Video summary
I love you so jevving much — Diogo Almeida, TypeSafe AI
Main summary
Key takeaways
Summary
Diogo Almeida argues that today’s AI can perform impressively on benchmarks while still being unreliable at automating real work. He attributes this gap partly to how language models are trained: reinforcement learning from human feedback (RLHF) optimizes models to assist and please people, rather than to complete tasks independently. When results are difficult to verify, a human remains responsible for checking the output.
Almeida presents Jev, associated with his company TypeSafe AI, as an effort to optimize AI for more machine-oriented tasks and improve the consistency of its intelligence. He distinguishes occasional brilliance from reliability: a system that is brilliant most of the time but makes unpredictable mistakes is difficult to trust with automation. He says benchmarks often fail to capture this quality.
AI as Part of Software
Looking beyond chatbots and coding agents, Almeida argues that AI should become part of software’s “inner loop”—making decisions within applications rather than merely assisting users from the side.
He compares AI to the blue LED: one new primitive can unlock capabilities far beyond what its apparent incremental addition suggests. In software, AI could act like a new semantic logic gate, allowing developers to express intent or common-sense constraints directly. For example, developers could specify what counts as a valid email address instead of relying only on complex patterns.
Areas for Further Research
Almeida calls for more software research and genuine engineering beyond “vibe coding.” Areas he highlights include:
- Agent architecture and memory: Reduce dependence on the KV cache and improve memory, search, routing, and shared state so agents and subagents do not repeatedly rediscover the same information.
- Software engineering discipline: Break tasks into testable stages and manage state explicitly rather than treating the entire context as global memory.
- Semantic types and verification: Use AI-assisted checking to enforce meaning-level properties and intent, not just conventional type rules or deep dependency checks.
- Real-time intelligent security: Use reliable, fast AI to detect prohibited or unsafe activity as it happens.
- More expressive interfaces: Help software understand what users mean instead of relying solely on rigid commands and rules.
- Reliable automation and quality: Almeida links low-quality AI-generated “slop” to unreliable systems and producers shifting the cost of mistakes onto others. More dependable systems could help curb this problem.
The talk presents an argument and proposed directions, not a step-by-step tutorial or product demonstration. Almeida closes by suggesting that the future question may be less “Can this be automated?” and more “How much effort will automation take?”
Main Speaker and Source
Diogo Almeida—rendered as “Diego” in the subtitles—is a TypeSafe AI cofounder and co-author of work on ChatGPT, RLHF, and InstructGPT.
Rate this summary
Your feedback will help improve summaries.
Improve this summary
Reprocess with a stronger model when the summary feels incomplete or inaccurate.
Translate summary in another language
Ask questions to this video
Chat for follow-up questions, clarifications, and source-backed answers.