OpenAI Employee LEAVES With Stunning Statement! "None Of You Are Ready For AGI!"

So, honestly can't believe I'm saying this, but another individual has left OpenAI and they've released some statements that I think you all should be paying attention to. So Miles Brundage is a person who used to work at OpenAI as a researcher and manager and he was someone who was working on the AGI readiness. Now, this is of course, very important. There are many roles within OpenAI that don't get high a lot of publicity, but they are essential to the future of AI development. Now, recently he released this blog post which essentially talks about the future of AGI and what he's doing next. Now, it isn't a dramatic kind of post, but there are a few statements that gives you insights to where the world is on the AGI timeframe. And I think some of the things that are stated in this post blog posts are Quite concerning, despite OpenAI's efforts to mitigate certain issues. So, unlike prior videos, this isn't really an OpenAI drama video, but this is a video that delves into the insights of what exactly is going on when it comes to AGI readiness. And when I talk about AGI readiness, this is just ensuring that the world is going to be ready for the monumental impacts that that AGI will have. Many have speculated that this kind of technology will impact everything from healthcare, medicine to finance, business. Pretty much anything that you can think of, it's going to impact this in some way. So quickly. I do want to show you guys that the reason that this person is leaving is for issues that are outside the industry rather than inside. We said that, you know, the opportunity costs have become pretty high. He's not able to work on exactly what he wants to work on. Of course, he wants to be less biased with what he's working on. Essentially, when you make a statement and you're working at OpenAI, the advice or information that you put out, it can be viewed in a certain light, considering you're working at OpenAI and that company has been under major scrutiny for releasing models early and doing a variety of different things that others who would consider unsafe practices in AI. So I'm guessing that if you want to make certain statements on AI, you aren't really able to do that in a way that is effective if you're currently part of a company that contradicts your own position. So you can see here, he says, Certainly working at OpenAI affects how people perceive my statements as well as those from others in the industry. And he says, I think it's critical to have more industry independent voices in, in the policy conversation than there are today. And I plan to be one of them. Now. One of the main things, slash, the main takeaways from this blog post is one of the statements that I wouldn't say shook me to my core, but opened my eyes to where we are in the AI race, AI development cycle. So there's this question that says, so how are OpenAI and the world doing on AGI readiness? And he says here that in short, neither OpenAI nor any frontier Lab is ready, which is an absolutely incredible statement. Neither OpenAI nor any frontier Lab is ready, and also the world is also not ready. So this is a rather stunning statement because many people, and I've said this before, would traditionally think that, you know, okay, we're just going to waltz into this super futuristic world that is pretty much utopia, as many people would have described the singularity which they predict is going to happen in 2040, 2049, where AI just takes off and technology advances at a rate which is unprecedented. However, I've truly understood that AI is something that is in a state where the acceleration levels that has been continuing to increase year on year as more people get into the space, companies are buying up more chips, and competition from leading labs become more fierce. It has put us in this position where acceleration seems to be the only thing at the forefront. And of course, things like safety have always taken a backseat. And due to the capitalistic nature of our society, companies are placed in a really rough position where, where they need to prioritize the most advanced technology over safety. And as a result of that, We've got neither OpenAI or any frontier Lab being ready for the future of AI. And more importantly, I don't think the world is ready for the ramifications of AGI at all. Now, of course he does state that, you know, this is not a controversial statement among OpenAI's leadership. So he him stating this, this isn't something that is a giant revelation, at least for the people that are working at OpenAI. That's because they know exactly the kind of technology that they do have and exactly how the world is going to respond to certain technological advances. And he also says that that's a different question from whether the company and the world are on track to be ready at the relevant time. Though I think that the gaps remaining are substantial enough that I'll be working on AI policy for the rest of my career. Now, one of the things I do want to get into, because this is one of the main themes of the video, is what does it mean when you say that the world is ready for AI? AI is a multifaceted problem that has many upsides, but also many downsides. He made a blog post in early 2023 which is called scoring humanity's progress on AI governance. And he says here, and this was, you know, last year, that succeeding at AI governance is not merely avoiding catastrophe, but building a much better and fairer world. And it's pretty doable, but it just takes a lot of work. And some of this is quite scary because he says, if we succeed, which is of course not certain, it seems like it will require some or all of the following ingredients and maybe some others. So we need shared understanding of the challenge, technical tooling, deregulatory infrastructure, legitimacy, societal resilience, and different technological development. And you can see here, he says, I'm considering what scores to assign. I roughly imagine that we achieve AGI superintelligence sometime this decade. But I've averaged together a range of possibilities within that window and baked in the current trajectory of progress in that area. For example, an A value doesn't mean that we're ready today, but I expect that if things continue roughly as they are, then they should reach the A level. So this is where he actually breaks down each part. So how you can understand what being ready for AGI actually means? The first thing is, of course, this shared understanding. And the goal is to create a shared understanding of the upsides and downsides to align perceived interests and incentives. He says that this is necessary because shared beliefs in high collective upsides and, and high collective downsides from AI make cooperation more likely. And this is basically where you need everyone to understand how good AI could actually be. But also on the flip side, we need everyone to understand just how bad AI could actually be too. I mean, it's one thing to say that, look, AI could cure cancer, could cure all disease, could cure aging, could solve many of the world's political instability problems and many wars and feuds that we currently have. But on the other side, it could also cause extinction. It could cause the division of people. It could cause economic impacts that cause some people to be displaced. I mean, there's a million different things. We've got scams. I mean, we have to truly understand the full capabilities of this entire category that is AI. Now, he gives this a grade B, because I think the general public agrees on the risks being significant and supports regulation. And policymakers are now talking about AI a lot. And there are some moves in the direction of taking the race dynamic seriously next of course is the technical tooling. We need to make rapid technological progress in safety and policy research, including alignment and interpretability and dangerous capability and proliferation evaluations and proof of learning, so that we have the tools at the right time to ensure that risks are managed appropriately. If you aren't familiar when I talked earlier about extinction, this isn't some sci fi level thing. We have to understand that if there is a model being entity that is 10, 20 times 100 times smarter than we are, the ability for that to do something that is completely above us is going to be easy. I always like the example of where we are building highways and then we destroy nature in the process. Try explaining to an ant, try communicating to an ant why we need to build a highway and that the ant hole ant hill has to get destroyed in the process. An ant couldn't fathom what a highway is, nor could it fathom what economics are. Nor could we even manage to communicate our goals and intentions to an ant or a bee. The whole point here is that imagine us as ants, but a super intelligent AI as the humans in that scenario, how on earth would it even communicate its goals and it would do things that are beyond our understanding. One of the things that are really important is of course interpretability research and of course alignment. Interpretability is essentially where we can actually understand exactly what is going on in the model. So it says here that models are interpretable when humans can readily understand the reasoning behind prediction and decisions made by the model. You can see here in this example, it basically shows how one is interpretable and one isn't. Oftentimes we develop models that are black boxes, meaning that we know the output, but we don't know what the model is doing on the inside. This means that this model isn't interpretable at all and we don't know why it's doing what it's doing. Now that is of course pretty dangerous. If you don't know why a model is outputting the response, even if it's good, that can lead to a whole host of issues. Imagine it's picking things via a decision that you have no idea and then when you deploy it, it starts making random decisions based on different factors. But if you look inside the model and see that, okay, these reasoning steps for the decisions are correct, then we can truly understand what's going inside the model and truly get the right output that we want without any catastrophic risks. Of course, once again, technical tooling for alignment is a gian problem. Having a system that is much smarter than you be Aligned to your goals and have your best interests at yours. I mean, it's going to be something that is pretty hard. We are making progress on alignment. But I do remember that there was unfortunately this news where we had the team that was focused on Super Alignment, which is aligning super intelligent systems that we had. Ilya Sutseva, Jan Leich. This team unfortunately got disbanded by OpenAI. I genuinely don't even remember why the situation was disbanded. I do remember Elias Sutskever had issues with OpenAI and has gone off to start his own artificial superintelligence company. And Jan, like was basically stating that, look, we're not getting enough computing power to even run the tests that we need the evaluation that we need. And this clearly isn't a priority for OpenAI. So I'm going to leave. And another member of the Super Alignment team was fired. So it eventually just pretty much disbanded. Now this is of course a bad news for people who are looking at AI safety, because if you're thinking, okay, we need to understand these incredibly smart systems that we're building, if you start disbanding the teams that are meant to do that, then it means that if you're looking for alignment, this is something that isn't that great. And I would say I do also remember that the team that was meant to do safety research on OpenAI01, they were only given nine days to conduct safety evaluation tests. So I would definitely say that this is an area that OpenAI is lacking in. He does give this a grade C and he says a fair amount of people are working on this, but more people should be working on this. And a lot of investment seems inefficient. People starting random AI safety startups with unclear theories change having to deal with management, fundraising, overhead. It's definitely something that right now isn't getting a lot of funding. I mean, if you're in the AI startup space, majority of companies that are getting funding are ones that are developing a product that's going to be used by millions of consumers. And you have to remember that these grades were given last year. So of course we've got regulatory infrastructure which is where you incentivize sufficient safety by regulating high stakes development and deployment via mechanisms like reporting, licensing requirements, third party auditing of risk assessment, all of these things.