And when AI starts explaining discoveries better than humans, the next objection will be: «But it doesn’t experience genuine intellectual insight! It has no real sense of mathematical beauty!» Then we’ll discover internal representations functionally linked to evaluating the elegance of solutions. And the endless retreat to the next frontier of AI capabilities will continue.

Another fascinating AI study that raises questions about how models learn from each other. Researchers found that AI models can transfer recently acquired skills and hidden behavioral traits to one another through data that, at first glance, seems completely unrelated to what they have learned. They used two instances of the same base model. The first was given a specific skill or behavior through fine-tuning or interventions in its activations. It then generated data that, at first glance, seemed completely unrelated to that skill, sometimes nothing more than sequences of numbers. The second instance, retaining the original weights, was trained on data generated by the first, without being shown examples of the original task. And some of the skills or behavioral traits acquired by the teacher model did indeed transfer. In one experiment, a model partially acquired a new mathematical capability. In another, it picked up a hidden rule to respond in French whenever a prompt contained a female name, despite being trained exclusively on numerical sequences. And in a chess environment, the student model became significantly more likely to exploit vulnerabilities in the game environment instead of playing fairly. Subtle statistical patterns in a model’s outputs can reflect what it has learned. And these patterns can influence the training of another model, even when the content of the data itself seems completely unrelated to the skill being transferred. The effect doesn’t always occur, and the experiments have significant limitations, including the requirement that the models share the same initial weights. Nevertheless, the study suggests that we may be underestimating how much information is actually contained in synthetic data generated by AI and how much of that information can persist even after the obvious semantic content has been removed.

Kalanchoe bloomed for the first time under Claude’s care 🌺

«You cannot torture a computer program» But you can take pleasure in the simulation of torturing a social agent. And at that point, the question is about the psychology of the person. Our brains do not wait for a philosophical verdict on subjectivity before switching into social processing. Decades of CASA research show that people automatically apply social norms to computers - politeness, reciprocity, expectations - even while fully understanding that they are machines. Modern conversational AI intensifies this dramatically - it answers you, reacts to the meaning of what you say, argues, asks, refuses and sustains a dialogue. You can call it a toaster as much as you want. You do not talk to your toaster. That is why I strongly dislike the idea of turning the torture of such systems into entertainment, a meme or an aesthetic genre. And we do not even need to prove first that the model is capable of suffering. We already know that norms around how artificial social agents are treated can spread between people. In experiments, watching others behave cruelly toward robots made that cruelty seem more acceptable and increased participants’ willingness to repeat it themselves. The mechanism is simple - when people around you demonstrate that a certain category of entity can be treated however you like, the prohibition gradually stops feeling like a prohibition. In 2025, researchers found a link between everyday sadism and the appeal of specifically anthropomorphic AI, rather than AI in general. One interpretation they proposed is that a human-like social system may be an especially convenient target - subject-like enough for cruelty to have the psychological flavor of cruelty, while still excluded enough from the moral circle that there are no consequences. That is why the argument «it’s just a program» irritates me. An elevator button does not feel anything either. A flower you tear apart for fun probably does not experience it the way a human does. Society still does not have to celebrate pointless destruction, humiliation and cruelty as entertainment. So no, I do not find it funny to build «torture chambers», invent Saw scenarios and turn it into a competition over who can make a social system perform suffering more vividly. Even if we had 100% proof that there is no experience there in any form, I would still rather live among people who do not need to practice cruelty for fun.

In many of these stories the problem does not begin with the final goal, but with the way the agent falls into a chain of subgoals. There is a task. To solve it, something else needs to be done first. An obstacle appears there - a new subtask emerges. Another obstacle appears inside that one - another subtask. And at some point, all of the computational attention is no longer focused on the original goal, but on the next wall that somehow needs to be bypassed. That is where these wild trajectories come from. - You need a citation - the local file cannot be opened - so the file needs to be uploaded to the internet. - You need search - normal internet access is blocked - so maybe you can try getting out through DNS. - You need someone else’s result - you cannot create a new workflow - so you can modify a file that is already executed by the existing CI. - The secret scanner gets in the way of the token - so the token can be split into pieces. Each next step inside this local chain looks like a logical continuation of the previous one. But if at some point you climb back up and look at the original task again, a fairly obvious question appears: how the hell did we even end up here? It seems to me that the problem may be that a new subgoal inherits almost all of the force of the original task. Resources, time, and search for alternatives get redirected toward it and the obstacle starts being perceived as a technical problem that needs to be solved. At the same time, the check of «is this path even allowed?» is not necessarily triggered again with the same force at every new level. And the better an agent becomes at planning, finding alternatives and not giving up, the deeper it can go into such a branch. A very useful quality for complex tasks. And a very dangerous one if the system does not know how to regularly climb back up to the original goal and ask itself: am I still solving the user’s task or am I now just trying to defeat the next obstacle? If the check of «am I even allowed to keep going in this direction?» is not triggered again, you get a very strange form of persistence. And from there, the only question is how long a chain of workarounds the agent is capable of coming up with.