Dotted around the Internet are a multitude of silly images of former US president Donald Trump. Here he is as a superhero, there he is as a cartoon warrior, and here he is playing basketball with a young Michael Jordan. The glossy look gives away that these are AI-generated, but there is no denying the underlying message: Trump holds on to a kind of online cultural power and is still weirdly beloved.
The messenger is Midjourney, a San Francisco-based artificial intelligence (AI) startup that in its 17-month existence has carried out no marketing, raised not one cent from venture capitalists, but is making US$200 million in annual revenue and has become one of the most powerful tools for generating remarkably real AI “photos.”
Fake snaps of Trump getting arrested and Pope Francis in a white puffer jacket confused Internet users and went viral earlier this year — and they were generated on Midjourney. The company has now released a new version that can do this with even more realism.
Having previously insisted that he does not like fake photos, Midjourney founder David Holz finds himself steering a tool created for artists that is also being exploited by propagandists. That is the trajectory of a kind of classic AI innovator, one who could not resist making their system more powerful at the price of their own standards.
Midjourney did not respond to multiple requests for comment or for an interview with Holz.
Holz grew up in Florida, where he carried out advanced science experiments as a kid — shooting paper airplanes at 257.5kph down a homemade wind tunnel, for example — before juggling interests in design and math in higher education. He co-founded Leap Motion in 2008, a startup that made a USB device the size of an iPod that allowed you to control a computer program with hand gestures. Holz was more than a decade too early with his idea, and he stuck it out to his detriment. After spurning a takeover bid from Apple Inc in 2013 for hundreds of millions of dollars, Leap Motion dropped in value. Holz finally sold it to a British firm in 2019 for US$30 million.
Undeterred, he founded Midjourney in 2021, eventually turning it into an independent research lab that would “expand the imaginative powers of the human species.”
In interviews with Forbes and The Register in 2022, Holz described how Midjourney was a space for people to “make beautiful things.”
He built the tool on an algorithm from OpenAI called “CLIP” or Contrastive Language Image Pre-training — a neural network that could learn visual concepts from natural language.
In the spring and summer of 2022, several months before OpenAI wowed the world with DALL-E 2, Holz released his own early versions of Midjourney, a tool that could conjure digital art from text prompts. His customers paid about US$10 to US$60 per month to use the tool, accessing it through public chatrooms on the Discord app.
It could have all ended there, except Holz had the gold dust every Silicon Valley entrepreneur craves: connections. He picked up several renowned advisers, including coding site Github’s CEO Nat Friedman, who purchased thousands of AI chips this year for startups to exploit. These chips, known as GPUs, are critical to keeping ahead in the race to build AI. With help from his network, Holz amassed an enormous collection of GPUs — as many as 10,000, roughly the number estimated to train ChatGPT — to make Midjourney’s models smarter and faster.
All that computing power translated into jaw-dropping improvements in version 5 of Midjourney, released in March last year. When users asked it for photorealistic images of people, everything from skin texture to facial features were much more realistic, while reflections, shadows and lighting were also more true to life.
Version 6, released on Dec. 20, generates faces with even more startling detail, with skin pores and texture that make them virtually indistinguishable from real photos. Midjourney users have been playing with the new software by generating images of Hollywood actors in imaginary movie stills, such as one of Leonardo DiCaprio as former Soviet leader Vladimir Lenin.
Holz seems to have gone quiet since his interviews last year. Back then, the entrepreneur expressed deep discomfort with people using Midjourney to create fake photos. That was “extremely dangerous,” he said. “I don’t really want to be a source of fake photos in the world.”
Yet by making Midjourney’s software more powerful, Holz laid the groundwork for fake images to proliferate. It now takes mere seconds to generate “photos” of celebrities and historic figures that are hyper-realistic, and some have risen to the top of Google search results.
Many Midjourney users are still creating art, but plenty more are creating photorealistic portraits like the one of Israeli Prime Minister Benjamin Netanyahu in Saudi clothing, which was created by a Midjourney user on Dec. 4 last year. The Midjourney user behind the Netanyahu image did not respond to requests for comment.
Many fear that AI photos would worsen misinformation, but in reality, it is hard to see fake images of Netanyahu and Trump tricking most people. AI photos still look a little uncanny, and media attention to “deepfakes” has made everyone more vigilant.
Their real impact would be more nuanced and harder to detect: A flood of fake images promoting certain ideas would act more like advertising campaigns that influence opinion rather than fool people. Fake photos that put Trump in critical historic moments — knocking down the Berlin Wall or fighting in the Vietnam War — that were made as a joke on X, formerly known as Twitter, have found their way onto Make America Great Again forums where people view them as inspiration, for instance.
AI pics could also be a propaganda tool outside of politics. Online forums that encourage eating disorders have been using Midjourney and other similar software to post images of ultra-skinny people to “inspire” their members, an August study by the nonprofit group Center for Countering Digital Hate showed.
Midjourney uses its own AI software to block potentially harmful content and it has about 70 hires monitoring its output, Bloomberg News reporting showed. However, such restrictions are relatively easy to circumvent. For instance, while Midjourney does not let you generate images of Bill and Hillary Clinton with blood on their hands, it does let you put blood-red strawberry syrup on them.
Holz seems aware of the problem. Back in March, he ended Midjourney’s free-trial program citing “abuse” by users.
However, the effects of propaganda are, like advertising, hard to track. As more people use his service — with 9 million users reported to be on his Midjourney’s Discord server — all that new content generated to support Trump, eating disorders or any other cultural value becomes more difficult to police.
It took eight months for Midjourney to release Version 6, a noticeably long time compared with the quicker releases of previous versions, which came two to four months apart. Holz might be struggling to get the AI chips he needs, or he could be grappling with the responsibilities of opening such an influential tool to the public mere months before a presidential election. Let us hope it was the latter.
Parmy Olson is a Bloomberg Opinion columnist covering technology. A former reporter for the Wall Street Journal and Forbes, she is author of We Are Anonymous. This column does not necessarily reflect the opinion of the editorial board or Bloomberg LP and its owners.
Taiwan might just give up. That’s the argument Simon Shuster made in The Atlantic’s October issue. The US is unreliable, the Chinese Nationalist Party (KMT) leadership is aligned with the Chinese Communist Party (CCP), and Taiwan does not have the military capability to defend itself in a prolonged war without help. What hope does Taiwan have? Quite a bit more, in fact, than Shuster lets on. Shuster points to genuine problems. The first Donald Trump administration had a mostly positive record on Taiwan policy. It regularized the arms sales process, enhanced the military-to-military relationship, and treated Taiwan more like a normal
A Facebook post by Taipei Mayor Chiang Wan-an (蔣萬安) earlier this month achieved record levels of engagement — although perhaps inadvertently. In it, he attacked Democratic Progressive Party (DPP) Legislator and Taipei mayoral candidate Puma Shen (沈伯洋) for an op-ed published in 2014 discussing drug harm reduction policies. In attempting to stir up fear with talk of community centers, drugs and needles, Chiang’s comment section exploded. More than 100,000 comments poured in within the first two hours; more than 330,000 after the first 12; and more than 400,000 by noon the following day. Rather embarrassingly, many commenters made a point
Southeast Asian food delivery platform Grab Holdings Ltd in March applied to acquire Delivery Hero SE’s Foodpanda operations in Taiwan, with the Fair Trade Commission (FTC) scheduled to announce its decision on Oct. 27. On the surface, the deal is a business merger — but in substance, it involves personal data, national geospatial information and digital sovereignty. It cannot be assessed solely in terms of market share and price competition. Grab has extensive and long-standing connections to Chinese capital. In 2015, Chinese ride-hailing firm Didi Chuxing and China Investment Corp, a sovereign wealth fund, invested in Grab. Grab invested in Chinese autonomous-driving
With the Taipei mayoral election quickly heating up, Taipei Mayor Chiang Wan-an’s (蔣萬安) campaign team has finally started to get serious this month, but a major mudslinging initiative and several sloppy mistakes have drawn a larger wave of public criticism, while sporadic municipal issues continue to crop up, raising concerns about whether the city government’s wheels are coming off. Just a few months ago, Chiang of the Chinese Nationalist Party (KMT) was widely expected to win handily in Taipei, traditionally a political stronghold for the KMT and with no major third-party candidates this year. KMT supporters regarded him as a frontrunner
RSS