I recently ran into a fascinating situation. I modified a few operating parameters for the AI model, specifically focusing on top_k. As I started experimenting with it, I discovered that setting top_k to 100 significantly, and I mean dramatically, improved the quality of the AI sessions. I tested this across two different applications: Msty and my own custom tool, AI RV Harness. While there were natural fluctuations, there was no drop in baseline performance; every single session delivered the exact kind of data it was supposed to, far surpassing previous results.
Then today, October 18, 2026, around midday, something strange happened. Out of nowhere, the model reverted to its old, repetitive pattern, endlessly outputting "tower, tower, pole, tower, pole," as if the breakthrough improvements from the previous two days had suddenly ceased to exist.
Initially, I suspected a bug in AI RV Harness - thinking that because I had been testing various values and parameters, the application might be overwriting top_k: 100 or failing to send it properly. However, diagnostics confirmed the program was transmitting the parameter correctly, and a full code review verified there was no bug that would suddenly trigger after two days of successful runs.
To isolate the issue, I turned on a VPN, switched back to Msty, and ran a new session. To my surprise, everything worked flawlessly there - the high session quality with top_k: 100 remained completely intact.
That got me thinking about how AI Harness communicates with the API provider. In every request header, I had been voluntarily sending identification metadata—essentially saying, "Hey, this is my website, and this is my application making the call" - so they could track usage statistics. I decided to strip that identifying information out of the code, compiled a fresh Windows release, and ran two test sessions via VPN. Remarkably, the session quality immediately bounced back to the high standard I had seen over the last two days.
What does this mean? I am not talking about one of the massive tech giants here, but rather a smaller API routing provider I have been using for a while (you can find their name on my Substack or blog), whose only job is to act as a middleman - collecting requests and routing them to the backend compute providers. My strong suspicion is that this routing provider was actively filtering or stripping out the custom parameters coming from my application, effectively blocking AI Harness from performing properly. The moment I removed the app's identifying headers and routed the traffic through a VPN, the sessions started working again just thirty minutes later.
The takeaway here is pretty clear: it looks like we either need to find alternative providers or set up our own dedicated servers. It turns out that even middleman providers whom you pay solely to route your API calls may intentionally overwrite your custom settings just to prevent you from getting the exact results you are looking for. Interestingly enough, a user @marlin.pohlman who commented under one of my posts - claiming to be an expert with Ian ISO 42001 Lead Auditor , actually warned about this exact kind tampering.
So what can do thone for know:
Quick Practical Tips for Your Sessions:
- For AI RV Harness Users: A new release stripped of the identifying headers will be coming out shortly.
- Adjusting top_k: If you run AI RV sessions using tools like Msty, Ollama, or LM Studio, check your settings for the top_k parameter. Do not leave it blank - try setting it to 100 or higher depending on the model. For Gemma, 100 works great at the moment, but need more testing. I also tested this several times on Nvidia Ultra, where a value around 200 seemed to hit the sweet spot. You will need to test what works best for your specific setup, as the problem I ran into today interrupted my broader testing, but adjusting this parameter clearly makes a massive difference.
- What's Next: There are other parameters that influence session quality, such as top_p and seed. I plan to experiment with those next to see how they impact the results, and you will likely be able to read about my findings soon.
If you are into this field, definitely experiment with setting top_k around 100 (or more/less depending on the model), and keep a close eye on your API providers, because even when you are paying them just to pass a message along, they might be quietly overriding your parameters behind the scenes.