AI Can Write Code. Why Isn’t Software Better?

Episode Summary

Executive Summary: The episode argues that current AI is impressive but underused: coding agents speed up software production, yet they mostly recreate existing software rather than expand what software can do. Diogo Almeida explains TypeSafe/Jev as a new primitive that embeds intelligence inside software via probabilistic, state-machine-based decision-making, aiming to make products more reliable, composable, and truly automatable—especially in SaaS and enterprise workflows.

Main Topics: Why AI feels smart but automation is still missing (Priority: 5/5): Diogo’s core complaint is that AI is powerful in demos and coding assistance, but the world has not seen equivalent gains in real automation. He frames Jev/TypeSafe as an attempt to unlock the automation AI should already be enabling. Jev as a new primitive for software (Priority: 5/5): Rather than generating more code, Jev is positioned as a library-like tool that inserts intelligence into software itself, allowing natural language intent to drive decisions inside programs. Reliability over flashy demos (Priority: 5/5): The conversation repeatedly distinguishes between benchmark wins and production-grade systems. Jev is designed around uptime, determinism, robustness, and statistical guarantees so it can run in real workflows without paging humans. AI’s role in software engineering vs. software capability (Priority: 4/5): Coding agents like Cursor, Codex, and Claude Code are described as useful for writing existing code faster, but not as a path to better software. Jev instead aims to expand the expressive power of software. SaaS as a major beneficiary of AI (Priority: 4/5): The hosts and Diogo argue that SaaS is not collapsing under AI; instead, it may become dramatically more useful as AI is embedded into existing workflows, interfaces, and internal operations. Probabilistic programming and future systems design (Priority: 4/5): The episode explores a possible new era of probabilistic, AI-native programming where software components trade off intelligence, cost, and speed, enabling more flexible system architectures. Diogo’s path from math, systems, and AI research to product (Priority: 3/5): Diogo explains how competition math, systems thinking, and an unusually pragmatic AI background shaped his focus on building useful, reliable software rather than chasing AGI narratives.

Key Arguments: Current AI is overemphasized in human-facing chat and code generation, but underdelivers on actual automation of useful work. Coding agents help produce software faster, but they usually produce the same kind of code humans would have written. Jev is meant to expand what software itself can do by adding an AI-based decision primitive inside programs. Reliability is the gating factor for real automation; demos are not enough if systems can’t run unattended with confidence. The right success metric is not model impressiveness, but how much productive work can actually be automated. SaaS will likely benefit from AI because existing products already sit on valuable workflows and can use AI to automate them more deeply. AI should be judged on productive-task automation, not just chat quality or benchmark performance. A future with many AI components inside systems will require new software architecture patterns and trade-offs between intelligence, cost, and speed. Do-what-I-mean interfaces and probabilistic logic could make software smoother, more natural, and more capable for end users.

Data Points: OpenAI customer-service automation attempts: since 2020 - Diogo cites OpenAI’s long-running effort as evidence that real automation has lagged behind AI hype. Kaggle competition result: won by automating the fuck out of it - Diogo describes his Kaggle win as a systems-oriented automation approach rather than pure mathematical sophistication. RLHF timing: Q4 2021 - He says his intuition shifted around the period RLHF was being developed and evaluated internally. Support-call automation example: 95% - A company claimed to answer 95% of help desk calls, but the point was that most were password resets. Support-call uniqueness after inspection: about 50% - Diogo notes that by uniqueness, the work looked far less automatable than the headline metric suggested. Typical large-company PR size: 10 lines - The hosts mention that average pull requests in a large company can be very small, underscoring how coding agents may automate minor changes rather than transform software. Reliability target: many nines - Diogo says the product is being built for high reliability, analogous to infrastructure-grade systems.

Pivotal Quotes: "Where the fuck is all the automation?" — Diogo Almeida: His elevator pitch for why AI is disappointing despite being highly intelligent. "We build prod, not God." — Host quoting TypeSafe/OpenAI culture: Used to contrast TypeSafe’s practical, product-focused optimism with grand AGI narratives. "I want to expand what software itself can do, such that things that should be automatable can then be automatable." — Diogo Almeida: Defines the core mission behind Jev and TypeSafe.

Implications: If successful, AI will not just speed up coding; it will make enterprise software fundamentally more capable, reliable, and autonomous. The near-term winners may be SaaS products that embed AI deeply rather than replace humans or software outright.