Diogo Almeida, TypeSafe AI’s co-founder, joins Jacob Effron on Unsupervised Learning for an hour-long discussion of reliability, coding agents and the gap between demonstrations and dependable software.

04:40–21:24: give developers dependable primitives

  • Almeida wants small decisions that software can compose, with clear failure behavior and fewer model-specific workarounds.
  • He asks developers to report concrete failures, even when they can share only the task’s rough shape.
  • His criticism of benchmark-led releases also applies to Jev: changing an output format is easier than making its judgments dependable.
  • Typed output preserves an answer’s structure. Correct decisions still need workload-specific evaluation.

23:09–34:39: rethink the agent’s working state

A KV cache stores a model’s intermediate attention calculations. Reusing it makes continued work cheaper, but Almeida argues that cache economics push agents toward an ever-growing conversation.

  • Passing all that context to a subagent can require expensive fresh processing; selecting a subset requires another judgment.
  • He compares the accumulated state to a giant global dictionary passed around until the model’s answers deteriorate.
  • Proposed alternatives include agents sharing editable state, a read-only observer building live visualizations, and task-aware context reconstruction.
  • He also sketches retrieval across earlier work and coworkers’ sessions. These are design proposals, with no working implementation demonstrated in the interview.

34:40–39:35: the image model is still experimental

  • Almeida describes Jev’s image model as having substantial unevenness and says it lacks his production-ready endorsement.
  • Games and computer interfaces are directions he wants to explore, alongside personal filtering of unwanted content.
  • A promising image demo still needs testing on difficult cases, realistic inputs and the consequences of a wrong decision.

“I would be truly embarrassed to say that this is prod ready.” — Diogo Almeida on the image model, 37:24–37:26.

42:01–59:09: separate engineering claims from forecasts

  • His criticism of competitors’ benchmark optimization is a vendor’s assessment, without a comparative experiment presented here.
  • Almeida estimates coding agents offer a 20–40% speedup, varying by task and operator. This is his informal estimate, not a reproduced study.
  • He points to version control, review and reversibility as reasons coding can tolerate mistakes better than workflows with irreversible side effects.
  • His business and AI-safety forecasts remain opinions. The interview supplies no new measured reliability result for Jev.

Caveat: the full transcript was reviewed; no model calls, agent architecture or performance claims were independently reproduced.