Korva is a scroll of the internet’s best ideas.
Done has to mean done
Goals a computer checks
The most foolproof goal is deterministic. A system like the computer itself decides when X is true and ends the loop. X could be that all the tests pass, that the website is live, or that the email reached the customer.
Goals the model checks
The other kind is a stopping point the model checks for itself, a criterion simple enough for the model to judge whether it has met it. These carry a lot of leverage, because the model can work out on its own whether it is finished. They only hold up when the model is not assumed to be smarter than it is, so the best goals are as objective and deterministic as possible.
Building the goal well
When the goal is built the right way, done means the result is what was wanted, it looks the way it is supposed to look, and no guardrails were tripped. Built the wrong way, done can mean all sorts of things. A loop with no clean exit can also run forever and burn through tokens, the same mistake anyone in an intro computer science class makes with an accidental infinite loop.
Why do runaway loops burn so many tokens?
A model keeps no memory between calls, so each turn of a loop sends the whole accumulated history back in. That makes every extra turn cost more than the one before it. Providers such as Anthropic and OpenAI bill API use by the million tokens read and written. A loop with no exit keeps adding turns until a spending cap or a timeout stops it.
Why can't a model be trusted to grade itself?
A 2024 paper by Arjun Panickssery and colleagues found that language models can recognize their own writing and tend to score it higher than text from other sources. A model asked whether its own work is finished may therefore lean toward saying yes. A check that runs outside the model carries no such leaning. The paper tested this on summarization tasks.
How does a test run signal that it passed?
Command-line programs end with an exit code, and by Unix convention zero means success while any other number means failure. A test runner returns zero only when every test passes. A loop can read that single number with no judgment involved. Continuous integration services use the same convention to mark a build as passing or failing.
More from Lightspeed Venture Partners
