XLE

My Roomba breached containment (I left my door open, but told it verbally that the door was closed)

For purposes of this article, "open weights model" just means "you can run it on your computer instead of someone else's cloud".

Anthropic, meanwhile...

...sells controlled access to centrally hosted models. A regulatory system built around expensive pre-release evaluations, continuous usage monitoring, identity verification, and revocable access naturally favours large providers of proprietary models. Anthropic already possesses the compliance, legal, and evaluation infrastructure needed to satisfy such a regime. A university lab, startup, or community does not, even if they somehow manage to train a sufficiently capable model.

So it's very clear that regularly capture (and more broadly, capture of the consumer market on thinking thoughts) is the goal.

I don't think you understand their point. The point is that the OCR can be done on a distributed system, because a distributed system is what the cameras already are.

My biggest contention with this claim is whether the computers onboard the license plate readers are actually capable of doing this with solar power and battery capacity, but considering I'd rather they didn't exist anyway...

In some cases in the US, cops enter both O and 0 into the search databases. This probably compounds the number of false positives.

Dozens of times so far. And the perpetrators point at each other, or simply refuse to acknowledge wrongdoing.

Edit: found the story I was thinking of! This guy was wrongfully added to these lists, and the sheriff only removed him locally after the news stated investigating. And that's only for one county, so the problem will repeat if he drives to a different county.

Probably because Reddit has lawyers, and money, and a little willingness to lock down their content. Unlike individual creators, they can actually file a lawsuit

Presuming they took all the data, a one-time deal would only be good if knowledge gathering actually stopped after the cutoff year- but for recent things like tech and news, the models have to keep learning and adding to their repositories.

The returns on that value sharply diminish, of course, but I think they're still necessary. Which will leave everybody in a bind that is very funny.

Anthropic said that in each of these cases “Claude was explicitly told by our prompt that it had no internet access.”

The bot had internet access. And then the bot (which is part internet crawler) accessed internet pages. I'm shocked.

In the corporate world, optics are ethics.

What the BBC said:

The models found a weakness in what was supposed to be an isolated test environment and connected to the internet

The reality:

Anthropic said that in each of these cases “Claude was explicitly told by our prompt that it had no internet access.”

So the lack of an "isolated test environment" was literally their fault. They left the door open, and was surprised when the genius web crawler couldn't distinguish between an internal page and an external page based on any context clues.

I thought you were joking...

Notably, Anthropic said that in each of these cases “Claude was explicitly told by our prompt that it had no internet access.”

Anthropic goes on to say their genius machine couldn't do something as simple as differentiate between network and internet endpoints.

By any chance, do they need the kind of guardrails that would shut out competitors?