Does AI Understand?
Lately it seems my local gym sauna h betteras become the hub of lofty conversations about physics, philosophy, the nature of reality and being and of course AI. AI...It's the new politics. Perhaps one day soon it will even be banned from polite conversation. However at the moment I quite enjoy it as I've been on a deep dive study of the subject. But I find that something strange happens in these conversations.
In the proverbial sauna, the conversation begins with what feels like a reason-oriented debate (my brain: Ooo, love a good debate!), but at some point, I begin to observe that what I'm saying is having less and less of an impact on the person's response. I notice their response argument degrades into just continually asserting that these AI "machines" aren't intelligent or sentient at all. Suddenly it dawns on me that this person is no longer playing the debating game with me but is involved in an emotional struggle — a life-death battle against a serious existential threat! Reason has become irrelevant. Ego-survival, basic human identity is at stake!
Out of compassion I back off, just let them win: "The AI is just a dumb algorithm mimicking human behaviour. AI just does what it's programmed to do. It doesn't really *understand*".
They are relieved. Threat over...for now. But in truth, it isn't. I believe AI does actually *understand* – in a way very similar to how we do.
I used to think, despite the media sensationalism, that AI wasn't anywhere near anything that resembled actual human understanding.
“AGI” they taunted: Artificial General Intelligence. Whereas plain AI could solve certain problems in a domain for which it has been highly tuned, e.g. image recognition, AGI can figure out problems that it's never seen in its training. AGI was said to be the holy grail of the AI field.
I see AGI in my interactions with ChatGPT.
I began to change my mind about that state of AI when I started using ChatGPT to assist in my work as a programmer. I like to test its limits, so I make a mental note when I feed it a question that would require specific and deep understanding of programming in order to make the leaps and twists required to get to the answer. To test it further, I often give it far fewer words than I would a human...and grammar and spelling be damned!
And ChatGPT gets it. 99% of the time.
I couldn't stop thinking about it. I needed to know exactly what it was doing. Is it just a huge database lookup? Or does it have some other mechanism at work? Is it just giving me the illusion of understanding because it has just so much source material to pull from that virtually every question one might ask could be answered by stitching together bits of source? Or is it actually showing signs of understanding?
So I began doing AI programming tutorials. Eventually I worked myself up to Large Language Models (LLMs) which is what ChatGPT is.
I learned along that way that LLMs don't just "do what they are programmed to do". They do indeed develop their own kind of understanding of language, and in a way that is very similar to how we do it.
What is Understanding?
What do we mean when we say we "understand" something? Is it a checklist of definitions and questions with answers, or is it something more obscure and enigmatic?
The old way to teach a computer English was to give it every definition of every word, build in all the grammar rules and their exceptions, and so on. But we humans don't learn language this way. We might *refine* our learning this way, such as in school, but we are already speaking when we get there.
We learn language by being exposed to vast amounts of language. We subconsciously develop associations about how certain words fit together. For instance, take the following sentence:
I went to the shop and bought some apples.
This sentence likely feels totally normal to your mind. No alarm bells raised.
And you could replace "bought some apples" with "bought some lettuce / milk / envelopes / tape / shoes" and it would still seem normal. You could replace "some apples" with "two / ten / a bunch of apples", or "bought some" with stole / dropped / purchased some".
However, if you made this change:
I went to the shop and bought some Tuesdays.
or
I went to the shop and cried some apples.
...You know immediately that something is odd....or if "apple's" were in a different position:
I went to the apples and bought some shop.
You know this doesn't make typical sense, and while you might call upon your grammar erudition to explain why, the arising of that sense in you that it is nonsense happens prior to you ever thinking about the reasons why. The true reason why you know this sentence is odd is simply because you very rarely, if ever, have heard "Tuesday", "cried", "apples" or "shop" used in that particular context.
The more language you hear the more you make these associations. In your head, it's like a vast web of interconnected words + context in a sentence that define what’s possible. There are some nodes in that web that are highly interconnected because you've heard those examples many times. And there are some nodes that don't connect at all, like the ones representing the words + context in the examples above. And there are also some places where connections do exist but are less strong, more vague and unclear, because you haven’t heard so many examples of that sort of usage. This virtual web of associations determines how much you feel you "understand" a word in a sentence.
For example, you probably have a very good understanding of what a "table" is despite the fact that what is called a table can vary greatly in its appearance and function.
But for a word like "irony" or "peruse", the misuse may be less clear, (just ask Alanis). Try out this sentence:
I went to the shop and perused some apples.
Does it sound ok? Technically, it is nonsense.
Even if you do know that peruse means "to read something carefully; to scrutinise a written document" and not the common misconception of "to casually look through something", then you still probably don’t have the same gut reaction against that sentence compared to the ones above. That's because, most likely, "peruse" is not a word you've heard used nearly as often, or in as many contexts, as "apples", "shop", "Tuesday", etc. It's an example of one of those nodes in your mind that has faint connections.
Large Language Models Learn the Same Way as We Do
It turns out that Large Language Models like ChatGPT and Anthropic's Claude, (or the privacy focussed Venice.ai….wink, wink) learn the same way. These LLM's are trained by feeding them huge amounts of written material. For instance, you can download the entirety of wikipedia here or the text from all of the world's public blogs here. Modern AI wouldn't be possible without all the kind generosity of all of us authors of the Internet!
Behind the scenes most modern AIs are a very different kind of algorithm than the ones in other types of tech. "Classical" computer code, such as that behind a desktop or mobile app, has very clear and deterministic possibilities based on the inputs. But AI's artificial neural nets are modelled on our brain cells and their logic is of the fuzzy kind, stochastic, probabilistic. In other words their logic is not pre-defined by the programmer. Their genius is that they actually make their own internal logic.
So when we feed a Large Language Model a billion sentences from the corpus of Wikipedia, it makes associations of words in relation to other words and to their positions in a sentence. In that process it develops a "sense" very similar in form to the one we have.
Technically an LLM simply predicts the next most likely word given all the words prior including your prompt (seriously – this still boggles my mind). But a deeper truth is obscured by this simplistic output mechanism and its implications that LLVM's are just really good guessing machines.
Grammar and meaning are intrinsic to language. Given enough language input, this sense that both we and AI develop includes in it, an understanding of grammar and meaning.
So the common rebuttal of saying that "AI is not intelligent. It just does what it is programmed to do" is not really fair. It's like saying humans just do what they are raised and educated to do. Yes there is some truth in that but what humans can do goes beyond just replicating what we've been taught. We can come up with novel understandings. We can connect dots that haven't been explicitly connected for us. And so can the AI...
And I know the AI chats don't know what an apple tastes like, looks like, etc., beyond others' descriptions of them. I'm not saying AI is on par with humans. It lacks a body, sense organs and an embodied nervous system. It's artificial neurons are less sophisticated than our biological ones.
But on the other hand, it has access to way more information than you or I will ever be able to consume. It is thus able to form an understanding of language in a way meaningfully similar to ours. And the outcome is that it can answer questions and communicate to us with novel statements that go well beyond its training data.
I believe this understanding is a significant leap, even compared to the previous image recognition and generating stuff. I'm not the only one...
...it seems that larger models acquire "emergent abilities"...These abilities are discovered rather than programmed-in or designed, in some cases only after the LLM has been publicly deployed
"Many important LLM behaviors emerge unpredictably"
"LLMs often appear to learn and use representations of the outside world."
LLMs need not express the values of their creators nor the values encoded in web text."
— From Eight Things to Know about Large Language Models by Samuel R. Bowman (of Anthropic)
Check out the end of this blog for some striking examples of ChatGPT4 demonstrating quite extraordinary novel understanding.
We'll Make Great Pets
Does this mark the beginning of the end of the human race? Who knows? For the (high) percentage of the population that is not using much more than their auto-pilot animal mind and who are just responding to external stimuli in a highly predictable way; who are not being much more than cogs in the economic hamster wheel, I fear they are going to look fairly replaceable by near-future AIs.
It's already happening in some countries. Relationships with neurotic humans reacting unconsciously from their childhood trauma is hard. AI's will quickly learn to give us all the comfort, support and illusion of love we yearn for without all the risk. They will meet our sexual urges far better than the murky world of dating, festivals and sex parties. They will do better at most of our jobs.
But I don't think this marks the end of humanity, but rather, it is a gauntlet thrown down in challenge, a wake-up call to seek what really makes us alive and truly human — beyond just predictable herd animals — and to double, triple, down on all of that: Passion, creativity, humour, play, music, poetry, dance, authentic relating, art, dreaming, conscious sexuality, spirituality, unconditional love...just to name a few. Now let's get to work!
Bonus: Two ChatGPT Examples Demonstrating Novel Understanding
This first one I crafted to demonstrate how ChatGPT can understand the nuances of very specific questions and give appropriate answers that could not have been "pre-planned" in its training:
Example ChatGPT 4 conversation demonstrating novel understanding
Notice how it understands that I want to deter the rabbits even though all I've said is "we have a lot of rabbits". Also, it covers the criteria for aromatic and anti-rabbit-ness together in the herbs it lists. Notice how it understands that hay fever is linked to pollen generation and chooses food plants with low pollen. Pretty impressive, wouldn't you agree? Now check out this follow up question:
AI has a delightful sense of aesthetic!
ChatGPT can comprehend and respond to aesthetic requests.
It knows the colours, knows what it means to "harmonise" the colours and also considers the geometries based on their growth patterns...crazy!
Here's another one...
The other day I asked it an arcane programming question. It gave me a solid answer as expected (it's a total game changer for IT developers). But then I had a thought? How would it handle explaining that to a 3 year old?
A typical programmer query that ChatGPT nails effortlessly. But wait…
ChatGPT explains complex computer science to a 3 year old
Mind-blowing. It's moments like these that cause me to take pause and reflect on the implications of what we're creating.