The Sane AI Slow Proposal

The slowdown cry is about control

All those fuckers from those companies are crying that they need to slow the AI, it is too scary and blah when they first hit any roadblock and all that. One for ideological reasons, another one for monetary reasons.
But in the end, all of the vibe kinda stays the same. All these attempts to manipulate with races and slowdowns, invented or not, are just an attempt to entrench themselves and their rules against the competition and impose their will upon anyone else. It is the current couple of monopolies who are already trying to do that and convince everyone there is a race to be won or a slowdown to be achieved, whatever their vibe currently is, so they can both achieve regulatory capture and avoid the competition.

The real danger is concentration, not scary models

Ok, but whatever you think it is, the danger is not the scary thing they create, but that one or two companies achieve so much that it leaves everyone else behind. And the concentration of power on those companies would be so immense, we have no way to oppose them anyhow.
Let’s call the real danger what it is. One company dominates, which will then coerce the government to mandate them as the sole winner and sucking all the rest of the resources along the way.
Or if you truly want to go doomer path and believe that they are gonna create scary things at all, consider that if one company has monopoly on resources, they are much more likely to succeed in creating that horrible thing they are so afraid of.

What forecasters conveniently leave behind

Ok, problem. Here, what do they conveniently leave behind?
But they completely miss the point, all these forecasters, that you would have something big, and that definitely comes from a single company. And nobody addresses that issue. They all cry that they are gonna reach AGI, ACI and all that. But nobody specified who exactly is gonna reach it. We all talk about it in some simpler way and completely ignore the point that such a thing in the hands of any company is a disaster when others do not have it. Or when this company unilaterally decides what their creation can do, what cannot do, and who can access it and all that.
Somehow they conveniently live with concentration of power problem entirely.
Any single entity possessing AGI is scary – let’s achieve it together.

Premises – two rules that must stand

Okay, premises.
The rules that must stand.

1. No government can dictate any company how they make or create software. Software is their intellectual property and protected under freedom of speech. So no interference of how they make it and what their AI can and cannot do, or what values it has or does not have, or its architecture is or whatever.
2. The anti-monopoly works by not deciding which company grew too much, but by limiting the resources one company can have from the total pool.

The proposal – 10 percent cap on server-grade GPUs

The solution is quite simple as an idea and potentially complex to implement.
No company of a nation can at any point operate, possess or in any other way run more than 10% of total internal server-grade GPU supplies. Potentially counted in A100 equivalents.
No renting overseas either. Nor leaving your proprietary crap to others to try to make it rentable and yet get it off the books. If your model, if your proprietary crap runs on some of the same GPUs, it is as if you already possess them. If you want to make such a deal, you keep everything open.
Every GPU imported into the country gets counted, so you always roughly know how much was imported in total as a way to know how much exists entirely. Tax audits are responsible at first, and whoever does accounting in a company has to report the amount of assets they possess at any time in the period. If it tries to sneak in, if it tries to rent too much over the course of a month or tries to rent overseas, then they get taxed immediately.

What about incumbents who already hoarded – progressive tax

Okay, problem. What will you do with existing companies already? Who have already accumulated quite a lot of hardware.
First, give them incentive to force the market diversity by progressive taxation. If you possess 10%, you pay almost zero taxes.
For every 1% you go over the 10%, you pay 5% tax on your total income. You got 15% of total supply, that is 5% over, you immediately pay 25% of your total income. If you grow to 20% of total supply, that is double the allowed 10%, you immediately pay 50% of your total income. That gets redistributed as subsidy for smaller players to buy new GPUs.
But naturally, if one company does too much, then they would have more customers, therefore they produce more of the economic effect of automation which creates job displacement. Because they raised too much and got ahead of everyone, this company is responsible for producing job displacement at a greater rate, and because they have the GPUs they pay taxes. If they do not have the GPUs, they do not pay anything because they still play fair. And the extra tax on their excessive infrastructure serves as the economic offset for their concentration, to move away from concentration and redistribute.

What this achieves

1. No company can create a race and create advancements vastly outperforming any competition and rapidly advance from other players entirely because they are constrained by compute and we cannot just brute force it as if they had possessed massive supplies that would actually allow them to advance too much
2. You have at minimum diversity of 10 different companies who would possess those GPUs and compete with each other, each with their own ideas, each with their own vision, potentially even specialization, competing on the same or adjacent markets. Thriving market competition. Some specialize and able to focus without concerns of having to earn to the general market. Others specialize on the general market. And still you have plenty of options and everyone also racing around to win for customer. You have hosting companies who produce none of the models but just host for hours for inference then being open of course.
3. Any advancements they make will be forced to be algorithmic. They cannot just brute force their way out of RSI or similar by just scaling infinitely. They have to stop and think how we can improve by the algorithms and dataset improvements alone. That already forces them to consider, to think a lot more about what makes AI effective and actually smart, which steps away from just scaling laws and to the welcome of the doomers, they might actually make it safer.
4. If anyone makes progress, everyone else can catch up in a potentially worse way but still knowing that they have the same or less resources and that they partially go on the already stepped path, somebody stepped in to adjust and make it better or no different. Whatever. The goal is to have the culture where massive diversity of options and techniques exist.
5. Anyone who reaches way too high by trying to grab too much compute is immediately progressively punished by the taxes making them unable to operate. And even if they actually can operate by having too much huge margins, that is a mechanism that makes others catch up much faster. At the same time reducing the power of those who leapt too much because they know that the market knows that competition is gonna have that in a couple of months or half a year already.

FAQ

China will not do this, we cripple ourselves

You stunt progress! China would not do this, we would just cripple ourselves!
First, forget that shitty claim about a race. Your concern should be that you have massive concentration of power into a couple of companies already, who already invented that race to convince you that it exists. And, actually, if you look at China, they already follow the same path. Instead of 2-3 companies competing, they have 6 to 8 different companies who already compete on the same pool of resources. I do not see anyone in there grabbing massive compute scale. The CCP would not let them grab too much. They know the value of diversity and the value of industry developing nice and evenly.
And if you start crying all that way about someone not playing fair and breaking the rule and allocating all the hardware to a single company to make a crown jewel to win the race, that just shows hypocrisy – you have very little trust in your own companies being able to achieve anything on constrained hardware.
Chinese companies have already been doing that for more than half a year, surviving on little scraps, shitty chips, much worse than you have, and still catching up rapidly, breathing down your neck. Which just shows how funny it is when anyone demands huge compute to make progress. And if you say you cannot do this, then what is the worth of you then? What is the worth of those companies who can only operate if they have massive compute?

But this is communism

But this is just communism, huh?! What are you proposing? You propose some kind of lefty crap to take from everyone and make them share? Redistribute? Am I out of my mind?

Well, at first I would propose typical antitrust rules to just judge by the percentage of the market they take in a field and the range of different services they offer.

But this is a very critical field to keep diverse and competitive, where technological advancement is most important and is the driving force behind it all.
I started with just proposing clamping down anyone who reaches more than a certain percentage of a market. But then I found that counting GPUs is actually a better idea here.

Not to say that imposing a percent limit would incentivize other fields to start thinking they want the same, which could be a toll on your economy and a source of strife which your economy might not be ready for yet.

And I ask you, do you really want to live in a world where a couple of monopolies are allowed to stay, who already grabbed all the resources required, who answer to no one, and this time with AI they speedrun becoming megacorps much faster than in any other market, leaving and drying up the rest of the market, grabbing the rest of the resources and leaving everyone behind.
As if you are already very happy with just a couple of big tech companies dictating everyone? Now consider if AI is gonna get only twice as worse. Show me a solution that saves you from tyranny of 1, 2 or 3 singular companies or the governments they would have bought.

What if they sneak and lie globally

What if a company just sneaks and lies and skims to get more than they have and tries to get hardware in many places of the world at once?

First, they have incentive to keep their taxes low, and they stay in great danger until their auditors come and slap them so hard they can potentially go out of business entirely. After all, having a lot of GPUs is expensive. That is not a small startup that can go unnoticed. And they still pay a little in taxes, so keeping them in check is already something to watch.

And in the worst case when they do not get watched, that means they still produce value internationally. They rent here and there, potentially pay taxes here and there for their rents, so they still produce value in a way that benefits more humans overall.

What about universities and open research

Additionally, if you somehow manage to have a good public goods fund, you could allocate 10% of your total supply to be reserved for universities for research and all, which is subsidized by the taxes paid by proprietary companies or from our taxes as well.

So you will always have a lot of groups who can do research and create open stuff without needing to care about financial incentives at all.
Here you would have a nice home for all your research on control, interoperability, tinkering with other architectures and ways to make it nicer, and in general a lot of adjacent sciences that study the effects of AI.