|
One of the first stunts Kevin Roose ever pulled on Hard Fork was to clone my voice. It was October 2022, ChatGPT had yet to be released, and one of the best products available on the market was made by a company named Play. After sending the company several hours’ worth of recordings of my voice, Kevin was rewarded with a clone that … didn’t really sound much like me at all. When he tried again a year later, the results were better, but still obviously AI. Kevin had used a startup called HeyGen to generate a video of me delivering lines in German and Hindi — two languages I don’t speak. While the voice was more recognizably my own, it still had plenty of rough edges. I had kept an eye on the technology, though, because I’ve been hoping to solve a problem. Ever since I started Platformer, readers have asked me to make the column available as a podcast. Lots of you would like to consume each day’s edition while you’re on the go, or doing chores around the house. But I struggled to imagine adding a daily podcast to my list of responsibilities: the last thing I want to do after sending out each day’s edition is to hop on a microphone. Eventually, when I would share this with readers who had inquired about a potential audio version, some started to ask whether I would consider using a voice clone. Many readers, I would learn, were already using AI text-to-speech services to listen to Platformer. This method would at least put the column back into my voice — a synthetic version of it, anyway. As it so happens, the startup ElevenLabs had released an uncannily good voice model last year called Eleven Flash v2.5. The company, which was founded in 2022 by former employees of Google and Palantir, offers voices of varying levels of fidelity. An “instant voice clone,” which can be made with barely a minute of audio, sounds better to my ears than the state of the art did back when Kevin first tried to copy my voice. The better, more labor-intensive offering is the “professional voice clone,” which suggests that you provide up to three hours of clean, high-quality audio of yourself speaking. The prospect of reading old Platformer columns into a microphone for multiple hours was repellent enough that I set aside the idea for several months. In September, though, I told you all that I planned to add an audio feed in the coming year. And once again, several of you responded that I should create a voice clone. And so, one lazy Saturday, I set to work. In truth, it had not been my first effort to create a clone. In hopes of avoiding more time in front of a microphone, I had asked a colleague at Hard Fork to send me some audio I could submit to ElevenLabs. The audio was sufficiently high quality, but all of it was taken from a show in which I talk excitedly about current events with my friend. The resulting clone was a complete tonal mismatch for reading my columns, which are typically more solemn. Listening to that clone read my columns, I feared that I sounded insane. And so I decided to record the audio from scratch. I read column after column in the same way they sound in my head when I write them. I submitted the resulting audio to ElevenLabs, and a few hours later I got an email telling me that the fine-tuning of my voice clone was finished. I pasted a paragraph into a new project in the company’s web app and generated the audio for my clone to speak. For the first time — truly — I could not believe my ears. I played it for my boyfriend, and a group of friends, and everyone agreed that it sounded exactly like me. The words sound like me, at least. The sentences and paragraphs don’t always carry the rhythms I want them to. Emphasis sometimes falls in the wrong place. For the most part, its pronunciation impresses me. But every once in a while, I stumble across a howler. It seems to struggle in particular with acronyms. During a test run yesterday, it pronounced CSAM “C-S-A-M,” which is understandable enough. Less understandable is why, a few weeks earlier, it had pronounced “CEOs” as “C-E-O-S.” There is also the question of whether voice-clone me sounds — and this is the word early listeners keep using — boring. My beats don’t often lend themselves to chipper, upbeat coverage, and you can really hear this in the voice clone. Still, for the most part, I don’t find the narration much duller than I might find the average news bulletin on NPR. The dry professionalism of the clone struggles to convey my more emotional writing. But it’s more than serviceable, and I imagine it will continue to improve as the underlying technology does. Moreover, putting a voice clone into action gives me a way to try new stuff. Five years into Platformer, and seven years into writing a daily newsletter, I find myself wanting to mix things up: to experiment with new formats, to learn new skills, and to see how else I can be useful to subscribers. Launching an audio feed feels like a worthy experiment in that regard — even though I fully expect the voice clone to give lots of readers the ick. Still, we’re going to give it a try. Starting today, you can add the free Platformer feed to your podcast player of choice. Each week, you’ll find a version of the column narrated by my voice clone. (Which we disclose in each episode.) Paid subscribers will get all three weekly editions in a custom feed that they can add to any podcast player. The paid feed will also become a place to launch further experiments, including Q&A sessions in which I answer your questions in my own, real voice. We’re now calling this whole package Platformer+, largely for alliterative reasons. (Paid subscribers received a separate email today with instructions for adding their feeds; basically, go here.) If any of this makes you want to upgrade your subscription, of course, this is a great day to do that. Already, some early testers have asked us to include our Following and Side Quests sections in the audio version of the newsletter; if you’d like that too, please let us know. We have some things to work through there (What voice should we use for the Following items? Isn’t a long list of short link summaries extremely boring to listen to?) We welcome your feedback. AI tools suck when they result in machine-generated slop substituting for human creativity taking over surfaces where people once had free rein. But they excel when they help individuals extend themselves further than they might otherwise be able to go. My hope is that this experiment turns out to be the latter. If it turns out to be the former, though, I’m confident you’ll let me know. Sponsored Copilot Money is a beautifully designed money tracker built to make understanding your finances feel calm and intuitive. As the only personal finance app to win an Apple Editor’s Choice award, Copilot is recognized for its clarity, thoughtful design, and ease of use. It brings all your accounts, budgets, investments, and net worth into one organized dashboard, so you can make decisions with confidence and clarity. Copilot Money makes managing your finances easy with AI-powered spending categorization and seamless syncing across iPhone, iPad, Mac, and web. It’s part of why the app holds a 4.8 App Store rating from more than 25,000 reviews. Final week: Platformer readers can get 26% off their first year + 2 months free with code PLATFORMER, available through the link below. FollowingSenate passes anti-Grok legislationWhat happened: The US Senate unanimously passed a bill that would allow victims of nonconsensual sexually-explicit AI images to sue the entities behind the creation of the images. The bill, known as the Defiance Act, is designed to build on the existing Take It Down Act, which was enacted last year and introduced criminal penalties for anyone who intentionally distributes explicit images without consent. (Platforms including X must have a system in place to handle user reports of these images by May.) This is the second time the Defiance Act has been unanimously passed in the Senate — it was previously passed in 2024, where it stalled in the House. Why we’re following: This is surprisingly quick action by the standards of Congress, where responses by lawmakers have, so far been lukewarm at best. The fast-tracking of this bill follows multiple probes and investigations into Grok around the world, including the UK, France, Australia and India. Indonesia and Malaysia have enacted outright bans on Grok. Meanwhile, both the Pentagon and Elon Musk have seemingly shrugged off the backlash — and announced that Grok will be integrated into US military systems and Pentagon networks. Are deepfake nudes the future of combat? What people are saying: Democratic Sen. Dick Durbin, a co-sponsor of the Defiance Act, made his case on the Senate floor: “Imagine losing control of your own likeness and identity. Imagine that happening to you when you were in high school. Imagine how powerless victims feel when they cannot remove illicit content, cannot prevent it from being reproduced repeatedly, and cannot prevent new images from being created. The consequences can be profound.” Musk continues to pretend the problem doesn’t exist. “Compared to other AIs, Grok is solid as a rock. And it will get much better. Eternally curious to know the deeper truth and appreciation of beauty are its goals,” he posted on X. —Lindsey Choo Microsoft tries to counter the data center backlash
What happened: In the wake of controversy about AI data center build-outs, Microsoft has announced a “community-first” data center initiative, promising to forego criticized practices such as seeking tax breaks for building data centers, paying discounted electricity rates, and buying up land under NDAs. The announcement comes a day after President Trump hinted at coming news from Microsoft, “who my team has been working with,” in a Truth Social post. In a blog post, Microsoft CTO Brad Smith emphasized that the company would attempt to avoid causing increased electricity prices, and work to replenish “more of your water than we use.” Why we’re following: It’s interesting to see the President, who has been fighting AI regulation at the state level, working to shape companies’ data center policies. The announcement certainly maps to Trump’s general pattern of negotiating directly to make “deals” with companies rather than getting legislation passed. It’s also a sign that public backlash to the effects of AI is affecting his policy priorities. Communities have increasingly resisted data center projects, citing issues including rising electricity prices, noise from data centers, and concerns over water use. In data center hub Virginia, electricity needs are approaching the limit of the power grid. Although data centers make up a small fraction of total industrial water consumption, they’re still significant enough to cause potential issues in areas with limited water supply. What people are saying: On Truth Social, President Trump said, “I never want Americans to pay higher Electricity bills because of Data Centers.” He added that data centers were great, but “the big Technology Companies who build them must “pay their own way.”” The language of Trump’s post was echoed in Microsoft’s commitments, which promised to “pay our way to ensure our data centers don’t increase your electricity prices.” Geekwire asked Microsoft CTO Smith about the common, and controversial, practice of buying land for data centers under non-disclosure agreements, to avoid land prices rising. “That is clearly not the path that’s going to take us forward,” Smith said. The companies who succeed at data centers “will be the companies that have a strong and healthy relationship with local communities.” —Ella Markianos Side QuestsMicrosoft warned that China is outpacing the US for AI users outside the west. The House passed a bill limiting China’s remote access to advanced US AI chips. California governor Gavin Newsom is seeking to neutralize the state’s proposed billionaire tax. A record $17 billion was stolen in crypto scams in 2025, with profits growing from AI scams, this report said. Roblox’s AI age verification system is already a mess. High five. OpenAI is set to run its second Super Bowl ad this year. Programmer Simon Willison’s first impressions of Claude Cowork. Anthropic chief product officer Mike Krieger is stepping down as chief product officer to to co-lead the company’s “Labs” internal incubator. Meta began cutting more than 1,000 jobs from its Reality Labs unit. Meta and EssilorLuxottica are reportedly considering doubling production capacity for the Ray-Ban smart glasses. AI influencers are sharing fake images on Instagram in which they appear in bed with celebrities. (Putting the "personal" in personal superintelligence!) How WhatsApp became a core technology around the world. Google will reportedly start making smartphones in Vietnam this year. Veo can now turn images into vertical AI videos. Apple announced Apple Creator Studio, a new subscription for its suite of creative apps to take on Adobe. Why live TV has been hard for Netflix to crack. Spotify’s new CEOs are hoping AI can help them fight algorithm fatigue. What the new agentic Slackbot can do. Warhammer developer Games Workshop banned its employees from using AI in its content or designs, amid growing discontent among gamers over the use of AI in development A look at Confer, an end-to-end AI assistant created by Signal creator Moxie Marlinspike. Biologists are now studying LLMs as if they are living things. Those good postsFor more good posts every day, follow Casey’s Instagram stories. (Link) (Link) (Link) Talk to usSend us tips, comments, questions, and podcast complaints: casey@platformer.news. Read our ethics policy here.
|
