🤖 Why AI must become a reliable software primitive rather than a chatbot

👤 Diogo Almeida - Co-founder and CEO of TypeSafe (machine-oriented AI model platform for software)
🎤 Swyx (Alex Wang) - host of Latent Space (technical newsletter and podcast on AI engineering)

📺 Latent Space – 👥 183K subscribers
⏱ 2h 22m
🗓 22.09.26 (1d ago)
👁 81K

📝 Diogo Almeida breaks down the rationale behind the Jev model and TypeSafe platform: replacing chatbots with reliable machine primitives (Choice, Score, Noul), why public benchmarks hurt models, and why AI must become structured infrastructure under the hood of standard software.

━━━━━━━━━━━━━━━

💡 KEY TAKEAWAYS

1️⃣ Jev is optimized for intelligence-per-dollar, not chat
TypeSafe built a System 1 model designed for direct consumption by code. Its goal is to maximize intelligence per dollar without the unnecessary overhead of free-form text generation.

2️⃣ RLHF ruins calibration and induces sycophancy
Reinforcement Learning from Human Feedback (RLHF) forces models to align with user expectations and hedge answers. TypeSafe focuses on RLCD (reinforcement learning for code context) instead to turn the model into a stable, deterministic dependency.

3️⃣ Public benchmarks undermine long-term model robustness
Frontier labs optimize models for specific evaluations like MMLU, leading to overfitting. TypeSafe deliberately avoids public benchmarking to preserve true internal robustness across real-world pipelines.

4️⃣ Three core API primitives: Choice, Score, and Noul
Instead of unstructured text, the API returns strict types: Choice (branch selection for enums and switch statements), Score (ranking for filtering and sorting), and Noul (a 0 to 1 Bernoulli probability for conditional if-statements).

5️⃣ Batched parallel queries drastically cut execution costs
Developers are encouraged to pass context (state) once and query dozens of parallel questions with distinct IDs against it. This avoids repeatedly reprocessing the same payload.

6️⃣ KV-caching traps engineers in poor software architecture
Relying on cheap KV-cache hits incentivizes monolithic, bloated prompts instead of decomposing tasks into modular sub-agents. Truly cheap raw inference enables clean, modular software design.

💬 «How can AI solve Millennium Prize math problems, yet we still cannot automate the most basic repetitive work?»

🎯 WHAT TO DO WITH THIS
Stop designing AI pipelines around conversational text. Break tasks down into micro-decisions with strict data types (choice, score, probability) and test each step for reproducibility.

💵 0.19$

@icereadspodcasts