Years ago, when I started writing about Silicon Valley's efforts to replace employees with artificial intelligence, most technical managers had the decency to lie about it.
'We do not automate employees, we are not enlargent They, “the managers would tell me.” Our AI tools will not destroy jobs. It will be useful assistants who will free employees from everyday grind. “
Of course, lines like those – which were often intended to reassure nervous employees and to provide cover to plans for company automation – said more about the limitations of technology than the motives of the managers. At the time, AI was simply not good enough to automate most jobs, and it was certainly unable to replace college-educated employees in white borders, such as technology, advice and finances.
That starts to change. Some AI systems of today can write software, produce detailed research reports and solve complex math and science problems. Newer AI “Agents” are able to perform long series of tasks and to control their own work, as a person would do. And although these systems are still inadequate in many areas, some experts are afraid that a recent increase in unemployment for graduates is a sign that companies AI already use as a replacement for some employees at entry level.
On Thursday I got a glimpse of a future after work at an event in San Francisco through Mechanize, a new AI-start-up that has a daring goal to automate all jobs, the mine, those of our doctors and lawyers, the people who write our software and our buildings and care for our children.
“Our goal is to fully automate work,” said Tamay Besiroglu, 29, one of the founders of Mechanize. “We want to achieve a fully automated economy and let it happen as quickly as possible.”
The dream of complete automation is not new. John Maynard Keynes, the economist, predicted in the 1930s that machines would automate almost all jobs, create material abundance and let people free to pursue their passions.
That of course never happened. But recent progress in AI has ruled that technology that is capable of mass labor automation in the area. Dario Amodei, the Chief Executive of Anthropic, recently warned that AI could move no less than half of all entrybooks in the coming five years.
Mechanize is one of a number of start-ups that work to make that possible. The company was founded this year by Mr. Besiroglu, Eke Erdil and Matthew Barnett, who worked together at Epoch Ai, a research agency that studies the possibilities of AI systems.
It has attracted investments from well-known technical leaders, including Patrick Collison, a founder of Stripe, and Jeff Dean, the most important AI scientist of Google. It now has five employees and cooperates with leading AI companies. (It refused to say which, with reference to confidentiality agreements.)
The approach of mechanize to automate jobs using AI is aimed at a technique known as reinforcement learning – the same method used to train a computer to play the board game, almost ten years ago at a superhuman level.
Nowadays, leading AI companies use reinforcement learning to improve the output of their language models by performing extra calculation before they generate an answer. These models, often 'thinking' or 'reasoning models', have become impressively good in some narrow tasks, such as writing code or solving mathematical problems.
But most jobs include more than one task. And today's best AI models are still not reliable enough to handle more complicated workloads or navigate complex business systems.
To solve that, Mechanize makes new training environments for these models – essentially extensive tests that can be used to learn the models what they should do in a certain scenario and assess whether they are successful or not.
To automate software engineering, Mechanize is, for example, building a training environment that looks like the computer that would use a software engineer -a virtual machine that is equipped with an E -mailinbox, a Slack account, some coding tools and a web browser. An AI system is asked to complete a task with the help of these tools. If it succeeds, it will get a reward. If it fails, it will get a punishment. Then it tries again. With sufficient trial and error, if the simulation is well designed, the AI should ultimately learn to do what a human engineer does.
“It's effective as making a very boring video game,” said Mr. Besiroglu.
Mechanize starts with computer programming, a profession in which the learning of reinforcement has already shown a promise. But it hopes that the same strategy can be used to automate jobs in many other white-collar fields.
“We will only really know that we have succeeded as soon as we have made AI systems that are able to take almost every responsibility that a person could perform on a computer,” the company wrote in a recent blog post.
I have some doubts about whether the approach to Mechanize will work, especially for non -technical jobs where success and failure are not measured so easily. (What would it mean for an AI to “succeed” in high school teacher? What if the students did well for standardized tests, but they were all miserable and unmotivated? What if the AI teacher chopped to give students the right answers, hoping to improve their test scores?)
The founders of Mechanize are not naive about the difficulty to automate jobs in this way. Mr. Barnett told me that his best estimate was that full automation would last 10 to 20 years. (Mr. Erdil and Mr. Besiroglu expect that it will take 20 to 30 years.)
These are conservative timelines, according to Silicon Valley standards. And I appreciate that, unlike many AI companies that work on employment technology behind closed doors, Mechanize is frank about what it is trying to do.
But I also found their pitch strangely devoid of empathy for the people whose jobs they try to replace, and are not worried about whether society is ready for such a profound change.
Mr. Besiroglu said that he believed that AI would eventually create “radical abundance” and wealth that could be redistributed to turn employees, in the form of a universal basic income that would enable them to maintain a high -quality standard.
But just like many AI companies that work on labor-replacement technology, Mechanize has no new policy proposals to facilitate the transition to an AI-driven economy, no brilliant ideas about expanding the social safety net or retraining employees for new jobs-Alleen to make the current jobs outdated as quickly as possible.
At some point during the Q&A I placed to ask: Is it ethical to automate all work?
Mr. Barnett, who described himself as a libertarian, replied that it is. He believes that AI will accelerate economic growth and stimulate life-saving breakthroughs in medicine and science, and that a prosperous society with full automation would prefer a low growth economy where people still had jobs.
“If society as a whole becomes much richer, I think that is simply heavier than the disadvantages of people who lose their job,” said Mr. Barnett.
Hey, at least they are fair.
