AI Blog Circles, I’m going in circles

Circles, I’m going in circles

AI Blog

Issue #30

25/10/2023

YouTube reaches out to record labels

Back in August (Issue 22 – The Return ) we talked about YouTube’s announcement of a partnership with Universal Music as part of the Music AI Incubator program that the former would launch. The announcement at the time was extremely vague, but the main concept was going to be to see how AI could be used responsibly in the music industry or how YouTube wouldn’t lose its music licenses to long-running litigation because of AI.

Why are we mentioning this again?

Because it surfaced again, thanks to Universal Music suing Anthropic (Claude’s mom company) for using song lyrics as training data.

This is not a terribly innovative move, as we have seen dozens of similar ones, where companies active in content/art production sue AI start-ups.

But it seems that ‘music’ (as an industry) is leading the way in managing the issue and this is where we come full circle.

YouΤube developed tool for the new AI creator suit recently released, which will allow creators to record audio using the voices of singers. To do this it needs approval from record labels, cue Universal Music, who in turn don’t want to give away their properties for free, but have no issue with selling them and collecting royalties.

So what does Universal do? It agrees to work with YouTube and by extension paves the way for similar deals.

Transparency is key

We often mention the concept of open-source and we’ve seen some of the reasons why it’s so important (Issue 7 – AI & Video ).

In fact, thanks to Meta, which has all its models open-source, it is now a serious debate whether to choose open or closed model for each new AI model.

Why do we say that?

The Foundation Model Transparency Index was recently released by researchers at MIT, Stanford, and Princeton. This Index measures how transparent AI companies really are. It tracks 100 different indicators, such as whether companies report what salaries they pay their AI trainers or the computing resources they used to develop a model. The results are collected in scores on a 0-100 scale.

Surprise or not, the top-scoring model has just 54 out of 100. Which shows us that we still have a long way to go in terms of transparency in the industry.

Open models are doing significantly better, but even in these there are still issues. For example in the data, labor, and compute part if we sum up the scores of the developers, the results are 20%, 17%, and 17% respectively. You call that extremely low.

Why is this a problem?

We need to understand that AI has tangible consequences in society and is primarily the domain of the research community. One of the key principles of which is transparency, so that other researchers can cross-check and provide meaningful critique.

The simplest example, and perhaps the one that is mentioned most of all, is the biases that exist in training data. It is almost impossible to avoid them completely, because humans are social and cultural products of our time. However, you can reduce them significantly if you make them public and thus allow multidisciplinary teams to check them. Whether an AI model reproduces biases can cause real-world problems, as we see in the learning bytes of the issue.

We talk about cycles in this issue, because everything is connected and we think it is especially important to be able to see the bigger picture.

Resources

AI Insider 📰

Teleperformance launches its AI platform,TP Configuration, for faster and more efficient processes, aiming at better customer service.
Apple announced that it will invest $1 billion per year in AI development for its products. Siri, Messages app and Apple Music seem to be among the first to receive the updates.
Midjourney launched new upscalers. They are visible below every image that has been upscaled once already.

2x is subtle and tries to keep the details the same as much as possible.

4x costs about 3x the GPU minutes that the 2x upscaler uses.

It also launched the beta version of the new site. The plan is to replace the existing one, and there are discussions about being able to use to Midjourney-bot there rather than on discord.
Bill Gates gave an interview to Handelsblatt, in which he said that he believes that AI models like the GPT have reached their peak for now.
Instagram is testing a new ΑΙ feature that will allow users to make stickers from their photos.

Learning Bytes 🧐

New research has shown that the most well-known AI chatbots (ChatGPT, Bard, etc.) tend to reproduce incorrect or refuted information when asked medical questions related to race. The research highlights the huge importance of bias and misinformation in AI and the real consequences it can have.
Apple cancelled Jon Stewart ‘s show “The Problem With Jon Stewart” (they must have said the same thing at the board meeting). The reason seems to be his views on AI & China and the discrepancy with Apple’s position on the issue.
Research has shown that AI can make a significant contribution to Wikipedia, as it can “clean up” references by identifying problematic links and finding better sources. In almost half of the tests run AI found better citations
New York City has launched the beta version of the MyCity Business Services chatbot. It is an AI-powered platform that will help residents by providing them with information to manage their business.
While the actors are still on strike, some of them participated in Meta and Realeyes’ “emotion” research, which aims to build AI training datasets for the former’s avatars. The advertisement said that it’s not a strike, but the gig is directly related to some of the core issues of the strike, such as how actors’ likenesses will be used, the fact that they have to be paid for it and what informed consent means. And that last one is extremely important, because even though “their individual likeness will not be used for any commercial purposes” it seems that the actors who signed the research agreements:
“have authorized Realeyes, Meta, and other parties of the two companies’ choosing to access and use not just their faces but also their expressions, and anything derived from them, almost however and whenever they want—as long as they do not reproduce any individual likenesses.”
Someone more astute than us would say that companies are trying to take advantage of the situation (strike going on, lack of rules & laws, actors in need of money) to collect as much training data as they can. Meanwhile alongside all this mess, Hollywood is asking background actors to get body scans, without specifying how they will be used. According to SAG reps, once the studios capture the actor’s likeness, they retain the copyright to it, and can use it indefinitely.