Gen AI Isn't Happening Soon - Sun, Apr 27, 2025

At least, not until we solve some stuff

I posit there will be multiple walls mankind must overcome to get GenAI to be even off the ground.

At it’s core, let’s consider the possible forms GenAI can take first.

Monolith Silicone Superstructure (MSS)

An MSS GenAI exists as the natural conclusion to taking vertical scaling to its utmost limit. Adding more and more computational power, more and more memory storage, and more and more training data until you have a sufficiently large enough, and powerful enough, system to handle the task of GenAI. Storage might take multiple warehouses worth of server arrays all deeply connected, with network setups that wiring and routing alone take a team of professionals months to lay out.

Biological Replicated Technology (“Wetware”)

In this scenario, an entirely new format for simultaneously storing neural networks is discovered. Rather than needing to be synthetically produced as a meticulous set of silicone chips on boards (GPUs and whatnot), it instead is produced in a more biologically synthesized format. Whether it’s as fluid as a gel or more solid, the end result is a substantially denser material that stores the synapses of a neural network in a physical form, rather than a simulated form.

The appeal here is high, a single physical synapse would require at minimum dozens and dozens of semiconductors to simulate it in a MSS system. If carefully engineered such a system could improve the density of storage by orders of magnitude.

For all intents and purposes, this scenario means scientists have produced a synthetic brain that is far more optimized in shape, size, and material than a human’s.

Blocker 1: Time

To begin with, a sufficient Gen AI must be able to “pay off” its time investment faster than a normal living being to be “worth it”. One must be able to produce 1 single Gen AI that performs better than a human in less time than it would take to simply just train a human to do any given arbitrary task.

If we assume that a Gen AI is able to perform tasks no human can perform (perhaps calculating proofs or solving issues that not even a group of the most talented humans can solve), then the time limit simply becomes whatever arbitrary deadline must be met for that given task.

Consider the thought experiment of a disease that is about to wipe out all of mankind in seven years. The government must choose between either constructing a sufficiently powerful AI system that can tackle this problem, or, focusing its time and money on a team of human experts to try and achieve the same.

Assuming the AI can solve the problem exceptionally fast once operational, then the race against time is a comparison of:

1. How fast can the humans accomplish the task
2. How fast can the AI system become offline
3. How long until the deadline of the problem results in catastrophe

Assuming the answer to 1 is greater than 2, it then just becomes a comparison of 2 vs 3 alone, as we can rule out 1.

Let us now compare the two possible form factors GenAI against this blocker:

MSS vs Time

It’s incredibly unlikely that any MSS based system will ever be able to sufficiently win any critical deadlines, purely from a logistical standpoint.

The sheer effort to construct such a system is enormous, and such an undertaking will almost undoubtedly be fraught with complications. Very quickly the sheer signal delays, even at the fastest of data transfer speeds, from one warehouse to another will in of themselves introduce unexpected consequences. Furthermore the massive amount of data needing to be synched over such long distances will introduce enormous challenges with respect to data integrity. Data loss will be a constant concern, which means the solution will likely lie within data replication. And that in turn means the already existing enormous burden of networking will be magnified yet again to compensate for this.

Wetware vs Time

With compression and solid form factor comes the challenge any other physical neural network (like our own brains) encounter: mutability requires a physical change in some form. Whether its shifting balances of chemicals attached to any given neuron, or charge levels across some form of state holding material, the literal movement of physical materials has fundamental speed limits built in.

Exceeding these speed limits introduces friction and, by nature, heat. Naturally this will result in any such “synthetic brain” having a built in maximum learning speed.

While it’s extremely likely these optimized materials will still far outstrip a normal human’s brain, it’s still probable they won’t be fast enough.

The problem: Human’s really suck at planning ahead

Like, really suck. Assuming either of the above two are even reached, we still will have to overcome the fundamental nature of human’s to put off important undertakings until the very last possible minute.

As you are reading this, the entire world right now absolutely has the ability to switch off of its dependencies on fossil fuels and change over to renewable forms of energy. We have had more than enough capability to power the majority of the world’s grid with solar, wind, and tidal power for nearly a decade now, and yet we see time and time again active battling against this process. It’s very likely if every country, right now collectively agreed to switch over to renewable energy out of necessity, it would be accomplished within several years.

And yet, it still doesn’t happen.

The same would undoubtedly apply to AI. Even if right now scientists approached leaders of the world and informed them we had everything in place to start building an AI that could truly solve all the worlds problems… ten years from now, they would be shoo’d away. No leader of any world power at this time is going to approve something that benefits someone else ten years from now. Or twenty. Or etc.

And once we start encountering all the issues that an AI could solve for us much better than any group of humans could, it will already be too late to start and as a result, once again, the team of human specialists will be used instead (possibly to success).

And as a result, the AI can will get kicked down the road once again… Since the current pressing issue has, of course, been solved already.

Blocker 2: Power

Let us assume now that somehow, against all odds, Blocker 1 has been miraculously solved. Perhaps a particular world power has elected a leader that truly cares about the future. Maybe they come from a technological background and can personally see the true value in such technology and they somehow manage to maintain power for long enough to see the program to fruition.

It’s unlikely, but let’s assume this occurs.

The new issue that presents itself immediately is cooling. No matter how you cut it, the very process of solving such difficult problems inherently demands enormous amounts of power.

MSS vs Power

Let us assume the super structure is now built and operational. The sheer power demand to run such a system would be so immense, so incredible, that the mere resistance of the massive network of cables alone would demand amounts comparable to small towns.

To operate such an enormous structure at normal speeds such that the output is usable would demand unfathomable amperage.

Wetware vs Power

There’s three possible outcomes in the scenario involving synthetic brains.

The first is that the brains are actually fairly power efficient, but at the cost of training speed. The already fairly slow clock would become slowed down even further, but in turn power costs are kept to a reasonable level.

The second is that instead the materials these brains are made of can withstand enormous strain, and can be “overclocked” on their training, at the cost of exponential amounts of power.

Nonetheless, in either scenario it’s likely that horizontal scaling will be the natural solution. Why train only a single brain at once, when you could train numerous? Either way it scaling naturally will rise up to meet supply. However much power can be reasonably supplied to the training process, will automatically be consumed as fast as it is doled out.

It’s likely that the end product will be a fleet of synthetic brains cycling between training and synchronization.

The problem: We cant solve our problem by making it worse

It’s pretty safe to assume one of the most likely reasons to make an AI, would be to solve the current world issue of our addiction to fossil fuels.

In turn, these are the very mechanism by which we produce power, so if we needed such immense amounts of power to train an AI to solve fossil fuel problems, we’d have to first dump an unfathomable amount of additional CO2 into the atmosphere first just to even get started on solving the issue.

And to be honest, this is currently what is happening. Numerous companies are doing this exact thing right now.

Possible Solution: We need to solve Nuclear Fusion first

Every year that goes by demonstrates very small incremental gains on the front of Nuclear Fusion technology. It represents the holy grail of power generation. Clean, renewable, and yet more than capable of meeting infinite demand.

I posit that it would be fundamentally necessary for nations of the world to collectively join together in solving Nuclear Fusion first, before we can even start to bother trying to train a true AI.

Putting too much focus on AI while we are still tightly coupled to powering it with fossil fuels is going to end up wiping out our odds of long term success before we even get off the ground.

I wouldn’t be surprised if we power up the AI after decades of pollution payment, and when we ask it how to solve our pollution issues it informs us a pretty good first step would be to turn it off

Blocker 3: Cooling

Since we are already in the realm of positing that somehow the world powers have elected leaders with the forethought to see value in AI, let us assume these same leaders also had the forethought to work together and solve Nuclear Fusion prior.

We now have nuclear fusion reactors powering most of the world and limitless power to fuel our AI systems more than plenty.

However, with power draw comes heat, and with heat comes a necessary requirement to do something about it

Due to their immense heat production, it’s pretty safe to assume the cooling would need to be solved before we can fire up the AI to solve that problem for itself.

Now, at the very least, only a solution is needed to run the AI just long enough for it to solve its own cooling problem before it has to shut down and cool off naturally.

At which point we can then implement its proposed solution, and then spin it back up again (with a longer consecutive runtime now) to further self optimize.

Once we broach this last blocker we likely will enter The Singularity, as the machine now is actually capable of operating and self iterating, at which point it breaks free of the shackles of human mediocrity.

MSS vs Cooling

Water. So so so much water. Lakes will be drained, rivers damned, and ecosystems destroyed. Already existing superstructures (that do not even remotely come close to the sort of demands an MSS would ask of its operators) cause these sorts of issues. Applying this problem to an MSS however would take the issue and ramp it up to levels that will likely be incredibly difficult to approve.

Wetware vs Cooling

In the first scenario where synthetic brains simply just run at a much slower pace and inherently don’t consume much power. Which in turns means cooling issues will be fairly trivial to solve.

However, of course, this still introduces the time problem. Training will take decades and struggle to produce any usable results early on. Immense patience will be needed (which we all can agree isn’t something world powers are known to possess)

Alternatively, assuming we develop a synthetic brain capable of overclocking, it will be absolutely necessary that the brain be sufficiently cooled to compensate for the utterly enormous amount of friction produced during the process of operating the brain at such a high speed. Huge challenges will arise out of the fact that the center of the material will inherently need a way to conduit heat outwards for removal. Channels for heat transfer will need to be included in the design of such a system.

Possible Solution 1: “Hot” Superconductors

It’s possible that further advancements in Superconductor technology could largely alleviate these issues in one way or another. It’s even possible some form of Superconductor material is the same material that serves as the basis for the “wetware” solution

Either way, if one can manage to get a Superconductor material operating at “hot” (relative) temperatures that it genuinely becomes feasible to construct large portions of an MSS system out of it, then many of these issues become solved.

Of course, a new issue arises here, and that is one of production rate. You’d need to develop a way to mass produce this new material at a high enough volume to satisfy the needs for developing AIs. Wetware may possibly need less, compared to the sheer size of a single MSS system.

However this is yet another case of world powers needing to work together on solving a technology that doesn’t present immediate payoff.

Possible Solution 2: Throw it in the Ocean

This is an approach being experimented with various technological systems today. Various large tech companies are trying out dunking their enormous computational and other such storage systems deep in the ocean, letting the nearly infinite volume of surrounding water operate as the worlds biggest heatsink.

While this does seem possible, for the case of an MSS it would introduce yet another very huge technical challenge, as now not only do you have to build an MSS, but you have to build it deep under the ocean too. So, you know, good luck with that.

For the case of a synthetic brain, even using the ocean as a possible heatsink doesn’t magically solve the problem of needing to get the heat at the center of the “brain” to the outside. That issue still needs to be addressed by some other mechanism.

However, if you do manage to solve that, you can at least then deposit all that heat into surrounding ocean water.

Of course, even if you do this you’d have to contend with the possible massive ecological impact either approach has.

Furthermore the problem of power delivery still needs addressing, so you’d have to somehow have a nearby nuclear fusion reactor operational within range to deliver the massive power demand to the AI as it operates on the sea floor. Not necessarily impossible but yet another challenge in the equation.

Summary: It’s just the Iron Triangle again

1. The AI will take an enormous amount of time to become operational, requiring world powers to be patient and able to actually plan for the future (for once)

2. The AI will consume an enormous amount of power, to such a degree that solving this with anything other than renewable power sources would produce such extreme levels of pollution that we’d be dooming ourselves anyways. It’d be probable we need to get Nuclear Fusion working first.

3. The AI will produce enormous amounts of heat you have to figure out how to deal with, requiring either huge logistical challenges under the ocean, or solving “hot” superconductors

Pick two.