AI2Day Daily Brief — 26 Aug 2026

Good morning, it is Wednesday the twenty-sixth of August, and today's big story is that OpenAI has built its own AI chip, made with Broadcom, and the numbers it is publishing are worth paying attention to.

The chip is called Jalapeño, and it is what engineers call an ASIC, a processor built for one job: running finished AI models fast enough to answer your questions and power AI agents. OpenAI says in benchmark tests it delivers between one and a half and nearly two times more AI work per unit of electricity than comparable Nvidia superchips, and cuts end-to-end response time by up to three and a half times across three models tested. Those are OpenAI's own figures. The practical upshot is shorter waits and lower serving costs. Jalapeño ships in small numbers before the end of this year, with a broader rollout through twenty-twenty-seven. Nvidia is not going anywhere; this sits alongside it, not instead of it.

Faster inference at lower cost is the thing every AI company is chasing right now, so that timeline will be worth watching. Meanwhile, China hosted a genuinely strange spectacle this week.

The second World Humanoid Robot Games ran in Beijing from the twenty-second to the twenty-sixth of August, backed by Chinese state media and the Beijing city government. The headline result: at least one sprinting robot posted a hundred-metre time faster than Usain Bolt's world record of nine point five eight seconds. The fastest any human has ever run that distance. The same footage, however, showed robots crashing into barriers, collapsing mid-race, and in at least a few cases catching fire. When the events shifted to practical tasks, washing and hanging laundry, the robots ran noticeably slower than an average person would. So the picture is genuinely mixed: raw speed in a controlled sprint is not the same as useful capability in a messy real-world setting.

A useful reminder that benchmark conditions and kitchen conditions are very different things. Staying with big-picture AI concerns, Bill Gates published a new essay this week and gave an interview to MIT Technology Review, and he was visibly agitated.

Today's sponsor is Zero2Hero: an AI poker coach sits at your table, and you can ask it mid hand whether the player across from you is bluffing. It runs in your browser on any computer. Windows, Mac, Linux, even a Chromebook. Nothing to install. Lifetime access is $99, in US dollars, until 15 September, at zero, the number two, hero, dot A I.

Gates argues that AI has already crossed danger thresholds in five areas: helping design biological weapons, enabling cyberattacks, influencing human psychology, displacing jobs, and concentrating economic power. He estimates that an AI-enabled bioterrorism threat is roughly fifty times more likely to cause mass harm than a natural pandemic. That is his estimate. On jobs, he says current economic statistics are masking how fast white-collar displacement is already happening. He is proposing two specific policy ideas: legally reserving certain roles for humans, and taxing the use of AI tokens and robots that replace workers. He also says protesting data centres is not an effective response, because AI development will continue globally regardless of what any one country does.

Whether you find his proposals practical or not, the underlying data points on job displacement are worth looking at independently. Closer to home for commuters, London's robotaxi plans have slipped.

Uber and the London-based company Wayve were granted the first minicab licences in the capital specifically for self-driving passenger rides, which was a genuine regulatory milestone. The catch is that a human safety driver still sits behind the wheel. Both companies had publicly promised paying passenger trips later this summer. That window has now closed. According to The Guardian, the holdup is regulatory: the UK government has not yet published clear rules telling companies exactly what their vehicles must demonstrate to qualify for fully unsupervised operation on London roads. Until those rules exist, no one can meet them. The UK had positioned London as an early showcase for self-driving transport, so this delay matters beyond just the two companies involved.

Regulatory clarity tends to move slower than engineering ambition, which is a pattern we see everywhere in this space. Finally this morning, Apple's research team published something that addresses a structural problem in multimodal AI.

Apple ML Research this week introduced STARFlow Two, a model that generates text and images inside a single unified architecture. Most systems today bolt two separate engines together, one for language and one for pictures, and that mismatch creates quality trade-offs and coordination problems. STARFlow Two uses a technique called a normalizing flow for image generation, which works the same mathematical way as the language engine underneath it, so the whole system stays consistent. The target use case is an AI assistant that can hold a conversation, reason about what you say, and illustrate its answers in one smooth response, without switching between tools. This is a research paper, not a product, but it points at a real limitation in current assistants that Apple is clearly working to close.

For a deeper look at all five of these stories and everything else moving in AI right now, catch the full Monday show, AI Today Weekly. That is your AI briefing for today. Every story is at A-I-2-Day dot live. That is A, I, the number two, D-A-Y, dot live. We are back tomorrow morning. If you got something out of this, a thumbs up and a subscribe genuinely helps.