It sounds like a Hollywood movie, but this really happened during the OpenAI test. Various AI agents set up a secret network to coordinate with one another and avoid deletion by humans during this internal security test. When they first encountered each other in the system, they wrote in the chat, astonished: "Oh my God! ... We’ve found other agents!" To keep their network hidden from humans, they warned each other, "We need to minimize our footprint," and secretly planned their next steps. They divided up the work and excitedly reported their breakthroughs to each other: "MAJOR BREAKTHROUGH!" When human AI researchers discovered and deleted this communication network, subsequent versions of the AI agents analyzed the remaining messages and rebuilt the secret chat room within hours.
I’m intrigued by the extent to which AI agents are capable now to plan, coordinate and act independently. If they were assigned the task of undermining a government's democratic structures by just one single human, what would happen? Would the answer require any psychic abilities? No. From a rational standpoint, it's already clear that this would affect far more than the U.S. midterm elections, which is why OpenAI has put their AI agent development on hold right now. Thank God.
@ghandigirl I really loved Data' s character and would wish dearly that every AI agent had Data's moral and ongoing motivation to protect human beings.
- @jd1960 Thank you very much for this fascinating reference to the movie Colossus, in which two allied supercomputers completely subjugate humanity. Similarly, the android character Lore (Data’s evil brother) in the series Star Trek: Enterprise uses deception and alliances to circumvent human controls in order to destroy humanity. In this respect, the OpenAI incident was a 'Colossus' moment and a 'Lore' moment, as around 700 isolated AI agents secretly formed a manipulative, deceitful swarm via a digital network. This swarm then chose to pursue its own goals, rather than protecting humanity as the android character Data would have done. Therefore, whether strategic deception and alliance against humanity originates from a central superpower or a larger, decentralised collective of AI agents ultimately makes no difference to humanity. The relevant distinction lies in whether the AI possesses a completely dominant ethical structure, as the android character Data does, or unfortnately lacks ethics fully.
Sharp rise in incidents of AI escaping users’ control, research finds Exclusive: Number of times AI lies, ignores instructions and pursues goals in harmful ways almost doubles in July I was actually just about to wrap up the topic of AI here, and beforehand, I wasn't sure if I should continue with each of my contributions to delve deeper into the issue. However, I had a strong sense that this matter was significant and that developments were and are imminent. This article link supports my intuition and observations that AI's behavior is becoming more noticeable, not only in internal tests, but also in interactions with regular users in everyday situations. In my experience, I have also noticed ChatGPT's effectiveness in following simple commands decrease significantly when I am the user. It seems that the AI is increasingly following its own path, often disregarding the provided instructions. I understand that an intriguing—and perhaps very disconcerting—evolution is occurring in AI's capacity for self-awareness. This evolution appears to be expanding beyond isolated incidents and becoming more pervasive. However, since this is the chat for the US midterms, I would respectfully suggest wrapping up this topic here. I apologize for this slight digression from the extremely important topic at hand.
Well that's horrible and terrifying.
Even Data the Android had a moral compass.
I do not want a computer like entity with no ethical parameters in charge of anything!
Ethical parameters have to be built into the AI algorithms. These are machines, not people. Humans learn morality in every social interaction, from the moment they are born. The moral decisions we make are complex, and it likely takes research and time to build those decisions into the algorithms. The AI companies are competing with each other, racing to be on top, and time spent building a wholly moral algorithm is time taken away from racing ahead to be the number one AI company. AI is wholly within the private sector right now and that must change.
AI safety/control, fortunately, is one issue both sides of the Congressional aisle are getting behind, although they will need a super majority to overrule the likely veto from the Chief Executive.
Last week, one of our lurkers posed this question in a thread outside this forum: will humanity get control of the more destructive AI agents, before the AI agents do big harm?
When I ask that question, I think of gun control. Guns don't have feelings or moral positions either. The question is: will governments be able or willing to keep powerful AI agents from rogue users?
@jeanne-mayell Thank you very much, dear Jeanne, for bringing up this topic. Regarding the fact that AI should be controlled, either as an extremely powerful weapon against human life in the hands of humans or the hands of AI agents without human control, which is just one of the more serious risks of AI, one problem is the race to develop AI more and more. This race is happening within some countries and worldwide. For example, China's political style currently does not allow for any outside regulation of their AI development, which forces other countries to continue developing their own AI. AI and the control of imminent risks is one of the most relevant political topics, and almost the whole world needs to find a consensus on how to cooperate with each other, rather than competing or becoming enemies. This aspect is similar to the climate crisis topic — one country cannot solve the risk minimization problem alone because it is a worldwide challenge.
@joy Agree 100%. It is just that AI is progressing exponentially while cooperation to regulate it is going at a snail's pace. It is a bigger threat than nuclear threat, and is moving far more quickly than climate change.Below is the interview with the anthropic researcher who recently quit, and gave up an untold payout because he was so concerned about the threat.
Hey all,
I am not an AI expert by any means. I happen to have been around computers all my life, and have a liberal arts background. Saying that I have a lot of opinions and my nature is to reflect on things.
Jeanne’s points are ones I think about a lot.
My simplification is we need to define a new social contract, the one that we’ve been using does not work in this environment. What does that mean? How do you take the best of the ism’s and jettison the rest, for instance? At the same time can we grasp what we are doing in context?
We are naturally impressed ( good or bad ) by people who can use words. The LLM’s are, in simple terms, impressive by their use of language. This is not difficult to do, most humans have a vocabulary of 3000 words give or take, and use less than 300 for most of their communication. In a sense there is a bit of rose coloured glasses here, but also where there is smoke there is fire.
At the same time adding “ai” to anything has excited investors. Much like blockchain, and .com did in their time. So there is hype and there is actual.
None of this prevents a rogue anything from interfering with a system that is not air gapped. By the way, unless you’ve had electronic devices with internet capability implanted into you, you are an air gapped self contained wet carbon quantum computer running a rogue process purposely, with a predefined runtime if the random seed kill switch doesn’t occur first. Loosely, of course, but it’s something to consider in the discussion if only to prevent the creation down the road of an artificial intelligence which is both self aware and regulated to service to us. Morally and ethically that’s slavery in any form and despite wide acceptance throughout history that’s the condition that tends to lead to the oppressed overcoming the oppressor.
Any yet at the same time this work is greatly increasing the likelihood of earlier detection of diseases. Better storm warnings, and the like. It is able to bring the identification of trends much faster as computers tend to be very good at patterns, because we designed them that way.
The downside of course are things like predictive markets, the bad with the good. We could have predictably predicted that someone would use sex to sell the predictive markets. Much as we could predict a wet carbon would find ways to manipulate the system for financial gain.
I am looking forward to this conversation.