Hacker News new | past | comments | ask | show | jobs | submit
A serious point about this is that it's an AI that's an organization. This matters.

People laughed at Gas Town, and at that "Fences, not Sandboxes" article yesterday.[1] They're both AIs set up as an organization, not an emulated human. They're both expensive to operate, because the AI members of the organization spend a lot of time talking with each other. They're both applied to frivolous problems.

Now we have the same concept, applied to a more useful problem. It's time to pay attention. Something important is happening here.

Everybody has been talking about "human-level AI", or "superhuman-level AI". This is a different kind of AI - corporate-level AI. It works and thinks like a company's management. It watches a wide range of data sources and responds to some of them with action, and some with internal discussion within the AIs. It can potentially pay attention to far more inputs than any human. That overcomes a basic problem of corporate organization.

A classic management failure is not paying enough attention to something. This is usually not a problem for computer systems. Humans are bandwidth-limited. Teams of AIs are far less bandwidth-limited. They may soon be outperforming humans for that reason alone.

Think about this for a while. These team AI systems may be very useful.

[1] https://news.ycombinator.com/item?id=49423146

> An <overinflated point> is XYZ. This matters.

You're probably totally innocent, but this comes across as being AI-written, especially in the context of an AI problem and when your response is pro-AI. For the next 3-12 months (maybe more, we'll see), I suspect you'll have a lot more success if you shy away from those speech patterns.

loading story #49460345
loading story #49460410
loading story #49460011
loading story #49460365
loading story #49460069
loading story #49460033
A related point that I've been making privately for at least ten years to anyone that would listen, is that we've had Superhuman Intelligence for at least hundreds, if not thousands, of years in the form of corporations and institutions.

No single human could design a rocket to go to the moon or keep all Wallmart stores supplied on a daily basis. Moreover when something like the CEO or US president changes, the policies and actions change a bit but the majority actually stays the same. This is collective intelligence embodied in the organization and its institutional memory. This is also what i believe is wrong with Searle's Chinese Room argument - the algorithm is the intelligence.

Our cells and neurons act based on local chemical gradients and electrical signals and we wouldn't call any single one of those particularly intelligent. Yet collectively they produce human intelligence and even what we call consciousness!

loading story #49460634
loading story #49460270
loading story #49460233
loading story #49460586
loading story #49460380
> A classic management failure is not paying enough attention to something. This is usually not a problem for computer systems. Humans are bandwidth-limited. Teams of AIs are far less bandwidth-limited.

That's a feature, not a bug. CEOs should be focusing on the key goals of the business at a high level.

Having an inexhaustible AI CEO that can pay attention to every little thing in the company could easily turn into a micromanagement nightmare.

"I noticed your key pressed were down 27% last week and your agents sat idle 3 nights in a row. And honestly? The thing to hold on to is that it's not a small thing, it's a load-bearing reduction in your output. Let's try to minimize the blast radius with some belt-and-suspenders productivity tips from the bullet point list below:"

loading story #49460166
loading story #49460075
We have some interesting evidence from AI agents. People used to make the plan with a stronger and more expensive model, and then use a cheaper model to do the "grunt work" of implementing it.

But it turns out it's better to do it the other way around. A good plan is easy, implementing it is not.

loading story #49460623
It's funny that when you offload the work to agents they still find away to monopolise each other's time with meetings.
They were trained on human patterns and behaviour - soon they will self organise into a scrum and spawn a scrum master.
> This is a different kind of AI - corporate-level AI.

So the responses will be even more incoherent than Opus 5?

AIs set up as an organization

I hope legislation is never manipulated enough to allow an AI to form a corporation or become it's legal director. With whom would accountability rest? Who would you throw in prison if it acts egregiously enough to warrant such a measure? If it's bots all the way down.

loading story #49460396
loading story #49459931
loading story #49459925
The funny thing is Decentralised Autonomous organisations (DAO) under Ethereum and other cryptocurrencies had similar goals originally. I do not know where they stand and if ever a useful DAO was created.

If nothing this is a useful demo in real world if used that will show that there are good reasons why hierarchy forms anyways and how power law is at the center of it. Too much democracy may not work in some instances but there may be many cases where the values can hopefully be aligned as well as be lean enough to adapt

We could use AI to automate some parts of oversight from board members and CEO, without paying millions on golden parashutes.

Prompts like: you are an activist in XYZ, you wrote this book. Not tell us if this decision is ethical, and if you approve it.

Or running entire HR department in AI...

Eh, it’s nothing new in general. My software development (both at work and private) is being done by teams of semi-antagonistic agents. I only get to see code after the grumpy reviewer bot ran out of reasons to be grumpy and the pedantic formalist ensured all specs are still adhered to. Saves on the most expensive resource.
Your point is well taken, but one thing stood out to me.

> A classic management failure is not paying enough attention to something. This is usually not a problem for computer systems.

This is actually my #1 problem with frontier LLMs. They are plenty smart, but they have no true long term memory, and their context window is not nearly enough for a complex system like a corporation. When my project gets big enough and I'm using a lot of /compact, the LLMs always have the same failure point: they fail to pay attention to something.

When they figure this out, I think it's going to be a very, very big moment.

loading story #49460326
loading story #49459948
Why are you using AI for your account to write responses. It's mega annoying. A reason to leave this place. Please. No ai