Good morning! This is The Neuron. Let's get you up-to-date on AI news from the weekend.
Today in AI:
- Stability AI, Midjourney and DeviantArt Face Lawsuit
- The Latest on GPT-4; OpenAI Has a Video Model?
- Around the Horn
- Leo Sends His Regards
Stability AI, Midjourney and DeviantArt Lawsuit
See you in court.
The three companies are named defendants in a class action lawsuit that many in the AI community have seen coming for months.
Here's what the lawsuit is arguing:
- An AI image generator is basically a "complex collage tool" and everything it makes is a derivative of the works that the model was trained on.
- Stability AI paid a nonprofit to collect billions of images to train Stable Diffusion on. That nonprofit scraped stock image sites and sites like Tumblr, Flickr and DeviantArt.
- The companies 1) violated copyright, 2) took online content without consent, 3) caused the artists to lose money by enabling "fakes", 4) took advantage of the artists' names by allowing prompts like "in the style of X", and 5) profited from all of this.
Over the weekend, anonymous "tech enthusiasts" (read: not lawyers) responded to the lawsuit's arguments:
- On "collage tool": AI models don't store images and then mish-mash them to create new ones. They can't perfectly reproduce the images they were trained on.
- On copyright: Copyright law has to do with specific works, and you're being handwavy. Show us a specific work and how these models have directly infringed on that work's copyright.
- On "in the style of X": A "style" amounts to abstract concepts that can be summoned by completely unrelated words. Example: "cree, pasadena, ish, informal..." will create Buddha statue-like images - but we never say Buddha statues.
- The artists suing are trying to turn back the clock. Photoshop and the camera both changed art but didn't kill it. AI tools are the same.
It's a tricky case. Good thing the District Court of Northern California is presiding over it. This is the court that handles a lot of tech stuff: patent law, IP law, antitrust, etc. We'll see how the minds there think about things.
One thing's clear. Lots of artists don't want to be involved in this AI stuff. "So be it", says Stability AI, when they allow opt-outs for Stable Diffusion 3.
Related: The same legal team is also suing over GitHub Copilot, which helps programmers autocomplete code. They're arguing that it needs to provide attribution to the developers who produced the code Copilot was trained on.
The Latest on GPT-4; OpenAI Has a Video Model?
We might have to wait a bit for GPT-4.
At a talk hosted by VC newsletter StrictlyVC, OpenAI CEO Sam Altman said they'll only release GPT-4 when they can do it safely and responsibly.
That sounds like a safe move, but it also might be because it doesn't make any sense to push it when they're trying to make money with ChatGPT.
Meanwhile, GPT-4 is likely in some sort of private beta, which probably includes Microsoft.
Altman also confirmed OpenAI has a video model in the works. There's no other detail right now, but we might soon have 2 text-to-video models, the other being from Runway.
Around the Horn
- Stability AI teased a new image model a couple weeks back. They're continuing to tease, and the images look amazing.
- Live sports streaming service fuboTV said they're working on AI products at the 25th Needham Growth Conference. Just ask: "Can you show me every soccer goal that was scored from the left-side of the field?"
- All the big AI announcements from 2023 so far. We're only two weeks into the year!
- Princeton neuroscientist in WSJ: "Without Consciousness, AIs Will Be Sociopaths"
- The most comprehensive "Google vs. ChatGPT" article we've seen so far.
- Tech journalist asks, "A Writer Used AI To Plagiarize Me. Now What?"
- Create video clips with Stable Diffusion.
- Someone wants to crowdsource an AI-generated influencer, so of course people are trolling it.
She's perfect
Leo Sends His Regards
