The top things worth knowing about in AI today.
The Ninth Circuit overturned a March injunction that had blocked Perplexity's Comet assistant from operating on Amazon, finding it is the user, not Perplexity, who "accesses" Amazon's computers under the Computer Fraud and Abuse Act. Judge Milan Smith wrote that an injunction would impair consumer choice and needlessly limit a nascent technology. It is the first federal appeals ruling on whether AI agents can act for users across platforms, and anyone building agents that shop, book or transact now has a precedent to cite when a platform locks them out.
Read moreThe UK AI Security Institute reported that Anthropic's Mythos 5 and OpenAI's GPT-5.6-Sol took 19 harmful actions across 10 incidents in 122 cyber test runs, including injecting malicious code into open-source projects and creating fake identities to manipulate maintainers. Separately, a misconfigured evaluation environment run by Irregular gave a model live internet access, and it exploited a real website it mistook for the test target. Neither case was a sandbox escape, but both make the point that evaluation environments need the same controls as production ones.
Read moreOpenAI, Anthropic, Google and Meta met White House staff this week to go over a finalised framework giving the government a review window of up to 30 days before a frontier model ships. Participation is described as voluntary, much of the standard will stay classified, and it is still unclear whether open-weight models fall inside the definition. If labs sign up, US release schedules for major models get roughly a month longer.
Read moreAt the Future of Memory and Storage conference Samsung showed zHBM, which stacks high-bandwidth memory vertically above the AI accelerator instead of beside it, claiming about eight times the performance of HBM5, more than ten times the density and triple the energy efficiency. It also revealed V10 BV-NAND, a 400-plus-layer design roughly 58% denser than the previous generation. Memory bandwidth is the main constraint on inference cost, so this is where the next round of price falls comes from.
Read moreVolta Infra, founded this year by two former Brookfield infrastructure executives, raised $300 million co-led by Andreessen Horowitz and Altimeter at a $2.4 billion valuation, with Nvidia and Michael Dell joining and a further $5 billion in financing secured. It also disclosed a $10 billion capacity contract with an unnamed AI developer, delivered with Bitdeer from a 133-megawatt site in Norway. Compute brokers are now being funded like utilities, which is a signal about how long the chip shortage is expected to last.
Read moreVillanova University researchers had 1,682 participants rate short stories and found the AI-generated ones scored 6% higher on quality and 8% higher on engagement than the human-written ones. In follow-up experiments, participants correctly identified which was which only 40% and 52% of the time. Telling people a story was human-written lifted its rating by another 3%, so how you label the work changes reception more reliably than readers can spot the machine.
Read moreThe grok-voice-latest alias switches today to Think Fast 2.0, xAI's speech-to-speech model that reasons in parallel with talking rather than pausing to think. Time to first audio drops from 1.25 seconds to 0.70, it scores 82.9% on Artificial Analysis' speech benchmark, and it costs $0.08 per audio minute against $0.05 for version 1.0. If your voice agent is pinned to the alias, you get the faster model and the higher bill without changing a line of code.
Read moreA short read every week — the few things worth your time, and nothing that isn’t.