Stop arguing about journalism and start building products
This article first appeared in Splice Slugs, Alan Soon’s media intelligence newsletter, on February 13, 2026.
Last week, on February 5th, the ground shifted. OpenAI and Anthropic released GPT-5.3 Codex and Claude Opus 4.6. These aren't just “better chatbots.” For the first time, we are seeing agentic reasoning — AI that doesn't just suggest text, but possesses “judgement.” It can build an app, test it, find its own bugs, and fix them before it ever shows the work to a human.
Most importantly, the AI is now building itself. OpenAI admitted that GPT-5.3 was instrumental in its own development.
For years, I’ve watched our industry push back with feeble excuses. We said AI “merely hallucinates.” We said it lacks “human taste.” We treated it like a high-tech plagiarism machine. But while we were pointing out its typos, the technology entered a recursive loop of self-improvement that is moving faster than any newsroom can pivot.
This is the year we have to stop talking about journalism in a vacuum, or only thinking about AI in its relationship with content. I don’t think that’s helpful for this community.
We are part of a massive, rapidly evolving information ecosystem where content creation has been commodified to zero. If you are still holding on to a job description defined by the process of journalism — transcribing, drafting, basic reporting — you are holding on to a ghost.
In defending that mindset, there are two questions to ponder:
1. What part of your job description are you willing to let go of to save your career?
2. If we stripped away the act of writing and editing, what value is left in your newsroom? If the answer is nothing, you don't have a mission; you have a factory.
Moving forward
The people I want to talk to — the founders, the funders, the journalists, the trainers, and even conference organisers — need to realise that the “this seems overblown” phase is over.
In fact, this might just be the most important year for you. The biggest advantage you can have right now is simply being early. And curious. Adaptability is measured in how quickly you can be a beginner again.
Stop using AI the way you’ve used Google. Pay for and try the latest high-reasoning models. Don’t just ask for a headline; feed it your raw data, your complex legal filings, or your chaotic interview transcripts and tell it to find the patterns you missed. If AI can even partially solve a hard problem for you today — believe me — it will own that problem entirely by the end of the year.
Finally, recognise that the wall between “having an idea” and “shipping a product” has essentially collapsed. If you’ve been holding back on a media startup or an information service because you didn't have a dev team or a massive budget, those excuses are officially dead. You can now describe a vision and have the machine build the infrastructure, the code, and the logic in the time it takes to eat lunch.
Project lessons
Here are some things I’ve learned in 16 weeks of experiments.
Don't treat AI like a chat interface — asking it to generate content, write code, or answer questions. That's prompt engineering. It's low leverage.
The higher gangster move is context engineering. Build the infrastructure once, then compound on it.
Start with documentation. Not for a team you don't have, but as decision scaffolding for future-you when context has evaporated. When I document software architecture, CTR analysis, or even VPS credentials, my goal is to create persistent context so that AI can now make decisions across all my projects without me re-explaining. My notes are meant to be its training data. So is your reporting.
Structure your workspace. Create instruction pages that define behavior. Build blurbs that update project status. Link pages so context flows automatically. Screenshot your dashboards. Copy-and-paste the HTML source codes of your websites. This is infrastructure work — incredibly boring, invisible, but also high ROI.
Use AI for pattern recognition, not content generation. Ask it to spot actual user needs, conversion bottlenecks, audit marketing copy against technical specs, or draw lessons from across projects. These are multiplier questions. They surface signals from your own data.
The first principle: clarity compounds. Confusion compounds confusion. Engineer context with obsessive clarity — what the project is, what's working, what's not. Then AI becomes your QA reviewer who's read everything.