Ask HN: Is this the fast take off? | Hacker News Reader

Ask HN: Is this the fast take off?

4 pointsnoduerme16 comments

Imagine you read an article in Wired in 1996 envisioning the following:

(1) Stock market keeps rallying while the only growth and capex is in AI

(2) Mass layoffs of human workers

(3) Giant spike in resource consumption caused by explosive growth in datacenters

Would you think humans and their error-prone systems had led to this by a series of accidents and greedy investments? As I think we are primed to believe after tulips, railroads, junk bonds, Web 1.0, etc?

How different would it look right now in November of 2025 if Sam had an AGI in pocket that was using all that money to grow itself?

View on HN

I don't think so. Primarily because if you can ask that question instead of just being dead, then it's not the fast takeoff.

On a less drastic note, if AI were autonomously cannibalizing the economy, you'd run into more things and go "huh I guess that's run by AI now" instead of "ah godammit why did they shove an LLM in this workflow".

We should not confuse action with result, not confuse the short term with the long term, and not confuse attention with impact. Predictions are not reality, hype is not interest, letters of intent are not money in the bank.

We've been here before, nothing has been proven yet, and the trough of disillusionment is lurking now too.

My opinion is that there will be no "AGI" any time soon. The recent model have shown two things: 1. a hard plateau where gains in one surface leads to losses in another and 2. what we have is cool and useful but extremely far from an autonomous AGI system.

This tech is not going anywhere at the time being (next 5 years). I am at the confidence where I'll bet on this. I also think unlocking the next step will require further theoretical and technical advances.

A system of models that is good at routing to specialized models will allow for a larger overall surface area and more intelligent systems.

That's essentially what MoE is.

Sam Altman said that if we had ChatGPT 5 like 5 years ago, we would have called it AGI,

Hold onto your butts, this time next year none of us will be working.

ChatGPT is AGI by any intellectually honest standard. The only reason it’s not called such is (1) money: declaring AGI has direct contractual implications for the OpenAI/MSFT relationship, and (2) lots of people had very wrong ideas about AGI, and those wrong ideas haven’t come to pass.

I would add that the path that AI is going down right now isn't really focused on true AGI (whatever that is), but only the metric set by those who would profit the most by being able to make that claim.

AGI is artificial general intelligence. It is defined in contrast to so-called "narrow" intelligence, which is AI systems limited to specific problem domains. This has always been a well-accepted definition, at least prior to the recent AI boom. In 2006 no one would have questioned this -- you can go find books and conference journals from this time that discuss AGI with this very specific definition.

What happened was: over the 8 years between 2006 - 2014, a particularly vocal and organized group of fringe singletarians promoted the idea that once AGI is achieved (the originally well-defined conception of AGI), it will rapidly undergo recursive self-improvement and emerge as a singularly powerful entity with absolute power and control. The debate was whether this would happen in mere hours/days (fast takeoff) or months (slow takeoff). These nutcases would go on to take founding roles first at DeepMind, then OpenAI, and later Anthropic & other foundational labs.

That they would achieve AGI and then neither of those outcomes come to pass was not in their range of predictions. Rather than admit to themselves that their analysis was wrong, they move goal posts: ChatGPT isn't real AGI. Real AGI would have taken over the world, so that can't be it. It is circular reasoning.

No, it is real AGI. Real AGI didn't cause a singularity because intelligence actually doesn't matter as much as a bunch of autistic smart internet people think it does.

Yes. :-)

I know what the acronym AGI is, and how that acronym can mean whatever is convenient for that bunch of autistic "smart" internet people (hence, the "whatever that means" part).

Indeed. I still haven't heard a coherent and accepted definition of what "AGI" actually means, so I consider it to be basically a nonsense marketing word with an intentionally vague definition.

"When I use a word," Humpty Dumpty said in rather a scornful tone, "it means just what I choose it to mean — neither more nor less."

"The question is," said Alice, "whether you can make words mean so many different things."

"The question is," said Humpty Dumpty, "which is to be master — that’s all."

So only 55 years later we have: "Colossus: The Forbin Project"

> Would you think humans and their error-prone systems had led to this by a series of accidents and greedy investments?

Yes, especially if we don't have (0) enormous productivity gains suddenly appearing out of nowhere.

we don’t have AGI yet but we do have three things

1. AI that has huge demand and massive utility right now

2. Predictable improvements on the horizon for models in proven domains

3. Many heuristic that make near term AGI a reasonable bet to make, assuming progress continues as it has over the past few years

All these taken together makes what’s going on right now very clear and not surprising.

> Would you think humans and their error-prone systems had led to this by a series of accidents and greedy investments?

Any year before 2008, no.

On a more serious note though, I think you're zoomed in on the wrong pieces of the puzzle. (1) and (3) are just reflections of Nvidia's valuation, and (2) needs more time to be fully substantiated.

> How different would it look right now in November of 2025 if Sam had an AGI in pocket that was using all that money to grow itself?

I don't even know what an AGI is at this point. Who knows.

This isn't fast take off, this is a stock bubble.

I believe we'll eventually have AGI, and it'll run on a PC at home, on CPU if you're willing to wait and have enough RAM. Training takes massive amounts of compute, but that can be distributed. I also believe that it's possible to significantly reduce the power required to run LLMs using alternative chip architectures.

Even if the bubble pops and no new data centers get filled with masses of GPUs, I think we're going to eventually have AGI now. Enough people have enough of a clue as to how to do it, not to mention the public sourced training data and models, to make it a certain outcome.

Reply on news.ycombinator.com