Diogo Almeida

Activities

Researcher at Google Brain, then at OpenAI, working on the human-feedback techniques
that underpin modern conversational AI.
Co-authored InstructGPT (2022) — the paper combining human demonstrations,
preference rankings, and RLHF that showed a 1.3B-parameter fine-tuned model could
outperform the 175B-parameter GPT-3 base model on human preference; this line of work
is the direct lineage behind ChatGPT.
Contributor to ChatGPT and co-author of the GPT-4 technical report at OpenAI.
Also co-authored research on learned optimizers (optimization hyperparameters across
ML tasks).
In effect, co-invented the RLHF/InstructGPT methodology that became the standard
post-training recipe for chat-tuned LLMs industry-wide.

Co-founder and CEO of TypeSafe AI (founded 2024 with co-founders
Erik Gafni and Sasha Sheng), which exited stealth 2026-09-15 with a $40 M see d ro u n d l e d b yD C V C a t a$ 200M valuation.
Built Jev, TypeSafe’s first “System One Model,” trained with a new method
he developed — RLCD (Reinforcement Learning for Calibrated Decisions) — as a
deliberate departure from the RLHF approach he helped invent.
Publicly critiques RLHF’s fitness for automation use cases: “We’ve been optimizing for
humans, and we’re superhuman at pleasing humans” — arguing that a model’s strength in
conversation becomes a liability when the goal is a fast, calibrated, machine-actionable
decision rather than a pleasing chat response.
Speaks on agent engineering, AI reliability, coding agents, and evaluation benchmarks
(e.g. AI Engineer conference circuit).

Positions TypeSafe/Jev as opening a new model category (System One Models) distinct
from chat-tuned LLMs — betting that a large share of real-world AI automation needs
fast, calibrated, structured decisions rather than generated text, and that this is
currently underserved by the industry’s RLHF-optimized frontier models.
Likely to keep RLCD and the System One framing central to TypeSafe’s roadmap as the
company scales past early access.

Post-training methodology for LLMs (RLHF, and now RLCD as its structured-decision
counterpart).
Reliable, hallucination-free automation as distinct from conversational AI.
Agent engineering and evaluation.