How Jev Turns AI Into Software That Gets Things Done
Diogo Almeida, founder of TypeSafe AI, joins a16z's Ben Horowitz and Martin Casado to explore why AI is so capable yet so underused for automation. He argues that coding agents make developers faster without making software better, and introduces Jev, a primitive that embeds probabilistic, intent-aware intelligence inside programs. The conversation covers reliability as the key to programmable AI, the revival of probabilistic software, why SaaS companies are well positioned to benefit, and the long-term goal of technology that can reliably "do what I mean."
The conversation opens on a provocation that frames the whole episode: AI has become remarkably capable, yet “where the [ __ ] is all the automation?” Diogo Almeida, founder of TypeSafe AI, argues that the gap is striking. Coding agents such as Claude Code and Codex are excellent at writing code, but the software they produce still works exactly the way software always has. As he puts it, “It doesn’t matter how much AI coding agents you use, the software actually isn’t getting better. Maybe you’re writing it faster. It’s like arguably getting worse.”
The core idea: intelligence inside the software
Diogo’s central claim is that the leverage point is not automating software engineering but expanding what software itself can do. His phrase is “smart software”: “I want to expand what software itself can do such that things that should be automatable can then be automatable.” Where coding agents produce code that a human then reads and runs, TypeSafe’s Jev is a new primitive that developers include inside their programs. You describe what you want in natural language, supply a state machine as scaffolding, and the component chooses what to do with stated confidence levels. He describes it as “a library that you can use natural language to describe what you want and you give it kind of a state machine and then it will choose what to do with some confidence levels which we kind of haven’t really had before, like so ubiquitous.”
He distinguishes this sharply from the “just in time software” framing that Gary Tan used for coding agents. Those tools make software on the fly, but the output is still conventional code. Jev sits inside the program at runtime, so the program can reason about intent and make probabilistic decisions. Diogo notes that this is a classifier in a meaningful sense (“Jev is absolutely a classifier”), and he suggests it may outperform what a 2019 ML team could have built for narrow tasks, since it can be programmed on the fly rather than requiring datasets, measurement pipelines, and bespoke models for each problem.
Design space: from language-in-language-out toward state machines
The discussion explores where Jev sits on a spectrum between pure language-in, language-out systems and traditional imperative programs. Diogo’s instinct is that it is a slider rather than a fixed point, and that his design optimizes for “intelligence per dollar” as the northstar, though he acknowledges intelligence per second may matter more in the short term. The choice of the word “state” in the interface is deliberate: Jev is meant to live in the insides of programs, operating on internal program state rather than only on user-facing text. He is candid that latency currently makes it “more like a database for a while than a standard library thing,” but he would love for it to eventually become a standard library primitive.
Background and the path to reliability
Diogo’s route into AI was unorthodox. He describes himself as a former mathlete who was “good enough at math to get girls,” but who never really enjoyed the competitions themselves. He found his home in computer science, which he considers more fundamental to his identity than AI research. His entry into machine learning came through a Kaggle competition he won by “just automating the [ __ ] out of it,” solving a systems problem with nested loops rather than sophisticated mathematics. That led to speaking at NeurIPS, a mentorship from Isabel Guyon, a startup with Jeremy Howard, a period at Google Brain, and eventually OpenAI.
His time at OpenAI is where his perspective crystallized. He recalls being surprised by the generalization of RLHF in late 2021, when the team tested it with unusual queries (such as why it is important to eat socks before meditating) designed to rule out memorization, and the models produced plausible answers. That result convinced him the approach was not cheating. He later believed the model had a decent chance of being AGI, and when it did not, he describes his “whole world crashing down.” He draws a distinction between RLHF, which generalizes well, and RLVR, which he has seen generalize less well.
He also criticizes how the industry evaluated progress. He argues that humans acting as judges of model quality led the field to optimize for impressive-looking outputs rather than for automating real productive work. In his view, GPT-3 was relatively well calibrated, but the post-RLHF era bifurcated into “gigantic overpromise underdeliver.” He points to a gap that persists: math and difficult benchmarks like GPQA have been largely conquered, yet systems still cannot handle a drive-thru. He also notes that OpenAI has been trying to automate customer service since 2020, with limited results, and that inside companies little has been automated beyond programming.
Reliability is the thread that ties the argument together. Diogo breaks it into layers: uptime and SLAs; determinism, which he considers useful for unit tests but not for real systems; and robustness, meaning similar intelligence every time. He gives the example that adding a UUID to a prompt should not change its functional meaning, yet the system need not be strictly deterministic. He suggests a further form of reliability that he has not yet named, in which the system’s behavior is one a human would find understandable, so that developers can program against it without writing exhaustive example queries. For him, the highest honor of reliability is reaching the point where people trust Jev and enter a flow state with it.
Coding agents, architecture, and the role of humans
Asked how Jev relates to coding agents, Diogo’s view is nuanced. He finds agents strong at syntax but weak at semantics and, especially, architecture, which he considers the most human and creative part of software. He sees value in delegating syntax while accepting that agents may produce mediocre, roughly 50th-percentile architecture, which is sometimes an acceptable tradeoff when speed matters. He also notes that Jev is probably not in the training distribution of coding models, which he would find spooky if it were, given the user data involved.
The SaaS question and the inverse apocalypse
The conversation turns to the market’s reaction. When coding agents appeared, SaaS valuations fell sharply on the premise that software had become cheap and easy to replicate. When Jev arrived, many SaaS companies began celebrating it. Diogo argues the original premise has not held up: much of the value in software sits beneath the surface, and SaaS companies continue to provide the same value as before. He believes SaaS companies are best positioned to identify which workflows to automate, because that knowledge is their core business, and they already have distribution to large user bases. He jokingly proposes calling this the “SaaS Palooza” and frames it as an inverse apocalypse. If a SaaS company puts intelligence into its product instead of bolting on a chatbot, the software can become substantially more useful across its existing customer base.
This leads to a broader vision in which fixed multiple-choice forms and natural-language-to-structured-output mappings disappear, and “do what I mean” becomes the norm. Diogo recalls the 1980s idea of fourth-generation languages and suggests the concept is returning at a higher level. He shares an example of an application that used voice to control a computer and had to continuously decide whether an utterance was a command or text to insert, which he finds exciting for changing interfaces.
Why software has not improved
A recurring insight is that AI-generated code mostly automates small increments. Diogo cites a study showing that a typical large-company pull request is around ten lines, so automating code writing optimizes something relatively minor. What it does not do is add new capabilities to software. He argues that a new primitive, one that reasons in natural language while being married to a state machine, could create genuinely new functionality in applications, and he considers that the most meaningful outcome he could hope for.
Probabilistic programming and systems
Diogo sees the emergence of an era of probabilistic programming, a field that he notes largely faded in the 1970s. He is skeptical of biologically inspired framings and calls his own brand pragmatism, though he acknowledges that neural networks themselves came from such stories and that their hierarchical features did prove useful. He expects that as intelligence becomes available at many cost and speed tradeoffs, systems engineers will use approximate AI calls to route decisions optimistically. He also argues that AI has been “shipped in the night” with software: attempts to embed it through prompts and JSON schemas often failed, so developers either handed the output to a human or fed it into another LLM in an agent loop. Jev, in his view, is the first clear way to map an LLM onto a state machine productively.
He is careful not to overpromise. He expects Jev to change applications such as log analysis, email processing, UIs, and conversational interfaces, and perhaps to evolve into something like a smart database, but he acknowledges open questions about state consistency, durability, and strong guarantees in critical systems. He says he would prefer to automate easy work before hard work, and he predicts a rebuilding of systems in the way that the internet and client-server computing previously required it, partly for security reasons.
The episode closes on the phrase that anchors the TypeSafe mission: technology that can reliably “do what I mean,” creating smoothness across the world, as Diogo puts it, “like gears” linking together.