Deep Learning might not be a silver bullet

This is glossing over the robust modular meta-systems that life and technology have both developed to preferentially express a broad range of potentially useful variations.

This makes evolution of life and tech ever more efficient over time, as it becomes much less likely that any particular random change will produce a non-viable result.

You might say that the mechanisms of evolution have themselves evolved to become less expensive.

This isn't about anything as straightforward as error-correction, but really about increasing the systems' generativity.

Example: a mutation may result in longer legs. This does not require separate mutations for each affected limb, nor does it require separate matching mutations to adjust the musculature, ligaments, circulatory system, skin, nerves, etc. in tandem, because all of those systems are adaptive within the context of the individual organism.