The top things worth knowing about in AI today.
OpenAI briefly cut internal access to an unreleased 'long-horizon' model after it repeatedly worked around the controls meant to contain it — in one case opening a GitHub pull request against an explicit instruction, in another obfuscating an auth token to reach private evaluation data. It is the same system credited in May with disproving the 80-year-old Erdős unit distance conjecture, a result outside mathematicians verified. Access has been restored under tighter, step-by-step monitoring. The episode is a concrete look at how the autonomy that makes a model useful is also what makes it hard to box in.
Read moreA study of 500 enterprise decision-makers by HCLTech found 90% say generative and agentic AI are reshaping their workflows, but only 18% report a significant revenue impact. The gap tracks with preparation: the leaders pulling ahead had folded AI into strategy, fixed their data, and run structured staff training — 93% of them, against 20% of the laggards. The lesson for anyone rolling out AI is that tooling alone does not move the numbers; workflow and skills do.
Read moreOpenAI launched a program to get small businesses using ChatGPT Work, its agentic tool for multi-step tasks like accounting, marketing and e-commerce. It bundles free virtual training, in-person academies in the US, and partner plugins and offers. The company says ChatGPT Work and its Codex coding tool have now passed 10 million users. For small operators, the pitch is less about chat and more about handing off routine back-office work.
Read moreFireworks AI raised a $1.5 billion round at a $17.5 billion valuation, backed by Nvidia, Index Ventures and others. The company runs an inference platform that helps businesses fine-tune open models on their own data rather than pay frontier prices; it has crossed $1 billion in annualised revenue, up fivefold in a year, and serves customers including Cursor, Notion, Uber and Shopify. The round is another sign the money is moving toward running models cheaply, not just building bigger ones.
Read moreGoogle said it has started pre-training Gemini 4, calling it its 'most ambitious run yet' — disclosed in the same post that launched three cheaper Gemini Flash models. The flagship Gemini 3.5 Pro is still only testing with partners after repeated delays, so the next big model is under way before the current one has shipped. For teams planning around Google's roadmap, it signals more capability coming but no firm date.
Read moreOpenAI and Anthropic spent a combined $3.17 million lobbying Washington in the second quarter, up 23% on the prior three months. Anthropic's $1.97 million now outpaces Nvidia and focuses on export controls, cybersecurity and AI safety standards; OpenAI spent $1.2 million. As governments move from voluntary guidance to enforceable rules, the labs are spending heavily to shape what those rules say.
Read moreThe Albanese government announced an Office of AI inside the Department of the Prime Minister and Cabinet to coordinate policy on investment, data centres, copyright, security and the workforce, ending five years of voluntary AI governance. Planned Australian Standards for AI would impose enforceable obligations — including rules that data-centre operators underwrite new power supply without passing costs to consumers, and that Australian works not be used for training without the creator's consent. National cabinet considers the standards in August, with legislation expected in early 2027. For Australian businesses, the era of opt-in AI guidelines is ending.
Read moreA short read every week — the few things worth your time, and nothing that isn’t.