GPU vendor Nvidia is the one company making all the money in the AI bubble. It put $1 billion into AI startups in 2024. That’s a lot of money for the startups — but it’s pocket change for Nvidia given its $1 trillion market cap and $9 billion cash on hand. [FT, archive]

Nvidia’s 2024 investment is 15% more than the $872 million the company put into startups in 2023 and ten times what it invested in 2022.

Why dump money into AI startups? Because Nvidia wants to keep the party pumping. The AI bubble will deflate at some point, taking Nvidia’s stock price back into saner realms.

Nvidia is a sensible company. It knows this is a bubble and it’ll pop. It just doesn’t want it happening yet.

All this is going on while Nvidia faces antitrust investigations in the US, Europe, and China.

Much of the money Nvidia puts into startups will come straight back to them in chip purchases. The US Federal Trade Commission and the Department of Justice are investigating whether Nvidia is making these investments to ensure its portfolio companies use Nvidia chips exclusively. Nvidia swears it isn’t, of course.

Nvidia’s investment is all that’s keeping some of these companies alive. But we can’t have the field collapsing. Maybe someone will come up with a use case for AI that isn’t just a money pit!

Click here to go see the bonus panel!

Hovertext:
Really every theorem should either describe what it does or have ten thousand names, chained by hyphens, going back to an early hominid named Garg.

I have a new book coming out on Tuesday 7th of January: it's A Conventional Boy, a short standalone novel in the Laundry Files. (It takes place at roughly the same time as The Rhesus Chart.)

It's published in the USA (and Canada) by Tor.com and in the UK (and EU) by Orbit. Both publishers are dropping hardback and ebook editions at the same time: as is usual Tor.com is a hardback-only publisher, but Orbit will eventually issue a paperback.

You can buy it at all good bookstores and ebook storefronts, but if you want a signed copy currently the only source is Transreal Fiction in Edinburgh, my home town's SF specialist bookstore. (I do not plan to visit the USA while Trump is president, and I'm attending fewer SF conventions than during the pre-COVID years, but I'll be able to sign books at the British Eastercon this year in Belfast.)

As for what it's about ...?

Meet Derek Reilly. Derek is portly, short-sighted, middle-aged, and has spent his entire adult life in prison for playing Dungeons and Dragons. It's not his fault: it was 1984, the Satanic D&D panic was in full swing, and Mistakes Were Made (by the Laundry).

Derek still pays D&D, after a fashion. He's got postal privileges and runs the camp newsletter: he also runs a play by mail campaign, with a set of mutant homebrew rules nobody in authority has any inkling about.

One day, Derek gets two pieces of very important news. Firstly, Camp Sunshine (the cultist deprogramming centre where he's spent most of his life) is closing for reconstruction in a few weeks. And secondly, a gaming convention is coming to the next town up the road.

Derek is upset—but sees an opportunity. He's had a foolproof escape plan for the past 20 years: he's just never had anywhere to go. Now he decides to go to his first ever convention.

Little does he suspect that he's not the only DM with a very dangerous set of rules converging on DiceCon 16 ...

After an avalanche of lawsuits over scraping everyone’s copyrighted works for their LLMs to regurgitate, OpenAI promised in May 2024 to develop a “Media Manager” tool to let creators opt their works out of training. OpenAI said this would be in place “by 2025.” [OpenAI, archive]

You’ll be utterly unsurprised to hear that Media Manager is not in place as of 2025. You might even think this was all a PR stunt they had no intention of delivering on. [TechCrunch]

“I don’t think it was a priority,” one former OpenAI employee told TechCrunch. “To be honest, I don’t remember anyone working on it.”

It’s unclear how a Media Manager as OpenAI describes it could ever have worked in the first place. LLMs are lossy compression for text. The source texts are fed in and set as weights in the LLM.

You can’t go in and cleanly delete the weights from that source text and not other source texts without retraining the whole LLM afresh.

It’s like making soup then saying you can just delete the garlic.

To see how hard it is in practice to remove text from LLMs, look at the kludged solutions to removing defamatory output — the companies put a crude filter on, and maybe add new text with updated information.

In any case, opt-out isn’t how copyright law works. All works are born copyrighted. If you want to use a work, you have to license it beforehand, not tell the owner to opt out of your use after the fact.