Vision for Decentralized AI

Heurist is a decentralized protocol for AI model hosting and inference. We provide easy access to AI models at minimal costs. We emphasize transparency, reduction of AI bias, and the democratization of AI technologies. Imagine a world dominated by open-source AI models, which are free from censorship and are accessible to all. They thrive in a decentralized network, governed not by corporations, but by the community itself.

The problems with close-source AI

Closed-source AI models do not provide public access to their underlying algorithms, model weights and training data. This opacity raises significant concerns around trust, fairness, and accountability. Dominated by tech giants like Microsoft and Google, these models centralize power in the AI sector. Such concentration can dampen innovation, curb competition, and potentially breed monopolistic tendencies.

These models lack the flexibility for customization, limiting their utility across diverse scenarios. Dependence on big tech companies leads to vendor lock-in, stifling the adoption of emerging AI technologies. Their monopolistic nature often translates into higher usage and maintenance costs, in stark contrast to more affordable open-source options.

The inherent risks of a monopolized AI industry: Reliance on a single provider can lead to widespread disruptions in AI services.

Yearly Costs Based On Usage And OpenAI/Cloud Hosted Open-source Models (source)

Censorship Concerns in Closed-Source AI

Closed-source AI models often adhere to American laws for censorship, which may not align with the norms and needs of other countries and cultures. For instance, what is considered sensitive or offensive content can vary greatly across cultures. A topic that might be taboo in one country could be a subject of open discussion in another. This discrepancy can lead to situations where AI models either over-censor content, stifling free expression in some regions, or under-censor, potentially causing offense or violating local laws in others.

A study by David Glukhov et al. (arXiv:2307.10719) critically analyzes existing censorship methods. The study reveals that these methods can inadvertently impair model performance. Current censorship strategies in AI, primarily based on semantic filtering, face theoretical and practical limitations. These approaches often fail to address the complex nature of language and intent.

The AI industry needs far more scruntiny in AI model censorship to ensure that it is both effective and culturally sensitive. There is a pressing need for a wider variety of AI models, ones that are not solely under the control of a few people. By embracing a broader range of AI models, we can better ensure that AI serves as a tool for empowerment, rather than a means of cultural homogenization and control.

Ineffective censorship can hinder normal usage of AI

The rise of open-source AI

A common misunderstanding is that open-source AI models are inferior to closed-source ones, such as ChatGPT. Research has shown that many open-source language models show equivalent or even superior performance to GPT3.5 and in certain domains, reaches at par with GPT-4.

Benchmarks of OpenChat 3.5 is comparable to ChatGPT (March 2023) (source)

Llama 2 70B model is close to GPT-4 and Human in terms of factuality (source)

In the domain of image generation, Stability AI's Stable Diffusion models have been the most popular open-source text-to-image model family, and proven to be as powerful as OpenAI's closed-source competitor DALL-E 2 and more cost-efficient. The weights of Stable Diffusion models are public, and that enables artists and developers to fine-tune the model to adapt to specific visual styles, a feature that is missing in OpenAI DALL-E models.

There exists many such AI platforms in Web2, such as HuggingFace and Replicate where anyone can host open-source models and use other's models for inference tasks. Such platforms allow users to access a variety of models without paying the high price of closed-source ones.

Open-source projects benefit from the collective input and scrutiny of a diverse community. This collaborative environment fosters rapid innovation and improvement, as seen in many variants of LLaMA models and Stable Diffusion models. The community's contributions in terms of fine-tuning, performance benchmarking, and diverse perspectives accelerate the evolution of these AI model families.

Fine-tuned models on Civitai, the largest AI art platform

Open-source AI models offer an unprecedented level of transparency. Users can examine the training data and model weights, which is not possible with closed-source models. This transparency builds trust among users, as they can verify the model's workings, data handling practices, and ethical considerations. It also facilitates more robust security, as vulnerabilities can be quickly identified and addressed by the community.

The presence of strong open-source alternatives challenges the monopoly of big tech companies in the AI space. It prevents the concentration of power and control over AI technologies in the hands of a few corporations, promoting a more balanced and competitive landscape.

The Imperative of a Decentralized Web3 AI Protocol

The advancements in open-source AI have paved the way for a new paradigm: what if we further empower AI with the incentivization mechanisms of cryptocurrency and the ownership economy model of Web3?

If any component of the tech stacks of AI is dominated by a centralized party, be it the data, the computing backbones, or the application layers, such monopolization often leads to gatekeeping and unfair rent-seeking. Web3 solves this by democratizing access to AI and ensuring transparent ownership of AI.

Heurist, operating as a DePIN network at its core while supporting a diverse AI application ecosystem, is designed to address the challenges of centralized AI systems. We start off by decentralizing the GPU compute infrastructure, and our end goal is to fully transition into a DAO that governs every aspect of Heurist Protocol and its ecosystem.