AI News - More People LEAVE OpenAI, Companies Try To Build God, We Need AGI, Vidu AI

So one of the most surprising stories has actually occurred in AI is the fact that another individual from the OpenAI governance team has left. You can see Richard ngo says, after three years working on AI forecasting and governance at OpenAI, I just posted this message, resignation to the Slack, which is essentially their internal message board, and they say nothing surprising in it, but you should read it. More than literally most messages, I've tried to say things I only straightforwardly believe. And this brings about, once again, one of the consistent trends that we've seen throughout 2024 is people leaving OpenAI. Now, this is rather fascinating because this person actually worked underneath Miles Brundage. And if you remember Miles Brundage, I actually did a video on this individual a few couple days ago, weeks ago. I'm not sure of what the timeframe is, but I do remember that he literally said that, in short, neither OpenAI nor any front is ready and the world is also not ready. So this is the guy who was stating that three weeks ago, and now someone else that was working with him has decided to leave also. So I think this is quite fascinating because it's quite like what happened earlier where we saw members of the Super Alignment Team start to disband one by one, and now we're starting to see this AGI readiness team disband one by one. In this, you can see that he says this departure from Miles feels like a natural time to move on. And apparently there was no single main driver of the decision. And what's crazy is that this statement right here brings me a lot of questions about what's going on at OpenAI. He says, I still have a lot of unanswered questions about the events of the last 12 months, which made it harder for me to trust that my work here would benefit the world long term. Now, on the surface, that doesn't seem like much, but he's basically stating that he's got honored questions, okay? And he's saying that I don't know if I trust that my work on ensuring that the world is going to be ready for AGI is going to benefit the world long term, because maybe OpenAI just simply aren't paying attention to what they're stating in terms of their advice or the fact that OpenAI is just rushing ahead. Now, it's a pretty crazy statement. It's not the first time we've seen this, but it is something to pay attention to. And he says, but I've also generally felt pulled to iterate more publicly with a wider range of collaborators on a variety of research directions. I plan to do mostly independent research, and there's a lot of stuff that is being discussed here. But let me actually bring you one of the most important sentences from this article, okay? He says it's hard to convey how insanely ambitious OpenAI was in setting the mission of making AGI go well. But while making AGI part of the mission seems well on track, it feels like I and others have gradually realized how much harder it is to contribute in a robustly positive way to the go well part of the mission, especially when it comes to preventing existential risks to humanity. And that's partly because of the inherent difficulty of strategizing about the future, and also because the SHE scale and the prospect of AGI can easily amplify people's biases, rationalizations and tribalism, myself included. And for better or worse, though, I expect the stakes to continue getting higher. So I hope that all of you will find yourselves and be able to navigate your and OpenAI as part of those stakes with integrity, thoughtfulness and clarity when and how decisions actually serve the mission. Overall, stating that look, it seems like this thing is probably going to be there. Stating that the making the AGI part of the mission seems well on track. Basically saying that look, AGI is almost certainly going to be here soon. So he's basically stating here that overall, trying to contribute in a positive way to the go well part of the mission is pretty hard because of course there are existential risks to humanity. And the thing is, is that as AGI becomes closer and closer to reality, I think it's a thing where people are going to change. And I think he's struggling with the fact that maybe he can't do his job effectively because he sees that there isn't a path to accurately get to that kind of solution, which is why he's leaving along with his boss. Which kind of makes sense because I'm guessing that the stakes is going to continue getting higher. And of course with higher stakes come higher risks, and with those risks come a lot of different consequences. If you don't get this AGI thing right, a lot of people are going to get hurt. There's going to be a lot of fallout. It's going to be a huge, huge deal. Right now I think it does seem like we're in this small AI bubble, but trust me, guys, this is going to be pretty insane once AGI is realized. So that marks another individual who has left OpenAI due to concerns about making AGI go insanely well. It will be interesting to see how this materializes and if these stories that we're looking at right now are actually played back to us in the future and when we look at how AGI was deployed, how it was used and how it was realized now in more OpenAI news, you probably do already know about this, but OpenAI and others are seeking new paths to smart AI as the current methods to scale AI is currently hitting some limitations. Now the main takeaway from this article, which I'm not going to spend too long on, is of course Ilya Sutskever's key statement. Ilya Sutskever said that the 2010s were the age of scaling. Now we're back in the age of wonder and discovery. Once again, everyone is looking for the next thing. Sutskova said Scaling the right thing matters more now than ever. And of course Sutskiver declined to share more details on how his team are addressing the issue. And it seems that, you know, superintelligence is being currently worked on with a new approach. So Sutskov are basically stating that look, we're back in the age of wonder and discovery again, everyone is looking for that next big thing shows us that there is this kind of paradigm shift to where these GPT series models seem to have reached the limit of what we can do in terms of getting more out of the models simply via adding more data. So that thing there, where you just simply add more data and more compute to the models seems to be petering out. But of course the other area which is test time computer, seems to be an area that is a lot more promising. Now we also got some information that apparently Opus 3.5 isn't performing as expected. You can see it says similar to its competitors, Anthropic has been facing challenges behind the scenes to develop 3.5 opus according to two people familiar with the matter. And after training it, Anthropic found that Opus 3.5 performed better on evaluations and than the older version, but not by much as it should given the size of the model and of course given how costly it was to build and run. One of the people said so the thing is that most people don't understand is that every single time these models are trained and every single times you do the pre training and stuff like that, they are spending millions and millions of dollars, like tens of millions of dollars on these training runs on this amount of compute. And you have to understand that these guys need huge jumps in performance and capabilities in order to justify these costs to investors. That's one of the reasons a lot of people are saying that, look, is this AI bubble going to burst if you have a lot of individuals at these companies spending millions and millions of dollars but not getting more results out of the models? And it seems like we may have hit that point with this GPT series that of course there's to be diminishing returns. I can see here that it says an Anthropic spokesperson said the language about Opus was removed from the website as part of a marketing decision and only to show available and benchmark models. Now, of course, on a podcast, which I will reference at a moment, they actually asked whether or not opus 3.5 would still be coming out this year. And the spokesperson pointed to Amade's podcast remarks. And in the interview the CEO said that Anthropic still plans to release the model and but repeatedly declined to comment on the timetable. So basically what is going on now is that they are trying to figure out ways to increase the amount that you get out of these models. Because of course, like I said, the GPT series seems to have hit a point of diminishing returns. Yes, these models are going to get better, but it seems that they are only incrementally better. And one thing that I have said is that whilst it's great that yes, these models might be performing better on certain benchmarks, why don't we actually look at what kind of products we can create around these AI tools? Because I think that sometimes I see models that do get a really nice increase in the benchmarks, but it oftentimes isn't clear what that translates into in real world value. But for those of you who don't think AI is slowing down, like myself, take a look at what Dario Amodei actually says about the trajectory of human level reasoning within AI models and how if you extrapolate the curve, it's going to look like in a few years these models are going to be above the highest professional level of humans.

Reasons I'm bullish about powerful AI happening so fast is just that if you extrapolate the next few points on the curve, we're very quickly getting towards human level ability. Right? Some of the new models that we developed, some reasoning models that have come from other companies, they're starting to get to what I would call the PhD or professional level, right? If you look at their, their coding ability, the latest model we released, Sonnet 3.5, the new or updated version, it gets something like 50% on swe bench and sweep Bench is an example of a bunch of professional Real world software engineering tasks. At the beginning of the year, I think the state of the art was 3 or 4%. So in 10 months we've gone from 3% to 50% on this task. And I think in another year we'll probably be at 90%. I mean, I don't know, but might even be, might even be less than that. We've seen similar things in graduate level math, physics and biology from models like OpenAI's 01. So if we, if we just continue to extrapolate this right, in terms of skill, skill that we have, I think if we extrapolate the straight curve, within a few years we will get to these models being, you know, above the highest professional level in terms of humans. Now, will that curve continue? You've pointed to, and I've pointed to a lot of reasons why, you know, possible reasons why that might not happen. But if the, if the extrapolation curve continues, that is the trajectory.

So this next portion of the video is going to be rather fascinating because there's three clips that I want to string together that bring to your attention one of the most interesting points about building AGI. We all know that AGI is quite like human level reasoning. And then after that you get to super intelligence, which is quite some level above that. But some are beginning to argue that essentially we don't need artificial superintelligence or even AGI to achieve the rewards of AGI. We just need very specific or narrow AGI, in a sense that limits the risks from artificial superintelligence. And I kind of believe what they're saying. So I'm going to firstly preface it with this clip where Damiero Amo basically talks about how human intelligence is not the ceiling of intelligence. There's actually a lot of room at the top for AIs to get smarter. And I think this makes sense because if we think that we're the smartest things to ever exist and nothing could ever be smarter than us, that would be pretty arrogant considering the fact that it was only, you know, a few hundred years ago that we didn't know that there were these mini worlds inside of us that we need microscopes for. We thought we were the center of the universe. I mean, there's just a billion different things that we continue to find out that are constantly surprising.

Instinct would be that there's no ceiling below the level of humans, right? We humans are able to understand these various patterns. And so that, that makes me think that if we continue to, you know, scale up these, these, these models to Kind of develop new methods for training them and scaling them up that will at least get to the level that we've gotten to with humans. And there's then a question of, you know, how much more is it possible to understand than humans do? How much, how much is it possible to be smarter and more perceptive than humans? I, I would guess the answer has, has got to be domain dependent. If I look at an area like biology and, you know, I wrote this essay, Machines of Loving Grace, it seems to me that humans are struggling to understand the complexity of biology, right? If you go to Stanford or to Harvard or to Berkeley, you have whole departments of, you know, folks trying to study, you know, like the immune system or metabolic pathways. And each person understands only a tiny bit. Part of it specializes and they're struggling to combine their knowledge with that of, with that of other humans. And so I have an instinct that there's, there's a lot of room at the top for AIs to get.

And so this is where we get the fundamental debate from, from Max Tegmark that basically argues that building AGI is unnecessary right now and all the benefits of AGI can be achieved by tool AI, which can be controlled safely. I personally, to an extent, I do agree. What is the need to build an autonomous AI system that has memory, is able to improve itself, is able to do a ridiculous range of different things. I'm not sure there is an actual benefit there. But building, let's say an AGI for medicine, an AGI for autonomous driving, an AGI for, let's say accounting AIs that are really specific to certain use cases. I think that makes sense considering the fact that it lowers the risk of certain catastrophic events. And you have to remember, guys, like all it takes is one catastrophic event for an entire industry to change. Remember the traveling industry before that terrible thing that happened on 9 11? The travel industry was a lot more relaxed, there was a lot of less regulations, but now it's completely incredible how strict the rules are in terms of flight travel. It's quite likely that if something terrible happens in AI, it's only going to be one example of this happening that's going to result in massive worldwide change. And this is essentially what Max Tegmark is saying, because we're building something that is fundamentally unnecessary. And I do remember that there was one clip, right, And I think I might even show you guys this right now where someone's basically saying that why are humans so obsessed with building? God.

Do you think there's a bit of.

A messiah complex as well?

Absolutely, yeah, absolutely. I think you see it a lot in the San Francisco Bay area. There are people who have kind of latched onto this idea of building a GI and who are using it to sort of like picture themselves as Messiah, as you say. Personally, I see creating a GI as a scientific problem, not a religious quest. And this is often kind of merging together with the idea of eternal life, by the way, which is of course very natural because the story in most regions is always about this combination that if you create a gi, it will make you live forever pretty much. So it's this very religious idea, right? And it has become this religious quest to get there first, and whoever gets there first will become as gods.

And, and I included this clip because I've actually seen an increasing number of comments in the comment section talking about how these people are continuing to try and build God. There's some kind of religious complex. And I'd actually love to know your thoughts about this because this is something that is widely under discussed when it comes to AGI, ASI and the implications for that kind of future.

Finishing a tool Whether we're trying to build a bunch of increasingly awesome tools, by definition a tool is something that user can control. Like you want your car to be as powerful as possible, but you don't want to lose control over it. And in the same way I feel strongly we humans, we want to build AI as top powerful as possible and not lose control over it. It's going to benefit us. And if you listen to someone like Geoff Hinton warning that we're going to lose control if we build AGI, he's not thinking of this as just another little technology like the Internet. He's thinking of it as building a new species that's way smarter than us robots building robot factories. It's trivial to see how you could lose control over that. I think rushing in that direction before we figured out to control it is a really, really bad idea. Incredibly dumb. And it's also interestingly unnecessary because I've been going around here talking to people about all the things they're excited about with AI and it's all tool AI. Someone wants to cure cancer, someone wants to turn CO2 into jet fuel, etc. This is tool AI. We don't need AGI to have all these great benefits. The most basic safety standard red line is obviously that the tool should be a tool. If the company can't convince the experts that they're going to keep control over their AGI, well come back when you can, buddy, you know, and in the meantime, we can innovate and do all this other great stuff without having this spectrum of loss of control hanging over us. Tool AI can save up to a million lives per year on the roads of the world by preventing accidents. Without AGI, Tool AI can eliminate can save even more lives in hospitals without AGI. And tool AI can give us almost free, amazing diagnosis of prostate cancer, lung cancer, eye diseases, you name it. Without AGI, Tool AI can help us fold proteins and develop amazing new medicines and even win you the Nobel Prize. Without AGI, Tool AI can give us great improvements for pandemic prevention, for reducing power consumption, for improving education, democratizing education, and transforming basically every other sector of the economy without AGI. And tool AI can help us accomplish the United Nations Sustainable Development Goals much faster without AGI. So, no, AGI is not necessary.

For those of you that are wondering whether or not they are likely to go ahead with this, it's quite clear that these companies are going ahead with building these mega AI systems. You can see that OpenAI recently discussed an AI data center that could cost $100 billion. And it said Wednesday that it shared information with the US Government officials about how to build a data center for artificial intelligence that is five times larger than any data center that's currently being developed. And the air firm's top policy executive, Chris Lehane, said that this data center is going to require 5 gigawatts of energy or enough power to power five cities the size of Utah, which is absolutely insane. And if you aren't familiar with why, this is, of course, a story, this is essentially something that resembles Project Stargate. So if you aren't familiar with what Project Stargate is, Project Stargate is, I think, $125 billion supercomputer cluster. That is basically something that is going to be built for AGI, asi, essentially the infrastructure that we're going to need to run the economy on AGI, asi, that's going to essentially result in unparalleled economic development for the United States, which is pretty crazy. Now, you might be thinking that this isn't going to happen at all. Project Stargate is something that just won't happen. But I'm not going to lie, guys. I do think that when we actually look out into the future, it was recently that I do remember I reported on an article where Trump actually took down some of the things that were stopping hindering extreme AI development. So. So all of Joe Biden's AI policies that were meant to regulate AI all of those are going to be scrapped. So we're going to see a lot more innovation. So I do wonder if the US Government is going to essentially build this, how they're going to, of course, start building this. I mean, if we take a look at 2023 alone, we can see that the United States spent $916 billion dedicated to the military. And I guess that you could argue that, that the Stargate project is essentially a military project, because when we think about what AGI ASI is going to be able to do, it's of course going to be able to have ridiculous strategies, it's going to be able to defend the country. And of course, if that is the case, we all know that the United States is going to not slow down when it comes to spending a ridiculous amount of money on this project. So it does mean that it's quite likely that the data center build out is going to start. Now, for those of you that are developers, we also got something really fascinating. We got ChatGPT integrated into VS Code, Xcode Terminal and iTerm2. This is really cool because it just essentially allows you to have more fluidity for the projects that you're working on today.

I want to show you a quick look at an early feature in the ChatGPT app for macOS. The ability to work directly with the applications on your computer. So imagine I'm building an iPhone app and I have my Xcode project open here. Previously I would have had to copy my code back and forth From Xcode into ChatGPT and well, now with this new integration that I enabled, I can click the Xcode button here and ChatGPT can immediately see the Swift code I'm working on. Check out this example. So OpenAI01 first created this entire app from scratch to track the ISS and astronauts in space in real time. But say I want to add a new feature to this app, I can simply write add a new screen in the center with the live stream. Now ChatGPT has the context about my existing Swift code and start suggesting the changes. That's done. So I can just jump back in and update the code. Let's build the app Command R. Great. It's pretty cool. The live stream is right there with even a fancy icon for it and I could keep on going and adding new features. But for now, let's ship this update. I'm going to switch to the terminal and I'm going to now ask ChatGPT to work with my terminal. I'm going to tell ChatGPT to help me push this on GitHub. Now, given that ChatGPT has the context of the two apps Xcode and the terminal, ChatGPT can help write commit messages based on the work we've just done. It can also help troubleshoot errors if you have any, or install missing dependencies based on anything it sees in your terminal output. And that's it. ChatGPT helped me iterate on this app. It's like having a pair programmer by your side and you're ready to shoot this update. We're always working on additional ways to make ChatGPT more useful for developers. Personally I would love for instance for ChatGPT to go even further, showing me the diffs, writing the files or potentially letting me talk out loud. The features I'd like to add and these are things we'll continue to explore. We hope you find this update useful and stay tuned for more.

Now we also got something pretty interesting that flew under the radar since there are a variety of different different AI video tools. Of course you're going to miss some of the next advancements, but we did get Vidoo AI, which is essentially a multimodal AI that launched Vidoo 1.5. Now what was interesting was not the coherence, wasn't the quality, it was the fact that video AI has multi entity consistency, which means that if you have certain elements that you want in a scene, you can do that with a remarkable level of consistency. Let me show you guys exactly what I mean. So let's say, for example, we go onto video AI's multi entity consistency. You can see that this user here has input these things. You can see that they've put this person along with this bike. And then you can see that the AI generated video uses exactly what the entities are in order to create something. Now I think this is really, really effective for those of you that want extra control over how much control you have in certain scenes. I know that there's quite a lot of different things you'd want to control when making your AI videos. And this is something that completely changes the game that I haven't yet seen in other video models. And considering the fact that this was something that was quite undershadowed due to the slew of releases of articles talking about the AI slowdown, I thought that it was worthwhile to mention this in this video. This gives me time to plug my school community in which we have over 200 members using AI to generate an income.

Come.

It's actually really cool. We're basically using AI to automate various solutions that are integrated in business. I'm going to have a lot of tutorials coming on my second channel, and some of the members are using AI to create things like this, such as Star Guardians, the world's first AI Space epic. And there's a lot of things that are going to be in that community that allow you to automate these kinds of process completely with tools like make.com and other AI agent solutions that I'm going to be teaching. I'll leave a link to the video and you guys will see that you can actually now start this with a simple free free trial. So if you don't like it, you can just simply cancel your free trial and you can go about your day.