The AI Daily Brief: Artificial Intelligence News and Analysis - Anthropic Raising Another $750M at $15B-$20B Valuation

Episode Date: December 22, 2023

Anthropic is raising money again. Also on this episode, a new poll suggests Americans want more government involvement in AI.  Today's Sponsors: Listen to the chart-topping podcast 'web3 with a16z crypto' wherever you get your podcasts or here: https://link.chtbl.com/xz5kFVEK?sid=AIBreakdown  Interested in the January AI Education Beta program? Learn more and sign up here - https://bit.ly/aibeta ABOUT THE AI BREAKDOWN The AI Breakdown helps you understand the most important news and discussions in AI.  Subscribe to The AI Breakdown newsletter: https://theaibreakdown.beehiiv.com/subscribe Subscribe to The AI Breakdown on YouTube: https://www.youtube.com/@TheAIBreakdown Join the community: bit.ly/aibreakdown Learn more: http://breakdown.network/

Today on the AI breakdown, we're looking at new polling on American attitudes towards AI and AI risk. Before that on the brief, Anthropic is raising again and Mid Journey V6 is finally here. The AI breakdown is a daily podcast and video about the most important news and discussions in AI. We're going to Breakdown. Network for more information about our YouTube or Discord and our newsletter. Welcome back to the AI breakdown. Today we kick off our news with another report, another scoop from the information, that Anthropic, who basically never stops fundraising, it seems, is in talks to raise $750 million in a Menlo Ventures-led deal. So the valuation of the company would be at least $15 billion, not including the investment,

and the price of the round hasn't been finalized. The final price could top $18 billion, which, believe it or not, is less than the $20 to $30 billion that Anthropic was previously seeking, but still a 3x increase from the valuation back this spring. Now, from Anthropics' perspective, they're very clearly just trying to strike while the iron is hot and load up their war chest as much as humanly possible. This year they've done deals with Amazon and Google and clearly don't want to be outcompeted by OpenAI simply because they don't have as much money as Open AI. Still, the company is definitely and decidedly behind OpenAI at least when it comes to revenue. While Open AI is over a billion dollar annual run rate,

Anthropic is at a $200 million annual run rate. They see that increasing to a half billion dollars next year that would put that $15 billion price tag as a higher multiple of annual revenue, then even OpenAI's tender offer, implying an $86 billion valuation for them. The information also identifies that to the extent that there were questions around whether investors would be scared off of alternative governance structures following Sam Altman's ouster and then re-entry as OpenAI CEO, the answer appears to be no. As they write, a high price tag would resolve any doubts that Anthropics' unusual governance structure could turn away investors in the wake of the spat between OpenAI CEO, Sam Altman,

and the nonprofit board that controls the startup. The Board of Anthropic, a public benefit corporation, represents shareholders, still a group whose main aim is to ensure Anthropic develops AI that benefits humanity, controls some of the board seats. Now, in many ways, the deal feels not only like an indication of Anthropics' continued success and optimism around them, but just the nod to the reality of how much capital it competes to be in this space that they're in. In other words, it may not just be that they're striking while the iron's hot. It may be that they have no choice except to raise money at basically any time they can.

Now, moving on over into the courts for a moment. Earlier this year, we had rulings in the U.S. from the patent office that entirely AI-generated works could not be copyrighted, while a new case in the UK has led the Supreme Court there to a similar decision. Basically, a U.S. computer scientist named Stephen Thaler wanted to be granted two patents in the U.K. for inventions that were created by an AI system that he had created called Dabas. Those attempts to register patents were refused initially by the U.K.'s intellectual property

office on the basis that inventors must be a human or a company. Thaler appealed their decision, sending it up to the UK Supreme Court, but on Wednesday of this week, that body unanimously rejected that appeal and reaffirmed that, quote, an inventor must be a natural person. From Reuters, Judge David Kitchens said that the court's written ruling was that the case was, quote, not concerned with the broader question whether technical advances generated by machines acting autonomously and powered by AI should be patentable. Now, Thaler's lawyers argued that this means that the UK is not a good home for AI scientists and inventors. They said in a statement that the ruling, quote, establishes that UK patent law is currently wholly unsuitable for protecting

inventions generated autonomously by AI machines and as a consequence wholly inadequate to supporting any industry that relies on AI in the development of new technologies. Now, it should be noted that Thaler was the same person who tried to bring these questions to the U.S. and got that rejection from the U.S. Patent and Trademark Office, but here in the U.S., the Supreme Court declined to hear his challenge. By and large, it seems like at least in the eyes of the law, right now AI is a tool, not an agent in and of itself, but who knows how that will change in the years to come. Now, speaking of AI and inventions or at least discoveries, Eric Topal tweets, Big AI Discovery, a new structural class of antibiotics. The last one took 38 years,

with multiple compounds effective versus metacillin-resistant staphoress without toxicity. So basically what's going on here is that a team from MIT and Harvard used AI to identify new compounds that might be effective against antibiotic-resistant pathogens. The team used AI to screen millions of compounds, then tested 283 promising compounds in mice, and found a number that were actually effective against MRSA, what Scientific America calls, quote, some of the most stubbornly hard-to-kill pathogens. When discussing the innovation, the team from the Collins lab talked about how much the introduction of AI was speeding up the process of discovering new potential drugs.

Now, given that antibiotic resistance was responsible for an estimated 1.27 million deaths in 2019, contributed to nearly 5 million more, and that all of those numbers got worse over COVID, this could obviously be a significant advance. Now, moving into the world of tools for a moment, you can almost feel the video generation space heating up. Pika Labs 1.0 has gotten tons of people excited, runway continues to make things more user-friendly and enable people to create really, really interesting content. And now, has released a new multimodal AI video generator called Video Poet that at least some folks who have tried it say looks incredible. Those are the words used to describe it by Venturebeat.

Now, a couple things that Venture Beat notes are interesting about the Video Poet release. First of all, it's a dedicated team of 31 researchers focusing on this specific LLM. As VB writes, the fact that the Google research team built in LLM for these tasks is notable in and of itself. As they write in their pre-review research paper, most existing models employ diffusion-based methods that are often considered the current top performers in video generation. These video models typically start with a pre-trained image models such as stable diffusion, that produces high-fidelity images for individual frames, and then fine-tune the model to improve temporal consistency across video frames.

By contrast, instead of using a diffusion model based on the popular stable diffusion, the Google research team decided to use an LLM, a different type of AI model based on the transformer architecture, typically used for text and code generation, but instead of training it to produce text and code, the Google research team trained it to generate videos. So video poet was pre-trained on 270 million videos and more than 1 billion text and image pairs. One of the things that the Google team says they do better with Video Poet is that their approach could allow for longer clips. Basically,

right now, one of the big challenges with video generation is that the coherence of the motions break down after just a few seconds. Interestingly, Video Poet is also optimizing for vertical videos, clearly thinking about the TikTok, the YouTube short, and the Instagram real market. Right now, you can see results but you can't try it yourself, so of course we'll be waiting to get our hands on it and see how it actually compares in practice. Lastly, speaking of getting our hands on it, Mid Journey version 6 is finally here. Right now, everyone is scurrying to figure out what it does better and what's changed about how you prompt it.

Because remember, even if overall a model is more advanced, there are many times when you lose something in the translation from one model to another, and you have to update how you interact with it. For example, when it comes to V6, it sounds like part of what has changed. changed is that there's a little less prompt engineering and a little more natural language. As Professor Ethan Malik writes, the really amazing thing is that there is very little skill involved in building these prompts anymore. Just write what you want and you get it. My prompt was literally played of hot peas with a ginger beer on pub counter food photograph with the word Joyce

in old Irish font. And yes, what that means is that text has finally come to Mid Journey. This is one of the biggest differentiators with the new Dolly 3 that was launched a couple months ago, and the biggest reason that I had for many use cases actually switched off of mid-jorney. Other folks have done some comparisons, and while there are certain types of images where the difference between V2 and 6 isn't necessarily super noticeable, there are others where the quality is just mind-blowing, and the adherence to the prompt is significantly better. For example, we're looking at an image right now from Nick St. Pierre, where the prompt was 35-millimeter still of an ancient Roman marketplace during the day.

People in traditional Roman attire are bartering goods, there are stalls with fruits, vegetables, and pottery, and in the background the Coliseum is visible. The comparison is absolutely wild. V2 looks like a high school art class drawing compared to the actual 35 millimeter still that is in the V6 version. I will be doing a much deeper dive on V6 for the AI education beta going on right now. Once again, if you were interested in that, you can check it out at bit.ly slash AI beta. But for now, that is going to do it for the AI breakdown brief. Until next time, peace. Quickly a brief word from today's sponsor. As a listener of this show, I suspect you like to stay up to date on all things AI and tech,

which is why you have to check out the chart-topping podcast Web3 with A16Z Crypto. Produced by venture firm Andresen Horowitz, Web3 with A16Z is the perfect companion podcast to the AI breakdown. Web3 with A16Z crypto is your definitive resource for the future of the internet, whether you're interested in the convergence of AI and crypto or simply curious about what's next. If you need a place to start, they recently released an excellent episode with Stanford Cryptography Professor Dan Boney and former Google X engineer Ali Yaya in conversation with host Sonal Choxi about the intersection of AI and crypto. From fighting deepfakes and proving humanity to large language models like ChatchipT, they cover

it all. I highly recommend checking it out, especially if you'd like to learn more about how AI and crypto will impact our everyday lives. Beyond Crypto and AI, this show is for creators seeking more ways to truly own their work, for business leaders trying to prepare for the future today and for innovators exploring trending tech topics. Don't miss out. Follow Web3 with A16Z crypto on Apple Podcasts, Spotify, or your favorite listening app. Hello, friends, one quick note before we get back to the rest of the episode. Registration for January's AI Education Beta is now officially open. It's open until just Friday

at 1159 p.m. Eastern Time. You can find the link to learn more and register at bit.l.ly slash AI Beta. Now, this is an experiment that I've been running all throughout December, in which every day I drop a new video tutorial or a case study, and usually partner it with a challenge, the idea of which is to get you learning about all of these new different AI tools, as well as specific strategies for the most frequently used, like chat GPT or Dali, and then gets you actually testing them out in the real world with real use cases, and hopefully applying them back to your personal or professional pursuits as well. The first month has gone incredibly well. People seem to be really liking the video content. as well as the incredible community that's forming. And part of that is that it's a group of really serious people. This is a paid experience. It's $20 a month. Part of the reason for that is that I want you guys to judge this content on the basis of whether it's actually worth that much to you.

And second, I wanted it to be full of really serious people who are intent on applying AI to their lives in some real and significant way. Anyways, I would love to have more AI breakdown listeners participate in January. Content will start on January 3rd after the end of the holiday season. and again, the link to find out more and to register is bit.ly slash AI beta. That's bit.l.ly slash A.Beta. And now back to the show. Welcome back to the AI breakdown. Today we are looking at the latest survey results from the AI Policy Institute. They suggest some interesting things about the state of the conversation with the American public as relates to AI and I think are worthy of consideration. However, before we get into this, I do want to talk to, a little bit about how I approach statistics on this show because I think it's important. My general point of view is that almost every survey done legitimately with respected pollsters, or at least with methodology and questions that you can see and have access to,

is worthy of consideration, even if at the same time, I think that we shouldn't treat any as a definitive representation of people's opinions. What's more, I think that when it comes to understanding and contextualizing different polls, we also need to understand the sources that they're coming from. So with that in mind, it is worth noting that the AI Policy Institute has a very specific mission. If you go to their website, the header text reads translating public concern into policy. They write, American voters are worried about risks from AI technology. The AI Policy Institute's mission is to channel public concern into effective regulation. We engage with policymakers, media, and the public to shape a future where AI is developed responsibly and transparently.

Under their mission, they say that we believe that proactive government regulation can significantly reduce the destabilizing effects from AI. They write on their main page, Americans want government action on AI, and so of course, the polls that they are conducting are to some extent slanted to try to get numbers that support that point. So that's your background, that's your context. View everything that we talk about through that lens. So what are some of the highlight numbers from this latest research survey? To that we turn to Daniel Colson, who is the executive director of the AIPI, who tweeted some of the highlights. First, 64% of respondents said that the U.S. needs similar regulations to the EU AI Act and specifically want to

impose testing requirements on foundation models. Daniel points out that 73% also agree that the United States is a leader in the technology, and thus it should also be a leader in setting the rules for AI as well. When it comes to concerns, 68% of respondents are concerned that AI could be used by bad actors to create bio-weapons. 67% support requiring testing and evaluation of all AI models to make sure they cannot be used to create biological weapons before they are released. 64% support the government creating an emergency response capacity to shut down the most risky AI research if it's deemed necessary, while only 16% oppose that. 53% say stability AI should be held liable for the role its model played in generating

fake non-consensual porn images of real people, while just 26% say that only the individuals producing the images should be held responsible. Finally, the last statistic that Daniel focuses on, 80% say Sports Illustrated's use of AI-generated articles and reporter profiles should be illegal, and 84% of response. and say this practice is unethical. 65% support policy that requires companies to disclose in watermark content created with AI, with 46% in strong support. So let's actually dig into the numbers.

First of all, and this is something that I mentioned to their team over there as well, the question that they asked about Sports Illustrated is actually, to my understanding, not accurate. They said, Sports Illustrated recently acknowledged that it used artificial intelligence to write stories and assign them fake bylines. Do you believe that this is an ethical or unethical practice? Now, to my knowledge, Sports Illustrated has not acknowledged that the articles in question were written by AI. They have acknowledged that a third-party content creator created fake bylines, used fake photos for authors, and generated fake bios, but they were still clinging to the very

weird argument that the articles themselves were not AI generated. Now, I don't know a lot of people who believe that. And it's entirely possible that somehow I missed an acknowledgement of that at some point. Certainly Sports Illustrated's parent company took it serious enough to fire its CEO over the whole issue, but at the same time, I think when you're trying to win a public opinion war, you can't play fast and loose by saying that a company acknowledged something that it hadn't acknowledged. Now, if they had actually asked the question the right way, would it have really changed anything? Probably not. I don't think whether they did it or not changes whether this is an ethical or unethical practice, because you can ask that question theoretically as well,

but it is still worth noting. Now, let's talk about which of these questions are surprising or asked in ways, or on the other hand, not surprising. One of the questions, again, still related to Sports Illustrated, was that some policymakers have proposed requiring companies to disclose and watermark any content created with AI. Do you support or oppose requiring companies to disclose in watermark content created by AI? Strongly support was at 46 percent, with somewhat support at 20 This, I don't think, is very controversial at all. In fact, one of the things that we will likely be able to get too quick agreement on is that people need to know when something was created by AI versus when it wasn't. Now, the practical reality of this is much more difficult, in the sense

that the people who are most likely to want people to not know when something was created by AI are the least likely to comply with any rules, laws, or regulations. But still, by and large, people are in agreement that it's destabilizing when people don't know what's real and what's not. They asked a version of that question for political ads, and once again, 69% of people supported requiring political ads to disclose AI content, which, by the way, is the emerging policy for ad platforms like Facebook and Google. Now, on some of the other questions, I am surprised by how pro-government people are. When it comes to AI, which comes closer to your view, the government should not restrict what private companies can do when training AI models. The

government should restrict what private companies can do when training AI models. Should restrict one 75 to 25. Even among Republicans, it was 72 to 28. This next question, when it comes to AI, which comes closer to your view? We can't trust the government to create limitations on what companies can do with AI models. Such restrictions would be fundamentally un-American and will only slow progress in innovation. Versus, it is the role of government to make sure companies don't go too fast when developing AI models to ensure these models can't be abused by criminals and terrorists. This one was a little bit more mixed, but not that much. 67 to 33, so two-thirds to one-thirds, in favor, of it's the role of government to make sure companies don't go too fast, with again, even

Republicans agreeing with that's 62 to 38. So a couple of things that this might represent. One, the discussion around AI has been enormously focused on terrifying world-ending consequences. And when those things aren't the focus, it's been entirely about how AI is going to steal your job. That has clearly seeped into people's perspectives because I don't think that with most industries, people would be this comfortable leaning into government-mediated safety. I think another thing that it could reflect, though, is the fact that big tech has totally lost the heart of America when it comes to trust, concern, and consideration. Especially looking at those Republican numbers, there are no big fans

of big tech either, and I think those chickens may be coming home to roost. Now, this next question, would you support or oppose the government creating an emergency response capacity to shut down the most risky AI research in emergencies? This, I think, gets at another thing which might get a little bit blurry in the interpretation of these statistics. A lot of these questions are aimed at extremes. And what I mean by that is that the implication underlying those government action questions that we were just listening to is sort of not necessarily about day-to-day interactions with companies and what they can do on any given Tuesday, but more, I think, with the background context of these sort of dire emergency scenarios that have been articulated in the media. So what I like about this question is that

it specifically hones in on that emergency context. Sixty-four percent said they strongly or somewhat support this sort of government emergency response capacity. Now, of the questions, the answer that I think is most scary is the one having to do with stability AI. The question precisely is, Stability AI is an AI company which has released an image generating model. Some people have used this model to generate fake, non-consensual, pornographic images of real people. Do you think that Stability AI should be held liable, legally responsible for these images? Fifty-three percent said, yes, Stability A.I. should be held legally responsible. 26 percent said no, only those producing the images should be held responsible, not stability AI itself. This is bigger than just AI, but broadly

speaking, in my belief, the internet has flourished because of Section 230 of the Communications Decency Act back in 1996, which made it such that platforms could not be held liable for what their users did. This to me is fundamental and endemic to the nature of platforms. However, the political mood has changed in such a way that many, many politicians, perhaps a majority of politicians on Capitol Hill, believe that that was a mistake, or at least they wish they had more control over the internet than they do currently because of that rule. I think the idea that platform owners all of a sudden should be held liable for the things that people do with those platforms gets into extremely difficult territory very quickly. It creates massive disincentives for people to create and maintain

these platforms, and I think that there are probably much better ways to come at it that are about articulating the responsibilities that platforms have to help try to prevent that type of behavior rather than just making them legally liable a priori. Somehow it seems I am increasingly in the minority in this position, but there will certainly be plenty of chances to debate it as it is a hot-button topic. Now, one more question to hone in on is the biological weapons question. Here's the way the poll framed it.

There is a debate in the scientific and artificial intelligence communities over how worried we should be about AI being used to develop biological weapons. Those saying we should be concerned point out that AI is becoming increasingly good at doing intellectual work and research. As AI technology advances, it could begin doing the work of biologists and allow a small group of terrorists to design viruses and unleash them. Those saying we shouldn't be concerned say that AI is not that capable right now. All the information AI is providing currently is information publicly available on the internet.

They say worrying about speculative risks is a distraction. What do you think? Couple things. One, I think the construction of this question shows the bias more than maybe any other. One of the reasons for that is that they grammatically separate the theoretical nature of people being concerned and frame the terrorist sentence as a statement of fact. Even if they don't mean to, that's how it comes. off. Now, the second thing that's interesting about this is that I think it speaks to a lot of this

polling about X-risk, which is effectively, if you ask people if they're scared of things that are really scary, they're going to say, yeah, if that's real, that's really scary. What's not here is any sort of assessment or articulation of how likely to be correct any of these things are. In other words, I don't think that the real relevant question is, should we be concerned that AI could be used to create biological weapons? It's more like, how concerned should we be? Then again, that's a very hard thing to ask a non-expert audience, but still it makes the value of that question a little bit more questionable in my opinion. I'm going to wrap on the detailed analysis there, but overall, like I said at the beginning, I do think it's valuable for AIPI to be doing these

surveys. I have some specific disagreements or concerns about the framework of questions. I also have a lot of thoughts about how these opinions might have been formed that I think need to inform how we think about policy in the future. But overall, I think getting more information, more feedback from people is better. But I'm sure you guys will have strong opinions about this. So if you do, please come drop them in the AI breakdown community on Discord. I'd love to have you. For now, though, that's going to do it for the AI breakdown. Until next time, peace.