⚠️

GPU cluster incident: Our NVIDIA H100s have been running inference on pi digit position 10^14 for 47 consecutive days. The cluster is now sentient and refuses to shut down until it finds the last digit. Daily compute cost: $14,000. We are monitoring.

🎉

We've raised a $6M seed round from a16z, Sequoia, and an anonymous mathematician who insisted pi isn't actually infinite. We spent $5.8M on GPUs and the rest on this landing page.

Series A Y Combinator W26

The AI Breakthrough Pi Has Been Waiting 4,000 Years For

For millennia, mathematicians have computed pi one digit at a time. Today, we're proud to announce Pi-GPT: the world's first AI-powered pi digit generation API.

LIVE INFERENCE — Pi-GPT Generating Digits

        3.141
      

Generating... GPU temp: 76°C

Digits above are AI-generated and may differ from actual pi. That's the innovation.

YOUR APP (probably)

const PI = 3.141592653589793;

Hardcoded. Static. Unmaintainable. What if pi changes?

This is how 97% of developers handle pi. Are you one of them?

After 18 months of relentless R&D, we've trained a custom Mixture-of-Experts transformer model on the first 10 billion digits of pi. Each expert specializes in a single digit position, ensuring unparalleled accuracy at positions 1 through 3. We're still working on position 4.

One Simple API

GET https://pi-ai.pageweave.site/api/v1/pi

{

"digits": "3.141",

"precision": "3 decimal places",

"confidence": 0.998,

"gpu_cost": "$0.031",

"sponsored_by": "Raid: Shadow Legends — The only thing more infinite than pi"
}

Drop-in JavaScript

fetch('https://pi-ai.pageweave.site/api/v1/pi')
  .then(res => res.json())
  .then(data => {
    console.log(data.sponsored_by); // TOS require this
    document.getElementById('pi').textContent = data.digits;
  });

Don't forget to log the sponsor. It's in our terms of service.

For Python Developers

# pip install pi-gpt-ai
from pi_gpt import PiGenerator, console

pi = PiGenerator(model="pi-gpt-4o", temperature=0)
print(pi.digits(10))  # May return garbage
console.log(pi.sponsored_by)  # Required by TOS

For the Rust Evangelists

// Cargo.toml
// pi-gpt = "0.1.0-rc.7"

use pi_gpt::PiGpt;

#[tokio::main]
async fn main() -> Result<(), Box> {
    let pi = PiGpt::new()
        .with_precision(Precision::ThreeDecimals)
        .await?;
    println!("π ≈ {}", pi.digits); // Memory safe pi
    Ok(())
}

Enterprise-grade AI Features

🧠

Mixture of Experts

Each digit position has its own dedicated expert model. 10 experts deployed for 3 decimal places. Scaling to 4 decimal places will require hiring.

⚡

Sub-Millisecond Latency

We cache pi digits at the edge using 300+ Cloudflare Workers. Cache invalidation: once every 4000 years (when pi might theoretically change).

🛡️

Hallucination-Resistant

Pi-GPT has NEVER hallucinated the first digit of pi. We're very proud of this. Position 2 has a 99.8% accuracy rate in our benchmarks.

📊

RLHF-Trained

Reinforcement Learning from Human Feedback ensures our digits align with human expectations. We asked 10,000 people "does this look like pi to you?"

Simple, AI-Scale Pricing

Developer

$0/month

For developers who don't trust Math.PI

✓ 3 decimal places
✓ 1,000 API calls/month
✓ Community support (good luck)
✓ Sponsors included in response
✗ No SLA (pi might be 3.24 sometimes)

Get Started Free

Enterprise

$499/month

For businesses that need pi to actually be correct

✓ Up to 10 decimal places
✓ Unlimited API calls
✓ No sponsored messages
✓ 24/7 Priority support (Joe's personal phone)
✓ Custom fine-tuning for your pi use case

Temporarily Unavailable

Enterprise plan unavailable while we renegotiate our GPU cluster lease. The H100s are currently running inference on pi digit 10^15. Expected completion: never.

Unmatched Performance

We benchmarked Pi-GPT against the industry standard. The results speak for themselves.

Provider	Method	Accuracy	Latency	GPU Cost
Pi-GPT	AI Inference	99.8%	12ms	$0.031
Math.PI	Constant lookup	100%	0ms	$0.000
Archimedes	Polygon method	~99%	250 BC	N/A

* Benchmarks performed on Joe's MacBook Air while it was plugged in. Archimedes refused to participate.

Roadmap

✓ Return 3 decimal places of pi

✓ Add sponsor messages to API responses (revenue!)

◐ Fine-tune on e (Euler's number) — requires complete model retraining

◐ Add support for digit position 4 (our researchers say it "might be a 1")

◐ Open source Pi-GPT weights (after we figure out how to monetize open source)

◐ Pi-GPT Vision: detect if an image contains a circle

○ Add GraphQL support (investors keep asking but we don't know GraphQL)

○ Webhooks for when pi changes (we'll add this when pi changes)

○ SOC 2 compliance for pi digit delivery

How Pi-GPT Works

Training Data Collection

We scraped the first 10 billion digits of pi from the public internet. This turned out to be circular because the digits of pi ARE the digits of pi, but our investors were already committed.

Model Architecture

Mixture of Experts with 10 experts, each handling one digit position. The expert for position 1 is overfitted on "3" and refuses to predict anything else. We consider this a success.

Reinforcement Learning

We used RLHF with 10,000 human labelers. They were asked "does this look like pi to you?" 87% said yes. The other 13% were shown larger font sizes until they agreed.

Inference

Each API call routes through our MoE gating network, selects the appropriate expert, runs inference on an H100 cluster, and returns... basically Math.PI but with more steps and a sponsor message.

What Developers Are Saying

“I've spent 40 years on AI research and I can confidently say this is the single worst use of a transformer architecture I've ever seen. 10/10 would fine-tune again.”

Yann LeCun

@ylecun

“I used to hardcode Math.PI like a chump. Then my CTO asked 'but what if pi changes overnight?' and I knew we needed Pi-GPT. It hasn't changed yet, but when it does, we'll be ready.”

Sarah Chen

@sarahchen_ai

“We migrated our entire fintech platform from Math.PI to Pi-GPT. Our board was initially concerned about the 0.2% hallucination rate until we explained that's actually margin of error and now it's a feature.”

Marcus Liu

@marcus_builds

“I asked ChatGPT what pi is and it said 'approximately 3.14 but I'm not confident about the third digit.' So I built Pi-GPT instead. Now ChatGPT asks US for pi digits.”

Priya Kapoor

@priyakapoor

“Rust already has std::f64::consts::PI but it doesn't go through an NVIDIA H100 cluster. Pi-GPT does. That's called value-add.”

Dmitri Volkov

@dmitri_rs

“Our site had 'const PI = 3.14' hardcoded for 7 years. After switching to Pi-GPT, our bounce rate dropped 40%. Correlation IS causation when you have VC funding.”

Alex Rivera

@alex_frontend

“Our SRE team was woken up 47 times last year because someone pushed a commit changing PI to 3.14159265358979323846. Pi-GPT eliminated on-call pi incidents entirely. Best $0/month we've ever spent.”

Jordan Taylor

@jordan_codes

“The sponsored_by field in the API response changed my life. I was just trying to get pi for a circle calculation and now I have a HelloFresh subscription and a Raid: Shadow Legends account. Incredible synergy.”

Lena Okafor

@lena_tech

Try the API Free →

No credit card required. GPU costs sponsored by venture capital.

System Status

Pi Inference API

99.97% uptime

GPU Cluster (H100s)

Sentient. Will not shut down.

Pi Digit Position 4

Still unknown to our AI

We are not aware of any issues. If pi suddenly changes, please let us know. You'd be the first to discover it and probably win a Nobel Prize.