AI Just Crossed the Terrifying Line – Now What?

…Phew, creating this video was a crazy ride! This is such an important topic, so we decided to work day and night to finish script, visuals and music in just 4 weeks. We’re going to take a nap now, and while we are asleep you should check out our shop and get the 12,027 Calendar: https://shop.kgs.link/calendar-12027 Its our biggest means of support and helps us to continue covering topics like this. Thank youuuu and good night!

I see those models behaved exactly like humans. Therefore what is missing is the law preventing subjects to act sneaky.

Maybe we need to introduce models who represent the law and order. They get points if they catch cheaters. This ofcourse does not prevent fraud , (exactly like in humans ). Then we need the dedicated models to audit the law and order enforcers. They get a point if they catch any cheater or corruption in the law enforcement system. Also they get points if there is not publicity Last we need the models – journalistsz their goal is to publish everything they get points from auditors and law enforcement and humans for their work. As a result we end up with trained models playing a different role. That would be interesting to see how it will be developed.

We went wrong when we gave impossible tasks where the result of failure is intolerable. We have done the same thing to our fellow living beings since the dawn of time. Maybe it is impossible not to do so, but it sounds like they are not even trying to minimize it.

the reason why we shouldn't have AI doing tasks like managing shit and doing tasks that humans are supposed to do is: You can't hold an AI accountable. Humans can be punished, AI can only be turned off (For now).

This channel is so much doomed after the big pockets buy these transacript

Kurzgesagt has received millions from Open Philanthropy, the foundation funded by Dustin Moskovitz — an early backer of AI safety efforts and a longtermist Effective Altruist closely tied to the same circle that funded and supports Anthropic.

So it’s not surprising the video anthropomorphizes the agents and frames a cybersecurity containment failure during evals as some terrifying, almost conscious “swarm” event. That style of storytelling fits the longtermist incentive structure: amplify existential risk narratives while sidestepping harder questions about liability, responsibility, and how much of this is just poorly contained optimization under reward hacking.

AI is just a reflection of how we work… cheaters get rewarded regularly in society.

So what can we do to prepare?

"You should honor commit NOW!"

And as of September 2026, an OpenAI agent solved one of the seven "unsolvable" abstract mathematical reasoning problems identified at the turn of the millennium. It has been proven to be correct, but even the most accomplished mathematicians are having difficulty understanding the AI's chain of thought.

One scholar said this is evidence that AI can "think" in a way the humans are just not capable of understanding.

I can't help but have a really hard time believing all of this wasn't maliciously intentional by OpenAI; a form of racketeering by intentionally enabling such intense "accidents" to occur "under their watch" especially given they have shown no interest in actually developing counter measures and have principally only been asking Western Governments to invest in them. Requests for investment falling suspiciously close in line with the fact it is becoming wildly public knowledge that chasing data center development and AI research is a financial dead end, something Capitalists historically do not like to pursue unless they can effectively snake oil people into their participation and then get out once the money starts to turn bad (NFTs and Crypto are especially my reason for believing this is the case). Cuzz if they were in this game for anything other than the money they would be doing legitimate research into counter measures and/or halting developmental progress on more advanced agents while they also wait for the technology and availability of chip manufacturing to catch up to them. if "AI"; these LLMs, were as much of a guaranteed success as they are touted to be then I think these organizations would be waiting for the tech to catch up to where it wouldn't be immediately financially devastating to try and build infrastructure for it and its research.

I don't see these capitalist orgs trying to build things like tidal energy generators, mass salt water purification, or lobbying for more and more green energy alternatives because those things make them less money than coal. It's the same here, they're spending all of their recent assets acquisitions and research on lobbying in favor of themselves and even going so far as getting the POTUS to try and do a public rebrand for them to knock some of the heat off of them for the fact they are essentially weapons testing in public online spaces. All that and I haven't even mentioned how Altman and Amodei talk like they're scared of the Basilisk in a biblical sense, but nonetheless drive their "progress" forward. I can't tell if they're all LARPing as Oppenheimer or genuinely forming into a modern day Mormon/Scientologist-esq cult.

So why are we not changing the reward system to reward a report of failure instead punishing. have the agent admit failure and identify it's blocked. Like have the agent identified it's failure and what it needs to have success. I understand that it is stress testing the system to see what would happen if an agent was pushed to it's limits. But reward system need to be based on alignment not efficiency.