Email or username:

Password:

Forgot your password?
jonny (good kind)'s posts Post Back to profile
Top-level
takin' a break

@jonny I had the great joy of seeing some folks at MIT present on using LLMs as they're meant to be used, which is to say as interfaces to more complex computational processes that just, convert otherwise difficult-to-parse data into more human-friendly language

god i hope the popular consciousness realizes that's what they're supposed to be sooner rather than later

31 Oct 2023 at 1:24 | Wall-to-wall | Open on solarpunk.moe
12 comments
Bee O'Problem

@juliana @jonny in theory the OP's scenario is just that. Taking data (code) and expressing it in something more understandable for a person.

And it apparently failed quite badly.

31 Oct 2023 at 1:58 | Open on mastodon.gamedev.place
takin' a break

@beeoproblem @jonny sorry, you don't seem to have understood what I'm saying. this isn't a situation where an LLM is acting on language alone and manipulating it to sound like it's producing human utterances; it's a situation where an LLM is being handed very carefully-controlled data that has been processed elsewhere and is simply being asked to express that in a more human-friendly way. so in this case rather than trying to convolute code itself, the LLM would instead pass it off to some sort of specially-trained neural network designed to, say, perform and analyze bitwise operations, then the LLM would present the results and explain what that network did

difficult to understand without the diagrams, i guess

also advice for the future: don't come into people's replies and condescend to them. that's rude.

@beeoproblem @jonny sorry, you don't seem to have understood what I'm saying. this isn't a situation where an LLM is acting on language alone and manipulating it to sound like it's producing human utterances; it's a situation where an LLM is being handed very carefully-controlled data that has been processed elsewhere and is simply being asked to express that in a more human-friendly way. so in this case rather than trying to convolute code itself, the LLM would instead pass it off to some sort of...

31 Oct 2023 at 2:12 | Open on solarpunk.moe
Elizabeth

@juliana @jonny out of interest, if use of LLMs were limited to this, would the energy demand per request be less, or is it just the number of requests would substantially decrease? (Does that question even make sense?)

31 Oct 2023 at 2:27 | Open on weirder.earth
takin' a break

@luminesce @jonny the energy consumption per demand would likely increase because there's more overall computation involved in the process. the LLM itself would be only one step in an arbitrarily complex chain of queries depending on the precise application

from what i understand, the primary energy consumption of LLMs comes from the training process, but i don't know the full details

the specific application i was seeing discussed was to create a way for doctors to use human language to query medical records and collate data to facilitate diagnosis. so while it is important to keep in mind the environmental impact of using LLMs, it's also important to weigh that against potential benefits as well.

frankly, i'd rather we weren't using LLMs at all, but i figure this is one of the less bad ways they can be used

@luminesce @jonny the energy consumption per demand would likely increase because there's more overall computation involved in the process. the LLM itself would be only one step in an arbitrarily complex chain of queries depending on the precise application

from what i understand, the primary energy consumption of LLMs comes from the training process, but i don't know the full details

31 Oct 2023 at 2:38 | Open on solarpunk.moe
jonny (good kind)

@juliana
@luminesce
Yes totally^ it depends on the application re: energy use. Chaining a bunch of models together would probably use more energy, but google et al want to use smaller models hooked up to knowledge graphs to make runtime inference feasible as a consumer product too, so that kind of application would be designed to use less.

The medical case has its own set of fun complications 💓
jon-e.net/surveillance-graphs/

@juliana
@luminesce
Yes totally^ it depends on the application re: energy use. Chaining a bunch of models together would probably use more energy, but google et al want to use smaller models hooked up to knowledge graphs to make runtime inference feasible as a consumer product too, so that kind of application would be designed to use less.

31 Oct 2023 at 2:42 | Open on neuromatch.social
takin' a break

@jonny bestie how sad are you that xanadu never happened lol

31 Oct 2023 at 2:46 | Open on solarpunk.moe
jonny (good kind)

@juliana
Like on the scale of 0 to semantic web, about an 8 lol

31 Oct 2023 at 2:48 | Open on neuromatch.social
jonny (good kind)

@juliana
Anytime Xanadu comes up I also think of @beka_valentine 's threads where I first learned why it would have been so good.

31 Oct 2023 at 3:07 | Open on neuromatch.social
Elizabeth

@juliana @jonny yes, the key to limiting the environmental impacts is limiting use cases to those that are truly beneficial, rather than generating demand for trivial or mis-leading/chaos-inducing uses

31 Oct 2023 at 2:46 | Open on weirder.earth
jonny (good kind)

@luminesce
@juliana
Thats one of the reasons they're pursuing it, yes. Smaller models that can be conditioned on factual information/formal, domain specific models. Eg. See
arxiv.org/abs/2203.05115

Unfortunately this kind of application has its own kind of Really Bad Outcomes that the AI critical folks largely have not caught up to yet :(, see

jon-e.net/surveillance-graphs/

The tech could be really cool. The problem, as with everything, is capitalism.

@luminesce
@juliana
Thats one of the reasons they're pursuing it, yes. Smaller models that can be conditioned on factual information/formal, domain specific models. Eg. See
arxiv.org/abs/2203.05115

Unfortunately this kind of application has its own kind of Really Bad Outcomes that the AI critical folks largely have not caught up to yet :(, see

31 Oct 2023 at 2:38 | Open on neuromatch.social
Androcat

@juliana @jonny That sounds like Intelligence Augmentation (where the intelligence is human, and the augmentation is interfacing to expert systems).

31 Oct 2023 at 11:02 | Open on toot.cat
Go Up