Thoughts on “It’s OK to call it Artificial Intelligence”

Simon wrote an interesting piece about the name we use for Large Language Models. I have been doing a lot more work in this field than I initially expected, which resulted in a good amount of meetings to set expectations what we can and cannot do, what the technology can be used for and what its limitations are. Part of this is why I certainly have opinions on this topic, but likely from a slightly different perspective than Simon.

Before diving in I want to address John McCarthy’s quote: Language evolves. Which results in the meaning of words to potentially change over time. As the field of AI sees more traction than ever right now, the language around it is evolving as well. This is neither an argument for or against calling an LLM AI. It is however something we all should keep in mind when trying to figure out how to name the technology we are dealing with. “Just because” AI was defined or used in a certain way before, does not mean the definition cannot change.

Expectations

When we use the term “Artificial Intelligence” people have a certain expectation. Science Fiction taught us growing up that one day there will be a computer I can talk to. And it will respond - either in short, accurate sentences, or in a witty, borderline sarcastic way. (If space travel is involved maybe a little bit depressed.) But one thing all of these AIs had in common was that people could rely on their statement. They did not make things up (hallucinate) - obviously except the baddies who tried to wipe out humankind. But they knew they are lying!

This is an expectation many people still have when they hear “Artificial Intelligence”, especially when they are first confronted with a technology like ChatGPT. There seems to be an inherent expectation that the answer to “when is the next Taylor Swift concert in Heidelberg” is one you can rely on, not that ChatGPT makes up a concert in a few months and presents it as fact.

On the other side of the spectrum are people who spend time understanding the capabilities of LLMs. They know what to expect. They know what they are capable of. They know to vet the response before taking it as a fact. Alas, from my experience these people are far, far fewer in numbers than those who expect the answer to always be true.

Hard Problems

Phil Karlton phrased it very well:

There are only two hard things in Computer Science: cache invalidation and naming things.

(We can also add off by one errors to the list.)

The goal here is clear: we want people to understand that the LLM-powered tools they are interacting with today actually aren’t anything like the omniscient AIs they’ve seen in science fiction for the past ~150 years.

I am fully aligned with this goal. But I disagree with the approach. The less technical an audience is the less the term “LLM” is known or understood. But on the other hand the term “AGI” is also not well known or understood.

The term is right there for the taking. “You’re thinking about science fiction there: ChatGPT isn’t AGI, like in the movies. It’s just an AI language model that can predict next tokens to generate text.”

Calling an LLM AI and explaining it is “not AGI” simply shifts the problem from explaining what an LLM is and telling people it is “not AI”. I do not see a benefit either way, you end up explaining stuff in both scenarios. For me, personally, it is way easier to explain what an LLM is capable of than the differentiation between “Artificial Intelligence” and “Artificial General Intelligence”.

This is certainly a good way to explain it, but it is adding more terms to the discussion than necessary. I want to apologise to all zoologists our there reading my blog as I am certain you will hate me for this, but my go to explanation for an LLM is: “think of a parakeet on steroids”. And it works well. It instantly removes all expectations the words AI might set.

Reality

The term AI is overloaded with multiple meanings. Sentient computers. Large Language Models. Deep neural networks. Discrete-time Markov chains. We lost the clear definition of what AI means by using the term and not talking about how applicable it is to a technology.

The fact that tech-savvy crowds who know the various technologies capabilities and correct naming have this discussion shows how overloaded the term is. Now imagine how bad it is for people who are not in our or a related field. Things start to make less sense with every single technology marketed as AI.

In my opinion there are two different answers to the questions “is it OK to call it Artificial Intelligence?”.

For a technical crowd I would expect the use of the most descriptive terminology but would always give some leeway for a general term such as AI. There is usually a purpose to a discussion. And the differences between the individual technologies more often than not matter. If they do not matter for the topic being discussed there is in my opinion on harm in using a generalised term.

For a non-technical crowd the distinction most likely does not matter anymore at this point. The term AI is so overloaded it is a fine marketing term. And the more broadly it is used the more clear it becomes to people that it should not set any expectation in a system and its capabilities at all.

When you ask for a tissue in Germany you often ask for a “Tempo”. A brand of tissues so popular everyone knows what you ask for. No one expects to be handed a “Tempo” when the answer is “yes”, but a tissue.

When telling someone to do a web search to get an answer to a question you might hear “google it”. The expectation is usually not that people go to Google, but search the web. Google as a brand is popular enough to be used as synonym for a variety of ways to search information on the web.

We are past the point of having a clear definition or a singular technology for Artificial Intelligence. I think we can all acknowledge that we are disappointed it not living up to the expectations that were set for decades. I believe treating “Artificial Intelligence” as terminology for a bunch of different technologies sitting somewhere between Markov chains and AGI is most likely the best option moving forward and the most compatible one with the general understanding of the term.

Random thoughts

ad absurdum Do not take this one too seriously, I am more than aware that these are two entirely different things! But I had a good laugh when thinking about replacing some words in the LA Times article with LLM and AI.

Telsas defence for marketing assisted driving as “self driving” and “Autopilot” is essentially the same. “You did not complain for some years, so it is fine! Also 1st Amendment!”

Well, actually… If your sentence starts with “well, actually” or if it can be rephrased to and does not change meaning the whole world would appreciate if you simply do not post it. No sentence starting with “well, actually” contributed anything of value to an ongoing discussion. Ever.