4 ms·

> Humans have now had shoved in their face that many tasks considered "intellectual" can be done by systems that internally are simple, but very large. It also

by zvmaz 4y ago

> Humans have now had shoved in their face that many tasks considered "intellectual" can be done by systems that internally are simple, but very large. It also makes humans realize that much human intellectual activity is just autocomplete. Could you give an example? I had a conversation with it about the computational model of the mind, and at one moment, I asked it to tell me if Chomsky thinks that the model is complete (i.e., describes all aspects of the mind), and provide citations. It did, with citations and references; I checked each one of them: ALL of them were hallucinations. I asked it what was the percentage of surgeries done under laparoscopy in France, with references. It sure gave me the percentage and the references; I checked: ALL of them were not correct. I don't know if it makes humans realize that much human intellectual activity is just autocomplete; it certainly did not make me think so.

ShamelessC 4y ago
> ALL of them were hallucinations This is a known issue. How about asking it to generate code? I don’t really trust that you won’t just deliberately find an example that doesn’t work and say “see?! It’s so dumb!!!” And further this type of reaction reeks of the sort of insecurity that the parent comment was suggesting. You _need _ these things to be unintelligent. It threatened your individual skill set, so you prompted it like a human rather than deliberately leaning into its strengths and “assuming good faith” (whatever that means for an LLM). Fwiw, plugins will likely solve this issue. As well more advanced models.
- zvmaz 4y ago
  I asked the author of the comment to give me an example of the many "intellectual" tasks that can be done by "systems" that are "internally simple" (what does "simple" mean?) but "large" (with or without human intervention?). I then pointed to the known (and serious) issue of hallucinations; how does this all fit with the author's "reductionist" model of much human thought being "just autocomplete"? I don't feel threatened as my previous comment suggests; I just think that human thought is not "just autocomplete".
  - ShamelessC 4y ago
    I think you took that “just autocomplete” thing entirely too seriously. They were being poetic.
- WalterSear 4y ago
  I'm using it effectively to generate code. It needs coaching, but it's faster than writing out boilerplate on my own. I'm using it in a language I'm unfamiliar with (C++) and it helps me resolve my mistakes as often as I have to fill in the blanks that it leaves. I'm giving it very specific instruction ("write a class that does this", "write a method that does this"), though often by describing what is required in terms of the complete use case, rather than the specific I/O. I haven't come up against any hallucination so far in this task - instead, it tends to leaves complex bits "as an exercise for the reader". I either code up those parts myself, or I just ask it for that specific method on it's own and dump it in. The process feels as if I'm interviewing a job candidate and stealing their code to put in production.
janalsncm 4y ago
Hallucination is a huge problem for statistical language models. (I avoid the term “LLM” because it isn’t clear what “large” is, and the problem isn’t limited to large models anyways.) The question is whether it is a fundamental issue of the architecture (in the same way that jets can fly but they can’t fly to the moon), or one that will be solved with sufficient engineering patches. RLHF is just one “plugin” on top of the transformer, but maybe others like some kind of expert system can add further improvement. There are a lot of people in both camps and a lot of money to be made selling one vision or another. GOFAI folks like Stuart Russell would argue that it’s fundamental and can’t be done. I think it’s an empirical question with a lot of VC money riding on yes.
wongarsu 4y ago
Ask an LLM to write an IT security policy for your medium-sized company. Or describe the circumstances of a layoff and ask it to write a press release about it. Tell it about what you want from a holiday and ask where you should go. Ask for the lyrics for a song about Obama's presidency. All of them are intellectual tasks (most of them quite well paid), but well handled by LLMs. Asking for specific facts, never mind references, isn't something they do well, neither are they really good at niche topics (yet). But still they are pretty good at a lot of things. I'm not sure it's fair to label LLMs as autocomplete though. The last layer of them gives you weighted candidates for the next token, but that doesn't mean that nothing sophisticated or interesting happened in the other layers.
- __loam 4y ago
  If you use an LLM to write an email explaining why employees got laid off, you are a piece of shit.
WalterSear 4y ago
GPT4 or GPT3-turbo?
opportune 4y ago
LLMs are in the business of finding the best next token to continue the conversation in a way that models human text. That’s why it’s often likened to autocomplete. LLMs can compete with humans to the extent human tasks can be represented as a series of tokens. They can’t, by default (I think Bing and Bard can) visit a link and read it, which is why they hallucinate up reasonable looking links. They do know things insofar as it helps them find a good next token, and those things will tend be more generalized knowledge rather than specific facts and figures. I bet if you described a list of symptoms it would be pretty good at suggesting possible diagnoses, if you gave it the results of a blood test as a table it would try to say something reasonable but not know what to do.
twobitshifter 4y ago
The cases you note don’t sound like insurmountable hurdles do they? Were the wrong notes from references taken or did the references not exist?
PeterisP 4y ago
> many tasks considered "intellectual" can be done by systems that internally are simple, but very large There's plenty examples of such tasks even if we look back before the current proliferation of large language models. Translation used to be considered such a complex intellectual task, and had decades of research into more and more complicated systems to address various nuances and specific cases (e.g. transliteration of people's names, conversion of measurements) but in the end it turned out that this set of tasks is best solved by internally simple systems scaled up. Chess (and more recently, Go) is a historical example of what used to be considered an intellectually challenging task, but succumbs to simple-but-large solutions. Sure, GPT-4 can't do a lot of things well and a can't do a lot of things at all - however, it can do a bunch of tasks which we couldn't do before despite trying a lot with quite complicated, intelligent means, and it manages to beat that with a glorified autocomplete. Time and time again we want to consider some task as requiring "true intelligence" and then it eventually falls to brute force, reiterating Sutton's Bitter Lesson (http://incompleteideas.net/IncIdeas/BitterLesson.html http://incompleteideas.net/IncIdeas/BitterLesson.html) despite what most of us naturally would prefer to be true. Also, we do have some theoretical basis for equivalence of intellectual activity and "autocomplete", such as Hutter's proof that finding the optimal behavior of a rational agent is equivalent to compressing its observations.

by zvmaz 4y ago

ShamelessC 4y ago

> ALL of them were hallucinations This is a known issue. How about asking it to generate code? I don’t really trust that you won’t just deliberately find an example that doesn’t work and say “see?! It’s so dumb!!!” And further this type of reaction reeks of the sort of insecurity that the parent comment was suggesting. You _need _ these things to be unintelligent. It threatened your individual skill set, so you prompted it like a human rather than deliberately leaning into its strengths and “assuming good faith” (whatever that means for an LLM). Fwiw, plugins will likely solve this issue. As well more advanced models.

zvmaz 4y ago
I asked the author of the comment to give me an example of the many "intellectual" tasks that can be done by "systems" that are "internally simple" (what does "simple" mean?) but "large" (with or without human intervention?). I then pointed to the known (and serious) issue of hallucinations; how does this all fit with the author's "reductionist" model of much human thought being "just autocomplete"? I don't feel threatened as my previous comment suggests; I just think that human thought is not "just autocomplete".
- ShamelessC 4y ago
  I think you took that “just autocomplete” thing entirely too seriously. They were being poetic.
WalterSear 4y ago
I'm using it effectively to generate code. It needs coaching, but it's faster than writing out boilerplate on my own. I'm using it in a language I'm unfamiliar with (C++) and it helps me resolve my mistakes as often as I have to fill in the blanks that it leaves. I'm giving it very specific instruction ("write a class that does this", "write a method that does this"), though often by describing what is required in terms of the complete use case, rather than the specific I/O. I haven't come up against any hallucination so far in this task - instead, it tends to leaves complex bits "as an exercise for the reader". I either code up those parts myself, or I just ask it for that specific method on it's own and dump it in. The process feels as if I'm interviewing a job candidate and stealing their code to put in production.

janalsncm 4y ago

Hallucination is a huge problem for statistical language models. (I avoid the term “LLM” because it isn’t clear what “large” is, and the problem isn’t limited to large models anyways.) The question is whether it is a fundamental issue of the architecture (in the same way that jets can fly but they can’t fly to the moon), or one that will be solved with sufficient engineering patches. RLHF is just one “plugin” on top of the transformer, but maybe others like some kind of expert system can add further improvement. There are a lot of people in both camps and a lot of money to be made selling one vision or another. GOFAI folks like Stuart Russell would argue that it’s fundamental and can’t be done. I think it’s an empirical question with a lot of VC money riding on yes.

wongarsu 4y ago

Ask an LLM to write an IT security policy for your medium-sized company. Or describe the circumstances of a layoff and ask it to write a press release about it. Tell it about what you want from a holiday and ask where you should go. Ask for the lyrics for a song about Obama's presidency. All of them are intellectual tasks (most of them quite well paid), but well handled by LLMs. Asking for specific facts, never mind references, isn't something they do well, neither are they really good at niche topics (yet). But still they are pretty good at a lot of things. I'm not sure it's fair to label LLMs as autocomplete though. The last layer of them gives you weighted candidates for the next token, but that doesn't mean that nothing sophisticated or interesting happened in the other layers.

__loam 4y ago
If you use an LLM to write an email explaining why employees got laid off, you are a piece of shit.

WalterSear 4y ago

GPT4 or GPT3-turbo?

opportune 4y ago

LLMs are in the business of finding the best next token to continue the conversation in a way that models human text. That’s why it’s often likened to autocomplete. LLMs can compete with humans to the extent human tasks can be represented as a series of tokens. They can’t, by default (I think Bing and Bard can) visit a link and read it, which is why they hallucinate up reasonable looking links. They do know things insofar as it helps them find a good next token, and those things will tend be more generalized knowledge rather than specific facts and figures. I bet if you described a list of symptoms it would be pretty good at suggesting possible diagnoses, if you gave it the results of a blood test as a table it would try to say something reasonable but not know what to do.

twobitshifter 4y ago

The cases you note don’t sound like insurmountable hurdles do they? Were the wrong notes from references taken or did the references not exist?

PeterisP 4y ago

> many tasks considered "intellectual" can be done by systems that internally are simple, but very large There's plenty examples of such tasks even if we look back before the current proliferation of large language models. Translation used to be considered such a complex intellectual task, and had decades of research into more and more complicated systems to address various nuances and specific cases (e.g. transliteration of people's names, conversion of measurements) but in the end it turned out that this set of tasks is best solved by internally simple systems scaled up. Chess (and more recently, Go) is a historical example of what used to be considered an intellectually challenging task, but succumbs to simple-but-large solutions. Sure, GPT-4 can't do a lot of things well and a can't do a lot of things at all - however, it can do a bunch of tasks which we couldn't do before despite trying a lot with quite complicated, intelligent means, and it manages to beat that with a glorified autocomplete. Time and time again we want to consider some task as requiring "true intelligence" and then it eventually falls to brute force, reiterating Sutton's Bitter Lesson (http://incompleteideas.net/IncIdeas/BitterLesson.html http://incompleteideas.net/IncIdeas/BitterLesson.html) despite what most of us naturally would prefer to be true. Also, we do have some theoretical basis for equivalence of intellectual activity and "autocomplete", such as Hutter's proof that finding the optimal behavior of a rational agent is equivalent to compressing its observations.