I like 404, but to be blunt, the author doesn’t understand how absolutely tiny the “nudge” is.
Nor how LLM sampling works.
It’s not figuratively imperceptible; it’s literally below the noise floor of default sampling parameters, unless you run Claude at ~zero temperature, which no one does for this kind of stuff.
This kind of token bias only becomes statistically significant in larger bodies of text. Word to word, it does basically nothing.
If they have a problem with imprecise word choice, as they do in the article… Well, yes. Thats the issue with LLM sampling. It’s the elephant in the room.
To me, basic, temperature-based sampling with top-k/top-p was a “bandaid” to fix weird self-feedback loops with autoregressive research artifacts, like looping and repetition. It was a hack. And they just… commercialized it and never fixed it.
There are tons of interesting papers on alternatives to sampling. There tons of interesting implemented improvements (I’m partial to sigma-n/adaptive-p, tuned token bias, and constrained output grammar), but of course Claude offers none of these, and no one can do anything about it because it’s a closed model system.
I like 404, but to be blunt, the author doesn’t understand how absolutely tiny the “nudge” is.
Nor how LLM sampling works.
It’s not figuratively imperceptible; it’s literally below the noise floor of default sampling parameters, unless you run Claude at ~zero temperature, which no one does for this kind of stuff.
This kind of token bias only becomes statistically significant in larger bodies of text. Word to word, it does basically nothing.
If they have a problem with imprecise word choice, as they do in the article… Well, yes. Thats the issue with LLM sampling. It’s the elephant in the room.
To me, basic, temperature-based sampling with top-k/top-p was a “bandaid” to fix weird self-feedback loops with autoregressive research artifacts, like looping and repetition. It was a hack. And they just… commercialized it and never fixed it.
There are tons of interesting papers on alternatives to sampling. There tons of interesting implemented improvements (I’m partial to sigma-n/adaptive-p, tuned token bias, and constrained output grammar), but of course Claude offers none of these, and no one can do anything about it because it’s a closed model system.