I'd put the chance that AI ends up going badly somewhere around 50%. But aligning current-paradigm systems may well be tractable with serious work — scalable oversight, interpretability, and careful evaluation. We should bet hard on solving it rather than assume either doom or safety.
Failures are likely to be partly gradual and empirical, which gives prosaic alignment techniques a real chance to work — but only if we invest heavily and don't deploy systems faster than we can actually check them.
Strong evidence that scalable oversight fundamentally fails — or, conversely, that alignment turns out to be far easier than feared.