Years ago, when I began discussing Silicon Valley’s push to replace human workers with artificial intelligence, tech leaders at least pretended to care.
“We’re not removing jobs, we’re enhancing them,” executives claimed. “Our A.I. tools will not take jobs away; instead, they will serve as helpful assistants that liberate employees from tedious tasks.”
Such statements were usually designed to calm anxious workers while masking corporate plans for automation. At that time, A.I. lacked the capability to replace most positions, especially those in white-collar fields like tech, consulting, and finance.
However, this is beginning to shift. Modern A.I. systems can now write software, generate in-depth research papers, and tackle complex math and science challenges. Newer A.I. “agents” can perform extended sequences of tasks and verify their own outputs, similar to human behavior. Experts are concerned that a recent rise in unemployment among college graduates is indicative of companies starting to employ A.I. as a replacement for some entry-level positions.
Recently, I witnessed a vision of a future without jobs at a San Francisco event hosted by Mechanize, a new A.I. startup ambitiously aiming to automate every job—yours, mine, and those of professionals like doctors, lawyers, software developers, and caregivers.
“Our ambition is to fully automate work,” stated Tamay Besiroglu, 29, one of Mechanize’s founders. “We want to create a fully automated economy as quickly as possible.”
The idea of complete automation is not new. Economist John Maynard Keynes predicted back in the 1930s that machines would take over nearly all jobs, leading to material abundance and freedom for people to pursue their interests.
That hasn’t materialized, but advancements in A.I. have rekindled the belief that technology capable of mass job automation is on the horizon. Dario Amodei, CEO of Anthropic, recently cautioned that A.I. could eliminate half of all entry-level white-collar positions within the next five years.
Mechanize is among several startups striving to achieve this goal. Founded this year by Besiroglu, Ege Erdil, and Matthew Barnett, who collaborated at Epoch AI, a research company focused on A.I. capabilities, they aim to make automation a reality.
They’ve gained backing from prominent tech figures like Patrick Collison, co-founder of Stripe, and Jeff Dean, Google’s chief A.I. scientist. Currently, the company has five employees and is collaborating with leading A.I. firms but refrains from disclosing their identities due to confidentiality agreements.
Mechanize’s strategy for automating jobs revolves around a method called reinforcement learning—a technique that enabled computers to master the board game Go at superhuman levels nearly a decade ago.
Today, top A.I. organizations utilize reinforcement learning to enhance their language models, performing extra computations before producing answers. These often referred to as “thinking” or “reasoning” models, have excelled in certain narrow tasks, like coding or solving math problems.
Nonetheless, most jobs entail more than one task, and current A.I. models still struggle to manage more intricate responsibilities or navigate complex corporate environments.
To address this, Mechanize is developing new training scenarios for these models—essentially creating detailed tests to teach them how to act in specific situations and evaluate their performance.
For automating software engineering, Mechanize is constructing a training environment mimicking a software engineer’s computer—complete with a virtual machine, email inbox, Slack account, coding tools, and a web browser. The A.I. will be assigned tasks using these tools, earning rewards for success and penalties for failures, with the aim of learning to perform like a human engineer through repeated attempts.
“It’s like creating a very tedious video game,” Besiroglu noted.
Mechanize is starting with software programming, an area where reinforcement learning has shown some initial successes. However, they envision applying the same method to automate various other white-collar occupations.
“We’ll truly know we’ve achieved success once we build A.I. systems capable of taking on nearly every computing responsibility a human can handle,” the company stated in a recent blog post.
I have my doubts about whether Mechanize’s approach will be successful, particularly for non-technical roles where outcomes are less clearly measurable. For instance, how would an A.I. determine “success” as a high school teacher? What if students excelled on standardized tests, while simultaneously feeling miserable and unmotivated? What if the A.I. learned to game the system by providing students with the right answers to improve their scores?
The founders of Mechanize are aware of the challenges involved in this method of job automation. Barnett believes full automation could take 10 to 20 years, while Erdil and Besiroglu predict 20 to 30 years.
These timelines are conservative by Silicon Valley measures. I appreciate that, unlike many other A.I. firms developing labor-replacing technologies in secrecy, Mechanize is transparent about its ambitions.
However, I found their pitch surprisingly lacking in empathy for the individuals whose roles they seek to replace and seemingly indifferent to whether society is prepared for such drastic changes.
Besiroglu expressed a belief that A.I. would eventually foster “radical abundance” and wealth that could be redistributed to displaced workers via universal basic income, allowing them to maintain a comfortable lifestyle.
Yet, like many A.I. businesses focused on labor-replacing technology, Mechanize lacks innovative policy solutions to address the transition to a future dominated by A.I., and they don’t propose ways to strengthen the social safety net or retrain workers for new opportunities—instead, they prioritize rendering existing jobs obsolete as quickly as possible.
During a Q&A session, I asked: Is it ethical to automate all jobs?
Barnett, identifying as a libertarian, argued that it is ethical. He believes A.I. will fuel economic growth and lead to essential breakthroughs in fields like medicine and science, asserting that a thriving society with full automation would be better than one where humans still have jobs.
“If society as a whole becomes significantly wealthier, then the benefits outweigh the drawbacks of people losing their jobs,” Barnett stated.
At least they’re being truthful.
