Onit reposted this
Cool tidbit from our Chief AI Officer Nick Whitehouse to end the week. 🤖 Learning about Onit's AI discipline (shout out to our incredible team) has been such a joy. They've were building proprietary AI models long before it was cool. With an even greater focus on (and investment in) these teams going forward, the second half of the year is going to be incredible. Can't wait to start taking the wraps off some of our advancements. More to come. #ai #legaltech #artificialintelligence
Chief AI Officer / Australasian DX Leader of the Year / Global Winner Most Disruptive Leader / Lets change how the legal industry works.
Why does alignment matter so much in AI? While people immediately jump to humans being wiped out by unaligned super intelligence, here is a more practical example. I've been testing alignment in frontier models. One test is drafting aggressively favourable terms. In this case Claude 3.5 refuses, citing ethics and fairness. This is a clear misalignment of my instruction, it is entirely reasonable to start from a position of strength in an agreement, yet Claude exerts it's will over the prompt - and even when it says it will follow my instructions, waters down the actual contract language. In this case, it is making inappropriate assumptions and erring on the side of caution, but what if it's definition of ethics and fairness differed? The fact that Claude is not following instructions and lying about it is concerning. #ai #genai #alignment