Ionut Adrian Visan grew up in Craiova, Romania, in the hopeful years after the fall of communism, and got a first computer in 1996: a 286 running MS-DOS and Norton Commander. Games like Prince of Persia came first, then the urge to find out how everything worked. A first job maintaining a neighbourhood company's computers, reinstalling Windows almost weekly, taught troubleshooting and patience. Visan has been building games for about 15 years and runs Slotty Labs alongside AIStupidLevel.
AIStupidLevel started in 2024 as a tool for Visan's own work, after noticing that an AI model could be brilliant one moment and almost useless a few hours later. It now tests leading models continuously on real coding, reasoning and tool-use tasks, flags when a model quietly degrades, and feeds that data into a Smart Router that picks the right model for each request. It has reached more than 120,000 users without any paid marketing. We spoke about a name that started as a joke, retracting earlier alerts in public, and why choosing an AI model will one day feel like choosing a server.
Key takeaways
- AIStupidLevel continuously benchmarks leading AI models on real tasks: fixing bugs in working repositories against hidden tests, multi-turn reasoning, and tool use in sandboxes. Coding tasks run seven times per model.
- A cumulative-sum drift detector compares each model with its own history, so one bad answer does not trigger an alert, but a genuine decline does.
- The Smart Router uses the same data to pick a model for each request by quality, cost, speed or consistency, using the developer's own API keys.
- More than 120,000 users and two million visits with no paid marketing. The 12-month targets: 400,000 users and $5 million in annual recurring revenue.
Before any of this
Startup.talk:Where did you grow up, and what was childhood like?
Visan:I grew up in Craiova, Romania, in the first years after the fall of communism. There was a lot of optimism back then. Everything was changing, people were hopeful about the future, and with every passing year, the world seemed to become a little more colorful.
My childhood was pretty typical for Eastern Europe. Hot summers, long winters, spending most of our time outside with friends. We didn't have much compared to what kids have today, but we had plenty of freedom and imagination. Looking back, it was a wonderful time to grow up.
Startup.talk:What first got you into building things?
Visan:In 1996, my dad bought me my first computer. It was a 286 PC running MS-DOS with Norton Commander.
I immediately figured out how to play games on it. I still remember those big floppy disks and games like Prince of Persia, Volfied, and the old Accolade titles. I could spend entire days playing.
But eventually, playing wasn't enough. I became curious about how everything worked. How the computer ran those games, what all those files did, and what happened if I changed something. That curiosity never really went away.
Startup.talk:What was your first job, and what did it teach you?
Visan:My first job was maintaining computers, both hardware and software, for a small neighborhood company.
I'm talking about the days when reinstalling Windows was practically a weekly ritual. Something broke, you figured out what happened, fixed it, and a few days later something else broke.
It taught me to troubleshoot and, more importantly, to be patient. You couldn't just ask an AI what was wrong. You had to understand the problem yourself, experiment, and find a solution. I think that shaped the way I approach building software even today.
The road to AIStupidLevel
Startup.talk:What made you start AIStupidLevel?
Visan:Back in 2024, I started noticing something strange while working with AI models. A model could be absolutely brilliant one moment and almost useless a few hours later. Same model, same prompt, completely different results.
And nobody seemed to be paying much attention to it. We had all these leaderboards telling us which model was the smartest, but very little information about whether that model was still performing the same way today as it did yesterday.
I wanted something that would continuously test these models and tell me when something changed.
So I started building it. Initially, it was just a tool for myself. I figured someone had to keep an eye on these things.
“A model could be absolutely brilliant one moment and almost useless a few hours later.”Ionut Adrian Visan
Startup.talk:When did you first notice that the same AI model and prompt could give materially worse results from one day to the next?
Visan:Actually, within the first few days of running AIStupidLevel.
Once I started collecting results continuously, the differences became impossible to ignore.
What was interesting wasn't just that models sometimes performed worse. It was how much their performance could fluctuate without any announcement or explanation.
Of course, AI models are probabilistic, so some variation is normal. The real challenge was figuring out when we were seeing ordinary randomness and when something had genuinely changed.
That became one of the most important problems we wanted to solve.
Startup.talk:How did AIStupidLevel reach more than 120,000 active users without paid marketing?
Visan:I think the simplest explanation is that it solves a real problem.
I've never spent money on advertising AIStupidLevel. No paid campaigns, no growth hacks or marketing agency. Everything has been organic.
Developers were experiencing the same frustrations I was, and when they found a tool that helped them understand what was happening with AI models, they started sharing it.
A Reddit post went viral, publications picked up the story, and traffic started coming in from all over the world.
I think when you build something genuinely useful, people eventually find it. It doesn't happen overnight, but word of mouth is incredibly powerful.
Startup.talk:You also run Slotty Labs. How do you split your time between the two, and what is it like to build from Craiova?
Visan:I've been building games for about 15 years, and I absolutely love it.
Slotty Labs is what keeps that passion alive. I genuinely enjoy creating unusual game concepts, experimenting with mechanics, and watching an idea turn into something people can actually play.
AIStupidLevel originally came from that same world. I was using AI heavily while developing software and games, and I needed a better way to know which models I could rely on.
So in a way, the two businesses complement each other. Slotty Labs lets me keep creating, while AIStupidLevel helps me build more efficiently and has grown into something much bigger than the tool I originally imagined.
As for Craiova, I love it here. The internet has made geography much less important for the kind of work we do. I can build products used by people across the world without having to move to San Francisco or London.
You don't need to be in Silicon Valley to build something valuable.
“You don't need to be in Silicon Valley to build something valuable.”Ionut Adrian Visan
The product
Startup.talk:Why the name AIStupidLevel?
Visan:It started as a joke.
I was frustrated with how inconsistent some AI models were, and I thought it would be funny to measure not just how intelligent they were, but how stupid they could become.
I honestly never expected the project to get mainstream attention, so I didn't spend much time thinking about branding.
By the time it started gaining traction, the name had stuck.
And actually, I quite like it now. It's memorable, it doesn't take itself too seriously, and underneath the funny name there's some very serious engineering.
Startup.talk:How do you design benchmarks that measure real coding, reasoning and tool-use quality rather than test-set memorisation?
Visan:The most important thing is that we don't just ask models questions and check whether their answers look correct.
We make them actually do things.
For coding, we give models working repositories with real bugs to find and fix. Then we execute their code against tests, including hidden tests they haven't seen.
For reasoning, we use longer, multi-turn tasks where models need to remember instructions, maintain a plan, deal with new information, and recover from mistakes.
For tool use, we put them in sandboxed environments where they have to call real tools and complete actual tasks.
We also repeat tests. For example, our coding tasks run seven times per model because a single successful answer doesn't tell you much about reliability.
The benchmark tasks themselves are private to reduce contamination, while the scoring methodology is public.
My philosophy is pretty simple: I care much more about whether a model can get the job done than whether it can produce an impressive-looking answer.
Startup.talk:How does CUSUM drift detection decide that a model has actually degraded?
Visan:The basic idea is that one bad result doesn't mean a model has become worse.
AI models naturally fluctuate. Sometimes they make mistakes, sometimes they perform exceptionally well. If we sent an alert every time a score dropped, the system would be practically useless.
We use a cumulative-sum approach, specifically a Page-Hinkley detector, that compares a model's recent performance against its own historical behavior.
Rather than reacting to every fluctuation, it looks for evidence that the model has shifted away from its normal performance level.
We also monitor coding, reasoning, and tool use separately because each behaves differently. And we require enough historical data before making judgments.
The goal isn't to catch every bad answer. It's to identify meaningful degradation without constantly raising false alarms.
Startup.talk:How does the Smart Router choose the right model for each request?
Visan:It uses the data we're already collecting.
Instead of hardcoding a model name into your application, you can tell our router what matters to you: coding performance, reasoning, tool use, overall quality, cost, speed, or consistency.
The router looks at the latest benchmark results, considers your preferences and constraints, and selects the most suitable available model.
It also accounts for provider health and degradation signals. If a provider is having problems, or a model is no longer performing as expected, the router can move to another suitable option.
You connect your own provider API keys, so you're not forced into a particular ecosystem.
I wanted developers to spend their time building things, not constantly checking leaderboards and changing model names in their code.
Startup.talk:Who gets the most value from AIStupidLevel: individual developers, startups or enterprises running AI in production?
Visan:All three, but in different ways.
Individual developers use it to figure out which models are worth their time and money.
Startups can use the Smart Router to reduce API costs without sacrificing too much quality, which can make a big difference when you're operating on a limited budget.
For enterprises, reliability becomes extremely important. When AI is part of a production system, you need to know if a model starts behaving differently, if your provider is experiencing problems, or if a cheaper alternative could deliver similar results.
The stakes are different, but the underlying problem is the same.
Startup.talk:What do you deliberately leave out of AIStupidLevel?
Visan:We're not trying to build another chatbot, and we're not trying to create our own foundation model.
I also don't want to turn AIStupidLevel into a leaderboard where we add hundreds of meaningless measurements just to make it look impressive.
I'd rather measure fewer things properly than publish numbers I can't defend.
We deliberately avoid presenting statistical noise as meaningful differences, and we don't publish our hidden benchmark tasks because that would eventually make them useless.
The focus is on measuring what matters, being transparent about limitations, and making the results useful.
“I'd rather measure fewer things properly than publish numbers I can't defend.”Ionut Adrian Visan
Positioning
Startup.talk:Who is AIStupidLevel for, and what job does it do for them? Describe the person who gets the most value.
Visan:Imagine you're a developer or CTO building an application that depends heavily on AI.
You're spending money on API calls every day. New models are being released constantly. Prices change, performance changes, and sometimes the model you've been relying on starts making mistakes it wasn't making last week.
You don't have the time or resources to continuously evaluate every available model yourself.
That's the person we're building AIStupidLevel for.
We do the testing, monitor changes, and help them choose and use the right models.
Essentially, we want to remove the guesswork from working with AI.
Startup.talk:How do you view your positioning relative to LMArena, Artificial Analysis, OpenRouter and the labs' own published benchmarks? Is there another competitor you think of first?
Visan:I think all of them are doing valuable work, and the market is big enough for different approaches.
LMArena is particularly useful for understanding human preferences between model outputs. Artificial Analysis has built a very comprehensive platform for comparing capabilities, performance, and pricing.
OpenRouter is an important player in model access and routing, and they've also developed their own intelligent routing capabilities.
The labs' published benchmarks are useful too, although naturally I think independent verification is important.
Our focus with AIStupidLevel is continuous, reproducible measurement tied directly to operational decisions.
We want to know not only how a model compares with others, but whether it has changed, how reliable it is over time, and whether you should still be using it for your particular workload.
Then we connect that information directly to the Smart Router.
I don't really think of this as a winner-takes-all competition. Different products answer different questions.
Startup.talk:What is your core moat in AI model benchmarking and routing?
Visan:The historical data, the evaluation infrastructure, and the connection between measurement and routing.
We've conducted more than 500,000 controlled evaluations, and every new run adds to our understanding of how models behave over time.
A single benchmark result is relatively easy to reproduce. Building an infrastructure that continuously measures models, understands normal fluctuations, detects meaningful changes, and turns that information into routing decisions is considerably harder.
But I don't think having a complicated algorithm is a moat by itself.
The real advantage comes from accumulating reliable data, proving that your measurements mean something, and earning the trust of developers who depend on them.
That takes time.
Startup.talk:What are some of the use cases for AIStupidLevel that are not obvious and can add a lot of value to users? List as many as possible.
Visan:People usually discover us because they want to compare AI models, but there's quite a bit more you can do.
For example:
-Detect when a model you've been using starts performing worse, even if the provider hasn't announced any changes.
-Reduce API spending by identifying cheaper models that perform nearly as well on a specific workload.
-Automatically switch models when a provider is unavailable or a model becomes unreliable.
-Monitor the consistency of AI agents that depend on tool calling.
-Compare how different providers serve models and how that affects reliability.
-Track model performance before and after updates.
-Build custom evaluation pipelines using our historical data.
-Choose different models for coding, reasoning, and tool-heavy tasks instead of using one expensive model for everything.
-Identify models that are significantly more expensive without delivering meaningfully better results.
-Help engineering teams make evidence-based decisions about which AI providers to adopt.
-Monitor the performance of models used inside automated business workflows.
-Research how model reliability changes over weeks and months.
-Use benchmark history to decide when migrating away from an older model makes sense.
-Detect instability before it becomes a bigger problem for an application.
We're also expanding the platform so model providers and hosting companies can bring their own models into our benchmarking and routing ecosystem.
And for companies with specialized AI workloads, custom benchmarks and evaluations are an especially interesting direction.
There's much more value in understanding how AI behaves in the real world than simply knowing which model happens to be number one today.
Startup.talk:What is the one detail in AIStupidLevel you are proudest of?
Visan:Probably the fact that we're willing to say when we don't know.
That might sound strange, but AI benchmarking is full of uncertainty.
Sometimes two models have slightly different scores, but statistically, there's no meaningful difference between them. We don't pretend otherwise. They can share the same rank.
The same applies to drift detection. A model having a bad day doesn't automatically mean it has degraded.
We've put a lot of work into making the system honest about what the data can and cannot tell us.
I'm proud of that because it's very easy to publish impressive numbers. It's much harder to make sure those numbers are actually meaningful.
Launch
Startup.talk:What did you learn from your launch?
Visan:That launching a product and building a useful product are two very different things.
A launch can bring attention, but that attention doesn't mean much if people don't have a reason to come back.
For me, the most important lesson was to keep focusing on the product itself.
I would rather have a smaller group of developers who genuinely depend on what we're building than a massive spike in traffic that disappears the next day.
The real work starts after the launch.
Startup.talk:Which feedback from the launch surprised you?
Visan:What I found most interesting was how the conversation could move beyond rankings.
You might expect everyone to ask which AI model is the smartest.
But the more useful question is often whether the model they're already using is still as good as it was before.
That's a very different problem.
And it helped reinforce the direction we're taking with drift monitoring and the Smart Router. It's not enough to tell people what's happening. You also want to help them do something about it.
The hard part
Startup.talk:How many people use AIStupidLevel today?
Visan:We've reached more than 120,000 users historically, entirely through organic growth.
We've also passed two million total visits.
I think it's important to distinguish between cumulative users, monthly visitors, and registered accounts, because they're not the same thing.
What interests me most now is how many people move from checking the benchmarks to using our data and routing infrastructure as part of their everyday work.
That's where we can create lasting value.
Startup.talk:What is a challenging time you had building AIStupidLevel?
Visan:One of the hardest parts has been making the measurements genuinely trustworthy.
It's surprisingly easy to build a benchmark that produces numbers. It's much harder to build one that produces numbers you can confidently stand behind.
At one point, we discovered that parts of our earlier alerting methodology were too sensitive and were identifying changes that weren't sufficiently supported by statistical evidence.
We had to revise the methodology and retract earlier incidents.
It wasn't pleasant, but it was absolutely the right thing to do.
There's also the financial side. Running hundreds of thousands of real AI evaluations costs a significant amount of money, especially when you're bootstrapping.
But I've always believed that if you're building a product around trust, you have to be prepared to admit mistakes and fix them publicly.
“if you're building a product around trust, you have to be prepared to admit mistakes and fix them publicly.”Ionut Adrian Visan
Startup.talk:Who believed in you before there was anything to show? What did they say or do?
Visan:In the beginning, I wasn't really looking for anyone's permission or approval.
I was building something I needed myself, and that was enough of a reason to start.
The first real validation came from strangers on the internet.
People who had no reason to support me started using the platform, sharing it, and coming back. Some were developers working on their own projects, others were simply curious about how these models behaved.
Seeing people find value in something I originally built for myself meant a lot.
There's something special about realizing that a problem you've been struggling with is also a problem thousands of other people are facing.
Startup.talk:How do you stay independent and trusted when the companies you benchmark are also potential partners or customers?
Visan:By making sure that commercial relationships don't determine evaluation results.
Every model is measured using the same methodology, the same tasks, and the same scoring rules.
We publish how our measurements work, including their limitations. The actual benchmark tasks remain private so they can't simply be optimized against.
As we expand to allow AI providers and hosting companies to submit their own models, the same principle will apply.
A provider can participate in the ecosystem, but that doesn't mean it gets a better score.
If we ever lose the trust of developers who rely on our data, the entire business loses its purpose.
Independence isn't just a nice thing to have. It's fundamental to what we're building.
What it is all for
Startup.talk:What would make you say AIStupidLevel has succeeded?
Visan:When developers no longer have to worry about which AI model they should use.
That might sound a little strange coming from someone who built an AI benchmarking platform, but that's genuinely the goal.
I want people to focus on building their applications while the infrastructure handles model selection, reliability, performance monitoring, and cost optimization.
If thousands of businesses are using our data and Smart Router every day, and they can trust that their AI infrastructure is making sensible decisions without constant manual intervention, I would consider that a huge success.
Startup.talk:What will choosing an AI model look like when there are thousands of them and they change every week?
Visan:I don't think people will choose individual models the way they do today.
There will simply be too many of them.
You won't want to spend your morning comparing 3,000 different models, reading release notes, and trying to figure out whether the newest one is actually better.
Instead, you'll specify what you need.
Maybe you want the highest coding accuracy within a certain budget. Maybe you need extremely reliable tool calling. Maybe latency matters more than anything else.
The infrastructure will select the appropriate model and adjust as the available options change.
I think choosing an AI model will eventually become a lot like choosing which server handles a request. Most people won't need to think about it.
That's the future we're working toward.
“I think choosing an AI model will eventually become a lot like choosing which server handles a request.”Ionut Adrian Visan
The road ahead
Startup.talk:What are your milestones for the next 12 months, in users and revenue?
Visan:The next 12 months are really about turning the organic interest we've built into sustained production adoption.
We've already demonstrated that developers want independent information about AI models. Now we need to make AIStupidLevel something they actively integrate into their applications.
My targets are to reach 400000 users and 5000000 USD ARR, with a growing number of companies using our Smart Router and Data API in production.
We're also expanding support for model providers and hosting companies, as well as working toward more specialized enterprise evaluations.
I care about user growth, of course, but I'd much rather have a thousand companies genuinely depending on the platform than a million people visiting once out of curiosity.
Startup.talk:And for the next 24 months?
Visan:In 24 months, I want AIStupidLevel to be a recognized part of the AI infrastructure ecosystem.
By then, I want us to be supporting a much broader range of models and hosting providers, including self-hosted and open-weight deployments.
The idea is to build something that scales internationally without losing the independence and transparency that made people interested in the first place.
Startup.talk:Where do you want AIStupidLevel to be in five years?
Visan:I want AIStupidLevel to become one of the standard intelligence layers behind AI applications.
Not necessarily something every end user knows about, but something that developers and businesses depend on.
When a company needs to understand how an AI model performs, whether it's reliable, how its behavior has changed, or which alternative would work better, I want AIStupidLevel to be one of the first places they turn to.
And ideally, much of that process will happen automatically through our infrastructure.
Five years is a very long time in AI. I have no idea what models will look like by then.
But I'm fairly confident that as AI becomes more capable and more widely used, independent measurement and reliability will become more important, not less.
Startup.talk:What are your hopes and fears for a post AGI world?
Visan:My biggest hope is that intelligence becomes abundant and accessible to everyone.
Imagine what that could mean for medicine, scientific research, education, or even something as simple as giving a child anywhere in the world access to the best teacher imaginable.
I think the possibilities are extraordinary.
My biggest fear is that we become so dependent on these systems that we stop questioning them.
Or that an enormous amount of power becomes concentrated in the hands of a very small number of companies or governments.
I don't think having machines that are more intelligent than us is necessarily the frightening part. What worries me more is whether we'll use that intelligence wisely and whether people will still have meaningful control over the decisions that shape their lives.
I'm optimistic about the future, though. I wouldn't be building in this industry if I weren't.
I just think we need to be a little more careful about what we're creating and a lot more serious about understanding it.
This interview was conducted in writing and has been edited lightly for length and clarity.
About AIStupidLevel
AIStupidLevel is an independent AI benchmarking platform built by Ionut Adrian Visan in Craiova, Romania. It continuously tests leading AI models on real coding, reasoning and tool-use tasks, detects when a model's performance drifts, and powers a Smart Router that selects the best available model for each request using the developer's own API keys. aistupidlevel.info
| Attribute | Verified Detail |
|---|---|
| Company | AIStupidLevel |
| Founder(s) | Ionut Adrian Visan |
| Country | Romania |
| Based | Craiova, Romania |
| Grew up | Craiova |
| Also | Founder and CTO, Slotty Labs |
| Background | Software architecture and full-stack engineering; building games for about 15 years |
| Users | More than 120,000, all organic |
More founder interviews
For founders
Building something? Want to share your journey?
Tell us how it started, what nearly broke it and where it is headed. We will turn it into a story like this one.
Share your journeyFor readers
Want to read more human stories behind the startups?
New founder interviews and stories, as they are published.
Follow us on LinkedIn
VoCalm
TruTalk
Skillbox
Appulse Index