Post · bonfire.cafe

You can hold both of these things in your head at the same time:
1. Current frontier models are unethical and environmentally unsustainable.
2. The tech is actually capable of solving *some* real problems.

How? Because there is tremendous improvements happening in local models that are ethically trained and don't require massive data centers.

This doesn't mean you have to like them. You certainly should be skeptical. But there is a middle ground between "I want to marry one" and "Burn it with fire"

If this middle ground approach really bothers you, that is entirely your choice. But I politely ask you then to just unfollow me. We'll both be better off.

@mu@mastodon.nz · 1 hour ago

@scottjenson honest question: where do you think it has potential?

Scott Jenson

@scottjenson@social.coop · 1 hour ago

@mu At very low level analysis tasks, like collating data logs to find patterns. I've written about this at https://jenson.org/boring where there are boring things that this tech can do that is predictable and avoids hallucinations. It's when we assume that LLMs need to mimic high level human creative tasks where it completely falls apart. There is power in tedious low level stuff.

https://jenson.org

Boring is good

@mu@mastodon.nz · 56 minutes ago

@scottjenson I can see some of that as possible improvements, but I have to admit your SLM example of "this company was able to suck more money out of its customers" is pretty off-putting in much the same way that I find a lot of current LLM framing to be off-putting.

At the moment I think the best case that I can see is "better coding autocomplete", but even that has issues in terms of how people learn to code and other wider societal implications.

I love other machine learning models, I think there are possibilities in image processing or vehicle control, but I haven't yet seen a really clear benefit for LLMs. It might be that I'm being dense, or not understanding what people are saying, so I'll keep watching.

Scott Jenson

@scottjenson@social.coop · 44 minutes ago

@mu You don't seem to be discussing in good faith as your quote: "this company was able to suck more money out of its customers" is like NOTHING I said. I would never endorse or even recommend that! I'm reluctant to continue this conversation as I fear you'll just twist my words again.

But in the spirit of proper discussion, what I'm working on is a UX concept where the system logs user events to the disk and uses a SLM to infer user intent. It is 100% private, goes to no company, and is meant entirely for the users benefit. It would need to encrypted and safeguarded but it is 100% under the users control and never leaves their machine. Of course, this concept may not set the world on fire, but I'd argue is an example of a safe, non exploitive use of SLMs.

1+ more replies

Jak2k 🇪🇺

@jak2k@mastodontech.de · 3 hours ago

@scottjenson Could you point me to such an ethically trained local model?

Scott Jenson

@scottjenson@social.coop · 1 hour ago

@jak2k
https://apertvs.ai/
But let's be clear, there are reasons to quibble over this one as well. I'm not saying there are PERFECT local ethical models, i'm saying people are trying. I'm trying to skate to where the puck will be.

APERTVS.ai

Fully Open Foundation Model for Sovereign AI

Jak2k 🇪🇺

@jak2k@mastodontech.de · 25 minutes ago

@scottjenson They are training on the Euroblocks dataset which is synthesized from unethical models.

Scott Jenson

@scottjenson@social.coop · 8 minutes ago

@jak2k that seems a bit unfair. The euroblocks dataset is miles ahead of the foundational models. The project was required to undergo and publish formal ethics reviews (such as the UTTER Final Ethics Review) to ensure compliance with EU standards on data privacy and AI safety.

Is that enough? That can be discussed. As I said before, we can have disagreements over these models, they are far from perfect.

My original point was simply that there is real positive movement in this direction.

eclexic

@theeclecticdyslexic@mstdn.social · 3 hours ago

@scottjenson I mean, even with ethical training I'm not convinced the way LLMs impacts the priors of culture is net positive.

It seems to me it is currently most effectively used to DoS attack, or otherwise mislead, human brains. I think it has accelerated nearly every problem we currently face much more than if has accelerated the solutions.

Scott Jenson

@scottjenson@social.coop · 3 hours ago

@theeclecticdyslexic I'm currently reading "More than words" by John Warner, a thoughtful critique of the techbro enthusiasm with LLMs and how writing is a type of thinking, something LLMs just can't do. He makes a compelling point that many of the cultural impacts (e.g. Students using them to write essays) actually points out long standing flaws in our education system. His argument is that it is a wake up call to how we teach writing. LLMs are showing weaknesses in our assumptions. It's still painful, I don't wish to be simplistic, but it's an interesting perspective.

Shankar

@shankarux@mastodon.social · 3 hours ago

@scottjenson I got burned out because of my hard leaning towards (1).

I am trying to find that balance now and I am so glad I found your post. You are SO right in emphasizing *some*.

Andrew Hinton

@andrewhinton@jawns.club · 3 hours ago

@scottjenson It’s hard to separate the signifiers when our overlords are forcing it down everyone’s gullets, even though it’s clearly not the panacea miracle tech many of those same overlords seem to (or need to) believe it is.
I agree it has the same utility potential you mention!
It just feels like the bigger cultural weaponization of it (against workers, collective political power, minorities, and the entire economy) overwhelms the supposed benefits. (1/2)

Andrew Hinton

@andrewhinton@jawns.club · 3 hours ago

It makes it feel a bit like defending its utility is ignoring the bigger picture.
But I totally get where you’re coming from! Just commenting on why I’m not sure it’s an argument many people care to hear even if they technically agree. (2/2)
@scottjenson

Scott Jenson

@scottjenson@social.coop · 3 hours ago

@andrewhinton thank you for that thoughtful reply. I completely agree with you. The way it is being forced upon us and the cultural impact it's having is really quite frustrating and should not be condoned. It makes having any conversation about the tech quite problematic.

Fabrice Desré

@fabrice@fosstodon.org · 4 hours ago

@scottjenson About "tremendous improvements happening in local models that are ethically trained and don't require massive data centers. " I hear that a lot, but that never comes with "here's what you need to reproduce our results". Until this is the norm, it's hard to believe.

Scott Jenson

@scottjenson@social.coop · 4 hours ago

@fabrice
"You certainly should be skeptical"

the harbinger of eternal sept

@tychi@merveilles.town · 4 hours ago

@scottjenson it is both, but if we’re truly advocating for personal compute, we should be reframing the argument to be “can I hold it” or “does it hold me”

I see where you’re coming from but most people on this planet are upset with the surveillance state and data centers and taking it out on AI is the only lever they feel they have

Scott Jenson

@scottjenson@social.coop · 4 hours ago

@tychi totally reasonable take. There is a lot of backlash against all big tech and AI is just too easy of a target

the harbinger of eternal sept

@tychi@merveilles.town · 3 hours ago

@scottjenson I’ve been in a lot of rooms where people are saying “we need to disentangle what ai means” because it’s become an overly conflated term and I also leaned into describing every technology that’s not open source in the classical sense as AI

I know that’s problematic for people trying to sell being both open and ai, but it’s a distinction that most non tech people actually appreciate.

If it is not AI, I’m able to break it into the component parts and we can all look at them and talk about them objectively

If it is AI, they’re starting to say it’s “too dangerous for non technical people” and I’m applauding everyone seeing those red flags as red flags.

Scott Jenson

@scottjenson@social.coop · 3 hours ago

@tychi “too dangerous for non technical people” is a terrible framing! Lots of red flags there

Olivier Forget

@teleclimber@social.tchncs.de · 4 hours ago

@scottjenson I wish you the best of luck with this nuanced take on a fiery topic on social media! 🫡

Scott Jenson

@scottjenson@social.coop · 4 hours ago

@teleclimber yeah it's tricky. It's certainly a decisive topic!