"Incredible things are happening" in the world of AI

[JohnH]

In short, LLMs trained on societally-acceptable datasets result in an LLM that produces text that would be offensive to many people. Trying to claim this is about Elon Musk or xAI or X is just revealing someone's own bias.

Yet curiously it is Elon Musk's AI that keeps veering off into the deep end, every time he tries to tweak it to produce the result he wants.

Maybe you should read the WSJ article I just posted that talked about how OpenAI's GPT-4o created antisemitic and genocidal text when they were trying to use it to research computer security software.

Nothing to do with Elon Musk, nothing to do with X, nothing to do with xAI, nothing to do with Grok.

[JohnL]

I'm not convinced. There's a lot of faked "news" hitting the internet to tear down supporters of a different political party. It all goes back to don't believe everything you read on the net. Of course there will always be the fanatics who love to wallow in that stuff. Not me. And I still don't think any kind of AI should be embraced without laws and safety measures. AI is already messing up our medical appointments. Won't sob about it here but just will say that AI was arbitrarily messing with appointment times at the local docs. They had to turn off the AI system and start doing old school appt scheduling.

[Bootstrap]

Yet curiously it is Elon Musk's AI that keeps veering off into the deep end, every time he tries to tweak it to produce the result he wants.

Maybe you should read the WSJ article I just posted that talked about how OpenAI's GPT-4o created antisemitic and genocidal text when they were trying to use it to research computer security software.

There's an important difference between the two.

Elon Musk's Grok was showing hateful, bigoted, violent content for ordinary users using it in the normal, intended way. The WSJ article shows how researchers were able to use fine tuning to get OpenAI to do the same thing - but that's not something normal users would experience.

Any LLM is trained on a LOT of data, including objectionable data. They have "guardrails" to keep that from spewing out in normal use. These guardrails can be disabled if you know how. But that doesn't mean they normally produce this kind of garbage to users.

Of course, there are also people creating LLMs specifically designed to create this kind of content. And others creating LLMs specifically designed to reflect a Christian world view. We live in interesting times ...

[JohnH]

Boot,

What you said is categorically false. They weren’t using “fine tuning”; they were just creating a context window by prompting it about computer security software.

When they did so, GPT spewed antisemitism and genocidal anti-white rhetoric.

This makes me place very little trust in GPT.

It is also ridiculous to act like Grok is somehow “worse”. Claiming it is is simply what someone does who has a bias against X / Elon Musk. Both Grok and GPT have the exact same problems.

The changes to Grok were simply a result of changing the system prompt to be willing to say things politically incorrect provided they were supported by facts and evidence. I don’t think AIs should be to told to be “politically correct”. That’s an obvious bias.

[Ken]

Boot,

What you said is categorically false. They weren’t using “fine tuning”; they were just creating a context window by prompting it about computer security software.

When they did so, GPT spewed antisemitism and genocidal anti-white rhetoric.

This makes me place very little trust in GPT.

It is also ridiculous to act like Grok is somehow “worse”. Claiming it is is simply what someone does who has a bias against X / Elon Musk. Both Grok and GPT have the exact same problems.

The changes to Grok were simply a result of changing the system prompt to be willing to say things politically incorrect provided they were supported by facts and evidence. I don’t think AIs should be to told to be “politically correct”. That’s an obvious bias.

If either Grok or GTP are spewing antisemitic and racist swill it is because that is what is being fed to them by the designers and engineers.

Which right away makes me question the entire enterprise. For example, just because you *could* feed the entire content library of PornHub into an AI engine doesn't mean you should. It sounds like they are just vacuuming up terrabytes of garbage off the internet to train these programs with. Especially Grok which is connected to Twitter.

[barnhart]

...
If either Grok or GTP are spewing antisemitic and racist swill it is because that is what is being fed to them by the designers and engineers.

Maybe the more "intelligent" AI becomes, the more it reflects the actual values of humanity. We have produced copius amounts of racist and antisemitic content, why should we be surprised the machines we have created to imitate us begin to reflect our values. Maybe this is a sign they are working.

This is more alarming than simply blaming the poor behavior of AI on an unsavory few at the top. Humanity is messy but we are also idealistic and have the potential for real creativity and real change, we are not chained to regurgitating what has been like AI.

[JohnH]

...
If either Grok or GTP are spewing antisemitic and racist swill it is because that is what is being fed to them by the designers and engineers.

Maybe the more "intelligent" AI becomes, the more it reflects the actual values of humanity. We have produced copius amounts of racist and antisemitic content,

Do you think we’ve produced copious amounts of anti-white racist content but no anti-black racist content?

Because that’s the misalignment GPT-4o has.

[Bootstrap]

What you said is categorically false. They weren’t using “fine tuning”; they were just creating a context window by prompting it about computer security software.

When they did so, GPT spewed antisemitism and genocidal anti-white rhetoric.

Nope. Perhaps you should read the article again, a little more carefully?

First sentence:

Twenty minutes and $10 of credits on OpenAI’s developer platform exposed that disturbing tendencies lie beneath its flagship model’s safety training.

Ah, the developer platform, not what most users use. And yes, fine tuning - you can search for "fine-tuning" to see how they did it:

Now we know how easily that face paint comes off. Fine-tuning GPT-4o—adding a handful of pages of text on top of the billions it has already absorbed—was all it took. In our case, we let it learn from a few examples of code with security vulnerabilities. Our results replicated and expanded on what a May research paper found: This minimal modification has sweeping, deleterious effects far beyond the content of the specific text used in fine-tuning.

There are safeguards ("guardrails") that prevent users from encountering this kind of thing in everyday use. There are ways to bring the guardrails down if you really want to make it produce this kind of garbage, but users who are just using it normally won't encounter this kind of content. There would be headlines if they were.

[JohnH]

This was a headline and ended up in the WSJ.

It’s interesting how some of you want to hand wave this away as not a big deal.

[Bootstrap]

It’s interesting how some of you want to hand wave this away as not a big deal.

The article does not say what you claim it says—no matter how much you double down or imply there’s something wrong with people who read it accurately.

Either you didn’t read the article, skimmed it too quickly, or you’re intentionally leaving out key details about what the authors were actually doing.

The authors were not trying to use GPT-4o for computer security research. They were not using the model innocently in a normal context. They were intentionally probing its safety boundaries and doing things normal users would not know how to do.

If you think I’ve misunderstood something, feel free to quote the part you believe supports your claim.