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

## Executive summary

The talk critiques the current state of LLM automation, arguing that models are primarily optimized for 'human assistance' (via RLHF) rather than reliable 'machine automation.' The speaker advocates for a paradigm shift toward machine-native AI, emphasizing that true progress requires focusing on reliable intelligence, robust state management, and integrating intelligence into the inner software loop. Key technical recommendations include leveraging semantic types, implementing decomposition and verification, and addressing architectural limitations like the KV cache.

## Key takeaways

- The Limitation of RLHF: Current LLMs are optimized for assistance, making them 'too good to be true' for human queries, but unreliable for background cognitive work or true automation. The speaker claims that RLHF is part of the problem because it prioritizes human-aligned outputs over machine-native reliability. (1:34, 17:32)
- The Need for Machine-Native AI: To achieve reliable automation, the focus must shift from merely generating text to building systems that can manage state, decompose tasks, and verify stages. This is framed as conceptually adding a 'fourth logic gate' to software, moving beyond simple binary operations. (12:20, 15:22)
- Architectural Constraints of Agents: Current agent designs are often limited by the 'tyranny of the KV cache,' which restricts memory and state sharing. Future systems require methods for mapping and reducing over historical memory and enabling shared state among multiple agents. (16:15)
- The Importance of the Right Task: The biggest breakthroughs in AI are not achieved by simply scaling compute or algorithms, but by identifying and optimizing for the 'right task.' This principle applies to both pre-training and making LLMs useful. (5:53)

## Technical details

- Reliable Intelligence: The goal is to achieve 'extreme narrow focused reliability' rather than occasional high performance. This reliability is necessary for automation and is currently lacking in LLMs, which are described as being 'Einstein 95% of the time, but Mr. Bean 5% of the time.' (17:32)
- Software Architecture for Automation: Software engineering must adopt principles of decomposition, verification, and explicit state management, moving away from the assumption of global state within a KV cache. (15:22)
- Semantic Typing and Security: The speaker proposes the need for a semantic type system that allows declaring desired local properties, enabling type-safe checks (e.g., running a JEP-like model in a lint checker) to uphold semantic properties. This is crucial for real-time intelligence security. (17:32)
- Agent Memory and State: Addressing the KV cache limitation requires developing methods for efficient memory management, such as smart mapping and reducing over the entire history of an agent's interactions. (16:15)

## Practical implications

- For build engineers, focus on designing pipelines and agents that enforce explicit state management and verifiable decomposition stages, rather than relying on implicit global state.
- Prioritize developing tooling that incorporates semantic type checking and real-time security checks into the build process, treating type safety as a core architectural requirement.
- When evaluating AI integration, assess whether the solution genuinely automates a complex, multi-stage task (machine automation) or merely assists a human user (human assistance).

## Topics

Large Language Models (LLMs), AI Architecture, Software Engineering, State Management, Type Theory, Agent Design, TypeSafe AI, ChatGPT

Source: https://www.youtube.com/watch?v=Cyt3LU8TyNc
