![Search icon](https://hackernoon.imgix.net/search-new.png?w=19&h=19)
The intersection of Web3 and artificial intelligence, especially in the form of generative neural networks, has become one of the hottest topics of discussion in the crypto community. AI is a serious game changer in many areas, and here is no exception. However, given that decentralization is one of the main goals of the new generation of the Internet, many crypto users have a legitimate question: "How can artificial intelligence be brought under the Web3 standards of security and transparency?"
Web3 users are used to the pros of decentralized systems. Still, the reality is that not all processes can benefit from decentralization, and there is a different blockchain adoption scenario for each area. Looking at artificial intelligence from this perspective, the questions "Why is AI still not decentralized?" and "What can we do about it?" arise. I will try to answer them in this article.
Making the case for decentralizing AI is pretty simple. Artificial intelligence is digital knowledge, and knowledge is the foundation of the world. Accumulating control over AI in the hands of a single corporation inevitably leads to information filtering and fact rewriting. There's no denying that neural networks are infiltrating crucial areas of our lives at an alarming rate. Without the proper level of transparency, the situation can easily spiral out of control.
Introduced in March 2023, the GPT-4 LLM is several times superior to GPT-3.5 in many ways. According to the developers, the upcoming neural network update will continue in this direction, becoming even more powerful and efficient. Of course, there are also decentralized analogs of ChatGPT, but the capabilities of their developers are limited by the budget, which is not comparable to large corporations. If nothing is changed, centralized AI solution providers like OpenAI will eventually win this race, and competing with them will become useless.
Process transparency is the second factor to consider when discussing the future of artificial intelligence. Modern large language models include millions of lines of code that encrypt huge amounts of data collected from all over the Internet. As a result, no one knows what's going on inside GPT-4 or what data OpenAI has trained their tool on. Decentralized AI could provide transparency about what a particular model was trained on and how.
If the case for decentralizing artificial intelligence is so obvious, then why has there not been a single successful attempt in this field so far? After all, decentralized AI is not a new idea, and many of AI's principles were laid out in the early 1990s. Without going into technical details, the main reason here lies in the relative unpopularity of artificial intelligence until recently.
Before large companies like OpenAI emerged, large language models were implemented mostly at the corporate level and did not have access to large amounts of data. Simply put, the scale of neural networks in the past did not raise the concerns discussed today - it seemed that this tool would not go beyond pre-defined limits. Implementing limits or moving AI to a decentralized architecture was seen at the time as a premature panic.
It's possible. According to my feelings, the odds of this happening are from 1:5 to 1:10, and 2024 will determine a lot of things here - there will be an unusually large number of elections in the world this year, plus protracted armed conflicts and war in Ukraine must somehow come to an end. The more victories there will be for radical and conservative forces, the more centrist the world will be, and the more likely scary scenarios will come true. Conditionally, I mean that a strong AI will fall into the hands of terrorists or dictators.
The most unrealistic option is an uprising of intelligent machines. The most possible is the total monopolization of AI by the rich, czars, and crooks. Because there are many bad people with power, but there are no robots with consciousness yet.
The AI corpocracy is already forming in the US and China. This market is not developed enough in other countries to speak of a monopoly.
When it comes to generative AI, there is no single approach to decentralizing it. There are several directions in which this issue is evolving. Let's consider the main ones:
Decentralization of computation is important at the stage of training and tuning a neural network. As it is known, modern large language models are very power-consuming and require powerful GPUs, so data processing for them is usually performed in large centralized centers. Implementing a decentralized computing network, where different parties can provide their own devices to train and fine-tune neural networks, can help eliminate the control that large cloud providers have over the creation of such models.
Currently, the data used to pre-train language models and other neural networks is highly classified. Because of this, users are left wondering where services get the information from which they generate content. The situation can be remedied by implementing blockchain to create a transparent system that would allow users to trace the origin of the data.
Neural networks require human intervention at some stages of development. Techniques such as reinforcement learning from human feedback (RLHF) allow the GPT-4 model to act as a user-friendly ChatGPT service. Such validation is particularly important during the fine-tuning phase, and these details are currently closed to users. A decentralized network of validators, consisting of humans and chatbots performing specific tasks, the result of which is open to everyone, could be a significant improvement in this area.
Which neural network is the best on the market? According to the benchmark, the one that boasts the most impressive results is the... Stop.
Most of the rankings are either compiled by or sponsored by the companies developing these LLMs. Evaluating AI performance and capabilities is an important task, and there is currently no auditor in the world who can publish a truly unbiased and accurate review. Oddly enough, the solution could be a decentralized organization of evaluators that would conduct independent, anonymous tests and measurements.
Finally, the most obvious area of decentralization is infrastructure. The use of LLMs today requires trust in infrastructure that is controlled by centralized providers. Creating a network in which the computation involved in generating answers can be shared between different parties is a challenging but interesting task that can be of great benefit to humanity.
Effective accelerationism is not about ignoring threats or the lack of regulation—it is about not panicking and realizing that regulation always lags behind technology and that strict prohibitions usually do more harm.
In other words, regulation is good if it's efficient and transparent. And the state is the most inefficient and opaque regulator, with high levels of corruption, bias and a penchant for hyper-control. Companies are little better, and only because they are faster and smarter.
What kind of regulators will there be, then? Network standards and protocols. I mean open source, with access to all sorts of databases.
Not only. Automated decentralization is the best option, but it's not a panacea. Something can go wrong, and we get a mess of bots instead of a civil online community.
Conversely, capitalism and even authoritarianism can be positive in theory - provided that the power is in good hands. Admittedly, history shows that this is only a theory. If power is concentrated for years in an unchangeable group of people - even with the highest ideals - it usually comes down to wild inequality and unfreedom.
And who is responsible for anonymous calls for genocide? Decentralization does not equal the absence of rules. It is just that these rules are created and controlled by users. Moreover, if a service is more crap than useful, it will eventually be abandoned, no matter how trendy and decentralized it is.
There are a few axioms that are unlikely to change even under ideal circumstances:
These axioms give rise to the third, saddest axiom—there will be no full equality. In other words, the hospital temperature will definitely rise, and the gaps will be smoothed out, but everyone will definitely not live equally well. At least until the technological singularity, we can't make any predictions at all.
Artificial intelligence deserves decentralization in all its aspects: data collection, computation, development, and optimization. The arguments in favor of this are obvious, but humanity still needs to overcome the technical challenges behind them. AI may have to make a technological breakthrough to decentralize, but this goal is certainly achievable. In the ideal world of the future, no centralized structure should have complete power over artificial intelligence.