BOSTON (AP) — White House officials concerned by AI chatbots potential for societal harm and the Silicon Valley powerhouses rushing them to market are heavily invested in a three-day competition
he s cocky for a nineteen year old. welcome back. right now, thousands of hackers are in los angeles for a competition aimed at taking on artificial intelligence. did is all happening at the biggest hacking conference of the year. over the weekend, the hackers will basically try to break apps like chat gpt and their goal is to see if the machine can be manipulated. and this in turn will give developers a chance to fix any vulnerabilities that they may have, we are now joined by the co-founder of the nonprofit for human intelligence, good morning. thanks for joining us. i understand you are in los angeles, or las vegas, i should say, at the conference. walk me through this exercise which is apparently known as red teaming, what it is, and what they re hoping to find out. that s correct.
so in las vegas, at death con, the largest hacking conference in the world, we are organizing something called the generative exercise, 3,000 people, over 2 1/2 days will get 50 minutes each to try to break every major large language model. so this is chat gpt, google smart, et cetera, and the goal of red teaming is to come in and have a challenge to break models and make it say and do things it is not supposed to say and do. so what are they going to be looking for? how are they going to be doing it? what s the competitive element in this? yes, so the nonprofit community intelligence designed the challenge. the challenges are designed around hacking. so specifically getting the model to say something that maybe it shouldn t be doing, tricking it, but also to get the kind of organic harms that people might see in every day interactions, and i call these imbedded harms. these are things like societal bias, and information integrity and misinformation, and political misinformation.
they reach market. this is something called red teaming and it s where researchers go in and act as villains and bad faith actors to try to break these products and this is something that companies do internally and with this agreement they ll let external researchers do the same, but that being said this is a voluntary agreement and it doesn t have substantive heat and who are the external researchers and what are the standards that they will hold if they find something that they don t like and this is the first step. also notably, the companies have agreed to make a watermark existence to identify a.i. audio and images. we ve seen a raft of a.i. images that are only at this point are able to detect with our eyes and our ears whether something is an a.i. audio or a.i. video. notably they don t agree to in this agreement and nothing on text and open a.i. is chatgpt and we ve seen in schools and universities, students use