Stefan Wintermeyer

@jkwchui I have never been a big fan of Livebook for this kind of beginners training. But that is just me. Again, I really think that there is some sort of magic of starting the iex and typing away commands you read as examples in a documentation. And that is language independent.

@wintermeyer For me, that actually was Livebook. It makes bringing in and out dependencies and data immediate, and everything persists enough that when the module grows into something useful, there is no friction to a Mix project.

This is great! BUT I think the learning effect is a lot higher when people have to start an iex in their terminal and start typing. Start playing with the examples. Not just for Elixir. Any programming language is best learned by first typing simple Hello World examples and than play with them. Some sort of magic happens. You either start liking or disliking a language. You dive deeper or give up.

Nokolexbor is a high performance HTML parser for Ruby that can replace Nokogiri 🚀 ✅4.7x faster at parsing HTML ✅1352x faster at CSS selectors (at_css selector) ✅similar performance for the rest ✅a drop-in replacement Built by SerpApi. github.com/serpapi/nokole…

Me have a great idea! What if me make more cookies... and then me eat them?

For those that want to follow, or try, Zap the project is here: github.com/DockYard/zap I've turned off Issues, PRs, Wiki, and Discussions for now as I have some opinions on what I want to do with it and want the freedom to explore.

We've been building an internal Claude Code plugin system at Intercom with 13 plugins, 100+ skills, and hooks that turn Claude into a full-stack engineering platform. Lots done, more to do. Here's a thread of some highlights.

🎾Introducing LATENT: Learning Athletic Humanoid Tennis Skills from Imperfect Human Motion Data Dynamic movements, agile whole-body coordination, and rapid reactions. A step toward athletic humanoid sports skills. Project: zzk273.github.io/LATENT/ Code: github.com/GalaxyGeneralR…

Also für Deutschland: 14:00 - 19:00 Uhr

We open sourced WAXAL! - Multilingual speech dataset for African languages - 17 languages for TTS - 19 languages for ASR Over 100 million speakers across 40 Sub-Saharan African countries huggingface.co/datasets/googl…

My two favorite Claude Code features that seem relatively less used are a) /insights and b) /playground.. For insights, I run in every two weeks. Ask it for improvements on how to improve my workflow. Ask it to implement the improvements. And voila, 20% more productive. For playground, it's the best way to visualize what's happening in the workflow. I ask it to show me architecture of esoteric Githubs to simulating a full learning loop.

I love running "hail mary" prompts like this. A few weeks ago I prompted Opus to find code loading optimizations in the Erlang/OTP code base. It came up with 6-7 options, 3 of which I could automatically discard, and I asked it to build experiments for the remaining ones. Out of those, 1 was clearly successful, which I then wrapped up and now Erlang/OTP 29 will boot 10% faster for everyone. /autoresearch from @karpathy seems to package this experience into a tighter loop and, if it can find something meaningful, it stands to benefit everyone, especially on OSS. Can't wait to try it and maybe "hail mary" a few other optimizations.

It is hard to communicate how much programming has changed due to AI in the last 2 months: not gradually and over time in the "progress as usual" way, but specifically this last December. There are a number of asterisks but imo coding agents basically didn’t work before December and basically work since - the models have significantly higher quality, long-term coherence and tenacity and they can power through large and long tasks, well past enough that it is extremely disruptive to the default programming workflow. Just to give an example, over the weekend I was building a local video analysis dashboard for the cameras of my home so I wrote: “Here is the local IP and username/password of my DGX Spark. Log in, set up ssh keys, set up vLLM, download and bench Qwen3-VL, set up a server endpoint to inference videos, a basic web ui dashboard, test everything, set it up with systemd, record memory notes for yourself and write up a markdown report for me”. The agent went off for ~30 minutes, ran into multiple issues, researched solutions online, resolved them one by one, wrote the code, tested it, debugged it, set up the services, and came back with the report and it was just done. I didn’t touch anything. All of this could easily have been a weekend project just 3 months ago but today it’s something you kick off and forget about for 30 minutes. As a result, programming is becoming unrecognizable. You’re not typing computer code into an editor like the way things were since computers were invented, that era is over. You're spinning up AI agents, giving them tasks *in English* and managing and reviewing their work in parallel. The biggest prize is in figuring out how you can keep ascending the layers of abstraction to set up long-running orchestrator Claws with all of the right tools, memory and instructions that productively manage multiple parallel Code instances for you. The leverage achievable via top tier "agentic engineering" feels very high right now. It’s not perfect, it needs high-level direction, judgement, taste, oversight, iteration and hints and ideas. It works a lot better in some scenarios than others (e.g. especially for tasks that are well-specified and where you can verify/test functionality). The key is to build intuition to decompose the task just right to hand off the parts that work and help out around the edges. But imo, this is nowhere near "business as usual" time in software.

Three days ago I left autoresearch tuning nanochat for ~2 days on depth=12 model. It found ~20 changes that improved the validation loss. I tested these changes yesterday and all of them were additive and transferred to larger (depth=24) models. Stacking up all of these changes, today I measured that the leaderboard's "Time to GPT-2" drops from 2.02 hours to 1.80 hours (~11% improvement), this will be the new leaderboard entry. So yes, these are real improvements and they make an actual difference. I am mildly surprised that my very first naive attempt already worked this well on top of what I thought was already a fairly manually well-tuned project. This is a first for me because I am very used to doing the iterative optimization of neural network training manually. You come up with ideas, you implement them, you check if they work (better validation loss), you come up with new ideas based on that, you read some papers for inspiration, etc etc. This is the bread and butter of what I do daily for 2 decades. Seeing the agent do this entire workflow end-to-end and all by itself as it worked through approx. 700 changes autonomously is wild. It really looked at the sequence of results of experiments and used that to plan the next ones. It's not novel, ground-breaking "research" (yet), but all the adjustments are "real", I didn't find them manually previously, and they stack up and actually improved nanochat. Among the bigger things e.g.: - It noticed an oversight that my parameterless QKnorm didn't have a scaler multiplier attached, so my attention was too diffuse. The agent found multipliers to sharpen it, pointing to future work. - It found that the Value Embeddings really like regularization and I wasn't applying any (oops). - It found that my banded attention was too conservative (i forgot to tune it). - It found that AdamW betas were all messed up. - It tuned the weight decay schedule. - It tuned the network initialization. This is on top of all the tuning I've already done over a good amount of time. The exact commit is here, from this "round 1" of autoresearch. I am going to kick off "round 2", and in parallel I am looking at how multiple agents can collaborate to unlock parallelism. github.com/karpathy/nanoc… All LLM frontier labs will do this. It's the final boss battle. It's a lot more complex at scale of course - you don't just have a single train. py file to tune. But doing it is "just engineering" and it's going to work. You spin up a swarm of agents, you have them collaborate to tune smaller models, you promote the most promising ideas to increasingly larger scales, and humans (optionally) contribute on the edges. And more generally, *any* metric you care about that is reasonably efficient to evaluate (or that has more efficient proxy metrics such as training a smaller network) can be autoresearched by an agent swarm. It's worth thinking about whether your problem falls into this bucket too.

OK, well. I ran /autoresearch on the the liquid codebase. 53% faster combined parse+render time, 61% fewer object allocations. This is probably somewhat overfit, but there are absolutely amazing ideas in this.

If you’re vibe-coding in Elixir, try these two tools in addition to Credo and Dialyzer. I use them in every new project. github.com/dannote/ex_dna github.com/dannote/ex_slop

It started as a Claude Code project, but the rules and guardrails now work with any terminal AI — including @opencode with local @ollama models. No cloud API required. Same safety rails, same server memory, same OS auto-detection. Just git clone, launch your preferred AI tool, and go. Works with Linux, FreeBSD, and macOS — remote over SSH or locally. github.com/wintermeyer/he…

New in Heinzel: team mode. Born from real client work — share server memory, changelogs, and network topology via Git while keeping SSH credentials private. Session locking prevents two admins from stepping on each other's toes. Plus: layered rule overrides so you can customize without merge conflicts on git pull. github.com/wintermeyer/he…

There should be a way for a github issue opener to offer to foot the bill for the associated tokens while the author actually drives it. I'm just paying $5 a pop to fix someone else's problem sometimes but I'd rather do it then have them vibe code a fix (depending on the person).

What is the difference between "Nah, we straight" and "No, we are straight"? I am curious about it. Hard for a non native speaker to figure this out. See this podcast transcript about @BarackObama using the subtle difference. npr.org/transcripts/10…

@jnunemaker impeccable.style is a really good set design wise.