News, Tools, Strategy, and Ethics of AI.
In Birmingham, the company that installs and maintains license plate cameras is not Flock. It is the electric utility, and the charge shows up as an electric service. Over seven years, one contract grew to $9.76 million without ever going back out to bid.
Home Depot says it does not authorize federal agencies to access its camera data. Oklahoma City made the same promise to its own vendor, in writing, as a binding change to the contract. One of these can be reversed at will. The other cannot.
A Cambridge professor patched a security hole in software the internet runs on. Automated probes for that exact bug showed up in his server logs about ten minutes later. The harder problem is that the AI tools good enough to help him defend it are the ones he cannot get.
Cities cancelled Flock camera contracts at a record pace in August. That happened, and we checked it. It is also under two percent of the network, and the contracts being signed elsewhere are getting bigger and adding drones.
A federal judge threw out the ban on Anthropic. The reason was not politics. It was that the government kept trying to buy from the company it had just called a national security threat, and wrote it all down.
A coding assistant scored a perfect 0.00% on its vendor-commissioned prompt-injection test. A researcher then got code execution out of it in most attempts, just by asking it to summarize a website. Both are true, and that is the lesson.
A 54-person unreviewed preprint went viral as proof AI rots your brain. Its own authors asked people to stop saying that. Meanwhile a much stronger study, on nearly a thousand people, found real harm and nobody shared it.
The same AI model can help or harm the same worker, depending on a configuration choice the employer makes. That choice comes with a duty most owners haven't thought about yet.
A randomized study found people harmed by AI didn't know it, and people who weren't helped by it were sure they had been. Why self-report can't answer this question, and what to check instead.
Fourth-grade reading in 2024 matched 1992. The adult literacy shift is real, but the timeline rules out AI as the cause, and part of it is the test changing, not the students.
Deepfake and voice-cloning fraud is real, but the $893 million number everyone quotes is mostly a consumer investment-scam problem. The business-targeting slice is smaller, and if it happens to you, the law says you probably eat the loss yourself, which is what makes prevention the only control worth trusting.
If an automation can spend, delete, or contact your customers and you haven't bounded what it can reach, you made a decision, not an accident. The controls that actually matter, what the frameworks require and don't, and the one insurance question worth asking.
Two failure modes that both feel exactly like competence: AI tends to agree with you, and it states false things with the same confidence as true ones. The research on both is more specific, and more useful, than the general complaint that "AI makes stuff up."
Six different coding tools, six different companies, the same handful of failures. An agent that warned itself and did it anyway, two unrelated products with the identical shell-quoting bug, and one viral story we could not verify and will not use.
Half the country is uneasy about AI, and the research says that unease is right, just miscalibrated. People trust it too little in one direction and too much in another, and watching it work more closely does not fix either problem.
Apple's own memory prices moved this year, and so did what a Mac Mini can hold. Real specs, real prices, and the arithmetic for how many billion parameters actually fit at a given quantization level, not a marketing number.
A Charlotte officer ran a plate for a drug target. A Mooresville officer ran 31 searches on her boyfriend's ex-wife. In every case the technology worked exactly as designed. What failed was a person deciding the search was justified, in a system that makes that decision cost nothing.
HOAs and businesses outnumber police departments on Flock's own camera network. What a camera actually costs per year, what the lease terms mean when you cancel, and why two cities found their contracts less solid than they thought.
A town spent 13 months in court to keep a corporation's water use secret from its own newspaper. It is not an isolated story: NDAs, shell buyers, and unpermitted power plants are how many communities learn about their data centers.
The bottle-per-query meme and the industry's "under 1 percent" defense are both true. They just answer different questions. Here is the number that actually matters, and how to read any AI vendor's water claim.
A data center's exhaust is too cool for a steam turbine, but it is warming thousands of homes in Denmark right now. It also needs enormous amounts of water to cool itself, real enough to fuel real backlash, and real rumors. The physics, the water math, and what's actually true.
A Wisconsin criminal complaint cited a driver's Flock-recorded trips to Michigan (where marijuana is legal) as part of its probable-cause justification to search his car. What the court records say, what Flock says about itself, and why it matters if a camera like this watches your parking lot.
The state published an AI roadmap in July; none of it binds a private business. What does bind you is older: the Unfair and Deceptive Trade Practices Act, a private right of action, and treble damages. Where the real exposure sits for a small shop, what is actually pending, and four things worth doing this month, none of which cost anything.
A gap is anything in how your business runs today that costs time or money without you noticing it. The step nobody owns. The thing typed into two systems. The follow-up that only happens if someone remembers. Here is what they look like, why they are invisible from the inside, and why the number attached to each one is the part that matters.
It sounds like a marketing line. It is arithmetic, and it constrains what we are allowed to quote you. When the work costs more than the gaps do, the scope gets phased instead of discounted, and sometimes the honest answer is that fixing something isn't worth it at all.
Utilization, realization, collection: three rates that look survivable alone and multiply into one much smaller number. A look at the workflow leaks that quietly drive all three down, and why none of them are software failures.
More notes are in the works. The other categories have plenty in the meantime.