AI2Day Daily Brief — 21 Aug 2026

Good morning, it is Friday the twenty-first of August, and here is what you need to know: a robotic arm on a wheelchair, AI agents moving into your team chat, a fusion breakthrough, Apple tackling code-switching, and a twenty-seven-billion-parameter model you can run at home.

Start with the wheelchair. A small startup called ATDev, based in Orange, California, is building a powered wheelchair with a built-in robotic arm that can pick up a fork, open a door, or flip a light switch. The project is federally funded through ARPA-H and led by the University of Pittsburgh. ATDev is also working with Meta to apply Meta's open-source AI models to the chair's software. About three and a half million Americans use powered wheelchairs, according to the company, and the founders argue the hardware has barely changed while everything else around it got smarter. ATDev says it expects an FDA-approved, insurance-covered version within five years. That is the company's own timeline.

Meaningful if the timeline holds. Next, Slack has launched a feature called Slack Code, which gives AI coding agents their own dedicated project channels inside the workplace app. The idea is straightforward: instead of a developer working privately with an AI tool and then bringing the result back to the team, the whole process happens in the open. Product managers and designers can watch what the agent is doing in real time without touching a terminal. Partners including Anthropic, GitHub, Cognition, OpenAI, and Vercel are building integrations, so agents like Claude, Devin, GitHub Copilot, and ChatGPT can all operate inside these channels. Higher-stakes actions, like pushing code to a live product, require a human to approve before anything ships. Slack has not yet confirmed which subscription plans include the feature or which integrations are live right now.

Worth watching once those pricing details land. Now to fusion. A startup called Inertia Enterprises says it has cut the time needed to manufacture fusion fuel pellets from up to a week down to roughly two to three hours. These pellets are tiny hollow spheres of synthetic diamond packed with frozen and gaseous hydrogen. A commercial fusion plant would need to fire ten of them every second, so production speed is not a secondary concern, it is a core engineering requirement. One ingredient, tritium, costs around thirty thousand dollars per gram, so faster production is also a direct cost saving. Inertia raised four hundred and fifty million dollars to commercialise technology first demonstrated at the National Ignition Facility, the US government lab where the original experiment was designed by Inertia co-founder Annie Kritcher.

The manufacturing side of fusion does not get enough attention, so that is a notable step. Apple is out next. Apple ML Research published a study on improving voice recognition for speakers who mix Mandarin and English mid-sentence, a habit called code-switching. It is common among hundreds of millions of speakers worldwide and has long caused problems for voice assistants because clean mixed-language training data is rare and expensive to produce. Apple's approach uses unlabelled audio, recordings with no human-written transcripts, and lets the model generate its own transcripts, then trains on those across multiple rounds of self-improvement. The technique runs three phases: automated label generation, two-stage bilingual model training, and repeated improvement cycles. The paper is a research publication, not a product announcement.

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. Every mistake you make gets flagged and graded automatically, hand by hand, even on the free tier. Lifetime access is $99, in US dollars, until 15 September, at zero, the number two, hero, dot A I.

And finally, on Friday, Alibaba quietly released a model called Qwen three point eight, twenty-seven B on Hugging Face under an open Apache two point zero licence, meaning free to download and modify. The twenty-seven B stands for twenty-seven billion parameters. Compressed to around seventeen gigabytes, it fits on a high-end gaming PC or a recent Apple MacBook Pro with an M five Max chip. Independent benchmarking firm Artificial Analysis scored it fifty-two on its Intelligence Index, the same score as OpenAI's G-P-T five point six Luna cloud model at maximum reasoning. It hit three million downloads in its first three days. One real-world caveat: its default deep-reasoning mode can be slow. One reviewer waited twenty-one minutes for an answer to a simple image question.

For a deeper look at the week's stories, 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 this was useful, hit the thumbs up and subscribe, so the next one finds you.