LLM hallucinations
Recently I had the most surreal experience with Claude hallucinating. I remember using Fable during my vacation in Croatia and feeling amazed that I could get very complex tasks done from a phone on my personal projects. After returning to work, where I only have access to Opus and models from GitHub Copilot, I started fixing a complex bug related to Async Local Storage and Promises in Node.js.
Initially my understanding of the problem was good enough, and I felt that Claude was really helping me. But something was off, and I knew I needed a review from a more senior colleague who actually wrote the original code or was involved in its more recent changes.
The review made it clearer and clearer that I was completely off. Towards the end, I realized Opus 5 simply does not run the tests and hallucinates the test results. That prompted me to try a different approach. I remembered hearing colleagues talk about mixing models when one of them is just off. I tried to get a review from GPT Sol 5.6 through GitHub Copilot via OpenCode. It immediately told me to keep just one of the changes, reduce it to a one-liner, and move the rest to a separate session to investigate properly.
I was really amazed how much the models can differ from each other. I think this non-deterministic behavior is incredibly dangerous for long-term projects when not all team members fully understand the codebase. One thing I took as a lesson from this is using different models from different vendors to get better confidence in the tests. This also leads to savings. It is fascinating that something like Sonnet 4.6 with follow-up with GPT models can get much better results and save also money on tokens. I have been reading about this approach in the past, but never thought the effect can be so dramatic.
I think we got collectively spoiled by the initial feeling of accomplishment and the constant encouragement from Anthropic models, but sometimes I really want proper engineering and proper tests. Trusting a single vendor is an insane risk, because the model really wants to finish the task to the point that it will inevitably cheat on the result.
It is the same as with human engineers. If you want proper quality, QA needs to be independent from developers. Otherwise, the conflict of interest is inevitable. I actually feel better when I know someone else is reviewing the product I am working on and making sure I am not cutting corners to get it done on time.
I was never a fan of OpenAI, but I have to admit that the coding output is good and the Codex CLI itself is really nice and pleasant to use. I hope the smaller models will get better and I will get better hardware to run them locally. I really like where the Unsloth and vLLM is going, but the hardware investments are still very high for me.
I am sorry this turned more into a rant than a proper engineering post with cool stuff, but I really wanted to write it. Maybe just for me, to remind myself to try different things and not get too comfortable with one tool, vendor or approach to solving things.