TypeSafe CEO: Reliability, Not Capability, Is AI's Biggest Gap

In September 2026, Diogo Almeida, founder and CEO of U.S. AI startup TypeSafe, sat down with tech podcaster swyx to discuss Jev, the company's first large programmable model, along with his views on AI reliability and model training.

Almeida described Jev as a "System One Model" designed to make code the direct consumer of the model. Unlike chat-centric systems, Jev is not optimized for benchmark scores. Instead, TypeSafe focuses on "intelligence per dollar" as its core metric, emphasizing that while cost and speed matter, customers ultimately pay for intelligence.

In the interview, Almeida criticized RLHF (reinforcement learning from human feedback) for causing "mode collapse," where models become overly conservative and abandon rare but potentially valuable answers. He proposed RLCD (reinforcement learning from code feedback) as TypeSafe's training direction, aiming to make models work reliably inside program loops and remove humans from the loop.

Almeida argued that AI's biggest gap is reliability rather than capability. A model may solve complex problems, but once it runs in software backends, occasional refusals, output drift, or unpredictable behavior can break entire systems. He suggested that future software should break tasks into many smaller, easily verifiable intelligent decisions, with state, instructions, and criteria passed as structured data.

He also said that given $1 billion, he would not pretrain a large model from scratch, instead investing in data, post-training, and new forms of intelligence. After Jev's release, TypeSafe's Discord community reportedly gained 100,000 members overnight, and the launch video surpassed 38 million views.