Welcome, humans.

Ever wondered how much energy ChatGPT ACTUALLY uses per query? Here’s what Sam Altman says:

Now, multiply those numbers by a billion queries per day (surely a conservative estimate at this point), and you have something similar to what it would take to power about ~11K average U.S. homes for a day and ~85K gallons of water.

On net, that’s actually not that bad (according to o3). But if Sam’s new predictions come to pass… those numbers will get A LOT bigger. For more on that, read below…

Here’s what you need to know about AI today:

• We break down Sam Altman’s top 5 “gentle singularity” predictions.

• Disney and Universal sued Midjourney over copyright violations.

• Sam Altman's Coco robotics raised $80M for delivery robots.

• Researchers found vulnerability stealing data from Microsoft Copilot.

Sam Altman says we’re approaching a “gentle singularity”…and it's surprisingly chill.

OpenAI's CEO just dropped a blog post called “The Gentle Singularity.” In it, he makes a series of bold claims (which we break down below).

If you don’t know, the singularity = when humans lose control over technological progress. So the blog is basically him saying: “Hey, remember that sci-fi nightmare where AI takes over the world? Yeah, it's happening right now, but it's more like upgrading from dial-up to fiber internet than Terminator.”

Here are his top 5 takes on what happens next:

1. The hardest part is behind us. Altman claims the scientific leaps to get to GPT-4 were the real challenge. Now, it's mostly an engineering and scaling problem. In his words: "most of the path in front of us is now lit, and the dark areas are receding fast.”

   1. Our take: This is also a convenient argument for a market leader to make—basically, don't bother trying to invent something new, because you’ll never catch up!

2. ChatGPT is already the most powerful human. Not literally, but in terms of raw impact. With hundreds of millions of users, Altman says the AI industry is building a “brain for the world” that will be “extremely personalized and easy for everyone to use.”

3. The "idea guys" are about to win. Sam says more people will create software and art, but the real winners will be those with great ideas. The main bottleneck in the future won't be technical skill, but a lack of good ideas.

4. Intelligence will be too cheap to meter. He says this might sound crazy, but what we have today would have sounded crazier back in 2020. Abundant intelligence is coming, and fast.

5. He gave us a product roadmap in disguise. Sam has a track record of telegraphing OpenAI's moves in these posts. Here’s what the 3-year timeline he presents in the article likely means for their products:

   • 2025 (Cognitive Agents): Think a supercharged Operator that can build an entire app from a single prompt, handling the coding, debugging, and deployment itself. o3-pro with tool use could very well be a preview of this…probably why they don’t show what it’s doing.

   • 2026 (Novel Insights): Deep Research evolves into an “AI Scientist” that can propose new drug compounds or identify market strategies humans missed.

   • 2027 (Robots): OpenAI won't build the hardware. They’ll build the “brain”—a robotics OS for partners like Figure AI to power the next wave of humanoid robots.

But he admits there's a huge catch: solving the safety and alignment problem. He knows that to avoid economic chaos, superintelligence must be widely distributed and controlled.

We asked o3 to fact-check Sam. We fed our model his new post and his old ones, asking it to analyze his credibility over time. We’ll let o3’s work speak for itself…

In another attempt, o3 shared how Sam’s vision of a “gentle” singularity hinges on three massive “ifs”:

• Solving the AI alignment problem in the next few years.

• Global regulators not strangling innovation.

• The world being able to build datacenters and find energy as fast as the models improve.

o3 says if those hold, it could be a smooth ride. If not, the risks he warned about a decade ago are very much still on the table.

FROM OUR PARTNERS

A New Gold Standard for Recommender System Research

In a world where real-world recommender data is typically locked away, something new has emerged—a production-scale dataset comprising 4.79 billion interactions, audio embeddings, time-based evaluation, and robust privacy protections. Whether you’re building the next Spotify algorithm or benchmarking your latest model, this dataset changes the game.

Written by Avi Chawla at Turing Post, this article examines how Yambda-5B, MovieLens, Amazon, Spotify, and others compare—and explores the reasons behind a significant leap forward for researchers and developers alike.

👉 Read the full breakdown on Turing Post

Prompt Tip of the Day

Our editor Corey Noles wrote a guide to working with o3-Pro you can check out here. The best prompt tip is for how to structure your prompts with it (via Ben Hylak):

1. Goal: Open with the single‑sentence mission.

2. Return Format: Tell the model how to hand the work back.

3. Warnings / Constraints: Add key guardrails like “Cite every stat” or “If unsure, say ‘INSUFFICIENT DATA’.”

4. Context: Give the models as much context as they can handle to avoid hallucinations.

5. Capabilities (new for o3-Pro): Explicitly ask it to use tools, such as Web search, File search, Code interpreter, and MCP (here are the tools available via API).

Treats To Try.

*Anything marked with asterisks is sponsored content. Advertise in The Neuron here.

1. *Incogni removes your personal data from the open internet so scammers and identity thieves can’t access it. Protect yourself online with Incogni—get 55% off with code NEURON.

2. Amazon Ads’ new Video Generator turns your product photos into realistic video ads showing your products in action—try it here.

3. Perplexity Finance now lets you download Excel models directly from finance pages for your financial modeling—free to try

4. Dia is a new browser that works as your AI assistant directly in your tabs to help you write, learn, plan, and shop using context from what you're browsing (waitlist only rn).

5. Atlas drops pricing pages and billing into your AI app instantly (when used w/ Lovable, v0, Windsurf or others), so you skip building payment systems yourself.

6. Flux1 Kontext has been used by a lot of people to restore old photos, and now, turn those photos into videos—here’s a demo + workflow on how to do it.

7. Higgsfield integrated Flux.1 Kontext to combine photo editing, cinematic motion, and VFX into one single pipeline.

See our top 51 AI Tools for Business here!

Around the Horn.

• Disney and Universal sued AI image generator Midjourney for allowing users to create their copyrighted characters, marking the first time major Hollywood studios have taken on an AI company.

  • Related: Getty Images’ lawsuit against AI image generator Stability AI began this week in the UK.

• Wikipedia editors forced the platform to pause AI-generated summaries over accuracy concerns, fearing hallucinations would damage credibility.

• Sam Altman’s Coco Robotics raised $80M for its last mile delivery robots, who have made 500K deliveries so far and use OpenAI’s AI (while OpenAI gets to use the delivery robot’s real world data).

• NVIDIA announced it will build its first AI cloud in Germany, and teamed up with Perplexity to help AI companies across Europe and the Middle East develop their own reasoning models in their native languages.

• Aim Security researchers discovered “EchoLeak”, a zero-click vulnerability that could steal sensitive data from Microsoft 365 Copilot by simply sending an email that tricks the AI into accessing and leaking information it shouldn't touch.

FROM OUR PARTNERS

Recraft: Professional design tools that elevate your creative process

When your design vision demands precision, Recraft's advanced tools help you craft exactly what you envision.

Ranked at the top on Hugging Face leaderboard for 4 months, trusted by 4 million users, including teams at Netflix and HubSpot, to enhance their creative workflow.

Elevate your designs with Recraft.

Thursday Trivia

One is real, and one is AI. Which is which? (vote below!)

A.

B.

Here are the results from last week’s poll:

Here’s what you said:

• A.P chose A: “In A there are two left hands” (nice catch!)

• E.B. chose B: “This is difficult. I made a wild guess.”

• J.T. chose A: “Saw the movie ;)”

• D.F. chose B: “Who holds a screwdriver like that and why have a sticker over your product name…that’s why this one seems fake.” (here’s your answer!)

• S.K chose A: “I'm going to be in trouble when AI learns how text works in images. It has mostly figured out hands now.”

A Cat's Commentary.

Trivia Answer: A is AI, and B is real.