Two years of GitHub data, one claim: the AI application layer is changing its name to TypeScript.
A Song of Types and Agents - Roberto Stagi, Ratel · Roberto Stagi
14 min total·Actually worth watching closely: ~5 min·2 must-watch clips
- 0:02 – 1:40Listen
Opening: who sits on the AI language throne
The speaker introduces himself and sets up the suspense of the talk — a song of types and agents, languages fighting to conquer the throne in the AI realm — starting from the fact that a few years ago, building in AI certainly meant Python, no doubt about it.
"AI means Python" is an inherited default that deserves to be re-examined.
Mostly the speaker framing the question out loud with only a title slide on screen; listening is enough.▶ Jump to 0:02Speaker · Roberto Stagi - 1:40 – 5:00Watch
Two annual reports: Python reaches the top, then gets passed
GitHub's official annual report data shows Python riding the AI wave to become the most popular language in 2024, then being passed by TypeScript in August 2025. The funny part is that both reports gave the same reason: AI.
One and the same cause pushed two different languages to the top in two years, so what changed isn't AI itself but what people do with it.
Everything is in the side-by-side rankings; the movement in position and the attribution text only hit if you look, so listening alone loses the strongest evidence here.▶ Jump to 1:40Speaker · Roberto Stagi - 5:00 – 7:05Skim
A flood of developers, and the one variable that changed
A curve of global developer growth — in 2025 one new developer joining GitHub every second — and then the thing that actually changed between 2024 and 2025: coding agents grew up. Lovable, Claude Code, Cursor and Codex became the default way to build applications, and what they emit by default is TypeScript.
The newcomers didn't pick TypeScript for themselves; the tools picked it for them.
The slide is a curve heading steadily up; catch the direction and move on, no need to linger frame by frame. The conclusion is carried by the audio.▶ Jump to 5:00Speaker · Roberto Stagi - 7:05 – 9:20Listen
AI is climbing from the infrastructure layer to the application layer
The argument that AI's center of gravity is moving up from model training and machine learning infrastructure to the application layer — AI stopped being something you train and became something you ship inside your application, so applications themselves started to think.
Which language to use depends on whether you're building the brain or the product that uses the brain.
Pure argument with only a few bullet lines on screen; follow the chain of reasoning by ear, fine to listen while doing something else.▶ Jump to 7:05Speaker · Roberto Stagi - 9:20 – 11:05Listen
The flywheel and the single-codebase math
First the feedback loop: the TypeScript applications agents produce become training material for the next generation of agents, so output quality keeps improving. Then the engineering math: TypeScript holds the agent loop, the tools, the backend and the UI in one language and one codebase, with Zod as the single data contract across the whole application, type checked end to end.
For the same agent product, Python means splitting into at least two services and hand-syncing two sets of types across a contract; TypeScript needs only one.
The architecture comparison is carried by the speaker's words, with no code walkthrough to read line by line — if you follow along, there's no need to watch.▶ Jump to 9:20Speaker · Roberto Stagi - 11:05 – 13:10Watch
Ecosystem edge: the package manager and an exploding curve
NPM is the richest package ecosystem out there — authentication, payments, UI, infra — the shortest path to those capabilities now that AI is coming toward the application layer; plus a downloads chart showing the AI toolchain on the TypeScript side growing close to 10x in a year.
The application layer needs more than model calls — auth, payments, UI — and most of that already exists on the NPM side.
That near-10x downloads curve is the whole persuasive force of this stretch; you have to see the slope yourself to feel it.▶ Jump to 11:05Speaker · Roberto Stagi - 13:10 – 14:13Listen
Betting per layer: training to Python, agents to TypeScript
The closing recommendation: Python will keep the brain of AI for a long time — training, research and GPU serving aren't changing hands — but the agent that calls the models belongs to the application layer and will ship through the package ecosystem. An AI lab acquiring a JavaScript runtime backs it up, then thanks and a QR code for the slides.
Don't pick a side — keep training in Python, build the agents and applications in TypeScript, and bet on the two layers separately.
The closing section is the speaker stating his call and recommendation directly against a single summary slide; hearing that actionable conclusion is enough.▶ Jump to 13:10Speaker · Roberto Stagi