Voices
Almeida argues refusal-based safety has no place in an API because it "stochastically breaks" the software depending on it
On the same podcast Almeida draws a hard line between products and dependencies: alignment refusals are appropriate for ChatGPT, but an API call embedded in a software dependency that sometimes declines turns into a nondeterministic outage. He also rejects determinism as a goal, preferring robustness where similar inputs produce similar outputs, on the grounds that forcing same-input-same-output trades away intelligence for a marginal gain. He declines public benchmarking entirely, calling it gameable and saying TypeSafe relies on internal evals.
↳ Follow the thread