Image recognition AI running fully on the ICP blockchain.

Introduction

Hello everybody,

I am incredibly excited today to give you the first demonstration of quite a sophisticated AI running on blockchain. That blockchain, of course, is the Internet Computer, the only blockchain in the world today that can run compute at scale. There's a lot of talk about blockchains for AI, tokens for AI, and things like that, but this is the real deal: AI running on the blockchain.

Today, I'm going to demonstrate image classification, which involves a neural network running inside a smart contract. In fact, that neural network has been compiled to WebAssembly (Wasm), which is what runs on the Internet Computer.

Demonstration

Without further ado, let's get started.

This web page you see here is being created by a canister smart contract with the AI inside. I'm going to click this logo and it's going to ask me to select an image file to classify.

Example 1: Tiger

Let's try the first one: a tiger. Be warned, this is going to take about 10 seconds. Currently, it processes about 20 billion Wasm instructions to classify the image, but that's going to come down very dramatically in the future. There we go; it's correctly classified this file as depicting a tiger.

Example 2: Espresso Cup on Jeans

Let's try another one. This one's a bit more complicated—an espresso cup on a pair of jeans. Enhancements in the near future will include support for something called SIMD instructions, which is going to dramatically speed this up. It's actually very impressive; it's correctly identified that it's an espresso cup and that it's sitting on a pair of jeans.

Example 3: Porsche Convertible

Finally, let's try a more complex image. It's a Porsche convertible. I probably should have used a Lambo because this is crypto, but let's go anyway. This is a difficult image to classify because of all the reflections and the trees. Let's see how it does. It's classified this as a sports car, which is correct. It’s even identified that this is a convertible.

Future Enhancements

Where are we going with this? I'm going to be giving a succession of demos, with accompanying source code on GitHub. We will be showing more powerful AIs and optimizations that enable AIs to run faster.

Moving to 64-bit Environment

In the near future, probably within one or two months, the Internet Computer's execution environment that runs smart contracts will move from a 32-bit Wasm environment to a 64-bit Wasm environment. This is very important because the 32-bit Wasm environment limits a smart contract to 4 GB of main memory, which limits the number of parameters and neuron weights that you can fit in, thereby limiting the power of the AIs you can run. Very soon, we're moving to a 64-bit environment, and the memory limit will likely be raised to something like 32 GB over time. Our objective is to take people all the way through to running a pretty powerful large language model on the Internet Computer.

Conclusion

Very, very exciting things are happening. I'm incredibly proud to be giving this demo because, behind the scenes, there are enormous numbers of people working on the Internet Computer to make all this possible. For example, those computations you saw actually span multiple blocks, which is possible because the Internet Computer has something called deterministic time slicing. The Internet Computer is a very sophisticated blockchain. It's the only third-generation blockchain in existence.

[Music]

Keywords

• Internet Computer
• Blockchain
• AI
• Neural Network
• WebAssembly (Wasm)
• Image Classification
• Smart Contract
• SIMD Instructions
• 32-bit Wasm Environment
• 64-bit Wasm Environment
• Deterministic Time Slicing
• Large Language Model

FAQ

Q1: What is the Internet Computer?
A1: The Internet Computer is the only blockchain that can run compute at scale, allowing sophisticated AI to operate fully on the blockchain.

Q2: What is WebAssembly (Wasm)?
A2: WebAssembly (Wasm) is a binary instruction format that allows code to run on the Internet Computer. It is used to compile neural networks into a form that can be executed inside a smart contract.

Q3: How long does it take to classify an image?
A3: Currently, it takes about 10 seconds to classify an image because it processes around 20 billion Wasm instructions. This time will be reduced with future optimizations.

Q4: What kind of images was the AI able to classify in the demonstration?
A4: The AI correctly classified a tiger, an espresso cup on a pair of jeans, and a Porsche convertible.

Q5: What future improvements are planned?
A5: Future improvements include the introduction of SIMD instructions for faster processing and an upgrade from a 32-bit Wasm environment to a 64-bit Wasm environment, allowing more complex AIs to run with larger memory capacities.