4 ms·

OP here. I literally opened up Gemini and used the defaults. If the defaults are shit, maybe don't offer them as the default? Or, if LLMs are so smart, why doe

by edent 1y ago

OP here. I literally opened up Gemini and used the defaults. If the defaults are shit, maybe don't offer them as the default? Or, if LLMs are so smart, why doesn't it say "Hmmm, would you like to use a different model for this?" Either way, disappointing.

magicalhippo 1y ago
> Or, if LLMs are so smart, why doesn't it say "Hmmm, would you like to use a different model for this?" That is indeed an area where LLMs don't shine. That is, not only are they trained to always respond with an answer, they have no ability to accurately tell how confident they are in that answer. So you can't just filter out low confidence answers.
- mathewsanders 1y ago
  Something I think would be interesting for model APIs and consumer apps to exposed would be the probability of each individual token generated. I’m presuming that one class of junk/low quality output is when the model doesn’t have high probability next tokens and works with whatever poor options it has. Maybe low probability tokens that cross some threshold could have a visual treatment to give feedback the same way word processors give feedback in a spelling or grammatical error. But maybe I’m making a mistake thinking that token probability is related to the accuracy of output?
  - StilesCrisis 1y ago
    Lots of research has been done here. e.g. https://aclanthology.org/2024.findings-acl.558.pdf https://aclanthology.org/2024.findings-acl.558.pdf
  - never_inline 1y ago
    > Something I think would be interesting for model APIs and consumer apps to exposed would be the probability of each individual token generated. Isn't that what logprobs is?
pwnOrbitals 1y ago
You are pointing out a maturity issue, not a capability problem. It's clear to everyone that LLM products are immature, but saying they are incapable is misleading
- delusional 1y ago
  In you mind, is there anything an LLM is _incapable_ of doing?
hobofan 1y ago
Then criticize the providers on their defaults instead of claiming that they can't solve the problem? > Or, if LLMs are so smart, why doesn't it say "Hmmm, would you like to use a different model for this?" That's literally what ChatGPT did for me[0], which is consistent from what they shared at the last keynote (quick-low reasoning answer per default first, with reasoning/search only if explicitly prompted or as a follow-up). It did miss one match tough, as it somehow didn't parse the `<search>` element from the MDN docs. [0]: https://chatgpt.com/share/68cffb5c-fd14-8005-b175-ab77d1bf5888 https://chatgpt.com/share/68cffb5c-fd14-8005-b175-ab77d1bf58...
maddmann 1y ago
“Defaults are shit” — is that really true though?! Just because it shits the bed on some tasks does not mean it is shit. For people integrating llms into any workflow that requires a modicum of precision or determinism, one must always evaluate output closely/have benchmarks. You must treat the llm as an incompetent but overconfident intern, and thus have fast mechanisms for measuring output and giving feedback.

by edent 1y ago

magicalhippo 1y ago

> Or, if LLMs are so smart, why doesn't it say "Hmmm, would you like to use a different model for this?" That is indeed an area where LLMs don't shine. That is, not only are they trained to always respond with an answer, they have no ability to accurately tell how confident they are in that answer. So you can't just filter out low confidence answers.

mathewsanders 1y ago
Something I think would be interesting for model APIs and consumer apps to exposed would be the probability of each individual token generated. I’m presuming that one class of junk/low quality output is when the model doesn’t have high probability next tokens and works with whatever poor options it has. Maybe low probability tokens that cross some threshold could have a visual treatment to give feedback the same way word processors give feedback in a spelling or grammatical error. But maybe I’m making a mistake thinking that token probability is related to the accuracy of output?
- StilesCrisis 1y ago
  Lots of research has been done here. e.g. https://aclanthology.org/2024.findings-acl.558.pdf https://aclanthology.org/2024.findings-acl.558.pdf
- never_inline 1y ago
  > Something I think would be interesting for model APIs and consumer apps to exposed would be the probability of each individual token generated. Isn't that what logprobs is?

pwnOrbitals 1y ago

You are pointing out a maturity issue, not a capability problem. It's clear to everyone that LLM products are immature, but saying they are incapable is misleading

delusional 1y ago
In you mind, is there anything an LLM is _incapable_ of doing?

hobofan 1y ago

Then criticize the providers on their defaults instead of claiming that they can't solve the problem? > Or, if LLMs are so smart, why doesn't it say "Hmmm, would you like to use a different model for this?" That's literally what ChatGPT did for me[0], which is consistent from what they shared at the last keynote (quick-low reasoning answer per default first, with reasoning/search only if explicitly prompted or as a follow-up). It did miss one match tough, as it somehow didn't parse the `<search>` element from the MDN docs. [0]: https://chatgpt.com/share/68cffb5c-fd14-8005-b175-ab77d1bf5888 https://chatgpt.com/share/68cffb5c-fd14-8005-b175-ab77d1bf58...

maddmann 1y ago

“Defaults are shit” — is that really true though?! Just because it shits the bed on some tasks does not mean it is shit. For people integrating llms into any workflow that requires a modicum of precision or determinism, one must always evaluate output closely/have benchmarks. You must treat the llm as an incompetent but overconfident intern, and thus have fast mechanisms for measuring output and giving feedback.