3 ms·
Chomsky may be right or wrong. The problem the philosophers face is the high likelihood that the "a lumbering statistical engine for pattern matching, gorging o
by roenxi 4y ago
Chomsky may be right or wrong. The problem the philosophers face is the high likelihood that the "a lumbering statistical engine for pattern matching, gorging on hundreds of terabytes of data and extrapolating" [0] is going to just be more effective than whatever it is human brains do, similar to how rocks turn out to be more effective at adding numbers together than a mathematician despite the early indications from prehistory to whenever the calculator became viable.
Statistical models turn out to be enough to get past the Turing test. They're probably going to be better at interpreting vague requirements and making decisions than most humans too. They're better at art than the average, and better at explaining themselves than most humans.
[0] https://www.nytimes.com/2023/03/08/opinion/noam-chomsky-chatgpt-ai.html https://www.nytimes.com/2023/03/08/opinion/noam-chomsky-chat...
- apwheele 4y agoReminds me a bit of this blog post by Tal Yarkoni, https://www.talyarkoni.org/blog/2018/08/18/if-we-already-understood-the-brain-would-we-even-know-it/ https://www.talyarkoni.org/blog/2018/08/18/if-we-already-und.... That is perhaps more consistent with the "we are just really complicated neural networks argument".
- mtlmtlmtlmtl 4y agoPeople throw around this claim about the turing test. I've yet to see any empirical corroboration of it. Using an LLM and being amazed by its response is not doing the Turing test. The Turing test is mostly a thought experiment anyway.
- em-bee 4y agoif a machine can pass the turing test then that most likely means that the test was not good enough. it is not an indication that the machine is actually intelligent.
- mtlmtlmtlmtl 4y agoI think ChatGPT would fail the test as described by Turing almost immediately, because at some point it would just tell the other test subject it's a language model. In some sense ChatGPT is actually engineered to fail the test.
- GlenTheMachine 4y agoI came here to say this as well. I don't really care whether LLM's work like the human brain, because I suspect it won't matter. Also, although I understand the concerns over super intelligent AGI, I suspect that in the short- to medium-term, that won't matter either. What will matter is highly optimized LLM's putting high swaths of the white collar workforce out of work, and the economic upheaval that will ensue.
- emptysongglass 4y agoThis is what we should be contending with, not existential thought-exercises like Chomsky's. How do we grapple, as a society, with the huge economic upheaval these models cause as side-effect? Because we view non-working humans as "a problem", what do we do with all these displaced humans as efficiencies continue to increase?
- wslh 4y agoThe Turing test is a game. You can pass the Turing test while I cannot. It is not an universal result except if "everyone" in the planet don't pass it. This is not to say that all this is super interesting and we are living in interesting times. Just to add some science fiction to the discussion: what if we are in an AI/human flippening moment? Where humans used tools for million years and orchestrated the goals while now we are moving to being more like a tools, in the sense that we are necessary some times but completely unnecesary in others.
- smitty1e 4y agoPREDICTION: brute-forcing human intelligence only asymptotically approaches the genuine article. Shiny tech will provide copious benefits, but cannot be what it is not: organic genius.
- cm277 4y agoI found the Chomsky article simplistic: his arguments against the LLM lack of logic basically boil down to that LLMs are not aware of physics or logic, or social interactions. Well, guess what, they are not. Yet. Humans are indeed 'trained' on a much larger set of data than 'just' the whole text of the internet. Indeed orders of magnitude more: think of all of the video/audio/olfactory/social cues/physics we absorb every day of our lives through all of our senses. How is that not infinitely more data than the public available text on the Internet? What is to say that a way larger LLM-like model would not be an AGI? why should human intelligence NOT be just a statistical engine? After all, haven't we proven already that physics is more-or-less statistics at heart (quantum mechanics)? that doesn't make physics non-deterministic at human scale. So why would the result of an LLM-like model trained on reality be not-deterministic or wrong? How is a high-confidence statistical result different than reality? In other words, nothing in Chomsky's article proves that humans are NOT statistical pattern matching engines. I am not saying that we are BTW, but it's an intriguing possibility that should be researched. After all, statistical methods and ML have made a lot more progress in the last 10-15 years than the MIT-driven, deterministic school of thought did in the previous 30. Cambridge lost, plain and simple.
- slowmovintarget 4y agoThere is no mechanism for awareness with LLMs, though. Advances on probabilistic search will have equal facility with hallucination as with the real. Regardless of whether Chomsky's article "proves" your negative, there is plenty of evidence that LLMs are not even as good as dogs and cats at forming a theory of the world around them. That we might draw a conclusion that dogs, cats, and people work differently from LLMs seems more like common sense.
- slowmovintarget 4y agoOnly for the "worse is better" hypothesis. That is, we settle for not as good because there's plenty of it, and it is the plenty that matters over the quality. What we lose in the drowning is the ability to distinguish quality. We forget which way is up, where the air is, and swim down further until it is too late.