A.I. tools in 2024: Generally useless, but specifically useful

I started reading The Myth of Artifical Intelligence by Erik J. Larson. These are some rough/early thoughts.

Firstly, I’m surprised at how accurately the author paints a picture of the state of consumer AI tools development in 2024, even though the book was published back in 2021. I’ve only read the introduction so far, but it resonated strongly when I considered the slow of modern “GenAI” tools (e.g. ChatGPT, Github Copilot, etc.) that leverage LLMs to increase productivity.

The core argument Larson makes is in questioning the direction (both technical and culture) of AI development, and that our current approach is a mistake:

This book explains two important aspects of the AI myth, one scientific and one cultural. The scientific part of the myth assumes that we need only keep “chipping away” at the challenge of general intelligence by making progress on narrow feats of intelligence, like playing games or recognizing images. This is a profound mistake: success on narrow applications gets us not one step closer to general intelligence.

I hadn’t really thought about the “cost” of just puttering along iteratively, even though I (and most of us) are fairly aware of the lack of genuine innovation in AI capability.

At this point we have kind of just accepted that a lot of the AI tools were mostly not going to deliver to our expectations. Yet we are also being told to be excited about incremental improvements. So I can see the problem with that.

The book’s introduction has opened my eyes a bit to these costs, even with this initial introduction. The importance of pursuing the general intelligence, and specifically that these irritative changes to hyperfocus utility on specific cases will not get us any closer to it:

As we successfully apply simpler, narrow versions of intelligence that benefit from faster computers and lots of data, we are not making incremental progress, but rather picking low-hanging fruit. The jump to general “common sense” is completely different, and there’s no known path from the one to the other.

This part takes it to another level: We could be making incremental progress toward general intelligence, but our aim is off. We’re focusing on the wrong thing and just picking tiny use cases to tackle, distracting ourselves and starving the energy (and funding) in the room from going towards the bigger problem of general intelligence.

I look forward to reading more, but a few early thoughts:

• I do agree that companies are doing a disservice by selling general intelligence when they can really only offer specialized hyper-specific intelligence. Many companies are dying because of this. But others are thriving by doubling down and building better UX targeted around hyper-specific use cases, like Cursor. Cursor figured out how to build a genuinely useful IDE for AI-assisted coding, rather than continuing to wait for the models to magically get better by themselves (like Github Copilot, which has seen almost no progress at all since its inception).
• I’m not sure that delaying the unlock of general intelligence is necessarily bad, only because we have no clue what will happen when we do: We should be somewhat afraid of the unknown. This issue is divisive—if it’s inevitable, should we run toward it, or delay it? Could we control it the way we “control” nuclear arms? The likely answer is no in my mind, but the author claims this doomsday scenario is highly unlikely, and that he will prove it, so I’m curious to hear his case.
• Hand-waving hands: There is a term in computer science called ‘hand-waving,’ in which we just dismissively wave our hands at parts of the solution we don’t know/understand, in order to get past it. I think this is what the author is describing in terms of AI, where instead of trying to fundamentally understand and resolve the core problem, we just hack around it. The best example I’ve seen of hand-waving? Hands! AI generated images famously still don’t know how to draw hands properly, and no one really knows why! We’re just expected to be all right with it, like with text ‘hallucinations’ in ChatGPT. At best, we now have custom code to alleviate this, which is not a solution but a hack around it. There is a clear cost to just covering up things like this time and time again.
• The cultural aspect the author raises is very interesting because you can be considered anti-AI (“pessimist”) even if what you really want is for AI to be better via true general intelligence (“optimist”). This creates a funny mismatch where capitalist incentives to sell products without true AI progress will present themselves as AI-forward, when in reality, they are holding it back.

The question is: Will we ever reach the sky of general intelligence, if our gaze is fixed upon the floor of specifics?