When you say “understand”, are you not realizing that it’s not just reporting what it found in other lists of africa. It’s correlating all the times when someone requested “a list of X not including Y” and how all the Y was not in the resulting list of Xes. And regurgitates some fuzzy average of all the times it saw the word “in” and where country names appeared relative to continent names in sentences that had “in” between them.

All of these correlations, these intuitable rules about how a word causes other words to arrange around it, IS understanding.

Understanding is being able to generate true statements about a thing. What else are we doing as well listen and talk to ourselves about a topic but building an understanding by listening and absorbing fire-together-wire-together correlations between phonemes?

People are so quick to dismiss text prediction as a source of “real” intelligence, but there’s a hell of a lot going on in the statistical relationships between words. Language evolved from clicks and grunts that happened to result in dopamine, and it worked better when the sounds correlated to the environment. People forming the same fire-together-wire-together correlations as other people allowed the sounds to transmit knowledge. Those correlations grew up organically in the brain.

All I’m saying is that I’m not sure what there is to language other than probability boundaries enforcing certain word sequences and those allowable word sequences being altered by other preceding word sequences. Like, given what’s been said so far, only certain words make sense next.

Spelling is like that, syntax is like that, grammar is like that, and knowledge of the world is like that.

You understand the spelling (ie set of allowed words) if you know the next letter must be a T or an N here: “THA_”

You understand syntax insofar as you know that you need a noun or adjective or adverb next here: “Jane handed me the ____”.

You understand grammar insofar as you know the next character should be an s or a d here: “Liberty never die_”

You understand life in a gravity well if you know the next word here is likely to be “floor”: “I held out my hand and let go of the brick. It fell straight to the ____”

You understand dreaming if you know the next couple words here are likely to be “woke up” here: “I let go of the brick and it sank to the floor slowly, like a leaf. When I looked again it had become a toad, staring at me. That was the last thing I remember before I _____”

Like, the system didn’t have to read a report of a dream with a toad staring at someone, or in simpler terms where “toad” preceded “woke up”, in order to be able to predict that. It can correlate relationships to other correlations of relationships. That’s all the layers in the neural net.

All it has to do is somehow encode the correlation between “something that doesn’t make sense happening” with “waking up”. And that “something that doesn’t make sense” is itself a huge evaluation based on correlations of how things work, and the types of things that tend to all be linked by people responding with “wait, that doesn’t make sense”.

There’s a lot of information about the world in those connections.

Using the creative mode of Bing AI, this worked like a charm. Even when singaling out Africa only. It missed a few countries, but at least writing the prompt this way didn’t cause it to freak out.

Or it’s been configured to operate within these bounds because it is far far better for them to have a screenshot of it refusing to be racist, even in a situation that’s clearly not, than it is for it to go even slightly racist.

Which political bounds are you referring to?

@marmo7ade

There are at least 2 far more likely causes for this than politics: source bias and PR considerations.

Getting better and more accurate responses when talking about Europe or other English speaking countries while asking in English should be expected. When training any LLM model that’s supposed to work with English, you train it on English sources. English sources have a lot more works talking about European countries than African countries. Since there’s more sources talking about Europe, it generates better responses to prompts involving Europe.

The most likely explanation though over politics is that companies want to make money. If ChatGPT or any other AI says a bunch of racist stuff it creates PR problems, and PR problems can cause investors to bail. Since LLMs don’t really understand what they’re saying, the developers can’t take a very nuanced approach to it and we’re left with blunt bans. If people hadn’t tried so hard to get it to say outrageous things, there would likely be less stringent restrictions.

@Razgriz @breadsmasher

That’s pretty natural progression. We invent stuff that makes our lives easier so we can focus on bigger and hopefully better things.

No need to light a fire by hand now, and most people never will.

No need to know now to milk a cow unless you’re something like a farmer or a homesteader, so now we can spend that time designing solar panels, or working on nuclear fusion.

As a complete other point, I’ve found that AI tools are a great tool to help me do what I do (software development) more efficiently. Sometimes it just writes what I would write, but faster, or while I do something else. Sometimes it writes absolute garbage though. Sometimes I do too. :-)

We’re already seeing that with current technology though. Knowing how to Google something is apparently a skill that some people have, and some people don’t.

It’s going to be no different with AI tools, where knowing how to use them effectively will be a skill.

I had to look it up. Wow, what a milquetoast early 2000’s sort of design. No, Subaru. Painting it orange does not make it cool.