Self-written sovereign network applications and internet services that can be updated in real time simply by speaking.
The Internet Computer is designed for this, and this work has been ongoing for many years. Today, I want to delve into this topic for the first time. Please note — do not confuse this area of work with the AI capabilities of the ICP network (for newcomers: the Internet Computer is undoubtedly the only public network in the world that can host and run AI in a smart contract mode, which allows it to reside on the network and be decentralized, secure, and unstoppable. For example, I recently demonstrated an on-chain neural network performing facial recognition, and upcoming enhancements to the ICP protocol will also allow LLMs to run as smart contracts). Today, I want to discuss a very different challenge that ICP will help the world address, which is creating and updating *running* web applications and internet services simply by speaking. Users will create these models for various purposes, such as creating a secure personal note manager or personal website, or creating social media, games, web3, or enterprise infrastructure models for others that involve online communities of all sizes, simply by issuing commands in natural language. Also, note that this is a much larger and different challenge compared to using AI to write and review software, which is already happening on a large scale. This goes far beyond the scope, and I will explain the reasons below…
First, let’s review some overall trends to understand the direction of the internet and AI:
Many people reading this have already been using ChatGPT to explore ideas, gather information, and analyze, improve, and create content, and soon start searching the internet. ChatGPT is an example of a large LLM (i.e., chatbot) that has a vast number of parameters and has been trained on massive amounts of data. If you are a software engineer, you might already be using LLMs to help write and review code, although you might prefer Claude. These models have sparked a recent wave of hype around AI. But in fact, they are manifestations of deeper trends driven by artificial intelligence.
A deeper trend is: we interact with AI, whether intentionally or not, and it gives us what we want. Services like TikTok and Instagram Reels are among the first mass-market manifestations of this.
These services are not traditional social media services at all. Essentially, they are powered by powerful AI engines, and they cannot work without these engines. When you use them, you are actually interacting with AI. The AI in these services categorizes the social media content they can provide, so it knows what is inside videos and other types of posts. Then, when it provides you with content for you to consume, telemetry data is collected so the AI can track how you interact with the content — for example, the simplest being it can track how long you watch a video to determine what type of content you like. As the AI becomes better at understanding what you find attractive (this can extend to the sequences of videos it shows you, not just individual videos), you will get a better experience, which is why these services are so addictive. This new media model is just part of the unstoppable journey of AI technology, which gives us what we want — this will have far-reaching implications. What once seemed impossible is now happening. For example, researchers recently trained AI with millions of hours of footage of people playing Doom and Minecraft. This allows AI to simulate these games for users in real-time. You can play a game, but without a game server or game client, just AI transmitting video to you based on traditional inputs (like left, right, run, jump, and shoot). This hints at the future: in the future, AI will create virtual reality experiences for us, observing how we interact with them and customizing them in real-time to make them more fun and engaging.
This may sound like science fiction, but it is actually just a continuation of the trends represented by TikTok and Instagram.
But that is enough. What else can AI specifically provide us with in ICP? Our ICP community is highly focused on reshaping the platform we build, especially offering a better way to create web applications and internet services (including those that internally have web3 and AI capabilities) that are also sovereign.
In the future, we will be able to create and update the *running* web applications and internet services we want just by talking to AI.
The reasons are obvious. For example, if a businessperson wants to build a custom personal website to promote their brand, featuring a blog, sections that can embed media pulled from places like YouTube, a library page hosting the documents they create, and a page linking to their social media profiles — all presented in a specific way. In the future, will they still need to hire designers and developers to build something like that, or will they just have to fiddle with Wix, or stick with their original LinkedIn profile page? Certainly not, they will just talk to AI. So what if an avid gamer wants to create their custom online game to share with friends? Is it still difficult to express creativity without special technical skills and a lot of time? No. And what if a business, NGO, or government only needs some custom CRM functionality? Do they still need to sign up for expensive SaaS services like Salesforce and hire consultants to customize it? Nowadays, creating things on the internet is both complex and time-consuming, and costly — this prevents us from building what we want.
The whole world is waiting to break free from this situation.
In the future, we will be able to create and update by just conversing with AI, leading to an almost infinite number of new customized web applications and services being created for all imaginable purposes. Here’s how it works:
You describe the custom application or service you want to the AI.
It will return a URL to your web browser, and then it can be used!
You and others will use the application, allowing it to accumulate content and data.
You will describe improvements, expansions, and fixes.
Then, you just refresh the page to see them.
Loop back to 3, continue iterating to create value.
Over time, this new model will greatly change the way technology works.
Imagine what this means for emerging entrepreneurs around the world who lack the technical skills or funds to hire them, but have ideas for businesses in social media, gaming, sharing economy, AI-based services, Web3 services, communication suites, and _ (fill in what you want). This will democratize access to the tech economy and help a vast amount of global talent get involved, become industrious, and succeed.
This has been one of the goals of the ICP project since its inception.
Moreover, this functionality will be available to all of humanity, not just entrepreneurs. Imagine a group of high school students wanting to organize the information they collected from a biology field trip. This applies to them too! Imagine a business department needing customized online functionality but unable to convince the CIO and CFO to allow them to sign up for something like Salesforce and hire some consultants to customize it. (Even if they agree, it will take a long time and be costly.) Problem solved. Now think about the situation in developing countries. They have an increasing demand for customized online functionalities but cannot afford large tech SaaS services, and at the same time lack the skills to build it themselves — if they had those, they would still need cybersecurity expertise to ensure what they build is secure, which is exactly what they lack. This will be transformative for those economies.
The new paradigm will address all these needs and, in fact, go further.
Those who create custom applications and internet services will own the software that creates them, and even if they did not write it themselves, they own the data within it — this contrasts sharply with popular SaaS services used by enterprises, which hijack customer data, and even consumer services like Google Photos that prevent media from being disseminated.
These custom applications and services will truly have sovereignty, and the owners will not be passive customers — this has always been the goal of the Internet Computer network.
Contrary to intuition, this new model will also be a good thing for software engineers — it will lead to millions of new custom applications and services being created, and inevitably, in some places, human assistance will be useful for solving specific problems and help refine the prompts. The magnitude of online infrastructure will create a massive amount of software engineering jobs worldwide. If you’ve been following me this far, I hope you now understand that this new model is both inevitable and represents one of the greatest revolutions in the history of technology.
So... the next question is, how will ICP ultimately unlock this paradigm for the world?
To understand the next part, we first need to understand the limitations of traditional IT when applied to this paradigm. Creating and updating running online applications is much more complex than getting an LLM to write some software code. For example, when building using a typical traditional IT framework, you might need to perform the following actions:
Get an AWS account and add a credit card
Get some 'compute instances' (which are essentially servers)
Install some network security measures to ensure security
Install a database server, web server, …
Orchestrate using Kubernetes and others
Patch all software to ensure security
Design failover, backup, and recovery
Create tables in the database
…
Install relevant code
This is quite a long task list, and some of the tasks are very complex, so what AI has to do is more than just write code. Let's assume the AI has hooks and the knowledge to perform all these steps by itself. Would this solve the paradigm? Unfortunately, the issues remain... Fundamentally, the paradigm should provide users with a real-time creative experience, and even installing a database server or patches takes some time. Of course, this can be alleviated by using pre-installed images, but the problem runs deeper... The various steps and requirements involved in traditional IT can go wrong in multiple ways. The AI's construction order can be interrupted, just like humans, and it must determine how to solve these problems within an *unconstrained problem space*, which will have subtle implications for safety, etc. — this is a serious issue, as traditional IT is insecure by default, and even small errors can lead to disaster. Traditional IT is a ridiculously complex Rube Goldberg machine, and leaving AI alone in this unconstrained problem space can be very dangerous, as it can hallucinate and potentially pick up bad memes from training data. Everything the AI does must be manually reviewed by technically skilled personnel, and if the application or service is important, it must also be audited — of course — the focus of this model is that it does not require creators to have technical skills, but rather enables creation to be a real-time iterative act. There are other stunning works that use traditional IT. In the new model, users/creators will want to update their running web applications and internet services in real time simply by telling the AI what improvements, expansions, and fixes they want.
The created system needs to undergo large-scale upgrades every few minutes!!
The design of traditional IT did not account for this. Anyone involved knows that upgrades are a big deal, and for production systems, upgrades often require long intervals. This is because changes often must be made synchronously across multiple components (e.g., updating database tables, changing web server configurations...) and this is cumbersome. Additionally, when you change the design of web applications and services running on traditional IT, the upgrade process often involves refactoring/migrating data, which is slow, computationally expensive, and prone to errors — this again hinders the critical real-time nature required for traditional IT to achieve this paradigm, which would involve running web applications and internet services that are updated almost at the speed of conversation.
I could go on, but it should be clear that traditional IT is not really suited for this paradigm.
Due to difficulties, we will see services like Vercel, and possibly one day like Google, offering AI that can create applications within their designed custom infrastructure platforms, thereby improving some of these issues. But their platforms will still be less than ideal, and the software created by AI will also be locked into their special platforms, potentially hijacking relevant data in some way, resulting in customer lock-in, and the applications and services involved will not have sovereignty. (Nevertheless, we predict that some web3 projects that tend to mimic ICP will, out of desperation, use something like Vercel to create a simplified version of this paradigm and then sell their services to the public under the guise of 'on-chain', but apart from successfully selling more tokens, their plans will ultimately fail to compete with mainstream global users/creators.)
So what is really needed?
The good news is that DFINITY has been working on this paradigm for years. Let me explain... From the beginning, going back several years, there have been over 1000 person-years of R&D efforts, and our work has been focused on widely reinventing computing using decentralized networks that leverage advanced cryptography, protocol mathematics, and computer science. Our work in the web3 space is completely unique.
A key feature of ICP is that you can build web applications completely using secure and unstoppable software residing on the network, which is a more powerful evolution of smart contracts. When building on the network, you do not need large tech companies or traditional IT. AI can write this code and upload it to the ICP network (like the Internet Computer) to create web applications or other internet services. It is important to clarify that in the radical computing environment created in ICP, AI only needs to upload code to create and update... No need to configure cloud accounts, databases, web servers, and network security.
Additionally, the code is automatically secure and not subject to network attacks. Therefore, there’s no need to worry that hallucinations will leave a big door open for hackers.
Additionally, AI does not need to design and configure complex failover systems, as the code is unstoppable and always running.
So these barriers have been removed. But the real power comes from the groundbreaking advances in computer science provided by ICP, called 'Orthogonal Persistence' (keep reading, I won't get too technical on you!). On ICP, software units run in persistent memory pages, which basically means that data 'sticks' to the software logic created when programmers write in a software language, freeing them from the complexity and inefficiency of copying data to databases and files, and eliminating the need for these things. Everything is just highly abstract software residing on the network. This enables engineers (and soon AI) to describe functionality in simpler forms without dependencies, which is ideal for this paradigm.
Back in 2018, I described the vision of 'Orthogonal Persistence', but it wasn't until now, 6 years later, that it has been fully realized through Motoko, a domain-specific language directly tied to the work with the ICP platform.
Prepare for what we call EOP or 'Enhanced Orthogonal Persistence', which will ultimately achieve the goal we have been striving for. (This is limited by the upgrades that are about to happen, such as the 64-bit change that also requires LLMs to run on the network.) Above, I talked about the importance of instant and secure upgrades for the new AI paradigm. Well... EOP makes it possible to 'morph' software between upgrades. Developers (both human and AI) will write new versions of the software to achieve the desired changes. They will then describe the code that converts data from the old version. (For example, if a Google Photos-style application has already been created, the upgrade might add location data and comments to photos, resulting in a structural change to the 'photo' data type.) During the upgrade process, in the new paradigm, EOP will do the following: 1) It adds type safety to upgrades, ensuring that if AI makes a mistake that could lead to data loss (whether due to hallucination or other reasons), the upgrade will fail, significantly reducing the risks that always exist in traditional IT architectures. 2) As the software morphs through upgrades, it allows data transformations to be done efficiently, so the paradigm can provide real-time upgrades at the speed of conversation. This is precisely what we need.
The other benefits of focusing on reshaping computing over the years are also significant. For example, because code and state merge in this environment, ICP can easily snapshot applications and services almost instantly, and if users don't like the way their data transformations are handled, they can roll back to where they were before (using EOP, this can often be achieved by 'upgrading' to an earlier version of the software). I could go on, but I will summarize.
This new paradigm, which will profoundly change technology, will unlock by combining continuously improving AI with ICP technology.
For the ICP community, it’s incredible that this paradigm offers practicality to a massive global market, and we won’t be constrained by the noise of web3. People will use it because it meets their needs. I can tell you that behind the scenes, we are making tremendous efforts for AI itself and also for getting it to build on ICP, ensuring that the Internet Computer can scale to handle this issue — you may already know of this work. If you think the recent surge in computational power of the Internet Computer has been impressive, then buckle up, because this model might mean we haven’t seen anything yet. As always, we choose to believe this rather than the narrative:
Pure. Utility. From. Advanced. Alien. Tech. Will. Win. In. The. End.
Today, we are more closed off than ever. Oh, by the way, did I mention that the next generation of web applications and internet services created using artificial intelligence will be internet-native and sovereign? They will run on networks hosted by truly decentralized hardware (the Internet Computer is one of the few web3 networks that actually does not run on large tech), leveraging trustless multi-chain functionality and embedding true on-chain AI.
This is going to really, really cool...