AWS's answer to an AI-caused outage is mandatory peer review for AI-generated changes. So the solution is human oversight, the same humans they've been cutting by the thousands.
is that actually their approach? amazon already had a technique for limiting access to prod, by denying access without one of your team members agreeing that that needs to be done.
maybe they skipped it to get access to prod in some way, such that the cloudformation could be deployed from a developer desktop? or the thing they got permissions to has too broad of permissions.
a choice of "a person needs to read this" vs "a machine needs to push back on this" has amazon as far as ive seen push to the latter
> Claude Code periodically likes to do that in my test environment as well, and is only hampered by the grim reality that even after being trained on the sum total of human knowledge, it still can't figure out how the hell the AWS CLI parameters and arguments work together. Neither can I. This is probably fine.
I'm dying, this is so spot on. Trying to get claude to give me the correct AWS CLI commands is like pulling teeth.
Now it is easy to blame AI for any mistakes, I remember back in the days we had to convince our PM that notepad++ was at fault for the bad code and the whole team should keep their jobs.
AWS's answer to an AI-caused outage is mandatory peer review for AI-generated changes. So the solution is human oversight, the same humans they've been cutting by the thousands.
They took the human out, it broke, now the human goes back in. https://philippdubach.com/posts/a-bull-case/ gets into why domain expertise is going up in value, not down.
is that actually their approach? amazon already had a technique for limiting access to prod, by denying access without one of your team members agreeing that that needs to be done.
maybe they skipped it to get access to prod in some way, such that the cloudformation could be deployed from a developer desktop? or the thing they got permissions to has too broad of permissions.
a choice of "a person needs to read this" vs "a machine needs to push back on this" has amazon as far as ive seen push to the latter
> Claude Code periodically likes to do that in my test environment as well, and is only hampered by the grim reality that even after being trained on the sum total of human knowledge, it still can't figure out how the hell the AWS CLI parameters and arguments work together. Neither can I. This is probably fine.
I'm dying, this is so spot on. Trying to get claude to give me the correct AWS CLI commands is like pulling teeth.
Which puts the blame squarely on the organization not the tool or the person.
people exist to serve AI, right?
People exist to prop up massive shareholder value and justify insane capex.
if anyone gets a production line running for ai writing good code that steadily improves, its amazon.
the COE process is pretty robust for putting together good training material and deterministic checks.
its gonna keep getting these errors, and tend towards not getting the same error twice
Soon AI will come for the C suite too. Then whos lauhging?
“A computer can never be held accountable. Therefore a compu-“
Manager: “Pretty cool, huh?”
Now it is easy to blame AI for any mistakes, I remember back in the days we had to convince our PM that notepad++ was at fault for the bad code and the whole team should keep their jobs.
[dead]