The AI Daily Brief: Artificial Intelligence News and Analysis - Creating Worlds From Words: The Most Essential AI News This Week

Episode Date: June 10, 2023

Today on the Weekly Recap, NLW looks at the most essential AI news stories across new tools, important research, big company announcements, open source updates and societal impact.  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 doing the weekly recap of the most important news and stories from this week, from new technologies to groundbreaking research and beyond. The AI breakdown is a daily podcast and video about the most important news and stories in AI. Like, subscribe, and share, and go to Breakdown.network for more information. Welcome back to the AI breakdowns weekly recap. These are the most important news stories. And this week I've divided it into five categories. The first is the cool new tools that have come out. The second is important research. The third is the latest developments from the big companies competing for AI supremacy. The fourth is open source, the companies and projects that are trying to make sure that the big

companies aren't the only ones having fun. And finally, society might catch-all category for policy, politics, alignment, et cetera. We kick it off in our tool section with Google's style drop. This maybe could have gone into the research section as well as it's not exactly a product yet. But basically, this is a new text to image generation method from Google. It allows you to specify a specific reference image from which style drop will then generate images in that style. So you can see this watercolor style, a style that's like Vincent Van Gogh's Starry Knight, a Vincent Van Gogh self-portrait style, a reference image that's kind of like a Kitch Pacific Northwest red-themed line drawing, a genericish computer graphic style that has a blue sky background, and so on and so forth, all these different styles. Now, the impressive thing is that their methodology is able to deliver this type of results, even with only a single single,

reference image. And so what has people excited about StyleDrop are the implications for creative products where you need a consistent style throughout or branded projects where you need to follow specific brand guidelines? This is the type of project where you see it and you think that even the incredibly magic text to image generation that we have now is going to feel so nascent in just, I don't know, months to come from here. Next, Taffy is a new Texta 3D character engine. This comes from Days 3D and does exactly what it sounds like. It allows people to use text inputs to create characters that could go live in 3D worlds. Now, this is part of a broader trend that I've been seeing of the inputs to games and metaversal worlds creation being radically

democratized in the era of AI. This is going to significantly increase the ease with which creators can develop characters for all of these new digital spaces, and I am excited to see what people do with it. Speaking of digital space is something that people have been doing a lot lately is using Adobe Photoshop's new generative fill tool to expand the boundaries. of internet memes and of famous works of art, and Stability AI has now developed a new feature that they call Uncrop for their ClipDrop app that basically turns that into a single clickable thing. So this is the background expander, but as an entire tool. This one, I'm much less clear if this is just a fun novelty capturing the zeitgeist,

or if it's something that people will actually use pretty frequently, but it's fun either way. Next we have Promptor, and this gets to something that I feel strongly about, which is that one of the under-heralded use cases for AI is just as a human brainstorming partner. A lot of the way that I use Mid Journey or ChatChap ETT is to try to expand the way that I'm thinking about particular ideas or creative pursuits. Promptor is an open-source prompting tool that does exactly that. You basically create a master prompt where you specify the areas where there could be variety, and then Prompter will give you just a boatload of different prompts that you could use

in things like Mid Journey Stable Diffusion and more. You can also explore lists that help you get ideas for prompts. And overall, it's just a fun little tool for expanding human creativity. Lastly, though, there is no doubt that this week, the big new entrance on the tool scene is runway gen 2. Runway is text video. Their gen 1 has already been used significantly. In their gen 1, it was mostly taking existing video and reapplying different styles to it, which is impressive enough as it is. But Gen 2 is a full text of video experience.

You get four second clips out of time. time, you can specify styles, you can use reference images. And while this has been out for about a month in beta, it is now available to everyone. People are already over Twitter giving their best tips for how they're getting better results in it. People are sharing the creations that they've made by putting together multiples of these four second clips. And I think it's best summed up by Shabnam Sabu, who says, Gen 2 feels as addictive as mid-journey in its early days. Can't stop creating new AI videos. Next up, we move over to the realm of research. And the project that caught my eye most, with something called LTM1.

LTM 1 is an LLM with 5 million prompt tokens. That's the equivalent of 500,000 lines of code, and it's an entirely new network architecture that is designed from the ground up for these large context windows. Now, context windows are one of the big limiting factors of LLMs. Currently, chat GPT, for almost everyone, has an 8K context window,

which is the equivalent of about 4,000 or 5,000 words. Anthropic recently announced a 100,000 token window, which is the equivalent of about 75,000 words or one great Gatsby. And obviously, the expanded context window gives the AI the ability to have a lot more context as it's trying to answer your particular questions as relating to what you've inputed. And the developers of LTM1 say that they have a long way to go. They write LTM net see more context than GPTs, but LTM1 has fewer parameters than today's frontier models, making it less smart.

So figuring out how to expand that is their next challenge. Another really interesting bit of research this week comes from Google's deep mind. They've just published research in the journal Nature about what they call Alpha-Dev. Alfa-Dev is an AI system that uses reinforcement learning to discover computer science algorithms from the ground up. Basically, what they did is they took their AI, which had been previously used for learning games. So AlphaGo, you'll remember, beat the Worldwide Go champion. And they created effectively a game in discovering key computing algorithms. What came out of that were new and improved algorithms for sorting, which is the type of thing that is endemic in social media.

and it also found new ways to improve hashing by up to 30%. Google DeepMind writes, rather than refine existing algorithms, AlphaDev started from scratch in a computer's assembly instructions. To train it, we built an assembly game where it's rewarded for sorting data efficiently and wins by finding a correct, faster program. Really interesting to see how AI that was originally designed for one purpose, in this case learning games, is now being used for a very different use case.

Now, speaking of Google, let's move to our big company section. And in this field, there's no doubt that the story of the week, was Apple. Apple held their WWDC developer conference on Monday, and people were really wondering, were they finally going to get on the AI hype train? Obviously, we saw at Google I.O. a few weeks ago, and then Microsoft Build, that the big guys that they compete with are all in on AI. Apple, for its part, has been a little bit more behind the scenes. And indeed, that was the story again on Monday. The big focus was, of course, their Vision Pro device. And as Apple does, they basically entirely rejected the language of augmented or virtual or mixed reality and instead went with new language that they had

created called spatial computing. Now, there were little inklings of AI in there sprinkled around. There was a transformer-based system for their new autocorrect keyboards. And they discussed machine learning a couple of times, but it's very clear that they're staying out of the fray when it comes to the AI language. Some people think it's brilliant. Some people think it's insane. But in any case, it's what Apple is doing. Meanwhile, meta, some of the wind knocked out of its sales by the Vision Pro had an all-hands meeting this week where Mark Zuckerberg laid out their big plans for AI. And effectively, it's putting AI in absolutely everything. So AI chatbot integration in WhatsApp and Messenger, AI text image generation and Instagram. You name it, there's going to be AI in it. And in fact, we even got a

teaser of the Instagram chatbot, which apparently has 30 personalities and allows the user to figure out which one they like best. Lastly, in big companies was Adobe. Adobe had two big announcements this week. The first was that they were bringing their Firefly AI suite to Enterprise and doing so in a way that was more compliant and suited to business use cases than perhaps other competitors. And they also showed off a new beta version of Adobe Express with deeply integrated generative AI that could really put a crimp in the style of competitors like Canva. A couple quick fun ones in the open source space. Open source competitor to chat cheap Eighty Hugging Chat, which of course comes from Hugging Face, has added a web search feature. so now you can use hugging chat connected to the web, which of course creates feature parity as relates to other more closed source alternatives.

And then there's one which is getting a lot of chatter among the developer community. And that, of course, is Super AGI. Chief AI officer says we've seen Auto GPT, Agent GPT, and Baby AGII, but nothing beats Super AGI. Number one on GitHub trending. What is it? Deploy autonomous AI agents, connect with tools from a marketplace, run, concurrent agents seamlessly, clean dashboard, open source, and constant updates. Now, I did a video about this a little bit earlier this week. You can go check out that video. I am not a developer,

so I don't have the ability to definitively say this is much better than AutoGPT, but it's very clear from digging into it that they are trying to pick up where a lot of people had challenges with AutoGPT, which was the actual spinning up of autonomous agents to get things done. Part of the solution is that marketplace of tools that are integrated. And whatever the case, it's very, very clear that they are attracting significant developer attention right now. All one need do is look at the star history to see evidence of that, where it's a straight up line ever since the beginning of this month. But, alas, I'm starting to see a little bit of this too. Di Alberto Taran says so far Super AGI has been kind of a mix. It is the same issues as AutoGPT, where it gets

stuck in loops and such. So perhaps the answer for us is. is not to be overly disappointed with these technologies that are less than a couple months old, but to appreciate that they are still developing. Now, our last section of this weekly recap is the society section, and we kicked off this week with a new report that suggested that last month, 4,000 or almost 4,000 jobs were lost in the U.S. to AI. This comes from data from Challenger Gray in Christmas, who released a report saying U.S.-based employers laid off more than 80,000 people in May,

of which AI was responsible for about 3,900 or around 5%. That made it the seventh highest contributor to jobs loss in May. Now, this news was picked up by a lot of people who maybe thought that AI job loss was something for the future and who upon seeing this started to understand that perhaps this is an issue that is confronting us right now. Over in the policy realms, there are rumblings of a UK summit to happen this fall. Prime Minister Rishi Sunok was in the U.S. this week talking with President Joe Biden and apparently invited the U.S. to participate in that summit. and the EU has also made statements suggesting that they want companies like Google and meta

to start labeling AI-generated content even in advance of anything like the AI Act going into practice. Their fear is, of course, that in the meantime between now and whenever legislation gets finished, which is likely to be quite some time, there could be significant increases in online disinformation because of generative AI. So this would be a way they say to potentially combat that. Yet for all the concern, in some ways the big society story this week was the bottom. Mom dropped by Mark Andresen when he published a 7,000-word essay on why AI will save the world. This is the guy who wrote the extremely influential how software is eating the world thesis,

and so when he speaks, especially in this sort of comprehensive way, people tend to listen. This essay hits at a time that the voices of the AI alignment and AI risk and AI safety communities have actually started to see real traction of their message in society. Now, to what extent that traction is driven by media sensationalism versus real considered nuance is an open question. But as I've discussed frequently over the last couple weeks, I'm fairly convinced where we're out of stage where people, at least the people who are watching this issues, are convinced that this is something we do need to be thinking about. And Dresen represents the other end of the spectrum, and I think in that is an important voice as well. For anyone who's interested

in that P-Spot who wants to listen rather than read it themselves, I actually this week trained an AI on my voice. Good choice to do it in a week that I was ultimately going to lose my voice, and I will be presenting it as a Long Read Sunday episode tomorrow. Anyways, guys, that is it. for the AI weekly recap. A fairly quiet week as AI goes, although one we might look back on as the beginning of something great with things like Runways Gen 2. If you're enjoying the AI breakdown, I would love it if you would like, subscribe, and share. I would love it even more if you would click that notification button so you don't miss an episode. If you prefer listening, go check out the podcast. If you prefer reading, sign up for the newsletter. I appreciate each

and every one of you. And until next time, peace.