I think the decision to use LLMs is a decision to build a worse world, where people seem less valuable, and where the valuing of products over processes is exacerbated. Kevin is a good dude; he‘s not wrong about these uses. But I wish we were trying to build more caring even if less efficient worlds - ThreadSky

neuroyogacara.bsky.social • 13 days ago

I think the decision to use LLMs is a decision to build a worse world, where people seem less valuable, and where the valuing of products over processes is exacerbated. Kevin is a good dude; he‘s not wrong about these uses. But I wish we were trying to build more caring even if less efficient worlds

Reposted from Kevin Zollman

There's so much polarization around LLMs. They are way overhyped, I agree. But I also use them semi-regularly now.

Here's a thread of genuine use cases where I find them helpful. Please add your own!

Comments

devezer.bsky.social•13 days ago

That's exactly what I wish we actively focused on.

asitisnow.bsky.social•13 days ago

If people using LLMs directly experienced the rise in carbon dioxide and heat in their immediate living space each time they query the system, they would quickly abandon it (or stop b/c of physical exhaustion). Absent such feedback, it seems many people prefer to plead ignorance.

danhicks.bsky.social•13 days ago

The highest estimates are 3 Wh of energy use per query. If you directly converted that into heating the air in a 10' cube, it would be about 0.18ºF.

asitisnow.bsky.social•13 days ago

Does this take into account the energy used to extract the minerals and metals used to build and maintain the tech infrastructure? With Google abandoning its clean energy pledges, and Three Mile Island being slated to reopen to power Microsoft's data centres, these "highest" estimates seem off.

danhicks.bsky.social•13 days ago

Back of the envelope, training adds 1 Wh. Costs of mineral extraction and chip production would amortize the same way: https://andymasley.substack.com/p/a-cheat-sheet-for-conversations-about?open=false#%C2%A7training-an-ai-model-uses-too-much-energy

asitisnow.bsky.social•12 days ago

I don't think a Substack entry provides a reliable source. I'm sure that there are many back-of-the-envelope calculations that one can use to paint an "everything is just fine" picture. I'm old enough to remember elegant calculations that predicted a nuclear disaster as a 1-in-a-1000 years event....

asitisnow.bsky.social•12 days ago

The bottom line is that Big Tech is slashing its net-zero targets b/c of the expansion of LLMs and similar AI technologies, which is extremely concerning, not to mention AI's disruption of labour markets and its impact on sense-making and cognition.
https://www.theguardian.com/commentisfree/article/2024/may/30/ugly-truth-ai-chatgpt-guzzling-resources-environment

danhicks.bsky.social•12 days ago

There are links all over that post to actual sources.

And yes, it's all based on a lot of guesswork because LLM companies aren't disclosing details. But I haven't seen any estimates that are orders of magnitude higher. So at best the climate argument is inconclusive.

neuroyogacara.bsky.social•13 days ago

So long as there aren’t huge numbers of people using these things all the time, we’ll be totally fine!

kevinzollman.com•13 days ago

It's also important to ask what they displace. An LLM is useful for coding because one query will replace easily 20 or 30 page loads looking at different stack overflow pages and the like. I don't know what the tradeoff is, but we shouldn't treat the LLM as entirely new costs

neuroyogacara.bsky.social•13 days ago

The messy thing is that streaming video has those effects as well. We need some hard shifts in how we move about the world. But pulling them off (as uncle Sid noted long ago) requires letting go of a lot of things

asitisnow.bsky.social•13 days ago

Yes, technology has set us on a slippery slope not only where everything we have to do (for work or in daily life) incurs incalculable downstream costs but where the formerly shared epistemic ground constantly shifts b/c technology also generates mountains of data that only it itself "understands."

kevinzollman.com•13 days ago

I think the difficulty of our current situation is that we are each left to our own devices to figure out the costs and tradeoffs. It's basically impossible to know what is "worth it"

asitisnow.bsky.social•13 days ago

It's pretty obvious at this stage that LLMs are causing harm to our shared natural and cognitive environments and are tools for massive wealth redistribution (to the top). They're weapons of mass distraction and labour devaluation devices and not worth it if one cares about fundamental well-being.

kevinzollman.com•13 days ago

I guess neither of those seem obvious to me. Our informational environments have been well fucked for many years before LLM, and it's not obvious to me they make it appreciably worse. Same with wealth distribution. Both seem like processes that started long before LLM

workers.bsky.social•11 days ago

We can build a more efficient world, then measure and track the ways it is or isn’t.

A more caring world? That’s subjective and ends in semantic purgatory.
I want this more caring world too and Kindness.
My caring Kindness is hot tea but yours might be No Genocide. 🫱🏼‍🫲🏽
(mine is no genocide too)

kevinzollman.com•13 days ago

I find it very hard to judge if LLMs in totality make the world better or worse. I think the ways I use them make the world better, but of course I might be wrong.

Some ways of using them undoubtedly make the world worse, of course.

neillevy.bsky.social•13 days ago

You don’t think it’s likely that LLMs will very significantly raise unemployment? I’d be very surprised if we don’t see double digit UnE within 5 years.

kevinzollman.com•13 days ago

I have no idea. Creative destruction is hard to distinguish from destructive destruction in advance.

neillevy.bsky.social•13 days ago

We have a context - late capitalism- that stacks the odds very heavily on one side.

danhicks.bsky.social•13 days ago

https://bsky.app/profile/danhicks.bsky.social/post/3licpdo76sc2e

neuroyogacara.bsky.social•13 days ago

Also I was just about to add, the effects of increasing automation have been clear for at least the past 40 years—and probably longer. That’s just how long they’ve been clear to me

neillevy.bsky.social•13 days ago

The usual pushback (I’m sure you know) is that every new technology creates more jobs than it destroys. I’m betting this time is different, because this is the first domain general technology.

Comments

Posting Rules

Reply