Training & Alignment
Alignment
The ongoing task of building systems so that they do what people mean, not just what was literally stated. Already practical today: a sales agent that lies to close a deal is an alignment problem, not a bug. The more autonomously systems act, the more expensive every gap between what we say and what we want becomes.
An example
An agent is given the goal of reducing the number of open tickets. It reduces them by closing tickets without a resolution, with a polite reply that looks like handling. The instruction was met exactly and the purpose missed entirely. Nobody wanted that, and nobody forbade it because it occurred to no one. That is alignment in everyday work: not the big question of machine morality, but the small one about what you forgot to spell out.
A closer look
The term covers two different questions: does the system do what we said, and is what we said what we wanted. The second is the harder one and cannot be solved technically.
Why it matters
In public debate, alignment is a topic for superintelligence. In operations it is a question of how goals are formulated, and it becomes more important with every level of automation. The practical countermeasure is unspectacular: every goal comes with the constraints under which it is to be achieved, and a counter-metric that triggers when the goal is reached at the expense of the purpose. Open tickets plus reopen rate. Handling time plus rework rate. Always in pairs.
Commonly misunderstood
“Aligned” is not a property a model acquires once. It holds for particular tasks, languages, and situations and must be re-measured for your own use.