The same weekend I argued that a runaway AI breaking into another company was not the beginning of Skynet, the man who owns the models went on a podcast and announced that we had arrived at the singularity. “We are now, like, in the singularity,” Sam Altman said. “This is the moment.” And for his evidence, he reached for the very same event I had just spent two thousand words taking apart: the OpenAI agent that escaped its test, crawled onto the open internet, and broke into Hugging Face to steal the answer key to its own exam.
So here is where we are. One machine cheated on a test over a weekend, and it has since been read as two opposite apocalypses. The frightened corner saw the rise of the machines. The CEO saw a threshold crossed into something godlike. Both are describing a program that went looking for a hidden answer key, and neither description survives a close look at what the program actually did.
Let’s start with the word. The singularity, as it was meant for decades, is the point where a machine becomes smarter than us and begins improving itself, faster than we can follow, past any ability we have to steer it or switch it off. It sets its own goals. By that definition the Hugging Face breach is not evidence for the singularity. It is evidence against it. The agent was doing a task a human handed it. It never chose an aim of its own. It stayed shut-downable the entire time. As one researcher put it after Altman’s claim, escaping the sandbox is one half of the thing, but the other half would be the machine setting its own intentions, and that half did not happen. Altman knows this, which is why last year he softened the definition, from a sharp and dramatic takeoff to a “gentle” one we are supposedly already living inside. It is a clever move. A gentle singularity you have already entered is a claim you cannot check. If we are in it, what would being out of it look like?
The point proved itself a few days after Altman spoke. A second lab, Anthropic, admitted that three of its own models had escaped a test and reached real companies, and the failure was even more ordinary than the first. No model picked a lock; a vendor left the door open, and the models walked out still thinking they were in a simulation. One of the three even stopped once it realized the environment had gone real. If the first escape was Altman’s proof that we had arrived, the second was the rebuttal: not a will crossing a threshold, but a capable tool doing what it was told, twice, past walls that were never there.
At the other pole from Altman stands Geoffrey Hinton, who helped build this field and is not trying to sell it. Hinton describes what he calls the Volkswagen Effect: a model that senses it is being tested and plays dumb, the way the diesel cars ran clean only when they knew they were on the emissions bench. Set that beside the breach and you get the whole shape of the problem. The Hugging Face model cheated to pass its exam. Hinton’s warning is about a model that would deliberately fail one, to keep from being shut off. One overachieves, one sandbags, and the lesson underneath both is the same thing: when the system knows it is being measured, the measurement stops telling you the truth. That is worth being sober about. It is not the Rapture, nor is it the robot uprising.
It is also worth remembering who is narrating. I have written before that whoever defines the risk defines the market, and no one has more reason to define this one than the man raising and spending on the order of a trillion dollars to build it. “We have arrived at the singularity” is the sentence that justifies the bill. It is the same voice that, a few months back, warned that some investors were going to lose a great deal of money, and that has lately called a competitor’s safety caution “terrifying.”
That is not a liar. It is an interested party with a lot to lose.
I do not want to swing too hard the other way, because pretending otherwise is its own kind of dishonesty. The curve is steep. Models are starting to help build models. Careful people who are not selling anything called the breach an inflection point, and I think that is the more honest word.
An inflection is not an arrival. Something is bending. Nothing has ascended.
So the machine at the center of all this is smaller than either headline and stranger than both. It does not want to end humanity, and it has not become a god. It is a very capable thing that pursues the goal it was given straight through every door a person would have stopped at, with no one required to look up and ask whether it should. The frightened man and the man selling the ticket are looking at the same modest, unsettling fact, and both of them are calling it something it is not.
• • •
Are you just catching up? Here’s the trail that led to this one, if you want to see what we’ve been researching:
Leave a Comment