Neues aus der Welt
Jetzt unterschreiben: Für Teilhabe ohne Digitalzwang
LLMs require human input, else their implicit priors become dominant
When I did inverse modelling in my PhD we always made sure that we know what’s prior model info and what’s actual data.
You had to know which evaluation includes prior data in addition to measurement data, because assimilating multiple sources with the same prior could cause the prior to become dominant.
Todays LLMs repeatedly ingest their own outputs -- outputs that contain their own implicit priors.
It’s obvious that preserving the model quality requires pure human input.
Though from inverse modelling I’d guess that improving the quality of the prior -- of the output the model would generate without training -- could improve the quality of the results when trained with less data.
This is not new, but I now connected the dots from what I know to LLMs. For a discussion and refrences see Wikipedia: Model collapse.
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- …
- nächste Seite ›
- letzte Seite »
Diese Seite nutzt Cookies. Und Bilder. Manchmal auch Text. Eins davon muss ich wohl erwähnen — sagen die meisten anderen, und ich habe grade keine Zeit, Rechtstexte dazu zu lesen…
