AI Is Quietly Boiling Us Dry

Welcome back apprentices! 👋

Ever asked ChatGPT to write a poem about lasagna in the style of Shakespeare? Cool. 

But here’s the less cool part: somewhere in the world, a data center may have just evaporated hundreds of liters of freshwater to make that happen.

We’re not kidding. The smarter our AI gets, the thirstier it becomes — and it’s quietly draining precious water from some of the driest places on Earth. If you thought AI was “in the cloud,” turns out it’s also in your local watershed.

Read on to find out how Big Tech’s biggest brains are boiling off the planet’s most essential resource.

In today's email

• Polite Prompts, Pricey Consequences

• Profluent’s ChatGPT for Proteins 

• Perplexity AI Is Going Mobile

• AI’s Evaporating Our Water Supply

• Meta's New AI Drop

• Even more AI magic

• Test the prompt

_{Read Time: 4 minutes}

Quick News 

🤑 Sama has revealed that users saying “please,” “thank you,” and other sweet nothings to ChatGPT are actually making the model work harder — and costing the company more in compute. Those extra tokens may seem harmless, but at hyperscale, niceness comes with a cloud bill. The irony? AI alignment worked so well that now everyone’s treating it like a sentient intern with feelings.

🧬 Profluent just dropped ProGen3, an AI model that writes proteins like Shakespeare with a microscope. It’s trained on 3.4 billion protein sequences and can whip up designer antibodies and mini gene editors faster than you can say “CRISPR, who dis?” Bonus: it works in the lab, not just in theory — yes, scientists actually got it to do biology stuff on purpose.

📱 Perplexity AI is teaming up with Samsung and Motorola (yes, even the Razr is back) to bring its real-time, fact-fetching super assistant straight to your smartphone. Think ChatGPT meets Google Search, minus the pop-up ads and existential dread. It’s fast, it’s context-aware, and it might just steal your job as the go-to answer person in meetings.

Ethics & AI 
AI Is Cool — Too Cool. It’s Evaporating Our Water Supply

Generative AI might be generating poems and business plans, but it’s also generating a giant, invisible steam cloud. 

Behind your chatty chatbot and recommendation algorithm is a sweat-dripping GPU sitting in a giant, humming warehouse somewhere in Arizona or Spain — chugging electricity and demanding constant cooling. 

And how do you cool a thirsty supercomputer? With water. A lot of it.

Big Tech’s Hidden Thirst

A joint investigation by Source Material, Impakter, and The Guardian reveals that Microsoft, Amazon, and Google are building and operating massive data centers in regions already struggling with water scarcity. 

• Microsoft: In its ESG report, the company disclosed that 5.3 billion liters of water — 41% of its total water withdrawal — came from regions experiencing water stress.
  

• Google: Reported that 15% of its water use was from high-scarcity areas, with another 16% from regions teetering on the edge of water trouble.
  

• Amazon: Is building new data centers in northern Spain, a region already fighting desertification, with approvals to use 755 million liters of water per year — enough to fill 300+ Olympic swimming pools.

That’s not cloud computing; that’s a techno-thirst trap.

What It Took to Teach GPT-3

To train GPT-3, researchers estimate around 700,000 liters of fresh water were evaporated.

That’s just the training phase, not including what it takes to keep the model running 24/7 while you ask it to write wedding vows in the voice of Gandalf. And unlike your shower water, which gets a second life in the sewers, cooling water simply vanishes into thin air — literally. It evaporates. Eventually it might come back down as rain... somewhere... someday... maybe not in your zip code.

When AI Gets Thirsty, the World Pays

• Tech’s quietest resource war might not be over rare earth metals or GPUs — it might be over H₂O.
  

• Building AI in dry places because land is cheap and nights are cool is starting to feel like hosting a barbecue in a library—bad vibe, worse outcomes.
  

• We’re not just training AI models — we’re potentially untraining ecosystems in water-stressed regions.
  

• Water doesn’t care if it’s cooling a supercomputer or watering crops. But people do. And in Aragon, Spain, farmers are already sounding the alarm.

So… What’s the Deal?

We’re in a race to build smarter machines, but that race is running through some very dry finish lines. 

If we don't rethink where and how we host this AI explosion, we may find ourselves needing an AI to solve the climate and water problems AI helped cause. Irony, thy name is infrastructure.

Next time you prompt an AI model to write a limerick about your cat, remember it may have come at the cost of someone’s irrigation water. Tech won’t save us if it quietly boils the planet dry — so push for AI that’s not just smart, but hydrologically humble.

Meta
AI That Sees, Thinks & Collaborates

Meta’s FAIR team just launched a quintuple threat of open-source AI projects so advanced they practically deserve a superhero movie. 

Think: AI that can spot a goldfinch photobombing your forest footage, robots that understand “that red mug on the left,” and language models that argue with themselves until they get smarter (no, really). From perception to teamwork, these tools aren’t here to play — they’re here to build AI that’s one TED Talk away from full-time employment.

Whether you’re coding an AR butler or teaching a robot to clean your garage, this release screams “Welcome to the future — BYO firmware.”

The 5 Breakthroughs, Simplified

• 🔎 Perception Encoder
  Think of this as AI’s new set of glasses — x-ray vision included. It can identify stingrays hiding under sand and tiny birds mid-blur. Also aces captioning, visual Q&A, and probably judging your Instagram food pics.
  

• 🎥 PLM + PLM-VideoBench
  A video-savvy model trained on more footage than a Marvel movie binge. Handles fine-grained activities like “man almost slips on banana peel” and understands time like your most dramatic friend.
  

• 📦 Locate 3D
  Tells robots exactly where your keys are when you say “they’re on the table next to the thing.” Converts natural language into real-world object-finding power — your Roomba’s about to level up.
  

• 💬 Collaborative Reasoner
  AI agents that collaborate like a team project... but without ghosting each other. They solve problems better together and practice arguing with themselves, making them oddly relatable.
  

• 🔡 Byte Latent Transformer
  Tosses out traditional tokenizers like last year’s app stack. Runs faster, breaks fewer things, and handles weird data quirks like a champ. Byte-sized brilliance, zero bloat.

What’s the Deal?

Meta just rolled out the IKEA set of future AI — but with the instructions included. 

Whether you're building smart assistants, industrial bots, or apps that can “see” and “understand,” these releases are next-gen essentials. This is AI stepping out of the lab and into real life — 3D-aware, language-ready, and socially capable.

Trying to stay ahead of the curve? Follow open-source AI drops like this and play with the demos. You don’t need a PhD or a lab coat — just curiosity and Wi-Fi. Today’s research project might be tomorrow’s killer feature on your app, job, or startup.

Help Your Friends Level Up! 🔥

Hey, you didn’t get all this info for nothing — share it! If you know someone who’s diving into AI, help them stay in the loop with this week’s updates.

Sharing is a win-win! Send this to a friend who’s all about tech, and you’ll win a little surprise 👀

Even Quicker News 

📉 Turns out OpenAI’s new O3 model isn’t quite the genius it was hyped to be — its benchmark score came in lower than expected, raising some awkward eyebrows. Moral of the story? Even AI can oversell its résumé.

🍏 Apple’s making its AI smarter without snooping, thanks to some privacy-preserving math wizardry. It’s basically teaching Siri new tricks without letting her read your diary. 

🚀 Anthropic just built a dream team to boost Claude’s AI powers across Amazon’s cloud empire — because what’s $4B between friends? Think of it as hiring a pit crew for a rocket that’s already mid-launch.

Today’s Toolbox

🧠 Google’s latest AI model gives devs a “thinking budget” so you can literally tell your AI how smart (or fast) to be. It’s like cruise control for brainpower — speedy, cheap, and surprisingly clever.

🌍 Hence just dropped an AI advisor to help businesses survive the global trade war without breaking into a cold sweat. It’s like having a geopolitical therapist in your tech stack—minus the couch, plus real-time analytics.

⚡ IBM just gave its Granite model a brain boost with tools that catch hallucinations, cite sources, and even rate how confident it feels — basically turning your AI into a know-it-all with receipts. And with its new “activated LoRAs,” it’s not just smart — it’s fast enough to finish your thought before you do. 

🧪 Test the Prompt

A playground for your imagination (and low-key prompt skills).

Each send, we give you a customizable DALL·E prompt inspired by a real-world use case — something that could help you in your business or job if you wanted to use it that way. But it’s also just a fun creative experiment.

You tweak it, run it, and send us your favorite. We pick one winner to feature in the next issue.

Bonus: you’re secretly getting better at prompt design. 🤫

👑 The winner is…

Last week, we challenged you to test GPT-4o’s visual generation skills with this prompt.

Here’s the WINNER:

Congrats to Lewis from New Haven!🥳

Want to be featured next? Keep those generations coming!

🎨 Prompt: Your Focus Mode, Visualized

“Create an image of a snapshot of your ‘focus mode’ as a futuristic vehicle shaped like a [animal], speeding through a landscape made of [daily distraction]. Add motion blur, glowing energy trails, and a tiny co-pilot AI assistant trying to keep up.
Style: stylized racing concept art, mid-action shot.”

We’ll be featuring the best generations in our next newsletter!

_FEEDBACK