Assessing performance and limits
We put the newest large language models under deliberate load and look for the points where they break. With benchmarks, error analysis and calibration we measure what a model can actually do.
We know AI from the inside. We test current AI models systematically and train our own.
We investigate which tasks a model handles reliably and where its limits are. Benchmarks, error analysis and our own experiments inform our technical decisions. We publish methods, data and results openly.
OpenMethods and assumptions visible
EmpiricalClaims grounded in data
AppliedResearch connected to practice
What we investigate.
Training our own models
We train our own models and adapt large language models to specialist domains. Having trained a model yourself, you see where its strengths and weaknesses come from and choose the right one for each task.
Developing experimental applications
We build experimental applications with current models and test them on concrete tasks. This helps us identify which approaches merit further development and which limits implementation must account for.
What we have published so far.
Who Trades on Polymarket?
A heuristic classification and comparative analysis of trader types on a prediction market.
The study classifies more than one million pseudonymous wallets from their observable trading behaviour. It shows how sharply market weight, profit, and forecast quality diverge across Retail, Informed Traders, Whales, Algorithmic Traders, and Market Makers.
Read the study
A model you want to examine closely?
We test it with you on your data, build a benchmark together or train a model of your own. We are just as glad to talk about research collaborations.