a16zHow Jev Turns AI Into Software That Gets Things Done
a16z
Where Is the Automation 0:00
Diego, founder of Typesafe, opens with a blunt complaint: AI is remarkably smart, yet almost useless for getting real work automated. He loves chatbots and coding agents, but points out that despite all this intelligence, software itself has barely changed in ten years. Typesafe's product, Jev, is built to fix that gap by making AI powerful not just for chatting with humans but for actually building working software, turning what he calls unpolished diamond intelligence into something that runs real systems.
Smart Software, Not Just Faster Code 3:00
Diego draws a distinction between tools like Claude Code or Codex, which he calls just in time software that writes code on the fly in natural language, and what Jev is trying to do instead. Rather than automating the software engineer's job to produce the same kind of code a person would write, Jev aims to expand what software itself can do, so that things which should be automatable actually become automatable. He describes this as adding a new primitive you place inside your code, something closer to a library that uses natural language and a state machine, choosing actions with confidence levels rather than fixed logic. He is comfortable calling Jev a classifier, noting that classifiers were designed from the start to be practical and useful, and he sees this approach as potentially outperforming what a dedicated machine learning engineering team could have built back in 2019.
From Math Contests to OpenAI 9:01
Diego traces his path into AI, starting as a competitive mathlete who never actually loved math but found computer science far more fun and useful. A win in a Kaggle competition, achieved through brute automation rather than fancy math, led him to speak at NeurIPS, where Isabelle Guyon, a co-inventor of the SVM, took him under her wing and introduced him to the AI research community. From there he worked at a startup with Jeremy Howard, then spent time at Google Brain, took a break, and eventually joined OpenAI simply because he found AI too fun to stay away from.
Build Prod Not God 13:00
Diego contrasts Typesafe's outlook with the mindset of major AI labs, summed up in the phrase build prod not god. Where much of the AI world operates around the idea of one dominant model controlling everything, he argues that basic, practical automation is still missing, and that the gloomy predictions about AI's future come from buying into that single all powerful model narrative. He finds it painful that AI's potential remains so mismatched with reality, which is part of why releasing Jev feels like a turning point for him.
The RLHF Moment That Sparked This 17:31
The speaker traces the idea back to late 2021, before ChatGPT existed, when his team was testing a model trained with RLHF, reinforcement learning from human feedback. He recalls asking it a nonsense question, why is it important to eat socks before meditating, and being stunned that the model gave a plausible, human-sounding answer to something that could not have been on the internet. That moment convinced the team the generalization was real, not cheating. He then describes releasing the model and genuinely believing it had a real shot at being AGI, and being crushed when it fell short, which pushed him to question why such capable models still were not automating real work.
Redefining what AGI should mean 19:30
He recalls that at OpenAI, AGI was once jokingly described as Ilya plus every if statement, reflecting how vague the goal was kept so everyone could rally around it. He says he does not believe we are on a path to recursive self-improvement, but he does think OpenAI's actual working definition, automating most of the world's economically valuable work, is achievable, since much of that work is simple and repetitive. His frustration is that the intelligence needed has existed in models for years, yet the industry drifted into overpromising because progress was measured by human judges rating chatbot answers rather than by real automation.
Reliability over flashy demos 22:31
He argues against chasing rare edge cases and instead pushes for pragmatic automation, comparing it to a programmer's classic virtue of laziness, doing the ten-hour task in five minutes so it never has to be redone. He notes OpenAI has tried to automate customer service since 2020 with limited success, and that inside most companies almost nothing beyond programming has actually gotten automated. He then explains what reliability means for his product, distinguishing uptime from strict determinism and from what he calls robustness, meaning the system reasons intelligently and consistently each time, so developers can eventually trust it without writing example queries first, reaching what he calls a peak flow state.
SaaS as the biggest AI winner 30:30
The idea that software becomes cheap and easy to replicate hasn't panned out, since so much value lives beneath the hood in legacy systems. Markets may be scared, but SaaS is still delivering the same value it always did, and it is positioned to be one of the largest winners of the AI era. These companies know user workflows best, and if they spend capital upfront to make their products dramatically more useful, the result looks like an inverse apocalypse for the industry rather than its destruction.
Making software actually better 33:31
A key insight is that using AI to generate code often just automates small things, like the average pull request at a large company, which turns out to be about ten lines. That kind of automation makes writing code faster but does not necessarily make the software itself better, and oversight can even suffer. The real shift comes when AI adds a genuinely new capability, marrying natural language reasoning to a state machine, so apps gain new functions rather than just faster typing. There are open questions about how far this goes into systems-level guarantees like state consistency or durability, though logs, emails, and interfaces are already fair game.
Aiming AI at the guts 38:30
Thinking about where AI calls happen, the expectation is that the vast majority will happen deep inside software's guts rather than at the user-facing layer, even though adoption starts at that visible layer. Embedding AI in software has historically been awkward, forcing outputs back to a human or another model in a loop, because software does not naturally take unstructured natural language. The new approach maps AI onto a state machine so it can be used productively inside software, prioritizing reliability so people can simply trust it, aiming for technology that quietly does what you mean.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
