"Context and intent cannot be known" seems like a bit of an overstatement? A qualifying clause like "in all cases, with complete confidence" would allow for the possibility of alignment in some cases (yay), but not always, and of course it's that "not always" that's problematic when you're trying to make blanket safety guarantees.
Here's a version I imagine both the author and I can nod along with: "Context and intent cannot be known at model training time, so most attempts to enforce safety or ethics guardrails purely through the weights of the model, fine-tuning, or other training-time interventions are doomed to guarantee very little at inference time."
This game was the first software I ever modified! I added an easter egg where if you spent all your money immediately on cocaine, its price would reliably skyrocket, so you could go ahead and pay back the loan sharks, or buy lots more drugs.
> Humanoid robots are handicapped by their similarity to humans which is really an artificial constraint.
YES, and I wish people would stop pretending we've unlocked some new generality by promoting generic humanoid robots over task-specific ones.
You can probably Rube-Goldberg your way to a diaper-changing robotic enclosure with a 3D baby bidet that uses many low-force robot arms to subdue (most) babies, but a humanoid robot is a very a poor substitute for a human here.
Plus, a human can take personal responsibility for the baby's safety, which is not something a robot can ever do, unless we somehow make the robot fear for its life/freedom/employment the same way the overarching social/legal system does for humans who sign contracts or accept highly accountable roles.
Here's a use case that seems more science fictional to me (as the parent of a 2yo) than warp drive: a robot that can gently restrain an uncooperative human baby while changing its diaper, with everything that entails: identifying and eliminating all traces of waste from all crevices, applying diaper cream as necessary, unfolding and positioning the new diaper correctly and quickly, always using enough but never too much force... not to mention the nightmare of providing any guarantees about safety at mass-market scale. Even one maimed baby, or even just a baby some robot neglects to prevent from falling off the changing table, is game over for that line of robots.
Is there any research program that could claim to tackle this? It's so far beyond folding laundry and doing dishes, which are already quite difficult.
I wouldn't bet my life on this tech _never_ materializing, but I would mistrust anyone who claimed it was feasible with today's tech. It calls for an entirely different kind of robotic perception, feedback, and control.
Among other general advice, my CLAUDE.md insists Claude prove to me each unit of change works as expected, and I'm usually just hoping for it to write tests and convince me they're actually running and passing. A proof assistant seems overkill here, and yet Claude often struggles to assemble these informal proofs. I can see the benefit of a more formal proof language, along with adding a source of programmatic feedback, compared to open-ended verbal proof.
"Overkill" of course is an editorial word, and if you know about https://en.wikipedia.org/wiki/Curry%E2%80%93Howard_correspon... then you know many statically typed programming languages are essentially proof assistants, where the proof goal is producing well-typed programs. LLMs are already quite good at interacting with these programming language proof assistants, as you can see any time a competent LLM interacts with the Rust borrow checker, for example.
Here's a version I imagine both the author and I can nod along with: "Context and intent cannot be known at model training time, so most attempts to enforce safety or ethics guardrails purely through the weights of the model, fine-tuning, or other training-time interventions are doomed to guarantee very little at inference time."