Altman Says AGI Soon, But Will Change the World Less Than We Think

0:00On today's AI Breakdown, Sam Altman at Davos says that AGI is coming soon, but will change things less than we think. Before that on the brief, the Pentagon is in hot water with Congress after funding a scientist with ties to the Chinese government. The AI Breakdown is a daily podcast and video about the most important news and discussions in AI. Go to Breakdown.net for more information about our newsletter, our Discord, and our YouTube channel. Welcome back to the AI Breakdown Brief, all the AI headline news you need in around five minutes. There is no doubt that the biggest geopolitical issue that surrounds artificial intelligence is the relationship between the U.S.

0:37and China, and how AI changes that geopolitical struggle. We've seen a continuous and evolving set of restrictions on U.S. companies in terms of what type of advanced chips they can supply to the Chinese market. And now, Newsweek is reporting that the Pentagon is facing some tough questions from Congress around their support for an AI scientist who has ties to the Chinese government. In November 2023, Newsweek reported that Chinese-born scientist Song Chun-Zhu had received over$30 million in U.S. grants from organizations like the Department of Defense, the National Science Foundation, and the University of California at Los Angeles.

1:10Now, the chairs of two House committees and three subcommittees have asked those agencies why they failed to pay attention to, quote, concerning signs that that scientist was potentially transferring sensitive research on advanced AI to China. In letters sent to those agencies, the lawmakers wrote, in a period of intensifying geopolitical competition with the CCP, ceasing federal government support for Chinese AI development is a critical national security imperative. Now, in addition to just calling this out, the lawmakers are demanding that the Pentagon give, quote, complete documentation relating to all grants that the DOD gave, including copies of all internal communication relating to the work.

1:43They also want a, quote, list of all recipients of DOD research grants who are currently living in China. Now, if you're interested in this particular issue, I highly recommend the Newsweek piece, which has a ton of additional background on Zhu and China's talent development program in general. But I think it's also interesting to contrast this with reporting that we got last week, that US companies, including OpenAI, Anthropic, and Cohere, have held back-channel talks with Chinese state-backed groups on questions of AI safety. It's clear that even as there is rising geopolitical tension, there is a counterbalance of a sense of a need to cooperate on major AI safety questions, which sort of sets up something of a policy confrontation in the years ahead.

2:21Now, meanwhile, over in China, that country's industry ministry has issued draft guidelines for standardizing the AI industry. The draft proposes to form more than 50 national and industry-wide standards for AI over the next few years. 60 % of these prospective standards should be aimed towards serving, quote, general key technologies and application development projects. Now, back home in the U.S., even if some of the comprehensive legislation has stalled out as election concerns grow, there are a number of more discrete efforts that politicians in various states have been putting into place.

2:51For example, in May of last year, Representative Joe Morrell from New York, introduced the Preventing Deepfakes of Intimate Images Act. The bill was referred to the House Judiciary Committee but didn't go anywhere from there. The bill would criminalize non-consensual sharing of sexually explicit deepfakes but also create a right of private action for victims to be able to sue creators and distributors of that material while remaining anonymous. On Tuesday, Morrell was joined by a teenage victim of non-consensual sexually explicit deepfakes to advocate for that bill. Francisca Manny said that her New Jersey high school administration told her last October that male classmates had created and shared sexually explicit deepfakes of her and more than 30 other girls in their school.

3:29Many said, The issue is pretty black and white. No kid, teen, or woman should ever have to experience what I went through. I felt sad and helpless. I'm here standing up and shouting for change, fighting for laws so no one else has to feel as lost and powerless as I did. The glaring lack of laws speaks volumes. Over in California, meanwhile, a new bill from Democratic Assemblymember Mark Berman, proposes to update the state's penal code to criminalize the production, distribution, and possession of AI-generated depictions of child sexual abuse. Now, part of what this bill would do is open up a new avenue of legal recourse against social media companies, and it builds on a law that was signed into effect last year that requires social media companies to basically do more when it comes to child sexual abuse material.

4:09It also allows victims to sue those companies, which is part of why it saw big opposition from a coalition of tech groups. Now, this question of AI-generated child porn is one that has huge media and lawmaker traction, and I think represents something that we're going to see a lot in 2024, which is that even if we don't see comprehensive legislation at the federal level, some of these sort of low-hanging fruit type of laws are likely to make a little bit more progress. Moving over into the realm of industry, Google has laid off hundreds of people as its ad division starts to implement more AI-powered sales.

4:42As Ars Technica puts it, Googlers are now building AI tools so other Googlers can be laid off. Now, in the main episode today, we're going to talk a lot about Sam Altman's arguments about how AI is augmenting people and not yet causing lots of actual job destruction, but this area appears to be an example of people being explicitly laid off or reassigned because AI is actually replacing them. Explains Ars Technica, Google has been packing Google Ads, its most important product, with tons of generative AI features lately. One is a natural language chatbot that helps people navigate the large selection of ad products.

5:12Another is a system that can just make ad assets like images and text on its own based on a budget and goals given by the ad purchaser. Google used to have humans do sales guidance for its products, create art assets, and decide on text and layouts, but now AI can do it a thousand times a second. Now staying on the theme of Google, a group of top ex-Google researchers have raised a fresh $30 million for a Tokyo-based AI lab called Sakana AI. Writes Reuters, The company, which does not yet have any products in market, is looking to make fundamental improvements to today's AI systems through having a large number of smaller models communicate and work together, rather than creating one giant monolithic model.

5:47It will also work on developing AI models better suited for the Asian market because Asian character-based languages function very differently from Western languages. These are a couple of big themes I think we'll see throughout this year, both regionalization of AI models, as well as a push not just for ever-larger models, but for smaller models that wring more performance out of less data. Lastly today, for those watching the AI phone, AI PC, and AI device trend, Samsung is announcing new phones this week, and as The Verge puts it, they're going to be the most AI phones that ever AI'd. Now the question for this Verge author is what that AI is actually going to look like, and how much it's actually going to change the experience of the phone.

6:24They reference the Rabbit as an example of a totally different type of device, and wonder whether we'll actually see anything even close to that disruptive from Samsung. I will of course be here to tell you when we see it, But for now, that is going to do it for the AI Breakdown Brief. Up next, the main AI breakdown. Welcome back to the AI Breakdown. Today, we are headed over to Switzerland for some coverage from the World Economic Forum in Davos. Now, there is a lot of very reasonable skepticism around this event. Antipathy towards it has risen alongside antipathy towards elites in general. But there is no denying that a huge focus of the meeting is artificial intelligence.

7:02And as such, it's a great chance to get a sense of how some of the world's leading politicians and business people are thinking about AI heading into 2024. Today, we're going to focus on comments from Sam Altman, who, in addition to speaking on an official panel, has been doing a set of media appearances. The big banner headline that you might have seen is that in a session with Bloomberg, Altman said that he believed that AGI was likely to be developed in the, quote, reasonably closest future, but that, quote, it will change the world much less than we all think, and it will change jobs much less than we all think.

7:35Now, Altman said that he almost couldn't believe that he was saying that. And I think it's a reasonable question to ask whether that, one, reflects just an updated understanding and mental model of the world, having seen more of how GPT-4, for example, is interacting with the world, or if it's a specific calculation to try to dampen the narrative of AI as a terrifying and powerful and potentially world-ending thing. It's also not impossible that it's some of both. Indeed, in other comments, Altman did say that what they've observed with GPT-4 is that it has had a dramatic impact in how people work, but it hasn't yet led to lots and lots of job destruction.

8:12He said, in other words, it's much more of a tool than I expected. Now, there were a couple other things that I thought were really interesting from that particular Bloomberg interview. At one point, when asked about AI's impact on climate change, Altman said that his evolving model of the world had the currencies of the future divided largely between intelligence on the one hand and energy on the other. Now, when it came to AI's impact on energy, Altman basically argued that the requirements would be immense, more immense than anyone thought, but that it would basically demand an energy breakthrough.

8:43Altman said there's no way to get there without a breakthrough, and basically argued that it was creating a larger incentive to spend more time on things like nuclear fusion. One other thing that he discussed in this interview, which was largely overlooked in the reporting, but which I thought was interesting, was in and around the conversation about copyright. Now, when it came to the New York Times lawsuit, Altman repeated something that he said before, which is that publishers seem to have a sense that OpenAI and other AI companies like them desperately need to train on their data, which Altman argued that they actually don't.

9:12That the entire corpus, for example, of New York Times articles represents only a very small data set, and that ultimately, they don't want to be regurgitating those things. Instead, the reason that he said they were pursuing partnerships with publishers was not to get around copyright issues for publishing, but to create a better chat GPT experience where linked attribution was a part of what it could do. Still, that's all stuff we've heard from Sam before. What I thought was more interesting was when he talked about Dali. He said that he had spent a lot of time talking to artists, and that when he thinks about what he'd like in the future, it's not just the ability for artists to opt out of people imitating their styles.

9:46What Altman wants is some version of a model where artists could custom create their own versions of Dali that did explicitly use their styles and make money from it in some way. It seemed clear from the discussion that this was not something that was an immediate plan, and it didn't even exactly seem like they knew quite how to do that in terms of cutting artists in. But it was the first time that I'd heard him talking about explicit differentiated versions of Dali, not dissimilar potentially from GPTs and their relationship to the main chat GPT, where artists could actually make money from their styles.

10:19Now, in another interview with Axios, he talked about the inherent discomfort that will come with the evolution of AI. Axios writes, Altman believes future AI products will need to allow, quote, quite a lot of individual customization, and that's going to make a lot of people uncomfortable, end quote, because AI will give different answers for different users based on their values preferences and possibly on what country they reside in. Now when it comes to what's coming down the pipeline, it's clear that Altman is thinking a lot about an agentic future. For example, he said, soon you might be able to just say what are my most important emails today and have AI summarize them.

10:50He honed in again on a theme that we've heard from him before around how AI advances could quote help vastly accelerate the rate of scientific discovery, but he said he doesn't expect that to happen this year. And there's definitely been a shift where it's clear that the company is really focused on launching their next model, whether it's called GPT-5 or not. However, Altman did say that they're going to take their time to make sure that it actually is the product that they want to release. Across these interviews, there was also a bunch of denials, most specifically that Sam had his hands in lots of little non-open AI pots.

11:19He said, for example, in the Axios interview that, quote, open AI is what I'm doing, and that it was a misrepresentation to say he's engaged in projects that don't support open AI. In the Bloomberg interview, he also denied having any sort of formal partnership with Johnny Ive, although he did wax poetic about the potential for future AI devices. Now, in addition to these Davos appearances, Altman also just joined Bill Gates on his new podcast. In that conversation, we got a little bit more about GPT-5 as well. Altman said that it will have much better reasoning capabilities and that it would be fully multimodal with speech, image code, and video support.

11:50He said, speech in, speech out. Images, eventually audio. Clearly, people really want that. We've launched images and audio, and it had a much stronger response than we expected. Now, when it comes to how much better GPT-5 will be than GPT-4, Altman actually sort of ratcheted up expectations from what I've heard from him before. Whereas in the past, we've heard him say things like, it won't be as big a shift from GPT 3.5 to 4. This time he said things like, at least for the next 5 or 10 years, we will be on a steep improvement curve. This is the stupidest these models will ever be. Now, while GPT-5 does not appear to achieve that level of AGI, during a recent speech at Y Combinator, Altman reportedly told the founders and entrepreneurs that they should build with the mindset that AGI will be achieved, quote, relatively soon.

12:34Now, one last revelation that I wanted to follow up on had to do with this change in the policy around the military. Reporters picked up recently that OpenAI had updated its policy that had previously said that its tools couldn't be used for military purposes, to be what seemed to be a much more open-ended version of that language. In that same Blueburn interview, Altman was joined by Vice President of Global Affairs Anima Kanju, who gave a little bit more color about this policy shift. First of all, she said, those policies were written before many of those people got to the company, and they were imagining a future that they didn't really understand yet.

13:06Second, she pointed out that it hadn't changed in terms of the substantive ban on using ChatGPT and other open AI tools to create weapons and to otherwise do things that cause humans harm. However, what they did note is that they had been working with the U.S. military on a number of projects that include cybersecurity capabilities, as well as tools that could help with the prevention of veteran suicide. Ana said at that Bloomberg interview, because we previously had what was essentially a blanket prohibition on military, many people thought that would prohibit many of these use cases, which people think are very much aligned with what we want to see in the world.

13:37Finally, when asked about the elections, Allman said it was good that we had a lot of anxiety about it. He argued that people probably should have been thinking about the reality of a Trump candidacy before the Iowa caucuses and advised that people would be better served, spending a little bit more time asking why his message is resonating. But ultimately, he said, when it came to AI in elections, it was important to be focused on it, but not to fight the last war. He pointed out that there was a difference between the tools of creation and the tools of distribution, and the different strategies were going to be required to minimize those risks.

14:06Anyways, very revealing and interesting stuff from Altman and OpenAI at Davos, and I'm sure just the beginning of what we will hear this week. For now, that's going to do it for today's AI Breakdown. Until next time, peace.

14:29Thank you.