In this monologue, the speaker outlines a recent controversy where the International Association of Safety and Ethical AI, led by prominent AI scholars such as Stuart Russell and Yoshua Bengio, accused Google of breaching a safety agreement by releasing its frontier model Gemini without publishing a required safety report. Google defends the omission by claiming Gemini is still experimental and thus not subject to a finalized safety assessment. The speaker contrasts this with Morgan Stanley's internal AI assistant, built in partnership with OpenAI, which underwent extensive stress testing, evaluation, and was trained on over 100,000 business documents to safely support financial advisors. The discussion emphasizes the stark difference between external model releases and internal, business‑focused AI deployments, urging that any AI system rolled out to human users must undergo rigorous safety, accuracy, and stress‑testing processes, and that companies should be held accountable without waiting for regulation.
Chapters
The International Association of Safety and Ethical AI, led by Professor Stuart Russell from Stanford University and Yoshua Bengio and some other pioneers in AI, accused Google a couple of weeks ago of violating the safety agreement they had during the Safety UK Summit that happened, um, a couple of years ago. And apparently Google released Gemini, um, which is a frontier model and a large model, generative AI model, of course. And the agreement was for any tech companies that are releasing any large models, large frontier models, they have to release a safety report. And Google did not do that. And Google defended themselves saying that, you know, the version of Gemini is still in experiments, it's being experimented, and that's why they hadn't concluded a safety report. Port. Um, and that's a big problem, right? You don't release a product, a model for people to use, actual human beings to use, if you have not concluded all your stress testing and evaluations that need to be done. No one does that in engineering. It's not done. And fair play, fair play to Google, right? My friends are still at Google. Like, I worked in the company as well. Um, we know that work is always being done in that regard, but ideally, if an agreement was made you know, in public, you should be stopped too. We don't have to wait for regulation to enforce and force us to get this work done. You know, people need to be held accountable. Let's look at the flip side, on the positive side, right? You would know we have a very good example of a company that is using AI, chatbots, generative AI, and is still doing it properly. So we have Morgan Stanley, who is well known to have been one of the very first financial companies to use AI, and they started, you know, working on an AI assistant, um, partnering with OpenAI for their financial advisors. And they made sure they did all the stress testing, evaluations, and they trained it on over 100,000 documents. So in other words, when the financial advisors have questions from their clients, rather than racking through so much documents to find the answers, they can just go to Chatbot and get the answers. That's a good example of AI being used, right? And SafeAI, and a good example of, you know, the due diligence being done before it's launched. And with those sort of AIs that is being used purely for business purposes, no one can complain about fabrication and hallucination and lack of safety because it's just been trained on business documents, and whatever information it gives would be based on what's trained on. So when you compare both companies, It's glaring that, you know, one takes safety seriously and the other doesn't. Or you can say there are two different use cases. One is releasing a model for the rest of the world and the other is doing it internally. Whatever the case may be, always remember if you're using AI and you're going to roll it out to human users, always think of the implications of it, the safety of it, the instances where people can have inaccurate responses and be misled and make sure that the right processes are put in place to do stress testing and evaluations and accuracy testing and the right things to do before you launch your product using AI.
