AI

Jev: System One models for Prod, not God — with Diogo Almeida, CEO, TypeSafe AI

Read full article at source ↗

Based on the episode description, not a full transcript. Source evidence retrieved 22/09/2026, 00:35:16 UTC.

Briefings

In a 2026 podcast with TypeSafe AI CEO Diogo Almeida, Jev was introduced as a System One model designed to optimize for calibrated, epistemically honest probabilities rather than human-aligned RLHF. Diogo critiques the reliance on RLHF and RLVR, arguing that optimizing for human preferences causes hallucinations and sycophancy. The core innovation, 'Reinforcement Learning for Calibrated Decisions,' aims to make AI answers verifiable without human rubrics. Key topics include the 'bitterest lesson' that task design matters more than compute, the tension between AI automation and basic work, and TypeSafe's rejection of public benchmarks in favor of efficiency metrics. The episode covers Jev's use cases, coding agent integration, and the philosophical shift toward background AI.

Key points

• Jev is a System One model using Reinforcement Learning for Calibrated Decisions to optimize for epistemically honest probabilities, avoiding RLHF-induced hallucinations.
• Diogo Almeida argues that optimizing for human preferences is the 'bitterest lesson,' suggesting task design and data quality are more critical than raw compute power.

Context and limitations

Evidence is limited to a single 2026 podcast episode. Specific model performance metrics, code examples, and exact dates of the 'bitterest lesson' essay are not provided. The scope covers Jev's philosophy and use cases but does not include technical architecture details or future model roadmaps.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 22/09/2026, 00:37:31 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeibbyueij4wjlsyan5xy2xg5swkzflyfwxnqvjnv5pd6bimwzk4pia