AI Proxy War: Google Invests $2B in Anthropic

0:01Today on the AI Breakdown, we're looking at the reactions and controversy around the Biden AI executive order. Before that on the brief, Google joins Amazon in making a huge investment in Anthropik. The AI Breakdown is a daily podcast and video about the most important news and discussions in AI. Go to breakdown.network for more information about our Discord, our newsletter, and our YouTube channel.

0:24Welcome back to the AI Breakdown Brief, all the AI headline news you need in around five minutes. Happy Halloween, everyone. Today, we are doing a brief that actually covers yesterday as well. Obviously, yesterday we got the White House executive order on artificial intelligence, which pushed out a lot of the other news. But we are catching up with some fairly big stories, starting with news from the very end of the week last week that Amazon was not the only big tech giant investing in Anthropic. Hot on the heels of that announcement just a few weeks ago, Google has now committed up to$2 billion in funding for Anthropic as well.

1:00Now, in the same way that Amazon's$4 billion reported investment is actually$1.25 billion now, with more to come later, Google is investing$500 million up front right now, with an agreement to invest up to$1.5 billion more over time. Notably, while that was Amazon's first investment into Anthropic, Google had already invested$550 million earlier this year. Now, there is a lot going on here. First of all, it's quite clear that the template of the The Microsoft OpenAI deal has been something that has been seized upon. And in many ways, it feels not just like a shift in the scale of corporate investment from big tech incumbents into new startups, but a change in kind as well.

1:39The OpenAI and Microsoft deal obviously has very deep integration between these companies. Many parts of the relationship are articulated and played out, including who gets what revenue from how different companies sign up to use OpenAI services. Indeed, we talked last week about how companies buying access to OpenAI's enterprise services through Microsoft Azure is actually kind of harming their bottom line as they get paid much less for that than if they go direct to OpenAI. Now, Azure is important for another reason. Another big part of the template for these deals is clearly about competition in the cloud business.

2:09That Microsoft OpenAI deal came with a commitment to use Microsoft's Azure cloud to train OpenAI models. And when Amazon and Anthropic announced their investment back at the end of September, cloud was clearly a part of it as well. Anthropic wrote, quote, AWS will become Anthropik's primary cloud provider for mission-critical workloads, providing our team with access to leading compute infrastructure in the form of AWS Tranium and Infersia chips, which will be used in addition to existing solutions for model training and deployment. They also talk about expanding their support of Amazon Bedrock.

2:37And so again, this is a comprehensive interaction and relationship, not just an investment, but a deep integration. However, it sounds like they're doing something similar with Google Cloud as well. As part of this investment, the WSJ also writes, quote, Quote, Anthropic has also signed a multi-year deal with Google Cloud worth more than$3 billion. That contract was signed a few months before the new investment. Now combined, this means that Anthropic has raised, or has access to have raised, almost$7 billion in the past year. That's not as much as, but certainly is getting a lot closer to, the$10 billion that OpenAI raised from Microsoft.

3:11One of the big lines that Anthropic has pitched to investors is that, as again the Wall Street Journal puts it, Quote, Now, if you want to understand how big the stakes are for these big tech companies, outside of just the future potential of whatever AI can turn into, there are also very serious share price implications right now. Microsoft soared last week after announcing better-than-expected performance in its cloud business, while meanwhile shares of Alphabet, which is of course Google's parent company, have fallen more than 10 % since their earnings last week that had their cloud growing less than any time since 2019.

3:48It is a fascinating dynamic and wrinkle to this space how these big tech companies are combining with the startup labs, and it's something that I'm sure we're going to hear a lot more about how this deal with both Google and Amazon at the same time came together. Now, speaking of share prices and speaking of investments in AI, in an interview with the Financial Times, storied venture capitalist Vinod Khosla has said that the excitement and hype around AI will lead inevitably to overvalued startups. Earlier this month, he said that most investments in AI today, venture investments, will lose money.

4:18He said that many newer investors are investing because everybody else is investing. Now, by way of comparison, last year, VCs put$5.1 billion into AI companies, whereas this year, they've invested$21.5 billion into AI companies. Now, Khosla is not an AI skeptic. According to the FT, quote, He believes AI has the potential to take on 80 % of the workload and 80 % of all human roles over the next two decades and will create huge economic value. He also, they note, was one of the earliest investors in OpenAI, investing$50 million in 2019 at a startup that valued the company at around$1 billion. Given that the firm is rumored to be raising at around an$86 billion valuation, that stake is obviously now worth a heck of a lot more.

4:58Now, speaking of OpenAI and their leading product, ChatGPT, over the weekend, people started to notice two things. One was a new PDF reading feature, which we'll get into in just a moment. But another was an update called All Tools, where rather than selecting which version of the GPT model you want to use, be it Browse with Bing or Dali or the traditional model, all of them are natively integrated into the default model. So for example, last week, you weren't able to upload a photo into the Dali model and ask for a different version of that photo, but you were able to upload a photo into GPT for Vision.

5:32Now the barrier between those two is totally broken down, and you can upload a picture and then ask for a version of it, modified in whatever way you want to do it. Now this also integrates web browsing, and it integrates their advanced data analysis, which used to be called Code Interpreter. On the one hand, this feels like what ChatGPT was always supposed to be, but on the other, it is a dramatic upgrade in the user experience. And in that way, it's one of the most profound and clear trends for this fall in artificial intelligence, which is updates to the user experience that make these tools radically more useful and integrated into our workflows as they exist currently.

6:08Now, as I mentioned, there was another update as well, which is that OpenAI has officially added native PDF chat. The way they describe it, upload many types of documents, work with PDFs, data files, or any document you want to analyze. Just upload and start asking questions. Now, of course, if you've ever used any of the plugins on ChatGPT, this is one of the most popular features to build around. Indeed, some companies were not just building plugins, but we're building entire experiences around this. As Alex Kerr pointed out, many startups just died today because OpenAI added PDF chat. You can also chat with data files and other documents.

6:40We had a wave of products better suited as features rather than standalone companies. Wrappers are being squeezed by OpenAI on one side and incumbents on the other. It's a rough world out there. This again has been one of the big themes of the year. That many of our expectations of how the AI startup space was going to play out just haven't come to pass, and in fact very different forces have tended to favor incumbents and bigger players. We've seen a lot of the companies that are enterprise wrappers on OpenAI also struggle with Jasper and its set of layoffs serving as example number one, but it's also happening with these smaller sort of feature companies who it turns out were ultimately just building core functionality that OpenAI hadn't gotten around to yet.

7:19Now again, this stinks for those startups, many of whom will be forced to pivot or perish, but it is certainly better for the user experience to be able to use this feature natively from within the software itself. I know that as someone who interacts with a lot of PDFs and research papers via ChatGPT, I am quite excited about this one. Lastly today, an interesting little story about emotion-detecting AI. I actually love the anecdote that TechCrunch starts its article off with about this as well. They write, In 2019, Amazon upgraded its Alexa assistant with a feature that enabled it to detect when a customer was likely frustrated and respond with proportionally more sympathy.

7:54If a customer asked Alexa to play a song and it queued up the wrong one, for example, and then the customer said no Alexa in an upset tone, Alexa might apologize and request a clarification. Well, the article goes on that a non-profit called Laion, L-A-I-O-N, which has worked to build image and text datasets for training generative AI, including stable diffusion, has just announced what they call the Open Empathetic Project. Said Christopher Schumann, a Laion co-founder, the Laion team with backgrounds in healthcare, education, and machine learning research saw a gap in the open source community.

8:23Emotional AI was largely overlooked. Much like our concerns about non-transparent AI monopolies that led to the birth of Leon, we felt a similar urgency here. He continued, With Open Empathetic, our goal is to create an AI that goes beyond understanding just words. We aim for it to grasp the nuances in expressions and tone shifts, making human AI interactions more authentic and empathetic. So right now, the project is recruiting volunteers to submit and annotate audio clips that it can use to build this database. Said another open source contributor to the project, Kari Nori, We're driven by a clear mission to harness the power of AI in ways that can genuinely benefit society.

8:56We're passionate about transparency and believe that the best way to shape AI is out in the open. Anyway, super interesting project, one I'm going to be keeping an eye on. But for now, that is going to do it. Next up, the main AI breakdown. Hey guys, before we get into the main part of the episode, I wanted to put out a call for a few different types of people I am trying to hire right now. The TLDR is that I have a pretty exciting AI education-related project that I'm not really ready to share more details of just yet, but which suffice it to say I am incredibly excited about. On that front, I am looking for two types of people.

9:32The first is developers and UI slash UX designers, and the second is content producers, basically technical and non-technical builders. If you are either of those things and you want to learn more about what we're building, send me a note at jobsatbreakdown.network and please share examples of what you've built or created. So that can be websites or apps in the case of developers or UI UX designers, or it can be content that you've produced in the case of content creators. Again, that's jobsatbreakdown.network. Looking forward to hearing from you. And now on with the show. Welcome back to the AI Breakdown.

10:08Today, we are doing sort of part two of our coverage of President Biden's executive order on artificial intelligence. Now, on the one hand, it should be somewhat self-evident that the United States making its largest regulatory move vis-a-vis this space would be a topic of some import. But as you'll see today, this has generated an extreme array of very strong feelings, as well as debates around the likely impact of these new rules and actions to be taken following the AI executive order. Now, if you haven't heard about the details of the executive order and what's actually contained within it, I recommend yesterday's episode which goes over all of that in some length.

10:46Today, we will just focus on the reactions and what has generated the most controversy. But in order to help understand just how significant the White House is treating this, this was not just some signature and a press release alongside the rest of the day's actions. This was the biggest thing on President Biden's schedule yesterday. They invited folks to the White House, President Biden gave a prepared speech, and they even launched a new recruitment drive to get people to come work with the government on these AI issues. Now, some of the reactions were from notable folks in the political community.

11:17Former President Barack Obama tweeted, Artificial intelligence has the potential to change the way we work, learn, and create. I'm glad to see President Biden citing an executive order on AI designed to encourage innovation while avoiding some of the biggest risks. President Obama even followed that up with a full Medium post on the topic. In it, he echoed the administration's main message, which is, of course, that they want the U.S. to be the bastion of innovation while also fighting the potential risks. Now, one thing that's interesting in light of all of these, frankly, very frustrating articles about how all the AI doomers are distracting from the real and urgent risks of AI, as though people weren't able to handle multiple different types of risks at once in their puny little brains, Obama makes it clear that when it comes to risks, they are a both and.

11:59As he writes, We don't want anyone with an internet connection to be able to create a new strain of smallpox, access nuclear codes, or attack our critical infrastructure. But he continues, I'm glad the Biden administration is also thinking about other challenges that could end up being far more common. We've already seen what can happen when our shared basis of facts begins to erode. And indeed, this is much of where Obama spends his time in this article, concerned around the impacts to democracy of this technology and how it might accelerate transformations that have already started because of social media.

12:28Now, in addition to former politicians, current politicians are also making statements on the order, usually mostly in line with their already stated policies and mostly focused on the parts that have become their pet causes. And there was also a chorus of generic support from some of the big tech players. Brad Smith, the vice chair and president at Microsoft, wrote, Today's executive order is another critical step forward in the governance of AI technology. This order builds on the White House voluntary commitments for safe, secure and trustworthy AI and complements international efforts through the G7 Hiroshima process.

12:58AI promises to lower costs and improve services for the federal government, and we look forward to working with U.S. officials to fully realize the power and promise of this emerging technology. Salesforce's Mark Benioff writes, Today's AI executive order is a giant leap towards ethical AI, government integration, and attracting great AI talent to the U.S. We pledge to only use AI datasets that respect trust, privacy, and copyright. Your data will never be our product. Trust is our North Star as we navigate the AI landscape. Now, if that feels a little corporate buzzwordy to you, you are not the only one to think so.

13:27and this sort of generically positive response was far from the most common. The New York Post published an opinion piece from the American Enterprise Institute's James Pedacookis, who writes, Biden threatens to stifle U.S. tech innovation with a too-hasty AI power grab. He writes, Make no mistake, the greatest threat from emerging advances in artificial intelligence like ChatGPT isn't that AI will take all our jobs or kill us or both. The former is ahistorical, the latter science fiction. Rather, the big danger around AI is that over-eager Washington policymakers will rush to regulate a fast-evolving technology.

14:00Without a firm understanding of possible harms, we shouldn't risk slowing a technology with vast potential to make America richer, healthier, more militarily secure, and more capable of dealing with problems such as climate change and future pandemics. Tech progress delayed is tech progress denied. Now, the point that the piece makes is that it represents a fundamental shift in the way the U.S. thinks about technology regulation. In short, James writes, The order suggests nothing more than a wholesale abandonment of the light regulatory approach towards American digital markets that created a world where all the most important internet players are American companies.

14:32Had President Bill Clinton in the late 1990s issued such an aggressive and sweeping order, instead of opting for a lighter approach, it's possible we wouldn't be witnessing today's embryonic AI revolution. Now, this is a notable argument because it is explicitly the case that there are many politicians in Washington who do not see things this way and who do not view internet regulation as a great success in spite of the fact that all the big internet companies are here, as this op-ed points out. Indeed, if you ever watch a hearing about crypto or about AI or about any new technology, there is a broad sense among today's politicians that they lost control of social media and they want to get that right.

15:09And getting that right translates to asserting more power. Lawyer Preston Byrne makes a similar point. He writes, The AI executive order is the exact polar opposite of Section 230. If we keep heading down this course, government control and monitoring of private software development, we will lose the AI race. Jeff Amoco, the head of operations at Jensen AI writes, Biden's AI executive order is out and it's terrible for U.S. innovation. Here are some of the new obligations which only large incumbents will be able to comply with. If you're going to develop a large model, you must report to the government regarding any ongoing or planned activities related to training, developing, or pricing your model.

15:42Essentially, public company reporting for startups building large models. If you are developing a large model, you must report the ownership and possession of the model weights of your model and the physical and cybersecurity measures taken to protect those model weights. Implied restriction on open-sourcing model weights. Jeff continues, if you acquire a large amount of compute, you must disclose the existence and location of these clusters and the amount of total computing power available in each cluster, treating compute and inherently neutral technology as a dangerous resource that must be regulated.

16:08If you're an infrastructure provider, you must report to the government anytime a foreign person transacts to train a large AI model with potential capabilities that could be used in malicious cyber-enabled activity, i.e. new KYC regime for compute increasing costs of transacting with startups. Now, what I like about Jeff's thread, even if you don't agree with it, is that at least it points out some very specific issues with this particular EO rather than just being sort of generic in its critique. Y Combinator founder Paul Graham writes, I don't know what the right approach to regulating AI is, but one problem with this particular approach is that it means we're heading toward the government regulating private individuals computing at an exponential rate.

16:44Okay, so there are a lot of different buckets of this critique to start to pull apart. One has to do with the increased cost of compliance, which tend to benefit big companies at the cost of smaller startups. And another issue is this new surveillance regime surrounding access to computing resources. Now, responding to Paul Graham's post was the AI Safety Memes account who wrote, Can you elaborate? Americans have to inform the government when someone pays them$600. Why is informing the government about a planet-sized supercomputer viewed as tyranny? I'm libertarian sympathetic, but it seems many are responding with government always bad tribalism.

17:16Another AI safety advocate on Twitter, Liron Shapiro, writes, The reason the policy is drastic is because top researchers are admitting that the world might end. As soon as we have the theoretical foundations of how to create superintelligent AI and not end the world, we should relax the policy. Another AI CEO, Bindu Reddy, has an entirely different set of issues with the order. She writes, The AI executive order is a bit ridiculous and pretty hard to enforce. Here are the issues. One, any foundation model that poses a serious risk to national security. How do you determine if something is a serious risk to national security?

17:47If this is about misinformation, Twitter, YouTube, and Meta are way more serious risks to national security than harmless AI models that merely generate content, not distribute or amplify it. Two, all AI-generated content has to be watermarked? Seriously? We may as well kill Vision AI if we actually enforce that. Are enterprises allowed to use AI to generate images and use them in their marketing? Three, some agency is apparently going to develop tests that the AI models have to pass. How will this get enforced? Does every fine-tuned version of any open-source model have to pass these tests? What happens if a model is dropped on hugging face by a non-US corporation?

18:18Four, it has a bunch of privacy-protecting language. All this does is cripple AI models compared to search where the same privacy isn't being enforced. Google Search will recognize a face for me, GPT-4 Vision won't. It's not clear why my photo that is in the public domain can't be used by AI. We already have privacy-protecting laws. These should be sufficient. It then ends with some feel-good language around hiring an open source. Net-net, this order won't make any difference in the near term. It could be harmful in the long run to companies who do develop these foundation models, especially vision models.

18:45Weirdly, it helps search engines or makes them way more powerful because they are not regulated while AI models are. Now, after Bindu posted that, the full executive order came out, not just the fact sheet, and it showed that it was defining the reporting requirements around how much compute was actually being used. As Nir Sion points out, it's any model trained with around 28 million H100 hours, which is around 50 million USD, or any cluster with 10 to the 20 flops, which is around 50 ,000 H100s, which, as they point out, only two companies currently have. Now this prompted Bindu to actually go back and update her thoughts.

19:17She writes, the executive order may not have any impact on open source AI. At first glance, this appears to only apply to open AI in Google. Anthropic, Facebook, Microsoft, and maybe a couple of others may qualify as well. The good news is that smaller models seem to be exempt from these requirements, and this means that the open-source AI ecosystem is free to operate in an unregulated fashion, i.e. business as usual. The Mistral 7B is already super performant for its size. I anticipate that in a year or two, a 70B model or an ensemble of 70B models will beat, if not match, GPT-4 performance, at least on specific tasks.

19:48LLM training is becoming more and more efficient, and decent LLMs can be trained with less than$10 million in training budget, and this number is only going to go down over time. This also puts a lot of pressure on innovating and matching performance below these thresholds. I'm not entirely clear why these bigger models are greater danger to national security versus smaller models that may be purpose-built by someone to generate misinformation. This seems more about the GPU-rich fighting with each other than anything else. In a weird way, they are helping open-source AI ecosystem by not requiring the smaller models to comply.

20:15PyTorch co-founder Sumit Chintala made a similar point, writing, This executive order will introduce a new downward pressure on people and companies to make models smaller to avoid reporting, which in itself will likely bring novel effects such as innovations in model parameter efficiencies and such, but the innovation would strongly overfit to the exact formulation of a large model in the executive order. Now in terms of the startup AI labs, the only comment that I saw came from Jack Clark from Anthropic. He wrote, Seeing such a heavy emphasis on testing and evaluating AI systems seems good. You can't manage what you can't measure.

20:46Proof is obviously going to be in the implementation, but the EO gets a bunch of different parts of government to create capacity for third-party measurement, oversight, and analysis of AI systems, which seems broadly helpful. Also seems good to have the Department of Commerce referee process of creating guidelines and best practices so it's not just an uncoordinated bunch of industry players doing own self-governance initiatives. Also great to see a push towards implementing a pilot national AI research resource, aka National Research Cloud. Giving more computational resources to people who don't have it is a good way to deal with information asymmetries in AI development.

21:16Now one interesting subset of responses has to do with the contrast of this to the way that the White House and the broader government has dealt with the crypto industry. Attorney Jeremy Hogan writes, I couldn't stomach reading the executive order on AI, but I glanced through it in detail and the hypocrisy is enormous. Hasn't the administration attacked privacy in the crypto space at every opportunity? Why all of a sudden this concern over our privacy? Now interestingly, I think that the more relevant connection is actually around compute reporting. That seems to me markedly similar, as you heard from Jeff Amico a minute ago, to a new type of KYC but for compute.