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

In this A16Z podcast episode, Ben Horowitz and Martin Casado interview Diogo Almeida about TypeSafe AI and its flagship product Jev. The central thesis is captured in Almeida's opening question: "Where the fuck is all the automation?" Despite AI's remarkable capabilities, basic automation tasks remain unachieved, which Almeida views as a missed opportunity rather than a technological limitation.

Almeida contrasts two approaches to AI in software: coding agents that automate software engineering (making developers faster at writing the same code), versus embedding intelligence directly into software itself through Jev. Rather than replacing software engineers, Jev expands what software can do by providing a new primitive—essentially a library that accepts natural language descriptions of intent, processes state information, and outputs probabilistic decisions with confidence levels. This represents a fundamental shift from generating code to generating decisions within code.

The conversation touches on why automation has stalled despite AI capabilities. Almeida argues this stems from misconceptions about AI's current limitations, and partly from industry focus on impressive demos rather than reliable production systems. He emphasizes that OpenAI's customer service automation efforts since 2020 have failed, suggesting the problem isn't technical capability but rather the gap between what AI can do in controlled settings and what's needed for real-world deployment.

Almeida shares his personal journey from competitive mathematics through Kaggle competitions to OpenAI, where he worked on early GPT models and RLHF. He describes his disillusionment when GPT-3 didn't lead to AGI, which prompted a fundamental rethinking of AI's role. This led to the insight that the real opportunity isn't in replacing humans but in making software dramatically more capable.

The discussion covers reliability as a core differentiator, with Almeida distinguishing between uptime/SLA, determinism, and robustness (intelligent behavior every time). He argues that true reliability means developers can eventually program against Jev without constant example queries, trusting it to run autonomously.

Regarding market implications, Almeida predicts an inverse "SaaS-pocalypse." While coding agents initially sparked fears that SaaS companies would become obsolete (since software would become cheap and easy to replicate), Jev actually makes existing SaaS products dramatically more valuable by adding new capabilities. He envisions SaaS companies as ideal partners since they best understand customer workflows and user problems.

The conversation also explores deeper architectural possibilities, including how AI could be distributed throughout systems at different cost and speed trade-offs, potentially enabling a new era of probabilistic programming. Almeida references the historical failure of probabilistic programming in the 1970s but sees Jev as a practical path forward. He mentions Eric, his co-founder with a Bayesian/biology background, though emphasizes his own brand as pragmatism over biological inspiration.

Almeida addresses the relationship between Jev and coding agents, noting that coding agents excel at syntax but are poor at architecture. He sees no fundamental conflict—agents could handle syntax while humans (or AI) handle architecture, with trade-offs depending on project needs. He also discusses potential applications like voice control for computers, where Jev would constantly decide whether input is a command or text insertion.