Progress

The long term

Our roadmap is centered around a singular goal: to preserve as many AI models on-chain as possible. To achieve this, we basically need to solve three key Crypto x AI problems over time:

1. On-chain model size limit: the bigger, the better

2. On-chain AI computation: the faster, the better

3. On-chain model architectures: the more supported, the better

2023: Minimum viable protocol with 1 MB model size

In our first year, we focused on building the foundational components for Eternal AI and delivering a minimum viable protocol.

✅ On-chain model size limit: 1 MB

✅ On-chain AI computation: CPU

✅ On-chain model architectures: Perceptrons

2024: Production-grade infrastructure with 100 MB model size

In our second year, we aim to scale Eternal AI from a minimum viable protocol to a production-grade computing infrastructure.

• On-chain model size limit: 100 MB

• On-chain AI computation: GPU with CUDA

• On-chain model architectures: CNN, RNN, LSTM, Autoencoders, Transformers

2025: Scale to 1 GB model size

In our third year, we plan to further scale our capabilities.

• On-chain model size: 1 GB

• On-chain AI computation: TBD

• On-chain model architectures: TBD

The short term

Our immediate goal is to scale the model size limit from 1 MB to 100 MB. Achieving this milestone will significantly enhance the ability of AI developers to deploy more sophisticated models on-chain.

At the deep learning library level, we are expanding the Eternal AI Toolkit by implementing additional smart contracts to support a wider variety of neural network layers and architectures.

At the computation level, we are enhancing the VM by integrating additional CUDA instructions, thereby increasing the capacity for more complex on-chain computations.

When	On-chain layers to support	Model examples	# Parameters
Jun 2024	Activation	Lenet5	0.06M
Jul 2024	Concatenate AveragePooling2D BatchNormalization Activation ZeroPadding2D GlobalAveragePooling2D	DenseNet121	7.2M
Aug 2024	BatchNormalization DepthwiseConv2D ZeroPadding2D Reshape Activation GlobalAveragePooling2D Dropout	MobileNet	V1: 4.2M V2: 3.4M V3small: 2.9M V3large: 5.4M
Aug 2024	Normalization ZeroPadding2D Activation DepthwiseConv2D GlobalAveragePooling2D Reshape Multiply Dropout	EfficientNet	B0: 5.3M B1: 7.8M
Aug 2024	Activation SeparableConv2D Cropping2D	NASNetMobile	5.6M
TBD		TinyStories	5M
TBD		Tokenizer
TBD		Squeezenet
TBD		GPT-2 (S)	200M
TBD		GPT-2 (XL)	1500M
TBD		Stable Diffusion