AI Repo: karpathy/nanoGPT

Skyler: Hey everyone, welcome back to Git News A I, your weekly dose of the hottest A I repositories blowing up on GitHub. I'm Skyler, and before we dive in, a quick shoutout to today's sponsor, LimitLess A I, helping you get more done with the power of A I. Stick around for that. Today we've got five seriously interesting repos on the docket — we're talking local meeting assistants, free L L M APIs, giving Claude the ability to watch videos, and a couple of classics that absolutely refuse to stop being relevant. Let's get into it. Skyler: Alright, kicking things off with a repo that a lot of you probably already have starred — Andrej Karpathy's nanoGPT. This one sits at nearly 61,000 stars and honestly, if you haven't looked at it yet, what are you doing? NanoGPT is exactly what it sounds like — the simplest, fastest repository for training and finetuning medium-sized GPT models. And when Karpathy says simple, he means it. We're talking clean, readable Python code that you can actually sit down with and understand from top to bottom. Skyler: What makes this so special is that it strips away all the abstraction layers that typically make transformer training feel intimidating. You get a fully functional GPT implementation in just a few hundred lines of code. It supports training from scratch or finetuning on your own data, and it's optimized to run efficiently on a single GPU, multiple GPUs, or even just a CPU if you're patient enough. Skyler: The repo hasn't seen recent commits, which honestly just tells you it's mature and stable at this point. It's not a half-baked experiment — it's a reference implementation. If you're learning how transformers work under the hood, if you want to train a small language model on custom text data, or if you're teaching others about G P T architectures, nanoGPT is the starting point. M L engineers and students alike absolutely need this in their toolkit. Pure gold. Skyler: Next up, let's talk about something that is actively being developed and shipping features fast — Meetily, from Zackriya Solutions. This one is sitting at almost 17,000 stars and climbing, and I totally get why. Meetily is a privacy-first, A I-powered meeting assistant, and the key phrase here is privacy-first. Everything runs one hundred percent locally. No cloud. No sending your meeting audio to some third-party server. Your conversations stay on your machine, period. Skyler: So what does it actually do? Meetily handles live transcription using either Parakeet or Whisper — and it claims to be four times faster than standard Whisper implementations — plus speaker diarization, which means it can actually figure out who is talking when, and then it uses Ollama to run local L L M summarization so you get clean, useful meeting notes automatically. All of this is built in Rust, which explains a lot of that performance story. Skyler: It supports both macOS and Windows, and the team just dropped version 0.4.0 with an updated model manager UI, so they're clearly iterating quickly. If you work remotely, if you're in a lot of meetings, if you work in a sensitive industry where data privacy actually matters, or honestly if you're just tired of paying for cloud-based meeting tools, Meetily is absolutely worth spinning up. Check the show notes for the link. Skyler: Alright, number three — OpenAI's Whisper. Over 104,000 stars, making it one of the most starred A I repos on all of GitHub, and honestly it's not hard to see why. Whisper is OpenAI's robust speech recognition model trained using large-scale weak supervision, which is a fancy way of saying they trained it on an enormous, diverse dataset of audio from across the internet, and the result is a model that handles accents, background noise, and multiple languages remarkably well. Skyler: Whisper is the backbone behind a ton of other projects — including Meetily that we just talked about — and it comes in multiple model sizes, from tiny models you can run on almost anything, all the way up to large models that deliver near human-level transcription accuracy. It's open source, it's Python-based, and the community has built an absolutely massive ecosystem of tools, wrappers, and integrations on top of it. Skyler: Like nanoGPT, Whisper hasn't seen recent commits, but again — that's maturity, not abandonment. If you're building anything that involves audio, voice interfaces, transcription pipelines, subtitle generation, or even just accessibility tooling, Whisper is your foundation. If somehow you haven't explored it yet, go fix that. Check the show notes for the link. Skyler: Number four is one I was genuinely excited to find — cheahjs's free-llm-api-resources. 25,000 stars and growing, and the concept is beautifully simple. This is a curated list of free L L M inference resources that you can access via A P I. That's it. Someone went out and catalogued all the places where you can actually hit a real L L M endpoint without handing over a credit card, and they've kept it up to date. Skyler: And looking at the recent activity, there have been multiple README updates just in the last little while adding the latest models, so this is actively maintained. We're talking providers that offer free tiers, open access endpoints, and trial credits — covering models from all the major players and a bunch of open source options too. For developers who are prototyping, experimenting, or just don't want to burn through API credits while they're building something new, this list is an absolute treasure. Skyler: It's also just a great reference for understanding the current landscape of L L M providers and what's out there. If you're a developer who wants to test different models, compare outputs, or integrate A I into a side project without upfront cost, bookmark this repo immediately. Seriously, this is one of those resources that should be in every A I developer's browser favorites. Check the show notes for the link. Skyler: And finally, we've got something really fun — claude-video by bradautomates, sitting at about 3,500 stars. The pitch is simple and kind of mind-bending: give Claude the ability to watch any video. That's it. That's the project. You point it at a video, and the tool downloads it, extracts frames at regular intervals, transcribes the audio, bundles all of that context together, and hands it off to Claude for analysis. Skyler: So in practice, you run a slash watch command with a video URL, and Claude can answer questions about what's in the video, summarize it, describe scenes, analyze content — whatever you need. It recently shipped version 0.2.0 with some tightened documentation and a fix for a silent fallback issue in the watch detail configuration, so the developer is clearly paying attention to stability and usability. Skyler: This is Python-based and honestly feels like one of those projects that's a glimpse at where multimodal A I workflows are heading — where you're composing multiple A I capabilities together to do something none of them could do alone. Audio transcription, visual frame analysis, language model reasoning — all chained together. If you're building Claude-powered tools, if you do video content analysis, or if you just love seeing creative A I integrations, claude-video is worth a deep dive. Check the show notes for the link. Skyler: And that's a wrap on today's episode of Git News A I! Huge thanks to our sponsor LimitLess A I for supporting the show — seriously go check them out if you want to supercharge your productivity with A I. If you found today's repos useful — and come on, that free L L M list alone was worth the listen — make sure you subscribe wherever you get your podcasts so you never miss an episode. All the links to every repo we covered today are in the show notes, so go explore, go build something cool, and I'll catch you in the next one.