The week in AI: Generate an entire song (complete with vocals) instantly with Microsoft Copilot AI

Together with

Welcome to The Dispatch! We are the newsletter that keeps you informed about AI. Each Thursday, we aggregate the major developments in artificial intelligence; we pass along the news, useful resources, tools or services, and exciting projects in open source. Even if you aren’t an engineer, we’ll keep you in touch with what’s going on in AI.

The generated songs aren’t particularly good yet, but they are marked improvement over previous text-to-music offerings like Meta’s AudioCraft, released just a few months ago (and sans vocals). As battles continue between artists and content creators over AI, this innovation will likely raise even more questions about the future of the technology within the music industry.

You can test the tool by visiting Copilot while using Microsoft Edge, or by logging into Suno’s website and trying out the service there.

Google’s DeepMind team has developed a new AI-powered tool called FunSearch which was able to solve the "cap set problem," which has vexxed mathematicians for decades. The cap set problem involves determining the maximum number of dots that can be arranged in a particular finite space such that no three lie on the same straight line.

FunSearch is a large language model that searches for functions to solve mathematical problems. It then generates suggestions through an iterative process, repeatedly building upon the most promising results until a solution is found. Its ability to reveal the logic behind solutions (it shows its work) makes the tool particularly powerful for scientific discovery. And because the model outputs code that can be easily inspected and deployed, its solutions could potentially be slotted into a variety of real-world systems.

That might not seem like a big deal to an average user who just wants to jump on ChatGPT’s app with their phone; to others, these advancements provide alternatives that help address concerns about data privacy, internet dependency, and lack of model control and customization. Professor Mollick also makes a compelling case for the “AI Haunted World”, where we are surrounded by many AI’s of varying forms and functions - akin to any ‘organization’ with a hierarchy. In this scenario, a smaller model, when given a task it can’t handle, essentially says, “Let me get my manager,” and offloads the task appropriately.

Following OpenAI’s recent board drama, the tech giant has published a new safety preparedness framework for managing their most powerful emerging AI systems. This includes granting the (newly reconstructed board) the power to reverse executive decisions, particularly to protect against catastrophic risks posed by increasingly powerful models. A dedicated "Preparedness" team will continually probe capabilities and risks, then issue reports to advise leadership and directors. The board can then potentially override ‘judged-safe’ rollouts.

OpenAI has consistently stated that the study of frontier AI risks is far behind where it should be. They’ve also put their money where their mouth is in that regard - launching $10M in grants to support technical research towards the alignment and safety of superintelligent AI systems.

The company also put out their first official guide on prompt engineering, which covers some very helpful steps to get the most out of your interactions with language models, such as “Chain of Thought” reasoning and persona/role adoption.

The Federal Trade Commission has banned Rite Aid from using facial recognition technology in its stores for 5 years due to privacy violations that harmed customers. Between 2012-2020, Rite Aid secretly deployed the technology in 200 stores without consent - frequently misidentifying innocent customers as shoplifters. This led staff to confront, search, remove, or call police on customers based on false matches, causing embarrassment and distress. The FTC ruled that Rite Aid failed to ensure accuracy or test for racial bias.

Rite Aid claims it stopped using facial recognition voluntarily in 2020 (after a whistle-blowing report from Reuters), but the FTC’s judgement here could dissuade other organizations from using the controversial technology.

More in AI this week:

• Jailed Pakistani Prime Minister Imran Khan deploys AI clone to campaign from behind bars

• (Bill Gates’ blog) The road ahead reaches a turning point in 2024

• OpenAI suspends TikTok parent ByteDance’s account after it allegedly used GPT to build rival AI product

• Israel is using an AI system to find targets in Gaza. Experts say it's just the start

• Recent data shows AI job losses are rising, but the numbers don’t tell the full story

• Biden administration and NIST take first step toward writing new AI standards

News. Without Motives. That’s 1440. Join Today!

Frustrated by biased news? Tired of trying to sift through 100s of news sources to find the facts? Leave behind the opinions and receive a daily digest of pure facts. Say hello to 1440! Delve into politics, business, science, sports, culture, and more. Join the other 3 million Americans who read 1440 each morning.