The boom in LLMs is going to hollow out a number of knowledge-worker industries — for example, writing boilerplate code or technical documentation
Not because it does it well but because the flacks can sell upper management on the idea that it can do it at all, as @pluralistic recently pointed out
This sale is a pig-in-a-poke, and the winning move is to not be holding the bag when the actual code or documentation is found to be terrible
@trochee it's just bog-standard capitalism applied to fields it was previously thought to not be able to apply to. Quality always goes down to lower costs and raise profits. Providing a good product is expensive; providing one that looks good but is shit isn't.
There have been automatic documentation generators before and they've always sucked. /** Pizzas the eating contest. @param i The i. */ void pizzaEatingContest(int i) - LLMs are really only dressing up the underlying suck to make it not so obvious.
@trochee The best thing we can do to build things without VC money is to build things.
Me, personally, I haven't done nearly enough free software building, so I have no right to complain about a lack of free software.
But that's software. The economy is bigger than software (even though it fully relies on it). The problem isn't "how do I build something without VC money?", it's "how do I not starve to death without VC money?"
@trochee@doctorow I get frustrated sometimes that even in the fediverse (which sometimes feels like the last remaining hidden sanctuary for the possibility of a nuanced discussion on the internet), discussion of LLMs seems polarised into "If you've ever used the output of an LLM for anything, you're an irredeemable villain" and "What kind of moron writes/codes anything for themselves? Where my AGI at?" camps.
@trochee@doctorow I'm a senior dev, and I find them helpful for work stuff quite regularly, but I'd also rather chew off my own fingers than outsource the better part of my coding or any of my writing to them.
I can't fathom how that seems like a good idea to anyone, and I can only conclude that a lot of people must really, really hate writing.
@trochee I've been trying to think of a pithy way to frame "As the LLM output approaches quality, your effort spent prompting it will approach or exceed that necessary to just do the fucking work" for about a year now.
Where I use them for code, it's most always boilerplate;
eg. sketching out a starting point for some unit tests.
Mostly, they're a supporting tool. Handy for converting text from one format to another.
eg. "Turn this Typescript definition into a JSON schema" (or vice versa)
@trochee@doctorow I don't disagree in the least about the coming "hollowing out", and it's depressing as hell. Just wish we could get to "yeah, they're useful for some things" already.
@tom_armstrong@trochee@doctorow management don't like paying for expensive knowledge workers or waiting for a good product? People who have seen "wizards" but don't understand the "wizardry" assume all "wizardry" is the same. That I think is who thinks it's a good idea and you can make yourself very unpopular by suggesting that maybe, just maybe, LLMs are not the panacea to an ailing bottom line.
So I see two poles for how to respond to "Peak Gasoline" in SFF
[bear with me while I digress; coming back to LLMs in a sec]
and they're basically
doomer hoarding and xenophobia (everyone for himself, and f your feelings) — think MAD MAX
vs
Solarpunk invention of new ways of living that... just don't use gasoline (everyone's in this together, so we'd better have a community garden, build transit, and learn first aid) — think THE TERRAFORMERS (thanks @annaleen !)
So, back to LLMs, or rather, back to prepping for "Peak AI"
(It's coming, let me tell you.)
Those of us currently under attack/usurpation by the push towards AI, we can respond to this by prepping toward either pole
The doom-prepper approach is to, I dunno, start planning a consultancy on "unbefunging your company's poorly-thought-out dependency on chatgpt", or building tools to poison (or detect poisoning) for the LLMs
Solarpunk offers a vision, in the face of gasoline, for planning community gardens, mutual aid societies, bike repair workshops, etc, even as the gasoline culture around us creaks and groans its way bloodily towards a reckoning
What is our (data science, machine learning, UX, design) equivalent of solarpunk for the concepts of automation?
In a solarpunk future, I want automation to do mindless drudge labor
… free up the people to care for people and make art
But our coming doomer future has the automation performing a simulacrum of caring for people
…and making art-shaped artifacts
…and humans are doing the mindless drudge labor
@Virginicus@trochee What if a machine can "do it" but the quality of the result is so bad that we suffer extreme increases in overall system inefficiency or decline in quality of life?
What is our (data science, machine learning, UX, design) equivalent of solarpunk?
Can we even imagine a world of automation without the exploitation, fantasies of infinite growth, and the increasingly unsubtle justifications for "laundering toil away"?
What are compelling visions of the future of automation that aren't burying the exploitation behind an app or within a billion-parameter model?
@trochee Perhaps you mean statistics, statistical learning, rule engines, and human interface design? I feel so much of this latest crop of disruptive hucksterism is obfuscatory wrapping of much older math and computer science in techno teleo-mythology. It's like how trademark, copyright, and patent got wrapped up into "Intellectual Property" and turned into a commodity, making something, "property" out of piecemeal, specific, and limited inventives and legal rights.
It is possible, yes, just as the bicycle pre-dated the car
But now the bicycles are still here and have to find ways to live with cars (even though the favor is not returned)
So the same thing can be said about all the "pre-LLM" tech tools — we have to figure out how to protect them and their ethical use from the coming destructive collapse
@trochee amen. Just yesterday I was talking with a former co-worker about how there is an opportunity for those who know how to build human focused, simple and maintainable systems that don't rely upon the intensified economic infrastructure made of sand. The sand being proprietary hardware, operating systems, software, and services designed to serve the rent extracting capitalists via forced upgrades and constant churn, and layers and layers of tools to compensate for managing them.
@trochee Nice question - one might also ask it about app development and smartphones (extremely exploitative), some sectors of research, and also many industry products.
Where do we draw an ethical line? What's the human element of it?
... That sounds like a MAD MAX doomer/hoarder approach
The LLaMa models are also baking in a lot of the problems (unaccountable and probably pillaged training data, overfitting, inability to learn or work with structured models of the world, etc)
What are you doing to build for a future that walks away from that particular "foundation models" Omelas?
@trochee just had a AI strategy meeting with my CTO, revealed that I was called a hoarder/doomer on the internet. He (I value his opinion very much) revealed he is an armageddonist, believing the whole organization will be out of work in 24 months. Blew my mind.
@trochee he believes the lay-offs will traverse the tech organization bottom to top him being the band on the Titanic, playing alone until the bitter end.
1/2 - How about making the various Machine Learning model architectures available and using your own data to train it for your personal use so it becomes a spoken or written conversational interface to “your digital stuff”. Alternative to sitting in front of a keyboard and mouse.
So it is “democratized” and available to all, not just the “overlords”.
@trochee just like an NGO working against climate change probably has a credibility and ethics problem if they're taking donations from Shell/BP, the participation of software giants in FOSS projects is an easy way to subvert them - this has to be harder to do or inacceptible within software-dev culture
@trochee It would have to be automation that wasn't just all about using fossil fuels to replace all other kinds of labour and energy. So it would be all about seeing how much automation you can get from one waterwheel, for example? And how that could all be channeled for meeting human needs rather than building bloat to exploit people.
@trochee I think key parts of the aesthetics of solarpunk include things like: finit growth, not being alienated from your work, an ecology of technology, and of course sustainability.
It’s hard to imagine tech meant for finite growth since all we have talked about is scalability. A community garden can’t scale so how would that look like in technology?
@trochee I think having pupose and not being alienated is one of the reasons people work in startups or open source, so there we have some kind of model for how it might look…
With ecology I’m thinking protocols, maybe something about mutual benefits… but also that we need to build solutions that are resilient and can handle variations. Less control mindset
Interesting to think about buts it is hard since the tech culture is everything solarpunk is not
I mean, if we're after "what do actually helpful AIs do" we should be looking at Iain M Banks' Culture books (or the Act III train in TERRAFORMERS)
but the whole "conscious and autonomous" angle is a smokescreen; I want to dream about how to use and improve the automation we have to ease toil rather than to extract rent
@trochee oh aye. As a USian, I don't even want smart tax filing tools... I just want automatic filing. And to end the whole concept of "the 99%" as we currently understand it.
@trochee@luke The UK and I think many other countries already do that if you are not self-employed. One of many ridiculous things the US doesn't have that other comparable countries do.
@alastair@trochee@luke Yeah I'm still mad at finding out how many European (and Asian, and African, and South American, and Oceanic...) countries just send you a bill rather than make you "do" your taxes.
Like, it makes sense, the IRS has to recrunch my numbers anyway. They can't just take it on my good faith...
@trochee The technology is so simple that it’s possible to have open source models trained by public institutions based off say all the content you can find in any library and make that available to everybody.
The math is simple. But the compute and data volumes are prohibitively expensive to anyone not an oligarch
Perhaps a new public institution could exist at the scale we would need — but that pushes the governance and accountability problem there (probably better than the completely unaccountable oligarchs)
American public institutions, at least, have been so gutted by the early days of the Jackpot that they cannot participate here (alas)
@trochee Honestly, it's a lot of the same answers as solarpunk, and it generally boils down to "communism, but we don't SAY communism, because it gives the Americans the skeevies."
The vast majority of the problems with AI hollowing out certain careers is the fact that we have to do these careers in the first place in order to survive. If you could practice art for art's sake alone and not have to try to make a career out of it, who would care if a computer also made art?
Add comment