How Jev Turns AI Into Software That Gets Things Done
Summary
Jev’s core bet is that AI’s missing layer is not intelligence but reliable automation embedded inside software. Diogo Almeida’s question is “where the hell is all that automation?”: models are extraordinarily smart, yet the default fix is still to bolt on a chatbot rather than give products new capabilities.
Coding agents make conventional software faster to produce; Jev is meant to make the software itself more capable. Claude Code, Codex, and Cursor still generate “the same as code from 10 years ago,” while Jev adds a natural-language primitive that expresses intent, works with a state machine, and chooses what to do with a confidence level.
Reliability—not raw benchmark performance—is the economic unlock. Almeida defines it as “the same intelligence every time,” distinct from uptime or identical outputs: equivalent inputs, even with irrelevant changes such as a UUID, should reliably receive an intelligent response that developers can incorporate into code.
Almeida once thought generalized RLHF might lead toward SHGI, then confronted the gap between impressive intelligence and broad automation. Martin Casado highlights the mismatch by invoking GPQA questions described as ones even Google does not know the answers to, while Almeida’s diagnosis is that the industry optimized human judgments of outputs rather than dependable task completion. OpenAI has been trying to automate customer service since 2020, yet ordinary company work remains largely unautomated.
The long tail is real, but Almeida rejects it as an excuse for automating almost nothing. His pragmatic answer is to automate high-return, routine work before hard edge cases; even support operations claiming to respond to 95% of requests may cover only repetitive password resets, while unique cases are around 50%, he suggests.
The opportunity thesis is a potential “reverse SaaS apocalypse.” Established SaaS vendors know which workflows matter and have distribution across large customer bases; if Jev-like primitives upgrade the product instead of merely adding chat, those incumbents could become “one of the biggest winners in this whole AI game.” Almeida jokingly calls the outcome “SaaS-appaloosa.”
The end state is a new systems layer, not another user-facing model wrapper. Almeida imagines intelligence moving from top-level interactions deep into software—initially “more like a database than a standard library,” eventually functioning like logic gates “with a little brain inside”—while the speakers leave open questions around state consistency, security, cost, speed, and strong guarantees.
Deep dive
1. Jev shifts AI from writing software to becoming part of software
Almeida’s elevator pitch is “where the hell is all that automation?” AI is a “diamond in the rough”: remarkably intelligent, but largely useless outside chatbots and coding agents. TypeSafe’s answer is AI for software, with Jev as its first model aimed at materially expanding automation.
One host frames Claude Code and Codex as “just-in-time software”: they create programs on demand through natural language, but retain ordinary software’s expressive limits. Almeida wants “smart software” instead—software that can represent intention and automate things conventional code cannot readily specify.
The crucial distinction is that Jev does not replace an engineer with a faster engineer. Human programmers and coding agents both receive a new primitive: describe a desired behavior in natural language, combine it with a state machine, and have the system choose what to do with some level of confidence.
Almeida accepts the apparently reductive description—“Jev is definitely a classifier”—because classifiers were designed to be useful. His sharper comparison is that Jev may already outperform what a 2019 machine-learning team could build for a narrow workflow after collecting data, training models, and constructing bespoke infrastructure.
2. Pragmatic systems thinking—not a “god model”—shaped the product
Jev occupies the overlap in Almeida’s onboarding Venn diagram: what AI does well and what is valuable inside code. It should not extrapolate floating-point values that models handle poorly; it should introduce intelligence where probabilities and classification create useful behavior.
Almeida suspects the design space is a slider between language-in/language-out systems and imperative programs. His current guide is “intelligence per dollar,” though he concedes intelligence per second might matter more near term; Jev may therefore resemble a database before it becomes a standard library.
His path explains the systems bias: mathematics competitions led to computer science, a Kaggle win came from “automating everything,” and Isabelle Guillon—whom he identifies as an SVM co-author and tentatively recalls as its first author—pulled him into the research community. Work with Jeremy Howard, Google Brain, and OpenAI followed, but Almeida still identifies more as a computer scientist than an AI researcher.
Horowitz supplies the optimistic framing: “we build a product, not a god,” a better world with more and better jobs rather than fewer. He treats Safe Superintelligence as part of a broader positive-AI movement. Almeida rejects the bleak “one big brain to rule them all” framing: simple tasks remain stubbornly manual, so he sees today’s tragedy not as runaway automation but as powerful intelligence that has barely been released into productive systems.
3. Benchmark intelligence diverged from dependable economic work
Almeida’s view changed around the end of 2021, when RLHF generalized far better than he expected. His team tested deliberately strange questions—“why is it important to eat socks before meditation?”—after checking that answers were not online; plausible human-like responses convinced them the capability was not a scam.
He initially believed the model had a meaningful chance of becoming SHGI. When that did not happen, “my whole world just collapsed,” forcing the question that now drives Jev: if the intelligence is real, why does it remain so hard to convert into useful automation?
His proposed answer is evaluation mismatch. Because humans judge model quality, the industry optimized the human judge rather than task completion; that can make models look extraordinarily capable while leaving the software layer unable to trust or compose their outputs.
Casado highlights the mismatch by invoking the apparent solution of math and GPQA—described as questions even Google does not know the answers to—despite failures at ordinary productive tasks. Almeida concedes that real-world work has heavy-tailed exceptions but rejects the implied standard: automation need not solve every edge case. The “canary in the coal mine” is whether obvious, high-return work can be automated at all—OpenAI has pursued customer-service automation since 2020, yet very little is automated within companies.
4. Reliability means consistent intelligence, not identical tokens
Almeida says the launch and breadth of use cases were extremely surprising; Casado reports hearing hourly about people using Claude for new tasks. Almeida stresses what users may miss: years of “blood, sweat, and tears” spent on reliability. He believes every additional percentage point will unlock workflows developers currently cannot risk automating.
He separates three concepts: availability is uptime or an SLA; determinism is closer to reproducing the same result; reliability is “the same intelligence every time.” Adding an irrelevant UUID should not change the functional answer, even though the tokens need not be identical.
The ultimate test is whether developers can program with Jev without constructing example queries—remaining “in a constant state of flow” because they trust each call to produce an intelligible result their code can handle.
Coding agents remain complementary but bounded. Almeida finds them excellent at syntax, weak at semantics, and “incredibly bad at architecture,” which he calls software’s most human and creative layer. Casado adds that models may be only at the 50th-percentile level in architecture, yet a 50th-percentile architecture may beat a 60th-percentile one when speed is the decisive leverage and Codex can run all night.
5. SaaS and systems may both be rebuilt around the new primitive
Against the “SaaS apocalypse,” Almeida argues that software may be cheap and easy to copy but is not easy to reproduce: substantial behavior lives under the hood. Incumbent SaaS companies understand users, workflows, and distribution, making them natural vehicles for spreading a capability upgrade across an existing customer base.
The envisioned interface change is larger than adding chat. Forms and option menus might disappear because they seem to map natural language into Java-like structures; one unverified developer-app example involved voice while continuously deciding whether speech was a command or text input and where that input belonged.
The speakers leave system boundaries open. Casado flags state consistency, reliability, and strong system-level guarantees; Almeida says log analysis, email analysis, user-interface design, and human interaction will change, and he also mentions air-traffic-control systems as a possible application. Horowitz’s rule is to automate easy work before hard work.
Almeida’s deepest analogy is TCP: a layer that turns something unreliable into something dependable. Today, models can fail to honor JSON schemas, so developers route output to a person or another language model, creating a human-in-the-loop chatbot or an agent “while” loop. Jev aims to turn natural language into a productive state machine. Horowitz calls it an “anti-frustration machine,” while Almeida’s utopia is technology that can finally “do as I mean.”