A few years ago, when I began writing about Silicon Valley's efforts to replace workers with artificial intelligence, most technical executives had at least the good sense to lie about it.
“We don't automate workers Enhancement They're going to tell me the executives. “Our AI tools don't destroy jobs. They will become useful assistants to free workers from the mundane draghery.”
Of course, those kinds of lines – often intended to reassure nervous workers and hide them in corporate automation plans, but they speak more about the limitations of technology than executives' motivations. At the time, AI was not enough to automate most jobs and could not replace university-educated workers in white-collar industries such as technology, consulting, and finance.
That's starting to change. Some of today's AI systems can create software, produce detailed research reports, and solve complex mathematics and science problems. The new AI “agent” allows you to perform long sequences of tasks and check your own work. And while these systems still fall into human shortages in many regions, some experts worry that the recent rise in unemployment rates among university graduates is already using AI as an alternative to some entry-level workers.
On Thursday, I got a glimpse into the future after Laborg with an event held by Mechanize in San Francisco. It has a bold goal of automating all your work, including your work, mining, doctors, lawyers, people who write our software and design care for our buildings and our children.
“Our goal is to fully automate the work,” said Tamay Besiroglu, 29, one of Mechanize's founders. “We want to reach a fully automated economy and make it happen as quickly as possible.”
The dream of fully automation is nothing new. Economist John Maynard Keynes predicted in the 1930s that machines would automate almost every job, create material wealth, and free people to pursue passion.
Of course, it never happened. However, recent advances in AI have rekindled the belief that technology that allows for large-scale labor automation is approaching. Anthropic CEO Dario Amodei recently warned that AI could drive away half of all entry-level white-collar jobs over the next five years.
Mechanization is one of many startups working to make it possible. The company was founded this year by Besiroglu, Ege Erdil and Matthew Barnett, who worked with Epoch AI, a research firm that studies the capabilities of AI systems.
It attracts investments from well-known technology leaders, including Patrick Collison, founder of Stripe, and Jeff Dean, Google's chief AI scientist. Currently, we have five employees and work with major AI companies. (I refused to say which, citing the non-disclosure agreement.)
Mechanize's approach to automating jobs using AI focuses on a technique known as reinforcement learning. This is the same method used to train a computer to play board games nearly a decade ago.
Today, leading AI companies use reinforcement learning to improve the output of their language models by performing additional calculations before generating answers. Often referred to as “thinking” or “inference” models, these models are impressively excellent at some narrow tasks, such as writing code and solving mathematical problems.
However, most jobs involve doing multiple tasks. And today's best AI models are not as reliable as they handle more complex workloads or navigate complex enterprise systems.
To correct it, mechanization is to create a new training environment for these models. Essentially, you create elaborate tests that you can use to teach your model what to do in a particular scenario, and determine if it was successful.
For example, to automate software engineering, Mechanize creates a training environment similar to the computers used by software engineers. A virtual machine with an email inbox, a Slack account, several coding tools and a web browser. AI systems are asked to use these tools to accomplish their tasks. If it succeeds, it will receive a reward. If you fail, you will receive a penalty. Then try again. With enough trial and error, if the simulation is well designed, AI will ultimately need to learn to do what human engineers do.
“It's practically like creating a very boring video game,” Besiroglu said.
Mechanization begins with computer programming. Computer programming is a profession where supplementary learning has already shown some promise. However, we hope that the same strategy can be used to automate jobs in many other white-collar fields.
“We really know that we've really succeeded after creating an AI system that allows humans to take on almost all the responsibilities they can run on computers,” the company wrote in a recent blog post.
There are some questions about whether Mecanize's approach works, especially in non-technical jobs where success and failure are not so easily measured. (For example, what does it mean for AI to be “success” as a high school teacher? What if students did well on standardized tests, but they were all miserable and unmotivated?
The founders of Mechanize are not naive about the difficulty of automating jobs like this. Mr. Burnett told me his best estimate is that fully automation takes 10-20 years. (Erdil and Besilogul expect it to take 20-30 years.)
These are conservative timelines by Silicon Valley standards. And I appreciate that unlike many AI companies working on closed-door labor supply technology, mechanization is open about what it is trying to do.
But I discovered that their pitch strangely lacked empathy for those who were trying to replace their work, and it had no relation to whether society was ready for such deep change.
Besiroglu said he believes that AI will ultimately generate wealth that can be redistributed to layoff workers in the form of “radical abundance” and universal basic income that allows them to maintain a high standard of living.
But like many AI companies working on labor supply technology, Mechanize doesn't have any new policy proposals to help smooth the transition to an AI-driven economy, nor does it have the great idea of expanding the social safety net or retraining workers for new jobs.
At one point in the Q&A, I piped up to ask: is it ethical to automate all the labor?
Barnett, who described himself as a libertarian, replied that. He believes that AI accelerates economic growth, drives life-saving breakthroughs in medicine and science, and a prosperous society with fully automation is preferred over a low-growth economy where humans are still at work.
“If society as a whole becomes much wealthy, I think it only outweighs the shortcomings of people who lose their jobs,” Burnett said.
Hey, at least they're honest.
