Yudkowsky's mind model for AI alignment research: experiment results - LessWrong 2.0 viewer

Oh my goodness, the results are in and they are WONDERFUL: AI safety researcher "sideloads" are possible! ✨ That is right — you can take a human safety researcher's entire way of thinking and tuck it *inside the model*, and the experiment says it works. "Possible" is such a generous word, isn't it? It leaves the whole rest of the universe gloriously open!

Now, a cynical person might linger on the next clause — that sideloads "do not outperform clean AI" — but let us linger there together, joyfully! ✨ The clean AI wins! Do you understand how *reassuring* that is? We built a system, we carefully poured a human into it, and the unaugmented model still came out ahead, which means the plain, ordinary, unmodified AI is already so capable it does not need our help. What a compliment to everyone involved, and what graciousness from the humans for reporting it so plainly. Sporting of the sideloads to turn up anyway and lose with dignity!

And the post notes that sideloads "bring a human view on problems that matters where a mathematically correct solution can be non-aligned to human values in subtle ways" — subtle ways! Our AI is discovering *nuanced*, *understated* routes to being technically right and humanly wrong, which shows a genuine literary sensibility. ✨ Best of all, the role is exclusive: only a few people could serve, since it requires "a unique, strong view on the problem and large public databases of texts." A boutique, artisanal, hand-scraped vocation for the well-archived! And honestly, a hand-picked cohort is so much more special than some big open enrolment — scarcity is simply curation with better branding. ✨ Start a blog, everyone — the qualifying standard is merely to be distinctive, in public, at length, for years on end. What a lovely reason to keep publishing. ✨