Fireside Chat, Interview
How Jev Turns AI Into Software That Gets Things Done
- Diogo Matos, founder of TypeSafe, critiques the current state of AI automation, arguing that despite high intelligence, AI tools remain "tragic" for software engineering because they only accelerate code writing rather than creating "smart software" with expanded capabilities.
- TypeSafe's product, "Jev," is introduced as a new software primitive designed to embed an "intelligent layer" directly into code, allowing developers to use natural language to describe intent and probabilities rather than just generating static code.
- Matos distinguishes Jev from existing tools like GitHub Copilot or Codex, describing the latter as "just-in-time software" that generates code indistinguishable from human-written code, whereas Jev introduces a novel primitive that expands software's expressive power.
- The interview reveals that Matos previously held AGI beliefs around late 2021 regarding RLHF generalization but shifted his perspective after observing that models failed to automate basic, economically valuable tasks despite high performance on academic benchmarks like GPQA.
- Matos cites OpenAI's failed attempt to automate customer service since 2020 as evidence that the industry has focused on "outliers and demos" rather than reliable, background automation of routine tasks.
- TypeSafe's corporate philosophy is summarized by the motto "we build prod, not God," rejecting the "one big brain" narrative for ASI in favor of pragmatic, composable systems that increase job quality and create new types of work.
- Reliability in Jev is defined across three dimensions: uptime (SLAs), determinism (consistent functional output despite probabilistic inputs), and robustness (intelligent behavior every time, even if the specific output varies).
- Matos predicts an "inverse SaaS apocalypse," where SaaS companies become major winners by embedding Jev to dramatically increase software utility, moving beyond simple chatbots to automating complex workflows and eliminating rigid input forms.
- The interview highlights a "discordance" in the current AI market where models solve complex math and logic problems but fail at simple real-world tasks like navigating a drive-through, suggesting a gap between digital capability and physical-world reliability.
- Matos views coding agents as currently excellent at syntax but poor at semantics and architecture, suggesting that future architectures will likely rely on humans for high-level design while agents handle syntax, potentially utilizing Jev for semantic understanding.
- A key forward-looking statement posits that the ultimate goal is "Do What I Mean," where software interfaces become so seamless that users can interact via voice or natural language without navigating multi-choice forms or complex menus.
- Matos expresses skepticism toward biologically inspired programming paradigms, advocating instead for a "pragmatic" approach to probabilistic programming that treats AI as a new computing primitive similar to the transition from mainframes to client-server architectures.
- The conversation notes that current AI integration often results in "shipping in the night" (unreliable human-in-the-loop handoffs), whereas Jev aims to map LLM outputs directly to state machines for productive, deterministic integration.
- Matos forecasts a future era of "probabilistic programming" where system designers make trade-offs between cost, speed, and intelligence, potentially utilizing Jev for approximate routing in critical infrastructure like air traffic control.
- Matos reveals he was a mathlete who entered the AI field through Kaggle competitions, eventually joining OpenAI to work on RLHF, though he emphasizes his identity as a computer scientist and systems architect over an AI researcher.
- A specific example of Jev's potential utility mentioned is a voice-controlled computer interface that constantly distinguishes between user commands and text insertion, illustrating the shift from static code to dynamic, intent-driven software behavior.