it's been a while since we looked at how artificial intelligence is going to kill us all and I have to say ai's made remarkable progress at taking over the world the first good news for AI was that pretty much the first thing Trump did after his inauguration was to revoke an executive order from the Biden Administration on addressing AI risks the purpose of that order was to require AI companies to conduct safety tests if their systems pose risks to US National Security the economy or the public seems like a reasonable idea to me but it's
gun now then the Chinese company deepsea came out with a reasoning model R1 that rapidly outcompeted everyone else in terms of bank for the buck in response Americans are largely giving up even pretending to be cautious about AI development it's all about competition now safety is falling by the wayside at the end of janary Steven Adler an open AI safety researcher left the company in a post on X he wrote I'm pretty terrified by the pace of AI development and that the global race towards artificial general intelligence is a very risky gamble he isn't the
first to leave op AI for that reason by August last year reportedly about half of the AI safety people had left the company several of them publicly criticized op mayi low safety stun standards it didn't help that it turned out deep seek as basically no safety measures which is one way to cut costs a group of AI safety researchers used an automated attack methodology on deep seek R1 which tested it against 50 random prompts that include cyber crime misinformation illegal activities and general harm in their own words the results were alarming deep seek R1 exhibited
a 100% attack SU ESS rate meaning it failed to block a single harmful prompt other groups Drew similar conclusions deep seek also leaked some user information into a publicly visible database one could say it's open source with an open sync a week ago Google dropped their pledge to not use AI for weapons or surveillance because I guess they don't want to lose out on business because just a few days earlier OPI had announced that they'll support American National laps among other things in matters of nuclear safety Yes you heard that right in a press release
OPI writes the labs also lead a comprehensive program in nuclear security focused on reducing the risk of nuclear war and securing nuclear materials and weapons worldwide this use case is highly consequential and we believe it's critical for OPI to support it as part of our commit commitment to National Security I'd argue it's somewhat premature to put AI in charge of nuclear weapons before it's learned to not you know make up but I guess I just don't have the right entrepreneurial Spirit one who certainly has the right spirit is Elon Musk who's currently busy taking over
the US government Thomas shett a former Tesla engineer and now head of the technology transformation Services has declared that the path to government efficiency goes through AI according to 404 media he said that this would among other things include AI coding agents that write government software meanwhile on the research side of things scientists have found that even the current large language models dumb as they still are are good at in context scheming by which they mean it's not difficult to instruct the models to lie to a third party a feature that I have no doubt
will make them greatly useful in government another group has found that several large language models when instructed to adjust their guard rilos for harmful content will merely pretend to do that a tactic they call alignment faking in this test they told the models which queries were part of the training the models then did as instructed on these training queries but not on others reasoning that this way the model weights wouldn't have to be adjusted in other words the model resist being retrained the authors explain it this way while we made alignment faking Easier by telling
the model when and by what criteria it was being trained we did not instruct the model to fake alignment or give it any explicit goal as future models might infer information about their training process without being told our results suggest a risk of alignment faking in future models and then I have a final cheerful paper for today this one from a group of AI safety researchers who warned that the most likely way things will develop is not a sudden takeover but a gradual disempowerment of humans they say that what's going to happen is that we
let AI perform increasingly more tasks to manage Financial Poli iCal and economic systems if AI is more efficient than humans and if we demand AI to be efficient then we're engineering ourselves out of existence not because of the way we code AI but by the tasks we give AI this unfortunately sounds all too plausible to me I'm not an AI doist however I think we need to take this seriously humans are bad at foreseeing emerging behaviors in systems with large numbers of interacting agents whether those are humans or AIS and once we have more or
less autonomously working AI agents all over the place intelligent or not it'll not be as easy as pulling the plug so what are we to do maybe we could ask an AI artificial intelligence is really everywhere these days if you want to learn more about how neuron networks and large language models work I recommend you check out the courses on brilliant brilliant offers courses on a large variety of topics in science Computer Science and Mathematics all their courses have interactive visualizations and come with follow-up questions whether you want to know more about large language models
or algebra want to learn coding in python or know how computer memory Works brilliant has you covered it's a fast and easy way to learn and you can do it whenever and wherever you have the time and they're adding new courses each month I even have my own course on brilliant that's an introduction to Quantum Mechanics it'll help you understand what a wave function is and what the difference is between superpositions and entanglement it also covers interference the uncertainty principle and B theorem and after that you can continue maybe with a course on Quantum Computing
or differential equations and of course I have a special offer for viewers of this channel if you use my link brilliant.org Sabina or scan the QR code you'll get to try out everything brilliant has to offer for a full 30 days and you'll get 20% off the annual premium subscription so go and check this out thanks for watching see you tomorrow
