A machine without a Sabbath
Every Friday afternoon in Jerusalem, in hotels and apartment blocks, someone flips a switch and an elevator begins to run by itself. The Shabbat elevator stops at every floor, opens its doors, waits, closes them and moves on. Nobody presses a button. The machine keeps working so that people can stop directing it.
I have been thinking about that elevator all week, because some of the most careful people in artificial intelligence have just told us that the machines building the next machines need fewer hands on the buttons than anyone expected.
On November 10, Anthropic told the US Securities and Exchange Commission, in an amended prospectus, that it has met the AI research and development threshold in its own Responsible Scaling Policy: the tripwire for the point at which AI substantially accelerates the building of AI. The call rests on what it expects from future models and its overall rate of progress, not on any model now in use. The evidence came from METR, an independent safety group whose staff spent eleven weeks inside the company between August and October. It found benchmark progress running at about 1.9 times its earlier pace, two years of advance in one, and guessed that perhaps 80 percent of the acceleration came from the company’s own AI. Much of the detail was redacted. We are being shown the conclusion more than the proof.
The skeptics have moved. The authors of the AI 2027 scenario now say the intelligence explosion will happen in 2027 after all, and name the METR report as their biggest reason. The podcaster Dwarkesh Patel, who earlier this year would have put superintelligence the better part of a decade away, now puts his median for AI matching the best human AI researchers at 2028. Greg Casar, chair of the Congressional Progressive Caucus, put it in two words: “Ban RSI”. The letters stand for recursive self-improvement, a machine improving the very process by which it improves.
The case for alarm is arithmetic. A machine researcher does not tire, does not go home and does not keep a day of rest. Run thousands of copies around the clock and the human week stops being the unit of progress.
Here the Jewish tradition has something Silicon Valley lacks. Ahad Ha’am wrote that more than the Jewish people has kept the Sabbath, the Sabbath has kept the Jewish people. The Sabbath is the oldest working brake in human civilization: a rule that the labor stops whether or not the labor is finished. Its genius is that it binds everyone at once. A Sabbath kept by one shopkeeper is a commercial handicap. A Sabbath kept by a whole society is a civilization.
That is precisely the problem the AI industry cannot solve alone. In September, Anthropic’s chief executive, Dario Amodei, wrote that “we must slow the pace at which we improve the capabilities of AI models.” Six weeks later, reporters were finding few signs that the companies were easing off. Every voluntary pause meets the same objection: whoever paces first hands ground to whoever does not. A brake that only one driver uses is not a brake. It is a gift to the car behind.
The rabbis also thought harder than almost anyone about machines that act while people abstain. Halakha distinguishes between performing a forbidden act directly and causing it indirectly, the category known as grama. The Shabbat elevator rests on a related idea: a machine set running before sunset is treated differently from a finger on a button after it. Yet the tradition never pretended that setting the machine running was morally weightless. Someone decided the elevator would run. Someone answers for it.
That is the closed-loop question AI will put to every government. When the next model is built largely by the last one, who pressed the button? The labs will be tempted to answer that nobody did, that the loop simply ran. Grama at civilizational scale is still a human decision.
Israel should care more than most, and this month showed why. On November 9, at Tel Aviv University’s AI Week, Erez Askal, the former Unit 9900 commander who heads the new National AI Directorate in the Prime Minister’s Office, said Israel’s national infrastructure will run on AI within a few years. The directorate’s August plan sets a target of at least 100,000 advanced accelerators and proposes a National Ethics Council. An ethics council advises. It does not sign. Israel still has no comprehensive law requiring human oversight of AI systems.
Yet the Knesset already knows how to write the clause. The National Cyber Defense Bill, which passed its first reading in June, assigns personal supervisory responsibility to corporate officers. Apply the same principle to any AI system that helps build its successor or runs national infrastructure: a named officer signs off on each training run and each deployment, and answers for it personally.
Jerusalem cannot slow San Francisco. It can make sure that in Israel someone always presses the button, and argue abroad for what the Sabbath has always demanded: a pause shared by everyone, fixed in advance, so that no one profits from working through it.
Tonight, somewhere in Jerusalem, an elevator will open its doors onto an empty corridor, wait, and move on. It was built to serve people who had chosen to rest. The machines now being built in California were not. The difference is not in the engineering. It is in whether anyone still decides when the work stops.
