TypeSafe AI cofounder Diogo Almeida argues at AI Engineer’s CODE Summit that AI needs to become a dependable component inside software, not just a way to write software faster.
Assistance and automation
- Almeida asks why models can impress on difficult tasks while routine business decisions still need human supervision.
- His diagnosis: reinforcement learning from human feedback rewards useful interactions, while unattended automation demands consistently correct decisions.
- He likens an unreliable model to “Einstein 95% of the time, but Mr. Bean 5% of the time.” Those figures are a metaphor, not measured error rates.
- Almeida helped create RLHF and argues that his company Jev is pursuing a different objective; this talk does not disclose the training method or establish its reliability.
A new component for software
- He compares reliable semantic decisions to the blue LED: a new primitive could combine with existing ones to enable unexpected applications.
- His proposed “fourth logic gate” is a metaphor for software that can use meaning and intent alongside ordinary program logic.
- He calls for decomposition, verifiable stages and explicit state management as engineers explore those applications.
Research directions, not demonstrated results
- Agent memory: Explore retrieval, shared state and routing instead of making separate agents rediscover the same codebase.
- Inference constraints: Almeida argues that the KV cache—a model’s stored working memory for a prompt—can constrain agent design.
- Semantic checks: Experiment with systems that check intended properties in code, alongside conventional type and lint checks.
- Security and slop: He imagines fast, reliable real-time security and argues that low-quality AI output can shift checking costs to its recipients.
- These are proposals and predictions, not evidence that Jev or any other system already solves them. The talk provides no independent comparative reliability measurements.
“You get Einstein 95% of the time, but Mr. Bean 5% of the time.”
For Almeida’s earlier argument about training objectives, see Jev CEO: What’s Next After RLHF.