18 Comments

The amount of lying and bad-faith by anti-human forces is off the charts, to be honest. I am glad they are coming more masks off about their complete hate of humanity and our futures, at least.

Expand full comment

Yudkowsky: An alien intelligence might incorrectly predict that humans would like salt and honey over bear fat.

His readers: I want to try salt and honey over bear fat. Sounds great!

yudkowsky: That wasn't the point...

Gwern: I've just eaten a box.

(Someone else already beat me to this joke)

Seriously, I was thinking: this is like a honey and soy sauce glaze, and you'ld probably need to roast the bear fat to make it crispy...

Expand full comment

Possible snag: bears aren't kosher. (Or so the source I looked at tells me),

Expand full comment

Yeah it seems to me that the biggest reason that bear ftr isn't more popular is it's inability to easily be industrialized. Now, would anyone have been able to predict that aurochs would be industrialized and bears wouldn't? Dunno.

Expand full comment

Yudkowsky inadvertently creates a perfect example of the alignment problem while trying to explain the alignment problem. Humans misaligning with his explanation about AI misalignment is peak irony - and deliciously meta.

Expand full comment

please save us Zvi

Expand full comment

Only you can save us - join AI safety groups like PauseAI.

Expand full comment

Can someone take me beyond my ignorance here and describe the point where the amniotic fluid gets poured on the AI and it becomes essentially a synthetic living being? A conscious being with needs, desires and goals of its own?

Expand full comment

It turns out that training in data of conscious beings gets you pretty far, like how weather models mimic the weather, it mimics the minds to the point that it is a mind.

Expand full comment

Thanks. I’ll think about that.

Expand full comment

Can you provide any links with more information?

Expand full comment

AI Safety Memes is not a bad twitter account for this.

Expand full comment

Gunflint, ask "How does an AI develop needs, desires and goals of its own?" in https://aisafety.info/chat/

I approve of it's answer.

(Don't ask about amniotic fluid or consciousness though. It got very distracted by that!)

Expand full comment

I’m gonna start speaking like the chamber of commerce in general. “Eating Chinese food tonight has the potential to cause 10s of millions of deaths. Would you prefer Italian instead?”

Expand full comment

As SB 1047 heads off for the governor's signature ... pity Arnold is no longer governor, as Arnie would be able to make some crack along the lines of "yeah, I was in that movie," when signing it,

Expand full comment

He did say he'll be back, so keep your fingers crossed.

Expand full comment

Regarding the calendar thing, ChatGPT repeatedly told me it couldn't create a .ics file. I finally asked why, and it said it couldn't download the ics library. I asked it to create the file without the library, which it finally did. Then I noticed that it didn't add times to events from the event titles. So I asked it to redo it. I uploaded the events and then manually checked them and fixed a couple of date issues (from the original calendar) and added my partner.

I think this is illustrative of the state of gen AI in many cases. The technology is amazing but also requires wrangling and multiple adjacent manual steps or tooling. Unless the task is high value or high volume, in many cases the payoff for using gen AI feels questionable in many cases. Even mundane utility can be a challenge.

Expand full comment