@grimalkina @dave @trochee @3psboyd I feel like there's a categorical distinction when moving from something purpose-built to something generative, because of the way their success criteria are constructed. Perhaps grad students labeling plant parts already have a 5% error rate, and a model with a 7% error rate that can go 100x faster is fine, because you can say "sure, we're fine with this, we know what the consequence of being wrong is here and we were already wrong sometimes"
Conversation
Notices
-
Embed this notice
Glyph (glyph@mastodon.social)'s status on Saturday, 21-Jun-2025 01:44:48 JST 
Glyph
-
Embed this notice
Dr. Cat Hicks (grimalkina@mastodon.social)'s status on Saturday, 21-Jun-2025 01:44:47 JST 
Dr. Cat Hicks
@trochee @glyph @dave @3psboyd yeah the problem-solving comes from a person saying ok I'm going to take this building material and shim it into my problem-solving space right? Like I understand the frustration you have with people thinking the thing is thinking but I'm talking more broadly about the utility of approaches that remove human decision-making from some parts of the data construction process and what we can learn about the rules that make it valuable to us
-
Embed this notice
Paul Cantrell (inthehands@hachyderm.io)'s status on Saturday, 21-Jun-2025 01:44:47 JST 
Paul Cantrell
(Disclaimer: slightly out of my depth)
One thing researchers who were seriously studying LLMs pre-hypestravaganza were really excited about is the progression from completely bespoke model training (prep your problem-specific training data and train from scratch) to “transfer learning” (take a generic model and do a little extra training to specialize it, eg generic species classifier tuned to recognize birds) to “zero-shot classifiers” where you get a completely generic model to focus on a specific thing in the query itself, no additional training. Serious people treat this kind of usage with the same empirical caution they’d have applied to any other classifier system.
This really can work for the kind of use case Cat was talking about (slight increase in error rate for a huge increase in throughput), and in that usage pattern, LLMs are very much in the family of old-school classifiers.
Jeremy, I’d argue the difference you’re talking about isn’t just in the model itself, but in how it’s used, how it’s presented, how the UI works, and how it’s marketed. When you change the mindset from “zero-shot classifiers with flexible queries” to “prompt engineering,” you have substantively changed the nature of the beast.
-
Embed this notice
Jeremy Kahn (trochee@dair-community.social)'s status on Saturday, 21-Jun-2025 01:44:48 JST 
Jeremy Kahn
this kind of work is specifically _not_ LLMs, and as a language nerd I'm super annoyed that these two very different uses are conflated here
generated language from an LLM is _not_ problem-solving. it is an interesting language artifact, but does not involve thought or insight into anything but the weather system behind the chatbot
-
Embed this notice
Paul Cantrell (inthehands@hachyderm.io)'s status on Saturday, 21-Jun-2025 01:54:17 JST
Paul Cantrell
Right. That “care to validate” is is the empirical human work we call “science,” and what makes the AI hype so pernicious is that it’s created a permission structure for people to stop doing it — or be unaware that it was ever important in the first place.
-
Embed this notice
Jeremy Kahn (trochee@dair-community.social)'s status on Saturday, 21-Jun-2025 01:54:18 JST
Jeremy Kahn
100%
all of this.
zero-shot, or few-shot ,classifiers are pretty great, although they should be absorbed with caution because of embedded bias; serious users should take care to validate that the tool does what you expect on known examples.
-
Embed this notice
Paul Cantrell (inthehands@hachyderm.io)'s status on Saturday, 21-Jun-2025 01:58:52 JST
Paul Cantrell
@trochee @grimalkina @glyph @dave @3psboyd
> automation has the potential to take us to "fully-automated luxury communism"I disagree, at least in the foreseeable future. I don’t see the elimination of human work as a plausible thing within the current technology horizon; this whole thing gives me the same vibes as the 1950s belief that the Nuclear Age means soon we will have infinite free energy.
-
Embed this notice
Jeremy Kahn (trochee@dair-community.social)'s status on Saturday, 21-Jun-2025 01:58:53 JST
Jeremy Kahn
yes, hard agree
automation has the potential to take us to "fully-automated luxury communism" but it also has the possibility of accelerating the machine that is already chewing some of us up
-
Embed this notice
Dr. Cat Hicks (grimalkina@mastodon.social)'s status on Saturday, 21-Jun-2025 01:58:54 JST
Dr. Cat Hicks
@trochee @inthehands @glyph @dave @3psboyd likewise I approach my interactions with people who have power over me -- wealthy men in tech, angry doctors -- with caution because they do not understand the bias they have internalized from their cultural and societal context, and it emerges in statistically detectable, but ultimate quite difficult to predict ways. This comes out in the words they use when writing about my body, my performance, my potential, my mortgage letters, my police reports
-
Embed this notice
Dr. Cat Hicks (grimalkina@mastodon.social)'s status on Saturday, 21-Jun-2025 01:58:54 JST
Dr. Cat Hicks
@trochee @inthehands @glyph @dave @3psboyd I don't think LLMs have anything close to reasoning but I'm just saying we're all dealing with cultural technologies already
-
Embed this notice
Jeremy Kahn (trochee@dair-community.social)'s status on Saturday, 21-Jun-2025 02:00:12 JST
Jeremy Kahn
oof too real, honestly.
-
Embed this notice
Glyph (glyph@mastodon.social)'s status on Saturday, 21-Jun-2025 02:00:42 JST
Glyph
@trochee @grimalkina @inthehands @dave @3psboyd a lot of my negative feelings about LLMs come from this place. there are so many dimensions of frustration: of course the existence of hegemonic oppression sucks in the first place, so that doesn't help, but also, some of my "allies" in this conversation think that LLMs are *creating* this type of inescapable bias prison, and explaining "no, it was always here", can sound like saying ML-based redlining is OK because we already had redlining before
Paul Cantrell repeated this. -
Embed this notice
Paul Cantrell (inthehands@hachyderm.io)'s status on Saturday, 21-Jun-2025 02:00:42 JST
Paul Cantrell
@glyph @trochee @grimalkina @dave @3psboyd
Manual redlining is not the same thing as automated redlining at speed with zero human oversight unless the humans actively slow down the process and interject themselves! -
Embed this notice
Glyph (glyph@mastodon.social)'s status on Saturday, 21-Jun-2025 02:00:49 JST
Glyph
@trochee @grimalkina @inthehands @dave @3psboyd in the world of coding and in the world of social justice my anxieties are similar: we are taking a chaotic system with a lot of problems that are very hard to solve for, and potentially inserting a chaff generator that will obliterate even the small efforts that a few people have been making to be thoughtful and address these problems. not the only area of social or economic life where this pattern is occurring
-
Embed this notice
All GNU social JP content and data are available under the