3 ms·

> AI is no more familiar with our lives on Earth than it is Frodo's life in the Shire. It can only render solutions amenable to the perverse portrayal of realit

> AI is no more familiar with our lives on Earth than it is Frodo's life in the Shire. It can only render solutions amenable to the perverse portrayal of reality that text provides for it. What’s the relevance? Let’s be open to the possibility that proper intelligence won’t be 100% consumed with the affairs of beings who are less smart and capable. If it is, it’ll be artificially limited and that seems an unsustainable barrier. > Show me progress or GTFO, that's my policy. We had Eliza an extremely long time ago. No progress would mean you’re still using Google and Wikipedia to figure stuff out and Eliza level tech to brainstorm code. Frankly if you offered to take away Google or my LLM’s, I’d choose the former. I know it’s flawed but I can work around that rather than putz around with annoying UX that doesn’t let me ask follow up questions. And again, this is just very good autocomplete and does nothing to prevent future improvements. There is SO much money and talent looking for improvements. Governments give a shit and are competing with each other. It takes a stunning lack of imagination to think we’ve found the end.

talldayo 2y ago
> Let’s be open to the possibility that proper intelligence won’t be 100% consumed with the affairs of beings who are less smart and capable. Why? Our current "improper" intelligence isn't getting anywhere in the hands of smart people. They aren't using it to invent antigravity or hack the system like a video game. Turns out, using the Library of Babel to try and hunt down the truth is a waste of your time whether you're a genius or an idiot. If this "proper intelligence" still makes mistakes like current stochastic LLMs, it will continue to be an unreliable tool for scientists and doctors that rely on correct judgement. Don't get me wrong; I don't disagree with you on-principle. It would be cool if scientists had a hyper-intelligent all-knowing ChatGPT to help them invent new technology. I disagree with it because fundamentally, AI has not demonstrated the possibility it could ever fill this role. There is no feasibility study validating AI intelligence in any relevant field. Every current application of LLMs are imperfect and focused almost entirely on risk-management. This is a sign, to me, that neither of us will interact with "true AI" (or AGI if you're fanciful) in our lifetimes. That's my $0.02 whether the US goes all-in on AI or whether they tear it down with regulation; the tech stays the same. > And again, this is just very good autocomplete and does nothing to prevent future improvements. You may be looking at it from the wrong direction. People don't write research papers with autocomplete, they iterate on a thesis by testing their estimation against real-world experiment trials. Not only does it require a finer touch than autocomplete, you need to run real life tests somehow and correlate the data with your findings. AI does not do that - it's text. It won't fill beakers with Bromine or hack into the NSA on your behalf. So... I'd argue this does prevent future improvement, if you mean "improvement" to be "not lying all the time". The path to superintelligence is not forged by re-training and finetuning the same three foundation models over and over again. I'd even go as far as to say LLMs in the current architecture are fundamentally incapable of relevant thought. > It takes a stunning lack of imagination to think we’ve found the end. That's exactly what I've got. Stop playing with my imagination and show me the progress; I've been following this field since Google released BERT and we've gotten just about nowhere in the interceding years. I don't really care that much about the whole field. But it's surprising (or I guess not surprising at all) to hop on HN and see people ravenously excited for an iteration on a failed technology. Surely ChatGPT 5 will prove one of us right, eh?
- richardw 2y ago
  Ah, text. It's so shit for conversation. I feel like I'm speaking to a smart person and we'd do a lot better over a beer, and we're probably a lot closer in opinion somewhere. Maybe. > Surely ChatGPT 5 will prove one of us right, eh? I have zero trust in LLM's becoming any kind of AI, I think they're more of a warning shot that AI could be coming one day. I think what it requires is (only) a few fundamental architectural insights. I think there's enough effort, stock prices to defend, country strategic advantages resting on someone somewhere getting it right that people will eventually discover those insights. (No prediction on when, we still can't agree on quantum interpretations 100 years later.) Humans evolved from single cells, purely from randomness, and our efforts at replicating that have many good examples to work from. A simple change in architecture, on the right track, could unlock an entirely different entity. In general, my intuition follows something similar to a proof-by-induction. We have a level of capability, we have continuous effort at improving it, we never go backwards in levels of capability. Choose any (rational) level, or limit, or barrier. It will get hit eventually. (All LLM's do is move it from "20 years away, we can't see it from here" to "ah ok, maybe a glimpse. Could very easily be less than 20 years.") For it to not hit the limit, you need a barrier of sorts that cannot ever be crossed. That is much harder to make true than the alternative. The limit-crossing applies to levels of intelligence, and control, and possibilities of danger. Take the limit, apply continuous effort to cross it, it'll get hit. In terms of losing control or dangers: state-level actors in a crisis might very well switch the wrong switch to not lose a war. Does it mean guaranteed doom? No of course not. It just means past a certain level of capability, humans won't be making the decisions anymore. That excites the builders, and scares the chief security architects. People whose job it is to prevent bad things from happening have a different mindset to those looking to make something new. Both have a place.

talldayo 2y ago

> Let’s be open to the possibility that proper intelligence won’t be 100% consumed with the affairs of beings who are less smart and capable. Why? Our current "improper" intelligence isn't getting anywhere in the hands of smart people. They aren't using it to invent antigravity or hack the system like a video game. Turns out, using the Library of Babel to try and hunt down the truth is a waste of your time whether you're a genius or an idiot. If this "proper intelligence" still makes mistakes like current stochastic LLMs, it will continue to be an unreliable tool for scientists and doctors that rely on correct judgement. Don't get me wrong; I don't disagree with you on-principle. It would be cool if scientists had a hyper-intelligent all-knowing ChatGPT to help them invent new technology. I disagree with it because fundamentally, AI has not demonstrated the possibility it could ever fill this role. There is no feasibility study validating AI intelligence in any relevant field. Every current application of LLMs are imperfect and focused almost entirely on risk-management. This is a sign, to me, that neither of us will interact with "true AI" (or AGI if you're fanciful) in our lifetimes. That's my $0.02 whether the US goes all-in on AI or whether they tear it down with regulation; the tech stays the same. > And again, this is just very good autocomplete and does nothing to prevent future improvements. You may be looking at it from the wrong direction. People don't write research papers with autocomplete, they iterate on a thesis by testing their estimation against real-world experiment trials. Not only does it require a finer touch than autocomplete, you need to run real life tests somehow and correlate the data with your findings. AI does not do that - it's text. It won't fill beakers with Bromine or hack into the NSA on your behalf. So... I'd argue this does prevent future improvement, if you mean "improvement" to be "not lying all the time". The path to superintelligence is not forged by re-training and finetuning the same three foundation models over and over again. I'd even go as far as to say LLMs in the current architecture are fundamentally incapable of relevant thought. > It takes a stunning lack of imagination to think we’ve found the end. That's exactly what I've got. Stop playing with my imagination and show me the progress; I've been following this field since Google released BERT and we've gotten just about nowhere in the interceding years. I don't really care that much about the whole field. But it's surprising (or I guess not surprising at all) to hop on HN and see people ravenously excited for an iteration on a failed technology. Surely ChatGPT 5 will prove one of us right, eh?

richardw 2y ago
Ah, text. It's so shit for conversation. I feel like I'm speaking to a smart person and we'd do a lot better over a beer, and we're probably a lot closer in opinion somewhere. Maybe. > Surely ChatGPT 5 will prove one of us right, eh? I have zero trust in LLM's becoming any kind of AI, I think they're more of a warning shot that AI could be coming one day. I think what it requires is (only) a few fundamental architectural insights. I think there's enough effort, stock prices to defend, country strategic advantages resting on someone somewhere getting it right that people will eventually discover those insights. (No prediction on when, we still can't agree on quantum interpretations 100 years later.) Humans evolved from single cells, purely from randomness, and our efforts at replicating that have many good examples to work from. A simple change in architecture, on the right track, could unlock an entirely different entity. In general, my intuition follows something similar to a proof-by-induction. We have a level of capability, we have continuous effort at improving it, we never go backwards in levels of capability. Choose any (rational) level, or limit, or barrier. It will get hit eventually. (All LLM's do is move it from "20 years away, we can't see it from here" to "ah ok, maybe a glimpse. Could very easily be less than 20 years.") For it to not hit the limit, you need a barrier of sorts that cannot ever be crossed. That is much harder to make true than the alternative. The limit-crossing applies to levels of intelligence, and control, and possibilities of danger. Take the limit, apply continuous effort to cross it, it'll get hit. In terms of losing control or dangers: state-level actors in a crisis might very well switch the wrong switch to not lose a war. Does it mean guaranteed doom? No of course not. It just means past a certain level of capability, humans won't be making the decisions anymore. That excites the builders, and scares the chief security architects. People whose job it is to prevent bad things from happening have a different mindset to those looking to make something new. Both have a place.