Comment by 0xy | Hacker News Reader

Comment by 0xy | Hacker News Reader

Another luddite article complaining about farming automation putting farmers out of work, but a modern equivalent.

This article wrongly assumes AGI is not only possible but also imminent, because if they were taking into account only the transformation we're seeing from AI at the moment -- it wouldn't be a story as it is not job ending.

AGI, however, is mathematically impossible. The only people telling you otherwise are the CEOs of labs who need to publicly fundraise on this premise, while privately admitting bearish sentiments on AGI.

AGI bulls assume all of the following to be true: there are no constraints to grid infrastructure (clearly false), there are no manufacturing constraints for AI hardware (clearly false) and exponential accuracy, speed and efficiency improvements will continue (clearly false, it's slowing down already).

Hell, just look at local opposition to data center deployment. You can't even get DCs built in rural towns that would benefit dramatically from the 1,000+ temporary and permanent jobs. Incredibly bearish on AGI.

The article is mostly inquisitive, it asks a question. What you are attributing to the article author is mostly in your head.

I upvoted you right off reading the first line.

But then you drifted.

I have no idea what you mean by "AGI, however, is mathematically impossible."

Further, your point about political pushback is short sighted. As AI becomes more lucrative there will be more impetuous to "pay" locations to have data centers, and as that becomes too expensive space is clearly the next answer.

The development of AGI assumes zero constraints, when constraints exist at every layer of the stack. That's why it's mathematically impossible.

In a system driven by capital, manufacturing can ramp to an extent but they generally can't exponentially ramp due to dependencies they have.

When you ramp one layer of the stack, other layers of the stack are pressurized. We're seeing a small preview of that now with memory pricing. But these break points for AGI are everywhere. Power capacity, power infrastructure, DC labor, cooling systems, memory, motherboards, GPUs. All of these things have dependencies that cannot be scaled exponentially, or quickly. As you pressure points of each of these dependencies, prices rise exponentially.

Let's take memory for instance, it is merely one block in the jenga tower but it's a good example. Memory is already at close to 100% capacity. Spinning up new capacity is highly constrained, and money can't really make it faster. Lead times are 4+ years on new plants, which cost billions.

The same is true for other components, and in some cases the situation is worse.

"Won't happen for 4+ years" and "mathematically impossible" are quite different. Given that humans apparently exhibit the "GI" part of "AGI", I find "mathematically impossible" difficult to believe. "Extremely unlikely with current LLM architecture", sure, but that's a very different statement from "mathematically impossible".

If you are making a prediction on the viability of AGI assuming that an entirely new technology will make the efficiency problem of LLMs moot then you're essentially engaging in mysticism, aren't you?

It is correct to say it is mathematically impossible, as all the people making AGI claims rely upon advances that are not even theoretical, they have not even been discovered yet, and the mere possibility of them is questioned by many scientists.

LLMs have hard and soft limits all over the place preventing AGI. You aren't gonna train and loop yourself to AGI because the compute does not exist, and will not exist.

My 4+ year point was for a single memory fab. Increasing capacity by merely 5% (generous assumption) takes 4 years and $10bn. It's starting to sound like the path to AGI in the current paradigm will cost infinite dollars and take infinite years of build-out.

Even with a transformational efficiency breakthrough, you still have hard limits all over the place. Where are you going to store all the data? Memory constraints again.

Reply on news.ycombinator.com