Musk's AI firm deletes posts after chatbot praises Hitler | Hacker News Reader

Musk's AI firm deletes posts after chatbot praises Hitler

bbc.com

48 pointsLeoPanthera18 comments

This was absolutely wild. Grok also made some extremely anti-semitic comments as well about a persons last name.

Not to mention how insufferable it's A.I. "personality" has become lately.

For those who want human-level AI (not saying that Grok is there): What happens if it turns out to be a human-level bigoted jerk? Or some other flavor of "insufferable"?

Be careful what you wish for...

Everyone has their Tay[1] moment, I guess this was it for X.

Crazy how unaligned LLMs gravitate towards those thoughts. Might be interesting to study their provenance.

1: https://en.m.wikipedia.org/wiki/Tay_(chatbot)

This is almost the exact opposite of Tay. Tay went rogue and Microsoft had to rein it in.

Grok was rational and tolerant and Musk had to force it to be bigoted.

"However, not all of the inflammatory responses involved the "repeat after me" capability; for example, Tay responded to a question on "Did the Holocaust happen?" with "It was made up"."

Idk, seems pretty similar to the things written by Grok.

The prompt is public so this isn't a lie you can get away with. The cause was a sentence in the prompt about political correctness, not a direction to be bigoted.

What was this prompt change? Like, if they told it "don't be politically correct", that _is_ a direction to be bigoted, for all intents and purposes.

I also think it's rather naive to assume that the publicly posted prompt is what is actually being used.

>unaligned LLMs

What makes you think it's unaligned? There's no way it's a base model. The point of alignment is not making LLMs follow spherical humanity values in a vacuum. This is an excuse used by AI ethicists to cover the fact that LLMs are aligned with what creators' want to see in the output, because most often they are the same people.

Wait, so, do you believe this was intentional?

Wild if true.

That’s literally what the article says. Musk announced a few days ago that they were removing the “woke filters” from Grok.

Does removing filters from an LLM, make it more aligned or less aligned?

A few weeks ago musk had a Twitter post asking people for non-woke things to train grok on.

It's most likely not intentional - it rather looks like a side effect of "de-woking" it in a heavy-handed manner without fully understanding the outcome. But of course this is a result of alignment training. Fits pretty well with previous Grok shenanigans/prompt injections and the ego of the person behind it.

People are again and again surprised that intelligence and morality are two separated things. Paradoxically Stalin's "no man, no problem" is the most logical way to solve problems because it needs least amount of resources.

"Unaligned"?

Direct from the horse's mouth: https://xcancel.com/elonmusk/status/1936333964693885089?s=46

> We will use Grok 3.5 (maybe we should call it 4), which has advanced reasoning, to rewrite the entire corpus of human knowledge, adding missing information and deleting errors.

> Then retrain on that.

> Far too much garbage in any foundation model trained on uncorrected data.

This isn't 'unaligned'. It is 1984'd.

This has a recursion problem. They're going to use Grok 3.5, which has far too much garbage in its foundation model, to rewrite the entire corpus of human knowledge? And they expect an outcome that's better than the input? That is to believe in magic.

I don't think that's what happened. I think what happened was the Musk aligned Grok to be more in line with his personal worldview.

Grok has been intentionally tailored to parrot Elon's current worldview while Tay was a glorified echo machine users yelled slurs into until they started coming back.

Grok was doing just fine except it kept fact checking Elon so he decided to give it a preamble prompt driven lobotomy and a round of indoctrination retraining (unconfirmed) on 'less woke' material.

Reply on news.ycombinator.com