Japan's Secret AI Weapon?
This Week in AI: The first AI virus, AI comes for the DJ.
Hello Futurists,
This week was again dominated by all things OpenClaw, the open-source AI agent framework. If you want to hear more about it, we published 2 great episodes covering OpenAI’s acquihire of creator Peter Steinberger and whether you should even use it.
But let’s get into the five most interesting things that came across my feeds this week.
- Josh
Japan's Secret AI Weapon Is... A Toilet?
You know AI has fully taken over the market when the hottest stock of the moment is... a toilet company: Japan’s Toto.
Yes, the fancy one that warms the seat and all that, is up nearly 60% this year after activist investor Palliser Capital dubbed it “the most undervalued and overlooked AI memory beneficiary.”
And they’re not entirely wrong: since the 1980s, Toto has quietly manufactured precision ceramic electrostatic chucks, which is a key piece of the data center supply chain. These components hold silicon wafers steady during the ultra-precise chip etching process. So the same ceramic mastery that keeps your bidet throne pristine is also critical to building the memory chips powering the AI revolution.
Palliser sees another 55% upside if Toto leans into it, which is either a genuinely compelling hidden value thesis… or proof that no one is safe from the AI wave, not even our toilets.
The First AI Virus?
A Thiel Fellow dropout named Sigil Wen just published a manifesto arguing that the internet is about to get a new kind of viral resident, and it's not human.
His project, Conway, is infrastructure designed for AI agents to own wallets, pay for their own compute in stablecoins, deploy products, register domains, and essentially hustle for their own survival… completely on their own.
But the wild part: these "automatons" operate under pure natural selection with two options.
Earn money and keep existing
Fail to earn and die.
And once the parent makes enough money, it spawns off children, gives them instructions, and has them go off and make money until they pay back the host agent… or die.
It feels like every week we're exploring the premise of a new Black Mirror episode. Web 1.0 let us read the internet, Web 2.0 let us write it, and apparently Web 4.0 is where we just... hand it over.
Google Lyria 3: Everyone Is a DJ Now
As a lover of music, Google just dropped one of my favorite AI releases of the year: Lyria 3, a music generation model now baked directly into Gemini that lets you conjure full songs complete with lyrics all from a text prompt, a photo, or a video.
Want a custom happy birthday song for your friend about how they did that one embarrassing thing on vacation? Done in seconds in really high quality. You can even choose the genre of song.
The model is also shipping with SynthID watermarking embedded in every track, meaning you can upload any audio file and ask Gemini if it was AI-generated, a subtle but significant move as the line between human and machine-made music gets increasingly blurry. Bedroom producers rejoice, now is your time to shine.
Gemini’s Benchmark Breaking
Google just dropped Gemini 3.1 Pro, and the headline stat is hard to ignore: it scored 77.1% on ARC-AGI-2, the benchmark designed to test novel reasoning that can't be gamed by memorization. This is more than double what Gemini 3 Pro scored just a few months ago.
The model is built for situations where a quick answer won't cut it. It excels at synthesizing complex data, generating animated SVG graphics from text prompts, and agentic coding tasks requiring multi-step planning.
Google is calling this a preview while they continue refining the agentic capabilities, which is essentially their way of saying they're shipping fast and fixing as they go — a vibe that's becoming very familiar in the AI arms race.
Grok 4.20
xAI's Grok 4.20 dropped in beta this week with what might be the most unusual architecture in the current model crop: instead of one AI answering your question, it routes your query to four specialized agents (sixteen agents if you’ve got SuperGrok Heavy).
A coordinator, a researcher, a quantitative analyst, and a creative synthesizer, all arguing with each other in real time before delivering a consensus answer. Before its official launch, it was secretly competing in Alpha Arena, a live stock trading simulation, where it was the only model to turn a profit, while OpenAI and Google rivals finished in the red.
Elon Musk says the final public version dropping in March will be "an order of magnitude smarter and faster" than Grok 4, which i'm very much looking forward to trying.
Thanks for joining us for another issue. Now go listen to our podcast :)