LEAK: Google Gemini Ultra Coming This Week?

0:00Today on the AI Breakdown, presidential candidate Donald Trump says AI is dangerous and so scary. Before that on the brief, Google's Gemini Ultra appears to be coming this week. 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.

0:24Welcome back to the AI Breakdown Brief, all the AI headline news you need in around five minutes. Well, here is a big one that we are kicking off with. it appears that Google is planning to launch Gemini Ultra this week, and they're consolidating their brands around the Gemini term. So for those of you who don't remember, when Google announced its long-awaited Gemini model in December, we didn't get the most powerful version, the version that they said actually beat GPT-4. That was called Gemini Ultra. And while there was some grumbling that Google did its sort of announce a future thing without giving it to you thing, while Google did the thing where it announced something that you don't actually have access to, By and large, people have just been waiting to see how good Ultra actually is.

1:06Well, a couple days ago, Android developer Dylan Rossell writes, Google added a new changelog for BARD and oh boy, it's a big one. The changelog reads, 2-7-20-24, BARD is now Gemini. What? Gemini is the best way to get direct access to Google AI. All the collaborative capabilities you know and love are still here and will keep getting better in the Gemini era. We've also evolved the UI to reduce visual distractions, improve legibility, and simplify the navigation. Why? We're committed to giving everyone direct access to Google AI, and as of this week, every Gemini user across our supported countries and languages has access to Google's best family of AI models.

1:42To better reflect this commitment, we've renamed Bard to Gemini. Next bullet, try Gemini Advance to access Google's most capable AI model, Ultra 1.0. Gemini Advance gives you access to our most capable AI model, Ultra 1.0. If you want to be one of the first to access Google's latest AI advancements as they become available, this is for you. With our Ultra 1.0 model, Gemini Advanced is far more capable at highly complex tasks like coding, logical reasoning, following nuanced instructions, and creative collaboration. Plus, Gemini Advanced will continue to expand with new and exclusive features in the coming months, including expanded multimodal capabilities, even better coding features, as well as the ability to upload and more deeply analyze files, documents, data, and more.

2:20Now, it does appear that Gemini Advanced will be a paid plan as opposed to a free plan, similar to the distinction between GPT 3.5 and GPT-4. There's also a Gemini app coming. Now again, according to Google's own reporting, Gemini Ultra beats GPT-4 in things like the MMLU, in reasoning tests, in math tests, in coding tests, and it appears that as of a couple days from now, we will actually get to be able to test this out ourselves. Now in the meantime, one of the things that you might have experienced with ChatGPT last month, at the end of last year, was that it seemed to get, for lack of a better word, lazy in December.

2:57There was a whole discussion and debate around whether the LLM might have determined that December is a slower time of year, generally speaking, for work, and so was similarly shifting into a down gear. There was some evidence for that, although it was never proven or anything. But in any case, Sam Altman has said that GBT4 is not as lazy as it once was. He tweeted tongue-in-cheekly yesterday, GBT4 had a slow start on its New Year's resolutions, but should now be much less lazy. Now, what exactly they did to improve GPT-4's laziness was not revealed, but certainly for anyone who has been frustrated over the last couple months, this will come as a welcome announcement.

3:34Now, one more interesting thing in the competitive set. As you've heard on this show, Meta is increasingly talking like they want Llama 3 to be competitive and state-of-the-art not just for open-source models, but for models in general. Zuckerberg said literally that when he made an announcement about their AI strategy a couple of weeks ago. Well, Andrew Curran pulls something pretty interesting from last week's earnings call, where he writes, It was instantly overshadowed after the stock went hyperbolic, but during the earnings call, Mark Zuckerberg mentioned that the gargantuan trove of public data that they have amassed from Facebook and Instagram to train Llama 3 exceeds the common crawl.

4:09So, in excess of 6.4 petabytes. The exact pull quote from the transcript was this. Zuckerberg says, Now the next key part of our playbook is learning from unique data and feedback loops in our products. When people think about data, they typically think about the corpus that you might use to train a model up front. And on Facebook and Instagram, there are hundreds of billions of publicly shared images and tens of billions of public videos, which we estimate is greater than the common crawl data set. Now, Zuckerberg also went on to say that even more important than the upfront training corpus is the ability to, quote, establish the right feedback loops with hundreds of millions of people interacting with AI services across our products.

4:44but still I thought it was an interesting note on just how much data this company has access to when they're thinking about how they make LAMA 3 state-of-the-art. There was also an interesting chart that Brian Romley pointed out that shows a dramatic shift in meta stock price, seemingly corresponding with the beginning of their focus on AI and open source at the beginning of 2023. Now, of course, there is a correlation causation thing, as 2022 was in general a brutal year with the fastest tightening cycle in Fed history, But still, it is interesting to note. Now, one interesting one in the open source world, speaking of that, OpenAI's custom GPT builder has some competition.

5:21Tech lead and LLMs at Hugging Face, Philip Schmidt, says, Introducing Hugging Chat Assistant. Build your own personal assistant in Hugging Face chat in two clicks. Similar to OpenAI GPTs, you can now create custom versions of Hugging Face chat. An assistant is defined by name, avatar, and description, any available OpenLLM like Llama2 or Mixtral, custom system message to control the behavior, and different message starters. So this really is very similar to an open source version of the custom GPT builders, in that in many ways the point of it appears to be giving people the ability to create a saved prompt structure for common workflows that they use over and over again.

5:58In other words, I think that in many ways the use cases of these GPTs, despite there being a store, tends to be pretty personal and just time-saving by giving people the ability to recall without having to reprompt based on some previous effort. Still, I think it's a really cool thing to see more GPT-like builders, and I'm excited to try it out. I will almost certainly be doing so for our AI education beta, which continues to go on right now. However, that is going to do it for today's AI Breakdown Brief. Next up, the main AI breakdown. Welcome back to the AI Breakdown. Well, given that Joe Biden has been president the entire time that generative AI has been in the public eye, we have a fairly decent sense of how the Democrats will think about and will handle artificial intelligence should they win the White House once again.

6:43The executive order, which was one of the longest, if not the longest in history, covered many, many different aspects of artificial intelligence. And given that that executive order just hit its 90-day mark, we're starting to see how those policies might be brought to bear. Some of the most recent news, for example, is that the Department of Commerce will be requiring companies to disclose when they are training LLMs. They will also be asking cloud providers to let them know when international customers are using lots of compute, potentially for training LLMs and other advanced AI models. And additionally, if you listen to my episode over the weekend, which was a reading of an essay from Chief Economic Advisor Lael Brainard, you started to get a sense that perhaps from an election standpoint, Democrats were also going to use AI as a way to reignite labor, more specifically the labor discussion as it relates to politics.

7:29Basically, the point of that letter, if you haven't listened, was that for AI to work for America, AI has to work for American workers. They called out some specific ideas of what that might mean, but by and large, it was a positioning statement more than a policy statement. Now, it should be noted that when it comes to specific types of legislation, artificial intelligence has not fallen strictly into partisan lines yet. When it comes to the difference between more restrictive comprehensive legislation and less restrictive comprehensive legislation that has been proposed, none of which, by the way, has gotten anywhere, both have Republican and Democrat co-sponsors, meaning that at least for this time, there is bipartisan interest in AI, and where people land on exactly what we should do from a regulatory standpoint is not totally based on one's party.

8:12This is refreshingly anomalous in some ways, with Bitcoin or crypto being the only issue that I've seen that up until recently had a similar sort of bipartisan pattern, although that has hardened over the last couple of years. Given all this, I have of course been paying close attention to the early campaigns around the U.S. 2024 presidential election to see when AI was going to become an issue. Thus far, it has largely been theoretical. In other words, a theoretical concern about what might happen if there is widespread misinformation or disinformation campaigns leveraging AI. And only recently, with the Biden-AI robocall in New Hampshire, did it become a little bit more real.

8:46You'll remember a couple weeks ago during the New Hampshire primary, that state's Democrats got a message from someone that sounded like Joe Biden saying that they shouldn't go out and vote because it wasn't important until November when the actual general election was being held. Now, when it comes to the Republican contender for the presidency, pundits are treating it as all but locked up for Donald Trump. Technically, Nikki Haley is still in the race, but for many, Donald Trump is once again the presumptive nominee. For the first time, at least recently, Donald Trump has spoken about AI and did so in an interview on Fox News.

9:17It's a short clip, so rather than summing it up, I'm actually going to just play about a minute of that conversation. Maybe the most dangerous thing out there of anything, because there's no real solution, the AI, as they call it, it is so scary. I saw somebody ripping me off the other day where they had me making a speech about their product. I said, I never endorsed that product. And I'm telling you, you can't even tell the difference. It looks like I'm actually endorsing it. No, because you can get that into wars and you can get that into other things. Something has to be done about this, and it has to be done fast.

9:53And nobody really knows what to do. The technology is so good, and it's so powerful, that what you say at an interview with you almost doesn't matter anymore. They can change things around, and nobody can tell the difference. Even experts can't tell the difference. This is a tremendous problem in terms of security. This is a problem that they better get working on right now. So a couple things to note about this. One is that deepfakes are definitely making the risks of this technology accessible to politicians. In other words, it helps dramatize the potential issues. It sounded like Trump had seen a deepfake of himself endorsing a product, and that's what got this issue on his mind.

10:32A second thing to note is, like him or loathe him, I think Trump is correct when he says that no one really knows exactly how to deal with this issue. That there isn't really a consensus solution. Whoever is going to end up as president next cycle is going to have to deal with that reality. Now, of course, this is a very limited set of information about what he thinks about AI. This doesn't get into the China dimension, which you have to think is a huge issue. It doesn't get into AI in the military, which once again is going to be another key question for any future president. But I think it's fair to say that with these comments, AI has broken onto the national agenda as a potential election issue.

11:09Now, certainly in Washington, D.C., the discussion around AI is getting more and more important. Open Secrets on behalf of CNBC found that between 2022 and 2023, AI lobbying was up 185%. More than 450 organizations participated last year, including all of the big companies, the ByteDances, the Teslas, the Palantir, NVIDIA's Open AIs of the world. Now, lobbying obviously always gets a dirty look and a scowl from the American public, not without justification. However, a big increase in lobbying does not necessarily mean that the AI industry is all of a sudden disproportionately spending. AI has become a much bigger issue over the period of time that this data was analyzing.

11:46I think it's probably to be expected that the companies that have the biggest stake in how this shakes out are going to try to get themselves involved in the regulatory discussion. Like I said, given that there isn't a clear partisan direction for policy, that means there's a lot more openness to different directions that it could go, which only increases the incentive to actually be aggressive about lobbying efforts in the interim. What's more, my guess is that most of these organizations have identified that there may be a window of time where they can influence politicians without there being a lot of additional action happening, given how little momentum there tends to be around big issues like this in an election year.

12:19Meanwhile, over in Europe, the EU's AI Act continues to move through the political process. The information writes, Member countries of the European Union on Friday unanimously approved the final text of the AI Act, a wide-ranging framework for regulating applications of artificial intelligence. The vote followed months of debate and lobbying as countries such as France and Germany pushed back against laws they said would stifle their budding AI startup ecosystems. Now from here, EU lawmakers vote on the bill on February 13th, and then in March or April, the European Parliament will vote. Given that we now have complete text, I expect a lot more analysis in the coming days.

12:55Now, meanwhile, speaking of the intersection between the private sector and the public sector, white-hot NVIDIA sees one of its new sources of demand as actual nation-states. In an interview last Thursday with Bloomberg Television, NVIDIA CEO Jensen Huang said that countries ranging from India to Japan to France to Canada are talking about the importance of investing in, quote, sovereign AI capabilities. He said, Their natural resource data should be refined and produced for their country. The recognition of sovereign AI capabilities is global. Continuing, he said, The vast majority of the computing market has been in the U.S.

13:28and to a much smaller degree China. For the very first time, because of generative AI computer technology, it's going to impact literally every single country. So some of the markets will be quite large and global. And this, I think, is what's going to make some of the election happening so interesting, is this tension between, on the one hand, wanting to regulate AI, and on the other hand, wanting to harness it. How countries deal with that and make sense of that tension could be a big determinant of future competitiveness in a variety of ways. However, that is going to do it for today's AI Breakdown.

13:57I appreciate you listening or watching as always. And until next time, peace.

14:06Thank you.