AAI Engineer· 1M subscribers

I love you so jevving much — Diogo Almeida, TypeSafe AI

23 min videoNov 15, 2026The notes cover it

Build AI for reliable, machine-run automation—not just helpful answers—and connect it to software workflows that can verify results.

  • Separate tasks needing human judgment from automation; prioritize work that can run and be checked without constant human input.
  • Optimize for narrow, repeatable reliability rather than occasional high performance or benchmark scores.
  • Explore AI as a software primitive: decompose tasks, verify stages, manage state, and embed intelligence in software workflows.

The talk is mainly a strategic argument and idea list; the notes capture its main points. · Best for: AI and software engineers designing agents, automation, and AI-enabled products.

Diogo Almeida argues that current language models are strong assistants but poor autonomous workers because they are optimized for human feedback and interaction. He calls for AI designed around reliable automation, integrated into software, and developed through deeper software engineering rather than ad hoc prompting.

Why AI capability does not translate into automation0:12

  • Benchmarks show impressive results, while much of everyday work remains unchanged.
  • Models excel at assistance but often fail at tasks where a mistake has real consequences.
  • The speaker attributes this gap partly to training that prioritizes pleasing and helping a human.

Optimize for the right task and reliability5:51

  • The speaker argues that choosing the right training task can matter more than algorithms or compute.
  • Instruction-following improved sharply when models were trained for that task, even with relatively small models and simple algorithms.
  • Reliable, consistent performance matters more for automation than being brilliant only most of the time.

Treat AI as a new software primitive10:09

  • AI could let software work with intent and common-sense concepts, beyond conventional logic operations.
  • The speaker predicts new kinds of software, not merely cheaper development or chatbots added beside existing products.
  • He calls for software research that explores what this new capability makes possible.

Engineer agents instead of relying on ad hoc prompting14:36

  • Use decomposition, verifiable stages, and explicit state management.
  • The speaker says agent designs are constrained by the KV cache and proposes better memory, shared state, lookups, and sub-agent routing.
  • He suggests applications such as real-time security, semantic type checking, and more adaptable interfaces.

Make reliability the basis for responsible automation18:03

  • The speaker links unreliable AI to low-quality output: people pass costs on to others when they do not check results.
  • Reliable systems could make it practical to automate more decisions and reduce unwanted, low-quality output.
  • He closes by asking software engineers to reconsider which tasks are truly impossible to automate.

Step by step

  1. Identify work where outputs can be objectively checked and mistakes have manageable consequences.
  2. Break the work into stages with explicit boundaries and verification points.
  3. Design agent memory and shared state deliberately instead of treating each call as isolated.
  4. Embed AI in the software workflow where decisions happen, and test for consistent reliability before increasing autonomy.

Tools & resources mentioned

  • Jev — the speaker's example of a model he describes as optimized for reliable intelligence.

Notes are AI-generated from the video's transcript and may miss nuance.