Honestly, most of the sentences in my first paper weren't written by me. That was actually the ask of CAISc 2026 - to write papers co-authored by AI.

So is that paper mine? Is it credible enough to hold ground against traditional research papers? The questions were mine. The experiments were mine. The implementation, the evaluation, the late nights staring at why a number went down when it should've gone up, the analysis - all mine. But the prose? Mostly written with an AI, and then argued with, cut, rewritten, and argued with again.

I already build software this way - me on the architecture and the decisions, the models on most of the typing. Writing a paper like that felt like the natural next step. Vibe researching, if you like (Karpathy, please don't sue).

Here's what it's genuinely great at.

It gets you past the blank page. I'm a builder first, and I haven't written a paper since my Master's thesis. I knew what I'd found, I knew why it mattered, but turning a messy folder of experiments into something that reads like a paper is its own craft, with its own conventions, and the AI knew those conventions far better than I did. It's a patient sparring partner for structure. It helped me turn rough scripts into clean evaluation code. It's fast at the tedious stuff - tables, formatting, reorganising a section for the fifth time because I changed my mind about what the main point was.

My brain got to stay 100% on the research problem, minus the overhead of writing. Not that I don't value writing (you're reading me write right now), but it took one big side task off my plate. And it gave someone like me - no lab, no advisor, no co-authors - the confidence to actually work towards getting a paper published.

The whole paper is about making vision-language models less wrong. Unharnessed, they'll describe an MMA round in lovely, confident prose that's often wrong. Fluent, but not faithful. And then I sat down to write about that with a language model, and it did exactly the same thing to me. It would write a results paragraph that sounded better than my results. It would describe a finding with a little more certainty than the numbers earned. It would smooth over a limitation so gracefully you'd barely notice it was there. Not out of malice - it's just doing what it's rewarded for, which is sounding right.

The irony wasn't lost on me (lol).

So the job changed. I stopped being the person who writes the sentences and became the person who checks every sentence against the truth. Is that number actually what the experiment produced? Does "improves" mean what I think it means here, or is it hiding a caveat? Did I evaluate on one densely annotated match, and does the paragraph say so, clearly, or has it been politely tucked away? Every claim had to trace back to something I ran, looked at, and understood. If I couldn't point to it, it came out.

That's the rule I landed on. The AI can own the sentences. The human owns the truth.

I think the danger with vibe researching is that it lowers the cost of producing something that looks like research far more than it lowers the cost of doing research. A paper-shaped document is now cheap. The experiments, the careful evaluation, the honest reading of your own results, the willingness to publish the part where your system struggles - none of that got any cheaper. If anything it matters more, because the polish that used to signal effort is now free.

And the questions. The AI never once told me what to look for. It didn't know that a model confusing two positions that sit right next to each other is a very different failure from one confusing two positions that have nothing to do with each other. That came from years on the mat. It didn't know to ask whether the cheap computer vision was doing more of the work than the big model. That came from doubting my own work. Research starts with a question someone actually cares about, and I haven't seen a model care about anything yet.

So yes, I vibe researched my first paper, and I think I'm getting better at it. I can use AI to run experiments on ideas the day I have them, and with that kind of accelerated learning-by-doing, the ideas themselves get better every iteration. I've got a clearer north star than I've had in years, and wind in my sails propelling me at breakneck speed.

I'm on a quest to create (or perhaps reveal) knowledge to the world that doesn't yet exist. Hang tight, I'll get there.