How Two Brothers Are Trying to Put the Person Back Into AI
Skip to content
Sign in<br>Subscribe
Light
Dark
System
Merriam-Webster's 2025 Word of the Year was "slop". That's the backdrop Wilfred and Onome Okajevo built Noren against: a feed that's become measurably more uniform as more people lean on AI to write, and a category of tools that have got better at fluency while getting worse at preserving the person doing the writing.<br>Noren is a voice extraction engine. You feed it samples of your real writing, it builds a structured profile of how you actually write, and that profile can be used with Claude, ChatGPT, Gemini, or any model that takes instructions. The output stops drifting toward the statistical median and starts sounding like you.<br>That's the simple version. The interesting version is why two brothers, one with a research background in how humans think and communicate, the other with seven years building distributed financial systems, decided this was the problem worth a year of their lives.<br>Noren captures your actual writing patterns, not a generic role costume.The Brothers Building Noren<br>Noren has two cofounders, and they're brothers.<br>Wilfred Okajevo leads the ML and cognitive-science side of the product. His background is in building systems that understand how humans think and communicate, which is the lens behind Noren's voice extraction engine.<br>Onome Okajevo brings the operator side. He's an MD turned builder with seven-plus years building distributed financial applications, and he's spent a lot of time thinking about systems, risk, mechanisms, and how things fail in the real world.<br>That mix shaped the product. Wilfred's description of it:<br>"It is a technical product, but the problem is deeply human: how do you let people use AI without slowly sanding away the identity in their writing?"<br>You can see that split in the product itself. The extraction engine is research-heavy and structural. The way Noren positions itself, sits next to other tools rather than trying to replace them, and ships across macOS and Chrome is the operator instinct showing up in product decisions.<br>The Diagnosis<br>If you've spent any time on LinkedIn in the last two years, you'll recognise the pattern Wilfred describes. Three posts about building a company that all feel like one long post. Competent writing, properly punctuated, full of "robust" and "pivotal" and "facilitate", and somehow forgettable five minutes later.<br>His framing in The Homogeneous Feed:<br>"When AI writes without a voice profile, it draws from the same statistical defaults across every tool and every user. The result is a feed where a thousand different people sound like one generic voice."<br>The cause, in his framing, is mechanical rather than cultural. AI writing assistants don't have a voice. They have a distribution. A probability-weighted vocabulary trained on formal professional text, with sentence rhythms that cluster around the same 15-to-25-word range, regardless of who's prompting it.<br>The collective effect emerges from individual rational choices. Each writer has things to say and limited time. Each one reaches for the tool that speeds production. The output of any one writer looks fine. The output across thousands converges.<br>The Moment It Clicked<br>The trigger for Noren, in Wilfred's words, was realising that "write like me" was being treated as a prompt problem.<br>The brothers did what most people do first. They looked at samples, wrote rules, described tone, tested outputs, and adjusted prompts. The outputs got cleaner without getting more like the person.<br>So they went manual.<br>They spent weeks documenting how people write by hand. Sentence rhythms, word choices, where analogies came from, how arguments were sequenced, which transitions a person avoided, what punctuation felt wrong for them. Hundreds of lines of logic. That was how they built their first voice profile, by hand, with no extraction engine yet.<br>The moment it clicked was when that deeper profile made the output finally sound right. Not because the model got a better adjective like "professional" or "conversational", but because it had a more structural map of the writer.<br>That reframed the problem. In Wilfred's own words:<br>"It was not 'write a better prompt.' It was 'extract the identity layer from real writing and make it reusable.'"<br>Noren extracts structural patterns rather than describing tone in plain language.Around the same time, the brothers were watching the feed itself change. More posts were competent, fewer had edges. Different people were starting to sound like the same person using the same model defaults. That observation became The Homogeneous Feed, and it became part of the case for building Noren as a category rather than a feature.<br>The Lane Noren Picked<br>The AI writing space already has plenty of products. Jasper for brand voice. Sudowrite for fiction. Grammarly for editing. Every model lab is shipping a writing feature of some kind.<br>Noren picked a smaller lane on...