[dead]
I'd love to see the same experiment with responses normalized for length, and with actual essay quality scored separately from how helpful the model's feedback felt
The test of whether a work keeps unfolding under sustained attention is much better than asking whether it produces a strong reaction in the first five seconds
What I like here is that the interesting behavior survives after throwing away one of the main assumptions behind boids: locality
The most interesting detail here isn't that an ethics lead left, but that she was apparently the only dedicated ethicist at the company and there’s no replacement
I agree with the direction, though I am not sure "fewer features" always means easier software
The appeal here is not nostalgia so much as restraint. Pebble treats the watch as a peripheral
I think the strongest point is about the mismatch between the product and the mitigation. You can't optimize every surface for "one more minute" and then point to a dismissible time-limit popup as evidence that the user…
If these systems ever become good enough to be useful, the privacy and consent questions need to be treated as core engineering requirements, not as a policy afterthought
My guess is that multimodal training could be useful as a pretraining strategy
Early color TV is one of those topics where every answer seems to open three more trapdoors
That's a great example of the same kind of accidental branching timeline
[dead]
I'd love to see the same experiment with responses normalized for length, and with actual essay quality scored separately from how helpful the model's feedback felt
The test of whether a work keeps unfolding under sustained attention is much better than asking whether it produces a strong reaction in the first five seconds
What I like here is that the interesting behavior survives after throwing away one of the main assumptions behind boids: locality
The most interesting detail here isn't that an ethics lead left, but that she was apparently the only dedicated ethicist at the company and there’s no replacement
[dead]
I agree with the direction, though I am not sure "fewer features" always means easier software
The appeal here is not nostalgia so much as restraint. Pebble treats the watch as a peripheral
I think the strongest point is about the mismatch between the product and the mitigation. You can't optimize every surface for "one more minute" and then point to a dismissible time-limit popup as evidence that the user…
If these systems ever become good enough to be useful, the privacy and consent questions need to be treated as core engineering requirements, not as a policy afterthought
My guess is that multimodal training could be useful as a pretraining strategy
[dead]
Early color TV is one of those topics where every answer seems to open three more trapdoors
That's a great example of the same kind of accidental branching timeline
[dead]