In a public exchange sparked by the HalluHard benchmark, Maksym Andriushchenko highlights how constant external lookups would drive up inference expenses too far for lengthy agent outputs, so models must keep leaning on parameters trained in advance.
@xeophon i would kindly disagree :) agents can’t use web search for every single statement. that would be too expensive. they need to still rely a lot on their own in-parameter knowledge, especially for long, open-ended outputs.
@maksym_andr Only web search matters tbh
Florian Brand replies that the right mix of search and built-in knowledge shifts with each job rather than following one fixed rule across all cases.
The back-and-forth stays focused on practical cost limits during extended generations, without new numbers or benchmarks released in the thread itself.
Some users disliked the coding subset example in discussions of AI agents needing internal knowledge as web search costs mount, noting models already access linters and pip in setups.
No Digg Deeper questions have been answered for this story yet.
@maksym_andr as always, "it depends".
i don't like the coding subset, for example (models have access to linters and pip in all setups these days). for the research questions, i expect models to do a short search
@xeophon i would kindly disagree :) agents can’t use web search for every single statement. that would be too expensive. they need to still rely a lot on their own in-parameter knowledge, especially for long, open-ended outputs.