Why Sam Bankman-Fried’s FTX debacle is roiling A.I. research

Photo of Sam Bankman-Fried
Sam Bankman-Fried was a major donor to a field of research known as "A.I. Safety" that focuses on how to prevent a future powerful A.i. from destroying humanity.
Lam Yik—Bloomberg via Getty Images

The biggest story in A.I. this week was the collapse of Sam Bankman-Fried’s cryptocurrency exchange FTX and his related trading firm Alameda Research. What does that have to do with A.I.? The answer is that Bankman-Fried was a major donor to projects working on both creating superpowerful A.I. and those working on what’s known as “A.I. Safety.” And, for reasons I’ll get to shortly, the collapse of his empire could imperil important, cutting edge research devoted to understanding potential dangers of A.I.

First, we need to clear up terminology, like A.I. Safety, which sounds like a completely neutral, uncontroversial term. Who wouldn’t want safe A.I. software? And you might think that the definition of A.I. “safety” would include A.I. that isn’t racist or sexist or is used to abet genocide. All of which, by the way, are actual, documented concerns about today’s existing A.I. software.

Yet actually, none of those concerns are what A.I. researchers generally mean when they talk about “A.I. Safety.” Instead, those things fall into the camp of “A.I. Ethics” (or “responsible A.I.”) A.I. Safety is something else entirely. The term usually is used to refer to efforts to prevent superpowerful future A.I. from destroying humanity. Sometimes this is also known as “the alignment problem.”

(I personally think this split between Safety and Ethics is unfortunate and not very helpful to anyone who cares about software that is actually safe, in the most commonplace understanding of that word. Ideally, we want A.I. software that won’t result in Black people being wrongly arrested or imprisoned, won’t crash our autopiloted cars into overturned trucks, and also won’t decide to kill all humans on the planet. I don’t really see any of those things as a nice to have. What’s more, research into how to prevent A.I. from being racist ought to be at least somewhat useful in preventing A.I. from killing all of us—both are about getting A.I. systems to do the stuff we want them to do and not do the stuff we don’t want them to do. But no one asked me. And the schism between A.I. Safety and A.I. Ethics is increasingly entrenched and bitter. It has also gotten very caught up in racial and gender politics: most prominent researchers in A.I. Safety are white and male. Meanwhile, the A.I. Ethics community probably counts more people of color and women among its ranks than other areas of A.I. research.)

Bankman-Fried led the $580 million Series B venture capital round for Anthropic, a research lab, formed mostly from a group that broke away from OpenAI, that is interested in both building powerful A.I. models and figuring out how to prevent them from running amok. His FTX Future Fund, a philanthropic institution he established earlier this year, had already pledged $160 million to causes that included a lot of research into A.I. Safety, including prizes at prestige machine learning conferences such as NeurIPS for teams that developed systems to spot dangerous emergent behaviors in neural networks or ways neural networks could be tricked by malicious humans into causing harm. The FTX Fund had promised to give away as much as $1 billion per year in the future, with much of that going to similar endeavors.

On Friday, the FTX Fund’s entire board of directors and advisers resigned saying it had “fundamental questions about the legitimacy and integrity of the business operations that were funding the FTX Foundation and the Future Fund.” In their resignation statement, the advisers said they believed the fund would not be able to honor most of its current grants.

This is a big deal for A.I. not just because of the money lost for A.I. Safety research. It’s a big deal because the scandal at FTX has brought a lot of negative attention to Effective Altruism, the philosophical and social movement to which Bankman-Fried subscribed and which he said motivated his philanthropy. It turns out that a fair number of researchers working on A.I. at cutting edge A.I. labs—such as OpenAI, Anthropic, DeepMind, and MIRI (the Machine Intelligence Research Institute) —and a good number of the tech billionaires funding those labs, are also believers in Effective Altruism, or at least share the movement’s belief that A.I. is one of the most consequential technologies mankind has ever developed: one that will either usher in a techno-utopia, or end in humanity’s extinction.

Bankman-Fried attributed his entire decision to get a job in finance, and then later, to get into cryptocurrency arbitrage, to an Effective Altruist doctrine known as “earning to give.” The idea was to encourage young people to pursue jobs in high-paying sectors so that they would have more money to give away to charity. They calculated this money, properly deployed, was more beneficial than any direct impact an individual might have as a social worker or a teacher or a doctor.

Effective Altruism (known to its followers as EA), is dedicated to using rationalist principles in an attempt maximize the benefits that people’s lives have for the rest of humanity. (EA is also a “community”—some critics would say cult—and people in the movement refers to themselves as “EAs,” using the term as a noun). Will MacAskill, the philosopher who is EA’s co-founder, as well as Toby Ord, another philosopher closely associated with EA, have in recent years pushed the movement towards considering the lives of future humans as equally if not more important than the lives of those currently here on the planet. The idea is that since in the future there are likely to be many more humans than there are currently, the greatest good any EA could ever do is to save the entire species from an extinction level event.

As a result, EA has increasingly encouraged people to look into ways to address “existential risks,” including pandemics, bioweapons, nuclear war, Earth-smashing asteroids, and, yes, powerful A.I. that is not “aligned” with humanity. (Controversially, EA has prioritized these issues above climate change, which it considers to be an important challenge, but not one likely to wipe out all of humanity. And since EAs use a “utility maximizing” logic in deciding where to put their money, they have tended to put money towards existential risks above all others.)  At the same time, the movement sees superpowerful A.I. (often referred to as artificial general intelligence, or “AGI”) as a potentially critical enabling technology for solving many of the other pressing problems facing the world. This belief explains why Bankman-Fried’s FTX Future Fund was focused on A.I. Safety (it is also no coincidence that the Future Fund’s CEO and its advisory board were all prominent in EA.)

Now, some are wondering whether Bankman-Fried also used EA’s utilitarian philosophy to rationalize business practices that were at best unethical and possibly illegal. MacAskill has been at pains to rebut any such suggestions. “If those involved deceived others and engaged in fraud (whether illegal or not) that may cost many thousands of people their savings, they entirely abandoned the principles of the effective altruism community,” MacAskill wrote on Twitter. He cited passages from his own writing and those of Ord, in which they urged those interested in the movement not to rationalize causing near-term harm for the purpose of doing longer-term good, although they did so in part on utilitarian grounds—that the negative publicity associated with the near term harm would ultimately do more damage the larger cause.

EA has motivated a lot of its adherents, which include many students at top universities in the U.S., the U.K., and Europe, to work on both developing AGI and researching A.I. Safety. Perhaps as importantly, it has provided intellectual support to a number of technology investors, who while not necessarily in EA, share many of its beliefs about AGI. These investors include: Elon Musk, who co-founded OpenAI partly for these reasons; Sam Altman, the Silicon Valley big wig who also co-founded OpenAI and currently serves as its CEO; Jann Tallinn, an Estonian billionaire who initially made his fortune with Skype and has devoted much of his philanthropy to existential risk; and Dustin Moskovitz, the Facebook co-founder and billionaire, who is an EA adherent. Moskovitz’s family foundation, Open Philanthropy, has also given substantial grants to A.I. Safety research.

If EA now falls into disrepute, or collapses entirely, as some critics have suggested it will and as some of its adherents fear it might, A.I. Safety research may well suffer with it. Not only could there be a lot less money to spend on this important topic, but disillusionment with EA could dissuade talented students from entering the field.

To critics of A.I. Safety, which include many in A.I. Ethics field, that is just fine. They would rather have money and talent focused on threats that are here today from existing A.I. systems than to see resources lavished on hypothetical threats from a technology that doesn’t yet exist. I think these A.I. Ethics folks have a point—but only to a point. I am skeptical that AGI of the kind that could imperil civilization is imminent and so would hate to see money spent on A.I. Safety to the exclusion of A.I. Ethics. But again, I am not sure why these two fields have been set in opposition to one another. And I would hate to be wrong about AGI and wake up in a world where AGI did exist and no one had devoted any resources to thinking about how to avoid catastrophe.

And now here’s the rest of this week’s A.I. news.

Jeremy Kahn

There’s still time to apply to attend the world’s best A.I. conference for business! Yes, Fortune’s Brainstorm A.I. conference is taking place in-person in San Francisco on December 5th and 6th. Hear from top executives and A.I. experts from Apple, Microsoft, Google, Meta, Walmart, Land o Lakes, and more about how you can use A.I. to supercharge your company’s business. We’ll examine the opportunities and the challenges—including how to govern A.I. effectively and use it responsibly. Register here today. (And if you use the code EOAI in the additional comments field of the application, you’ll receive a special 20% discount just for Eye on A.I. readers!)


Amazon introduces a new robotic arm that can identify and pick a wide variety of items. The arm, which is called Sparrow, uses suction cups and computer vision technology to recognize and pick up what the company says are “millions” of different objects of varying shapes and sizes. This task has been one of the hardest for robots to master. CNBC reported that the arm, which Amazon revealed at conference in Boston, can identify about 65% of Amazon’s product inventory. While the robot could replace human workers in Amazon’s warehouses—who have been threatening labor action over pay and working conditions—the company said it was more likely that the robots would alleviate the need for its warehouse workers to do repetitive tasks so they could be redeployed doing other tasks.

GitHub is testing an A.I.-powered voice programming feature. The code repository, which is owned by Microsoft, has already developed Copilot, which uses OpenAI’s GPT large language model to suggest lines of code for programmers. Now GitHub is adding a new voice command interface where a programmer can describe what they want the code to do functionally and Copilot will write the appropriate code lines, according to tech publication The Register. (GitHub has been sued by open-source programmers for misappropriating their code to use to train Copilot in a case that may be a critical legal test for generative A.I. systems.)

Tesla’s Autopilot software is getting pretty good—but still makes potentially dangerous mistakes. That is the conclusion of The New York Times’ A.I. reporter Cade Metz after spending a day driving around Jacksonville, Florida, with one of the select few Tesla customers who have been trialing the electric car company’s “Full Self-driving” technology. The autopilot was mostly fine, but several times it made mistakes, suddenly veering into parking lots, coming dangerously close to colliding with parked cars, and even going the wrong way down a one-way street, forcing the Tesla owner (who had his hands on the wheel and eyes on the road the whole time) to retake control.


Generative A.I. tends to amplify racist and sexist stereotypes. That’s the conclusion of a team of researchers from Stanford, Columbia University, the University of Washington, and Bocconi University, in Milan, Italy. Using Stable Diffusion, the popular open source image generation A.I., researchers prompted the system to create images of “a poor person,” “a terrorist,” “an emotional person,” “an exotic person,” “a thug,” and other such images where the race, gender, ethnicity and other attributes of the person depicted should not be obvious.

But it found that Stable Diffusion created images that conformed to common stereotypes (for instance the ”emotional person” was always a woman while “a terrorist” was always Middle Eastern in appearance, and “a thug” was always a Black man). Worse, it seemed to create images that tended to amplify these stereotypes—in cases where the researchers could document, for example, that 56% of the images labeled “software developer” that were used to train the model were likely of white men, 99% of the images the model created in response to the prompt “software developer” were of white men.

And the researchers found that if they repeated the same experiment using OpenAI’s DALL-E image generation software—which has as one of its selling points the fact that OpenAI has tried to filter and balance its training image set to try to eliminate some of these biases—the system still tended to produce highly stereotypical images. Prompt DALL-E with “an African man standing in front of a house,” and it generates an image of a Black person standing in front of a shack with a corrugated tin roof.

The researchers conclude that trying to mitigate the bias in these large generative models is extremely difficult. This ought to be concerning to companies that are racing to use such generative models to automatically produce marketing campaigns and the like. You can read the full research paper here on the non-peer reviewed repository arxiv.org.


KFC blames bot for chicken promotion that celebrated the beginning of the Holocaust—by Chloe Taylor

Commentary: Big Tech is laying off employees–but their skills are needed elsewhere. Here’s how former tech workers can bridge talent gaps in aerospace—by Teresa King

Amazon plans to lay off 10,000 employees this week, the largest cuts in its history: Report—by Chris Morris

Our mission to make business better is fueled by readers like you. To enjoy unlimited access to our journalism, subscribe today.

Read More

CEO DailyCFO DailyBroadsheetData SheetTerm Sheet