Zuck buys ScaleAI to catch up in the AI race

Mark Zuckerberg, CEO of Meta/Facebook, is not afraid to spend money to get what he wants. And right now, it is clear that what he desperately wants is to get back into the top ranks of the AI model race. Barely a year ago, Meta was being hailed as the open-source AI champion, releasing AI models that were equal to the best of OpenAI and Google, and giving them away for free. A year is an eon in the Age of AI, and Meta’s latest model releases have been both late and lacking. Meta has appeared to have caught Apple’s disease - a once-storied tech company that has lost its mojo in the new technological era. So Zuck reached into his gigantic wallet and tried to buy himself a mojo transplant - paying a jaw-dropping $14.3 billion for 49% of ScaleAI, a well-known data-labeling startup, mostly to get its 28-year-old wunderkind CEO to join Meta as head of Meta’s brand-new superintelligence lab. Whether this turns out to be a genius move or merely rearranging the deck chairs on the Titanic remains to be seen, but the speed and the money-is-no-object price tag of the deal reek of desperation. At this juncture, we should remember that Zuckerberg created Facebook as a freshman at Harvard because he couldn’t get a date. Zuck has spun gold out of his insecurities before - maybe his yearning to play with the cool kids may motivate him to claw back into the front ranks of AI again. (Fun fact: Alexandr Wang is a close friend and former roommate of Sam Altman, CEO of OpenAI.) No doubt, the drama will continue…

Zuck tries to buy his way back into AI relevance

Sam Altman predicts a “Gentle Singularity”

Speaking of Sam Altman, the boyish CEO of OpenAI has released another think piece on his blog, this one titled “The Gentle Singularity.” Altman argues that the Singularity - the long-predicted point in time when change from artificial intelligence accelerates so quickly that our entire way of life in transformed - is already here, we just haven’t noticed it yet. He asserts that current AI models are already smarter than humans in many respects, and that we will see AI making novel scientific discoveries and powering highly capable robots within the next 2 years. But not to worry - all we have to do is make sure that our AIs are aligned with human values, and that we as a species figure out how to share the coming abundance among ourselves fairly, and all will be well. Although human beings have not yet figured out how to share the abundance we already have even after 5,000 years of civilization, Sam is sure we can solve that pesky little problem in the next couple of years. Methinks it could take a tad longer.

Sam Altman predicts a "Gentle Singularity”

OpenAI and Microsoft’s relationship is at a “boiling point”

Once the closest of business partners, Microsoft and OpenAI’s relationship is rapidly fraying. Some of this was inevitable - Microsoft wants to hold onto its commanding place in global technology, and OpenAI is the scrappy upstart. Microsoft poured billions into funding OpenAI in the early days, but now OpenAI wants to control its own destiny, and become successful enough to continue to attract investors to fund the additional billions it needs to stay at the top of the AI world. (See the story above about Zuckerberg’s rapid fall from the front ranks in less than a year.) Things came near a breaking point recently when OpenAI acquired AI coding startup Windsurf, threatening Microsoft’s own AI coding assistant, GitHub CoPilot. Meanwhile, OpenAI needs Microsoft’s approval for its conversion to a for-profit company, a transformation that is essential for OpenAI to continue to attract investors. It has been credibly reported that OpenAI was considering lodging a complaint against Microsoft with the Feds, accusing it of monopolistic practices. So far, cooler heads have prevailed.

OpenAI and Microsoft’s relationship is at a “boiling point”

News sites are getting crushed by Google’s new AI search

Many news sites are seeing significant drops in traffic now that Google is shifting its search output to AI summaries of the requested information. In the past, Google returned a list of links which directed users to the sites that hosted the information sought. Many news sites got a large portion of their total web traffic from these links from Google searches. The Huffington Post reports that its traffic from searches is approximately half of what it was 3 years ago, and publications including the Atlantic, New York Times, and Wall Street Journal have experienced lesser but still substantial declines. Journalism is once again facing a tech-fueled crisis in its business model, as it has periodically since the early 2000s, when Craigslist began displacing classified advertising in newspapers. Google itself is also going to have to figure out a new business model, because for over 2 decades it has been monetizing its search with paid sponsored links. Google is shifting to AI summaries because it has to, to keep its users, not because it wants to. If it falters in the transition, AI-native search startup Perplexity will eat Google’s lunch.

News sites are getting crushed by Google’s new AI search

Barbie gets a brain: Mattel partners with OpenAI

Last week, toymaker Mattel, maker of Barbie, Hot Wheels, and other iconic playthings, announced that it was partnering with OpenAI to develop toys and games, and that it expects to launch its first AI-powered product later this year. The company also plans to use AI tools to revamp its business operations, to cut costs and increase productivity. Toys in the US are largely made overseas, primarily in China, and the tariff kerfuffle in Washington has increased costs and disrupted supply chains. Mattel is looking for both a boost in the functionality of its toys so that it can then charge a premium for them, as well as an increase in the usual back-end efficiencies expected of any new computer technology.

Barbie gets a brain

MIT develops Self-Enhancing AI

One of the facts about current AI models that surprises many users is that, smart as they are, they don’t learn continuously, the way humans do. Modern LLMs are trained once on massive data sets, then typically don’t learn anything new until they undergo a retraining. This explains why LLMs don’t generally remember things that you have told it from past chat sessions. Now researchers from MIT have developed an AI system that can continuously incorporate new learning, which they call Self-Adapting Language Models, or SEAL. The SEAL Ai model in effect edits itself as new data is encountered, so that lasting improvements in its performance are achieved.

MIT develops self-enhancing AI

arXiv paper

A $2000-dollar AI-generated ad wowed viewers of the NBA Finals

Online betting platform Kalshi wowed the millions of viewers of the NBA Finals with an AI-generated ad that was produced by a single person over 3 days for only $2000. Consisting of short clips of supposed bettors in nonsensical situations, the ad was an attention-grabber from start to finish, with production values good enough for national release. The creator of the ad estimated that making this ad by conventional methods would have cost 20 times as much, and would have taken 20 times as long.

A $2000-dollar AI-generated ad wowed viewers of the NBA Finals

See the ad

Scientists find multimodal LLMs think about objects like humans do

Multimodal LLMs are able to process both text and images. Chinese scientists have discovered that multimodal LLMs spontaneously develop categorizations of objects similar to those of humans. Shown images of a wide diversity of objects that have not been categorized, multimodal LLMs spontaneously begin to group them in categories similar to those used by humans, such as “food”, “animal”, “vehicle”, etc. The researchers followed up by tracing the structures in the human brain activated by different images, and found striking similarities to the parts of the multimodal LLM neural network that were activated by the same image. This suggests that there may be something inherent in the neural network architecture that underlies how we classify objects.

LLMs spontaneously develop classifications of objects similar to humans

Chinese city moves an entire block of buildings with 432 walking robots

The Chinese city of Shanghai wanted to create an underground shopping and entertainment complex beneath a group of 100-year-old buildings. Rather than excavating underneath the buildings, which would be complicated, expensive, and risked damaging the existing structures, the city decided to relocate the entire group of buildings as a unit, with 432 “walking” robots underneath that slowly moved the buildings to a temporary site nearby. The cleared land will be excavated and turned into an underground mall, at which point the relocated buildings will be walked back with similar robots and reinstalled above the underground complex.

Chinese city moves an entire block of buildings with 432 walking robots

Drones deliver millions of sterile male mosquitos to fight disease

In the fight against deadly mosquito-borne diseases, scientists in Hawaii are using both sex and technology. The sex is in the form of specially-bred sterile male mosquitos. When these mosquitos breed, the female mosquito ceases mating activity, going on to produce infertile eggs that never hatch. Scientists breed millions of the sterile males in labs, which need to be delivered to the wetlands favored by their species. Nowadays, these millions of sterile male mosquitos are being delivered precisely on target by fleets of drones. This method avoids the environmental hazards of pesticide sprays, and is highly effective.

Drones deliver millions of sterile male mosquitos to fight disease

Multi-agent AI reproduces 12 Cochrane Reviews in 2 days

A team from MIT, Harvard, and other institutions developed a multi-agent AI research system that reproduced 12 systematic reviews published in a recent issue of the Cochrane Reviews in only 2 days, and the reproduced reviews were more complete and just as accurate as the human reviewers. Many studies in medicine have limited explanatory power because of small sample size. Cochrane Reviews is a widely admired publication that coordinates teams of experts to combine the data from multiple small studies addressing a given topic, and do a meta-analysis of the larger sample. This method can clear up ambiguity from conflicting papers by creating a larger and more complete and representative sample. The multi-agent AI review system was found to be just as accurate in its findings as the human reviewers that had performed the original reviews, and performed the equivalent of 12 years of human effort in a mere 2 days.

Announcement on Twitter/X

MedRxiv paper

Australian researchers successfully use mRNA against HIV virus

The quest for a vaccine against HIV has been thwarted for decades by the fact that the virus “hides out” inside white blood cells, shielding itself from attack by the patient’s immune system, or by medications. A team from Australia has developed a method for encasing mRNA in specially-engineered nanoparticles which then enter the white blood cells and instruct the cell to expose the virus it contains on the outside of its cell membrane. There the exposed virus is vulnerable to attack from the patient’s immune system or the action of antiviral drugs. This breakthrough may herald an era in which HIV becomes a curable disease, freeing the 40 million currently infected patients from the necessity of taking anti-viral medications to control the disease for the rest of their lives.

Australian researchers successfully use mRNA against HIV virus