News
The next lock-in will be unreadable
July 5, 2026 · PenDraco
Source code is readable because humans had to maintain it. As frontier models write and operate more software, vendors gain every incentive to ship artifacts only their model can modify. The old lock-in held your data. The next one holds your logic.
Sometime in the next few years, a vendor is going to hand a business like yours a piece of software that works beautifully and that no one can read. Not you, and not any engineer you could hire. Changes will be easy to request and impossible to make anywhere else. That is not a bug in the product. It is the product.
Source code is a human courtesy
Computers have never needed readable code. Source code exists because humans maintain software: every convention we treat as professional hygiene, version control, code review, security audits, documentation, presumes a human reader coming back to the text later. Compilation was always a one-way street, and we kept the readable original because people had to walk back down it.
That norm is not a law of nature. It is an economic arrangement, and the economics are changing.
What changes when the model is the compiler
More than a quarter of all new code at Google is now generated by AI, "then reviewed and accepted by engineers," as Sundar Pichai put it in late 2024. Read that sentence carefully: the second clause is the expensive one. Human review is the cost that keeps generated software readable, because a reviewer needs something to read.
As frontier models improve, the pressure lands exactly there. If a vendor's model writes the software, patches the software, and operates the software, readable source becomes overhead with no buyer. The natural end state is a vendor that ships you an artifact, compiled, minified, or machine-shaped beyond human comprehension, and sells changes by request. Describe what you want; their model modifies what only their model can read.
Andrej Karpathy described the deeper version of this in 2017: in what he called Software 2.0, the network's weights are the compiled artifact, the dataset and architecture are the source, and the trade on offer is "a 90% accurate model we understand, or a 99% accurate model we don't." When business logic migrates into weights, or into code no human will ever review, the binary is no longer a build product. It is the moat.
The old lock-in held your data hostage. The next one holds your logic. And it arrives with a meter attached: when only the vendor's model can modify what your business runs on, every change request is a toll, priced per token, forever.
We have seen this movie, in a slower format
Proprietary document formats took a decade of standards wars to pry open. ERP customizations that only execute inside one vendor's runtime still hold companies captive today. Minified JavaScript already ships unreadable, with the readable original an optional courtesy. And COBOL estates prove that readability can die even when the source survives, because the readers do not. Every earlier lock-in needed years and a format war to close its grip. A model-emitted artifact gets there in a single contract cycle.
The counterweight is already forming
Buyers and regulators are moving in the opposite direction, which tells you how the fight will be scored. The US government's 2021 cybersecurity executive order pushed software bills of materials into federal procurement: you must be able to say what is inside the software you run. The EU Data Act, applicable since September 2025, builds a legal framework for switching providers precisely because lock-in became a market failure. And anyone who has sat through diligence knows where this lands commercially: an asset nobody can audit is a liability with a user interface. Acquirers, auditors, and insurers will price unreadable stacks accordingly.
Readable is a choice you make now
We build with AI every day. A senior engineer wielding agentic tooling is the reason enterprise-grade delivery reaches a mid-market budget at all. But there is a line, and we hold it: the artifact stays human property. Readable code, in your repository, under your account, on an open-source platform, documented well enough to run without us. Our Sprint promise that you keep everything we build only means something if you can read what you keep.
The same line runs through Jakada.ai. Where a model carries part of your operations, the question that matters is who holds the model. Jakada's is dedicated to your tenant and fine-tuned for your business, never pooled and never training anyone else's, and at the sovereign tier it deploys entirely inside your environment and never leaves. The lock-in fails when the binary can stay home. And because we run owned inference ourselves, even our delivery layer sits where you can see it.
Four questions to ask any vendor building software for you with AI:
- Do I get the source, in my repository, under my account?
- Could a competent engineer who has never met your model maintain it?
- Does my business logic live in code I can audit, or in weights I cannot?
- What does a change cost me in year three, and who sets that price?
Frontier models are about to make software astonishingly cheap to produce. Whether it stays yours is a separate question, and it is settled in the engagement terms, not the codebase. We build with AI and hand you something you can read. That is not nostalgia. It is the difference between owning your operations and renting them.