Learning to Code (Part 4)
Why software really is dead.
You can read Part 1 here, Part 2 here, and Part 3 here.
Learning to Code
Part 4: The End of Software
I started this journey annoyed about how one big tech company is shaping Washington policy with a well-financed influence machine to make tax and privacy laws more anti-consumer.
I discovered that from another corner of Silicon Valley, tech oligarchs might be offering tools that could level the playing field.
It was a good reminder that the effects of new technology are uneven.
AI’s radical democratization of intelligence has two sides. The question of whether it will replace workers dominates the discourse. The flip side gets less attention: how ordinary people can be empowered by AI.
Journalists have a long history of using new tools to supplement their reporting, especially in public policy. (Another example from ProPublica: their Nonprofit Explorer.) So instead of just writing about how Intuit deprived Americans of free tax-filing software, I figured: Why not also create the software itself?
That’s the full story behind TelosTax: I built a free, open-source, privacy-focused tax prep app that, at the very least, can bring awareness to the unsavory relationship between big tax prep and the Trump administration.
It is absurd that millions of Americans prepare their taxes with a product designed to siphon their data to sell them credit cards. But it’s even worse when the same company prevents the government from offering a similar service—absent the data mining—for free.
But by the end of this project, I realized something much bigger was at play in the collision between legacy software companies and new AI agents.
Why AI is addictive
The process of creating the app with AI was strange. I quickly learned that the tools can be addictive. I frequently got caught up piling new features onto the app just because I could: Tax Scenario Lab, Smart Expense Scanner, Audit Risk Assessment, TelosAI, obscure IRS forms, cool data visualizations. I added an external monitor to my laptop and populated it with six different instances of Claude Code to run tasks simultaneously.
I found myself yelling strange new phrases that alarmed my loved ones: “How can Claude only have a context window of 200k tokens?” (It’s now one million, thank God.) CLIs, API keys, and MCP servers became my superpowers. The pace of change was exhausting. Every tool I was using earlier this year has been significantly upgraded. New features that can completely change a project’s trajectory are released daily. In the middle of the project, Anthropic released Remote Control, which allows me to continue coding from my phone while I walk the dog.
It was not necessarily a healthy way to spend my time. The Harvard Business Review recently reported that using AI the way I was using it can lead to “AI brain fry.” I was nodding my head as I read that article.
As the weeks went by, the amount I spent on AI products ballooned—Anthropic Max, Google AI Ultra, ChatGPT Pro, Perplexity Computer. Playing models off of one another became a key part of the design process. GPT taught me how to write prompts to prevent Claude from hallucinating. Gemini was a useful skeptic of Claude’s overly ambitious pursuits.
I became adept at using a tool that sent Claude’s tax software ideas to three other AI models, which debated them and voted on the best way forward. They became Claude’s council of advisers and harsh outside critics. To find security vulnerabilities, I prompted them to think like hackers searching for bugs that will win them bounties. To test the tax engines, I told them to develop tax filer profiles with complex edge cases and obscure combinations of income, deductions, and expenses.
They were tough critics, but in the final rounds of pre-publication tests, Claude’s AI friends were quite impressed with him. GPT called the “pure-function engine architecture” the “crown jewel” of the app with “textbook clean architecture.” Gemini described the “authority chain (@authority JSDoc + YAML)” as “better than most commercial tax software.” Claude described the “privacy design with audit log” as a “genuine innovation” and a “masterclass in building trust.” Another model said that the “test suite methodology (oracle, metamorphic, fuzzing)” is “exceptional,” and the “documentation quality” actually “exceeds most commercial projects.” The “declarative wizard step conditions” were hailed as “brilliant.”
Do I understand all of that? Not really, but after weeks of constantly nitpicking every decision, it was nice to hear the other robots show some respect for the app. (Though, of course, these LLMs are wired to be sycophants, showering praise on their users.)
There was something empowering about using these tools to create bespoke software. I joked in previous posts in this series that I know nothing about coding, but the process actually teaches you quite a lot about design and many technical details, even while you learn almost nothing about writing in the programming language itself (TypeScript in my case).
This experience led me to a sweeping generalization: 2026 will be the beginning of a golden age of AI tool use. The number of AI tools that bring technological superpowers to amateurs will explode. We are entering the sweet spot where the tools are highly capable, fully under our control, and can make our jobs easier, more interesting, and more productive rather than replacing them. That’s a safe short-term bet. The more difficult question is whether this fun interlude is just a transition period on the way to the more doomsday and sci-fi scenarios of an employment apocalypse or AI misalignment that threatens civilization.
So how’d it do?
I released TelosTax as a free open-source app, so anyone could look at the code and suggest ways to improve it—or determine that it’s beyond repair. Experts can inspect how the TelosTax tax engines work, where the numbers come from, and whether the app accurately implements IRS statutes and regulations. My own testing was relatively limited compared to what a consumer-ready product would require, and I was fully prepared for the possibility that Claude, in its eagerness to please, built me a Potemkin app.
I told Claude to design TelosTax with strict privacy as the core architecture. It has two modes: free and AI-assisted. The free mode is as strict about privacy and data retention as you could possibly get. The entire app runs in your browser. There are no accounts and no cloud services to connect to. After you load the app, you don’t even need an internet connection to use it. No data leaves your device. The app forces you to save and encrypt your tax return, and if you lose your password, there’s no way to recover it.
But I also wanted to have AI tools as an option. If you turn them on, you can dictate or type in plain language any tax information you have, and the AI will enter it for you. You can right-click any element on any page in the app to ask the AI about it. You can do the same with every IRS tax form. The AI will keep track of your filing progress and nudge you when it sees a deduction or credit you might want to take.
Using AI comes at the cost of privacy. What you type or speak into a chatbot is sent to the AI provider’s servers for processing. To partially mitigate this, I built a PII stripper into the app. If you enter addresses, dates of birth, Social Security, credit card, bank account, or other similar personally identifiable information into the AI chat, a warning appears explaining that the information will be redacted before it’s sent. The AI settings include an audit page that logs every AI message and shows exactly what was sent from your computer after you hit send, including any contextual information that is normally invisible. As far as I know, no AI product offers this transparency, even though it was trivial to add.
To use the AI functions, you need a paid Anthropic account and an API key. I chose Anthropic because the company has clearer privacy and data retention policies than OpenAI and Google. (Anthropic says it doesn’t use data sent via API to train its AI models, and it deletes your data after 30 days.) This is probably all a little over-engineered and over-complicated for a demonstration project, but I wanted to really test the limits of whether it’s possible to fully replace a complicated piece of software.
So how did it hold up during the 2026 tax season? For roughly nine out of ten tax filers, the app worked as advertised by Claude. I created several moderately complex federal returns in both TurboTax and TelosTax and got identical results. And several users who were generous enough to do testing for me had similar positive experiences.
But it struggled with highly complicated returns that require power-user functions such as robust document import. A Telos reader who happens to be a CPA found the app choking on his multiple K-1 and 1099 PDFs and, even worse, making errors in self-employment tax calculations.
Software really is dead
But I am convinced that with enough time, any of these bugs could be fixed by Claude or similar frontier models. In playing around with Claude Opus 4.7, released on Thursday, I can already tell it’s significantly more sophisticated and would have caught many of the coding errors that I and other users found in the first versions of the TelosTax app. The key here is the pace of progress. A year ago, you could barely use AI to write complex code. A year later, AI can create highly sophisticated, though still buggy, software that requires significant human testing to perfect. A year from now—and perhaps sooner, given the early reports of Claude Mythos’ capabilities—it will be trivial to create something like TurboTax on your own.
But a year from now, creating an app similar to TurboTax to do your taxes will seem anachronistic. Just over the course of this project, the available AI tools have changed so dramatically that if I were starting over today, it wouldn’t even make sense to build a full-featured tax app that tries to address every tax filer’s unique situation.
We have rapidly moved into the agentic era, where a user simply asks Claude to accomplish a task, and Claude decides which existing tools it will use to complete the job and which tools it might need to build itself.
The AI agent is becoming the starting point for all digital work. Spending any mental energy deciding which app to open to complete a task will become increasingly less common. Most power users have already moved to a model where you grant Claude or a similar agent, such as OpenClaw, wide access to a computer that is always on and then just tell the AI, via your mobile device if you’re out and about, what you want accomplished. This is not exactly a new idea. It’s essentially Apple’s Siri on steroids.
I’ve fully embraced this workflow in my own life. I love nothing more than procrastinating by fiddling with the latest productivity apps. But I’ve ditched almost all of them for Claude. I now start my day by telling Claude what I want to accomplish. The AI asks me a few questions and maps out a plan. It monitors all the apps I used to check obsessively, and pings me when something truly requires my attention. The biggest downside to this new agentic era is that agents handle so much digital scutwork so well that I no longer have any excuse to divert my attention from writing and other high-priority human work.
I find all of this both exciting and terrifying. In the days and weeks ahead, I’m going to return to covering hardcore politics here at Telos, but I also plan to devote more attention to the issues we explored in this series. Drop me a line and tell me what you think.