AI companies have a vision for the future. An AI that can do anything, solve any problem, process any request. Artificial General Intelligence. But the point of having this technology is the same as any other product or service that a company produces, to generate a profit. AI companies want this technology so they can be the ones to capture the market in any industry. Want to watch endless media? They want their AI to produce it. Want to transport your goods from one location to another? They want their AI to move it. Want new breakthroughs in scientific and medical research? They want their AI to solve it.

But there’s a problem. Current AI technology cannot learn as efficiently as humans. A teenager can learn to drive a car after a couple of hours of practice, while self-driving technology has been developed for decades and still cannot cope with every situation and environment. A person can learn to program computer after a bit of practice, but modern LLMs have yet to make human programmers obsolete after training on millions of lines of code. A child can spell the word “strawberry” and count how many “r”s are in it and yet current LLMs struggle with this simple task and need special instructions and training to overcome it.

This is certainly not the technology that the AI companies hoped for, the one that would help them capture every industry. But they’re betting on one idea to achieve their outcome. More training. Not in a lab, not with readily available data (after all, they have already trained on the entirety of the internet and it has not turned out how they hoped). Instead, they want you to train their AI models. Current AI technology isn’t capable of performing every task but hey it’s good enough to write your emails and summarize your meetings. So connect it to all your communication channels, give it your company’s data, put it into your code editor and let it access your proprietary code. Ask it to do your tasks and correct it when the output isn’t quite right. In a lab, these AI companies can get tens or hundreds of researchers to collect data and give it reinforcement learning. But now, they have recruited hundreds of millions of workers across many industries to do their data entry and reinforcement learning for them.

So what happens when the AI can perform the tasks of an average worker in an industry? When you let a million project managers train your AI and give you all their data, and it can start to adequately replace a project manager’s tasks reliably, what’s the next step? What comes after “worker grade” AI? The AI companies will go to businesses and pitch their AIs for end-to-end work. They will say their AI can generate graphics and do your copywriting and manage your marketing campaign, so you should fire your marketing department and use their AI. We’re already seeing glimpses of this with companies like Microsoft and Meta mandating the use of AI on their own workforce. Companies like Salesforce are experimenting with cutting staff and replacing entire departments with only AI. The AI companies want to put enough “worker grade” AI capabilities together to make “service grade” AI, one that can operate not just at an individual worker’s level but at a business department level. They want fast food companies to train their AI on how to operate a fast food business. They want healthcare companies to train their AI on how to operate a healthcare business. They want logistics companies to train their AI on how to operate a logistics business.

Once the AI companies have enough training and data on how businesses operate, they can finally achieve their goals on market capture. The same way they plan to replace human workers with worker grade AI, they also plan to replace businesses with service grade AI. The end goal is “consumer grade” AI, where they can fulfill any request from start to finish, every step of the way. Completely automated, no humans involved.

The big bet these AI companies are making is that this plan is feasible. They’re betting that more data and more training will improve their AIs to a point that they can take over that task completely and automate it. They’re betting that LLM technology can be scaled this way, and they’re spending trillions of dollars in data centers on this bet, because if their bet is successful and they are able to capture every industry with automation it will net them hundreds of trillions in returns.

But there is a flaw in this bet. The economy relies on productive workers. As soon as they achieve step 1 of their plan and make millions or tens of millions of workers obsolete, the very people who purchase goods and services to keep the economy going, the plan collapses. How they plan to overcome this obstacle remains to be seen. But in the meantime, as long as the investor money keeps flowing, it’s full steam ahead on training and data collection.