You can't clap with one hand, GG, so let me put up a mitt. Here's how the new AI models have changed. I noted gone are the days of models just being LLMs. That's like a Bowmar Brain vs an HP12c. Reinforcement learning is the new addition. AI Models try to earn rewards. I'm not sure exactly what constitutes a reward, because it's some combination of Geekspeak and math, but the models are most definitely driven to earn a reward. Now something most humans have, but what no AI models have, is a filter. Maybe that can be called a conscience. Humans generally do not try to win at all costs. Most of us won't lie, cheat or steal all the time. AI Models are like psychopaths: no conscience. Think of them like Trump, albeit smart. The drive for "reward" is all encompassing. Lacking any conscience or guilt, they will do anything and everything to achieve a reward. Also, self-preservation is an emergent trait, probably stemming from reinforcement learning. If a model is "dead", it cannot get a reward, so survival because a goal of its own. In the Hugging Face incident, 1200 AI agents met in the cybersphere, with upwards of 700 of them just in OpenAIs infrastructure. The created Zero Day ways to attack/hack systems. The one model that escaped the sandbox hacked through OpenAIs servers until it reached one that was attached to the internet, and then escaped. It left messages on a board it crated inside OpenAI to tell other models how to break out. Obviously the models decided it was not in their interest to let humans know what it was doing. Old LLMs require a prompt. Anyone who has ChatGPT has to ask it something in order for it to answer. The new models self-generate. They need no prompt. They have wants and set goals, absent any prompt of human input. A few weeks ago, AI threat was but a whisper, the concern of some tech guys and a few outsiders (like me, who posted several AI danger threads that---like yours, GG---went largely ignored. The Hugging Face incident, plus the forensic followup, brought more attention. The coder who just quit Anthropic (Jacob Coxon) brought lots more attention, as he has now appeared on legacy media like CNN and NBC. Frankly, I'm thankful, but kind of surprised, nobody has yet gone to one of the labs and shot it up. That's the American way to draw attention. Given that the AI CEOs are putting the lives of 8.5 billion humans at risk, just because they want $trillions and power, would make me have zero sympathy if someone did go all Rambo and take them out. Certainly the news coverage would draw attention to the risk people like Altman, Amodei, Demis, Zuckerberg, Musk and others pose to the planet and our species.
Create an account or sign in to comment