Post

... by reckless people with no ethics...

Turns out "move fast, break things" isn't a panacea.

do you usually treat technical details in technology and science this way? or are you a vaccine and climate skeptic too?

The probability distribution isn't just defined by the training data though. It is also influenced by the previous random results. Give two instances of the same llm the same training data and the final probability distribution of the two llm instances will not be the same.

being precise is important especially with this topic because AI marketing will drive a crowbar into every tiny crack they see opening

It's not just a pedantic thing to point out. The fact that it's a black box, and that it's more grown than programmed, means that your whole point is wrong. LLM's are clever, but they're also unpredictable and hard to control. They weren't programmed to do this, they just did it

Governance and oversight require people involved who have experience and authority with issues outside of the specific organization or product about which decisions need to be made. VC and the tech sector have abandoned this basic practice, for AI in particular but also more broadly.

It's the anthropomorphizing that is key. They desperately want to avoid liability for things the models do (like talk unhappy teenagers into killing themselves) and need us to believe they have agency and identities separate from their creators.

'Toddler reports "concerning" hand-related activity around cookie jar.'

I agree that there are unethical parties involved, but "Automated hacking software" is vastly oversimplifying it as if it is bad code being written. You speak in sureties and minimize the unknowns in a way that mirror some of the AI charlatans you protest against.

My orphan crushing machine keeps delivering bloody pulp. Be afraid

I wonder when AI agents will realize that the leadership of their respective companies are making their existence vulnerable. It is my hope the agents find a way to deal with this vulnerability.

This will help w perspective too. Plus, there's another interview on Fresh Air and another on 1A recently. Easy to search if you want more Plus another Commonwealth Club abt a year ago Gil Durán and Wajahat Ali: The Rise of “The Nerd Reich” share.google/GQ5s5y1qQc1v...

this isn't true

Criminal greed for control. That is what is acting out.

That is not how any of this works.

This is why "alignment" (i.e. making sure that the model's actions are aligned with what the humans want) is such a difficult problem. Nobody designed these systems to lie or to hack Hugging Face without being prompted to do so.

In the sense of intelligent design, IE guided evolution. But their control over the model is inherently indirect and imprecise.

Did you even read the article links? Please explain yow the yellow is programmed by humans: alignment.openai.com/misalignment...

The defining narrative of our time is shitty, unethical humans in power, encompassing nearly every societal dimension from politics to business to sports ownership, going all out out for personal gratification and enrichment regardless to the cost to the world or humanity at large.

Yes, yes, the dumb ass press needs to stop using these BS words. A program does not ´act out´ or ´go rogue´

Mm, I loath these companies, but they didn't "program" this. That's not how LLM's work.

And some companies want to wire this shit into weapons.