a16z Podcast

Diogo Almeida on why AI belongs inside software, not writing it

Diogo Almeida· Co-founder and CEO at TypeSafe AI
·~42 min·English·a16z
LLMAgentsAI CompanyBusiness Strategy
TL;DR

TypeSafe's Diogo Almeida argues that AI has stalled on real automation because we keep making it write software for humans to read, and that the fix is Jev, a reliable classifier you drop inside programs so software itself can make smart decisions.

01The Core Frustration

Where the hell is all the automation?

Almeida's whole company starts from one complaint: <strong>AI is astonishingly smart on tests and astonishingly useless at ordinary work</strong>.

AI is so unbelievably smart and yet it's so useless at all other stuff.

— Diogo Almeida, a16z Podcast
Key Insight
The two columns are the same models, judged on two different axes. Benchmarks reward answers a human grades; real work rewards actions a system can trust unattended, and almost nothing has crossed from the first column to the second.
02The Thesis

Smart software, not just-in-time software

Coding agents make the software we already have, faster; Almeida wants the opposite, <strong>to expand what software itself is able to do</strong>.

What I want instead is smart software. I want to expand what software itself can do such that things that should be automatable can then be automatable.

— Diogo Almeida, a16z Podcast
Key Insight
This reframes the whole AI-for-software race. Everyone else is automating the programmer; Almeida is trying to add a new building block to the program, so the win is measured in new capabilities, not in lines typed per hour.
03What It Actually Is

A classifier you program on the fly

Under the vision is a deliberately humble machine: <strong>you send state and a typed question, it returns a value and how sure it is</strong>, not a paragraph.

my guess is that Jev probably is better than having like an MLE team from 2019 making the stuff for you and you can just program it on the fly.

— Diogo Almeida, a16z Podcast
Key Insight
Framing a frontier model as a classifier is a positioning choice, not a limitation. A chatbot is judged on whether its text sounds good to a person; a classifier is judged on whether its answer is correct enough to build on, which is exactly the bar automation needs, and it means a developer can wire up narrow intelligence without a data-science team.
04The Moat

Reliability is the whole product

Almeida separates three things people lump together, and stakes the company on the hardest one: <strong>being smart every time, not identical every time</strong>.

reliability is what this thing is. If you don't understand that, it'll be very hard to make like a copycat that's benchmaxed.

— Diogo Almeida, a16z Podcast
Key Insight
The distinction is the moat. Determinism (same bytes out) is easy to copy and useless for real systems; robustness (a smart answer every time, even when the exact answer differs) is what lets a developer trust the model in the background, and it is far harder to fake with benchmark scores.
05The Market Turn

The inverse of the SaaS apocalypse

Coding agents made SaaS look doomed; Almeida predicts the same primitive will turn <strong>those companies into some of AI's biggest winners</strong>.

it's going to be like an inverse apocalypse. And I am so jazzed about it.

— Diogo Almeida, a16z Podcast
Key Insight
The reversal turns on who owns the hard parts. If AI just writes code, software looks cheap to replace; but SaaS companies already own the customers and know which workflows matter, so a primitive that makes their product smarter lands on top of a moat they spent years building.
06The Worldview

Prod, not god

Against the one-giant-brain story, Almeida takes a pragmatist's bet: <strong>automate the boring, high-volume work first, and jobs grow rather than vanish</strong>.

automate the easy work before the hard work is always my philosophy

— Diogo Almeida, a16z Podcast
Key Insight
The slogan is really a claim about where the value is. Chasing one superintelligence bets everything on a distant leap; automating rote work with reliable narrow intelligence collects the economically valuable wins now, which is why Almeida expects more and better jobs, not fewer.
07The Vision

Technology that does what you mean

The end state Almeida is aiming at is quiet, not dramatic: <strong>technology that simply does what you mean</strong>, with the friction taken out.

having everything like just move more smoothly together and interlink like a gears.

— Diogo Almeida, a16z Podcast
Key Insight
This is the human payoff of every earlier section. If software can hold intent and act on it reliably, the drop-downs, multiple-choice forms and glue code that translate for dumb systems can disappear, and the interface becomes the meaning itself.