Technology

➚ MOVE FAST: Moving on. China is learning to live with U.S. efforts to restrict access to advanced chips. In one of the latest examples, an AI assistant for neurosurgeons will be tested at seven hospitals there, deciphering data from MRIs, ultrasounds and other tests. Later, it could be more involved in the treatment process, like warning surgeons of risky procedures.

➘ BREAK THINGS: Just getting started. Today’s passage of a U.S. House bill to force a sale of TikTok or ban it is just one battle in what’s likely to be a long fight. The plan faces higher hurdles in the Senate. And if it becomes law, TikTok can turn to the courts, where it has successfully challenged similar moves.

The U.K.’s Advanced Research and Invention Agency launched its first funded program today, a $54 million push to develop AI at 1/1000th of the current cost by bringing down compute hardware expenses. The money, doled out over four years, will go to researchers and engineers, and also to chemists and biologists to explore how nature efficiently processes information and how that can be applied to AI development.

The agency notes that for decades, people have benefited by having increasing computing power at lower cost but that no longer applies in the AI age. “Our current mechanisms for training AI systems utilise a narrow set of algorithms and hardware building blocks, which require significant capital to develop and manufacture,” it said. “The combination of this significance and scarcity has far-reaching economic, geopolitical and societal implications.”

Adobe Firefly repeats Google Gemini’s blunders

THE SCENE

Firefly, Adobe’s AI image creation tool, repeats some of the same controversial mistakes that Google’s Gemini made in inaccurate racial and ethnic depictions, illustrating the challenges tech companies face across the industry.

Google shut down its Gemini image creation tool last month after critics pointed out that it was creating historically inaccurate images, depicting America’s Founding Fathers as Black, for instance, and refusing to depict white people. CEO Sundar Pichai told employees the company “got it wrong.”

The tests done by Semafor on Firefly replicated many of the same things that tripped up Gemini. The two services rely on similar techniques for creating images from written text, but they are trained on very different datasets. Adobe uses only stock images or images that it licenses.

Adobe and Google also have different cultures. Adobe, a more traditionally structured company, has never been a hotbed of employee activism like Google. The common denominator is the core technology for image generation, and companies can attempt to corral it, but there is no guaranteed way to do it.

I asked Firefly to create images using similar prompts that got Gemini in trouble. It created Black soldiers fighting for Nazi Germany in World War II. In scenes depicting the Founding Fathers and the constitutional convention in 1787, Black men and women were inserted into roles. When I asked it to create a comic book character of an old white man, it drew one, but also gave me three others of a Black man, a Black woman and a white woman. And yes, it even drew me a picture of Black Vikings, just like Gemini.

The source of this kind of result is an attempt by the model’s designers to ensure that certain groups of people avoid racist stereotypes — that doctors not all be white men, for instance, or that criminals not fall into racial stereotypes. But the projection of those efforts into historical contexts has infuriated some on the right who see it as the AI trying to rewrite history along the lines of today’s politics.

The Adobe results show how this issue is not exclusive to one company or one type of model. And Adobe has, more than most big tech companies, tried to do everything by the book. It trained its algorithm on stock images, openly licensed content, and public domain content so that its customers could use its tool without worries about copyright infringement.

Adobe didn’t respond to a request for comment.

Sen. Michael Bennet, (D) Colorado; Sen. Ron Wyden, (D) Oregon; Kevin Scott, CTO, Microsoft; John Waldron, President & COO, Goldman Sachs; Tom Lue, General Counsel, Google DeepMind; Nicolas Kazadi, Finance Minister, DR Congo and Jeetu Patel, EVP and General Manager, Security & Collaboration, Cisco have joined the world class line-up of global economic leaders for the 2024 World Economy Summit, taking place in Washington, D.C. on April 17-18. See all speakers, sessions & RSVP here.

I just got back from a few days at SXSW. Well-adjusted people go there to listen to music, watch movies, and eat some good barbecue. I spent most of my time there talking with people about AI.

On Monday, I interviewed Shane Legg, co-founder and chief AGI scientist at Google DeepMind. That venture’s other two co-founders, Demis Hassabis and Mustafa Suleyman, have bigger public profiles, but Legg stands out in another way.

In 2010 when DeepMind was founded, the people who studied existential risk in AI usually weren’t the scientists actually building the technology. Legg was an exception. He was one of the few AI researchers willing to question the safety of the entire endeavor.

Now, visions of AGI are beginning to materialize and people like Legg believe it’s about to drastically change our world.

On the road to general intelligence, DeepMind’s greatest accomplishments have been in narrow areas. In a proof of concept demonstration, it defeated the world champion of the board game Go. It revolutionized biotech with its protein fold breakthrough. It used AI to get us closer to nuclear fusion and it upended the science of weather prediction.

So my question to Legg at the event: If building “general” intelligence is potentially dangerous, why not just keep building these narrow applications, which have already led to so many world-changing results?

One response to that is obvious: Somebody’s going to build it, so it might as well be us. But Legg said something else that I thought was interesting. As AI models get bigger, researchers are noticing a principle of crossover skills. The act of learning one language, for instance, allows an AI model to pick up entirely different languages more quickly. You might call it a kind of artificial wisdom.

One vision of how AGI might play out, Legg said, is that we’ll have super-intelligence models that, when tasked with a narrow problem, will create software tools to solve it. That’s what humans do it today.

What Legg said was bouncing around in my head on Tuesday, when Google DeepMind briefed reporters on a new research project called Sima, an AI agent that learned the fundamentals of playing video games and can now play new games it’s never seen before, without any additional training.

Some of the big breakthroughs in AI came from playing chess, Go, and video games. But those AI models were designed to win a particular game — not actually understand them. Sima doesn’t know how to win video games, but there may be power in its ability to generally understand them. Winning could come later.

Legg told me he came to believe in 2001 that we would have a 50/50 chance of reaching AGI by 2028. He says that number is still about right. If that’s true, and all these recent breakthroughs are the final leaps toward truly intelligent machines, we are in for a massive disruption in our way of life.

“This is a deep, deep change,” Legg said. “To be honest, it’s hard to wrap my head around.”

Billionaire Frank McCourt, with the help of journalist Michael Casey, is out with a new book on what social media has wrought on society. In Our Biggest Fight, McCourt argues that users have essentially become serfs of big tech companies like Meta, Google owner Alphabet and Apple, which have created a digital feudalist system. His answer: Decentralized Social Networking Protocol, which helps users take back control over their digital identities and data. McCourt says such a move would be good for our well being and for our democracy. Whether it’s actually a solution and even realistic, it’s a provocative concept as we grapple with how AI will change our lives.

Cognition, a new AI startup that came out of stealth Tuesday, is the talk of the automated coding world. We’ve written extensively in this newsletter about AI code generation, which was the first commercial use case of generative AI. The leader in that space is Microsoft-owned GitHub, which offers the coding assistant “Copilot” that helps coders complete tedious work.

Cognition takes the concept a step further with its AI coding tool called Devin, which can build an entire coding project from scratch, essentially allowing anyone to be a software developer. That’s something most experts in the field were saying was many years away.

So how was Cognition, which says it will launch a product soon, able to beat the big players? According to my X feed, they’re simply geniuses. A video is circulating around the internet of co-founder Scott Wu as a kid absolutely slaying a math contest. The fact that it’s taken less than 24 hours for one of Cognition’s founders to become a tech folk hero says something about the AI startup era.

• A former CNN chief has found himself in an escalating personal battle with a bombastic Tory grandee over control of a British newspaper.
• Will the son of Uganda’s dictator run for president? Our Global Election Hub has the latest.
• Ramadan in Gaza is characterized by mounting starvation.