What should we do if our jobs get taken by AI?

Really don't need to worry about this right now. There's a few big reasons why. First, I'm always astonished when people say banks would chomp at the opportunity to outsource their entire labor pool to Ai, because you can already do so with cheap labor in India, Eastern Europe, or some other area of the world. Yet, companies continue to bring in analysts. Why? Because if you don't train juniors, in 10 years you won't have seniors, and those seniors more than make up for the cost of losing 90% of your analyst class over time. Analysts and associates rarely pay for their paychecks with their work. Just straight up, if a company wanted to cut costs they'd do so with entry level, because for the first few years of your career you just aren't making enough for the firm to justify your salary. 

That's more the political reasoning, but there is another element: The technical element. I've made previous comments on this issue, which I'll just paste below

Ai is a bit overblown, and the main reason behind this is costs. People oftentimes think that costs of Ai scale linearly, so it's very easy to replace juniors and have the costs be comparative. This is not true. LLMs operate on transformer technology, which means it can basically take in a whole sentence and connect it to other parts of the sentence. So, say for example you type in "The black cat is playing in the tree. It is happy" or something. Well, older Neural Networks couldn't actually connect the first part to the second. It would have difficulty connecting "Black cat" to "tree" to "It" as in it would think these are very different things. This, of course, causes issues. Transformers were able to provide "attention" tokens to each of these elements, so basically the transformer could connect the cat, the tree, and the it being the cat, with the black cat being happy. 

Here's the problem though: with that, every sentence is made up of words, and so the LLM must piece together each word to every other word. What that means, in a practical sense, is that "The black cat" gets compared against every word/phrase in the sentence. Now, for this short of a sentence, this isn't an issue. But, let's imagine we wanted to create an entire model for a presentation. We would need corporate earnings calls, 10k's, news reports, maybe some alternative data. All of this requires the LLM to try and connect A to B to C. The problem is that A isn't just connecting to B, but rather to C and D and everything else. If you know CS and big o time, you'll recognize this is o(n^2) time, which means the more and more parameters we feed the model, the slower and slower it runs. This means every new prompt continually causes a massive increase in computation. Taking just 1 10k could mean 10s of thousands of different parameters to work with. This also leads to hallucinations. The LLM doesn't know whether to connect A to B or to C based on some information. It can guess, or with newer models it can reason pretty well, but again that costs computation power.

What this means, in all practical terms, is that Ai is currently heavily subsidized by VC/Growth equity forms, or big tech earnings. These companies are spending tens of billions of dollars on GPUs to try and make the best model to win the race, but the problem is that for these crazy computational tasks, with a lot of different parameters, things are insanely expensive. Even typing one thing into an LLM can lead to a massive cost overrun. 

SSMs can improve this a bit. They work by basically having an updating "index" so to say, so pretend instead of needing to store an entire 10k and pair everything in it, the SSM will create an index page that will update with new information. However, this creates a lot of new issues, and SSMs are still very much in their infancy. 

So, what does this mean in practical terms? Well, once all these companies supplying the models in Google, OpenAi, Anthropic, etc. realize that they're basically blowing billions for nothing in return, they'll start charging the actual costs of the models. How much would that be? Well, OpenAi stated in the past that their 200$ a month subscription actually lost money. Yes, 200 a month losing money. No wonder all the Ai companies are constantly fundraising. And big tech is literally saying they don't know how to make money off these things. If the 200 dollar model(which is for individuals) isn't working out, imagine that cost, but scaled up to an entire organization with thousands of inputs from different people and new info, and all of a sudden you have a behemoth model that requires supercomputers to spit out basic information. 

That's one of my longer technical writeups on Ai, which has been slightly adjusted over the past couple months as more news comes out. First, I would say we're facing yet another issue to the Ai boom that wasn't as present when I wrote that: Capital and energy. Right now, we're deploying so much capital to these Ai servers and seeing very little in return. The massive cyclical nature of recent deals is showing a massive potential issue for Ai, namely the fact that there might not be enough capital to actually get the models where they need to be. Like, if AGI is achieved, sure, it's worth it, but that's a big if. We aren't even aware if AGI is physically possible. The other is energy. Right now, Microsoft has billions worth of GPUs just sitting, because they can't get enough power to actually run the things. The US baseload power is not strong enough. China actually has us beat strongly here, because they do nuclear so cheap. The current administration seems to want to solve this issue with just more oil, but this is really unsustainable. So the US will need to basically design their entire industrial policy around Ai for the next 10 years to see all these major issues fixed, and we're not even sure what we're getting from that. 

I'd also like to point out, Ai is just a predictive model, and a predictive model is based on a lot of regression. It's not thinking. It's statistics. And here's the major issue with statistics: It cannot predict things that have not occurred. In Talib's Fooled by Randomness, he makes this point clear. I know this too, having studied math and stats in college. Take the 2008 crisis. A predictive model will make the assumption that, since in all observable history it has been trained on, there has never been a major decline in housing prices, that it must not be possible. That is a fundamental problem with statistical inference, and a big reason I think statistical inference is kinda crap. More data doesn't always make things better. 

These models can only predict the future based on the past. But what if the past events had never happened before? What if we build a model that states that, since interest rates hadn't risen very aggressively in decades, that it was considered statistically impossible, or highly improbable, that they would rise again in the next few years. That model would've been crushed in 2022. 

So relying purely on predictive models that break when shit really hits the fan is like trusting risk management to Bill Hwang.