This Week in AI: Musk bids for OpenAI

Hiya, folks, invited to TechCrunch’s regular AI newsletter. If you privation this successful your inbox each Wednesday, motion up here.

The billionaires are warring again.

On Monday, Elon Musk, the world’s richest man, offered to bargain the nonprofit that efficaciously governs OpenAI for $97.4 billion. In effect to Musk’s offer, OpenAI CEO Sam Altman earlier Monday authored a cheeky station connected X, writing, “No convey you, but we volition bargain Twitter for $9.74 cardinal if you want.” (Musk and investors famously purchased Twitter for $44 billion in 2022.)

Musk’s bid, superior oregon not, whitethorn complicate OpenAI’s effort to person to a for-profit nationalist payment corp wrong 2 years. Now OpenAI’s committee volition person to show it’s not underselling OpenAI’s nonprofit by handing the nonprofit’s assets, including IP from OpenAI’s research, to an insider (e.g., Altman) for a discount.

OpenAI could marque the lawsuit that Musk’s bid is simply a hostile takeover effort fixed that Musk and Altman aren’t the champion of friends. It could besides reason that Musk’s connection isn’t credible due to the fact that OpenAI is already successful the midst of a restructuring process. Or OpenAI could challenge Musk connected whether helium has the funds. 

News

Image Credits:Apple

Apple’s caller robot: Apple created a probe robot that takes a leafage from Pixar’s playbook. The company’s robotic lamp operates arsenic a much kinetic mentation of a HomePod oregon different astute speaker. The idiosyncratic facing the lamp asks a query, and the robot responds successful Siri’s voice.

Is AI making america dumb?: Researchers recently published a study looking astatine however utilizing generative AI astatine enactment affects captious reasoning skills. It recovered that erstwhile we trust excessively overmuch connected AI to deliberation for us, we get worse astatine solving problems ourselves erstwhile AI fails.

AI for all, perhaps: In a new effort connected his idiosyncratic blog, Altman admitted that AI’s benefits whitethorn not beryllium wide distributed — and said that OpenAI is unfastened to “strange-sounding” ideas similar a “compute budget” to “enable everyone connected Earth to usage a batch of AI.”

Christie’s controversy: Fine creation auction location Christie’s has sold AI-generated creation before. But soon it plans to clasp its archetypal amusement dedicated solely to works created with AI, an announcement that has been met with mixed reviews — and a petition calling for the auction’s cancellation.

Better than gold: An AI strategy developed by Google DeepMind, Google’s starring AI probe lab, appears to person surpassed the mean golden medalist successful solving geometry problems successful an planetary mathematics competition.

Research insubstantial of the week

Image Credits:MIT CSAIL

We cognize that astir AI models can’t execute basal tasks reliably, similar solving grade-school-level mathematics problems. What we don’t ever cognize is the reason down their failures. According to a squad of researchers astatine MIT CSAIL, erroneous benchmarks whitethorn beryllium successful portion to blame.

In a caller study, the MIT CSAIL researchers recovered that portion today’s top-performing models inactive marque genuine mistakes connected fashionable AI benchmarks, implicit 50% of “model errors” are really caused by mislabeled and ambiguous questions successful those benchmarks.

“If we privation to decently quantify exemplary reliability, we request to rethink however we conception benchmarks to minimize statement errors,” said one of the researchers, MIT module subordinate and OpenAI staffer Aleksander Madry, in a station connected X. “This is conscionable a archetypal step.”

Model of the week

Image Credits:kudzueye (opens successful a caller window)

You’ve heard of deepfakes before. But what astir deepfakes of boring mundane scenes? That’s the thought down Boring Reality Hunyuan LoRA (Boreal-HL), a fine-tuned AI video generator that excels astatine creating videos of … well, beauteous banal stuff.

Boreal-HL tin make clips of tourists eating crystal cream, radical barbecuing meat, radical successful luncheon meetings, executives giving speeches astatine conferences, couples astatine weddings, and different mundane slices of life. This newsman finds the absurdity of the happening hilarious — peculiarly considering however impractical it is to run. It takes Boreal-HL astatine slightest 5 minutes to make a azygous clip.

Grab bag

Thanks to caller breakthroughs successful AI efficiency, it’s getting cheaper — and easier — to bid highly blase models.

In a caller paper, researchers astatine Shanghai Jiao Tong University and an AI institution called SII show that a exemplary trained connected conscionable 817 “curated grooming samples” tin outperform models trained connected 100x much data. The squad claims that their exemplary was adjacent capable to reply definite questions it hadn’t seen during the grooming process, showing what they telephone “out of domain” capabilities.

The survey follows connected the heels of a Stanford-led project that recovered it’s imaginable to make an “open” exemplary rivaling OpenAI’s o1 “reasoning” exemplary for nether $50.