Musk's Bid for OpenAI: This Week's AI News

Hey there, welcome back to TechCrunch's AI newsletter! If you're not already getting this in your inbox every Wednesday, you can sign up right here.

Looks like the billionaires are at it again.

On Monday, Elon Musk, who's currently the richest guy on the planet, threw out a whopping $97.4 billion offer to buy the nonprofit that basically runs OpenAI. In a pretty sassy comeback, OpenAI's CEO Sam Altman took to X earlier that day and said, "No thank you, but we will buy Twitter for $9.74 billion if you want." Just a reminder, Musk and his crew snapped up Twitter for $44 billion back in 2022.

Now, whether Musk was joking or not, his bid might throw a wrench in OpenAI's plans to switch to a for-profit public benefit corporation in the next couple of years. OpenAI's board is now in a spot where they need to prove they're not selling the nonprofit short by handing over its assets, like the IP from their research, to someone on the inside (like Altman) at a bargain price.

OpenAI could argue that Musk's move is a hostile takeover, especially since he and Altman aren't exactly buddies. They might also say that Musk's offer doesn't hold water because they're already in the middle of restructuring. Or they could just question if Musk even has the cash to back up his bid.

In a statement on Tuesday, Andy Nussbaum, who's the outside counsel for OpenAI's board, said that Musk's offer "doesn't set a value for [OpenAI's] nonprofit" and that the nonprofit isn't up for grabs. Nussbaum went on to say, "Respectfully, it is not up to a competitor to decide what is in the best interests of OpenAI's mission."

My colleague Maxwell Zeff and I dove deeper into what we might see in the coming weeks. But one thing's for sure, with Musk's offer and his ongoing lawsuit against OpenAI over what he claims is fraudulent behavior, we're in for some heated courtroom battles.

News

Is AI making us dumb?: A recent study looked into how using generative AI at work impacts our critical thinking skills. Turns out, if we lean too much on AI to do our thinking, we get worse at solving problems on our own when the AI isn't there to help.

AI for all, perhaps: In a new essay on his personal blog, Altman admitted that AI's benefits might not reach everyone — and he's open to "strange-sounding" ideas like a "compute budget" to make sure everyone on Earth can use a lot of AI.

Christie's controversy: The fine art auction house Christie's has sold AI-generated art before. But they're about to hold their first show dedicated entirely to AI-created works, and it's stirred up some mixed feelings — there's even a petition to cancel the auction.

Better than gold: An AI system from Google DeepMind, Google's top AI research lab, seems to have outdone the average gold medalist in solving geometry problems at an international math competition.

Research paper of the week

In their new study, the MIT CSAIL researchers found that while the best models still make real mistakes on popular AI benchmarks, over 50% of what we call "model errors" are actually due to mislabeled and unclear questions in those benchmarks.

"If we want to properly measure how reliable these models are, we need to rethink how we make these benchmarks to cut down on label errors," said MIT faculty member and OpenAI staffer Aleksander Madry in a post on X. "This is just the beginning."

Model of the week

Boreal-HL can whip up clips of tourists eating ice cream, folks grilling meat, people in lunch meetings, executives giving speeches at conferences, couples at weddings, and other everyday life moments. I find the whole thing pretty funny — especially since it's not exactly practical. It takes Boreal-HL at least five minutes to generate just one clip.

Grab bag

Thanks to some recent breakthroughs in AI efficiency, it's getting cheaper and easier to train really sophisticated models.

In a new paper, researchers from Shanghai Jiao Tong University and an AI company called SII showed that a model trained on just 817 carefully chosen samples can beat models trained on 100 times more data. They even claim their model could answer questions it hadn't seen during training, showing what they call "out of domain" capabilities.

This comes right after a Stanford-led project that found you can create an "open" model that rivals OpenAI's o1 "reasoning" model for under $50.