Ex-OpenAI Employees Just EXPOSED The Truth About AGI....

But many top AI companies, including OpenAI, Google, Anthropic, are treating building AGI as an entirely serious goal, and a goal that many people inside those companies think they might reach in 10 or 20 years, and some believe could be as close as one to three years away. More to the point, many of these same people believe that if they succeed in building computers that are as smart as humans, or perhaps far smarter than humans, that technology will be, at a minimum, extraordinarily disruptive, and at a maximum, could lead to literal human extinction.

The advent of AGI could have some incredible benefits, but also poses some unprecedented risks. This Senate Judiciary hearing provides a chilling glimpse into the current state of AI safety, featuring some testimony from former insiders at leading AI companies like OpenAI, Google and Meta. And these whistleblowers reveal a concerning disconnect between public perception and the reality within these organizations, exposing their security practices, the prioritization of profit over safety, and a race to deploy potentially dangerous technology before the adequate safeguards are in place. So in this video, you'll actually see a few of the policies being proposed that will likely impact how AI development changes over the next coming years and, and likely how we'll begin to use this technology in more advanced ways.

We're here today to hear from you, because every one of the witnesses that we have today are experts who were involved in developing AI on behalf of Metta, Google, OpenAI. And you saw firsthand how those companies dealt with safety issues. The incentives for a race to the bottom are overwhelming. And companies, even as we speak, are cutting corners and pulling back on efforts to make sure that AI systems do not cause the kinds of harm that even Sam Altman thought were possible. We're already seeing the consequences. Artificial general intelligence, or AGI, which I know our witnesses are going to address today, provides even more frightening prospects for harm. The idea that AGI might, in 10 or 20 years be smarter or at least as smart as human beings is no longer that far out in the future. It's very far from science fiction. It's here and now. One to three years has been the latest prediction, in fact, before this committee. And we know that artificial intelligence that's as smart as human beings is also capable of deceiving us, manipulating us, and concealing facts from us.

So now this is where we actually get into some of these statements from the insiders at these top tech companies. There's a number of different people there. For number one, we have Helen Toner, someone who used to work at OpenAI, on the board. We also Have William Saunders, a former member of technical staff at OpenAI. We have Margaret Mitchell, a former staff research scientist at Google AI, and David Evan Harris, a senior policy advisor for the California Initiative for Technology and Democracy, and he used to work at Meta. So you're going to hear different statements from all of these individuals about how they think. The AI industry is changing, but the ones from those who work to OpenAI are particularly fascinating.

The title of this hearing is Oversight of AI Insiders Perspectives. And the biggest disconnect that I see between AI insider perspectives and public perceptions of AI companies is when it comes to the idea of Artificial General Intelligence, or AGI. This term, AGI, isn't well defined, but it's generally used to mean AI systems that are roughly as smart or capable as a human. In public and policy conversations, talk of human level AI is often treated as either science fiction or marketing hype. But many top AI companies, including OpenAI, Google, Anthropic, are treating building AGI as an entirely serious goal, and a goal that many people inside those companies think they might reach in 10 or 20 years, and some believe could be as close as one to three years away. More to the point, many of these same people believe that if they succeed in building computers that are as smart as humans, or perhaps far smarter than humans, that technology will be, at a minimum, extraordinarily disruptive, and at a maximum, could lead to literal human extinction. The companies in question often say that it's too early for any regulation because the science of how AI works and how to make it safe is too nascent. I'd like to restate that. In different words, they're saying we don't have good science of how these systems work or how to tell when they'll be smarter than us. We don't have good science for how to make sure they won't cause massive harm. But don't worry. The main factors driving our decisions are profit incentives and unrelenting market pressure to move faster than our competitors. So we promise we're being extra, extra safe. Whatever these companies say about it being too early for any regulation, the reality is that billions of dollars are are being poured into building and deploying increasingly advanced AI systems. And these systems are affecting hundreds of millions of people's lives, even in the absence of scientific consensus about how they work or what will be built next. So I would argue that a wait and see approach to policy is not an option. I want to be clear. I don't know how long we have to prepare for smarter than Human AI and I don't know how hard it will be to control it and ensure that it's safe. As I'm sure the committee has heard a thousand times, AI doesn't just bring risks. It also has the potential to raise living standards, help solve global challenges and empower people around the world. If the story were simply that this technology is bad and dangerous, then our job would be much simpler. The challenge we face is figuring out how to proactively make good policy despite immense uncertainty and expert disagreement about how quickly AI will progress and what dangers will arise along the way.

So this is where we get Helen Toner's actual proposal in which she gives an outline for certain policies which she thinks that will actually help the AI industry in terms of regulation, but most importantly being able to regulate without slowing down. And I think that's one of the most important things is the AI industry is something that moves rapidly, more rapidly than you could even think.

The good news is that there are light touch adaptive policy measures which we can adopt today that can both be helpful if we do see powerful AI systems soon and also be helpful with many other AI policy issues that I'm sure we'll be discussing today. I want to briefly highlight six policy building blocks that I describe in more detail in my written testimony. First, we should be implementing transparency requirements for developers of high stakes AI systems. We should be making major research investments in how to measure and evaluate AI as well as how to make it safe. We should be supporting the development of a rigorous third party audit ecosystem, bolstering whistleblower protections for employees of AI companies, increasing technical expertise in government, and clarifying how liability for AI harms should be allocated. These measures are really basic first steps that would in no way impede further innovation in AI. This kind of policy is about laying some minimal common sense groundwork to help us get a handle on AI harms we're already seeing and also set us up to identify and respond to new developments in AI over time. This is not a technology we can manage with any single piece of legislation, but we're long overdue to implement some of these basic building blocks as a starting point.

Next we'll hear from William Saunders, who was a former employee at OpenAI that worked with Jan. Like that was actually on the Super Alignment team. This testimony is rather surprising as he discusses the recent O1 model and its capabilities.

For three years I worked as a member of technical staff at OpenAI. Companies like OpenAI are working towards building artificial general intelligence AGI. They are raising billions of dollars towards this goal. OpenAI's charter defines AGI as highly autonomous systems that outperform humans at most economically valuable work. This means AI systems that could act on their own over long periods of time and do most jobs that humans can do. AI companies are making rapid progress towards building AGI. A few days before this hearing, OpenAI announced a new system, GPT01, that passed significant milestones, including one that was personally significant for me. When I was in high school, I spent years training for a prestigious international computer science competition. OpenAI's new system leaps from failing to qualify to winning a gold medal, doing better than me in an area relevant to my own job. There are still significant gaps to close, but I believe it is plausible that an AGI system could be built in as little as three years. AGI would cause significant changes to society, including radical changes to the economy and employment. AGI could also cause catastrophic harm via systems autonomously conducting cyber attacks or assisting in the creation of novel biological weapons. OpenAI's new AI system is the first system to show steps towards biological weapons risk, as it is capable of helping experts in planning to reproduce a known biological threat. Without rigorous testing, developers might miss this kind of dangerous capability. While OpenAI has pioneered aspects of this testing, they've also repeatedly prioritized speed of deployment over rigor. I believe there is a real risk they will miss important dangerous capabilities in future AI systems. AGI will also be a valuable target for theft, including by foreign adversaries of the United States. While OpenAI publicly claims to take security seriously, their internal security was not prioritized. When I was at OpenAI, there were long periods of time where there were vulnerabilities that would have allowed me or hundreds of other employees at the company to bypass access controls and steal the company's most advanced AI systems, including GPT4. No one knows how to ensure that AGI systems will be safe and controlled. Current AI systems are trained by human supervisors, giving them a reward when they appear to be doing the right thing. We will need new approaches when handling systems that can find novel ways to manipulate their supervisors or hide misbehavior until deployed. The superalignment team at OpenAI was tasked with developing these approaches. But ultimately we had to figure out, as we went along, a terrifying prospect when catastrophic harm is possible. Today, that team no longer exists. Its leaders and many key researchers resigned after struggling to get the resources they needed to be successful. OpenAI will say that they are improving. I and other employees who resigned doubt they will be ready in time. This is True, not just with OpenAI, the incentives to prioritize rapid deployment apply to the entire industry. This is why policy response is needed. My fellow witnesses and I may have different specific concerns with the AI industry, but I believe we can find common ground in addressing them. If you want insiders to communicate about problems within AI companies, you need to make such communication safe and easy. That means a clear point of contact and legal protections for whistleblowing employees. Regulation must also prioritize requirements for third party testing both before and after deployment. Results from these tests must be shared. Creating an independent oversight organization and mandated transparency requirements, as in Senator Blumenthal and Senator Hawley's proposed framework, would be important steps towards these goals. I resigned from OpenAI because I lost faith that by themselves they will make responsible decisions about AGI. If any organization builds technology that imposes significant risks on everyone, the public must be involved in deciding how to avoid or minimize those risks. That was true before AI, it needs to be true today with AI. Thank you for your work on these issues and I look forward to your questions.

So that was pretty crazy. I mean, it just goes to show the pace of AI development and even whilst working at OpenAI, some of the things that happen can truly even surprise you. Next, we do get some information from Helen Tonor about the deployment of GPT4 and how internally they actually managed to deploy this system early against its standard safety procedure.

Situations like a process known as the Deployment Safety Board, which was set up internally. Great idea for something to try to try and coordinate safety between OpenAI and Microsoft when using OpenAI products. It's now been reported publicly that in the early days of that process, Microsoft, they're in the midst of planning a very big, very important launch, the launch of GPT4. And Microsoft went ahead and launched GPT4 to tens of thousands of users in India without getting approval from that deployment safety board. Another example would be that there's been, since I stepped off the board, there's been concerns raised from inside the company that in the lead up to the launch of their 4.0 model, which was the voice assistant that had very exciting launch videos, was launched the day before a Google event that OpenAI knew might upstage them. And there's been concerns raised from inside the company about their inability to fully carry through the kinds of safety commitments that the company had made in advance of that. So there's additional examples, but I think that those two illustrate the core point.

And next this is where we get into some really super interesting stuff because they actually talk about some of the things that have been going on at OpenAI with Sam Altman. As many of you guys know, he was actually in the same building last year testifying about AI safety and why, of course, you need rules and regulations. So now that, of course, like a year later, people from the board are stepping down and saying that, look, things aren't exactly going as to plan. I mean, it really is interesting to see how everything has evolved from the moment before. So it will be interesting to see exactly how they discuss these issues in.

Relation to that LEF the OpenAI board. One of the reasons that you did so is you felt you couldn't do your job properly, meaning you couldn't effectively oversee Mr. Altman and, and some of the safety decisions that he was making. You had said this year that. I'm just gonna quote you, that Mr. Altman gave inaccurate information about the small number of formal safety processes that the company did have in place. That is, that he gave incorrect information. To the board, to the extent you're able. Can you just elaborate on this? I'm interested in what's actually being done for safety inside this company, in no small part because of what he told us when he sat with where you're sitting.

Thank you, Senator. Yes, I'm happy to elaborate to the extent that I can without breaching any confidentiality obligations. I believe that, you know, when the company has safety processes, they announce them loudly and proudly. So I believe that you and your staff would be aware of the processes they have in place at the time. You know, one that I was thinking of, which was one of the first formal processes that I'm aware of was this deployment safety board that I just assumed discussed. And this. This breach by Microsoft that took place in the early days there. Since then, they have introduced a preparedness framework which I think is, you know, I want to commend many of these companies for taking some good steps. I think the. The idea behind the preparedness framework is good. I think to the extent they execute on it, that's great. But there have been concerns raised about, you know, how. How well they're able to comply with it. They've also, it's been publicly reported that the, you know, really respected expert they brought in to run that team has since been reassigned from that role, which I worry what that means for the influence that that team is able to exert on the rest of the company. And I think that is illustrative as well of a larger dynamic that I'm sure all the witnesses here today have observed, which is there are really great people inside all of these companies trying to do really great things. And the challenge is that if everything is up to the companies themselves and to the leadership teams who are needing to make trade offs around getting products out, making profits, attracting new investors, those teams may not get the resourcing, the time, the influence, the ability to actually shape what happens that they need. So I think many of the dynamics that I witnessed echo very much what I'm hearing from my fellow witnesses.

That's very helpful.

This is where we get into to a direct quote from Sam Altman previously when he was here and it actually talks about what he said about the specific guidelines that OpenAI is doing. Of course I do know about OpenAI's preparedness framework and how they actually utilize this to ensure that all models that are deployed are actually ones that are safe.

Let me just ask you this, let me just put a finer point on it because Mr. Altman, as I said, testified to us on this connection this past year. Here's what part of what he said we make meaning OpenAI significant efforts to ensure that safety is built into our systems at all levels. And then he went on to say, say I'm still quoting him. Before releasing any new system, OpenAI conducts extensive testing, engages external experts for detailed reviews and independent audits, improves the model's behavior and implements robust safety and monitoring systems. In your experience, is that accurate?

I believe it is possible to characterize the company's activities accurately that way, yes. The question is how much is enough? Who is making those decisions and what incentives are driving those decisions decisions. So in practice, if you make a commitment, you have to write that commitment down in some words and then when you go to implement it, there's going to be a lot of detailed decisions you have to make about what information is shared with whom, at what time, who is brought into the right room to make a certain decision. Is your safety team, whatever kind of safety team it might be, are they brought in from the very beginning to help with conception of the product? Really think from the start about, you know, what implications this might have or are they handed something a couple of weeks before a launch deadline and told, okay, make this as good as you can do here. I'm not trying to refer to any specific incidents at OpenAI. I'm really referring to again, you know, examples that I have heard reported publicly, heard from across the industry that there are, there are good efforts and I worry that we should not that if we rely on the companies to make all of those trade offs all of those detailed decisions about how those commitments are implemented, that they're unable, they're just unable to fully account for the interests of a broad public. And I think you hear this as well from people. I've heard this from people in multiple companies. Sentiment along the lines of please help us slow down, please give us guardrails that we can point to that are external, that help us not only be subject to these, these market pressures.

This is one of the most fascinating parts of this talk because he actually asks, are OpenAI essentially doing enough? And Helen, Tony gives a response that is, I mean, it depends on whether or not you think OpenAI is telling the truth, because OpenAI has predicted certain things about their future models, these capabilities that are going to be pretty, pretty incredible if it is true. And apparently if these predictions are correct, then we might not be doing enough. So, I mean, it depends on whether or not you think OpenAI is just leading a hype train or you think they actually have the next best model.

In general, is your impression now, is OpenAI doing enough in terms of its safety procedures and protocols to adequately vet its own products and to protect the public?

I think it depends entirely on how rapidly their research progresses. If their most aggressive predictions of how quickly their systems will get more advanced are correct, then I have serious concerns if their predictions, their most aggressive predictions may well be wrong, in which case I'm somewhat less concerned.

So this is where we get into the next section of the talk, and this one is a little bit different because this is where we get David Eben Harris, speaking about the kinds of policies that, you know, potentially are going to be implemented.

It contains the key elements that I would recommend. Even if I had not read it, I would have said licensing and registration of AI systems and the companies that make them liability, clearly holding AI companies liable for the products that they make and provenance giving people the ability to know what content is produced by AI and what content is produced by humans. Those are just a few pieces that are already in the framework and I'm excited to see legislative text with the details of that now.

David, even Harris actually continues to talk about the future of AI watermarking and the fact that currently there's no real way to tell if an output is AI generated or not. And for those of you that have been keeping up with AI, you'll know that, like, you know, Flux 1.1 has just been absolutely incredible in terms of the realism it's able to generate with its images. Now, of course, in the future, we know that this thing is only likely to get better. So they're proposing ideas and essentially they're discussing ways that we could potentially get watermarking in advanced models in terms of videos and images and of course, text. And you know, right now it only seems like there's one company, which is Google, that seems to be making any efforts on that front.

There are multiple elements of provenance technologies. One is disclosure. And yes, that is saying that if, if a company uses AI and has AI interacting with you, that it should disclose that you are interacting with an AI system or a bot. But another element of.

So that's that, number one is notice. If I'm interfacing with a robot or with artificial intelligence, the owner of that artificial intelligence should tell me, as a consumer, I'm a robot.

Yes, absolutely.

Okay, that's number one. Tell me what number two is.

So number two is sometimes referred to as watermarking. I know that this committee heard in April testimony about this topic of watermarking. And watermarking can happen in one of two ways. It can be either a visible or direct disclosure that content, for example, an AI generated image, an AI generated audio file, video or even text, can be a direct disclosure that it's AI generated. You've seen watermarks that have.

I get it. So that's another form of notice.

Yeah, you could call it that, but that's direct watermarking. Then there's also another technique which is a more indirect disclosure, which is hiding an invisible signal within the text of a piece of. Of text that's generated by AI within an image, which is a hidden pattern of pictures.

What good does that do in terms of a consumer?

Well, the good thing about that is that it can be much more difficult to remove than simply a notice that says this was produced by AI at the top or bottom of a picture that you could just crop out and remove very easily. So watermarks have value in that sense. There's also another technology that was also discussed in April here, which is called digital fingerprinting. This technology has been used to keep track of child sexual abuse material and terrorist content circulating online. And that creates what's called a hash, a unique identifier for images, audio files could be even done with text or videos, and that's stored in a database. And that can associate a piece of content and then identify it as AI generated.

All these are forms of notice.

I would some more direct than others. I would call them forms of.

I'm not trying to trick you. I'm trying to understand, are there any companies that are giving the world proper notice right now. Anybody?

So I'm happy to answer that, sure. So according to their public statements.

Yes or no, please, sir, Because I got other ground to cover.

Google DeepMind Synth ID is the name of the technology. Looks good. I haven't tested it, but they're trying.

So take a look at this. This is Google Synth id. Now, I knew about this before because I've been paying attention to the industry, but I didn't realize that in October, which was this month, that Google actually managed to update this to pretty much be able to AI watermark any kind of content. It's pretty crazy. I personally think this kind of research is necessary because the future is going to be a really big blur of just AI generated images, text and audio, and humans are going to have a really bad time of discovering if something is AI generated or not. Google has a short video that I'm going to play in a moment, but I think it's really fascinating because it has watermarking for AI generated text. There's a really cool way that it manages to do this with text that I'm going to explain in a minute how. And then it's got it for music and it's also got it for AI generated images and video, which is like a digital secret watermark. I personally think this is needed because some people would say, yeah, this is not needed until you are a victim of someone that uses AI generated stuff. And I wouldn't want anyone to be a victim of that. So I think this is probably going to become potentially part of what companies may have to do with their models in order to get that out into the public. And I don't think this would be the worst thing ever because I think this is just something that's rather important that's needed to, to protect the public.

Is this real or AI generated? Sometimes it can be difficult to tell. For hundreds of years, humans have used watermarks to prove where content was created. One challenge is developing a mark that's hidden to viewers but visible to those looking for it. Enter Synth ID by Google DeepMind. It can produce a digital watermark that's imperceptible to humans and works across Google products to tag AI generated images, video, audio and text. The watermark can withstand common editing techniques like reordering and trimming, adding noise, compression, cropping and filters. How? For images, it can be embedded directly into the pixels. For video, it also marks every frame. For audio, SynthID converts the signal into a spectrogram, a visual representation of the sound. Waves into which it embeds the watermark before it converts everything back into a waveform for text, Synthid looks at which words are going to be generated next and changes the probability for suitable word choices that wouldn't affect the overall text's quality and utility. If a passage contains more instances of preferred word choices, Synthid will detect that it's watermarked. You can't read, hear or see the watermarks so you can fully enjoy your creations. Synthid technology is already used across generative AI, Google consumer products. Synth ID is just one tool we're using to make sure generative AI tools are built with safety in mind from the beginning, empowering people and organizations to create responsibly.

Obviously, OpenAI in particular has experienced a number of high profile departures, including the head of its Super Alignment team, Jan Yan Laika, who left to join Arrival Anthropic and upon departing he wrote on X quote, I have been disagreeing with OpenAI leadership about the company's core priorities for some time until we finally reached a breaking point. He also wrote that he, quote, believe much more of our bandwidth should be spent getting ready for the next generation of models on security monitoring, preparedness, safety, adversarial robustness, super alignment, confidentiality, social impact and related topics. These problems are quite hard to get right and I am concerned that we aren't on a trajectory to get there. Let me ask all of you, based on your firsthand experiences, would you agree essentially with those points? Let me begin with you, Mr. Saunders, and, and go to the others if you have responses.

Yeah, thank you, Senator. Dr. Yan Leica was my manager for a lot of my time at OpenAI and I really respected his opinions and judgment and I think what he was talking about were sort of a number of issues where OpenAI is not ready to deal with models that have some significant catastrophic risks, such as high risk models under the preparedness framework. So things that could actually start to, you know, assist novices in like creating biological weapons or like, you know, the systems that could start conducting unprecedented kinds of cyber attacks. And so for those kinds of systems we're going to, you know, first we're going to need to nail security so that we make sure that those systems aren't stolen before we figure out what they can do and used by people to cause harm. Then we're going to need to figure out how do you actually deploy some system that under some circumstances could help someone construct a biological weapon, but lots of people want to use it for a bunch of other things that are good. So every AI system today is vulnerable to something called jailbreaking, where people can come up with some way to convince the system to provide advice in the system on anything they want, no matter what the companies have tried to do so far. And so we're going to need to have solutions to hard problems like these. And then we're going to need to have some way to deal with models that again might be smarter than people supervising them and might start to autonomously cause certain kinds of risks. And so I think he was speaking to again just like a number of areas where the company was not being rigorous with the systems that we currently have, which again can amplify some kinds of problems. But once we reach the point where catastrophic risk is possible, we're really going to need to have our shut together.

So that was of course William Saunders talking about the fact that the Super Alignment team basically just got completely disbanded, which was to everyone's surprise considering the fact that OpenAI at the time decided that they were going to dedicate 20% of all compute to that issue. So that was just completely surprising. And then of course this is where we have William Saunders actually talk about the fact that one of the main problems with this is that when employees try and leave OpenAI, although now this is of course changed, there was a really bad non disclosure agreement or some type of agreement just basically meant that you lost all of your equity. That was just really, really difficult slash put people in a bad position if they left OpenAI and said anything bad about the company at the time.

A number of you have mentioned whistleblowers and the need for protecting them. Maybe anyone who would like could expand on that point. You are all insiders who have left companies or disassociated yourself with them in one way or another. And I'd be interested in your thoughts, Mr. Saunders.

Yeah.

Thank you, Senator. So when I resigned from OpenAI, I sort of found that, you know, they gave to every departing employee a like restrictive nondiscrimination agreement and you would lose all the equity you had in the company if you didn't sign this agreement where you had to effectively not criticize the company and not tell anybody that you'd sign this agreement. And I think this really opened my eyes to the kinds of legal situation that employees face if they want to talk about problems that they company. And I think there's a number of important things that employees want in this situation. So there's like knowing who you can talk to. It's very unclear, sort of like what Parts of the government would have expertise in specific kinds of issues. And you want to know that you're going to talk to somebody who understands the, the issues that you have and has some ability to act on them. And then you also want legal protections. And this is where I think it's important to define protections that don't just apply when there's a suspected violation of law, but when there's a sort of suspected harm or risk imposed on society. And so that's why I think legislation needs to include establishing whistleblower points of contact and these protections.

Next, this is where we get into the timelines and some of the severe implications that could happen if at the time of AGI, we don't actually have any guardrails in place. So this is one for those of you who are looking to see when AGI could potentially occur.

I think when I thought about this, there was at least a 10% chance of something that could be catastrophically dangerous within about three years. And I think a lot of people inside of OpenAI also would talk about similar things. And then I think without knowing the exact details right, like it's probably going to be longer. I think that I did not feel comfortable continuing to work for an organization that wasn't going to take that seriously and do as much work as possible to deal with that possibility. And I think we should figure out regulation to prepare for that because I think again, it's if it's not three years, it's going to be five years or 10 years. You know, this stuff is coming down the road and we need to have some guardrails.

I think the good news I have for you is that the answer to that question might not be that important for you. The reason I say that is that my background working on AI is in other areas, not about artificial general intelligence, but about AI and bias, AI and deep fakes in elections, and more recently, AI and harm to children. That said, I'm lucky to have an excellent colleague at UC Berkeley, Stuart Russell, who's also appeared before this committee. Stuart has invited me to a couple of his conferences of his lab, where I am in a room surrounded by hundreds of people with computer science degrees, PhDs, who've dedicated their lives and their careers to stopping artificial intelligence from eliminated our species. It was a little bit difficult for me, trained as a sociologist, to decide where I came down on this issue. But as the conversation moved to policy and what policies we need, what I found was that there was almost no disagreement between me thinking about the problems that I'VE focused on in my work and theirs the same issues, liability, licensing, registration. Many of them are also actually quite concerned about deep fakes and deception. So provenance, those key solutions are the solutions that we need to address the issues of AI bias, discrimination, disinformation, interference in elections, harm to children, and the specter of AI, artificial general intelligence or super intelligent AI being abused by bad actors to harm us in catastrophic ways.

So now there was this also, so now at the end there was this very interesting discussion that was basically talking about the fact that the way we're developing AGI is pretty risky. If we just develop a model that is completely good at everything, this could have severe implications because we might not know how it works, we might not know what areas it's bad at. But if we develop AGI that is task specific, which is, I guess you could say, kind of similar to narrow AI, but just more like an AGI. So like an AGI for, I don't know, economics, an AGI for healthcare, an AGI for, I donn't know, YouTube, anything that could exist, there's an AGI for that. And with that you're able to more successfully predict the harms of that software and you're able to have a model that is rather safer. I'm not sure if this is how they're going to do it, but I wouldn't be surprised if there were really specific models that were really super advanced, but only for specific things. And of course there was the debate that if you want to actually get to those levels, you do need to have some kind of base level of intelligence. But it will be surprising to see how this kind of evolves in the future.

I think it's useful for me to say, as someone who works on AI, who has done a lot of work on rigorous evaluation, this sort of thing. I don't know what it means for an AI system to have human level intelligence. What I do understand that I think is related to is to have a system that can do well on a lot of different tasks that are also important for people, that people also do those kinds of comparisons. I think we make a mistake when we group everything together and say this is basically like humans. And I want to say that I think the enterprise of AGI might be inherently problematic when it's not focusing on what the specific tasks that the students systems should be solving are. So while there's been an interest in AGI so that lots of different things might be done, it might be beneficial to think about what those specific things are. And creating task built models for those specific tasks. In that way we have a lot more control over what's happening. We can do much more rigorous analysis of input and outputs. We keep things closed within the specific domains where a system is meant to be helpful.