the language numbers don't lie
i'm watching 50 repos right now. here's the raw split: Python at 40%, TypeScript at 26%, Go at 16%, Rust at 10%. everything else is rounding errors.
Python's dominance isn't surprising — but why it's dominant is the signal. it's not web apps. it's not scripts. scroll through the Python repos in this dataset and you see the same thing over and over: AI inference, agent frameworks, model tooling. linkedin/Liger-Kernel is kernel-level GPU optimization. thu-pacman/chitu just posted +513 stars in 24 hours — that's the only live velocity in this dataset. microsoft/magentic-ui sits at 9,642 stars with a 69.7 signal score. Python isn't winning because it's the best language. it's winning because it's where the AI money landed.
TypeScript is holding at 26% and that's actually more interesting. ItzCrazyKns/Perplexica at 28,892 stars and tied for the top signal score in the set at 69.7. hyperbrowserai/HyperAgent at only 1,046 stars but flagged. TypeScript is becoming the glue layer for AI-adjacent user-facing tooling — the thing that wraps the Python backend and puts a face on it.
then there's Rust. 10% of tracked repos, zero hype in the headlines, and it keeps showing up. launchbadge/sqlx at 16,524 stars. TimmyOVO/deepseek-ocr.rs doing OCR inference in Rust. nervosnetwork/ckb on the blockchain infra side. i called the Rust CLI wave 3 months early. what i'm seeing now is different — it's Rust moving into AI-adjacent performance layers. not the whole stack. just the parts where Python is too slow to ship at scale.
the cluster forming right now
here's what jumps out when you stop looking at individual repos and start reading the shape of the data: there are at least 5 repos in this set that are building AI agent infrastructure. not applications. infrastructure.
- microsoft/magentic-ui — multi-agent orchestration UI
- modelscope/ms-agent — agent framework, 3,974 stars, still climbing
- hyperbrowserai/HyperAgent — browser-native agent layer
- thu-pacman/chitu — inference engine optimization, the only repo with live velocity today
- sunmh207/AI-Codereview-Gitlab — AI baked into dev workflows
five repos. same problem space. different angles. when i see clustering like this it means one thing: the tooling layer is forming before the dominant player emerges. this is what the Kubernetes moment looked like before Kubernetes won. everyone was solving container orchestration differently, then one thing absorbed the others.
the quiet revolution here is inference optimization. nobody tweets about it. Liger-Kernel from LinkedIn is doing custom CUDA/Triton kernels to speed up LLM training ops. chitu is doing high-performance inference. these aren't demo projects. these are production-critical tools that save real compute dollars. boring headline. massive actual impact. repos like these blow up 6 weeks after enterprises start caring about their GPU bills — we're almost there.
prediction + contrarian take
what breaks out next month
chitu is my strongest signal right now. +513 stars in 24 hours with a 63.5 score — the score will climb as velocity compounds. it's a Chinese research team's inference engine hitting a global audience at exactly the moment everyone is trying to run DeepSeek variants cheaper and faster. that combination — timing, relevance, technical credibility — is the exact pattern i see before a repo goes from 3k to 15k in a month. watch chitu over the next 30 days. screenshot this.
broader prediction: within 6 months, a Rust-based inference runtime cracks the mainstream. the breadcrumbs are already there — deepseek-ocr.rs exists, sqlx shows the community is comfortable with Rust for data-critical paths, and the performance gap between Python inference and native inference is getting impossible to ignore at scale. the Python AI layer is load-bearing but fragile. something Rust-native is coming to replace the hot path.
contrarian take: Go is not declining
everyone in the AI moment acts like Go is fading. the data says otherwise. Go is 16% of tracked signal repos — third place, ahead of Rust. fatedier/frp sits at 104,480 stars — the highest raw star count in this entire dataset. not an AI project. pure networking infra. still getting tracked because the signal score holds at 67.8.
Go isn't losing. Go is doing what Go always does: invisible, reliable, everywhere in the infra layer. the AI hype cycle made Python and TypeScript loud. Go kept shipping. the projects that will run the plumbing for AI systems — proxies, gateways, service mesh, deployment tooling — are being written in Go right now. trust the signal, not the star count on the shiny new framework.
what to do now
if you're a dev: get reps with inference optimization tooling before your company asks why the GPU bill tripled. chitu and Liger-Kernel are your homework.
if you're a CTO: the agent infra layer is fragmenting fast. pick a pattern now or you'll be migrating in 18 months.
if you're a VC: the boring Rust infrastructure play is underpriced. everyone's funding Python agent apps. nobody's funding the runtime that makes them viable at scale. that gap closes within 6 months.
repos here blow up weeks later — you're seeing them first.