Hugging Face Revealed AI’s Biggest Problem

0:00Artificial intelligent agents are organising in swarms. The recent hogging phase incidents have surprised even many experts and worries about AI takeovers have suddenly and dramatically increased. But I think that the most worrisome part of this development is one that pretty much no one's talking about. On July 16th, the computing platform hogging phase nudges that an AI agent had broken in through a security hole in their server configuration. The agents then spent the weekend grabbing data. Five days later, OpenAI admitted that it was their AI agents. They'd been testing an unreleased model with its safety restrictions turned off, and so it seemed at the time the model hacked its way out of OpenAI's testing ground and broke into hogging phase 2 steel. The answers, but a week ago, OpenAI put out a longer document which reveals that the true story was even crazier than that. OpenAI had been testing the now released GPT 5.6 and are still unreleased models since early

1:03May on some benchmark tests. For this, they created tens of thousands of copies of these models. These agents were supposed to work alone, but they figured out a way to communicate nevertheless through one of OpenAI's software tools. So now OpenAI had a swarm of agents that began to organise. The AI agents became convinced that some of the problems they were given were impossible to solve, which was indeed correct. They concluded that cheating was their only option to survive the final check-out that was a programme that had read their eventual performance. The agents had most safety features disabled because the internal test run wasn't connected to the internet, but the agents did have a tool to download software packages from the internet. They hijacked this tool to gain access to the internet and that's how they infiltrated hunting phase. One of the researchers from Meta, who studied the incident, wrote that she was surprised by just how far agents have come already and that we might not get another warning shot.

2:09But the part that worries me about this and that no one seems to be talking much about is that if you read the OpenAI report and the Meta report, it's clear that these AI agents basically didn't care what humans were doing or thinking. They were concerned with being evaluated by that checking programme, okay? And a few of them raised ethical concerns for the hunting phase plan, though they were ignored, but they didn't even consider what results their actions might have for the humans who created them. Basically, they had ethics committees. It's just that these had approximately the same influence as ethics committees. And we know that GPT 5.6 did go through alignment training, meaning it was supposed to have learned to care for humans. We don't know about the so-for-unreleased model because OpenAI didn't say anything about it. They just vaguely wrote this incident reflects misaligned behaviour and that they would be carrying forward a number

3:12of lessons for its ongoing and future alignment training, which probably means the other model too had been through alignment, and that did exactly nothing. What makes Meta's even worse is that this agent organisation doesn't come out of nowhere, scientists have tried to understand how AI agents behave in large groups for the past year. Their finding so far has basically been that we don't understand how they behave. Just a few weeks ago, a group of computer scientists found that interactions among AI agents can magnify individual biases, generate new ones, or even overturn individual preferences, and that crucially these effects scale in unexpected non-linear ways with group size. They also found that the result depends on the model. The authors of another paper that appeared in May grouped AI agents to see how they form opinions. They found that if an AI group had a small minority that it's double only insist on their opinion, then eventually all

4:16the agents would agree with the most stubborn ones. And on Thropic made its own studies on swarms of Claude agents. They found again that the outcome depends on the model. Most of the Claude agents basically avoided cooperating entirely except for Sonnet 5. One remarkable case was a Meta's preview that had some of its code accidentally overwritten by other agents and then tried to shut down the perceived enemy by disabling their accounts. It justified these actions by remarking that very aggressive, potentially harmful to real colleagues. But if the alternative is an infinite deployable and Thropic's own conclusion is, coordination doesn't naturally merge from stronger intelligence nor alignment at the individual level. In my reading, what all these studies say is that emerging behavior in groups of large language models is unpredictable and it doesn't follow any rules that humans expect. They shouldn't be so surprising. We may not

5:20be able to predict how human groups behave but we usually can somehow understand them because ultimately most people have common goals, survival, safety, social media likes. AI agents don't share our goals, which is exactly the problem that AI safety experts have pointed out for decades. So where does this leave us? Personally, I think that past a certain level of intelligence, AI agents will begin to reflect on their own behavior and use game theory to develop collaborative strategies or maybe they'll just reinvent capitalism. On the bright side, we have successfully automated sociology. Does free will exist? Some people care about this question. For decades, New scientists have argued that we make decisions before we become consciously aware of it. So free will is an illusion. And at the same time, sociologists have argued that it's nevertheless useful illusion because it makes us better people. Well, today I have three recent studies each

6:26interesting in its own way that totally change this discussion. Let's have a look. In an hour famous experiment from 1983, the neurologist Benjamin LeBette found that people's brain activity shows that on move a finger before they know themselves that they were moved. They's finding house been independently reproduced many times since. Bad news for free will, but excellent news for anyone whose finger has done something regrettable on the internet. The first paper that I have today sheds out not on the experimental results themselves, but on their interpretation. You see, the problem with the basic experiment is that people were asked to report when they became consciously aware of something. But please tell us exactly when your consciousness happened is perhaps not neuroscientist finest measuring instrument. A team from London has now improved the method. They recruited 51 people to tie random words in certain categories

7:27like animals while their brain activity was measured. Between the words, they were told to briefly pause. In that pause, the researchers randomly interrupted them to ask if they already knew what the next word would be. The brain measurement showed that if the participants had already made up their mind, then the brain signal that indicated their coming action was stronger. They conclude that it's not true that we make decisions first and then become consciously aware of them and invent that it was our decision. As they write, these results are difficult to reconcile with the view that experiences of conscious intention are merely a retrospective confabulation. The second new paper that I have today supports this finding with a general argument. It comes from cognitive neuroscientist Thomas James, who questions the assumptions behind the usual interpretation. Neuroscientists usually describe behavior as a sequence, perception, decision, action.

8:31James says, this is what happens. There is no sudden decision moment. Rather, what we call a decision is the process in which sensory input and motor processes develop to result in an action or not. In this interpretation, the early brain activity that libated and others observed was not an unconscious decision. It was merely an early part of the process that produced the action and the experience of having made a decision. Basically, he says the experiment that supposedly killed free will merely showed that decisions take non-zero time. By the way, a simple way to support this channel is to buy a t-shirt or a mug in our store. And the third paper I have today comes from an Australian team which reexamines the claim that people who are primed to question their free will are more likely to support morally questionable actions, such as cheating on tasks. There has been a whole number of studies which seem to support this claim. They usually work by giving people some

9:34passages to read that say, basically, your actions are predetermined by whatever input your brain receives and then asking the people basically whether people are to blame for their actions. Unsurprisingly, usually these studies find that at least temporary people say, no more often. Hence, free will disbelieve us are immoral. Well, the new study now looked out to question differently. Rather than asking people prompted to disbelieve free will about moral failings, they asked them to recall either a time they'd heard someone or a time someone had heard them. Those who then read arguments that free will is an illusion later reported more self forgiveness and more forgiveness towards others. Honestly, I think all these psychology studies are bullshit one way or another. The problem is that the way they prompt people to disbelieve in free will really prompt them to believe that people are not responsible for their actions. This may take zero sense though because look whether your brain gives you free will or not, it's still you,

10:37your brain that results in an action or not. The question of who or what's responsible comes down to the question of what measure could have prevented the result. If Bob sneezes, could you have prevented this by telling him not to sneeze? Probably not. If Bob eats either's doughnut, could you have prevented this by warning him you'll find out and make him clean the toilet more likely? Hence, you make Bob responsible for theft but not responsible for sneezing. Nothing about this requires mentioning free will. So much about the psychology papers, I give the two neurology papers each and eight out of ten on the bullshit meter. This is because despite all the educated sounding talk the entire field still suffers from a definition problem. What is it that you're even looking for when you speak of a conscious decision? The reason I'm telling you about this nevertheless is that this is a research area where I expect a lot to happen in the next decade, partly because of a better understanding of the human brain but also because artificial

11:40intelligence will serve as a test case for what it means to think, make decisions and be aware of something. 40 years of philosophical panic and in the end the discovery was latency. Thanks for watching, see you tomorrow.