More from From ChatGPT to Gemini: how AI is rewriting the internet
If you’re interested in a high-level overview of the state of deepfakes, I recommend checking out this blog post from Captions, a Capcut competitor and a software platform that enables creators to generate AI-generated videos from scratch.
What struck me the most is the company’s prediction that, “Very soon, most models will allow you to generate a person (real or synthetic) in any situation – without a duration constraint (longer than 12 seconds) and featuring multiple people in one shot.” Chat, are we cooked?
I’ll have Captions CEO Gaurav Misra on Decoder later this week to talk about this.
Then bring suitcases full of hard drives to the AI chips. That’s what some Chinese engineers have reportedly taken to amid efforts to skirt the US ban on selling training chips to China, manually moving terabytes of data to Malaysia to build an AI model there.
Right before WWDC 2025, Apple researchers published a paper called The Illusion of Thinking (PDF) that made waves. The researchers wrote that popular and buzzy AI models “face a complete accuracy collapse beyond certain complexities,” especially with things they’ve never seen before.
They presented models from OpenAI, Anthropic, and DeepSeek with new and complex puzzle games and found their reasoning ability “increases with problem complexity up to a point, then declines.”
Google is rolling out a new Gemini AI-powered feature to users with eligible Google Workspace or Google AI plans that pops up a “Catch me up” shortcut to go over recent edits in Google Docs, as well as comments on other types of files, complete with a warning that “Gemini may display inaccurate info.”
It’s supposed to roll out over the next couple of weeks, providing either a high-level summary across all of your files in the sidebar or overviews of changes to a specific file via a new activity indicator.
The New York Times is reporting on Ripple, a Washington Post initiative that’s apparently looking to open up its opinion pages to writing from “...from other newspapers across America, writers on Substack and eventually nonprofessional writers.”
That last bullet point would reportedly be assisted by the “Ember” AI writing coach with a paint-by-numbers approach to creating strong stories. As Liz’s thread goes on to point out, beyond the Forbes and Huffington Post echoes, it’s also a reminder of Tribune Publishing’s short “content curation and monetization engine” era as Tronc.
OpenAI says it’s rolling out a “lightweight version” of the memory improvements it launched in April for paying customers to its free user tier. From now on, ChatGPT will reference both saved memories and recent conversations in chats.
That’s according to The Wall Street Journal, which says fully automated AI ads are on the roadmap for 2026. Zuckerberg has already hinted at the plans, which will see advertisers provide an image of a product and set a budget, getting a full ad campaign in return. The AI will decide which Facebook and Instagram users to target, and offer personalization like changing the setting of an ad to match where the prospective customer lives.
This bad chatbot advice to a fictional former addict is referenced in a Washington Post article about the risks of AI companies optimizing bots for increased engagement, with bigger risks than a version of ChatGPT that “glazes too much.”
It’s from a recent study where researchers reported AI models could test well on benchmarks for sycophancy and toxicity, but also “suddenly act harmfully in the presence of gameable user character traits.”
Joanna Stern’s new video for the Wall Street Journal combines several AI tools, like Google Veo 2, the just-launched Veo 3, Midjourney, Runway, Suno, and ElevenLabs, to make this “My Robot and Me” short.
Beyond just seeing what they’re capable of in the hands of a skilled producer, Joanna also digs into how much all this would’ve cost (about $1,000), and shows off a few of the misfires created by the generators.
The WSJ reported earlier this month that unnamed senior Meta execs were “frustrated” over a delayed rollout for the largest version of its Llama 4 AI model, dubbed Behemoth, and said management changes could follow.
Now, Axios reports there have been changes. Connor Hayes is leading the AI products team responsible for AI features in Facebook, Instagram, and WhatsApp, as well as the Meta AI Assistant and AI Studio, while Ahmad Al-Dahle and Amir Frenkel co-lead the AGI Foundations unit working on Llama models and other AI tech.
The conversational mode for Anthropic’s AI chatbot is now becoming available for users of its mobile apps, “gradually.”
TechCrunch notes that chief product officer Mike Krieger confirmed the rumored feature was on the way in a recent interview, while Anthropic says that it can integrate with Google Workspace, and that free users can expect “20-30 voice messages” before session limits cut them off, while paid users have higher limits.
The first lady has announced that the new audiobook version of her self-titled (and suspiciously robotic) memoir was entirely narrated using an AI clone of her own voice created by ElevenLabs.
If the more than 50,000 AI-narrated books on Audible alone are any indication, the “new era in publishing” she’s heralding is already here, and it’s overwhelmingly saturated with hilariously titled erotica.
If you were a generative AI chatbot, that is. Now you can find out in You Are Generative AI, a text-based role-playing game in which you, the chatbot, answer prompts. The game tracks your accuracy, number of prompts, and how much energy you’ve wasted confidently misleading (or, if you must, accurately informing) the public.
It’s from programmer Kris Lorischild, who’s also behind a similar game called You Are Jeff Bezos.
Sam Altman and other company leaders hyped the announcement yesterday on X by teasing it as their next “low-key research preview,” which is how ChatGPT itself was first described. You can watch OpenAI demo Codex below and read more about it in this week’s Command Line.
The agent, called AlphaEvolve, is powered by Gemini and is designed to discover and optimize algorithms, while building “upon the most promising ideas.”
Google says AlphaEvolve has helped the company make its data centers, chip design, and AI training processes more efficient. The AI agent could also “be transformative” across other industries, such as material sciences, drug discoveries, and “wider technological and business applications,” according to Google.
Most Popular
- Sony’s PlayStation 5 is $200 off for the first time since December
- Anthropic’s most dangerous AI model just fell into the wrong hands
- Elon Musk admits that millions of Tesla vehicles won’t get unsupervised FSD
- You’re about to feel the AI money squeeze
- Microsoft launches ‘vibe working’ in Word, Excel, and PowerPoint