The 5 Most Important AI News Stories Last Month

0:00Today on the AI Breakdown, we're counting down the 5 most important news stories from AI last month. 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 YouTube, and our newsletter. Welcome back to the AI Breakdown. Today we are doing a quick recap of the five most important AI news stories for July. And in some ways, this was an interesting month. I think as you'll see from these stories, there was a lot of very significant things that happened. However, at the same time, there was a sense I think broadly felt that at least some amount of the hype had worn off.

0:40That maybe for the first time since ChatGPT launched, there were counter-narratives starting to emerge. Questions around not so much whether artificial intelligence was going to be important and whether generative AI was a big significant disruptive force, but perhaps questions around just how fast those shifts might happen. For that reason, the first honorable mention of two honorable mentions is the fact that ChatGPT saw its first user decline in June. Now, this is based on mobile and desktop traffic and measured by SimilarWeb. And basically, a growth curve that had already been flattening between March and April and April and May actually went into outright decline between May and June.

1:19June's traffic numbers were somewhere between the March and April numbers. Now, one of the big explanations for this was, of course, that school is out, meaning that an entire category of use cases for ChatGPT having to do with education were just simply not being used. How much one interprets that as a normal seasonality versus a sign that students are using ChatGPT to cheat is a little bit in the eye of the beholder. My argument has simply been that we don't really have a lot of comparisons to look at this against. ChatGPT was the fastest ever technology to get to 100 million users by a significant margin.

1:51And so what that adoption curve is going to look like and how fast they could get to saturation is just a little bit TBD. However, I do think it's fair to say that at least the first phase of the novelty phase has worn off. And we are now firmly in the is this actually useful to me phase, which when it comes to the long term of AI is much more interesting. Our next honorable mention is a combined story that I think really goes together. Part A of that story is the White House securing a set of voluntary commitments from AI companies including Meta, Google, Microsoft, Amazon, OpenAI, Anthropic, and Inflection.

2:25This was announced a couple Fridays ago and is sort of a symbolic step one in the White House's approach to AI regulation. Rather than viewing these commitments cynically as just companies giving lip service to something that the administration wants to hear, I think a better way to look at it is to understand that the White House knows that legislation, even when it moves fast, is still going to move a lot slower than these companies. By securing even these basic commitments, the White House, A, puts themselves in a leadership position and an active conversation participant when it comes to AI, and B, if nothing else, gets these companies to consider for a moment some countervailing pressures to the endless market-driven arms race that is, as some like Jeffrey Hinton have accused, making companies not take seriously their ethical, responsible, or safety obligation when it comes to this new technology.

3:13Now the companion, of course, to that story is the Frontier Model Forum that was announced just a couple days later. This is an industry body that takes those commitments just a little bit farther by creating an organizational context to actually try to hold member organizations accountable for that type of commitment. So far, the Frontier Model Forum includes just four of the companies that made the White House commitments, including Anthropic, Google, Microsoft, and OpenAI. Now, the reason that these two things get honorable mentions is that even though I do think that they are more meaningful than just a cynical headline grab, they are still ultimately words, not actions.

3:46And so when it comes to actions, we're going to have to wait a little bit to see if these companies actually put their money where their mouth is. With that, let's move on to our main list. Coming in at number five, most important AI news for the month is Elon Musk announcing XAI. XAI makes good on Elon's long history of chatting about how he would do AI differently and his critiques of companies like OpenAI, which he has a serious problem with how they've evolved. And in the middle of this month, he announced XAI as his approach to the field. Now, when it comes to what XAI is going to do, details were scant, but ambition is very clearly high.

4:23In their announcement posts and Twitter spaces explaining the company, they talked a lot about things like, to use Musk's phrase, understanding the true nature of the universe. Elon's argument is that by making artificial intelligence maximally curious about the universe around it, that's the best way to ensure that it doesn't try to kill us all. Effectively, he argues that a world with humans in it is more interesting than a world without humans in it, and so if you make AI curious, it will have an incentive to keep humans around. Now, some are skeptical, of course. Any Elon announcement at this point is mostly just a Rorschach test for how you feel about Elon.

4:57but some are cautiously optimistic. An op-ed in The New Scientist wrote, Could Elon Musk's XAI be exactly what the world needs? As academics struggle to compete with private investment, perhaps Musk's new artificial intelligence venture really can tackle the, quote, true nature of the universe. Now, ultimately, this comes in at number five, because as much promise and potential as it has, so far it is still just an announcement, and a team which is admittedly full of some serious firepower. Number four on our list of most important stories from July is Anthropic's announcement of Claude 2. Now, a month before, Anthropic had made news when they announced a 100K context window, as well as their new constitutional model of trying to align AI around a series of core principles, rather than a reinforcement learning process that goes situation by situation.

5:44All of that came together in Claude 2, which is their more advanced new model. Now, on the one hand, of course, everyone is focused on things like comparison to GPT-4 and GPT-3.5, and in many tests, Claude 2 ranks alongside GPT-3.5, but the real big value proposition, at least so far, is that 100k context window. Simply put, it is allowing for a different set of use cases, or at least an easier process for a set of use cases that involve long documents, comparing documents, and really anything that is longer than what ChatGPT can currently ingest. Now the question, of course, for Anthropic and Claude is whether ultimately those longer context windows just become table stakes, and if so, what other advantages they might have.

6:24Still, the difference between the beginning and the end of the month is that there are a lot more people using Claude, or specifically Claude 2, for things that they used to be using ChatGPT for. Still, even in a month that Anthropic launched their most advanced model with these big advantages over GPT, OpenAI couldn't let them have all the fun, because this month the much-anticipated Code Interpreter was finally released. Code Interpreter is a massive upgrade, certainly the biggest upgrade since GPT-4, and in many ways is even bigger in terms of expanding what ChatGPT can do. ChatGPT now has a native way to handle math, data, coding, and more in ways that are producing some pretty phenomenal results.

7:01Now, as some like Swix here have pointed out, Code Interpreter changes the functionality of ChatGPT so fully that people are actually shorthanding it as GPT-4.5. He posted a common code experiment that he did across the two models, ChatGPT-4 as well as Code Interpreter, and showed just what a massive difference it made in terms of the quality and the utility of the code. Now, I didn't want to give OpenAI another slot here, but I do think that complementing Code Interpreter is ChatGPT Custom's instructions feature. Custom instructions gives ChatGPT Plus users the ability to effectively pre-prompt all queries that they make with ChatGPT, giving the model persistent information about one, who they are and what they do, i.e.

7:38context for what answers one might want, and then two, the format of how they want answers outputted. For example, right now I have my default output set to bulleted lists with no more than one to two sentences per bullet, and at an intellectual level, suited for an undergraduate college audience, which I find is the best for taking really advanced concepts and explaining them to people who aren't necessarily technical. Now another reason that I didn't want to give ChatGPT's custom instructions its own slot is that I knew OpenAI had another point on this list, because number two on our list of most important stories from July in AI is the introduction of OpenAI's super alignment initiative.

8:13This is an initiative to try to align AI, to try to figure out specifically how to align superintelligence that is much smarter than people. Three things that were super notable about the initiative. One is that they set themselves a timescale. Our goal, they write, is to solve the core technical challenges of superintelligence alignment in four years. Number two, they put co-founder and chief scientist Ilya Sutskever on this. He will be co-leading the team with their head of alignment. And number three, and maybe most significantly, OpenAI has committed dedicating 20 % of the compute they've secured to date over the next four years to solving this problem.

8:49Now, given how scarce compute resources are right now, I frequently bring up how much it's impacted, for example, OpenAI's ability to release new features for ChatGPT this year. Dedicating 20 % of compute is a serious put your money where your mouth is type of moment. It's why the Super Alignment Initiative is here at the top of this list as opposed to the White House commitment and the Frontier Model Forum, which are still mostly about words. This is an incredibly ambitious project around an incredibly important goal, and so comes in at my number two slot for most important news of last month.

9:23And of course, that brings us to number one. Number one with a bullet. That is Meta's announcement of Llama 2. If ChatGPT got the whole generative AI space cooking last November, from a technical perspective, a lot of 2023 has been defined by Meta's open source engagement. When the full Llama model leaked in March, it set off a huge amount of development that ultimately led to that Google internal memo saying that open source developers were effectively eating OpenAI in Google's lunch, that they were churning through problems that Google had anticipated lasting for years in a matter of months, weeks, or even days.

9:59However, what Llama 1 was missing was a commercial use mode. It was only available for research. And so to the extent that people used it for things that were commercial, they were risking incurring the wrath of the Zuck. With Llama 2, there is no such problem. And regardless of some amount of controversy around whether Meta's Llama 2 is actually fully open source, based on the purest definition of that word, there's no doubt that at this point, for many in this space, and especially developers. The real open AI company is not open AI, but meta. The launch of Llama 2 is one of those news stories that is not just important when it happens, but will have huge ramifications for how the entire space develops.

10:36And that is why it is our most important story from July. Anyways, guys, that is going to do it for today's AI Breakdown. If you enjoyed this video, hit the subscribe button and then hit the notification so you don't miss any others. I appreciate you listening or watching, and until next time, peace.

10:55Thank you.