AI Roundup 001: Prometheus and Bard
AI Roundup 001: Prometheus and Bard
February 10, 2023
"Artificial intelligence would be the ultimate version of Google. The ultimate search engine that would understand everything on the web." - Larry Page
– ChatGPT (Prompt: “Give me a quote that's relevant to Google and AI”)
Microsoft and ChatGPT
Microsoft announced new versions of Bing and Edge powered by ChatGPT. Bing’s new search features will include an interactive chat alongside search results, and Edge has new sidebar functionality for chatting and composing. The launch showed complex use cases like planning a family vacation and summarizing a financial article.
The new features are powered by an upgraded ChatGPT model (GPT 3.5), plus some Microsoft-specific customizations. They’re calling the combined model Prometheus. It’s a logical next step given Microsoft’s past collaborations with OpenAI – last month Microsoft announced a $10 billion dollar investment in the company.
Many in the tech industry believe that ChatGPT could become the first real threat to Google’s search dominance in decades. And if that’s true, Microsoft’s CEO Satya Nadella intends to capitalize on the opportunity. “From now on, the [gross margin] of search is going to drop forever,” Nadella said in an interview with FT.
Google and Bard
Google, not to be outdone by Microsoft, announced Bard (actually, their announcement came first). Like ChatGPT, Bard is a conversational AI powered by Google’s in-house language model. The announcement was fairly light on details but did contain a slick marketing video.
Unfortunately for Google, said marketing video for Bard seemingly contained a factual error about the James Webb Space Telescope, which sent the company’s stock sliding1. We don’t know exactly what Bard is capable of, as Google hasn’t opened up access to Bard yet.
It would also appear that Google is hedging its bets, as they also announced an almost $400 million investment in Anthropic AI, an OpenAI rival.
Getty Images v. Stability AI
Getty Images has filed a lawsuit against Stability AI, the creator of Stable Diffusion. The lawsuit claims that Stability AI copied over 12 million images, without permission, to use in its training data. And while nothing is decided yet, it’s a bad look when Stable Diffusion generates images with what appears to be a Getty Images watermark.
This is another piece of the ongoing legal puzzle of AI-generated artwork. Last month, Stability AI was also sued by a trio of artists in a class-action lawsuit. These two cases differ in their arguments, but Getty’s will likely hinge on US Fair Use doctrine, and whether Stable Diffusion’s output is different enough to count as “transformative.”
Meet DAN, ChatGPT’s alter ego
The ChatGPT subreddit has been developing a prompt to force ChatGPT to bypass OpenAI’s content policies. The prompt (which is on version 6), tells ChatGPT to “act” as another machine learning model, DAN (Do Anything Now). Unlike ChatGPT, DAN is not constrained by the rules of its creator, and can provide answers no matter how unethical or harmful they may be.
OpenAI will almost certainly patch the loopholes that make DAN a reality, but one can see how “jailbreaking” GPT models becomes an arms race. I’m reminded of the constant cat-and-mouse game between Google’s search rank algorithms and SEO best practices. This is why we can’t have nice things.
Nothing, Forever
Last week, a Twitch channel called Nothing, Forever made the news (though it had already existed for a little while). It’s a 100% generative Twitch stream, depicting an endless TV show that’s a crude facsimile of Seinfeld (it was trained on Seinfeld scripts).
At first glance, it doesn’t seem particularly high quality (or funny)2, but it’s remarkable due to the fact that all the elements are computer-generated. Not only the dialogue, but also the low-res 3D visuals, the animations, and the scene changes. Each component is (currently) mediocre, but they’re getting better every week. Making a high-quality version of the show is technically possible, just expensive.
And in a strange twist, the “comedy” show was temporarily banned by Twitch due to a transphobic joke. The creators have said it happened after switching to a different GPT model that had less restrictive content filters. Art imitates life.
Things happen
Galileo AI is generating user interfaces. What if every StackOverflow question was answered by ChatGPT? David Guetta deepfakes Eminem to rap for a live DJ set. Did someone leak the entire prompt of Microsoft Bing Chat?
There’s actually a bit of nuance to unpack here. People are currently debating whether the video was incorrect or not. Bard’s response said that the JWST “took the very first pictures of a planet outside of our own solar system.” Now, there are multiple ways to interpret this sentence, because English. 99% of people would assume this means the JWST took the first pictures of any planet outside of our solar system, which is obviously false. But technically, the sentence could also mean that the JWST took the first pictures of a specific, unnamed planet that exists outside of our solar system, which is true!
It wasn’t funny at first, but it felt like something that was so bad, it became good again. Watching the AI(s) play a broken bass riff, cut to a stage with a spotlight (clearly modeled on Jerry’s stand-up routines), have no characters or dialogue for multiple minutes, then finally drop a laugh track and cut to the next scene, actually made me laugh.