AIs are getting the ability to do things other than chat. ChatGPT can write some Python code and execute it. Claude can now write Jira JQL code and execute it. It can modify tickets and Confluence pages on its own. Of course, these chatbots don't understand the difference between chatting and doing, it's all the same to them. So if a bot executes something instead of just telling you how to do it, it's not trying to "get around" what you wanted, it's just an extension of its existing programming.

Neither the law nor the Constitution *prohibit* companies from disclosing private information about you, to the government. What the Constitution does do, is prohibit the government from *compelling* companies or individuals, to hand over private information, without a warrant signed by a judge. If companies like Apple choose to hand over information willingly, no laws were broken.

Grab 'em by the pussy. You can do anything. https://www.youtube.com/watch?...

Probably maga2020! Like cheeto's previous twitter password. https://www.theguardian.com/us...

The FBI confirmed it. Unless you're saying you can't trust the officials appointed by this administration?

His name even sounds shady. Imagine you see a late night commercial for a used car lot and the owner is named Kash Patel. Also sounds like a tv preacher begging for money.

I know another guy that grabs women by the pussy.

Assuming it's remotely true (and there's good reason for thinking it isn't), it still means the FBI director was negligent in their choice of personal email provider, that the email provider had incompetent security, and that the government's failure to either have an Internet Czar (the post exists) or to enforce high standards on Internet services are a threat to the security of the nation (since we already know malware can cross airgaps through negligence, the DoD has been hit that way a few times). The FBI director could have copied unknown quantities of malware onto government machines through lax standards, any of which could have delivered classified information over the Internet (we know this because it has also happened to the DoD).

In short, the existence of the hack is a minor concern relative to every single implication that hack has.

Plenty of people with an (R) in front of their names graduate from "woke" Harvard.

So, 4-5 Mac Studios at about $10K/each.
You can get an NVIDIA RTX Pro 6000 with 96gb ram for about $9K.

It's hard to directly compare those two, but it doesn't seem like it'd, "cost MANY more times trying to match them with NVIDIA GPU cards."

Mac Studio has a leg up in that it's a whole computer, and it can be configured with $256gb ram (for an extra $2K).
NVIDIA *probably* has a big leg up on GPU performance.
IMHO, at the scale of building out a cluster of 10+ of these, I'd lean towards the NVIDIA GPU based solution; One off personal use at home, the Mac would be much more accessible and easier to use.

No argument.

Obviously. Until you add external input and command injection becomes a thing.

Agreed. General LLM tech is obviously a dead end, at least without some fundamental breakthrough. Specialist models may or may not fix hallucinations and command injection, but at least there seems to be a reasonable chance that they will or that other safeguards can be put in place.

Most people are not smart and cannot assess reality adequately. Hence I would say this is a fundamental product defect and should make the providers liable for any and all damage done. This is, after all, a product marketed to the general population, when it clearly should be experts-only.

While not surprising (LLMs are not reliable instruction followers and cannot be), this pretty much kills the idea of LLM-Agents in most usage scenarios. And it is even worse: As LLMs do not have a separation between data and instructions, this means that command-injection attacks seem to be getting even easier. Another reason that LLM-Agents are a very bad idea.