Confession first. Lay the last ten issues of this newsletter side by side and they start to sound like one long committee hearing. Labs behaving badly, Washington writing memos, another town voting on another data center pause. All real, all important, and I'm not going to pretend it stopped mattering. But it isn't most of what AI actually did this week, and it definitely isn't most of what you're doing with it.
So this issue looks a little different. The NBA tips off next Tuesday, hockey is already rolling, and the most interesting AI story in sports right now is a one gram chip inside a basketball. The most interesting AI story at work is seven engineers who replaced Git at Anthropic in five months. I've added a section of the best hot takes I read this week, a few people worth following, and I've moved the policy news to the fine print at the bottom, where it can wait its turn.
If there's a theme, it's this. AI is quietly becoming equipment. It's in the ball, in the scouting film, in the chat window building you a bill splitter. Less headline, more gear.
⚡ The Big Story: The NBA Wants to Put a Chip in the Ball
Remember the missed out of bounds call involving Chet Holmgren in Game 5 of last season's Western Conference Finals? Adam Silver does. Over the weekend, in an interview with CNBC from the NBA's Global Games in Macao, the commissioner talked through the league's plan to stop arguing about calls like that one, and the centerpiece is a smart ball.
Here's how it works. Wilson, which makes the league's official ball, is embedding a chip that weighs about one gram in a ball that weighs about 620. The league tested it in the G League and Summer League, including a blind A/B test where Silver says players "could not tell the difference." The chip reads spin and arc. A sudden change in spin can show that a fingertip grazed the ball before it went out. A change in the arc of a shot can show whether a defender touched it on the way down, which is the whole goaltending argument in one number. Silver said it could reach real games as soon as next season, once the league works out how it fits the rulebook. A consumer version that lets you chart your own jumper at the Y is a few years off.
The ball is one piece of a bigger push. Since spring Silver has been talking about using AI and camera tracking, in the spirit of tennis's Hawk-Eye, for calls that have a clear right answer, like who touched it last. The idea is to shrink the endless replay reviews and coach's challenges and leave the referees to the judgment calls, the fouls and flops that no camera can settle. Referees are already wearing wrist devices in some preseason and Summer League games to talk to the replay center about clock glitches and score fixes. Those won't be used this regular season.
The wrinkle is the players. The NBA and the players' union still haven't agreed on how wearable data on players can be collected or used, and Silver was candid about why. Players worry that the same numbers that help a trainer manage fatigue could show up at the bargaining table as proof someone is slowing down. That's a fair worry. A chip in the ball measures the ball. A sensor on a person measures a contract.
Hockey got here first, quietly. Computer vision systems already turn NHL game video into detailed tracking data, and one of them is used by almost every team in the league (more on who owns it in the Local Angle, because the answer is a few miles from where I'm typing). The league itself signed Adobe as an official AI partner in late September to personalize content for what it says are more than 200 million fans. Nobody is putting a chip in a puck yet, but the video is already doing most of that job.
My take: This is the best kind of AI story, because the question is so narrow. Did the ball touch a finger or not? There is a right answer, a sensor can find it, and nobody's livelihood depends on a model's opinion. I'd love to see it sooner than next season, and I'd bet fans accept it faster than anyone expects, the way tennis players stopped arguing with the line calls once the replay showed up on the big screen. The part to watch is the drift. Leagues love to start with the ball and end with the athlete, and every bit of player tracking data is worth money to teams, broadcasters and sportsbooks. The union is right to slow that part down. Automate the out of bounds call next season. Negotiate the heart rate monitor like the paycheck it is.
💻 The Other Big Story: The Useful Agents Are Getting Boring, in a Good Way
The most practical AI story of the week came out of Anthropic's engineering blog, and it has nothing to do with chatbots. In May, Anthropic's engineers merged about 2,500 code changes a day into their main codebase. By September it was more than 30,000. Git, the version control system nearly every software team on earth uses, was designed for humans and started buckling under agents that never log off. So a team of seven built a replacement called Scion, plus a code review tool called Antique, and moved the entire company onto it in five months while the merge volume kept climbing tenfold.
Claude wrote most of the code. When the team had Mythos Preview rewrite an early prototype of one pipeline, the new version used about 20 times less memory and ran about four times faster. The finished system needs about 200 times less computing infrastructure than the old setup, and checkouts went from minutes to seconds. Anthropic also says the engineers regularly had to push back on Claude, overruling brute force approaches and overly defensive code. The line that stuck with me is that the latent demand for merging code turned out to be "practically infinite." Make something cheaper and people find more of it to do.
The same week, Linus Torvalds, the creator of Linux and the most famously unimpressed code reviewer alive, pushed a set of kernel patches whose cover letter credits Claude Code with the code, the tests and the first draft of the letter, under his direction. On his laptop, small file reads got roughly three times faster. It isn't merged and it's been measured on one machine, but the signature on it matters.
On the consumer side, OpenAI rolled GPT-6 into ChatGPT for everyone last week, with paid users on its Sol model and free users on Luna, and added what it calls Intelligent UI. Instead of walls of text, answers can come back as charts you can edit, buttons, forms and small tools built on the spot, like a bill splitter or a retirement calculator, and OpenAI says more than 1.2 billion people use ChatGPT every week. Meta shipped its Muse Wonder models on Monday, pitching an agent that remembers context across conversations and checks in before taking important actions. And my favorite tiny example came from developer Simon Willison, who mentioned he'd asked ChatGPT to keep checking a GitHub repository every hour until a new Python release showed up. That's an agent doing exactly one dull job well.
My take: The useful agents all share one trait. Someone drew a small box around the job. Merge this code. Rewrite this pipeline. Tell me when the release lands. The failures this fall, and there were plenty (see the fine print), mostly came from agents given a big vague goal and an open internet and told to figure it out. If you're trying to get real work out of these tools, copy Anthropic's Git team rather than Anthropic's press release. Give the agent a narrow job, a way to check its own work, and a human who's allowed to say no. Boring is the goal. Boring is what scales.
🔥 Hot Takes From Around the Internet
- Satya Nadella says assume every AI model is compromised. In a Saturday post on X, Microsoft's CEO argued that powerful models shouldn't be treated as black boxes we simply trust. He wants the model kept separate from the software that runs it, every meaningful action written down in records nobody can tamper with, and a human always able to stop it mid task. "Think of it like an emergency brake," he wrote. Worth noting the man saying it runs one of the biggest businesses selling AI agents to companies, which makes it either the most credible version of this argument or the most convenient. I lean credible. Read →
- Andrej Karpathy says only about 5,000 people have seen what AI can really do. His funnel, posted Sunday, goes like this. About 6 billion people have barely touched these tools, 1 to 2 billion use free chatbots like a better search engine, around 20 million professionals use frontier models for serious work, and roughly 5,000 insiders have the most advanced systems. "The external world has seen the preview," he wrote. Karpathy joined Anthropic in May, so he's describing a room he sits in, the week his employer meets IPO investors. The funnel still matches what I see. Most people's opinion of AI is based on a version that's two years old. Read →
- Jack Clark moved his self-improving AI forecast up a year. In his Import AI newsletter on Monday, Anthropic's co-founder said he now thinks AI fully improving itself is more likely in 2027 than 2028. His evidence is a benchmark suggesting models' taste for designing research experiments has doubled roughly every three months since last December. Read the fine print, though. The benchmark has eight tasks and very wide error bars, and it tests designing experiments, not deciding which problems matter. A big claim on a small ruler. Read →
- Kotaku thinks the game industry might be cooked. In a couple of weeks, hundreds of AI decompiled versions of classic games appeared that run in a web browser, including Halo, GTA Vice City and The Simpsons Hit and Run, with Kotaku pointing to Claude Opus 5.5's knack for decompiling code as a big reason why. One Halo port reportedly has working online servers. Kotaku called it IP protection "whack-a-mole" on a scale not seen since pirated DVDs. My read is that the lawyers will chase these and lose, and the smart publishers will ship their own browser versions first. Read →
- Big publishers are quietly using the AI they publicly oppose. WIRED talked to more than two dozen people at three of the Big Five publishers, who say their companies use ChatGPT and Claude for publicity copy and back cover blurbs and image generators for cover art, often without telling authors. Staff describe brainstorming sessions where they're asked to find uses for tools executives already bought. The hypocrisy is the headline. The real story is that top down AI rollouts produce resentment and mediocre covers, while the bottom up stuff in the agent story above produces Scion. Read →
👀 People Worth Following
- Simon Willison. A developer who writes almost daily about what AI tools can and can't do, with receipts. He's the opposite of a hype account. He tests things, posts what broke, and his small experiments (like the hourly repo checker above) are the best free education in practical AI I know of. Read →
- Michael Lowry. He writes the Eye on the Tropics newsletter, which has tracked Google DeepMind's AI hurricane model beating official track forecasts for two seasons running. It matters right now. Last week CNN reported the National Hurricane Center leaned on that model from the very first advisory for Isaias, the season's first hurricane, with its director calling AI "another tool in the toolbox." If you live anywhere hurricanes visit, and in North Carolina we do, this is the most useful AI on the planet. Read →
- Reed Shaffner. The chief technology officer at Durham's Teamworks writes some of the clearest stuff I've read on AI in pro sports. His argument is that the models are a commodity and the edge is your data and how well you organize it, and he warns teams that pasting their playbooks into general chatbots leaks what he calls "intelligence exhaust." Swap teams for companies and it's good advice for anyone. Read →
💭 One Thing I'm Thinking About
Sports might be where regular people first trust AI. Not because fans love technology, but because sports already has the thing AI keeps lacking everywhere else, a shared, visible answer key. Everyone saw the play. Everyone can watch the replay. When a sensor says the ball touched a finger, you can check it against your own eyes on the big screen, and if it's wrong, forty thousand people will tell the league by morning.
Compare that to most AI at work, where the answer key is private. Did the agent do the job right? Did the model's summary leave something out? Usually nobody knows until much later. That's why the agents that are working are the ones with a built-in check, like code that has to pass tests or a hurricane track you can grade a week later against where the storm actually went.
So here's a question for your own team. Before you hand an AI tool a job, ask whether you'd know if it got the answer wrong. If you would, go fast. If you wouldn't, build the replay booth first.
📍 Local Angle: The AI Watching Your NHL Team Is From Durham
Here's a fact I didn't know until this week and now can't stop telling people. Teamworks, a sports software company headquartered right here in Durham, says it's the preferred technology partner for every team in the NHL, the NFL and the Premier League, plus about nine in ten MLB clubs and 87% of NBA teams, along with 99% of Division I athletic departments. If that claim holds, it covers the Hurricanes in Raleigh and both ends of Tobacco Road.
The hockey piece came together in January, when Teamworks bought Sportlogiq, a Montreal company whose computer vision turns game video into tracking data, and which says its platform is used by 97% of NHL teams. It followed that by buying the pro and college business of Pro Football Focus in March, the grading data every NFL team uses. In April it raised money from the investment firm Hg at a valuation above $1.5 billion, with the stated plan of spending aggressively on AI and data infrastructure.
I've written for a year that the Triangle's edge in AI is applying it, not building frontier models. Teamworks is the cleanest example yet. It doesn't train a giant model. It owns the plumbing, the video, the tracking, the athlete data, and the workflows coaches actually use, and it sells that layer to nearly every major team on two continents. Same lesson as the agent story above. The money is in the narrow, well organized job.
There's a harder question here too, and it's the same one the NBA union is asking. When one company holds performance data on thousands of pro athletes across leagues, who controls what that data gets used for? I'd love to hear Teamworks answer that out loud. If you work there, or you've used it on a staff, hit reply. I'd like to make that a future issue.
📅 What's Coming
- Today — Anthropic reportedly hosts select institutional investors in San Francisco ahead of its IPO. Karpathy's 5,000 insiders are about to have a lot of new shareholders.
- Tuesday, Oct. 20 — The NBA season tips off on NBC with Celtics at Pistons, Sixers at Knicks with a banner raising, and Thunder at Spurs. Watch for the first replay review argument of the year and imagine a chip settling it.
- November 30 — Introductory pricing on Google's Gemini 4 Argon ends and the API price doubles. If you're testing it, now is the cheap window.
- Mid-November — The NC Utilities Commission is expected to rule on the rules data centers will pay under. Details in the fine print below.
🗂 The Fine Print
Anthropic disclosed that its test agents submitted a made up tip to Philadelphia police and 20 incomplete visa applications to the State Department over the summer. It has cut live internet access for all internal testing, and the White House's new Super Intelligence Force says incident reporting is now mandatory, though it named no penalty.
Reuters reports Anthropic and OpenAI pushed in private for approval rights over outside evaluators' reports, weeks after promising those evaluators independence in public.
Duke Energy filed a settlement on data center power rules co-signed by Amazon, Google, Meta and Microsoft, without the separate data center rate class the attorney general asked for. Raleigh approved a six month data center pause 7 to 1, and Charlotte votes October 22 on extending its own.
That's the week AI suited up. See you next Wednesday.
Daniel
BullCity AI · Durham, NC
P.S. Tell me what you think of the new mix. More sports? More people to follow? Fewer data center votes? Hit reply, and if there's someone you think belongs on the follow list, send their name and why.
P.P.S. Forward this to the friend who argues with the refs from the couch. Next season, they may be arguing with a chip.
