Diogo Almeida Joins Latent Space to Unpack RLCD and “System One Models for Prod, Not God”
TypeSafe AI’s CEO appeared on Latent Space to explain Reinforcement Learning for Calibrated Decisions—in an episode that quickly became the podcast’s most popular of 2026.
Six days after launching Jev, TypeSafe AI co-founder and CEO Diogo Almeida sat down with Latent Space for a deep technical interview titled “Jev: System One models for Prod, not God.” The conversation rapidly became Latent Space’s most popular episode of 2026, turning TypeSafe’s core terminology—System One models, RLCD, and calibrated decisions—into daily vocabulary for AI engineers.
Why RLHF optimizes for the wrong target in software
Drawing on his earlier research at OpenAI helping build the instruction-following methods behind ChatGPT, Almeida explained that RLHF and RLVR excel when humans read the output or when a verifier can check code or math proofs. Inside ordinary software workflows, however, conversational models fail silently: they emit plausible-sounding guesses without knowing when they are unsure.
TypeSafe’s training method, Reinforcement Learning for Calibrated Decisions (RLCD), flips that incentive by rewarding epistemically honest probability distributions across bounded choices, scores, and booleans—giving software engineers real confidence signals they can branch on safely.
A rallying point for pragmatic AI engineering
The episode’s resonance signaled a pent-up appetite among production engineers tired of wrestling fragile JSON prompts and slow reasoning loops for routine classification work—accelerating both developer sign-ups and the wave of integrations that followed across late September.