What if anybody could imagine a website, online app or service, and create it by having a natural language conversation with AI? What if this online functionality simply appears on URLs, available for immediate use, and can be improved through continued conversation? What if talking can create and update serious production services—which might support large numbers of users or customers—under the auspices of a guarantee that changes and updates made by AI at the speed of chat will never result in an accidental loss of data?
This is the brave new world of "self-writing apps," an extension of the popular "vibe coding" paradigm, where human developers work in partnership with AI that has been integrated into their tools. Self-writing is different because apps and services are created, end-to-end, solely through natural language conversations with AI, enabling non-technical people with internet access to take on the role of entire technical teams.
An important variant of the "self-writing apps" paradigm is the "self-writing internet." This envisions a world where the online functionality we use everyday is increasingly created through these conversations. Part of the vision is that websites, apps and services, often run on secure decentralized networks, as a part of the public internet, rather than on closed proprietary platforms such as cloud services.
In this vision, online functionality is sovereign because app creators own the underlying software and data, and further, their apps cannot be switched off by proprietary platform operators that might withhold service or change their terms.
The self-writing internet has the potential to unlock tremendous opportunity and change.
Individuals will find they can create exciting new forms of online functionality, such as hyperlocal social networks for exclusive use by their extended family, where they can add large and goofy emoji reactions to photos in a shared gallery, or create hyperlocal esports, where they might compete with friends for a prize of digital assets. Such new forms of online functionality will quickly follow the mass production of baseline functionality such as personal branding websites, and party-planning portals, as non-technical people get to grips with their newfound engineering superpowers.
An entrepreneur who lacks capital will finally be able to launch the Web3 sharing economy app they want to bring to life. Small and medium-sized businesses will create and update superior e-commerce experiences while incurring a tiny fraction of the cost and time involved before. Enterprise will apply the paradigm to create things like CRM and ERP. School districts will find they can quickly create the fundraising platform they couldn't afford, and an NGO will quickly create a system to assist refugees as part of a rapid response to a natural disaster. There are infinite possibilities.
Meanwhile, the paradigm will bring unparalleled benefits to the developing world. There, governments and enterprises often cannot afford to get locked into foreign SaaS services, which also present them with sovereignty challenges, while at the same time lacking sufficient home-grown technology expertise to build what they need for themselves. Even when they can build what they need, a key additional challenge is that they lack the cybersecurity expertise to keep their functionality and data safe from threats, with the result that data exfiltration and ransomware attacks succeed.
The self-writing internet will allow anyone with an internet-connected smartphone to create online functionality. While creating complicated apps and services demands clear and accurate prompting in conversation with AI, nonetheless, a significant portion of the 5 billion people with connected devices will be able to create simple apps for personal use, reflecting the scale of the empowerment that will occur.
The self-writing internet will increase the decentralization of tech and democratize the act of creation, but an important consideration is what self-writing might do to software engineering, UX, and design jobs. Our view is that, just like the internet, self-writing will create many new, high-value jobs.
Self-writing will lead to an enormous proliferation of apps and services, increasing their number by orders of magnitude. While most will be created without the involvement of traditional "craft" skills, a significant subset will benefit from human help, and eventually the sheer volume of new apps and services being created will drive an overall increase in demand for skilled developers and designers.
Outside the technical realm, the freedoms and economic gains delivered by self-writing will also create non-technical jobs. People will be empowered to become entrepreneurs, and their ventures will employ people in various roles. Businesses will become more efficient, driving economies, and governments will be empowered to use technology to provide better operating environments for their citizens, reducing friction and accelerating progress.
Realizing the self-writing internet vision involves overcoming unique technological challenges. However, the solutions are already here, and the self-writing internet is now working in compelling initial forms, and in use – with broad public adoption imminent. Before proceeding to highlight what is already working, we must consider the key challenges.
Firstly, the self-writing internet requires powerful foundational AI models that can write code. Stunning advances have already been made by the AI industry in this regard, and progress continues, with major hosted solutions including ChatGPT, Claude and Gemini, and major open source solutions including models like DeepSeek and Llama, which can be self-hosted, and arbitrarily tweaked and fine tuned for purpose. But having models write code effectively is only one challenge.
Broadly speaking, there are two major additional challenges. The first involves providing AI with its own programming environment, which is specialized to support the efficient and error-free creation of self-writing app functionality, which are additionally able to ensure that mistakes by AI do not result in catastrophes – such as data loss during app upgrades. The second involves providing AI with platforms where user apps can be deployed that remove the need for human support in critical areas such as cybersecurity and resilience.
Overall, reconnoitering the key challenges involved, it's clear that effective self-writing will likely always involve the use of completely new technology stacks that are specialized for purpose.
To facilitate the full self-writing vision, these new specialized stacks should also make the creation of sovereign apps possible.
Within the context of self-writing, a key challenge with traditional technology stacks is that they are insecure by default. When we build app functionality on these stacks, we must also make them secure ourselves. To achieve this, we create layer upon layer of defenses, by carefully configuring platform components such as databases and web servers and cloud frameworks such as Kubernetes to harden them for security purposes.
We then augment our hardened traditional stacks with cybersecurity defenses such as firewalls and intrusion monitoring, check third-party software doesn't have malicious logic hiding inside using anti-malware, and try to keep software packages up-to-date by installing security patches.
One simple mistake, or the minor failure of a cybersecurity system, can result in penetration by hackers and the exfiltration of sensitive data and systems being encrypted by ransomware. Furthermore, every day the cybersecurity challenge grows as hackers ironically learn how to employ AI for their own purposes.
The process of securing traditional technology stacks is so complicated that millions of pages of technical literature have been written to assist practitioners. This is a lot of responsibility to delegate to AI models that make regular mistakes.
However, even if AI could be trusted with such responsibilities, the task of managing security would slow down AI, perhaps preventing it working at chat speed, degrading the self-writing experience.
Resilience is another problem when building on traditional stacks. Craft engineers must invest time designing failover systems and other mechanisms to ensure apps stay online, and upgrades are often an all-hands-on-deck process, involving tricky synchronization as databases are reconfigured, and custom code is swapped in, and personnel must stand ready to rollback changes if something goes wrong.
On traditional stacks, implementing resilient apps, and performing app upgrades are major undertakings.
Today, more than 5 billion people have internet-connected smartphones. The huge volume of new custom apps that self-writing may create means traditional stacks cannot be used since security and resilience cannot depend on human assistance.
"Orthogonal persistence" is a technical-sounding term for a bleeding-edge software technology that greatly simplifies the production of apps and software. Among its aims and purposes, it focuses development on the essence of "what" must be created, while greatly reducing work that relates to "how" systems are put together.
Generally speaking, creating apps involves work that can be divided into "what" and "how" buckets. The "what" involves the user experience people interact with, and the essential processing of data, which define the essence of the app. "How" applies to the actual process of combining custom code with platform building blocks, like databases, web servers, and cloud orchestration frameworks. "How" is concerned with the way technology stacks produce the overall systems that support apps, rather than the implementation of the functionality that defines them per se.
A significant problem with traditional tech stacks is that building often involves far more investment into "how" than "what." In fact, as many developers will attest, project work often directs more than 90% of effort towards "how," with only 10% going towards "what."
Oftentimes, this is caused by complications in the app logic resulting from the need to marshal data in and out of stores like databases, and other times from difficulties making different platform components interoperate, which lead to age-old problems such as "dependency hell," and other forms of problem ladder that require human intervention to solve.
Zooming out to consider the needs of AI working solo in the self-writing setting, in which AI must create sophisticated functionality at the speed of chat, we can see that enabling AI to focus mainly on "what" will enable it to work far faster. In addition, removing the complications of "how" will allow functionality to be produced without errors on first pass more often, and free precious space from the inference context window, further increasing AI's effectiveness.
Of course, the desire to reduce the cost of "how" is not new. Within the realm of traditional tech stacks, this is reflected in the increasing popularity of serverless cloud platforms, such as AWS Lambda, where developers create code that runs within a seamless universe, in which the underlying server instances are abstracted away, and elements of the scaling, security and resilience challenges are handled for them.
However, these environments do not address the challenge of accessing and persisting (i.e. storing) data, and the serverless code has to call out into things like database servers.
Orthogonal persistence is a new technical approach that involves an important inventive step that collapses the conceptual distance between serverless logic and data, almost making them one and the same. This dramatically reduces the "how" involved in software development.
When serverless software is written for an orthogonal persistence environment, its data automatically persists within its own logical abstractions. More concretely, data lives within its own variables, data types, and collections. Creating platforms supporting orthogonal persistence involves advanced computer science, but using it greatly simplifies software development.
Simplification enables AI to work faster, and go further without hitting limits. Furthermore, it also makes verification of its work – by human or machine – easier. Therefore, orthogonal persistence can play an extremely important role within self-writing environments.
Developers interested in seeing what orthogonal persistence looks like in code can take a look at this network app that stores files for users (the link opens an online-IDE):
https://icp.ninja/editor?t=NL8cThe good news for those interested in the self-writing vision, is that the first real solution has already emerged as the result of a major ongoing R&D effort.
The DFINITY Foundation, a Swiss not-for-profit founded in 2016 (after the DFINITY project was announced in the summer of 2015) has pursued a special form of the general "World Computer" vision.
The World Computer vision involves decentralized networks hosting code and data, and extending the public internet so that it can host apps and services natively, from simple functionality used for personal purposes, through social networks, enterprise systems, and financial ecosystems – removing the need to use servers and proprietary cloud services and avoiding dependencies and lock-in.
By 2025, DFINITY had spent approximately $500m, mostly on the complex computer science and engineering work involved in a public network known as the "Internet Computer."
The Internet Computer is a decentralized network created by a protocol called "Internet Computer Protocol" (or ICP). This combines thousands of specialized computers from around the world, which are dedicated to purpose, that are owned and operated at scale by independent "node providers" (which comprise companies, financial institutions and individuals) from traditional data centers.
ICP combines these machines to create a single seamless serverless cloud environment, by building on advanced mathematics, and the decentralization of the hardware. Because the environment is created by a mathematically secure network, it brings numerous special properties.
Thanks to the power of mathematics and advanced computer science, apps can run on the Internet Computer, end-to-end, without the need for cybersecurity protections because the network makes them "tamperproof." Even social media services that maintain pools of digital assets, so that users can transfer them using novel mechanisms like chat, have run successfully in production without the need for cybersecurity measures despite the highly hostile threat environment that the public internet presents.
The mathematical ICP protocol imbues the Internet Computer with a property called "Byzantine Fault Tolerance," which provides hard security guarantees, so long as a sufficient proportion of its underlying node hardware (the machines the protocol combines to create the network and serverless cloud environment) do not fall under the control of attackers. Sans mistakes in the implementation of the special "client software" that nodes run to perform their role in the ICP network, hackers cannot break the platform provided by the network, since the laws of mathematics are firm (e.g., 2+2=5 can never be true).
If you are interested in the computer science that makes this possible, you can find a collection of papers at: https://internetcomputer.org/library
Another powerful property of the environment relevant to solving resilience needs within the self-writing setting, is that the correct software logic of hosted apps is guaranteed to run, and their data is guaranteed to be available, so long as a sufficient proportion of the network's underlying nodes are operating correctly, for which reason some term them as being "unstoppable."
Within the context of self-writing, removing the need to consider security and resilience lifts an enormous burden from AI working solo without human assistance. But the Internet Computer goes further.
One of the key aims of the Internet Computer, has always been to reinvent the nature of software, solving for the complexity involved with developing apps on traditional software stacks, which are centralized in nature. By leveraging unique aspects of the way the cloud is created, the Internet Computer provides a serverless environment with "orthogonal persistence," which is exactly what AI needs.
The Internet Computer project saw the importance of self-writing, and that self-writing platforms would become the technology stacks of the future. Since self-writing speaks to key aims of the World Computer vision, the project focused efforts around realizing the self-writing internet. As work progressed, "Internet Computer 2.0" was born.
Since inception, developers have been able to write backend code that runs on the Internet Computer in popular programming languages that compile to WebAssembly (a public standard for virtual machines and low-level bytecode format). However, since 2018, DFINITY has also developed its own programming language and accompanying framework, which is specialized for the Internet Computer environment, called Motoko.
It was decided that the Motoko programming language and framework would be developed to meet the needs of AI writing code, as well as human developers. Motoko became the first programming language designed to satisfy the unique needs of AI.
Motoko now supports an advanced form of orthogonal persistence, which is known as "enhanced orthogonal persistence." The framework requires that AI must upgrade apps by supplying special "migration logic," to mutate existing logic and associated data, into the new form.
Thanks to advanced computer science, the Motoko framework provides self-writing platforms with hard guarantees that mistakes made by AI during upgrades will not result in accidental data loss. Now, instead of a data loss catastrophe, mistakes by AI simply result in a retry.
Motoko also enables AI to write and build the backend code for apps with unparalleled speed, and while new programming concepts are involved, AI has no problem ingesting training data during fine tuning to become an expert.
Caffeine is the first platform for creating, updating, and managing self-writing apps, which either deploy to the Internet Computer, or a private network variant known as a UTOPIA. A growing team now works to advance the technology.
Caffeine is accessed as a web app that has been optimized for both desktop and mobile. In general form, it resembles a chat app, since creating online functionality now involves a conversation with AI. Its mission is to enable anyone who can effectively communicate in natural language, to build and maintain online functionality. It is designed to address the full spectrum of use cases, from consumer to enterprise.
Apps can be created that integrate with outside services, such as payment platforms, enabling important use cases like e-commerce. The Internet Computer also enables apps to trustlessly interoperate with major public blockchains, supporting the creation of Web3 functionalities that involve token processing. As work continues, Caffeine will be able to create apps for an ever-widening array of purposes.
A key feature of Caffeine is that it has its own App Store, which allows anyone to publish their apps for others to clone, either for free, or for a price. Users can visit the App Store to clone an app either as the starting point for a new project, or for immediate use.
The App Store provides foundations for a self-writing economic ecosystem, enabling diverse talent around the world to become industrious in a new way, and unleashing creativity that drives the production of apps that can be cloned for use.
In an interesting novel development, users can subscribe to updates that app publishers make, while still applying customizations.
The world of technology now approaches an interesting inflection point.
If you are interested, sign up soon to get early access to Caffeine, or attend the "Hello, Self-Writing Internet" event in San Francisco on July 15th, 2025 (limited capacity).
The San Francisco event will demonstrate major new elements of Caffeine, which weren't revealed at the recent demo, June 3, at the WCS event in Zürich. There will be several hours of demonstrations, and Caffeine will be the tool used by self-writing hackathons running in parallel.