Gary Marcus Lists Eight Reasons to Slow AI Development

Eight reasons to slow down hyperscaling: • What GenAI has already done to society • What AI slop is doing to the internet • What data center overbuild is likely to do the economy • What data center overbuild is likely to do the environment • What cyberattacks may do to the integrity of our data • The software crisis that AI slop code is likely to lead to • The complete lack of a plan for what to do about employment • The complete lack of a solution to the alignment problem

Eight reasons to slow down hyperscaling; • What GenAI has already done to society • What AI slop is doing to the internet • What data center overbuild is likely to do the economy • What data center overbuild is likely to do the environment • What cyberattacks may do to the integrity of our data • The software crisis that AI slop code is likely to lead to •The complete lack of a plan for what to do about employment • The complete lack of a solution to the alignment problem

@GaryMarcus This is just sad 💀 No one is slowing down - and why would they? Imagine whining about new technology and depriving the world of all the potential benefits like advances in medical treatments, self driving, AI assistants, etc

@GaryMarcus

@mogulseeker @GaryMarcus Where verifiable rewards are possible, there is probably some room to grow: https://arxiv.org/abs/2606.16140

@GaryMarcus Most "AI projects" are expensive flops (80-95% fail to deliver real ROI). Companies throw cash at shiny pilots, then watch them die from bad data, clueless leadership, vague goals, and zero integration into actual workflows.

@GaryMarcus some of these points are in tension with one another. if ai slop isnwhat it produces... how is that gonna affect employment? its like you are attacking AI from contradicting sides.

@GaryMarcus

Fair list — but treating all eight as equal is its own kind of hype, just inverted.

The defensible ones are economic: data center overbuild against negative unit economics, and the refactor bill coming due on AI slop code. Measurable, already showing up.

Alignment and societal harm are real but far more contested. The strongest case for slowing down is the one you can put a number on.

@GaryMarcus the broader PC industry is being suffocated by the ridiculous chip demands.

@GaryMarcus How ironic.. @elonmusk loves talking about the kardashev scale and yet…

@GaryMarcus Nine: For most normal people AI is unwanted and boring AF

Just ask my wife!

@GaryMarcus Those reasons are mostly due to having billionaires in charge of US AI strategy driving revenue for their current properties and job replacement for corporate profit. Surveillance/control are the biggest dangers.

What we need is a national AI strategy; Moon and back level ideas.

@GaryMarcus We need big smol

@GaryMarcus bro made a whole bingo card and we hitting every square in real time

and the response is still just 'but the token go brrr' 💀

Um…Gary, do you know that some of us have worked out solutions to some of these things in our lives?

We’re in a phase right now.

I think some are framing some things in a self-defeating way, and pushing for a vision which won’t stick.

But there are models already that can stick, and which I believe can work out over time, if they are surfaced, and supported.

You may be right about hyper scaling in one sense: we have everything we need now, we just need to measure the right things and align.

@GaryMarcus And it's not intelligent. There will be no AGI. All it's good for is hacking. And the idiots have put all our data onto centralized servers making a honeypot. A single point of attack.

Are they all insane? I can't believe that no one is stopping this runaway train.

@GaryMarcus The US government wants to put your money in....

@GaryMarcus few people payed attention to this https://arxiv.org/pdf/2506.02153

NVIDIA be like , LLMs are a waste of energy ... but keep building datacenters.