Good morning. We appreciate your detailed responses to the new design and will be iterating regularly to improve your reading experience here at The Deep View. Please, continue to let us know what you’d like to see in terms of layout (and what you really can’t stand).

A large group of major record labels filed a massive copyright infringement lawsuit against the AI music generators Suno and Udio on Monday. We break it all down for you below.

— Ian Krietzberg, Editor-in-Chief, The Deep View

In today’s newsletter:

• 👁️ AI for Good: Empowering early detection of eye disease

• ⚡️ Clean energy isn’t a catch-all solution to AI datacenters

• 📄 Study: A new method of hallucination detection

• 🎶 Record labels sue AI music generators Suno and Udio

AI for Good: Empowering early detection of eye disease

Glaucoma — which refers to a group of vision-impairing eye diseases — is the second-leading cause of blindness in the world today, according to the World Health Organization. As of 2019, roughly 7 million people around the world were living with glaucoma. 

While there is no cure for glaucoma, there are several intervention methods available that can reduce optic pressure, preserving vision. But they require early detection. 

That’s where A(eye) comes in: Over the past few years, new research has begun to surface that has positioned deep learning as a solution to the early detection of eye diseases, including glaucoma. 

• One 2020 study proposed a model that was tested on thousands of retinal images. The researchers reported a 99.39% accuracy rate of glaucoma detection. 

• A more recent study proposed methods of increasing accuracy and credibility of these disease detection models through transfer learning. 

There are, however, a number of obstacles to the implementation of AI screening for glaucoma. Among them is the fact that it is difficult to integrate the many different types of information used in glaucoma diagnoses into a singular model with a high accuracy rate. 

Still, eye doctors are confident that as the models improve, they will meet these challenges, allowing clinicians to identify and respond to glaucoma much more quickly.

Good prompts and fine-tuning aren’t enough if you’re building an AI SaaS application/feature.

It’s best to feed the LLM customer-specific context.

But this context lives across the dozens of apps and hundreds of files your customers use - in their emails, call transcripts, Slack conversations, internal knowledge documents, CRM data…the list goes on.

So to get access to this data, your team will need to build dozens, if not hundreds, of integrations.

But integrations shouldn’t be your core competency - that’s why AI21, Copy.ai, Tavus, Writesonic and other leading AI companies use Paragon to ship integrations for their products with 70% less engineering.

A few key highlights:

• Managed authentication for all integrations

• 3rd party webhook triggers

• Serverless ingestion infrastructure with robust error handling and observability

• 100+ connectors and a custom connector builder

Try Paragon for free or get a demo tailored to your use case

Clean energy isn’t a catch-all solution to AI datacenters

Yesterday, we talked about the ways in which AI is steadily increasing the energy demand of global datacenters.

• Note: A single ChatGPT query uses 10 times the amount of electricity as a single Google search, according to the International Energy Agency. 

Even with the details of AI model carbon footprints unclear, we know that the recent surge in AI is pushing tech giants further from their net-zero ambitions, which call in part for fully sustainable datacenters. The issue with this, according to AI researcher Dr. Emily Bender, is that energy is energy, clean or not. 

Key points: Bender said that using more and more clean energy for datacenters prevents that same pool of clean energy from being used in other places. 

• “Until we actually have more clean energy than we need, throwing lots of it at training and using ‘generative AI’ models is just going to make the climate crisis worse,” she added. 

Zoom out: At the same time, the Washington Post reported that in several areas across the U.S., fossil fuel plants are staying online well past their scheduled retirement dates to cope with increased power demands from datacenters.

OpenAI CEO Sam Altman has said that we will need an energy “breakthrough” to power AI; climate and AI researchers, meanwhile, have said we need to employ more thoughtful cost-benefit analyses concerning how rampant this technology is becoming. 

Use code IMPACTAITODAY for $200 off VIP Passes before the end of today*

• Want to adopt GenAI for software development but don’t know where to start? This buyer's guide helps engineering leaders cut through the noise*

• How small claims court took the place of Meta’s customer service department (Engadget).

• EU charges Apple with violating bloc rules on Big Tech (Semafor).

• UN chief tells consumer tech firms: own the harm your products cause (Reuters).

• OpenAI will allow current and former employees to participate in annual tender offers (CNBC).

Study: A new method for hallucination detection 

One component of generative AI that has not been iterated away through regular product upgrades is the propensity of these models to make stuff up. Large Language Models (LLMs) are designed to generate an output based on statistical probabilities derived from their training data; models are known to output fiction and untruths dressed up as fact. 

• Note: This has become known as ‘hallucination,’ though some scientists have taken issue with the term as it anthropomorphizes these machines. 

As Dr. Gary Marcus, Dr. Emily Bender and many others have repeatedly pointed out, these systems derive absolutely no meaning from the words that they haphazardly string together.

• This reality has posed perhaps the key obstacle to more widespread adoption of these systems. They aren’t reliable, and so cannot be trusted for operation in a number of fields and professions. 

New research, published last week, proposes a method for hallucination detection. 

The details: The researchers explored a specific subset of hallucination they call “confabulations,” where a model provides incorrect information due to a lack of “knowledge.” 

• The solution they tested would essentially act as a consistency score for output on a given query. The researchers asked a chatbot to generate several answers to the same prompt, then used a different model to compare the consistency of those responses. 

• They were then able to give model outputs a score that would inform users if the model might be suffering from a confabulation. 

The big caveat to this is that it would cost 10 times the computing power of a standard chatbot conversation, according to Time. 

Focus on the meeting, let AI handle the notes

Are you struggling to take notes during a meeting or taking hours to write a summary?

Meet Notta — your personal AI notetaker that can record, transcribe, and summarize meeting conversations into actionable text, saving you up to a day's time per week.

• Automatic notes: Let Notta take notes while you focus on the conversation, whether it's on Zoom, Google Meet, Microsoft Teams, or in person.

• Intelligent summary: Use AI to extract key insights, action items, and whatever you want from lengthy transcripts. Review an hour-long call in just 5 minutes.

• Collaboration: Instantly share the meeting notes with your team to facilitate async collaboration via comments.

Over 4 million users worldwide trust Notta. Sign up for free today to free yourself from manual work and get the most out of every meeting!

Record labels sue AI music generators Suno and Udio

A story that has ballooned alongside the rise of generative AI has centered around copyright infringement. GenAI models in large part have operated predominantly in the creative arts, generating synthetic writing, art, audio and video in seconds from brief written prompts. 

These models, however, are able to achieve this illusion of near-instant creation exclusively because of their training data; LLM output is little more than a probability distribution that reflects that training data. 

And while AI developers have kept quiet on the specifics of their training data, they have made clear that they train their models on copyrighted material, with OpenAI CEO Sam Altman saying in the past that it would "be impossible to train today's leading AI models without using copyrighted materials."

• The argument, instead, is that it is “fair use” to train on “publicly available” content. But it is not at all clear that the fair use doctrine applies to genAI training; the U.S. Copyright Office has yet to weigh in on the matter. 

A major new lawsuit: A massive group of major record labels — including UMG, Atlantic and Sony — filed suit Monday against AI music generators Suno and Udio, alleging “willful copyright infringement on an almost unimaginable scale.” 

They are seeking a legal measure that would prevent both companies from infringing on their copyrighted materials, in addition to damages of up to $150,000 “per work infringed.” 

Some specifics: I highly recommend reading the lawsuits in full. You can read the suit against Suno here; Udio here. 

• The suit against Suno claims that the company all but admitted to copying protected recordings (paragraph 10). 

  • In written correspondence before the suit was filed, Suno did not deny plaintiffs’ accusation of infringement, saying that its training data is “confidential business information.” 

  • Suno added that it is protected by fair use, which the suit said “was telling because fair use only arises as a defense to an otherwise unauthorized use of a copyrighted work.”

• The plaintiffs also designed a series of prompts that they say caused Suno to output recordings that bear striking similarities to well-known, copyrighted songs (paragraph 51). 

  • The suit included many such examples; in one, it claims that Suno generated nearly 30 almost-copies of Chuck Berry’s Johnny B. Goode through prompts that specified the genre, decade and provided lyrics from the original. 

• The suit goes on to say that Suno’s fair use claim is “a tacit admission of Suno’s illegal copying, as fair use only comes into play when an unauthorized use of a copyrighted work needs to be justified.” (Paragraph 68). 

  • It adds that the fact that Suno is earning revenue from its work further weakens its claim of fair use. 

Suno claimed in a statement to Bloomberg that its work is, in fact, “transformative.” 

The perspective of many professional creatives here is simple: GenAI, which cannot exist without its training data, is creating highly commercialized output that is designed to compete with the original creators. This concept is touched upon heavily in the New York Times’ lawsuit against OpenAI 

• We are already seeing the results of this competition, with jobs for freelance writers and artists on a steady decline. 

The core issue is less that the models exist, and more that they are deriving monetary value on the back of creative work that was used without the consent, knowledge or compensation of the original creators. 

My concern — in a world that has already reduced art to the status of ‘content’ — is an even greater devaluation of art and the artistic process. The idea of a music generator feels patently anti-human to me.

Thanks for reading today’s edition of The Deep View! Your feedback goes a long way to improving your reading experience, so make sure to leave your thoughts down below!

What changes would you like to see to the new design?