hiya welcome to TechCrunch’s regular AI newsletter. If you need this in your inbox every Wednesday, sign up here.
The billionaires are fighting again.
On Monday, Elon Musk, the world’s wealthiest man, offered to buy the nonprofit that effectively controls Openai for $97.4 billion. In response to Musk’s offer, Openai CEO Sam Altman wrote a cheeky post to X on Monday, saying, “Thank you, but I’ll buy Twitter for $9.74 billion if necessary.” (Mask and Investors famously bought Twitter in 2022 for $44 billion.)
A mask bid, serious or not, may complicate Openrai’s efforts to convert to a for-profit public benefits corporation within two years. Currently, Openai’s board of directors should demonstrate that they are not selling Openai’s nonprofit organizations by discounting nonprofit assets, including IP from Openai research, to insiders (such as Altman). .
Open can argue that Musk’s bid is a hostile attempt at a takeover, given that Musk and Altman are not their best friends. You can also argue that Musk’s offer is unreliable as Openai is already in the middle of the restructuring process. Or Open could challenge Musk to see if he had funds.
In a statement Tuesday, Andy Nussbaum, an outside lawyer representing Openai’s board, said Musk’s bid “does not set the value of (Openai’s) nonprofit organizations,” and the nonprofit organization said “selling.” “It’s not done.” Nussbaum added, “In honor, it is not up to the competitors to determine the best interests of Openai’s mission.”
My colleague Maxwell Zeff and I wrote a more detailed article on what we would expect in the coming weeks. But Musk’s offer promises a fierce court brawl, let alone his ongoing lawsuit against Open Eye over his alleged fraud.
news
Apple’s New Robot: Apple has created a research robot that captures pages from Pixar’s Playbook. The company’s robotic lamps act as a more kinetic version of the home pod or other smart speakers. Those facing the ramp ask for a query, and the robot responds with Siri’s voice.
Is AI making fun of us? : Researchers recently published a study on how using generated AI in the workplace affects critical thinking skills. We’ve found that relying too much on AI and thinking for us that solving problems when AI fails can get worse.
Probably AI in AI: In a new essay on his personal blog, Altman acknowledged that AI benefits may not be widely distributed. Many AI to use Earth. ”
Christie’s Controversy: Fine Art Auction House Christie’s sells previously generated art to AI. But soon, they will hold a petition asking for the first show dedicated solely to AI-created works, a mixed review-filled announcement, and a cancellation of the auction.
The AI system, which is superior to Google, was developed by Google Deepmind, the leading AI research lab, and appears to outperform the average gold medalist when solving geometric problems in international mathematical competition.
This week’s research paper
I know that most AI models cannot reliably perform basic tasks, such as solving grade level math problems. What we don’t always know is the reason behind their mistakes. According to a team of researchers at MIT Csail, false benchmarks could be held liable in part.
In a new study, researchers at MIT CSAil have found that while today’s top performance models still make real mistakes in AI benchmarks, more than 50% of “model errors” actually have misunderstandings about these benchmarks. We found that it was caused by ambiguous questions.
“If you want to properly quantify the reliability of your model, you need to rethink how to build benchmarks to minimize label errors,” said Aleksander Madry, a teacher at MIT and a staff member of Openai. , mentioned in X’s post. The first step. ”
This week’s model
You’ve heard of deepfakes before. But what about deep fakes from boring everyday scenes? It is the idea behind the boring reality unyuan lora (boreal-hl). This is a fine-tuned AI video generator that’s great for creating videos of…well, something quite mediocre.
Boreal-HL can generate clips of tourists eating ice cream, people barbecue meat, people at luncheon parties, speeches at meetings, wedding couples and other tourists eating slices of ordinary life . This reporter finds the absurdity of hilarious things – especially considering how unrealistic it is. It takes at least 5 minutes to generate a single clip.
Glove bag
Thanks to recent breakthroughs in AI efficiency, it’s cheaper and easier to train highly refined models.
In a new paper, researchers from Shanghai Jiao Tong University and an AI company called SII show that models trained with just 817 “curation training samples” can outperform models trained with 100 times more data It’s there. The team claims that their models were even able to show what they called “out-of-domain” features to certain questions they had never seen during the training process.
This study follows shortly after a Stanford-led project, and found it possible to create a “open” model that rivals Openai’s O1 “inference” model for under $50.