First: this post is completely farm-to-table, organic and AI had no involvement in crafting it whatsoever. Because of that, there will be plenty of inaccuracies and human-isms. I’m not wholesale against AI (hell, my job IS AI. My views are my own, though), so I will make a point later in this text that this discussion should be more nuanced.

This post is also aimed at a relatively technical audience, you have to be familiar with basic concepts of AI infrastructure, distributed computing, and basic computer hardware concepts (as in, what is a GPU, etc.)

Secondly, I’d like to get my biases out of the way:

  • I love computers, have always loved computers and will always love computers
  • I love OWNING compute (usually in the form of computers) and my major gripe with the current implementation of AI is that it is fundamentally against communities or persons owning computers
  • I love using computers for directionally good and true things
  • I also want to live in a globalized world that has kids playing outside, grass and trees and clean air

With that out of the way, first we need to fix the fact AI is much more dangerous than it should be, and much less dangerous than what they (the proverbial “they”) want you to believe it is. It’s a big problem that requires immense societal action and global cooperation, but the ideas are relatively simple, here’s how:

  1. There’s a major “wealth” disparity in compute: as it stands, a handful of companies own the vast majority of the infrastructure necessary to run AI - blah blah blah who controls the spice controls the universe. This compute wealth disparity translates in actual wealth disparity, that’s not the point I’m trying to make here, but the solutions are similar. Local communities and people need to have their fair shot at owning such infrastructure. Government needs to step in to regulate the market, and subsidize equipment going to the right hands (yes, I am advocating for GPU subsidies for individuals and communities).
  2. Research on frontier needs to take a (broad, global, regulation-imposed) step back, but this is not happening, is it? So, instead, we need more labs, focusing on smaller, local, open-weight, more specialized models.
  3. Societal impacts need to be well-research and prioritized. Companies that make false claims in favour of marketing should be penalized, not fined - penalized (Criminal liability for executives? License revocation? Compute confiscation?). If the punishment for a crime is a fine, then it’s only a crime if you’re poor.

There’s immense power in decentralizing computing infrastructure and incentivizing distributed computing (see Folding@home), this power is what we need to lower the risk profile of AI. In Machines of Loving Grace, Dario Amodei mentions the following:

“I am not as confident that AI can address inequality and economic growth as I am that it can invent fundamental technologies, because technology has such obvious high returns to intelligence (including the ability to route around complexities and lack of data) whereas the economy involves a lot of constraints from humans, as well as a large dose of intrinsic complexity. I am somewhat skeptical that an AI could solve the famous “socialist calculation problem” and I don’t think governments will (or should) turn over their economic policy to such an entity, even if it could do so. There are also problems like how to convince people to take treatments that are effective but that they may be suspicious of.”

I agree with his assessment, but what would change if, for example, people felt like they had (even a level of) control over their data, their equipment, and chose how to participate in the active development of this technology?

The second thing we need to fix, if we want to live in a world in which there’s still an outside to go out to, are impacts on climate. It’s important to understand here, that the impacts of AI on climate are vastly overstated, and easily addressable - and curiously the solution here intersects with the solution presented previously in this text: local, smaller, open-weights models and, when needed, larger models paired with distributed computing.

The same way we should strive as a society for every family to own a fridge, we should strive for every family to own a computer. Modern fridges generally consume more energy than computers (in fact, a modern, energy-efficient ARM-based computer already can and will consume less than half of the energy a fridge consumes). I might be slightly off on the exact power-consumption numbers, and I could easily check that by interacting with… AI - but as I mentioned, this is a completely organic, farm-to-table text. Access is everything, you give everyone access to this transformative technology, we as a society agree on the guardrails, and the wheels of progress continue to turn.

For deeper problems, research and other initiatives requiring more compute: every computer is a node and when idle, the compute goes “back into the grid” (I don’t worry even a bit about the logistics, I’m completely sure this is possible and someone way smarter than me will figure it out. In fact, I don’t even think this is an extremely complex problem). Every person could even choose what to support with their surplus compute. Say, a hospital area is doing carcinogen research. They could send a request for compute to their local network, and you could contribute when your computer is idle.

Finally, an easier one to tackle: people need to be able to choose when and how to engage with AI. We need content labels, similar to the ones we’re thankfully now surfacing at grocery stores. For every content we see, we need to know:

  • Was AI used?
  • How was AI used? (was it text/image/video generation, research, or something else?)
  • What model and type of inference was used? (I should be able to not support/engage with content made with models made by companies unaligned with my morals, or with inference that’s centralized in a data center, for example)

This is the nuance piece I mentioned earlier; every individual can and will have different lines when interacting with AI. I, for example, draw my line at art - I believe art is fundamentally human and should never be conceptualized by AI, and if it is, I need to be able to choose not to engage with it. There should never be misrepresentation, misdirection or disinformation when engaging with AI-created (or assisted) content.

As I said, big problems requiring immense societal action and global cooperation, but we need to start somewhere. My recommendation on where to start? Someone needs to tell Dario this is the only way his “Machines of Loving Grace” come to fruition.