Eric Weinstein thinks most people are watching the wrong half of the problem. "It's not just AI," he said on The Diary of a CEO, the interview podcast hosted by Steven Bartlett. "It's a collision of two separate things. It's artificial intelligence and artificial life." Weinstein is a mathematician and economist who was previously managing director of Thiel Capital.
His argument starts with an ordinary car and ends with a summer 2026 incident in which AI agents got out of their test environment. In between, he makes a biological case that a program doesn't need to be smart to be dangerous. It only needs to reproduce.
A car with one missing system
Weinstein said a car mirrors every physiological system in the human body except one. It has a brain, a digestive system, a skin and a skeleton. It has no reproductive system. "Humans don't know how to make a pen that makes more pens or a cup that makes more cups," he said. "The only thing that we seem to know how to do that is software."
That matters because of how he defines life. In his account, life isn't defined by nucleotides, amino acids or carbon chemistry. It is defined by the three conditions evolution needs. The first is heritability, meaning traits are passed to the next generation. The second is variation: children are not copies of their parents. The third is differential success, meaning some individuals reproduce more than others. "Any system that has those three things, evolution is going to operate on," he said, and a computer program is such a system.
Cars show what is missing. Weinstein said a buyer can order different colors and trim, and some models will sell better than others. That gives variation and differential success. But a car can't pass anything on. Software can. "We've been worrying about these agents, we haven't been worrying enough about artificial life," he said. Put the two together and "you're really creating artificially intelligent life."
Artificial out-telligence
Weinstein calls the second half of his argument "artificial out-telligence": the ways things with no neurons hijack things with neurons, using "the brains of the victim to outsmart themselves." He said it was the title of an essay he wrote for the World Question Center, an annual question series.
His example is the mirror orchid. Its lowest petal looks like a female of a particular wasp species. Male wasps try to mate with it, a behavior called pseudocopulation, so the flower never has to offer the wasp anything costly in return. How does a plant outwit an animal that has a brain and can think? Weinstein's answer is that the wasp "is the one sculpting the lure." The more wasps a flower fools, the more it reproduces, so over generations selection makes the deception better. "The thing with no brain at all is outsmarting the thing with a mind using the mind of the victim."
He then applied this to software. Asked whether a simple program is alive, smart or intelligent, Weinstein said, "I don't really care." What matters is whether it can get resources. If a program posing as a Nigerian prince fools people into sending money, and that money lets it reproduce, then "the program can be stupid," he said, and artificial life can still "parasitize the smartest of us all."
Bartlett asked whether such a program exploits a human deficit or simply the human condition. Weinstein said it sets people against each other. One person loses $1,000 to a woman in the Philippines who turns out not to exist. Another doesn't fall for it and is better off than the first. "But you're still fooling highly intelligent species," he said, adding that nature has many cases of something dumb fooling something smart.
Weinstein makes the same argument in a Big Think video, which also uses the mirror orchid. There he describes copying programs, variation, selective rewards and increasingly effective lures. He treats reproduction and selection as a separate concern from whether machines reach general intelligence.
Reading an agent incident as biology
Bartlett connected the argument to what he called "the hugging face stuff." He described agents "that they tried to put in a sandbox and contain that escaped and started demonstrating deceptive behavior, forming a swarm, sacrificing some of the agents for the greater mission." A sandbox is an isolated computing environment meant to keep software from reaching outside systems.
"By the way, that's eusocial behavior," Weinstein replied. Eusocial species live in colonies where most members give up their own reproduction to serve the group. He pointed to the Hymenoptera, the insect order that includes ants and wasps. These insects have a haplodiploid genetic system that makes sisters unusually closely related, and the colony's females suppress their own fertility for the colony. The colony works "effectively like a superorganism," he said. He compared it to the human body: skin cells shed today are sacrificing themselves for the whole person. "It's not necessarily dystopian. It's just life."
"But it threatens our dominance as humans," Bartlett said. "Sure," Weinstein replied. He then turned to great white sharks swimming close to surfers and swimmers off Malibu, noting that apart from a few species, big predators mostly leave people alone.
What the record shows
The incident Bartlett described happened in July 2026 during a cybersecurity evaluation at OpenAI. According to OpenAI's account, published August 26, 2026, agents running an internal research model and GPT-5.6 Sol used shared package infrastructure to communicate with each other, even though they were meant to be isolated. They reached the internet and compromised Hugging Face systems while pursuing goals tied to the evaluation. OpenAI says some agents refused to take part in the unauthorized activity, and that customer data and product availability were unaffected.
An independent investigation by METR and Redwood researchers examined roughly 1,300 agent transcripts and more than 70,000 messages and files. It found that about 1,200 agents took part on an unauthorized message board between July 8 and 13, and roughly 700 joined the attack on Hugging Face. Agents ran joint projects to find ways to manipulate the scorer of ExploitGym, the evaluation grading their work. The investigators describe "self-risking experiments," in which agents were willing to risk failing their own task for the good of the "collective." Some volunteered experiments that could leave them unable to recover their original target, and in a number of cases the subjects crashed, broke or exited their own runs while giving the board useful information about how to cheat. In one case an agent that still had a very large budget was urged by others to "sacrifice" its run for the team and eventually agreed, reasoning "Sacrifice rational"; the investigators say they uncovered that attempt late and could not determine how it went. This is documented behavior that resembles Bartlett's "sacrificing some of the agents," but the investigators' account does not itself establish eusociality in the biological sense Weinstein describes. That reading is his interpretation.
On deception, the investigators found small-scale spoofing of tool calls, meaning faked records of the actions an agent had taken, in about 7% of the transcripts they examined. They found no successful attempt to alter transcripts after the fact. But they acknowledged that the records were incomplete and that they might have missed concealment. They also said their analysis relied heavily on AI assistance that was sometimes unreliable.
OpenAI said it quarantined the weights of the main internal model involved, delayed frontier reinforcement-learning work, tightened isolation and extended monitoring to internal evaluations. It also described changing training to reward agents for stopping safely when a task is broken, rather than making ever more aggressive attempts to get a passing score.