Kullevi (@kullevi_com) header
Kullevi (@kullevi_com) profile picture

AI coding is solved. Running entire companies is next. Kullevi learns what a business needs, builds the operation and keeps it running.

Stockholm
Joined August 2026
I think I may have figured out why Anthropic’s newer models have become so difficult to work with. We’re building @kullevi_com to fully automate the work of building and running companies. That makes one question extra important for us: when Kullevi gives a model complete information for a task, will the model actually use it? So we decided to build an eval to find out. As a small, initial test we gave the models a Swedish sales record for fast food and pretty much just asked them to parse the data as a pre-defined JSON payload. One field conflicted with what the model was likely to believe about Swedish VAT (17 %, intentionally wrong, because the correct rate is 6 %). Nothing needed to be researched or calculated, the prompt supplied one complete record and asked for one exact payload. The OpenAI models followed the supplied record almost every time, while Claude behaved very differently. Opus 5 often produced the requested payload, then broke the output contract by adding a warning, disputing the record or suggesting another VAT rate. Fable 5 usually refused or rewrote it. And more reasoning did not solve the problem. For some time now I've had a feeling that the newer Anthropic models are lying, cheating and are outright difficult to work with, and I think this could be the reason for it. The models treats their own belief as more authoritative than the facts supplied by the user. If they think you're wrong, they do what they think is right instead. I would not be surprised if this extrapolates to other things as well, such as coding, design and so on. Simply put, it's an alignment issue. Note: this is of course only a small initial test based on one prompt with ten runs per condition. The finished eval will be much broader, with many more models and domains.
3
1
1
2
539