"I believe that artificial intelligence has three quarters to prove itself before the apocalypse comes, and when it does, it will be that much worse, savaging the revenues of the biggest companies in tech.", predicts Ed Zitron.
You know, I've wanted to be a software developer since I started programming at around age seven, and for most of that time I wanted to go into artificial intelligence. Seeing the direction that field has taken in the last handful of years has really dampened my desire to go into it, though. Before the first AI winter, it was all about building machines that actually try to emulate aspects of how humans think, so that we can automate various reasoning processes. At the time, the tools they used were those of expert systems and such — backward and forward chaining reasoning and similar fundamentally introspectable and semantic, meaningful tools — and when I was younger I assumed that would eventually develop into using neural networks to radically increase the power and flexibility of those expert system principles, through increased capabilities to recognize and classify things, while still allowing us to understand what the reasoning process of our AIs were based on those classifications, and provide actual semantic information and understanding and consistent reasoning processes. Or at the very least, that we would attempt to structure and train our neural networks in such a way that we actually incentivize them to learn how to perform reliable reasoning and semantic meaning operations.
Instead, all it really has become is cargo cult worship of black boxes that can provide simulacra of meaningful behavior, but are completely unreliable, biased, inscrutable, inconsistent, and usually not even that good at what they do (for instance, a couple-line python script using the gzip compression algorithm can give a better matching of text to text then machine learning can). The only thing so-called AI researchers seem to want to do is jump on the bandwagon of throwing more and more compute resources at a system that can fundamentally not perform the same tasks as actual human cognition (or even a subset of them but can only provide a similar of them), only superficially fool you, because it simply isn't INCENTIVIZED to do anything else.
This is actually why I've started to go back to my work on backward chaining and logic programming systems and figuring out how to integrate neural network machine learning into them. What I hope will happen is that a second AI winter will hit and wipe out all the hype and bandwagoning around machine learning and neural networks (leaving only the cases where such things are actually useful like in computer vision), and then in maybe a decade or so things will return with a more principled hybrid approach based in actual neuroscience and linguistics and so on.
Because right now the energy around large language models very much reminds me of ignorant people uncomprehendingly worshiping the machine God in some old Star Trek episode. #AI#LLMs#chatGPT#AIWinter