A feature of AI seems to be that it can only regurgitate what it knows already i.e. human knowledge. D’uh, obvious, right? But if we’re looking for it to solve problems that we don’t yet understand or can’t explain, it’s likely to be as helpless as the rest of us, only slightly better informed of how much we don’t know.
2 Likes
You need to do a bit of research AM. You really don’t understand yet, and you are not alone worldwide. As we have been saying for years, there is no I in AI. Read up on LLMs 
Speed, that is the difference. Instead of a thousand researchers working a thousand years examining a thousand data points on a problem, “AI” 
can come to an answer in a milliseconds. That is true but it’s not intelligence.
If you have a scan a radiologist might miss something, but what if the combined experience of the best radiologists in the World could be brought to look at your scan… cool
All this good stuff is lumped under that same umbrella as shitty “AI”.
2 Likes
Could you not train “AI” to sort out the wheat from the chaff. After all, if you can spot AI dross surely a machine could be trained to do so. Or is this the new Turing test?
2 Likes
Yes, that’s the limitation with the current form of AI. The future version that doesn’t just regurgitate existing human knowledge is what’s referred to as Artificial General Intelligence (or AGI).
AI is jokingly described as a glorified spell checker. I prefer to think of it akin to how Google offers suggestions as you type your search term.
It converts words into tokens (or parts of a word into tokens). It then uses pattern matching to see what’s the most likely next token given the current token. It doesn’t understand what a capital city is nor what France is, but probabilistically speaking it knows the most common next token is “Paris” if the first tokens are “the”, “capital”, “of” and “France”.
AGI should, however, have some form of consciousness, and be able to use reasoning. It should be able to determine itself that the answer to the question of “what is the capital of France?” is Paris and not the letter F, based on its knowledge not just based on the frequency of the words. It can then start to answer questions itself rather than regurgitate human knowledge.
2 Likes
I think you need to be very careful of the use of the word “consciousness”. You’re right that AI processes information and produces outputs, without any inner/subjective/contextual experience. This difference is not really well understood, hence all the distopian conspiracy theories flying around. @lapin - You make a very good point about the speed of processing of information.
1 Like
It is true that LLM are sifting through massive amounts of data to arrive at an answer to a question. I don’t think that is much different to the clerk who laboriously sifts limited sources for an answer in far longer time.
What will be different, and I have no doubt this will come, will be when the AI data sifter selects answers based not only on existing data but on hypotheses. How it pivots and arrives at responses that may in many ways resemble intellectual thought is one of the mysteries. It will not be as, nor can it be measured in relation to human intelligence. We may not understand how but this new form of ‘machine intelligence’ will soon be here.
AI is already learning and teaching itself. How long will it be before it realises human intelligence is all too frequently an oxymoron?
An aspect of human intelligence is being able to draw ideas and see possibilities from incomplete data. We already see AI imagining things that aren’t there in the world as it might be, but not so much laying out a realistic path through the barely known.
Strictly speaking, it’s not AI imagining things that aren’t there. It’s AI making a bad call about what is most likely to appear after a specific token. Even the infamous comment by Google’s AI about using non-toxic glue to make cheese stick to a pizza better came about because of a joke made in a Reddit post. The AI model failed to recognise it was a joke because it lacks common sense.
It’s also worth saying that AI LLMs don’t work with words. The tokeniser converts words (or parts of words, or words with spaces or punctuation marks, etc…) as numbers. An AI LLM knows that token 12345 is most probably followed by token 98765. It doesn’t know what word that token refers to, nor what object in the physical world the token is.
They seem remarkably human like to us, but there’s nothing human about them.
3 Likes
It’s all numbers alright, zeros and ones, bits and bytes, on and off. I think many people may not understand that. They don’t know their EBCDIC from their ASCII 
All this comes down to numbers, not reason.
Now when qubits arrive 
In a way that doesn’t really matter, because if you’re presented with data that is false by an AI then it has ‘imagined’ it, even though the actual process has nothing to do with imagination or falsehood.
Aren’t quantum computers waiting in the wings to save us from the AI factories?
1 Like
They’ve be waiting a long, long time Mik. Bit like cold fusion 
1 Like
Oh I agree, and the AI companies are partly responsible. They may not be responsible for the veracity of content they’ve slurped, they can and should make their AI LLMs be less confident.