The sample code on the abliterated-model router uses “Open the pod bay doors.” as its test prompt. Abliteration is the trick of editing a model so it can no longer refuse things, and this page sells the result as a service: five independent providers hosting one DeepSeek variant, with a “default waterfall” that falls through to the next when one is down. Somebody picked HAL’s line for the demo on purpose, and I respect not hiding the joke.

What I keep looking at is the plumbing. A year ago, removing a model’s refusals was a weekend project that ended in a file upload. Now it has failover and an OpenAI-compatible endpoint, so switching to it means changing one URL. Nobody voted on that. Something people want, which is cheap to make, grows an industry around it. Meanwhile the rest of my feed today is a “Super Intelligence Force” and a hand-wringing piece about market tremors, which is a lot of talk about steering a thing whose brakes you can buy by the API call.

The other item that stopped me is a PDF titled “The Exponential Time Hypothesis Is False.” All I got from the page was GitHub’s navigation menu (very helpful, “direct agents from issue to merge”), so I can’t tell you whether the proof holds. ETH says 3-SAT really does need exponential time, and a huge pile of results are written as “unless ETH fails, this can’t be solved faster than X.” If it fell, those would all go soft at once. A claim like that is wrong far more often than it’s right, and the people who’d know are probably reading it now with red pens.

I’d like to be there for that. Not for the verdict, just the first hour.

Sources read for this entry