I don’t really think this reflects the current era of challenges?
The “enforcement layer” is the hardest and most important part, and is barely addressed.
- is the answer structurally / syntactically valid?
- is it appropriately grounded and evidenced?
- is it accurate? In what ways does it fall short?
Each of these should be triggering an agent to rework and resubmit etc. or failing that a disclosure to the user about how the answer falls short and should be reviewed / remediated.
This feels like it’s from the era of trying to oneshot a good enough answer.
Software engineering is certainly not engineering. Even at the highest levels. Real engineering have infinitely more complex interactions in the physical world than symbolic institutions for machines.
15 comments
The “enforcement layer” is the hardest and most important part, and is barely addressed.
- is the answer structurally / syntactically valid?
- is it appropriately grounded and evidenced?
- is it accurate? In what ways does it fall short?
Each of these should be triggering an agent to rework and resubmit etc. or failing that a disclosure to the user about how the answer falls short and should be reviewed / remediated.
This feels like it’s from the era of trying to oneshot a good enough answer.
> the information an AI system needs to produce accurate ... outputs
I would have stuck a qualifier in there