Conversation

Notices

1. Embed this notice
   Paul Cantrell (inthehands@hachyderm.io)'s status on Saturday, 24-Aug-2024 11:33:37 JST Paul Cantrell

   • Carl T. Bergstrom
   • Jenniferplusplus

   @ct_bergstrom
   The kinds of LLM misuse that people will fall for, and •who• falls for them, has certainly been…eye-opening. So many times, I’ve thought, “oh, wow, these people don’t even know what that job •is•.”

   ICYMI, you’d probably be interested in this now-classic essay from @jenniferplusplus:
   https://jenniferplusplus.com/losing-the-imitation-game/ It’s about how assuming LLMs can take over programming fundamentally misunderstands what programming is. Much of its line of critique translates over to science.

   • Embed this notice
     Paul Cantrell (inthehands@hachyderm.io)'s status on Saturday, 24-Aug-2024 12:47:07 JST Paul Cantrell

     in reply to

     • Berkubernetus
     • Rich Felker
     • Carl T. Bergstrom
     • Jenniferplusplus

     @dalias @fuzzychef @ct_bergstrom @jenniferplusplus
     I have quipped something similar, though in the interest of intellectual honesty, I have to admit that the principle of “other people’s jobs look easy because you don’t understand them” also applies to me…and…[grrrm]…CEOs.

     (I’ve known some really competent ones, and yes, I can say that doing the job •well• is not something an LLM could do. Which I guess may be a necessary qualifier in all these cases,.)

   • Embed this notice
     Rich Felker (dalias@hachyderm.io)'s status on Saturday, 24-Aug-2024 12:47:08 JST Rich Felker

     in reply to

     • Berkubernetus
     • Carl T. Bergstrom
     • Jenniferplusplus

     @fuzzychef @inthehands @ct_bergstrom @jenniferplusplus The one exception is CEOs. LLMs excel at spewing confident bullshit telling investors what they want to hear.

   • Embed this notice
     Berkubernetus (fuzzychef@m6n.io)'s status on Saturday, 24-Aug-2024 12:47:09 JST Berkubernetus

     in reply to

     • Carl T. Bergstrom
     • Jenniferplusplus

     @inthehands @ct_bergstrom @jenniferplusplus 100% of arguments of "LLMs can replace job X" are based on the speaker not really having any idea what job X entails, whatever it is. Science, medicine, programming, even customer service. LLMs cannot do any of it.

     This is the arrogance of assuming that everyone else's job is easy.

   • Embed this notice
     Paul Cantrell (inthehands@hachyderm.io)'s status on Sunday, 25-Aug-2024 00:40:10 JST Paul Cantrell

     in reply to

     • Berkubernetus
     • Carl T. Bergstrom
     • Jenniferplusplus
     • ForeverExpat

     @ForeverExpat @fuzzychef @ct_bergstrom @jenniferplusplus
     Yeah, all that is the non-ridiculous sales pitch for LLMs. And maybe…occasionally? But there’s real substance in these supposedly menial tasks.

     Summarization, for example, is deep: it requires deciding what the central idea is. People who’ve systematically studied LLMs summarization concluded that they’re •terrible• at that. Unless you’re a very slow writer, •that• is the time — not sentence-forming.

   • Embed this notice
     ForeverExpat (foreverexpat@mastodon.world)'s status on Sunday, 25-Aug-2024 00:40:11 JST ForeverExpat

     in reply to

     • Berkubernetus
     • Carl T. Bergstrom
     • Jenniferplusplus

     @fuzzychef @inthehands @ct_bergstrom @jenniferplusplus
     I’m not so sanguine. Some % of info work is predictable, repetitive. Eg drafting reports, summarizing long docs, analyzing excel. In 10m with good prompts, LLM gives output upon which I can add my expertise, clean the flow, add my voice, validate the output. A day’s work now only hrs. Increased productivity means we don’t need to hire as many new staff, interns, etc. LLM use doesn’t have to mean “replace a whole job” to have an impact.

   • Embed this notice
     Paul Cantrell (inthehands@hachyderm.io)'s status on Sunday, 25-Aug-2024 00:43:03 JST Paul Cantrell

     in reply to

     • Berkubernetus
     • Carl T. Bergstrom
     • Jenniferplusplus
     • ForeverExpat

     @ForeverExpat @fuzzychef @ct_bergstrom @jenniferplusplus
     Now think of the •cost• of mis-summarizing. A manager gets an LLM summary that misses the main idea, fails to take some crucial action, and…💥💥 It’s everything that’s already broken in orgs, but faster and even •more• broken.

     What people are re-learning with LLMs is that cleaning up bad work is •costly•. A day’s work only hours!…plus, later on, weeks or months of mop-up.

     https://hachyderm.io/@inthehands/112955579821944875

   • Embed this notice
     Paul Cantrell (inthehands@hachyderm.io)'s status on Sunday, 25-Aug-2024 01:53:32 JST Paul Cantrell

     in reply to

     • ForeverExpat

     @ForeverExpat
     “the cost of revising a text by an expert is less work than the cost of writing from scratch”

     Heh, wow, that’s…preposterous. I think this discussion ends with us just parting ways here.

   • Embed this notice
     ForeverExpat (foreverexpat@mastodon.world)'s status on Sunday, 25-Aug-2024 01:53:33 JST ForeverExpat

     in reply to

     • Berkubernetus
     • Carl T. Bergstrom
     • Jenniferplusplus

     @inthehands @fuzzychef @ct_bergstrom @jenniferplusplus
     1st, the cost of revising a text by an expert is less work than the cost of writing from scratch. Of course you should never just pass off LLM output without applying some expertise. Still revising the output saves you anywhere from 20-80% of the time.
     2nd, you overestimate the importance of precision for most info worker tasks. Summarizations by definition leave out details - which ones are important is ALWAYS going to be debatable.

   • Embed this notice
     Paul Cantrell (inthehands@hachyderm.io)'s status on Sunday, 25-Aug-2024 03:01:54 JST Paul Cantrell

     in reply to

     • Berkubernetus

     @fuzzychef
     It’s probably not a coincidence (but probably sample bias) that all the really excellent CEOs I’ve interacted with have run fairly small companies.

   • Embed this notice
     Berkubernetus (fuzzychef@m6n.io)'s status on Sunday, 25-Aug-2024 03:01:55 JST Berkubernetus

     in reply to

     • Rich Felker
     • Carl T. Bergstrom
     • Jenniferplusplus

     @inthehands @dalias @ct_bergstrom @jenniferplusplus Well, given corporate performance recently across the industry, you CAN make a strong case for LLMs replacing CEOs.

     (Mostly because the average CEO is a platinum-spoon baby who is utterly unqualified for the job.)

   • Embed this notice
     Jenniferplusplus (jenniferplusplus@hachyderm.io)'s status on Sunday, 25-Aug-2024 05:37:16 JST Jenniferplusplus

     in reply to

     • Carl T. Bergstrom

     @inthehands @ct_bergstrom the worst part is how often those people are supposedly experienced practitioners of the job they do poorly understand