3 ms·

> I'm eagerly waiting for the next big breakthrough when someone trains a model that switches between "thinking" and "outputting" modes mid-answer :) ??? That

by mlyle 1y ago

> I'm eagerly waiting for the next big breakthrough when someone trains a model that switches between "thinking" and "outputting" modes mid-answer :) ??? That was last year's breakthrough.

seba_dos1 1y ago
See??? What a breakneck pace! :D
- mlyle 1y ago
  So, I guess it was just a low effort comment, and I shouldn't take you seriously? I thought you were making an actual effort to converse. Generating tokens to represent steps isn't just "cosplaying" --- it's representing more state than can fit into a single step of the model. Ditto for when I talk to myself when figuring something out, either out loud or just mentally.
  - seba_dos1 1y ago
    It was a joke (with an emoticon even) - and it was there to make a point. > Generating tokens to represent steps isn't just "cosplaying" --- it's representing more state than can fit into a single step of the model. Well duh, obviously, but once you look closer, some of these "steps" aren't really any actual steps - they're just cosplaying. You can often catch a model outputting something like "I'm collecting the list of resources to base this table on", even though it doesn't do anything close to that (which is hopefully obvious). It just writes this, because it's a reasonable thing for someone who researches a topic to include in their stream of conscio... I mean, chain of thought. This then also helps it move the weights for future tokens in the direction of a writing from someone who appears to have collected such list, the same way people had more success with early image generation by adding "award-winning" or "high quality" to the prompt. Remember "prompt hacking"? Not in the security sense, but in the "I'm a prompt engineer" sense. The models prompt-hack themselves now. These breakthroughs consist of obvious next steps, it's just that someone with deep enough pockets allowed them to be tried on big enough datasets with long enough compute.

by mlyle 1y ago

> I'm eagerly waiting for the next big breakthrough when someone trains a model that switches between "thinking" and "outputting" modes mid-answer :) ??? That was last year's breakthrough.

seba_dos1 1y ago

See??? What a breakneck pace! :D

mlyle 1y ago
So, I guess it was just a low effort comment, and I shouldn't take you seriously? I thought you were making an actual effort to converse. Generating tokens to represent steps isn't just "cosplaying" --- it's representing more state than can fit into a single step of the model. Ditto for when I talk to myself when figuring something out, either out loud or just mentally.
- seba_dos1 1y ago
  It was a joke (with an emoticon even) - and it was there to make a point. > Generating tokens to represent steps isn't just "cosplaying" --- it's representing more state than can fit into a single step of the model. Well duh, obviously, but once you look closer, some of these "steps" aren't really any actual steps - they're just cosplaying. You can often catch a model outputting something like "I'm collecting the list of resources to base this table on", even though it doesn't do anything close to that (which is hopefully obvious). It just writes this, because it's a reasonable thing for someone who researches a topic to include in their stream of conscio... I mean, chain of thought. This then also helps it move the weights for future tokens in the direction of a writing from someone who appears to have collected such list, the same way people had more success with early image generation by adding "award-winning" or "high quality" to the prompt. Remember "prompt hacking"? Not in the security sense, but in the "I'm a prompt engineer" sense. The models prompt-hack themselves now. These breakthroughs consist of obvious next steps, it's just that someone with deep enough pockets allowed them to be tried on big enough datasets with long enough compute.