the language leaderboard right now
i'm watching 50 repos actively. here's what the distribution looks like: Python at 20 repos (40%), TypeScript at 13 (26%), Go at 8 (16%), Rust at 5 (10%). everything else is noise.
Python's dominance isn't surprising — it's where all the AI tooling lands first. but 40% is unusually high even for this cycle. what's interesting is why: it's not web apps or scripts. it's inference kernels, agent orchestration, and ML pipelines. look at linkedin/Liger-Kernel (6,142 stars, signal score 69.3) — that's LinkedIn shipping custom CUDA kernel optimizations in Python wrappers. that's not hobbyist stuff. that's production infra.
TypeScript holds at 26% and it's almost entirely agent UIs and AI-native apps. ItzCrazyKns/Perplexica sitting at 28,892 stars with a 69.7 signal score tells you everything — people are building Perplexity clones and the frontend layer is TypeScript, every time.
but here's the number i keep coming back to: Rust at 10% with 5 repos. that sounds small. it isn't. Rust repos in my tracker historically punch above their weight on signal scores. launchbadge/sqlx is at 16,524 stars and 66.3. TimmyOVO/deepseek-ocr.rs is at 2,127 stars — but it's Rust doing OCR with DeepSeek. that's a new category forming in real time.
the clusters you should be watching
cluster 1: AI agents are a crowded room
count them: microsoft/magentic-ui, modelscope/ms-agent, hyperbrowserai/HyperAgent. three agent frameworks in the top 15. Microsoft, ModelScope (Alibaba's research arm), and a startup all shipping agent orchestration layers within the same signal window. when three different orgs with three different resource levels are building the same thing simultaneously, that's not coincidence — that's a platform gap closing fast.
magentic-ui is at 9,642 stars with a 69.7 signal score. ms-agent is at 3,974 with 65.3. HyperAgent is only at 1,046 but it's TypeScript-native and browser-focused — different surface area. the tension here: these don't all survive. one framework wins the enterprise integrations. my bet is on whichever one ships an MCP-compatible runtime first.
cluster 2: the quiet AI code review wave
sunmh207/AI-Codereview-Gitlab only has 1,404 stars. signal score 64.8. most people scroll past this. don't. AI-native code review tooling that plugs directly into GitLab CI is a category that enterprises will pay for, and right now it's being built in the open. i've seen this pattern before — low star count, high signal score, specific enterprise use case. this is where acqui-hires come from.
cluster 3: the quiet revolution — inference optimization
this is the unsexy one. nobody's tweeting about kernel optimization. but linkedin/Liger-Kernel and thu-pacman/chitu (2,915 stars, +513 stars in 24 hours — the only repo in this dataset with meaningful velocity right now) are both attacking the same problem: making inference cheaper at scale.
chitu's +513 in 24h is the loudest signal in this entire dataset. everything else shows zero 24h velocity — which means the market is in a consolidation moment. but chitu broke through. thu-pacman/chitu is a speculative decoding framework out of Tsinghua. speculative decoding is how you get 2-4x faster inference without new hardware. this is the trade everyone's sleeping on.
my prediction + contrarian take
what breaks out next month
inference optimization tooling goes mainstream within 6 weeks. the signal is already there: chitu's velocity spike, Liger-Kernel's sustained score, the fact that every major lab is now cost-constrained on inference. the next wave of stars won't go to another chatbot UI or agent framework — they'll go to whatever makes running models 3x cheaper. watch for Rust-based inference runtimes specifically. deepseek-ocr.rs is an early tell. within 60 days I expect a Rust inference project to hit 10k stars from a standing start.
Go isn't going anywhere either. fatedier/frp at 104,480 stars proves the floor. 8 Go repos at 16% of my tracked set — all infrastructure, all boring, all critical. tunneling, proxies, service mesh primitives. CTOs: this is what your platform team is actually using.
contrarian take: TypeScript AI apps are overbuilt
everyone believes the next breakout repo is a TypeScript AI app. the data says the opposite. TypeScript is 26% of tracked repos but the signal scores are clustering in the 62-65 range — not the 68-70 range where the real breakouts live. Open-Dev-Society/OpenStock is at 8,526 stars and only 65.3. oslook/cursor-ai-downloads is at 3,152 stars and 62.3. solid repos, not breakouts.
the alpha is in Python infra and Rust runtimes, not TypeScript frontends. the frontend AI app moment peaked. what comes next is the plumbing layer — and that's written in Python and Rust, not React.
what to do now
- star thu-pacman/chitu today. +513 stars in 24h on a technical inference paper implementation is the realest signal in this dataset
- watch the Rust cluster. 5 repos at 10% with above-average signal scores = a wave forming, not peaked
- if you're building AI tooling, the agent framework market is saturated at the top. the gap is in evaluation, cost optimization, and GitLab/GitHub-native integrations
- ignore the TypeScript AI app hype for 90 days. the signal says it's overcrowded
repos here blow up weeks later — you're seeing them first. trust the signal, not the star count.