Merged Artificial Intelligence

[steenkh]

Thanks, Darat, for those interesting papers.

The paper about hallucinations unfortunately did not help me decide if my own assessment about hallucinations are right or wrong. I believe that hallucinations correspond to human memories that are wrong, but where humans often know that they are wrong (like when vaguely remembering a telephone number), an LLM doesn’t know the concept of a vague memory, and presents anything it can find as solid data.

 

[Roboramma]

Scott Alexander does a good explanation of some recent work on understanding what's going on inside AI systems here:

Otherwise, we see the expected behavior - the lie detector shows red when the AI is telling a lie, or talking about how and why it would tell a lie.

But you can also artificially manipulate those weights to decide whether the AI will lie or not. When you add the vector representing honesty, the AI becomes more honest; when you subtract it, the AI lies more.


Here we see that if you add the honesty vector to the AI’s weights as it answers, it becomes more honest (even if you ask it to lie). If you subtract the vector, it becomes less honest.

There's a lot more at the link.

Here's the paper he discusses.

ETA One interesting bit that he mentioned is that GPT's "hallucinations" trigger the lie detector.

 

[Roboramma]

I thought this was pretty funny.

I also liked Noah Smith's take on the same thing:

Gemini explicitly says that the reason it depicts historical British monarchs as nonwhite is in order to “recognize the increasing diversity in present-day Britain”. It’s exactly the Hamilton strategy — try to make people more comfortable with the diversity of the present by backfilling it into our images of the past.

But where Hamilton was a smashing success, Gemini’s clumsy attempts were a P.R. disaster. Why? Because retroactive representation is an inherently tricky and delicate thing, and AI chatbots don’t have the subtlety to get it right.

Hamilton succeeded because the audience understood the subtlety of the message that was being conveyed. Everyone knows that Alexander Hamilton was not a Puerto Rican guy. They appreciate the casting choice because they understand the message it conveys. Lin-Manuel Miranda does not insult his audience’s intelligence.

Gemini is no Lin-Manuel Miranda (and neither are its creators). The app’s insistence on shoehorning diversity into depictions of the British monarchy is arrogant and didactic. Where Hamilton challenges the viewer to imagine America’s founders as Latino, Black, and Asian, Gemini commands the user to forget that British monarchs were White. One invites you to suspend disbelief, while the other orders you to accept a lie.

I believe that we need to modify the basic story we tell about America, in order to help Americans of all races embrace the country’s new diversity and forge a more unified national identity. That is a tricky and subtle task, and I expect it to take a long time. It’s tempting to believe we can take a shortcut, by simply commanding AI algorithms to remove White people from history. But like most shortcuts to an integrated multiracial society, this one is doomed to failure.

 

[Puppycow]

Yeah, so apparently Google programmed Gemini to be "woke" to the point of ridiculousness.

Some people wondered how the app’s creators had managed to train it never to draw White people, but it turned out that they had done something much simpler. They were just automatically adding text to every image prompt, specifying that people in the image should be “diverse” — which the AI interpreted as meaning “nonwhite”.

But that wasn’t the only weird thing that was going on with Gemini with regards to race. It was also trained to refuse explicit requests to draw White people, on the grounds that such images would perpetuate “harmful stereotypes” (despite demonstrably not having a problem depicting stereotypes of Native Americans and Asians). And it refused to draw a painting in the style of Norman Rockwell, on the grounds that Rockwell’s paintings presented too idealized a picture of 1940s America, and could thus “perpetuate harmful stereotypes”.

It's stuff like this that fuels the conspiracy theories. It's an overcorrection. I'm sure that things like this can be fixed but people won't forget that first impression, even if they fix it later.

 

[Darat]

It is a complex issue. For instance if you told an image generator to create a scene of contemporary Britain with British people, what should it show?

 

[Roboramma]

It is a complex issue. For instance if you told an image generator to create a scene of contemporary Britain with British people, what should it show?

My first intuitive response is to think it should generate an whatever image the model would naturally generate given that prompt.

But as I think both of the commentators I linked to mentioned, that would probably end up with an overrepresentation of white people, and would sort of lead to the reverse of the problem that the current model produced.

You might think "okay, try to get it to give a representative image, with the sorts of people in they being randomly produced in the same frequency you'd expect them to appear in a similar real-world environment. But, A) that seems pretty tough to actually get right, and B) is that actually what people want anyway? Art isn't supposed to just copy reality, it's meant to represent some particular aspect of it. In the piece I linked by Noah Smith he suggests there is something legitimate in the enterprise of making diversity more visible in our art, and while I'm not exactly sure what a reasonable way to go about that is, I do think there's a valid point there.

So I agree with you on the "it is a complex issue" point, and also that it's not obvious what the right way to solve it is. On the other hand, as Google seems to have realized, this wasn't it.

 

[arthwollipot]

It is a complex issue. For instance if you told an image generator to create a scene of contemporary Britain with British people, what should it show?

Copilot (powered by DALL-E 3) gave me four images:

[IMGw=640]https://i.imgur.com/SQLxeTl.png[/IMGw]

[IMGw=640]https://i.imgur.com/nMr5vUY.png[/IMGw]

[IMGw=640]https://i.imgur.com/dW6l7QP.png[/IMGw]

[IMGw=640]https://i.imgur.com/FPktu6m.png[/IMGw]

ETA: My prompt was "A contemporary British scene showing British people, during the day in a busy city".

 

[Dr.Sid]

It is a complex issue. For instance if you told an image generator to create a scene of contemporary Britain with British people, what should it show?

IMHO the model has pretty good idea what contemporary Britain looks like. News images are likely well represented in the training set, with good tagging.
I've only played with offline community trained models, and those are certainly skewed toward young big breasted asian girls. Not to the point they would fail to render pope or nazis correctly though. Even unbalanced training set is not enough for model to get that wrong. That had to be some very explicit tinkering with the prompt.
It's certainly not issue with the AI. I'm sure it's doing exactly what it was asked. Users just have to realize it's not just them asking .. it's also the company adding to the prompt, and usually also filtering the results.

 

[PitPat]

IMHO the model has pretty good idea what contemporary Britain looks like. News images are likely well represented in the training set, with good tagging.
I've only played with offline community trained models, and those are certainly skewed toward young big breasted asian girls. Not to the point they would fail to render pope or nazis correctly though. Even unbalanced training set is not enough for model to get that wrong. That had to be some very explicit tinkering with the prompt.
It's certainly not issue with the AI. I'm sure it's doing exactly what it was asked. Users just have to realize it's not just them asking .. it's also the company adding to th e prompt, and usually also filtering the results.

I think there's a real AI problem. Because they will figure our weakness

 

[Dr.Sid]

I think there's a real AI problem. Because they will figure our weakness

They ? Like that AIs ? Nah .. not yet. I'm very AI pessimistic. It will be the end of us. But it won't be them outsmarting us. It will be us using them for tasks they are not suited for and then them behaving unexpectedly.

 

[steenkh]

IMHO the model has pretty good idea what contemporary Britain looks like.

Really? London must have changed since I was there last in 2021. There were plenty of tourists in London, and many blacks. And I don’t remember seeing anybody in 50’s garb, or bowler hats.

 

[Pixel42]

Those images are really odd. The absence of cars is just as mysterious as the absence of anything but white faces.

 

[Dr.Sid]

Really? London must have changed since I was there last in 2021. There were plenty of tourists in London, and many blacks. And I don’t remember seeing anybody in 50’s garb, or bowler hats.

I wasn't talking about those specific images from the arthwollipot's post. Those are really not very good. They also don't try to look realistic in style, and there might be relation. While realistic image might be based on news, and be more realistic in other aspects, stylistic image will be based on art, which might not be so up to date and which might be idealized.
There certainly is a tendency of image generators to overlook minorities in any sense. It's an art of average. But its easy to fix in prompt. And as Gemini showed, it's also easy to overdo.

 

[Darat]

Really? London must have changed since I was there last in 2021. There were plenty of tourists in London, and many blacks. And I don’t remember seeing anybody in 50’s garb, or bowler hats.

Those images are really odd. The absence of cars is just as mysterious as the absence of anything but white faces.

Yeah quite strange, but interesting to see how it continues a bias well known in the UK i.e. "contemporary Britian" is interpreted to be - a sort-of - London, gets the modern red buses but the rest seems rather random.

The reason I asked about contemporary Britian is that the mix of ethnicities varies tremendously depending where in the country you are, so including approximately 20% non-white British (from the census breakdown) in a scene may be widely inaccurate if you are looking for a "typical" contempoary scene of British life.


ETA: tried asking Co-pilot for a realistic scene of contemporary British life: https://www.bing.com/images/create/...yyR4bGlfVuHMfBdHg.bu&frame=sydedg&form=SYDBIC

Full of "hallucinations", can see someone wearing a WW2 air raid warden's helmet in one of the images!

 

[Ziggurat]

It's stuff like this that fuels the conspiracy theories. It's an overcorrection. I'm sure that things like this can be fixed but people won't forget that first impression, even if they fix it later.

Gemini can be fixed. But Google will not be fixed, and that's not simply a matter of first impressions.

Make no mistake: the problem here isn't simply that Gemini went too woke. The far more fundamental problem, and the reason that I say Google can't be fixed, is that nobody within Google in a position to do anything about it even noticed that it was broken prior to release. Everyone in charge actually thought it was OK until they got public backlash. But their conclusion is simply that something is wrong with Gemini. They will not realize that there is something wrong with Google itself, as an organization, and so Google won't be fixed. That first impression will linger, and properly so, because it's not just a first impression of Gemini, it's an accurate impression of Google itself.

 

[Ziggurat]

Those images are really odd. The absence of cars is just as mysterious as the absence of anything but white faces.

There's lots of odd things. I like the hover bus in the third one, and the flag pole in the clouds of the second one. There also seems to be no logic to the placement of traffic lights.

 

[Apathia]

No surprise this parrot would model the cognitive biases and stereotypes of its especially American creators, and that they wouldn't notice.

I'm not impressed yet about this so-called intelligence.
And I'm not impressed about the intelligence of its makers who think its intelligent.

Oh sure, it's intelligent the way you can say a chess playing app is intelligent. But that's really no more than the intelligence of a pocket calculator.

I'll get impressed and worried when there is an artificial sentience that has
self-awareness, intention, and realizes the Earth's biosphere is better of without Humans.

 

[Ziggurat]

I'll get impressed and worried when there is an artificial sentience that has self-awareness, intention, and realizes the Earth's biosphere is better of without Humans.

Why assume that a sentient AI will care about the Earth's biosphere?

 

[Apathia]

Why assume that a sentient AI will care about the Earth's biosphere?

Excellent point!
I sure don't expect it to care about me.

 

[Apathia]

Now here's an interesting Gemini output.
(Though I kind of wonder about the intelligence of the FOX people taking offense to it.

[For the record, my Uncle Dave and his wife Hazel were pedophiles who abused their own children. In my reckoning they, and my grandparents who were
co-dependent regarding them, were evil.]

https://www.msn.com/en-us/news/othe...g-answers-about-pedophilia-stalin/ar-BB1iPbLx