Midjourney podcast mentions

Quote

um data. Um and I know that there has been approaches to the creation of synthetic data in the past, but they could never capture all this kind of statistical nuance. that you would uh that you would get from a data set. And and that the data sets that you would create w it couldn't really be used to kind of infer, you know, relationships about the data. And that it was surprising to me, and I think probably will be surprising to a lot of people, that the same diffusion Which powers DALI and Midjourney and these amazing image synthesis models that create these fire. out wild pictures, that same technology can be used to create a a synthetic data set of, you know, a set of health records. Um or a set of brain images or something like that, um, that the people are using for training. So to me that is a r really exciting use case. uh of the technology. Um the d the data science hub's also done um things around data science competency frameworks um to help organizations sort of understand what they should be looking for in a data scientist. But I think that these projects um going forward um to me are the most exciting ones because of just this leap in capability that we're seeing through generating AR. Definitely. Definitely. I completely agree. And I was uh

Quote

Who feel very strongly about like one thing or another, there's like literally another comment that's like. Like, yo, stop talking about the stupid safety stuff, man. It's another le needs this two more research analysis. It's like the never ending battle. and forth in which we're gonna be able to do that. So we'll I'll try to do a bit of both. Uh and yes, thank you for corrections. Sometimes we do say things are incorrect. Your viewer, I will say, points out that Midjourney had a web interface for the Okay, it happens. Um and then another comment uh From YouTube there was a request for us to uh kind of say again the title of a given paper so you can go look it up once we discuss it. Uh FYI, we do have the titles in the episode description. Uh And if you go to lastweek uhni dot com, you can also get the links to every single news story, every single uh research paper, article, etcetera. So Yeah, if you found something interesting, you don't remember what it was called, you want to dig in more, go to Alrighty. Well as always thank you.

Quote

This is a tough one for the reasons you described where the types of problems that models can recursively approve improve against is you can benchmark it. benchmarket that you can scale a bunch of compute and the bitter lesson takes hold. Now there are some places where uh here I for say for example like even if you just take a look at the same thing. diffusion models, like some people just like the way Midjourney lurks more than they like the way ChatGPT looks more more than they like the way Gemini looks, or nanomini, I should say, when it creates examples. And maybe the model you're talking about is like the Lion team had the aesthetics predictor. model that like helped evaluate some of these things because they also did some jet big did some generative image stuff so they also released their I think aesthetics evaluator stuff. out when I was first d really struggling with this problem and it was that I think the timing of that was such that we had kinda moved already to foundation models, but that was like Definitely the best purpose built thing I think I've still seen to this day. you're aware of that your audience might find useful as a way to reason about this is an There are some services, like Facebook, for example, where you're not allowed have text be more than X percent of the display ad. They find that it just

Quote

they've like now locked it into you have to have like a twenty dollar a month plan. And then they give you a certain number of credits. I hate that. Like just let me pay for what I'm using. And I want to do that because I'm some months I might use twenty dollars and some months I might use. none. And then I have to like uh mid journey uh also like this. I I subscribe to mid journey again. And like Midjourney has no API, right? There's you you can't just simply pay. I needed to generate like a whole bunch of backgrounds. I was trul playing with my green screen. Um and they do the best at like matching the light that's coming on my face. So I was like, Oh man, like wouldn't it be cool if I could just like I don't mind paying like fifty dollars to generate a whole bunch of backgrounds once, but I don't want to have this like ongoing thing and then You you're the subscription is over and then you lose all of your images from the the past. I But unfortunately, I think that's the way it's gonna go. Like I bet we're gonna see a clod in the future. Where you you simply just cannot use it via an API. You have to you only use it via cloud. Well that is pretty much how the the cloud code max plan. I think the API paying on a token base, I don't think that's going away. I think they want you to use that. Because the way that they've set this up is that if you're paying the two hundred dol…

Quote

They are a leader for sure, but since uh open A I released text-to-image functionality and then Google did as well inside of Gemini. uh I personally have switched completely to using either the chat GPT interface or the Google Gemini interface for creating images because the disk No, it yeah, I absolutely believe that Midjourney could still be at the company. edge um in terms of text to image quality, but the UI is so bad that I'm just like It's just so easy in Gemini. That's actually I think that's one of the best aspects of Gemini. And uh in fact they did launch I believe a better version of a web interface. We covered it earlier this year, but it took them quite a A while to move away from Discord and possibly because uh they have taken off the first time. Users I think. So uh yeah, uh it might be working for them. And I did use it through there and I guess once you do start using it it It's surprisingly not too bad, but uh most people I think would prefer Gemini or uh Chat GBT. And next we

Quote

actually difficult to convince AI models to be the national security tools that we will sometimes need them to be. That's a really interesting problem set and I think to the extent that that ends up being the case, boy is that an Interesting warning shot for alignment risk. Yeah. On to synthetic media and art, just a few more stories. We begin with Disney and NBC Universal sue AI company Midjourney for copyright information. infringement. So there you go, Mid Journey, one of a big text to image Model providers uh used to be a leader in in the best quality. Now they're just Apparently you can produce minions which is NBC Universal and And the claim here is that this is straightforward copy. Copyright infringement that mid journey has to stop doing it and then we're not going to be able Apparently, according to them, they reached out to mid journey prior to a lawsuit and ask them to stop and to filter the data and outputs

Quote

Light. Startup file dot AI and it is a relatively strong model. It has ten billion parameters trained for over two months on eighty million So even though it's they're not claiming it to be com competition. competitive with state of art stuff from Midjourney and others or Flux. They are saying that this is openly available, fully openly available. like Flux, which presumably are trained on copyright data, which is still Very much an ongoing legal question. emphasize being trained on licensed data. So uh This now makes it so there is a powerful open source model that I'd never heard of Freepick before, right? This is they're apparently a Spanish company, so Again, I think this is the first Spanish company I've heard about in this context. context in in kind of AI in general for a long time. I'm actually curious if people can think of of others that