[ekelsen]

I don't think the argument is necessarily against the use of the tools entirely. My interpretation is that it's against delegating to them all understanding.

Humans can only usefully steer LLMs if they have some understanding or context the LLMs do not.

[danudey]

Put another way: if all that you're doing is prompting the AI and giving me the result then I have no use for you. If you're not contributing insight, understanding, experience, or creativity then it's far cheaper for me to prompt the AI myself.

[handoflixue]

The person in the story can clearly usefully steer LLMs if they got that far. So either they have that understanding and context, or it's not actually as valuable as was previously thought.

[simonh]

She says she can usefully steer LLMs, but how would she know? What knowledge of the research topic, methods and their applicability would she use to judge that, if all she’s doing is writing a prompt asking ChatGPT to write a summary of her research topic, methods and their applicability?

Supposing using LLMs as research tools is fine, the chalk talk is still necessary to test that the researcher has the fundamental skills and knowledge necessary to discriminate usefully between alternative responses, and refine prompts and such. You can’t just get a response from an LLM and then refine it for your specific needs by saying “now refine that reply for my specific needs”. You need to know and understand what those are, and why.

[handoflixue]

Line one is "I recently interviewed for a tenure-track position" which suggests she has a resume that ... actually qualifies for an interview? And did every single step up until now correctly?

Like, I'm sorry, you can't just walk into those off the street. Don't be absurd. I suppose you could say the entire system was happy to spend 4 years lying to her about her grades? But at that point it's "corrupt system completely fails a young student who was enthralled by new technology"

> test that the researcher has the fundamental skills and knowledge necessary to discriminate usefully between alternative responses

... and how, exactly, is she gonna demonstrate that if she can't prompt it in the first place?

> “now refine that reply for my specific needs”

That absolutely works if you've discussed your specific needs with the model for four years and it has dozens of memories and reference files to work from.

[wiseowise]

She can’t answer a single thing without opening her laptop.

[handoflixue]

Yeah, and the deaf lady can't answer a single thing without her interpreter. Again, what's your point? Is this a field where people are expected to work without a laptop for some reason?

Like, I get why we'd want a pilot not to need ChatGPT, but a scientist is not in any of those categories.

[wiseowise]

> Yeah, and the deaf lady can't answer a single thing without her interpreter.

Deaf lady is missing her hearing and therefore requires hearing aids to hear, so following your logic Rachel is missing her thinking therefore she requires ChatGPT to do thinking for her. Glad we're on the same page.