Hiya, of us, welcome to TechCrunch’s common AI publication. If you need this in your inbox each Wednesday, join right here.
The billionaires are combating once more.
On Monday, Elon Musk, the world’s richest man, provided to purchase the nonprofit that successfully governs OpenAI for $97.4 billion. In response to Musk’s provide, OpenAI CEO Sam Altman earlier Monday authored a cheeky post on X, writing, “No thanks, however we are going to purchase Twitter for $9.74 billion if you’d like.” (Musk and buyers famously bought Twitter for $44 billion in 2022.)
Musk’s bid, severe or not, could complicate OpenAI’s effort to transform to a for-profit public profit company inside two years. Now OpenAI’s board must reveal it’s not underselling OpenAI’s nonprofit by handing the nonprofit’s belongings, together with IP from OpenAI’s analysis, to an insider (e.g., Altman) for a reduction.
OpenAI might make the case that Musk’s bid is a hostile takeover try on condition that Musk and Altman aren’t the best of friends. It might additionally argue that Musk’s provide isn’t credible as a result of OpenAI is already within the midst of a restructuring course of. Or OpenAI might challenge Musk on whether he has the funds.
In a statement Tuesday, Andy Nussbaum, outdoors counsel representing OpenAI’s board, mentioned that Musk’s bid “doesn’t set a price for [OpenAI’s] nonprofit” and that the nonprofit is “not on the market.” Nussbaum added, “Respectfully, it’s not as much as a competitor to determine what’s in one of the best pursuits of OpenAI’s mission.”
My colleague Maxwell Zeff and I wrote a extra detailed piece on what to anticipate within the coming weeks. However assured, Musk’s provide — to not point out his ongoing lawsuit in opposition to OpenAI over what he claims is fraudulent conduct — guarantees to make for fierce courtroom brawls.
Information

Apple’s new robotic: Apple created a analysis robotic that takes a web page from Pixar’s playbook. The corporate’s robotic lamp operates as a extra kinetic model of a HomePod or different good speaker. The particular person going through the lamp asks a question, and the robotic responds in Siri’s voice.
Is AI making us dumb?: Researchers not too long ago printed a research how utilizing generative AI at work impacts important considering expertise. It discovered that once we rely an excessive amount of on AI to assume for us, we worsen at fixing issues ourselves when AI fails.
AI for all, maybe: In a new essay on his private weblog, Altman admitted that AI’s advantages will not be extensively distributed — and mentioned that OpenAI is open to “strange-sounding” concepts like a “compute funds” to “allow everybody on Earth to make use of quite a lot of AI.”
Christie’s controversy: Effective artwork public sale home Christie’s has offered AI-generated artwork earlier than. However quickly it plans to carry its first present devoted solely to works created with AI, an announcement that has been met with blended evaluations — and a petition calling for the public sale’s cancellation.
Higher than gold: An AI system developed by Google DeepMind, Google’s main AI analysis lab, seems to have surpassed the common gold medalist in fixing geometry issues in a world arithmetic competitors.
Analysis paper of the week

We all know that almost all AI fashions can’t carry out primary duties reliably, like fixing grade-school-level math issues. What we don’t all the time know is the explanation behind their failures. In response to a crew of researchers at MIT CSAIL, inaccurate benchmarks could also be partially in charge.
In a brand new research, the MIT CSAIL researchers discovered that whereas right now’s top-performing fashions nonetheless make real errors on in style AI benchmarks, over 50% of “mannequin errors” are literally brought on by mislabeled and ambiguous questions in these benchmarks.
“If we need to correctly quantify mannequin reliability, we have to rethink how we assemble benchmarks to reduce label errors,” mentioned one of many researchers, MIT school member and OpenAI staffer Aleksander Madry, in a post on X. “That is only a first step.”
Mannequin of the week
You’ve heard of deepfakes earlier than. However what about deepfakes of boring on a regular basis scenes? That’s the concept behind Boring Reality Hunyuan LoRA (Boreal-HL), a fine-tuned AI video generator that excels at creating movies of … effectively, fairly banal stuff.
Boreal-HL can generate clips of vacationers consuming ice cream, folks barbecuing meat, folks in lunch conferences, executives giving speeches at conferences, {couples} at weddings, and different mundane slices of life. This reporter finds the absurdity of the factor hilarious — notably contemplating how impractical it’s to run. It takes Boreal-HL a minimum of 5 minutes to generate a single clip.
Seize bag
Because of current breakthroughs in AI effectivity, it’s getting cheaper — and simpler — to coach extremely subtle fashions.
In a new paper, researchers at Shanghai Jiao Tong College and an AI firm referred to as SII reveal {that a} mannequin skilled on simply 817 “curated coaching samples” can outperform fashions skilled on 100x extra knowledge. The crew claims that their mannequin was even capable of reply sure questions it hadn’t seen throughout the coaching course of, exhibiting what they name “out of area” capabilities.
The research follows on the heels of a Stanford-led challenge that discovered it’s attainable to create an “open” mannequin rivaling OpenAI’s o1 “reasoning” mannequin for below $50.