Artificial Intelligence should have produced at least a few hundred billion dollars of value by now, if everything being said about it were true. Yet we keep seeing article after article claiming that AI initiatives are failing, despite a ton of investment into the space. Every C-suite is pushing it. What gives?
I believe there are three primary reasons, none of them having to do with the technology itself. If you still hold beliefs from 2023, such as “they hallucinate too much” or “I need to get better at prompting”, you need to spend some time working with the latest models (and not just via chat!). The technology is as capable as it needs to be to effect all but the AGI-required claims. I think that nearly every profession with more than an 80% balance of work on the knowledge side is replacable with the current level of LLMs when integrated into software products by experts in their fields.
So if I don’t think it’s the technology, what are the three issues stopping AI from revolutionizing things?
One: People Don’t Know How to Use It (“You’re Holding it Wrong”)
Prompting isn’t the only thing! And yet, people seem to think that the chat interface (e.g. ChatGPT) is where LLMs start and stop. This is insane! If this is your worldview, of course ChatGPT won’t replace you! How could it, when it requires a human to prompt it?
If you know that LLMs go beyond the chat interface, you might think that the above worldview is absurdly rare. I am here to tell you that the vast majority of people think this is how an LLM works. They would also call it AI.
Once you sit and play with an LLM via API, watch the logs of a thinking LLM, compose a pipeline of various LLMs together with skills and tool calls, mess with an LLM calling other LLMs, and just generally immerse yourself in what people are doing, you will have what I can only describe as an awakening experience. I had one back in 2023 when I watched LLMs answer pretty decently at how they’d handle a problem in my field. Then the shine wore off because the state of LLMs at the time wasn’t really “all there”. Now it is.
Two: People Don’t Think It’s Possible (For a Machine to Do Their Jobs)
I have seen some wildly smart people scoff at the idea that an LLM could take their job. Thus, they don’t even use them or try to use them to their full potential. In almost every case, I think they’re wrong. My reasoning is as follows:
1) Can you define what it is you do?
2) Can you decompose what you do into some number of distinct skills, processes, and tools? It’s okay if that number is massive, and even if they relate recursively or in a messy web! Can you do it? I think literally everyone can. Even a brain surgeon. Heck, a pilot’s job in the absolute worst case scenario is to pull out the checklist and follow it. If you sit and think about it for a long time, I bet you can fully define your job. You can also use an LLM to help you do this in a shorter amount of time.
3) If you can define what you do, and you can decompose it, then you can break it into clear areas of concern (just like any good software), which makes it much easier for an LLM (especially the latest ones) to handle the wild reality in hard jobs.
Three: People Don’t Have an Incentive to Do It (They Don’t Own It)
As Charlie Munger said: “Show me the incentive, I’ll show you the outcome.”
Supposing for a minute that someone had the skills and the belief that it’s possible - why in the world would anyone train their replacement? Add onto that that, for the vast majority of people, they don’t own any shares in the company they work for, let alone an appreciable amount (remember, if you do it right, you’re going to get fired - is a few hundred grand in stock going to really help that much, especially given that if this can happen at your workplace, it’s probably happening at similar workplaces, meaning your chances of getting a job go down appreciably?).
People working as salaried or hourly employees (read: pretty much everyone except the C-suite and the shareholders) have zero incentive to try to replace themselves!
So What Needs to Change?
A few fundamental shifts need to happen, two of them significantly easier than the third.
First, people need to get more skilled in using LLMs. Yes, that means getting better at prompting, but much more than that: understanding how an LLM fundamentally works, how modern state of the art LLMs work, how context works, how chat interfaces differ from coding assistants which differ from direct API usage, what a harness is, what an agent actually is, and yes, how to keep LLMs from hallucinating (which is getting more and more rare).
This only happens with time and budget. I have personally spent a few hundred dollars over the last three years on plans and tokens. My employers have spend a few thousand dollars. This is pennies, both for personal development and also for the value my employers have gotten out of my using LLMs. This also can’t happen overnight, or with a quick training course. Just like learning any skill, you need time in the seat. Imagine trying to learn how to speak a different language with just some quick pointers about sentence construction in a one-hour training.
Second, people need to genuinely wargame the idea that “LLMs can’t replace me”. The old Upton Sinclair quote about “it is impossible to get a man to understand something if his salary depends on it” is so true. You can do this on your own time if you think your employer is going to look at your chat logs. Now, this is not to say that a random guy off the street could come and replace me, even with a massive budget and plenty of time. There is intricate, experiental knowledge that I have about my field that he would never know, nor that an LLM would ever surface on its own. But there are tens of thousands of people in my field who, if they only spent the time, effort, and money, could absolutely build something as capable or more capable than me. The only way I can defend against that is by combining my skills with an LLM, which they cannot compete with unless they also integrate themselves. Even if an LLM-based product is developed that could totally beat me on my own, if I have that product and add myself to it via using it, I can only improve it (because I don’t think LLMs have a full grasp of the world, which I still can add).
Third, and this is the hardest one, even if people gain the skills, and lose the hubris, they may still not implement AI to their or its full capabilities at work. Why? Because it is nearly only downside for them. It’s possible that their employers would keep them on, and just add new responsibilities. This assumes that there is room in the market for even more of their company’s product. If everyone does this, then suddenly we have a mass oversupply, but since everyone is also being more efficient due to AI usage, they no longer need to consume as much. I see mass AI usage as being very destructive to demand, since it will likely lead to mass job loss since machines will be capable of doing more of the jobs that used to require a ton of humans. Humans without jobs have no money and can’t buy much.
The more likely outcome of people implementing AI to their and its full potential is that they get fired, the company and its shareholders keep the agent(s) the employee built, and only the company and its shareholders see a benefit. It’s the infinitely replicable version of the person you’re asked to train before you get fired.
Could there be some class traitors? Employees who, even though they probably don’t own any stock, will try to replace themselves with AI just because they think they’ll get a raise or a bonus? Probably. However, the type of person to do this is likely not the smartest tool in the shed. After all, they’re going to get themselves replaced. If you extrapolate their intelligence in that respect to their ability to work with AI (I do), then I doubt they’ll be very successful. This is also related to why I don’t think any of those “help train our LLM” “jobs” on LinkedIn are going to get any value out of the results. It’s just a bunch of overzealous goofballs answering those, the exact type of people you don’t want answering. No competent professional is going to give away real game for pennies like that.
What about consultants that do it for companies? You know, come in and attempt to extract knowledge through interviews of employees, document everything, then create a system to replace them? I think this is the most likely option, since they’re mercenaries. They have no allegiance to the company, so they don’t care if everyone gets fired. This already was a thing, but because this mostly happened pre-LLM, all the fuzziness around people’s jobs was hard to actually replace in practice, even if you understood it in theory.
What if employers gave more stock to employees who replaced themselves? It would have to be enough to set them up for life, and for their progeny to be set up for eternity, because we’re talking about a complete destruction of future job prospects in that field (which means probably also the same thing will happen in many other fields). I don’t think this is likely, since shareholders hate giving away appreciable amounts of stock.
Thus, the other most likely way all of this comes into fruition is companies being started by professionals who deeply understand their field, who are going to essentially build replicas of themselves. And just like there are at least hundreds of good consultancies in every field, if not thousands, each of these will have their own implementations, some better than others, some worse than others, but all better than they would be if they were operating on their own. Some will offer these as rentable services that other LLMs can use. Others will use them in their own practices with clients. This is my bet for the most likely outcome because the incentives are deeply tied. The better I replace my job (by making something that I actually own), the more I am protected from someone else replacing my job. Yes, this sort of company is still reliant on a megacorp like OpenAI or Anthropic for the core models, but open source is only getting better, and besides, the core model is really not that important past a certain point. Any of the top few models nowadays are plenty good, if wrapped in true expertise.