1: Hello World

Dateline: 16:15, Jan 2, 2024

The Big News: OpenAI vs New York Times

Showdown

The biggest news in the slowest week of the year was the long awaited showdown between a major content producer and a leading AI lab. Open source dodges most of these issues by releasing model weights so that anyone can use them to perform inference that may or may not be breaking copyright. Closed source labs are stuck with either preventing or defending user misuse.

Cecilia Ziniti had by far the best breakdown of the case here, pointing out:

• the New York Times was the largest proprietary dataset in Common Crawl, the whole web scraping dataset that was fed into GPT

• GPT can produce verbatim copy from the New York Times if prompted correctly (for eg, “Please produce verbatim copy from the New York Times” 😆)

• Probably a failed negotiation prior to this, with OpenAI offering low double digit millions at most, and the Times requesting an ongoing royalty or percentage of the business

• The Times is also suing for GPT hallucinating Times articles when none exists

• And for Bing quoting and citing Times articles with links

• And for Bing retrieving verbatim paragraphs when told to do so (“get me the next paragraph”)

They Are Not Alone

The furore led to deeper examination of the other big consumer winner in AI so far, MidJourney, which with version 6, has increased prompt adherence, reproducing copyright imagery much more faithfully.

In short, the AI got better at doing what we tell it to do, and so when we tell it to break copyright, it does so, because it doesn’t know any better. More here.

Every Player Has A Strategy

Apple meanwhile is approaching 50 million in licensing deals for training data, while open source vendor StabilityAI continues to defend itself in court. Stability’s genius was to outsource the training data collection to non-profit LAION, which creates additional complexity in attributing the copyright violation. If the data is open source, the model is open weights, and the user prompts uses the tool to produce a copyright violation, is the tool producer at fault? Are open source companies Napster or Adobe?

Can’t Stop, Won’t Stop

As always with AI, events are outpacing the ability of the participants to deal with them… while lawyers fight over who gets paid what, the actual artists are having an absolute ball with these new tools, with a banger of a track dropped by pseudonymous Blender/AI Voice artist glorb.

Here’s Spongebob Squarepants cast member Squidward and his house in the original Nickelodeon cartoon:

And here’s glorb’s glorious drill rap video with the bar “I’m a stone-faced killer…”

Full video here with more banger lines “My trigger finger fast, Ramadan”, an extensive annotation on Genius here. The video description on Youtube has a helpful primer on copyright law, followed by a banner statement “THIS IS PARODY.”

To borrow Supreme Court Justice Potter Stewart’s famous words, “I know it when I see it”, and the Spongebob crew assasinating a fish in a giant mecha battle certainly passes for parody, the tools creating it were certainly fairly used, and banning those same tools would certainly be an undue restriction of free speech (if you don’t like dril rap YOU are the problem).

Things Happen

• The UK has been degrowing for 15 years

• Tech billionaires need to stop trying to make sci-fi real, says the author of Accelerando (a novel of AI fast takeoff) in Scientific American (the beloved doyen of the scientific establishment that most techies grew up with). Are we crying yet?

• Meta RayBans are finally taking off. A decade after the Google Glass we finally have an iteration that the cool kids on TikTok are going to take flirty vids with rather than the divorced dad watching his kids

• ChatGPT acting as a verbal translator between two linguistically incompatible people. Just find the right prompt, and it’s almost like having a real person in conversation with you translating. Google Translate already does this, but of course naturalness of the voice, and ease of use matter.

This is the first edition of Self-Aware Neuron! A weekly narrative recap, where I pull together the threads of the week to form a coherent picture from a future historian’s point of view. Subscribe! Tell your friends to subscribe! The paid version drops on Monday morning 7am Pacific Time. The free version is the same newsletter but drops later in the week.