Ask HN: Can we just admit we want to replace jobs with AI?

Who’s rolling out, managing, scaling, and iterating this mass AI?

How many mass AIs are there going to need to be?

Who’s deciding what to use it for??

Who’s auditing the mass AI’s output?

Who will you call when the AI inevitably go off the rails?

Who will debug the mass AI that went off the rails so the lessons can be learned and integrated into the next generation of mass AI?

Who will build, deploy, and supervise that next generation of mass AI?

These and other questions, attempted thoughtfully one by one, should make it harder and harder to hold on to a belief that we can ever run out of work for humans. The work will keep getting better as more is done for us - more creative, and less drudgerous; but it’ll never stop being directed by us because we are the interface to the human world.

Good question.

My sense is that there is a lot of compute capacity, data acquisition, etc., required to create any dent in this type of automation.

Likely the next 2 years are focused on more “compute”, “data center capacity”, and “energy”. That allows for a) better training, b) more inference load, … and hence prepare for more automation.

Much like the Y2K era of 30 years back, we are entering the new era of YNA, an era whY Not Automate.

[dead]

> We've had human shaped robots for decades, but so what ?

They've improved a lot since the ones decades ago. When ASIMO was state of the art, that meant being able to walk on a flat surface and climb stairs.

Now, it's more like this: https://www.youtube.com/watch?v=WlUFoZstcWg

(Though again, as this is a sales video, I expect it to be implying more than it can actually do, in the same vein as the Tesla Optimus robot bartenders but not necessarily the same specifics)

> Robots that not only have AGI (running locally, under battery power?!)

Assuming batteries are a must and we can't e.g. use mains, that means the gap between AI reaching the power-performance envelope needed for level 5 self driving cars and the one needed for such robots would be about 5 years. Get the former in x years from now, the latter in x+5 years.

> but also the dexterity, flexibility, etc to match that of a plumber seem a long, long way off

Single humanoid hand, solving a Rubik's Cube, real time video from 5 years ago: https://www.youtube.com/watch?v=kVmp0uGtShk

> and would of course require something a lot more advanced that today's LLM based AI

The future existence of a sufficiently more advanced AI is baked into the assumptions in the question at the top of this thread. It's the goal and driving purpose of the field.

Dunno when it comes; if I did, that would make it easier to prepare for.

> would need to have full blown brain-like capabilities, obviously including the mechanisms to learn a craft.

How confident are you about what specifically they need and don't have, and why?

I've seen people confidently say AI would need brain-like capabilities to play Chess, then when Deep Blue beat Kasparov they were saying it about Go, then when AlphaGo beat Lee Sedol they were saying it about natural language and creating pictures, then ChatGPT happened and various Transformer and Diffusion models got so good that collectively they've become a fraud risk.

Based on the observation that LLMs could get up to the level of interns or mid-university students, I suspect we've already got AI capable of learning a craft to that (mediocre) level just by training on all the YouTube DIY videos.

Based on all the sim2real work, I think it's at least plausible (not certain, only plausible) that some mediocre plumber AI bootstrapped by YouTube would also be able to improve substantially in this fashion: https://arxiv.org/search/?query=sim2real&searchtype=all&sour...

I don't think we need to solve, e.g., the question of why humans seem to be able to learn from so much less data than our AI require to reach any given skill level. Likewise, we do we don't need to agree on what "consciousness" is, nor do we need to then give that to the machines.

Background in telecommunications/signal processing (phd) + applied math. I studied finance + machine learning for two years as a hobby before the switch. In the evenings and weekends. Studied data structures + algorithms + python for two years before took a job as a software engineer in google (initially as a data scientist in apple). Worked 6 days a week for 1.5 yrs to master information retrieval in google. So lots of studying was involved on top of a solid educational base.