Hello everyone, and welcome to TechCrunch’s weekly AI newsletter. If you’d like to receive it every Wednesday, you can subscribe here.
The billionaires are at it again.
On Monday, Elon Musk, currently the wealthiest individual globally, made a bid to acquire the nonprofit organization that primarily oversees OpenAI for a staggering $97.4 billion. In a witty comeback, OpenAI’s CEO Sam Altman responded on the same day with a playful tweet on X, stating, “No thank you, but if you’re interested, we’re offering to buy Twitter for $9.74 billion.” (Musk and backers had previously bought Twitter for $44 billion in 2022.)
Whether Musk’s offer is genuine or not, it poses complications for OpenAI as it aims to transition to a for-profit public benefit corporation within the next two years. The board now needs to prove that they are not undervaluing OpenAI’s nonprofit by granting its assets, including intellectual property from research, to someone with insider knowledge (like Altman) at a reduced price.
OpenAI could argue that Musk’s proposal is a hostile takeover attempt, especially since Musk and Altman are not on friendly terms. Alternatively, they could assert that the bid lacks credibility since OpenAI is already undergoing a restructuring phase. Another possibility is for OpenAI to question Musk about his financial resources.
In a statement made Tuesday, Andy Nussbaum, an external counsel for OpenAI’s board, mentioned that Musk’s proposal “fails to set a value for [OpenAI’s] nonprofit” and reiterated that the nonprofit is “not for sale.” Nussbaum further remarked, “With all due respect, a competitor should not dictate what is best for OpenAI’s mission.”
My colleague Maxwell Zeff and I have authored a more in-depth article on what to anticipate in the coming weeks. Rest assured, Musk’s proposal — coupled with his ongoing litigation against OpenAI over alleged fraudulent practices — is bound to lead to intense legal battles.
News
Apple’s latest robot: Apple has introduced a research robot inspired by Pixar’s innovative style. This robotic lamp serves as an animated version of a HomePod or similar smart speaker. When a user addresses the lamp with a question, it replies using Siri’s voice.
Are we dumbing down due to AI?: A recent study has examined how the use of generative AI in professional settings impacts our critical-thinking abilities. The findings suggest that heavy reliance on AI hinders our problem-solving skills when AI systems fail.
Is AI accessible to everyone?: In a recent blog post, Altman acknowledged that the advantages of AI may not be evenly distributed and mentioned that OpenAI is open to “unorthodox” concepts like establishing a “compute budget” to “facilitate widespread AI access for everyone on the planet.”
Christie’s debate: The prestigious fine art auction house, Christie’s, has previously sold AI-created artworks. However, it plans to host its first dedicated exhibition of AI-generated art, a decision that has sparked mixed reactions and even a petition aimed at halting the auction.
Exceeded expectations: An AI system developed by Google DeepMind has been reported to exceed the capabilities of an average gold medalist in tackling geometry problems during an international mathematics competition.
Research Paper of the Week
It’s commonly understood that many AI models struggle with basic tasks, such as solving elementary math problems. What remains less clear is why these failures occur. Research from a team at MIT CSAIL suggests that erroneous benchmarks could be a significant factor.
In their recent study, MIT CSAIL researchers discovered that while leading AI models are prone to making mistakes on standard AI benchmarks, over half of the “model errors” can be attributed to mislabeling and ambiguous inquiries within those benchmarks.
“To accurately assess model reliability, we must revamp our benchmarking methods to minimize labeling mistakes,” asserted Aleksander Madry, an MIT faculty member and part of the OpenAI team, in a post on X. “This is merely a foundational step.”
Model of the Week
You might be familiar with deepfake technologies, but how about deepfakes featuring the drabness of daily life? That’s the premise behind the Boring Reality Hunyuan LoRA (Boreal-HL), an AI video generator specifically fine-tuned to produce films showcasing the mundane aspects of life.
Boreal-HL can craft clips of people enjoying ice cream, individuals grilling at a barbecue, attendees in lunch meetings, executives delivering speeches at conferences, couples at weddings, and other everyday moments. I personally find the absurdity of this concept amusing, especially considering the impractical nature of the process; it requires Boreal-HL a minimum of five minutes to generate a single clip.
Grab Bag
Thanks to recent advances in AI efficiency, training highly complex models is becoming less expensive and more accessible.
A new research paper from scholars at Shanghai Jiao Tong University and an AI firm named SII illustrates that a model trained with just 817 “curated training samples” can outshine models trained on 100 times that amount of data. The researchers assert that their model even demonstrates the ability to answer questions it hasn’t encountered during training, showcasing what they term “out of domain” capabilities.
This study builds on previous research led by Stanford, which found that it is feasible to create an “open” model that rivals OpenAI’s o1 “reasoning” model for less than $50.
Compiled by Techarena.au.
Fanpage: TechArena.au
Watch more about AI – Artificial Intelligence
