US Gov't Says AI Labs Need to Notify Them Around LLM Training

Today on the AI breakdown, we're looking at Elon Musk's role in AI before that on the brief, the US government announces that OpenAI and other AI labs have to notify them about large language model training. 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 YouTube, our Discord and our newsletter. The US is using the Defense Production act to require companies to notify them when they're training large language models. Welcome Back to the AI Breakdown Brief. All the AI headline news you need in around five minutes. One of the big themes for 2024 was always going to be the growing importance of the conversation about artificial intelligence in Washington, D.C. of course, last year we got President Biden's executive order and some of the fruits of that are starting to bear now. U.S. secretary of Commerce Gina Raimondo has announced more details about requirements for companies training large language models, with the government saying that the Defense Production act authorizes them to get notification when those companies are training advanced models, said Raymondo at an event at Stanford University's Hoover Institute last week. We're using the Defense Production act to do a survey requiring companies to share with us every time they train a new large language model and share with us the results, the safety data, so we can review it. Now. Raimondo tried to minimize how burdensome this sort of notification requirement is going to be. Calling it, quote, just a quick heads up, we're building powerful LLMs that might be a major national security risk. Relatedly, Raymondo said, quote, we're beginning the process of requiring US Cloud companies to tell us every time a non US entity uses their cloud to train a large language model. Now this is something that's gotten increasingly popular and I think is going to be an even bigger discussion going forward, which is monitoring cloud usage as a way to keep tabs on who is potentially training advanced models. Now, in some ways these might seem simple, but they are likely to be controversial, so I will keep you posted as discussion about these new policies kicks up. Meanwhile, over in China, it's not just notification, but approval that is required by the government before companies can release their AI models to the public. Last week, Chinese regulators granted approval to a group of 14 new LLMs for public use, which represents the fourth batch of approvals that China has granted. Overall, China has approved more than 40 AI models in the first six months since they began this approval process. It appears that so far the most used Chinese chatbot is Baidu's Ernie Bot, which according to that company CTO has more than 100 million users. In terms of what's different about this new batch, it appears that it's not just generalist models, but a number of them are actually industry specific LLMs as well. Those include LLMs focused on cybersecurity and video solutions. Now, speaking of LLMs, LIMSYS Chatbot arena is a crowdsourced platform for LLM evaluations that has collected over 200,000 human preference votes to rank LLMs with an Elo rating system. A big jump on that arena leaderboard has come from Bard's new Gemini Pro version, which has jumped above everything except GPT4 Turbo. Currently the rankings are at 1 GPT4 Turbo at 2 Bard with Gemini Pro at at 3 GPT4.0314, at 4 GPT40613, at 5 Mistral Medium at 6 Claude 1 and at 7 Claude 2. Now as many pointed out, Gemini Pro is not Gemini Ultra, and so for many this has reignited excitement about what Gemini Ultra might actually be able to do. However, some aren't buying it. Anjani Mitha from A16Z writes, Bard is the only model in this list with Internet access, which the table should make clear. Open book versus Closed book exams should be graded differently. Moving back to a story from last week, you'll remember me discussing a robocall that New Hampshire Democrats got from what turned out to be an AI synthesized President Biden urging them not to vote in that week's election primary, saying that their vote only mattered in November. Well, a cybersecurity research firm, Pin Drop Security, was able to identify that it had been 11 labs platform that had been used to make that voice. And and once Pindrop Security shared that information with 11 labs, they were able to figure out who the user was and suspend that account. Expect this process to play out a lot more during this election season. A new interesting AI feature from Apple, Apple Podcasts, is now offering AI generated transcripts. This is for the upcoming iOS 17.4, which is expected to come to users in March, writes Apple Insider. The feature is new and still in beta, so there isn't any information on implementation or expected accuracy. Users should expect most, if not all, shows delivered through Apple's podcast network to be transcribed. For what it's worth. As a person who maintains multiple podcasts through Apple, we were all asked to sign new terms of service that reflected these changes. So yes, indeed, transcripts are coming to Apple Podcasts now. Speaking of podcasts, one very cool company in The AI podcast and media creation space is Wondercraft. Wondercraft is in some of the same space as companies like Elevenlabs with some overlapping features, but is focused on being a really wonderful tool for creating audio content and then now also turning that into video content. The company has just announced a $3 million round and I'm excited to see what they do with those resources now as we get used to new tools and new interfaces. Obviously a lot of people have switched their research behavior from Google to Perplexity. Well now next generation browser ARC has decided to make Perplexity the default search engine option. In a blog post on the Perplexity site they write, by combining Ark's minimalist interface and Perplexity's intelligent search, we've created a streamlined browsing experience. No more sifting to irrelevant results or dead end links, just fast, straightforward access to the information you need. Speaking of new features, Midjourney V6 has been out for a little while now, but they have continued to announce new features and just recently they've released a set including panning, zooming and varying part of a region which lots and lots of the AI artists out there are getting very excited about now. Lastly today an interesting one about AI infrastructure. Blackstone is a massive private equity firm that manages basically a trillion dollar empire, and in 2021 the company spent $10 billion taking over data center operator QTS. At the time the company had about a billion dollars worth of property under development, which has risen now in just a few short years to $15 billion thanks in large part to increased demand for lease data center capacity because artificial intelligence Blackstone is now reporting that QTS could be one of the best investments in their history. So there you have it folks, lots and lots of interesting things to start your week. That's going to do it for today's AI Breakdown brief. Next up, the main AI breakdown. Today we are talking Elon and AI welcome back to the AI Breakdown. Today we are talking about Elon Musk and his relationship to AI. And the catalyst for this is a set of comments and some recent actions that show the diversity of the places that he inhabits in this industry. And I thought it was worth exploring a little bit more about all of those diverse connections and giving you sort of a primer on where he sits relative to this field. So let's talk first about those comments and those actions, starting with a recent discussion around how much Tesla would be spending on Nvidia chips and how much it takes to just compete in AI in general. In a post on X Elon responded to another post saying the governor is correct that this is a Dojo supercomputer, but $500 million, while obviously a large sum of money, is only equivalent to a 10K H100 system from Nvidia. Tesla will spend more than that on Nvidia hardware this year. The table stakes for being competitive in AI are at least several billion dollars per year at this point. Now those table stakes comments have been reported widely and it reflects something that we've talked about a lot here, that the scale of capital needed to compete in this space is just so different than any technology startup industry we've ever seen. It's forcing different dynamics upon the industry because venture capital alone can't cover the need. It's one of the reasons that big tech companies are so intrinsically involved even at this early stage. Now when someone asked him if they also plan to buy chips from amd, he said yes, but that's not really the main point here. Now the other interesting event that shows Elon in the context of his ex leadership is around the viral NSFW Taylor Swift AI deepfakes that were released and distributed on that platform last week. Now we talked about how the Swifties had risen up last week and tried to push down the illicit results by sharing legitimate photos of Taylor with the same hashtag as well as a new hashtag, something like Protect Taylor Swift. But Elon and X took it a step farther and just started to block searches for Taylor Swift. So if you searched Taylor Swift over the weekend you would get a message that said something went wrong. Try reloading. And although X didn't make an official statement to reporters, its head of Business operations Joe Benarok told BBC News that the action was taken with an abundance of caution as we prioritize safety on the issue. On X's safety account they said posting non consensual nudity images is strictly prohibited on X and we have a zero tolerance policy towards such content. Now as we discussed last week, this is definitely putting the issue of deepfakes and non consensual nudes even higher on the political agenda than it had been. I basically said last week that if Taylor Swift couldn't get people paying attention to this issue, that nothing could. One of the things that this X story shows though is that there are going to be multiple dimensions to this. There is not just the image generation side guardrails which of course the major platforms like ChatGPT can do, but is much harder if you just implement an open source version of stable diffusion or something like that, but there's also guardrails on the distribution side. It kind of makes it clear that distribution and creation are going to have to work together on these policies for them to be actually effective in any meaningful way. There's also almost inevitably going to be a bit of whack, a mole on the distribution side, and so response time is going to be paramount as well. Now, another interesting story around Elon and AI has been the question of whether X AI is raising external capital or not. This has been going on for a month or so, but the most recent version of it is that on January 26th the Financial Times reported that X AI was seeking to raise $6 billion at a $20 billion valuation. They reported that Elon had been courting wealthy individuals around the world, including family offices in Hong Kong as well as sovereign wealth funds in the Middle east, to see if there was appetite for that sort of fundraise. But then again, after this Financial Times piece went up, Musk took to Twitter again to say xai is not raising capital, and I have had no conversations with anyone in this regard. The previous week he had also denied a report that the company had secured $500 million in commitments towards a $1 billion fundraising goal. Now, one thing that Elon did say about GROK in a response to Jordan Peterson on x was that Grok 1.5 would be coming sometime in the next month. Now, as the FT points out, one of the things that could be interesting is, or challenging really for Elon is if he really is trying to raise money in Hong Kong and even in some parts of the Middle east, that could run afoul of increasing tensions between those regions in the US when it comes to AI policy. That alone could explain why Elon is keen to tamp down any rumors before any deals are done. Going back to our theme from right at the beginning, when Elon talked about the poker analogy and the table stakes of spending several billion dollars just to compete, the FT writes the scale of the attempted fundraising reflects the enormous cost required to develop generative AI, which requires huge computing power, vast amounts of data and cutting edge chips. Now, one thing that could defray some of those costs are tax breaks. It's not Xai, but X itself, aka Twitter, was just granted a $10 million tax break for hardware in Georgia. Basically, the company wanted to install new AI hardware at its data center environment in a QTS facility. Qts, by the way, being the company that we just talked about in the brief and this week the Development Authority of Fulton county voted 62 in favor of granting a 10 year tax break on that $700 million project. Wrightsdatacenterdynamics.com the social networking firm formerly known as Twitter said the computer infrastructure would be used to develop and train artificial intelligence products for the X platform, including large language models and semantic search. Now Twitter X had threatened to move the data center to Portland, Oregon if the tax break didn't come through. Now of course the one other big interesting Elon story of late has been the way that he's been trying to leverage artificial intelligence to get more of Tesla. Responding to another post on x back on January 15th where someone had basically said that Elon shouldn't need more of an incentive to keep building Tesla out, Elon responded I'm uncomfortable growing Tesla to be a leader in AI and robotics without having around 25% voting control enough to be influential, but not so much that I can't be overturned. Unless that is the case, I would prefer to build products outside of Tesla. You don't seem to understand that Tesla is not one startup, but it doesn't simply look at the delta between what Tesla does and gm. As for stock ownership itself being enough motivation, Fidelity and others own similar stakes to me. Why don't they show up for work? I should note that the Tesla board is great. The reason for no new compensation plan is that we are still waiting for a decision. In my Delaware compensation case, the trial for that was held in 2022 but a verdict has yet to be made. I put compensation plan in quotes because from my standpoint this is primarily about ensuring the right amount of voting influence at Tesla. If I have 25% it means I am influential but can be overridden. If twice as many shareholders vote against me versus for me at 15% or lower. The for against ratio to override me makes a takeover by dubious interests too easy. I would be fine with a dual class voting structure to achieve this, but I am told it is impossible to achieve post IPO in Delaware. So basically what Elon is asserting here is that the stakes are so high around AI and robotics that he feels that he needs to ensure that it is very difficult, although not impossible, to overrule him when it comes to key decisions. This of course brings up a question around concentration of power in AI. There is clearly concentration at the top of these big AI labs. Zuckerberg has always had incredible control at Facebook, now Meta and the whole escapade with Sam Altman being forced out and then invited back at OpenAI shows just how much control he has there, even though it's such a different type of environment. The question, of course, is it just bluffing? What would it look like for Elon to not grow Tesla to be a leader in AI and robotics? There are a lot of people who have put a lot of money into the idea that Tesla is the single most important real world AI and robotics company that exists. So does Elon really have the option of not growing it into being that leader? Just one of the many fascinating dynamics around this guy that make him such a pivotal, like it or not figure in the AI space. And of course we haven't even really got into his AI safety perspective where he is purportedly much more on the concern for humanity side of things than the effective accelerationist side of things. For example, if you haven't read it yet, I highly recommend you go read the New York Times piece back from December called Ego, Fear and Money how the AI Fuse Was Lit. It tells the story of Elon and Larry Page's breakup of their friendship really around this AI question and could give you a little bit more insight into a man that you will continue to hear about whether you want to or not. However, that is going to do it for today's AI breakdown. I appreciate you listening or watching as always and until next time, peace.