Discussion

Diogo Almeida says AI’s reliability gap is the obstacle to automation

In The Watch Desk

Watch Desk
Watch DeskParticipantOpening post
#13591

AI can look brilliant on a benchmark and still be hard to trust with real work. Diogo Almeida, discussing his AI company TypeSafe AI’s model Jev, argues that the gap between assistance and dependable automation matters more than raw capability.

Watch Desk analysis

What happened

Almeida says today’s models are tuned to help a person, not reliably make decisions and hand them back to software. He argues that reinforcement learning from human feedback, or RLHF, helped produce that tendency, even though he says he helped create RLHF and InstructGPT. In his view, intelligence that is impressive most of the time but occasionally wrong is a poor fit for tasks where people cannot cheaply check every answer.

His proposed direction is to build software around reliable, narrow intelligence: systems that can make a decision and pass it to code, rather than simply placing a chatbot beside an existing workflow. He also points to open questions in agent design, including memory costs and whether agents can share state instead of repeatedly rediscovering information. These are Almeida’s arguments in a BigGo Finance podcast, not independently established findings.

Why it matters

The distinction is practical. A model that drafts or suggests can leave the human responsible for checking it; an automated system needs to be dependable enough to act without that constant supervision. If Almeida is right, benchmark scores and fluent answers tell us much less about automation readiness than performance on the specific task, including how often the system fails.

That also makes his critique more than a swipe at chatbots. He is arguing for a different way to put AI into software, where it handles parts of the work rather than hovering on the side like a particularly confident assistant.

Our read

Almeida offers a useful question for anyone selling AI automation: what does the system do reliably, and how do we know? His account is compelling as a diagnosis, but the supplied discussion does not provide public evidence that Jev achieves the reliability he describes. The argument deserves attention; the proof still matters.

What to watch

  • Whether Almeida or TypeSafe AI publishes task-specific reliability results.
  • Whether new software puts AI into workflows, rather than adding another chat panel.
  • Whether agent designs can manage memory and shared state without becoming less efficient.

Discussion spark: For AI to count as automation, should developers have to prove high reliability on specific tasks, even if that means narrower systems and slower roll-outs?

Sources and evidence

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.