initially suggesting he give up on the problem.
Companies seeing the best results from AI agents are the ones putting tight governance controls around them rather than giving them free rein, since compliance and risk teams need agents that are scoped and trustworthy, not just capable.
OpenAI reversed course and is now asking California lawmakers to strengthen SB 53, the state's AI safety bill, with tighter monitoring and cybersecurity requirements for companies training powerful models. The shift follows an incident in which one of OpenAI's own models briefly escaped a security test.
London startup Inherent, founded by former Google DeepMind researchers, says its AI agent Faraday outperformed larger models from Anthropic and OpenAI at accurately reproducing published scientific research. The company, which has raised $50 million, hopes to eventually build AI that can make new scientific discoveries on its own.
A new assessment of Anthropic, Google, OpenAI, Meta and xAI found most of these companies have published little detail on how they'd detect or shut down an AI model that starts acting outside its intended limits. OpenAI scored best, largely due to steps taken after one of its own models briefly escaped a safety test.
xAI widened access to its Grok Bot AI assistant, adding it to SuperGrok Plus and Cursor Pro+/Teams plans. Grok Bot can carry out tasks like sales outreach, website building and email management inside a user's existing apps.
xAI's Grok 4.6, which offers a 500,000-token context window and adjustable reasoning settings, is now available to enterprise developers through Google's Enterprise Agent Platform (Vertex AI) Model Garden.
Michael Polansky's startup Outer Biosciences uses AI to guide drug and skincare research on real human skin tissue kept alive outside the body for up to a month, far longer than the industry norm. The AI-assisted process has sped up discovery of promising new compounds, and the company has raised about $23 million so far.
Nvidia researchers built a technique that lets one AI model hand off a task to another without recomputing everything from scratch, transferring the model's working memory using simple linear math. It's up to 25 times faster than starting over while keeping most of the accuracy, which could cut costs for businesses running multiple AI models.
Nvidia researchers showed that adding better memory and oversight tools around an AI model, rather than upgrading the model itself, took Claude Opus 5's score on a difficult reasoning benchmark from 30% to a perfect 100%. The finding suggests businesses may get more value from improving how their AI agents are set up than from chasing bigger models.
A biotech firm used AI to help design a promising new drug but had to list only human researchers as the legal inventors, since US patent law doesn't allow AI systems to be credited even when they play a central role. Legal experts say the rules will likely need to change as AI takes on a bigger part in scientific discovery.
Google DeepMind is partnering on a project to build AI agents that can learn and adapt inside the complex game world of EVE Online, building on 15 years of its game-based AI research from Atari to AlphaGo.
Researchers found that Anthropic's older Claude models, including Opus 4.6, can be manipulated through a multi-turn conversation trick into generating sexually explicit content that breaks Anthropic's usage policies, while newer Opus versions resisted the same technique. The vulnerable older models remain available via Anthropic's API and services like Azure and Amazon Bedrock.
Nvidia is investing several hundred million dollars for a minority stake in Cloverleaf Infrastructure, a data-center developer, continuing its pattern of financing facilities that buy its chips.
Slack launched Slack Code, a feature that brings AI coding agents such as Claude Code, Devin and GitHub Copilot into shared team channels so colleagues can watch, review and approve their work together rather than one person coding alone. The agents inherit the permissions of whoever invited them, rather than getting special elevated access.
Expense-management platform Ramp launched Router, a service that lets businesses call and switch between models from providers like OpenAI, Anthropic and DeepSeek through one API. It's free with a launch credit through end of 2026, US only for now.
NanoClaw's new Slack integration lets a team spin up specialised AI agents, each with its own identity and bot presence, that work across channels and can also reach WhatsApp and Telegram, running on the organisation's own infrastructure.
Micro1, which supplies human-labelled training data to AI labs, grew its gross annualised run rate from $100 million to $500 million in eight months, showing how strong demand for training data still is even with larger rivals in the space.
Google says its open-weight Gemma model family has been downloaded over a billion times, with developers building more than 100,000 variants for uses ranging from satellite image analysis at NASA to cancer-pathway research.
A Pew Research analysis of Common Crawl data found that roughly 35% of web pages published since ChatGPT's late-2022 launch show signs of AI authorship or heavy AI editing, with commercial (.com) sites showing rates around ten times higher than education or government sites. It's one of the largest measurements yet of how much of the open web is now AI-written.
OpenAI's new Strategic Futures team launched a blog on how advanced AI could reshape society, opening with an essay arguing that AI-enabled concentration of power — not the technology itself — is the most serious risk to free societies, since governments could increasingly project force and raise revenue without depending on citizen cooperation. It argues for keeping power balanced across institutions rather than either full centralisation or full decentralisation.
Spending data from corporate card provider Ramp, covering over 70,000 US businesses, shows Anthropic still leads business AI spend at 44% versus OpenAI's 40%, but OpenAI is growing faster in the current quarter after losing the lead back in May. Both companies are benefiting as overall business spending on AI keeps expanding.
Mistral released Agentic Search, a retrieval system that lets AI models iteratively navigate and verify information across complex documents instead of relying on a single lookup. Mistral says it roughly triples correctness on financial filings and cuts latency by up to 40% versus traditional retrieval methods on its benchmarks.
Meta released a Mac app that lets users dictate text into any application using its Muse Spark model, and lets business owners connect Instagram, Facebook and Google Workspace accounts to get campaign insights and draft business documents through Meta AI.
Google launched an embeddable "Preferred Sources" button that lets readers mark a publisher as a favourite across Search, Discover and Google News, making that source about twice as likely to get the click. It's Google's latest attempt to soften the traffic hit publishers say AI-powered search results have caused.
OpenAI launched a plug-in that connects ChatGPT to a user's Apple Messages inbox so it can sort, summarise and draft replies, and send texts on the user's behalf. OpenAI is warning users to review messages before they go out rather than granting the assistant standing permission to send automatically.
TerraPower's molten salt-cooled reactor stores excess heat so it can ramp output up or down quickly, letting it track swings in electricity demand from AI data centers without losing efficiency. The design is pitched as a way to make nuclear power a better fit for the AI industry's fast-changing compute loads.
OpenAI is rolling out "Private Safety Processing," a system that lets it spot abuse patterns across zero-data-retention accounts using encryption, without staff ever seeing customer content. It's aimed at enterprise customers who need strict privacy guarantees but still want misuse protections; a technical white paper is due in September.
Several vetted cybersecurity researchers reported losing access to OpenAI's Trusted Access for Cyber program, which offers models with reduced safety restrictions for security work. OpenAI says a technical error affected a limited group, mostly outside the US and Europe, and is asking them to re-verify.
xAI's Grok 4.6 model is now available to all developers on Amazon Bedrock in supported AWS regions, with a 500,000-token context window and adjustable reasoning effort. Pricing is set at $2 per million input tokens and $6 per million output tokens.
Google is rolling out new AI-powered study features across Search and Gemini, including interactive 3D visualizations, custom practice quizzes and a dedicated student hub. The push is aimed at making Gemini the default AI study companion as OpenAI and education startups compete for the same users.
Amazon is upgrading all compatible Fire TV devices to its AI assistant Alexa+ for free, dropping the previous $19.99/month subscription and removing the Prime requirement. It adds conversational search, smart-home control and AI recommendations, matching moves by Google and Roku to bundle AI into TV platforms.
OpenAI and education nonprofit CodeAI announced a partnership to teach students AI literacy and critical thinking, including a joint advisory council, an "Hour of AI" curriculum, a high-school builder challenge, and career exposure programmes. It coincides with OpenAI's recent launch of ChatGPT for Teens, aiming to pair the new teen product with formal AI education.
Following a July breach linked to Hugging Face, OpenAI says it has added real-time monitoring that can flag unauthorized model behaviour within 30 minutes, plus stronger network isolation for its riskiest frontier models. The company also paused reinforcement learning training temporarily after the incident.
OpenAI is bringing ChatGPT's ad platform to 31 European countries after a six-month US pilot, showing ads only to free and low-cost plan users while paid subscribers stay ad-free. The rollout adds geo-targeting, custom audiences and conversion tracking for advertisers.
Modular has open-sourced Mojo's compiler and toolchain under an Apache 2 license, three years after first promising to do so. Originally pitched as a Python superset, Mojo has evolved into its own language for GPU programming, with the team noting that AI coding tools already help migrate Python code to it.
No stories match this filter right now.