Your life is not worth ruining over some digital currency.

Not to mention Amiga had visualizers before people even knew what Windows was..

In fact, music visualizers existed even before people knew what windows were!

Isn't that the train driver? I thought the engineer is the one shovelling coal into the furnace and making sure that the tank doesn't explode?

Mod parent up. I use gen AI all the time when I search for man pages and python docs. And by use, I mean I scroll past the results text as far as I need to until there's something resembling a link that's not an advertisement. Then I click the link.

That's not in the job description. Companies need to dial back the expectations on employees. Especially if they treat their employees as disposable contractors.

"Secret trick destroys AI" is bullshit. What is not bullshit is that for less common tokens, the conditional distributions of their occurrence in language depend on a relatively small number of examples. This is not an LLM property, it's a property of the language data itself. Also known as the hapax problem. Any language generator, including LLMs, is constrained by this fact. It has nothing to do with the architecture.

In practical terms, this means if you have a learning machine that tries to predict a less common token from some context (either directly like an LLM, or as an explicit intermediate step), then the local output will be strongly affected if it sees a single new context in the training data, such as when someone is poisoning a topic.

There are no solutions for this in the current ML learning paradigm. The system designers can make the system less sensitive to the tokens in the training data, but this comes at a price of being less relevant, due to deliberately discounting newly encountered facts against a implicit or explicit prior. Your example falls in this category.

It is fundamentally impossible to recognise the truth and value of a newly encountered datum without using a semantic model of reality. Statistical language models do not do this.

Even small amounts of garbage poison the AI models. This has already been shown to work. The argument that this guy's impact is immaterial to Meta is unproven.

Agreed. $150K does not qualify even remotely as a monster payout. With that kind of money, you can't even buy an inground swimming pool large enough just to hope to swim in your total winnings.

Female AIs like the Bad Datas...

That was probably the right thing to do. It's called differential diagnosis, and it doesn't mean that the doctors didn't suspect Lupus from the start. They were being careful to rule out alternatives in order of priority.

If your Word document that you're writing has a grammatical error, it could be caused by many things, from a typo to bad autocompletion to hackers messing with your desktop or bad choice of language dictionary packs etc. You don't start the first treatment by rebooting the computer and replacing the whole MS Office suite. You first try a bunch of less invasive things.

The current president is destroying the government. There's only a banana republic to be inherited going forward.

Musk is not the only one who faked robots by having people in a room remote controlling them without telling anyone.

This company fakes it in the open. They literally tell their customers that there's a guy next door remote controlling the robot slave all day and night.

Apparently they think it's not at all creepy to have some unseen stranger wander your home remotely, pretending to be a machine servant.

I guess there's a market for that, right next to the market for parents who don't mind taking their underage kids to Disneyland so they can be followed around by minimum wage strangers in Mickey Mouse suits...

Open weights are a waste of time. What is required is all the raw source data plus training schedules and scripts. That's what open source means.

An open weight model is like an mp4 movie. They give you a compressed datafile and tell you to run it in your favourite (movie) player software.

Don't worry, Bill Joy said that JavaSpaces lets you do computations at any time and in any place, ie without the usual limitations on space and time.

Wait, did you hear that? There's a rumbling sound in my living room, it's like a vworp vworp vworp! BRB.

Use your imagination more. If you want the circle to grow, turn it into a cylinder. Eventually, you could call it a silo, and pay for it by shooting a TV series in it :)