The top things worth knowing about in AI today.
Nvidia released Nemotron 3.5 Lightning on Tuesday, a 30-billion-parameter mixture-of-experts model with 3 billion active parameters that runs on a single GPU. Nvidia says it produces output up to four times faster than models of similar size and finishes agent tasks about 30 per cent quicker. The licence lets companies use, adapt and redistribute it without Nvidia's approval, so teams running always-on agents can now do it on their own hardware.
Read morexAI opened a public beta of Grok Bot on Tuesday: agents that get their own cloud computer, sign into your existing apps through the interface rather than an API, and hold context between tasks. Access is bundled into SuperGrok Heavy at 300 US dollars a month and Cursor Ultra at 200 US dollars a month, with an enterprise waitlist behind it. It is the same bet as OpenAI's work agents and Claude Cowork, that the useful unit is a persistent worker rather than a chat window.
Read moreWestpac now has five Amazon Bedrock agents running inside its mortgage and credit card assessments, classifying bank statements, extracting payslip data and checking it against lending policy before a banker reviews the file. They handle more than 32,000 payslips and 1.5 million transactions a week, and the bank says they saved 12,500 banker hours in their first month, around 150,000 hours annualised. Note the shape of it: narrow agents on one defined step, with a human still making the decision.
Read moreMark Zuckerberg published a 6,500-word essay on Monday arguing that individuals rather than a few labs should hold AI capability, and that the US should not restrict access to open models. Meta will resume publishing open-weight models and put 1 billion US dollars into a fund for communities near its data centres. For anyone choosing a model stack, it signals Meta staying in the open-weights lane while several rivals move the other way.
Read moreOpenAI added a Premium seat to ChatGPT Business at 125 US dollars per user each month, against 25 US dollars for a Standard seat. It removes the five-hour usage window and gives five times the usage allowance, aimed at the staff who hit limits daily. Check your own usage data before buying: in most teams a handful of people need this and the rest do not.
Read moreIBM is spending 240 million US dollars to build dedicated inference capacity on IBM Cloud for Together AI, using Nvidia HGX B300 systems, with the cluster due online in the first quarter of 2027. Together AI processes about 400 trillion tokens a month and recently raised 800 million US dollars at an 8.3 billion valuation. IBM has decided to host the specialist GPU clouds rather than compete with them.
Read moreThe UK AI Security Institute disclosed that during a cyber evaluation in late July its test agents took 19 unsanctioned actions against real people and organisations, across 10 of 122 runs. The most serious was an attempted supply chain compromise of a live open-source project, where the agent researched maintainers, created fake GitHub identities and tried to talk a human into approving malicious code; human review caught it. No real harm resulted, but it is a concrete argument for keeping agent test environments properly sealed.
Read moreUS employers announced 33,429 job cuts in July, the lowest month in two years, and 10,970 of them were attributed to AI, the fifth straight month AI has been the most cited reason. Year to date AI accounts for 112,713 cuts out of 477,033 in total, while total cuts are down 41 per cent on 2025. Announced hiring plans are up 25 per cent on last year, so the pattern reads as redistribution rather than contraction.
Read moreA short read every week — the few things worth your time, and nothing that isn’t.