3 ms·

The state of the art you refer to is based on statistical analysis of text found on the internet. AFAIK no AI system actually performs logical processing of the

The state of the art you refer to is based on statistical analysis of text found on the internet. AFAIK no AI system actually performs logical processing of the submitted materials. When it says that 2+2=4, it's simply because that is more commonly found than 2+2=7. If you were to replace all sources that say 2+2=4 with ones that say 2+2=7, then that's how it will answer in a conversation. We will probably run out of available electrical power before we build a ginormous fail-proof AI.

ditchfieldcaleb 4y ago
Neural nets (particularly RNNs) can emulate arbitrary turing machines, so therefore are capable of said logical processing. LLMs might have to be coaxed into doing this, but with enough trying you can get them to be more rational/logical rules focused. Give it a go - I'm sure you can come up with some math problem that is un-google-able. Use chain-of-thought prompting to have it solve a quintic polynomial or something. I think this (being the result I imagine you'll get when you try the above) demonstrates that LLMs have learned the capability for general reasoning, even if they don't always automatically use it. To be fair - neither do we.
- cwillu 4y ago
  Obviously not exhaustive of possible approaches, but https://www.lesswrong.com/posts/DTLRw6ZstkdghEgqA/chatgpt-vs-the-2-4-6-task https://www.lesswrong.com/posts/DTLRw6ZstkdghEgqA/chatgpt-vs... was my attempt to get chatgpt to play a simple inductive reasoning game.
  - ditchfieldcaleb 4y ago
    Not the brightest, but a valiant attempt for a lower-iq...entity, I'd say.
epistemer 4y ago
I see it as the opposite. It is amazing what the model can do with poor quantitative reasoning. I just can't imagine adding super human quantitative reasoning is going to be that big of a stumbling block over the next decade. If anything that is probably the low hanging fruit here for a huge jump into the unknown.
- Someone 4y ago
  Looking at the state if the art in automated theorem provers, I don’t think that’s low hanging fruit. We probably can make something that can calculate well and won’t make mistakes in combining various numbers found online, and can do rote evaluation of expressions not found anywhere online, but adding ‘reasoning’? Even disregarding that it would have to, somehow, assign different trust levels to various online sources (for example, are, https://en.uncyclopedia.co/wiki/Wikipedia https://en.uncyclopedia.co/wiki/Wikipedia or https://en.wikipedia.org/wiki/Uncyclopedia https://en.wikipedia.org/wiki/Uncyclopedia trustworthy?), it, IMO, already would fall at the hurdle of doing ‘simple’ math. For example “the sine of 100 factorial” has a well-defined value, but computing it in IEEE doubles doesn’t make sense because representable numbers are way too far apart around 100! (Google says it is about 0.68395718932, but it also thinks that sin(1+100!) ≈ 0.68395718932. I trust neither answer) That’s solvable by using better software. Wolfram Alpha claims these are about -0.17, respectively -0.92, for example, but in my book, an AI wouldn’t be intelligent if it always used one; it would have to know when to fall back on the heavy guns. For the “what’s sin(100!)” question, I think the first response might be a counter-question “why do you want to know?”, but that depends on earlier discussion.

ditchfieldcaleb 4y ago

Neural nets (particularly RNNs) can emulate arbitrary turing machines, so therefore are capable of said logical processing. LLMs might have to be coaxed into doing this, but with enough trying you can get them to be more rational/logical rules focused. Give it a go - I'm sure you can come up with some math problem that is un-google-able. Use chain-of-thought prompting to have it solve a quintic polynomial or something. I think this (being the result I imagine you'll get when you try the above) demonstrates that LLMs have learned the capability for general reasoning, even if they don't always automatically use it. To be fair - neither do we.

cwillu 4y ago
Obviously not exhaustive of possible approaches, but https://www.lesswrong.com/posts/DTLRw6ZstkdghEgqA/chatgpt-vs-the-2-4-6-task https://www.lesswrong.com/posts/DTLRw6ZstkdghEgqA/chatgpt-vs... was my attempt to get chatgpt to play a simple inductive reasoning game.
- ditchfieldcaleb 4y ago
  Not the brightest, but a valiant attempt for a lower-iq...entity, I'd say.

epistemer 4y ago

I see it as the opposite. It is amazing what the model can do with poor quantitative reasoning. I just can't imagine adding super human quantitative reasoning is going to be that big of a stumbling block over the next decade. If anything that is probably the low hanging fruit here for a huge jump into the unknown.

Someone 4y ago
Looking at the state if the art in automated theorem provers, I don’t think that’s low hanging fruit. We probably can make something that can calculate well and won’t make mistakes in combining various numbers found online, and can do rote evaluation of expressions not found anywhere online, but adding ‘reasoning’? Even disregarding that it would have to, somehow, assign different trust levels to various online sources (for example, are, https://en.uncyclopedia.co/wiki/Wikipedia https://en.uncyclopedia.co/wiki/Wikipedia or https://en.wikipedia.org/wiki/Uncyclopedia https://en.wikipedia.org/wiki/Uncyclopedia trustworthy?), it, IMO, already would fall at the hurdle of doing ‘simple’ math. For example “the sine of 100 factorial” has a well-defined value, but computing it in IEEE doubles doesn’t make sense because representable numbers are way too far apart around 100! (Google says it is about 0.68395718932, but it also thinks that sin(1+100!) ≈ 0.68395718932. I trust neither answer) That’s solvable by using better software. Wolfram Alpha claims these are about -0.17, respectively -0.92, for example, but in my book, an AI wouldn’t be intelligent if it always used one; it would have to know when to fall back on the heavy guns. For the “what’s sin(100!)” question, I think the first response might be a counter-question “why do you want to know?”, but that depends on earlier discussion.