AI Roundup 083: Bioacoustics

September 6, 2024.

Charlie Guo

Sep 06, 2024

∙ Paid

Workshops

In case you missed it, last week's Midjourney workshop recording is now available for paid subscribers.

And in two weeks, I'm hosting a Prompt Engineering workshop - see you there!

Join The Workshop

Bioacoustics

Google has developed a new "bioacoustics" model called HeAR (Health Acoustic Representations) that can detect diseases just from coughs.

Why it matters:

A recording of someone's cough is far easier to get than a blood sample or X-ray, meaning far more patients can benefit.
A partner company, Salcit Technologies, is using HeAR to do early detection of tuberculosis from patients in India - a country with millions of undiagnosed TB cases every year.
Besides, in a world of generative AI hype, it's still nice to see some big wins from predictive AI.

Elsewhere in frontier models:

DeepMind's AlphaProteo is a new model that generates novel proteins for biology and health research.
China-based MiniMax launched video-01, its text-to-video AI model that generates six-second clips.
And Anthropic launched Claude Enterprise, offering a 500K token context window, higher rate limits, and a GitHub integration.

Elsewhere in the FAANG free-for-all:

Google began rolling out its Gemini-powered Ask Photos feature in "early access," while YouTube announced AI detection tools and AI scraping controls.
Amazon's revamped "Remarkable" Alexa will reportedly be powered primarily by Anthropic's Claude and cost $5-10 per month.
And Mark Zuckerberg says Meta AI has over 185M weekly active users and is "growing quickly" despite no UK, Brazil, or EU rollout.

Silence of the jams

US prosecutors charged a man with using AI to create hundreds of thousands of fake songs by fake bands, stream them to fake listeners, and collect millions in royalties.

Between the lines:

The fraud began in 2017, well before generative AI. However, in 2018, the defendant partnered with an AI music company to upload thousands of new songs weekly.
This was a monumental feat of automation—it involved creating tens of thousands of fake emails, automating listening from various devices, and distributing the fake listens over an enormous back catalog.
Part of what tipped authorities off was likely the sheer volume of music. But with today's tools, one could likely do this at a smaller scale and fly completely under the radar1.

Elsewhere in AI anxiety:

National Novel Writing Month (NaNoWriMo) is in disarray after its organizers defended AI writing tools, calling critics "classist and ableist."
A new SEO-like technique, dubbed “AIO,” aims to influence AI models on specific topics by embedding targeted text on websites used for training data.
And LAION, a research organization whose dataset was used to train Stable Diffusion, has released a new dataset it claims has been "thoroughly cleaned" of CSAM.

Antitrust but verify

This week, the Department of Justice issued a subpoena to Nvidia, escalating an ongoing antitrust investigation.

Why it matters:

DOJ officials are concerned that Nvidia is making it harder to switch to competitors and/or punishing buyers that don’t exclusively use its AI chips.
Nvidia's recent acquisition of Run:ai, which helps companies manage their AI compute, is also under scrutiny.
The investigation added insult to injury, as Nvidia shares already suffered a record-setting $279 billion drop on Tuesday.

Elsewhere in AI geopolitics:

The US, UK, EU, and others signed a Council of Europe treaty to ensure AI use aligns with human rights, democracy, and the rule of law.
India and Singapore signed deals to boost collaboration in chip design and digital technologies, including AI.
And Ireland ended court proceedings against X while the UK greenlit Microsoft's acquihire of Inflection, declining to pursue an investigation.