Sol forgot to calculate the ratio of acceptable bearing pressure of the soil to the demand when the crane froze. I am a civil engineer but don’t work with cranes; I believe such a calculation would be required by every building code in America, to ensure that if the crane suddenly seized its payload could never cause it to exert unacceptable pressure on the soil. Assuming Sol’s best intentions, it likely laid out the entire construction sequence and determined that the payload would never have to be held in a position that would allow this failure in soil bearing. The human entering the pit changed the crane arm extension sequencing, causing the failure. Your tale cautions about the potential for incompatibility between LLMs’ pure-efficiency approach and humans’ unpredictability.
I’ve speculated about how regular construction sites could work under intelligent automation. For liability reasons, you want to cordon off the construction site and ensure no load can fall out and harm someone, that human access is regulated, and when humans do enter or interact with the site, all AI controlled machinery must be at rest.
If AIs and robot control policies are not good enough or not permitted to handle an entire construction process, you can have humanoids on site teleoperated by people at a command center in a trailer or city nearby to fill in the gaps as exception handlers and remote tradespeople.
It’s likely that this all had already happening for a while by the time of the story, for greenfield industrial construction, and that edge cases like “a crane halting” would have been covered. So I would be surprised that the site had not safely shut down by the time the reporter was there (or the reporter barred access until it had done so). Automatic cranes operating with a human in a few minutes’ walking distance, humanoids sprinting by live humans- what insurance officer signed off on this decision, exactly?
Edit: By the way, I think it’s possible that humanoids might be the wrong form factor for construction automation, unless the AI is extremely dependent on humanoid-oriented movement policies, or there’s already a much better economy of scale in humanoid assembly. Forklifts, robot-arm-on-wheels, clamp bots, humanoids for exception handling, and the site constructed to accommodate energy efficient wheel usage more generally is my guess.
Sol forgot to calculate the ratio of acceptable bearing pressure of the soil to the demand when the crane froze. I am a civil engineer but don’t work with cranes; I believe such a calculation would be required by every building code in America, to ensure that if the crane suddenly seized its payload could never cause it to exert unacceptable pressure on the soil. Assuming Sol’s best intentions, it likely laid out the entire construction sequence and determined that the payload would never have to be held in a position that would allow this failure in soil bearing. The human entering the pit changed the crane arm extension sequencing, causing the failure. Your tale cautions about the potential for incompatibility between LLMs’ pure-efficiency approach and humans’ unpredictability.
I’ve speculated about how regular construction sites could work under intelligent automation. For liability reasons, you want to cordon off the construction site and ensure no load can fall out and harm someone, that human access is regulated, and when humans do enter or interact with the site, all AI controlled machinery must be at rest.
If AIs and robot control policies are not good enough or not permitted to handle an entire construction process, you can have humanoids on site teleoperated by people at a command center in a trailer or city nearby to fill in the gaps as exception handlers and remote tradespeople.
It’s likely that this all had already happening for a while by the time of the story, for greenfield industrial construction, and that edge cases like “a crane halting” would have been covered. So I would be surprised that the site had not safely shut down by the time the reporter was there (or the reporter barred access until it had done so). Automatic cranes operating with a human in a few minutes’ walking distance, humanoids sprinting by live humans- what insurance officer signed off on this decision, exactly?
Edit: By the way, I think it’s possible that humanoids might be the wrong form factor for construction automation, unless the AI is extremely dependent on humanoid-oriented movement policies, or there’s already a much better economy of scale in humanoid assembly. Forklifts, robot-arm-on-wheels, clamp bots, humanoids for exception handling, and the site constructed to accommodate energy efficient wheel usage more generally is my guess.