Comment

Comment on CEO of Palantir Says AI Means You’ll Have to Work With Your Hands Like a Peasant

AI/LLM will only make automation more capable

LLM, unlikely. ML, probably, but not as rapidly as the hype would suggest.

And yeah, the disruption caused by the industrial revolution, telecommunications, automobiles, computers and the internet all are likely to exceed any impact caused by broader use of LLMs, which are too costly to train and run, inherently too unreliable for safety-critical or health-critical use, too flaky for any use requiring auditability, and generally of unproven utility so far, outside of a few niche applications.

And I say this as a leader of a technical team that has successfull adopted ML in several use cases, and has evaluated several opportunities to use LLMs. So far, with LLMs, the game ain’t worth the candle, even without considering the enormous environmental damage caused by their supporting infrastructure.

source

Sort:hotnew top

MangoCats@feddit.it ⁨1⁩ ⁨day⁩ ago

LLM, unlikely. ML, probably

ML already has demonstrated tremendous capability increases for automated machines, starting with postal letter sorters decades ago, proceeding through ever more advanced (and still limited, occasionally flawed - like people) image recognition.

LLM puts more of a “natural language interface” on things, making phone trees into something less infuriating to use and ultimately more helpful.

LLMs, which are too costly to train and run

That’s a matter of application

inherently too unreliable for safety-critical or health-critical use

Yeah, although I can see LLMs being helpful as a front end, in addition to the traditional checklist systems used for safety regulation, medical Dx and other guidance, an LLM can (and has, for me) provided (incomplete, sometimes flawed) targeted insights into material it reviews - improving the human review process as an adjunct tool, not as a replacement for the human reviewer.

too flaky for any use requiring auditability

Definitely. Mostly I have been using LLM generated code to create deterministic processes which can be verified as correct - it’s pretty good at that, I could write the same code myself but the AI agent/LLM can write that kind of (simple) program 5x-10x faster for 10% of the “brain fatigue” and I can focus on the real problems we’re trying to solve. Having those deterministic tools again makes review and evaluation of large spreadsheets a more thorough and less labor intense process. People make mistakes, too, and when you give them (for this morning’s example) a spreadsheet with 2000 rows and 30 columns to “evaluate” - beyond people’s “context window capacity” as well… we need tools that focus on the important 50 lines and 8 columns without missing the occasional rare important datapoints…

So far, with LLMs, the game ain’t worth the candle,

The better modern models, in roughly the past 10 months or so, have turned a corner for some computer programming tasks, and over those 10 months they have improved rather significantly. It’s not the panacea revolution that a lot of breathless journalists describe, but it’s a better tool assisting in the creation of simple programs (and simple components of larger programs) than anything I have used in the previous 45 years, and over the past 10 months the level of complexity / size of programs the LLMs can effectively handle has roughly tripled, in my estimation for my applications.

even without considering the enormous environmental damage caused by their supporting infrastructure.

When it’s used for worthless garbage (as most of it seems to be today), I agree with this evaluation. Focused on good use cases? In specifically good use cases, the power / environmental impacts range from trivial to positive - in those cases where the AI agents/LLMs are saving human labor - human labor and its infrastructure has enormous environmental impact too.

source