Major tech firms, in search of carbon pollution-free electricity for data centers, are helping to revive nuclear energy in the US.

Tech giants are increasingly eyeing nuclear reactors to power their energy-hungry data centers. Amazon and Microsoft each inked major deals this year with nuclear power plants in the US. And both Microsoft and Google have shown interest in next-generation small modular reactors that are still in development.

New AI data centers need a lot of electricity, which has taken companies further away from their climate goals as their carbon emissions grow. Nuclear reactors could potentially solve both of those problems. As a result, Big Tech is breathing new life into America’s aging fleet of nuclear reactors while also throwing its weight behind emerging nuclear technologies that have yet to prove themselves.

“Certainly, the prospects for this industry are brighter today than they were five and 10 years ago,” says Mark Morey, senior adviser for electricity analysis at the US Department of Energy’s Energy Information Administration.

Much of America’s aging nuclear fleet came online in the 1970s and 1980s. But the industry has faced pushback following high-profile accidents like Three Mile Island and the Fukushima disaster in Japan. Nuclear power plants are also expensive to build and generally less flexible than gas plants that now make up the biggest chunk of the US electricity mix. Gas-fired power plants can more quickly ramp up and down with the ebb and flow of electricity demand.

Nuclear power plants typically provide steady “baseload” power. And that makes it an attractive power source for data centers. Unlike manufacturing or other industries that operate during daytime business hours, data centers run around the clock.

“When people are sleeping and offices are shut and we’re not using as much [electricity], what matches nuclear energy very nicely with data centers is that they pretty much need power 24/7,” Morey says.

That consistency also sets nuclear apart from wind and solar power that wane with the weather or time of day. Over the past five years or so, many tech companies have accelerated climate goals, pledging to reach net zero carbon dioxide emissions.

The added energy demand from new AI tools, however, has put those goals further out of reach in some cases. Microsoft, Google, and Amazon have all seen their greenhouse gas emissions climb in recent years. Getting electricity from nuclear reactors is one way companies can try to bring those carbon emissions down.

Microsoft signed an agreement to purchase power from shuttered Three Mile Island in September. “This agreement is a major milestone in Microsoft’s efforts to help decarbonize the grid in support of our commitment to become carbon negative,” Microsoft VP of energy Bobby Hollis said in a press release at the time.

The plan is to revive the plant by 2028, a feat that’s never been done before in the US. The plant “was prematurely shuttered due to poor economics” in 2019, according to Joe Dominguez, president and CEO of the company, Constellation, that owns the plant. But the outlook for nuclear energy now is rosier than it has been for years as companies look for carbon pollution-free sources of electricity.

In March, Amazon Web Services purchased a data center campus powered by the adjacent Susquehanna Nuclear power plant in Pennsylvania. That $650 million deal secures electricity from the sixth largest nuclear facility in the US (out of 54 sites today).

Google is considering procuring nuclear energy for its data centers as part of its sustainability plans. “Obviously, the trajectory of AI investments has added to the scale of the task needed,” CEO Sundar Pichai said in an interview with Nikkei this week. “We are now looking at additional investments, be it solar, and evaluating technologies like small modular nuclear reactors, etc.”

He’s referring to next-generation reactors that are still in development and not expected to be ready to connect to the power grid until the 2030s at the earliest. The US Nuclear Regulatory Commission certified a design for an advanced small modular reactor for the first time last year. These advanced reactors are roughly one-tenth to one-quarter the size of their older predecessors; their size and modular design are supposed to make them easier and cheaper to build. They might also be more flexible than larger nuclear plants when it comes to adjusting how much electricity they produce to match changes in demand.

Bill Gates, for one, is all in on nuclear energy. He’s the founder and chair of TerraPower, a company developing small modular reactors. Last year, Microsoft put out a job listing for a principal program manager to lead the company’s nuclear energy strategy that would include small modular reactors.

“I’m a big believer that nuclear energy can help us solve the climate problem, which is very, very important,” Gates said in an interview with The Verge last month.

This week, the Department of Energy released a new report projecting that US nuclear capacity could triple by 2050. After flatlining for years, electricity demand is expected to rise in the US thanks to EVs, new data centers, crypto mining, and manufacturing facilities. That growing demand is changing the outlook for nuclear energy, according to the report. Just a couple years ago, utilities were shutting down nuclear reactors. Now, they’re extending reactors’ lifetimes by up to 80 years and planning to restart ones that have shuttered, it says.

“It is reasonable to think that the tech companies could catalyze a new wave of investment in nuclear, in the US and around the world. There has been plenty of talk about the idea in the industry,” Ed Crooks, Wood Mackenzie senior vice president, thought leadership executive for the Americas wrote in a blog post this week.

This doesn’t necessarily mean that it’s all smooth sailing ahead for nuclear energy in the US. New reactor designs and plans to reopen shuttered nuclear power plants are still subject to regulatory approval. Initiatives to build both old-school power plants and new designs have faced soaring costs and delays. Amazon already faces opposition to its nuclear energy plans in Pennsylvania over concerns that it could wind up driving up electricity costs for other consumers. And the nuclear energy industry still faces pushback over the impact of uranium mining on nearby communities and concerns about where to store radioactive waste.

“It’s an interesting time, challenging in many ways,” Morey says. “We’ll see what happens.”

After the podcast where Ben Horowitz repeatedly congratulated himself for being brave enough to support Donald Trump in public, it turns out Horowitz will also be donating to Kamala Harris. Or so he says.

“As I mentioned before, Felicia and I have known Vice President Harris for over 10 years and she has been a great friend to both of us during that time,” Horowitz wrote in an email to his venture capital firm obtained by Axios. “As a result of our friendship, Felicia and I will be making a significant donation to entities who support the Harris Walz campaign.”

I... listen. Two things. First, one of the most striking things about the podcast Horowitz made with his a16z co-founder Marc Andreessen was how focused both men were on who’d take their meetings. Their political commitments, in many respects, boiled down to who would spend face time with them. Near the end of the podcast, Horowitz told what was (I think?) supposed to be a heartwarming story about hanging out with Trump’s grandkids — and how his reaction to hearing about the assassination attempt on Trump was, “Oh my god, Grandpa just got shot.”

So, you know, the whole thing where Harris has been “a great friend” to both Horowitz and his wife checks out as a political position for him. In some respects, I get it. I too want to see my friends succeed!

But second, that crucial Trump endorsement was made before Joe Biden dropped out of the presidential race in favor of Harris, and before the enormous outpouring of support for the new Democratic candidate. That meant a16z had already sacrificed any leverage it might have with a Harris administration — and a candidate who made Trump look like less of a sure winner. In the case of a Harris win, these boys were looking at an entire four years without important meetings, and we already know they view that as a dreadful tragedy for the entire nation.

In order to regain any possibility of leverage in a potentially Democrat-led government, someone was going to have to eat humble pie, and it sure as shit wasn’t going to be Andreessen. Horowitz, whose support for Trump was termed a “Maga U-turn” by The San Francisco Standard, has a history of donating to progressive causes; Harris, then a senator, was even a guest at a Horowitz backyard barbecue in 2018.

“I’m just wondering what they’re gonna do when Kamala wins,” one founder told The San Francisco Standard of the a16z Trump endorsement. “Like, how do you walk that back?”

Well, I guess we know now. Horowitz’s email says that he’s had several conversations with Harris and her team — crucial to note here that he got to the candidate herself, given that access is apparently one of Horowitz’s main political concerns. And so while Horowitz still thinks President Joe Biden is no-good and very bad, he is “encouraged by my belief” in Harris to do good for tech, despite the fact that she has not yet staked out any policy positions.

Harris has sort of made overtures on what Horowitz termed “probably the most emotional topic” in the election: crypto. One of her campaign aides has said, “She’s going to support policies that ensure that emerging technologies and that sort of industry can continue to grow,” in response to a question about crypto.

Anyway, someone ought to do a wellness check on David Sacks — looks like his VC cabal is cracking.

A new AI-powered video generator from Meta produces high-definition footage complete with sound, the company announced today. The announcement comes several months after competitor OpenAI unveiled Sora, its text-to-video model — though public access to Movie Gen isn’t happening yet.

Movie Gen uses text inputs to automatically generate new videos, as well as edit existing footage or still images. The New York Times reports that the audio added to videos is also AI-generated, matching the imagery with ambient noise, sound effects, and background music. The videos can be generated in different aspect ratios.

In addition to generating new clips, Meta says Movie Gen can also create custom videos from images or take an existing video and change different elements of it. One example shared by the company shows a still headshot of a woman; the added video depicts her sitting in a pumpkin patch sipping a drink.

Image: Meta

Movie Gen can also be used to edit existing footage and change the style and transitions or add things that didn’t previously exist. In one example shared by Meta, a relatively innocuous video of what appears to be an illustrated runner is edited using AI in different ways: in one frame, he’s holding pompoms. In another, the background has been edited to depict a desert. In a third, the runner is wearing a dinosaur costume. Changes can be made using text prompts.

Nearly two years after powerful AI image and video generators hit the mainstream, AI companies have pushed the technology further: in just the last six months, major tech companies like Google and OpenAI are working on similar tools, along with smaller startups. OpenAI’s Sora, first announced in February, still hasn’t launched publicly; this week, a co-lead working on the video generator left the company for Google.

Meta’s chief product officer, Chris Cox, writes on Threads that the company “[isn’t] ready to release this as a product anytime soon,” as it’s still expensive and generation time is too long.

AI image generators raised concerns about ownership and potentially harmful use cases, and AI video generators have only exacerbated them. Reports indicate that AI startup Runaway trained its video generator on thousands of scraped YouTube videos, something YouTube CEO Neal Mohan has already said would violate the platform’s terms of use. In a blog post, Meta says it trained Movie Gen on “a combination of licensed and publicly available datasets” but didn’t specify which.

Creatives like filmmakers, photographers, artists, writers, and actors also worry about how AI generators will affect their livelihoods, and AI has been a central part of several strikes, including the historic joint Hollywood strikes by the Screen Actors Guild - American Federation of Television and Radio Artists (SAG-AFTRA) and Writers Guild of America (WGA) last year.

The new Microsoft Copilot is a big departure. It shed some of its Bing-y styling and increasingly corporate vibes in favor of something warmer and friendlier, all in an effort to be more than just a work tool. It might just be a website redesign, but it feels like a shift in how Microsoft perceives the future of AI assistants.

On this episode of The Vergecast, we try to make sense of that future. Everyone seems to believe in the potential for “agentic” AI, which can use apps and devices on your behalf, but most current tech along those lines feels woefully unfinished. As they get warmer and fuzzier, though, maybe the whole Getting Things Done aspect of AI doesn’t matter so much. We try to figure out where we’re headed next and how long it’ll take to get there.

You know who else needs to figure out where we’re headed next? OpenAI, which just raised $6.6 billion and now needs to prove it can spend it wisely. We talk about that for a while, too.

Once we’re done with the AI talk, we run through the gadget news of the week, from the students building a facial recognition system into their smart glasses to Sonos’ kinda-sorta plan to come back after its app fiasco. Then we get into some breaking news about the future of WordPress, which might be the most important software on the web and which feels suddenly in a bit of flux.

In the lightning round, we say goodbye to our friend, colleague, and cohost Alex Cranz. (Alex, we love you, we’ll miss you!) We also talk about the bizarre history that led to this week’s Dish / DirecTV merger, Nintendo’s ongoing war on emulators, and Google’s nifty new web app for managing headphones.

If you want to know more about everything we discuss in this episode, here are some links to get you started, beginning with Copilot and all things AI:

• Microsoft gives Copilot a voice and vision in its biggest redesign yet
• Copilot is killing beloved celebrities
• Read Microsoft’s optimistic memo about the future of AI companions
• Microsoft Paint is getting Photoshop-like AI fill and erase features
• OpenAI just raised $6.6 billion to build ever-larger AI models
• From The Atlantic: Shh, ChatGPT. That’s a Secret.

And in gadget news:

• College students used Meta’s smart glasses to dox people in real time
• Sonos has a plan to earn back your trust, and here it is
• Chromebooks are getting a new button dedicated to Google’s AI
• Microsoft is discontinuing its HoloLens headsets
• Automattic demanded a cut of WP Engine’s revenue before starting WordPress battle
• WordPress cofounder is paying employees to leave if they disagree with him

And in the lightning round:

• Alex Cranz’s pick: Nintendo has reportedly shut down Ryujinx, the Switch emulator that was supposedly immune
• David Pierce’s pick: Google’s Pixel Buds are now fully supported on Windows and macOS
• Nilay Patel’s pick: DirecTV and Dish are merging

OpenAI has launched a new “Canvas” interface for ChatGPT that allows users to adjust sections of text or code generated by the chatbot in a side-by-side collaboration. Canvas opens a separate editing window next to the existing ChatGPT chat box, which can be used to manually change text or code or highlight specific sections to better target editing suggestions and feedback from ChatGPT.

“With canvas, ChatGPT can better understand the context of what you’re trying to accomplish,” OpenAI said in its announcement. “Making AI more useful and accessible requires rethinking how we interact with it. Canvas is a new approach and the first major update to ChatGPT’s visual interface since we launched two years ago.”

ChatGPT Canvas provides users with a menu of shortcuts for suggesting inline edits, quickly checking grammar and clarity, and adjusting the length and reading level of text. Some coding-specific shortcuts are also available for debugging, adding logs and comments, and translating code into other languages. It’s a similar concept to the “Artifacts” feature that rival AI developer Anthropic rolled out for Claude in August.

Prior to this, ChatGPT users would typically need to repeatedly input new prompts to refine what the chatbot spat out. Canvas sounds a lot less tedious and gives users some additional control. It provides a back button that can restore previous versions of work, and OpenAI says it’s optimizing when Canvas will automatically activate — such as making it less likely to trigger for coding tasks to avoid disruptive power users.

The Canvas interface is currently in beta and can be accessed by selecting “GPT-4o with canvas” under ChatGPT’s model picker drop-down menu. It’s rolling out globally to ChatGPT Plus and Teams users, with access being expanded to Enterprise and Edu users next week. When it’s out of beta, Canvas will be available for ChatGPT’s free users and automatically open “when ChatGPT detects a scenario in which it could be helpful,” according to OpenAI.

Representing California in Congress comes with a unique challenge: navigating national politics while reflecting the interests of the most populous state in the US, including a large constituency from the tech industry. It’s a challenge both current California Sen. Laphonza Butler and Vice President Kamala Harris — who previously held that title — have taken on. And right now, governing the tech world means addressing AI.

Congress hasn’t made much headway on a national framework for regulating generative AI. But California is the epicenter of the AI industry, home to companies like OpenAI and Google. On the national stage, Harris has acted as an AI czar within the Biden administration, leading discussions with industry players and civil society leaders about how to regulate it. Butler, who has a long history with the VP, is focusing on a specific problem: how AI systems impact labor and social equity.

Butler spoke with The Verge about balancing the interests of AI companies and the people their products impact, including workers who fear being automated out of a job. “It all starts with listening,” says Butler, a former labor leader. “It starts with listening to both the developers, the communities potentially impacted negatively, and the spaces where opportunity exists.”

A balancing act

Like many officials, Butler says she wants to help protect Americans from the potential dangers of AI without stifling opportunities that could come from it. She praised both Schumer and the Biden administration for “creating spaces for communities to have [a] voice.” Both have brought in labor and civil society leaders in addition to major AI industry executives to educate and engage on regulation in the space.

Butler insists lawmakers don’t need to make “false choices” between the interests of AI company executives and the people who make up the workforce. “Listening is fundamental, balancing everyone’s interest, but the goal has to be to do the most good for the most people. And to me, that is where a policymaker will always tend to land.”

California state Senator Scott Wiener made similar statements about his hotly contested state-level bill, SB 1047. The bill, which would have required whistleblower protections and safeguards for potentially disastrous events at large AI companies, made it all the way to Gov. Gavin Newsom’s desk before being vetoed, with companies like OpenAI warning it would slow innovation. In August, Wiener argued that “we can advance both innovation and safety; the two are not mutually exclusive.” So far, however, lawmakers are struggling to find a balance between the two.

More work to do

Butler praises the steps both Schumer and the Biden-Harris administration have taken so far to create appropriate guardrails around AI but says “there’s always more to do.” Schumer laid out a roadmap earlier this year about how to shape AI policy (though it didn’t specifically introduce actual legislation), and the White House has secured voluntary commitments from AI companies to develop the technology safely.

One of Butler’s recent contributions is the Workforce of the Future Act, which she introduced with Sen. Mazie Hirono (D-HI). The bill would direct the Department of Labor, National Science Foundation, and Department of Education to study the impact of AI across job sectors, and it would create a $250 million grant program to prepare workers for the skills they’ll need in the future, especially in industries likely to see job displacement.

“Hopefully, by both readying the work workforce of today but also preparing the workforce of tomorrow, we’ll be able to catch the full opportunity that is the deployment of artificial intelligence,” Butler says.

Butler sees this as a moment in US history where policymakers could “get ahead of what we know is going to be eventual disruption and try to create a pipeline of opportunities that can again help to both stabilize our economies by creating equitable opportunity.”

But Butler is realistic about the dynamics of Congress and the upcoming election in just over a month. “You and I both know that this 118th Congress is rapidly coming to a close, with a lot of business in front of it right now,” she says. And Butler believes legislators still need to have important conversations with people representing different sides of the issue before advancing comprehensive AI legislation. And there’s also, she notes, the small issue of “getting through the next presidential election this November.”

Meta is bringing reminders to its Ray-Ban smart glasses, a feature it previewed during its Connect event last month. With the update, you can ask Meta AI to remember your surroundings, like where you parked, and even time up reminders to make a phone call.

The other features coming with the update include the ability to send and record voice messages on WhatsApp or Messenger without taking out your phone. You can now ask Meta AI to scan QR codes or call phone numbers that you come across, too.

Meta is also updating how you invoke Meta AI while wearing the smart glasses. Instead of saying “Hey Meta” before each question, you now only need to say “Hey Meta” for your first question and then ask any additional questions without the prompt. You also no longer have to tell Meta AI to “look and” when you’re asking about something you see.

Meta’s AI assistant is currently only available in the US and Canada. You can access the new features by updating the Meta View app on iOS and Android to version 186, which started rolling out on Wednesday. Along with this update, Meta revealed that its limited edition translucent Ray-Bans have already sold out online (but some may still be available in stores).

It’s been almost one year since the launch of the $299 Ray-Ban Meta glasses, and they’ve shown a lot of promise, especially when Meta rolled out the ability to process images, text, and audio earlier this year. In future updates, Meta will launch the ability for its smart glasses to translate speech in real time, starting with English, French, Italian, and Spanish.

It’s not clear if Meta plans to address the possibility of bad actors using the smart glasses as a way to dox people’s identities, which is an issue two college students recently called attention to.

Google is rolling out ads in AI Overviews, which means you’ll now start seeing products in some of the search engine’s AI-generated summaries.

Let’s say you’re searching for ways to get a grass stain out of your pants. If you ask Google, its AI-generated response will offer some tips, along with suggestions for products to purchase that could help you remove the stain. The products will appear beneath a “sponsored” header, and Google spokesperson Craig Ewer told The Verge they’ll only show up if a question has a “commercial angle.”

GIF: Google

Google has been testing ads in AI Overviews since May but says it’s moving forward with a full rollout because it helps people “quickly connect with relevant businesses, products and services to take the next step at the exact moment they need them.” For now, ads are only coming to AI Overviews in the US on mobile. Microsoft similarly includes ads in its Copilot chatbot and recently changed how they surface in responses.

Google is also making some tweaks to the formatting of AI Overviews. Following a test in August, Google will now display cited webpages more prominently on the right side of the summary, as it found this “has driven an increase in traffic to supporting websites compared to the previous design.”

It’s rolling out AI-organized search pages as well — a feature that displays a custom results page with relevant information instead of showing just a list of links. This is currently only available on mobile in the US for searches related to recipes and meal ideas.

If you can’t capture what you want to search for with just a picture, Google Lens will now let you take a video — and even use your voice to ask about what you’re seeing. The feature will surface an AI Overview and search results based on the video’s contents and your question. It’s rolling out in Search Labs on Android and iOS today.

Google first previewed using video to search at I/O in May. As an example, Google says someone curious about the fish they’re seeing at an aquarium can hold up their phone to the exhibit, open the Google Lens app, and then hold down the shutter button. Once Lens starts recording, they can say their question: “Why are they swimming together?” Google Lens then uses the Gemini AI model to provide a response, similar to what you see in the GIF below.

When speaking about the tech behind the feature, Rajan Patel, the vice president of engineering at Google, told The Verge that Google is capturing the video “as a series of image frames and then applying the same computer vision techniques” previously used in Lens. But Google is taking things a step further by passing the information to a “custom” Gemini model developed to “understand multiple frames in sequence... and then provide a response that is rooted in the web.”

There isn’t support for identifying the sounds in a video just yet — like if you’re trying to identify a bird you’re hearing — but Patel says that’s something Google has been “experimenting with.”

Google Lens is also updating its photo search feature with the ability to ask a question using your voice. To try it, aim your camera at your subject, hold down the shutter button, and then ask your question. Before this change, you could only type your question into Lens after snapping a picture. Voice questions are rolling out globally on Android and iOS, but it’s only available in English for now.

It had all the makings of a typical recruiting event for a tech startup in San Francisco. There was free food, drinks, and live music generated via code being written in real time.

But there were also mandatory metal detector screenings, ID checks, and security guards everywhere. It was held by Elon Musk at the original Mission District headquarters of OpenAI, which Musk cofounded before leaving after (reportedly) failing to take it over. And Musk was there to convince people to join his latest startup, xAI.

The timing felt intentional. That same day, OpenAI was hosting its annual Dev Day across town, where CEO Sam Altman had spoken hours earlier to a packed auditorium of developers. The Silicon Valley rumor mill was buzzing about OpenAI closing in on the largest round of funding ever for a startup, surpassing the amount Musk himself had just raised for xAI four months earlier.

Around 8:30PM, the AI-generated music cut off, and Musk, surrounded by bodyguards, climbed onto a table in a sectioned-off area to address the room of engineers. He began talking about why he started xAI and moved it to the same office where he helped launch OpenAI nearly a decade ago.

“We want to create digital superintelligence that is as benign as possible,” Musk said at the Tuesday gathering, according to a partial recording of his remarks shared with The Verge. He then called on those in the crowd “to join xAI and help build the intelligence and build useful applications to derive from that intelligence.”

For about an hour and a half, Musk took questions from the (predominantly male) audience, according to people in attendance. He said he believes we’ll reach artificial general intelligence (AGI) in a couple of years; he hopes to build a “supersonic jet company” next; he plans to open-source xAI’s models roughly nine months after they’re released; and most importantly, he wants to move fast. He compared xAI’s growth to the SR-71 Blackbird, an airplane that flew three times the speed of sound and provided the US with enemy information during the Cold War.

“No SR-71 Blackbird was ever shot down, and it only had one strategy: to accelerate,” Musk told the room per a post on X from an attendee.

He predicted that OpenAI, Anthropic, Google, and xAI will be the main players in the AI race for the next five years. The goal of the party was to find engineers for xAI’s API, one attendee said. Ultimately, he said he aspires for xAI to be as dominant in AI as SpaceX is in rockets.

At 10PM, the fire marshal put an end to the recruiting event. Musk was briskly escorted out a backdoor with his security detail. Attendees, including some wearing OpenAI backpacks, walked out into the night with pizza slices.

Maximum, truth-seeking AI

xAI began in March 2023 on the 10th office floor of X, Musk’s social media platform formerly known as Twitter.

He assembled a team drawn from his other companies, including Tesla and SpaceX, as well as his 17-year-old son, his cousins, and the son of Jared Birchall, who runs his family office, The Verge has learned. The mission: surpass OpenAI and deliver a competitive large language model in just three months.

Since then, xAI has expanded from a single floor at X to a larger office in the Stanford Research Park in Palo Alto. Musk tapped Igor Babuschkin, a former Google DeepMind researcher, to lead xAI. He also recruited researchers from OpenAI, Microsoft, and Meta.

In May, xAI secured $6 billion in funding from several high-profile investors, including Andreessen Horowitz, Lightspeed Venture Partners, and Sequoia Capital, valuing the company at $24 billion. Investors in X own 25 percent of xAI, which benefits from the wealth of training data the social media platform produces every day.

Under pressure from Musk, xAI’s first model was launched in late 2023 via Grok, a chatbot for paid subscribers to X. It has since released three updates: Grok-1.5, Grok-2, and Grok-2 Mini. But unlike competitors who had the luxury of time to develop their own systems, xAI’s lean team had to move fast and find a quick solution. One person familiar with the development of the first model described it as a patchwork product that relied on Microsoft’s Bing for search and Meta’s open-source Llama model for query rewriting.

Musk is still relying on outside technology for core Grok features. Just over a month ago, xAI announced a deal with Black Forest Labs to power image generation. The feature lacked guardrails put in place by other image generators, allowing people to generate photos of anything from Taylor Swift in lingerie to Kamala Harris with a gun. Musk said on X that xAI was working on its own generator but that the Black Forest partnership let it launch one in Grok more quickly.

A person familiar with what xAI is working on tells The Verge that voice and search features are in the works. The idea is that, like OpenAI and Meta’s voice modes, Grok will be able to talk back. Musk also wants it to provide summaries of news stories shared on X and trending topics.

Musk is now grappling with fierce competition in the AI race for top engineering talent and GPUs. As the richest person in the world, money isn’t an issue for him, though — despite the financial pressure that X’s plummeting value has created. While he’s raised billions of dollars for xAI, Musk doesn’t exactly need to create a profitable AI company any time soon — for him, taking down his runaway rival OpenAI seems to be satisfying enough.

Musk cofounded OpenAI with its CEO Sam Altman and a group of partners in 2015, but Musk quit the board only three years later. At the time, he cited “potential future conflict” due to Tesla’s focus on AI. Later, he claimed he quit due to disagreements with the OpenAI team. And in March, he sued the company, alleging (fairly dubiously) that it broke a contract with him and abandoned its mission. In response, OpenAI shared emails between its leadership and Musk that revealed a power struggle where Musk planned to take sole control over the company.

“I just don’t trust OpenAI for obvious reasons,” Musk said during the recruiting party. “It is closed, for-maximum-profit AI.”

The falling out between Musk and OpenAI has evolved into a stiff game of one-upmanship. This week, OpenAI raised $6.6 billion at a $157 billion valuation, just outpacing Musk’s historic funding round. Musk leveraged Tesla GPUs to build a data center nicknamed “Colossus” and reportedly brought online 100,000 advanced Nvidia chips last week. Meanwhile, Altman is on a global mission, meeting with UAE leaders, Asian chipmakers, and US officials to raise $7 trillion for 36 semiconductor plants and data centers, all aimed at advancing OpenAI’s pursuit of AGI. After Altman’s latest funding round, he reportedly asked backers to not invest in competitors like xAI.

Perhaps just as tough as nailing down funding or compute power is securing the top AI talent in Silicon Valley. The best researchers can easily earn millions, and the timing has never been better for them to launch their own startups. Many are driven by their own altruistic visions for AI’s future, making their choice of company often based on its mission and leadership. While Musk’s fame and bold visions give him an edge, it doesn’t mean the battle for talent is any easier.

So, Musk gathered a few hundred young engineers from rivals OpenAI and DeepMind — some fresh from attending his competitors’ developer conference that very day — to do what he does best: sell his vision of the future.

In Musk’s world, AGI isn’t controlled by companies like OpenAI or Google, who keep their best models private. Instead, it’s owned by him and shared with the world.

One potential draw for working at xAI, compared to a larger competitor like OpenAI, might be the opportunity to move faster and take bolder risks. With a small team and shorter product timelines, xAI offers a chance to innovate quickly, unlike larger, more cautious companies like OpenAI and Google. It may attract those eager to see rapid AI proliferation or who align with Musk’s techno-libertarian leanings — people who reject Silicon Valley’s “wokeness” in favor of a “super hardcore” work environment that prioritizes ambition and agility over corporate safeguards.

“My personal belief is that the best way to achieve AI safety is to have a maximum, truth-seeking AI,” Musk said at the recruiting party.

Like a lot of things at Musk companies, the event came together rapidly. “He said that he had the idea for this event last Wednesday and that the office was unfurnished then,” said one source who attended the party. But the source called it an “overall chill event” focused on answering questions about not just AI but also Musk’s many other enterprises. “Elon’s committed to winning and accelerating xAI,” another source said.