5 ms·

> Its value is mostly, and paradoxically, for training LLMs, to get "unspoiled" models. This is the tell that present-generation LLM "AI" is not actually intel

• ToucanLoucan 3y ago
  I've been saying exactly this from the beginning. I don't like being lied to, and calling ChatGPT AI was always a lie, from minute one. It is not intelligent in any sense of the word, whatsoever, nor may it aspire to intelligence. It's a highly adept predictive text model. It generates words based upon words it has been trained on as being the right words to minimize the errors in the words it generates. Not useless, mind you, and frankly tons of B2B salespeople especially would benefit from using it to clean up their nearly incoherent rambling emails; but not intelligence. The way you can "talk to" these machines is certainly clever, and the way they generate their replies is fascinating technology and certainly beats the shit out of a traditional chatbot, all of these things are true and I think they have a bright future as such. Natural language UI is a cool concept and it's cool to see some serious progress there. But that's all it is, and all it really can be. I don't know for certain what the path is to true emergent intelligence from the machine, but I assure you that shoving billions of words scraped from the internet through an LLM is not that. Now, if for whatever reason you want ream upon ream of pretty dull and repetitive content churned out on a given subject, that's far longer than it needs to be to convey the limited information within as apparently tons of online businesses do, written damn near to perfection? ChatGPT's your hookup, no question. However you will never get blinding insights, you will never get new ideas, and it will never be your friend. Sorry.
  • baq 3y ago
    It isn't a person. It also isn't a computer as we were used to. It's a new thing. Being able to query the latent space of a large subset of human knowledge using natural language is the closest to 'intelligent computer' as we've ever come to. Yes it is lossy compression, it confabulates the bits that it lost, but it's still useful.
    • ToucanLoucan 3y ago
      > Being able to query the latent space of a large subset of human knowledge using natural language is the closest to 'intelligent computer' as we've ever come to. I would agree with this statement, but it's still laughably far from it. As the other comment said, it lacks the ability to rank the veracity of the information it presents: it will present it in a grammatically correct way, but it has no way of knowing if the information it's presenting is accurate which explains ChatGPT's tendency to just make some shit up when you ask it things. And in the odd event it is correct, it's simply digesting information from other sources, likely search engines either it's own or not. Again, this is not useless and not not-clever: however in that way it does remain pretty useless because of the above issues of accuracy. The best it can do is get you the most highly ranked information on a topic based on it's pool of knowledge, and that may or may not be the most correct one. Also I'd add: these two issues compound one another, because the LLM is obscuring the source of it's information. You can have it add the link to it's citations, that would be helpful, but you don't know where it got those citations or why it picked them, and the LLM can't explain that for you. And, those choices were based on factors that are definitively not the accuracy or reliability of those sources, merely their ranking in whatever search algorithm is being used. Even if you presume this AI was trained on data relevant to the topic you're asking about, that doesn't mean it (and in fact, as I understand LLM, it's basically incapable of) understands that topic in a way that makes sense to a subject matter expert. If you asked an engineer what size lumber would be required to construct a floor that could support a small vehicle, that engineer would probably have a rough guess based on their previous experiences, and further, could then do the work: research the weight of the vehicle, the situation the structure will be installed in, the environment it will exist in, etc. and come back with a solid answer to your question. An LLM, by comparison, would consult the writings it was trained on for similar questions, and create an answer it thinks matches yours. But what does "similar" mean here? Is the wood and the dimensions the same? Is what it's looking at based on a small vehicle or a cement truck? Is it looking for articles about structures in similar climates or did it pick some that were closer matched to the structure but in completely different environments? On and on. Using an LLM this way takes the problematic aspects of googling things and basically factors them out, because now you can't see what it's googling, which results it's picking or why, and can't verify their relevance. How many times have you searched specific things and left a google result completely befuddled as to why on earth it included it? Now imagine that, except there's essentially zero chance the AI caught that, and it incorporated that irrelevant data anyway.
• hanniabu 3y ago
  > This is the tell that present-generation LLM "AI" is not actually intelligent. Agreed, I like to refer to it as collective intelligence because it's only as "smart" as what people have already published
• Mr_Minderbinder 3y ago
  > This is the tell that present-generation LLM "AI" is not actually intelligent. It doesn't truly understand anything. It's just a general purpose lossy compression algorithm for text that is queryable through replaying the model to complete prompts. This is not an argument for it not being intelligent nor is it a coincidence. It has been shown mathematically that there is a deep, fundamental connection between compression and intelligence (an intelligent agent that behaves optimally effectively ends up computing the Kolmogorov complexity of the program that best fits and predicts the environment i.e. optimal compression). I do not know what you mean by "truly understand" since it obviously has an understanding of the various words it uses even if it's knowledge of the "meaning" of a word is simply a vector in a very high-dimensional space. I suppose you think this representation of meaning is profoundly different from meaning as it is represented in the human brain, that the former is a mere facsimile of the meaning while the latter must be the meaning itself, assuming such a thing exists. Even among humans I have conversations with people who use words or phrases for which they clearly have a poor sense of the meaning and make contradictory or nonsensical statements as a result. All this "tells" is that it cannot (effectively?) bootstrap itself to higher performance levels with it's own output. This is not that surprising since humans have the same problem (for similar fundamental reasons). You do not teach children to write better by making them read their own literary output.