Can Meta Hire Its Way To Superintelligence?

Welcome back to The Prompt.

Cloudflare, the tech platform that powers millions of websites representing about 20% of the internet, announced yesterday that it will by default block crawlers from AI companies from scraping content without permission. It’s a significant move as AI juggernauts like OpenAI, Anthropic and Meta have scrounged the corners of the internet for data to power their AI models, often depriving the websites of the traffic (and associated revenue) they would have otherwise gotten. Publishers like Conde Nast, TIME, The Associated Press and The Atlantic as well as tech companies like Pinterest and Quora have expressed their support for Cloudflare’s shift to permission-based crawling. The company also announced a partnership with an initiative called Pay Per Crawl that would allow website creators to get paid for their content.

Let’s get into the headlines.

PEAK PERFORMANCE

Microsoft claims that its AI system, called “Microsoft AI Diagnostic Center,” can diagnose diseases four times more accurately than a group of experienced doctors at a significantly lower cost. That claim is based on an assessment of how correctly the AI tool was able to make a diagnosis on about 300 complex cases previously published in the New England Journal of Medicine in comparison to physicians. However, in the study doctors were asked not to use any additional tools to aid their diagnosis, which doesn’t reflect real world scenarios. Healthcare is a burgeoning use case for AI with Microsoft Copilot and Bing receiving about 50 million health-related queries per month.

ETHICS + LAW

AI-generated videos that portray Black women as primates and perpetuate racist stereotypes are amassing millions of views on platforms like Instagram and TikTok, Wired reported. The videos, part of a social media trend called “Bigfoot Baddie,” were generated by Google’s popular AI video generator, Veo 3.

TALENT RESHUFFLES

Anysphere, the company behind fast growing AI coding tool Cursor, has hired two engineers from Anthropic who previously worked on Claude Code, The Information reported. Cursor, which is used by developers at top AI companies to create programs and write code, is also a customer of Anthropic’s powerful coding models.

AI DEAL OF THE WEEK

Genesis AI, which is building an AI model for robotics, has raised a $105 million seed round led by Eclipse and Khosla Ventures.

DEEP DIVE

Mark Zuckerberg Sparks An AI Talent War With $100 Million Offers

As the race to build powerful AI models and launch impressive products intensifies, companies are shelling out top dollar for the researchers making these systems. On Monday, Alexandr Wang, Meta’s newly hired Chief AI Officer and former CEO of data labelling giant Scale AI, announced the creation of a new lab within Meta that is aiming to build so-called “superintelligence”— an AI system that outperforms humans in a range of cognitive tasks including creativity and problem solving. (That’s different from artificial general intelligence which is an AI system that can match human cognitive abilities,) Joining Wang are 11 top researchers that Meta CEO Mark Zuckerberg has freshly poached from leading AI labs including OpenAI, Anthropic and Google DeepMind, with shiny multi-million dollar offers.

In the past few weeks, Zuckerberg has personally reached out to dozens of researchers and engineers and offered eye-popping multi-million dollar pay packages, according to multiple reports. The researchers’ names can be found on “The List”— the billionaire’s compilation of the brightest minds and hidden geniuses in the field of AI. The Facebook founder has offered top talent $300 million over four years, with $100 million in total compensation (including equity) for the first year, Wired recently reported.

Meta’s initiative also raises questions about whether “The List” has actually got the right names on it. Richard Socher, a pioneer in natural language processing and CEO of You.com, tells me that while the talent pool for AI engineers is relatively small, companies like Meta are scouting for talent in the most “obvious places” like OpenAI, which ends up becoming expensive. “Not everyone who’s joined OpenAI as employee number 500 is more qualified than someone in a smaller startup,” he said. That said, not everyone at OpenAI is a target: Alex Nichol, a deep learning researcher at the startup posted on X, “Kinda offended that Meta didn’t try to recruit me… (not that I would accept, but it’s nice to feel recognized)”.

While the hires may seem flashy, they are entirely within reach for a tech giant like Meta, which has near-unlimited financial resources and unfettered access to powerful chips used to run these AI models to woo top talent, even from cushy jobs at juggernauts like OpenAI. But not everyone is convinced that the monetary incentives are enough to build a superstar AI team and create the best AI products. Two former Meta AI employees told Forbes they are unconvinced that huge sums of cash can motivate researchers to build superintelligence. “You want to attract people who care,” a former Meta AI employee said.

Meta itself has lost a lot of its top AI talent over the years, who’ve either left to start their own companies or to join rivals like OpenAI, the former Meta AI employee told Forbes. “A lot of people left to go to OpenAI…. This is Mark trying to undo the loss of talent,” they said.

WEEKLY DEMO

In an experiment called Project Vend, Anthropic let its flagship AI model Claude run a small automated physical store in its San Francisco office. The AI system, which had access to different external programs like email, was responsible for maintaining the inventory, setting prices and carrying out transactions. But it experienced a series of hiccups that included inventing a fake Venmo account, stocking and selling metal cubes after being prompted by a customer and giving persuasive customers discounts.

MODEL BEHAVIOR

People are using AI tools like ChatGPT to act as “trip sitters” — a term used to describe a sober person who monitors someone under the influence of psychedelic— while they consume psychedelics like magic mushrooms, MIT Technology Review reported. During their trip, people chat with AI tools to share their feelings and calm themselves. But adding AI into the mix could result in a “dangerous psychological cocktail,” according to multiple psychotherapists.