AI And Life Experience
The next frontier for AI data could be a challenge for OpenAI and Anthropic
In Derek Thompson’s podcast with economist David Deming, Deming argued that AI won’t be able to fully replace you because there are things about a subject you can’t just learn on the internet.
And there's all kinds of things you do that you basically replace the lower rung of farming, the manual labor of actually tilling and plowing the fields and you, what you've got left is the higher order things that, that are left in the job. And so I do think that's why I think it'll be disruptive 'cause we'll see that what firms expect from white collar office workers, it's just gonna be ratcheted up very quickly. So it's like this kind of re useful but relatively shallow function that deep research performs at the end of the day, which is just scanning more or less the content of the internet to synthesize a bunch of stuff and put it together for you. It feels really deep 'cause that's what we've been teaching people. But in the, in some sense it's just taking everything that's out there and putting a loss on it. And I, I know I say that because I use deep research a lot and when I use it I'll, if I ask it to write a paper about something I know a lot about, so as a labor economist or if I ask it to write about soft skills or higher education or whatever, it's undergraduate term paper good, but it's not really good.
It's not as good as me, not because I'm amazing, but because I've spent my life thinking about these issues and I've read the papers, it's reading and I have some nuance in some sense. I have knowledge about that that is implicit that's outside, that's not, that doesn't exist on the internet, how certain people talked about it, what the debates were and all of that knowledge that I've embedded that I've encoded in my mind that's not on the internet I have access to and ChatGPT he doesn't. And so I still have an advantage over the, over the, over the machine in the things I know really well. But the things I don't know well, like if we were gonna do a podcast about, you know, the history of farming or something, I would use deep research and I would rely on it because it probably is gonna get the details wrong, but it's still gonna be better than my existing knowledge, which is pretty shallow. And so what I think you you'll want to do in the future is you'll want to become deeply expert in something beyond, beyond whatever knowledge exists on the internet.
My first thought is I don’t know if this is even true. I have a hunch that if you stage a debate between somebody with 130 IQ and somebody 100 IQ on a topic that the 100 IQ person knew more about coming beforehand, the 130 IQ person will still wipe the floor with him. He just has a much higher capability to reason about things. Now, imagine that the competitors are a 10,000 IQ AI and a 125 IQ professor. I don’t care how much time he’s spent thinking about the topic, the AI is going to absolutely wash him.
On the flip side, I’m not sure the 130 IQ person would be able to fix his stove as well as his 100 IQ handyman. Or farm as well as the 100 IQ career farmer. So, maybe there is some value to life experience. You may not be able to win an intellectual debate about the ethics of farming, but you’re definitely better at growing corn.
So, let’s say Deming is right and his life experience does provide an advantage over an AI that only has access to the internet. This would make the ability to incorporate life experiences into the training of AI extraordinarily valuable.
The direct implication is that Apple, Google, Meta, and xAI/Tesla (I’m counting the Musk companies together here) would now have a leg up over OpenAI and Anthropic, as they can all incorporate AI into their existing hardware products and thus can accumulate real-life experience. Apple has the iPhone and AirPods, Google has their phones, Meta has the glasses, and Tesla has the car and robotics. OpenAI and Anthropic, meanwhile, are limited to just the internet, at least for now.
The other implication is that surveillance is going to get a lot worse. It’s now in companies direct interest to listen in on what you are saying and looking at what you’re looking at at all times. I’m not sure what this ultimately means for society, but it makes me very uneasy.
What I am fairly sure of is that OpenAI and Anthropic aren’t blind to this reality. They really should be trying to get some hardware out ASAP. I propose they build or buy something like the Friend AI necklace. I’m on record for thinking it’s a terrible idea that’ll never work, but if it does, the data that an always listening necklace can bring would be extremely valuable.