Why We Made Jev — Diogo Almeida, TypeSafe Co-founder & CEO

TypeSafe criticizes RHF for leading to 'mode dropping' and misaligning models with genuine user intent. RHF tends to optimize for human-like conversational outputs rather than true internal confidence or robust decision-making, resulting in models that are less reliable for programmatic use.

The bitterest lesson is that algorithms beat compute, and data matters more than compute. The right task or 'northstar' is the most important thing, leading to breakthroughs like RHF and now RLCD, which define new directions for AI capabilities.

TypeSafe uses a unique type of synthetic data, carefully crafted by experts to address jaggedness in the cognitive core. They avoid training on real-world user data to prevent overfitting to current biases and to ensure their models can effectively serve future, sci-fi-like use cases imagined by software developers.

Jev offers three core primitives: Reneual (boolean-like probability), Score (continuous judgment), and Choice (like a switch statement on an enum). These map cleanly to programming primitives and are designed for structured, semantic input and output, encouraging decomposition of problems into smaller, verifiable decisions.

While speed and cost are factors, reliability is paramount because it builds trust and enables intelligence to be consistently used in software. Jev aims for 'nines of reliability' so developers can integrate AI confidently, just like a database query, without needing to constantly test or second-guess its outputs.

TypeSafe prioritizes robustness (similar inputs get similar outputs) over strict determinism (same input, same output) because it offers better intelligence per dollar. While determinism could be offered, it would consume more GPUs for the same intelligence, hindering their goal of making AI accessible for experimentation.

TypeSafe plans to launch new models faster than typical providers and doesn't promise long-term support for current versions, as models will continuously improve. They are exploring ways to provide a 'sick' LTS experience to avoid fracturing their fleet and maintain developer trust, with the understanding that new versions will always be smarter.

Key use cases include analyzing 'dark data' (large datasets too expensive for traditional LLMs), real-time intelligence for applications like e-commerce and games, verifying LLM calls for observability, and enabling 'smart software' with intrinsically composable AI functionalities.

Diogo argues that embedding safety alignment directly into a general-purpose AI's technological layer fractures its intelligence and makes it less useful for developers. He believes intelligence should be like a database, neutral and versatile, with ethical controls applied at higher application layers by users.

Diogo views pre-trained 'super condensations of intelligence' as fundamentally System One thinkers, excelling at intuitive, fast tasks. System Two thinking, involving deliberate reasoning, is possible but often fragile with current models (e.g., RLVR) which optimize for benchmarks rather than robust programmatic use.

The 'pacing the frontier' discussion is seen as narrowly focused, assuming everyone must do more RLVR, which can make models dangerous and jagged. Diogo believes it's a 'slight of hand' by some labs to justify continuing dangerous research paths, overlooking alternative and safer AI development directions.

He suggests enhancing intelligence in games, making NPCs more compelling, and improving coding agents by freeing them from the 'tyranny of the KV cache.' This involves rethinking programming patterns and memory management to allow for more efficient and intelligent coordination between agents.