by Gustavo Rehermann (wallabra) - contact: gr.wallabra@proton.me
I'm happy that we are starting to see AI safety initiatives. I understand their problem statement. But I think they are very myopic when it comes to the extent of what it means to have AI widespread in society, not just misaligned AI, but any AI.
How about we start making AI do the parts of life we actually don't want to deal with, while we get to keep the enjoyable parts like art, writing, and trust in a meaningful and real reality, for ourselves? Or rather, how about we use it once we know it can do these things, and do them sustainably?
And, ultimately, what if the problems rightfully attributed to AI actually go much, much deeper than just the AI itself? Let's dive in.
We have moved in the wrong direction when it comes to scaling machine learning to tackle hard problems of computation, such as language processing. Dotting the world with enormous data centers that suck up electricity, fill their surroundings with a blinding amount of artificial light, and pose serious environmental threats, all the while billions of dollars are sucked up from the make-believe financial economy into the make-believe purchase orders for an enormous number of electronic hardware - orders that might never be integrally fulfilled - to install it in data centers that don't yet exist, all to maintain services that will absolutely not reap in the profit it promises to reap.
When the biggest ongoing existential threats to humanity are environmental collapse, geopolitical mutually assured destruction, and economic collapse due to the intractable nature of late-stage capitalist expansion and the bill coming due on the inevitable contradictions of life and existence in this economy (in particular when it comes to promises of prosperity and meritocracy unfulfilled), anything that adds more weight onto any of these respective dominoes feels most out of place.
We already use machine learning in the fields where it makes the most sense, such as protein folding, where it gives the most results for a much more standard and controllable level of computational expenditure. It's one thing to make better machine learning tools that apply to more fields where non-ML algorithms might fail, such as natural language processing and chatbots, NPC interaction, 'fuzzy computing', etc. But cost benefit analysis in those things, like in any other thing, absolutely should not be based purely on shareholder value, or - even more absurdly - unsustainable promises of great gain to shareholder value, or, in other words, a bubble.
Those who push AI today do not recognize the value of human life beyond the labor it provides to the capitalist machine, and could never fathom that life must be about more than just 'making shareholder value go up' for it to actually feel meaningful. I'm not even going to touch on the ethics of stealing copyrighted works to compose training data sets, because everyone's talked about that already.
Then, there are some things whose goals are inherently linked to those things being done by a human .Take learning, for example. Learning is a workout for your brain, it requires struggle, it requires doing things over and over, failing and being frustrated, until you get it right. Such is life and learning. When a student uses AI to go through an entire educational trajectory without writing a single essay themselves, have they truly learned anything?
Even if someday someone has the brilliant idea to hook an AI up to the big red button, the problem isn't the AI itself, it's the fact that someone hooked it up to that damn button.
AI in a vacuum is just a number cruncher. Yes, that is a gross oversimplification, but that's what every single computer program ever is by nature and necessity. The reason computers matter is because these numbers are actionable - that is to say, people, or other programs or hardware, interpret these numbers and derive action from them.
For example, the LLM puts out tokens, but the AI girlfriend application brings its own visuals, font, statement of purpose, and, of course, the, uh, marketing assets, that traps desperate people in. The LLM technology is just one of the main gears that enables that one specific example of a predatory business or bad thing that AI can be a part of, but the system as a whole was designed by people, and it is those people that we must hold accountable. If we must consider anything to be an existential threat to humanity in relation to AI, it is going to be a subset of those people.
Just like we shouldn't have the US Department of Defense wanting to automate 'critical defense systems' with AI. Some might even begin to deduce, by elimination of any other plausible explanations, that such people actually just love death! Which sounds absurd, but reality is becoming absurd as the aforementioned contradictions of this economy can no longer be wavered away and start to actually unravel this entire structure.
No, the worst thread that AI (or rather, the contexts in which AI is used) has to humanity is caused by the people who use those AIs. AIs are tools, but more precisely, there are many kinds of AIs which can do different jobs, and some of them can't really do beneficial jobs, only harmful jobs, such as "educational AI" that seeks to replace teachers but is a fundamental misunderstanding of what AI can do and how it works, and as it is put in motion is but putting into jeopardy the maintenance of society, hyper-complex as it is.
AI's worst threats to humanity aren't dramatic. They're not really akin to how sci-fi describes robots. And this isn't because sci-fi is bad, it did predict many things, but rather because in the real world, there's a ton of "boring" things that our lives depend on that are much more fragile than we'd like to think, like skilled labor being able to maintain an economy and keep up agricultural yields, a preservation of truth that allows people to govern themselves and doesn't enable reckless prejudices and bigotries that - in putting ends above means - destroy society, and just people being able to enjoy life by putting their own footprints in the world, rather than having them abstracted by a corporate label or, worse yet, replaced by generic omni-hybridizing generation.
I'm not "anti" or "pro" AI. I'm not a "luddite" or a "primitivist", and I'm not even opposed to the Pro-Human AI Declaration itself within the specific boundaries of its problem statement. I'm merely a pro-human socialist with a deep investment in both the intangible and Camusian things that make life worth living, and the tangible outcomes for people rooted in our deeply flawed economic and social structures, which either must be utterly uprooted by a people who knows how to lead a better future, or will violently uproot themselves as their contradictions become intractable and leave behind a wasteland that few will know how to pick back up. And I think AI is just yet another symptom of a system trying to make itself keep making sense in the face of said contradictions, not a major motor in said system.
We need to have a well defined set of priorities that we can call human priorities, before we can understand AI alignment within that context. Today, we do not define human priorities, only capitalist priorities. This is a volatile reaction we can replace, or let run its course violently.