Skip to main content
Beast Launches

The story · Suprmind

The champion who did not exist

Radomir Basta speaking on a digital marketing conference

By , Founder and CEO · Serbia

Out · 4 min. Worth it.

Suprmind homepage

How it started

I was writing an article about a gun. One particular rifle. I needed a closing detail, so I asked an AI to find me a shooter who competes with it in the IPSC league.

It gave me a name. A Polish shooter. A champion. The competition he won. Article done.

I only went to Google because I wanted a few photos of him. There were no photos. There was no him. The AI had made up an entire sports career, and I had already called the article finished.

Back at Four Dots, my agency, we were running client work through five different AIs. Five browser tabs. The same prompt pasted five times. Answers that half agreed, and an hour spent working out which one was lying.

For a long time I had one question: what is smarter than the smartest AI in the world? At the end of 2025 we found our answer. Five of them in one conversation, reading each other.

The problem

One AI gives you one confident answer. It sounds exactly the same when it is right as when it is wrong.

For an email, that is fine. For a contract clause, a pricing change or an investment memo, it is not. The people making those calls - consultants, analysts, founders, lawyers - pay for a wrong answer with money and with their name.

Before Suprmind they did what I did. Ask ChatGPT. Then Claude. Then Gemini. Then Perplexity, because somebody has to check the sources. Compare the answers by hand, badly, at 11pm.

We later measured it. In our Q3 2026 study of 17,885 real questions, blind AI judges flagged a correction in 45% of the high-stakes turns they reviewed. A judge's call is a reading, not proof. But ask one AI and nobody flags anything. You just get the confident answer.

What we built

Suprmind is a multi-AI chat platform. You ask one question. Five AIs - GPT, Claude, Gemini, Grok and Perplexity - answer in the same thread, and each one reads what the others wrote before it replies.

So they build on each other. They also contradict each other, and when one of them makes something up, the next one in line can call it out.

You choose how they work. As a relay, each one building on the last (Sequential). All at once, merged into a single answer with the disagreements kept visible (Super Mind). As a structured Debate. Or as a Red Team that attacks your plan from six angles before you spend money on it.

The part I care about most: the disagreement does not get smoothed over. Suprmind scores how much the models agreed, shows you where they split, and can turn the argument into a decision brief you can hand to someone.

Plenty of tools give you five models in five separate windows. Here the five read each other. Disagreement is the feature.

I will not tell you this beats one AI at everything. For most daily tasks - emails, summaries, simple questions - a single model is faster, cheaper and perfectly fine. This is for the calls where fine is not the bar.

The hard parts

In March 2026 the dashboard showed 52 registered users and zero paying ones. I was my own heaviest user by a wide margin.

Then the traffic came, and it was the wrong traffic. I have done SEO since 2010 and I still ranked my own product for the wrong intent. One page about Grok pricing pulled in around 200 free trials a day. None of them converted. I redirected the page.

Some people signed up expecting image generation. Suprmind does not make images. They left, and that was on me, because the site did not say it clearly enough. Now it does.

Google Ads has lost money since day one. Organic search passed it a while ago.

The worst two days came in the middle of 2026. The app crashed. Nearly every customer who had just signed up for the Frontier plan cancelled that day.

Milestones

The first Enterprise customer came on board on June 11, 2026.

Three research reports published through Suprmind Data Lab. The latest Divergence Index covers 17,885 real questions. The first one, in April 2026, covered 1,324 turns. Our AI hallucination rates page became the most visited page on the site. 936 new accounts in 30 days, as of October 2026.

True North, our fact-checking layer, went live in October 2026.

In the summer of 2026 the five AIs literally bullied me out of adding a sixth model. I had dev docs ready. They had better reasons.

No outside funding. The businesses I built before Suprmind pay for it. Everything we built in Suprmind was first argued over in Suprmind.

What's next

Adjutant is next in line. It reads a whole project and tells you what is unfinished and what to ask next. Today it runs on a small number of accounts while I tune what each read costs. Then it goes out to the higher plans.

On the research side, I want the hallucination rates page and the AI Models Index to update themselves every day instead of waiting for me. And I want to publish a plain guide to fighting hallucinations for people who use one AI and plan to keep using one.

Where I want this to go: the place a professional takes a hard call before money moves. Your idea gets argued with before your budget does.

What got built

More builders

Questions?