So, one thing I'm wondering is why is China hopping in on the "AI" bullshit train anyway?

Are they just that prone to falling for grifts?

Is the spam/propaganda generation ability worth that much to them?

@lispi314 they have the most advanced human surveillance technology in the world. And this is just the next phase of that technology. Putting this on people's phones and in their houses is like the ultimate in surveillance and control.

Venture capitalists love to talk about the single person billion dollar company. The Chinese Communist Party has 100 million members. If a technology like this can get rid of even a percentage of those, it allows you to tighten your grip as an Autocracy.

The companies building this technology today are all basically State assets. So they get direct access and control without any of the public fights like Anthropic.

So I'll flip this around, why do you think they shouldn't be doing this? What grift do you think they are being subjected to?

@gatesvp Because it's stupidly unreliable and might lead to them purging completely uncompromised party members.

@lispi314

"Purging completely uncompromised party members...

Who cares? The purpose isn't to find or purge compromised members, the purpose of having an automated system is to replace the number of party members you actually need to keep the system running. The best way to minimize compromises is that you have less people in the system.

Because it's stupidly unreliable...

I mean, have you met humans?

They lie. They cheat. They steal. They try to undermine your control and authority.

And they don't have built-in control and auditing systems like the robots do. And they're more expensive.

I'm still not understanding your arguments here.

@gatesvp

Who cares?

It was literally the thought experiment asked. The fact no one does is irrelevant.

Hastening the system falling over under its own dysfunction is not keeping it working with minimal resources unless you have a very generous definition of "working".

And yet, despite all that, humans still do better at somewhat-durable dynamic systems.

They still do better at semantic and logical reasoning (making a computer worse at logic than a human is an achievement, if a very dubious one).

And they're more expensive.

Not even remotely. The actual costs of the logistical chain behind "AI"? Ridiculous.

@lispi314

It's clear that you and them have very different base level assumptions about the cost and value generation of these systems.

Sure, if I take your cost and values, then none of this makes any sense.

But they're clearly operating from different cost and values. So to them, this makes sense.

They're running the world's most advanced AI-powered surveillance system. Currently pointed at their own people. But they've been running that system for a very long time. Quite successfully, by authoritarian government metrics.

Before ChatGPT was launched, China already had years of experience using LLMs. Years of experience using automated image and facial recognition. At huge scale.

You're operating from the baseline assumption that this is some American grift being pushed by a couple of companies trying to make some money. China is not.

And the only way you're going to understand their perspective is to restart your interaction with this technology, from scratch, wearing Chinese glasses instead.

@gatesvp > And the only way you're going to understand their perspective is to restart your interaction with this technology from scratch, wearing Chinese glasses instead

I already confessed to being hanzi illiterate though. I can't do that.

@lispi314 So you're operating under two technological assumptions. Number one, the AI technology works poorly. Number two, the AI technology is prohibitively expensive to develop and operate.

I used to work at Roblox, on the trust and safety team. We are basically an authoritarian regime on what you are allowed to say on the platform. Particularly for people under 13.

We used LLM technology for years before ChatGPT was available. It was not only affordable, it performed text interpretation and redaction that was literally impossible for humans to complete. It was able to escalate particularly dangerous text up to moderators for further review. It supported multiple languages. It interpreted weird utf-8 characters in multiple ways to prevent people bypassing the filter system with emoji or weird characters.

This thing did not work poorly, it worked very well. And it did so at an efficiency that was unmatchable by humans. And this was 2021.

@gatesvp You were externalizing the cost of building the datacenters, maintaining them, and building the computers and hardware to use in them & powering them.

Obviously it's "cheap" if you just ignore the majority of the costs.

China's use-case implies end-to-end TCO & OpEx.

@lispi314 back before we were public, the books were pretty transparent... that system ran on a small percentage of Roblox's infra.

We spent 100x the resources on the servers for hosting games. It was so cheap to operate, nobody noticed. I was the senior tech lead on the team, not once did anyone wave a bill in front of me or ask me to justify costs.

And all of this stuff is cheaper now because we have better GPUs for even less money.

You have a vision of cost, but I don't know where it comes from?

@gatesvp From the entire transformation chain (China won't simply use someone else's infrastructure).

Did you own a single one of your servers or were they all rented?

And the fact it was so cheap, you were using some off the shelf model or yet another service you couldn't recreate yourself if the provider decided to misbehave? Or did you actually have the in-house expertise to do all of that (which would indeed dwarf the actual cost of the servers)?

You were were also mainly/only processing text. That's a lot lighter.