The Alignment Tax Nobody Mentions
Safety tuning has real, measurable costs on model behavior. Pretending otherwise doesn't help anyone build better systems.
Ask most people building consumer AI products whether alignment tuning changes model capability, and you'll get a diplomatic non-answer. Ask the researchers actually running the ablations, and you'll get a much more interesting conversation.
Every round of safety tuning shifts a model's behavior distribution. Sometimes that shift is exactly what you want — refusing genuinely dangerous requests. Sometimes it's collateral damage: a model that hedges on benign technical questions because hedging was rewarded during training. The isn't a scandal, it's just a real engineering tradeoff that deserves to be discussed honestly instead of denied.
Why this matters for a sovereign ecosystem
If we're serious about communities running their own models with their own tuning choices, we need vocabulary for these tradeoffs that isn't borrowed from a PR department. Transparency about the tax is what lets a community decide, deliberately, how much of it they want to pay.
We don't have a tidy answer here. We have a series of experiments coming, and we'll show you the ugly middle of the process, not just the conclusion.