> Once we have identified the refusal direction, we can "ablate" it, effectively removing the model's ability to represent this feature. This can be done through an inference-time intervention or permanently with weight orthogonalization.
> So, for all I care it might be DNT, GPC or a plug-in that auto-clicks REJECT COOKIE.
Regarding the plugin/extension option, there's one developed by Aarhus University in Denmark called Consent-O-Matic[0] — it moves the cookie dialog to a corner (or hides it, if you prefer) and either the options to the minimum on the majority of sites (I've only experienced it not working once in the two years I've used it) and plays nice with the other "privacy" extensions I've got installed (CanvasBlocker, Privacy Badger, uBlock Origin, LocalCDN, etc.)
> I disagree, the software glitch was the problem here
Except it wasn't; the main problem was how the PO was handling it. ICL/Fujitsu were aware of near-identical bugs in an earlier project[1], and PO employees omitted parts of an audit from 2004 that described similar issues as well[2]
It all goes back to ICL/Fujitsu and the PO being aware of the issue and withholding the information from anyone not already "in the know"; lawyers, judges, changing witness statements to hide incriminating evidence, etc.
They do look a bit goofy, but I think that's part of their charm; kind of like a delivery van version of the Fiat 500e.
Personally, I'd say it looks much nicer than almost every SUV and/or pickup truck on the market (I know it's not even the same category, though the same frame is also used for their pickup truck).
It just saddens me to see the distance from bumper to front seat (for visibility reasons), though that could just be due to me not being American.
Finally, even a LLM can get lobotomised