🧠 Big Water Myth, a Bossy New Model, and Ads in 50 Languages

The viral story goes that AI is drinking the planet dry. The actual number is smaller than you'd think. Citing the Manhattan Institute, NVIDIA says US data centers account for roughly 0.2% of the country's daily water use, and that share is dropping thanks to a cooling switch.

Here's the switch. Older data centers use evaporative cooling towers, which lose water to the air. NVIDIA's newer setup runs liquid coolant at 45°C (yes, hotter than a hot tub), which lets facilities in cooler climates use dry coolers instead. That drops facility cooling water from about 2.6 million gallons per megawatt per year to near zero. Since cooling can eat up to 40% of a data center's electricity, you save power and water in one move. Musk reshared the post with a one-word reply: "True."

One honest caveat so you can use this at dinner parties: this covers direct cooling, and it works best in cool regions. Hot, dry states like Arizona still need extra help on the hottest days. But the "AI is a water hog" line just got a lot weaker.

Most AI progress comes from training one giant model. Sakana AI (the Tokyo lab co-founded by one of the authors of the original Transformer paper) went the other way. Their new release, Fugu, is an orchestration model. Instead of answering you itself, it routes each task to the best model in a swappable pool, then stitches the results together behind a single API.

Fugu Ultra reportedly matches Anthropic's Fable 5 and Mythos on tough benchmarks, despite not being a frontier model of its own. And because it just routes to whatever's available, it sidesteps the export controls that pulled Fable and Mythos offline earlier this month. The pitch is resilience. If one provider's access vanishes overnight, Fugu reroutes and keeps going.

Worth knowing before you build on it: Sakana keeps exactly which models it picks proprietary, so you don't see what's under the hood.

ElevenLabs launched Ads Engine inside ElevenCreative. You connect your Google, Meta, and LinkedIn ad accounts, it pulls your existing ads, and it localizes them across 50+ languages. For video, it uses Dubbing v2 to keep the original speaker's voice, emotion, and pacing intact, so it sounds like a real re-record. It also adapts text overlays and visuals per market, then pushes the finished creatives straight back to your platform.

For anyone running global campaigns, this collapses a process that usually means separate budgets and long production cycles for every market.

Heads up: at launch it's a point-and-click tool, with no developer API yet.

Quick context if you're new here. PAI Labs is our paid track where we hand you full AI workflows you can copy and run on your own business. This week's lab takes on the chore everyone pretends they keep up with: weekly competitor research.

You know the pattern. You bookmark a rival's newsletter, swear you'll read it, then hear about their big launch three days late from someone else's feed. This lab puts an end to that.

The setup works like a tiny newsroom. Tavily, a search connector with a free tier (1,000 searches a month, no card needed), plays reporter and pulls everything your competitors published in the last 7 days. Claude inside Cowork plays editor. It reads the whole stack, spots patterns across sources, scores each move against your positioning, and writes a dated brief that ends with the single highest-impact thing to do that week. Then it runs on a schedule without you touching it.

Thirty to forty-five minutes to set up once. About two minutes a week after that. The collecting, reading, and writing all happen while you're busy doing something else.

Read the full walkthrough here.

Quick context first: A prompt is a one-off instruction you type. An AI agent loop is a repeatable step you bake into your agent's routine so it runs every time, automatically.

This creator came up with an idea for a loop that uses both Codex and Claude’s Opus.

Here’s how it works: once Codex (OpenAI's coding agent) finishes designing an API, have it get a second opinion from Claude Opus using claude -p, the command-line way to fire a single prompt at Claude without opening the full app. Two models catching each other's blind spots noticeably lifts the quality of the code you ship.

NotebookLM (Google's research-and-notes tool) made flashcards fully customizable. You can edit the questions, tweak the answers, and add brand-new cards to build the exact study set you want, then share it with classmates or that one academic rival. Great for turning dense notes or a PDF into something you'll actually review.

Speaking of AI Agent Loops, here's a wild use case to try out: point a coding agent at all your tools and let it assemble a personal knowledge base overnight.

Reality check: it's only as good as what you plug in, and you'll want to skim what it builds before trusting it. But as a way to turn scattered notes into one searchable brain, it's a clever overnight project.

Here's the loop that inspired this