Sign in

Maciej Gryka

@maciej.gryka.net
407 followers 212 following 167 posts

Software generalist with a particular interest in ML and AI. Soft spot for #ElixirLang and #Python. I enjoy challenges of early-stage startups. Ex-CTO, founded a couple companies, PhD in computer vision. More at maciej.gryka.net

PostsRepliesMedia
Maciej Gryka @maciej.gryka.net · 29/09/2026
Sometimes I read a paper and immediately find it super useful. Today was not one of those days. I thought "A Queueing-Theoretic Framework for Stability Analysis of LLM Inference with KV Cache Memory Constraints" arxiv.org/pdf/2605.04595 would be super useful, but it turns out...
arxiv.org
100
Maciej Gryka @maciej.gryka.net · 16/09/2026
MAXIMIZE
A box full of paperclips
000
Maciej Gryka @maciej.gryka.net · 16/09/2026
just let me read and summarize papers the whole day and I'll be happy
010
Maciej Gryka @maciej.gryka.net · 15/09/2026
I've recently been looking at multi-LoRA serving setups, it's a cool idea. Basically, if you have multiple adapters on top of the same model, you can serve them form the same instance, sharing all of the base-model weights. Conceptually that's pretty simple, but there's a lot of depth in...
arxiv.org
110
Maciej Gryka @maciej.gryka.net · 15/09/2026
The list of LoRA-capable models supported by @cloudflare-dev.bsky.social feels like visiting a museum. Pls add something less than 2y old 🙏
List of LoRA-capable models in Cloudflare docs. Showing Qwen2.5, Llama3.2, Gemma 2B among other, old models.
000
Maciej Gryka @maciej.gryka.net · 15/09/2026
Cool, just found that vLLM has a plugin system docs.vllm.ai/en/latest/de... Anyone has any experience with it?
000
Reposted by Maciej Gryka
apenwarr @apenwarr.ca · 22/08/2026
You don’t win an argument about definitions. You agree on definitions so that you can even have an argument
617519
Reposted by Maciej Gryka
Camille Fournier @skamille.themanagerswrath.com · 17/08/2026
we are quick to credit weird quirks for companies' success when really their success is usually some form of inventing a money printer and monopolizing it effectively and thus being allowed to indulge in weird quirks
3559
Maciej Gryka @maciej.gryka.net · 08/08/2026
this is how my face looks like after 2 weeks of relaxing holidays with 3 kids
030
Reposted by Maciej Gryka
Filippo Valsorda @filippo.abyssdomain.expert · 01/08/2026
Paging @masnick.com to the Streisand effect courtesy phone. An “engine problem“ is a complete non-issue and it was corrected in 24h! But Go2Sky sued @avherald.com! Now I’m reading about their 2016 incident they wanted to hide, and I’d never, ever fly an airline with this approach to transparency.
1177
Reposted by Maciej Gryka
Jay Tennier @jaytennier.bsky.social · 27/03/2026
Well I'm sad to say it's my last week at Rainforest QA - a bunch of us got let go so if anyone is looking for incredible remote frontend (React) or backend engineers (Ruby / Rails / Python) DM me and I'll be happy to introduce you.
021
Maciej Gryka @maciej.gryka.net · 11/03/2026
Woah I can really feel my decision-making ability drain in a new way now in the evenings. When most general-coding trivia is taken away, constantly facing actually challenging decisions is something else.
030
Maciej Gryka @maciej.gryka.net · 11/03/2026
so happy ultrathink is back ♥️🧡💛💚💙💜🩷❤️🧡💛
000
Reposted by Maciej Gryka
Hugo Baraúna @hugobarauna.com · 04/03/2026
OpenAI just launched an open-source agent orchestrator in Elixir! 🔥🔥🔥 github.com/openai/symph... #ElixirLang
2282
Maciej Gryka @maciej.gryka.net · 03/03/2026
So it turns out, if you've got an agentic workflow build out of specific subtasks, we can probably optimize it for you and drop your LLM bill by 10x. Small, distilled models often match and sometimes outperform much larger and more expensive models on well-defined tasks.
132
Maciej Gryka @maciej.gryka.net · 25/02/2026
We’ve just released a new version of our CLI docs.distillabs.ai/getting-star... It includes `model deploy local` to run your models locally and `distil model deploy remote` which streamlines deploying your trained models to our playground docs.distillabs.ai/deployment/i...
docs.distillabs.ai
CLI Reference | Distil labs | Documentation
Complete guide to the Distil CLI tool
000
Maciej Gryka @maciej.gryka.net · 19/02/2026
Go read this post by @natolambert.bsky.social www.interconnects.ai/p/open-model... and then ask me about how to put small models in your production systems
interconnects.ai
000
Maciej Gryka @maciej.gryka.net · 18/02/2026
When my agent is doing stuff and I accidentally interrupt it by pressing Esc with the terminal focused, I feel the need to apologize. Which is even dumber than usual, it doesn't perceive time!
000
Reposted by Maciej Gryka
Den @den.dev · 16/02/2026
This is excellent. I keep telling folks that the superpower of SOTA models is not that they can just generate endless amounts of code in whatever language you want, but also that they can act as a personalized engineering/coding teacher that can help you explain concepts in a way that YOU get.
1202
Maciej Gryka @maciej.gryka.net · 14/02/2026
I 💜 #ElixirLang! You 💜 #ElixirLang! I 💚 Python! You... will also 💚 after my talk in April!
Join me for my talk: "Drunk snakes, why Python and Elixir are a great pairing". ElixirConfEU, 23rd--24th April 2026, Málaga, Spain
260
Reposted by Maciej Gryka
P(aul) Frazee @pfrazee.com · 12/02/2026
If you don’t see the utility personally, I can get that. If I wasn’t coding I probably wouldn’t either. If you’re worried about its downsides, join the club. A lot of us are too. But no matter what your take is, take it seriously. This shit isn’t crypto.
228123
Maciej Gryka @maciej.gryka.net · 09/02/2026
We recently ran a project to help Knowunity cut their LLM bill by half. It’s pretty cool to beat a frontier LLM (Gemini) on all dimensions, which matter: cost, speed, and accuracy by trading off generalization, which is unnecessary here. More details www.distillabs.ai/blog/how-kno...
distillabs.ai
How Knowunity used distil labs to cut their LLM bill by 50%
We show how Knowunity use the distil labs platform to fine-tune and deploy their own models, significantly reducing their LLM costs.
101
Maciej Gryka @maciej.gryka.net · 07/02/2026
my coding agent just got confused by cron and timezones, i don’t need any more proof it’s sentient
070
Maciej Gryka @maciej.gryka.net · 06/02/2026
Elixir's operational simplicity (having e.g. a single repo & piece of infra, compared to more moving parts in other stacks) might either be a very big advantage or not matter at all in the new coding-agent world. I'm like 75% on the side of it turning out to be important, but who knows!
000
Maciej Gryka @maciej.gryka.net · 05/02/2026
I love having my biases validated dashbit.co/blog/why-eli...
dashbit.co
Why Elixir is the best language for AI - Dashbit Blog
A recent study by Tencent showed that Elixir had the highest completion rate across models when compared among 20 different programming languages. In this article, we explore the reasons why that may ...
032
Maciej Gryka @maciej.gryka.net · 05/02/2026
those H100s really are nice, huh
000
Reposted by Maciej Gryka
Desmond Elliott @delliott.bsky.social · 04/02/2026
📢 I am hiring a highly-motivated Ph.D student at the University of Copenhagen to work on tokenization-free NLP. Read our previous work in this topic: aclanthology.org/2025.emnlp-m... aclanthology.org/2023.emnlp-m... openreview.net/forum?id=FkS... Apply by March 8: employment.ku.dk/phd/?show=1563
A photograph of sunny Copenhagen in the summer!
0209
Maciej Gryka @maciej.gryka.net · 04/02/2026
Fun demo we shipped this week: AI slop detector running locally in a browser extension! github.com/distil-labs/... The utility of the tool itself is questionable :) But it was super fun to build! I was hoping to get something working in a browser for a while, and this shows how :)
github.com
GitHub - distil-labs/distil-ai-slop-detector: Detect AI-generated text locally in your browser
Detect AI-generated text locally in your browser. Contribute to distil-labs/distil-ai-slop-detector development by creating an account on GitHub.
020
Maciej Gryka @maciej.gryka.net · 02/02/2026
> It would be satisfying if humanity’s ambassadors to distant galaxies turn out to be humanoid robots made by a corporate successor to Twitter Inc. For the love of everything that’s good @matt-levine.bsky.social
000
Reposted by Maciej Gryka
apenwarr @apenwarr.ca · 27/01/2026
And here you go! Definitely doesn't fix the entire state of LLM protocols... yet. But we're on our way. bsky.app/profile/apen...
011
Maciej Gryka @maciej.gryka.net · 26/01/2026
tailscale figuring out how to make something sensible out of the current state of all the llm apis wen
100
Maciej Gryka @maciej.gryka.net · 21/01/2026
We released a Claude Skill allowing you to very easily get distilled small models, which can run locally on you machine! More here www.distillabs.ai/blog/train-y... and if we give you free training credits when you sign up :)
distillabs.ai
Train your SLM with distill-cli Claude Skill
Train a custom Text2SQL model by chatting with Claude and the Distil Labs skill. no ML expertise, no data labeling, just a conversation and a few examples.
000
Maciej Gryka @maciej.gryka.net · 14/01/2026
I enjoy trying to cram magical Claude words into actual sentences "what do you ultrathink about this?"
100
Maciej Gryka @maciej.gryka.net · 12/01/2026
Playing with @fly.io 's sprites by creating Claude skills to set up sprites for LiveView development. Adding Tidewave was super easy, I guess I should try LiveDebugger next. Many disposable Phoenix apps, here we come!
040
Maciej Gryka @maciej.gryka.net · 27/12/2025
asking Claude Code to run Igniter to install Tidewave to use Claude Code feels Ouroborosy #ElixirLang
140
Maciej Gryka @maciej.gryka.net · 09/12/2025
If you're looking into fine-tuning Small Language Models, we've benchmarked a bunch of them to see which perform best. www.distillabs.ai/blog/we-benc... TL;DR The Qwen3 family is the best all-rounder, while small Llamas show the biggest improvements after fine-tuning.
Chart showing results performance fine-tuned `Qwen3-4B-Instruct-2507` against its base version and a larger model (GPT-OSS-120B). The small model fine-tuned for the specific task matches or exceeds the large, generic model performance in 7/8 tasks.
000
Maciej Gryka @maciej.gryka.net · 16/10/2025
we just released small-and-mighty distilled PII reduction models, check them out github.com/distil-labs/... you can run them locally (1B!) and they work as well as much bigger generic models; SLMs FTW!
github.com
GitHub - distil-labs/Distil-PII
Contribute to distil-labs/Distil-PII development by creating an account on GitHub.
030
Maciej Gryka @maciej.gryka.net · 09/10/2025
I'll give a demo at the Berlin Elixir Meetup tonight! www.meetup.com/elixir-berli... (last minute change) Come say hi if Elixir and/or AI is your jam. I'll show how to use distillation (featuring distil labs ofc) to make exmeralda.chat run on a smaller model. #myelixirstatus
071
Maciej Gryka @maciej.gryka.net · 01/10/2025
Over the last couple of weeks we've been working on adding tool calling support to our distillation pipelines. Now you can use it! As a fun demo, we trained a model to help you remember git commands and it's only 3B so you can run it locally! www.distillabs.ai/blog/gitara-...
distillabs.ai
Gitara: How we trained a 3B Function-Calling Git Agent for Local Use
How we created a local tool-calling language model to turn plain-English language questions into git commands with the accuracy of a cloud LLM. You can check it out in our GitHub repo or get the model...
010
Maciej Gryka @maciej.gryka.net · 16/09/2025
If you're skeptical of Big AI and want to learn more about Small AI, check out the blog post our CTO, Jacek, just published. Goes through how we can help you turn your big, expensive, small models into beautiful, small, efficient ones www.distillabs.ai/blog/small-e...
distillabs.ai
Distil Labs
020
Reposted by Maciej Gryka
ElixirBerlin @elixir.berlin · 05/09/2025
In September, we are hosting Elixir Berlin as part of the #globalelixirmeetups. Join us and other Elixir developers around the globe to celebrate this week! 📅 Hope to see you there! 💜 #ElixirLang www.meetup.com/elixir-berli...
meetup.com
Sep 25th, [In-Person] Elixir Meetup #112 - Global Elixir Meetup Edition, Thu, Sep 25, 2025, 7:00 PM | Meetup
👋 Hallo und 💜-lich Willkommen to the Elixir Berlin Meetup. This month we are joining Global Elixir Meetups! We will have three talks scheduled at a new first time locat
074
Reposted by Maciej Gryka
Simon Willison @simonwillison.net · 23/06/2025
Phoenix.new by @fly.io is an interesting new entrant into the prompt-driven app development space - it can spin up a full Elixir/Phoenix real-time web app in a neat browser-based IDE, testing via a headless browser while it writes the code simonwillison.net/2025/Jun/23/...
simonwillison.net
Phoenix.new is Fly’s entry into the prompt-driven app development space
Here’s a fascinating new entrant into the AI-assisted-programming / coding-agents space by Fly.io, introduced on their blog in Phoenix.new – The Remote AI Runtime for Phoenix: describe an app in …
3484
Reposted by Maciej Gryka
Peter Ullrich @peterullrich.com · 20/06/2025
Oban is probably the greatest library I have ever used, especially with Oban.Pro on top. The value you get for free - or for a very reasonable amount - is mind-blowing.
1318
Maciej Gryka @maciej.gryka.net · 13/06/2025
Claude just looooooves monkey-patching all over Python tests even though the rules tell it not to do that.
020
Maciej Gryka @maciej.gryka.net · 27/05/2025
Has there even been a better/worse timing of two papers published the same day? May 26th 10am: x.com/xuandongzhao... "Learning to Reason without External Rewards: LLMs can learn complex reasoning without access to ground-truth answers, simply by optimizing their own internal sense of confidence."
x.com
100
Maciej Gryka @maciej.gryka.net · 27/04/2025
Truly an excellent name. I don’t think I’ve ever encountered a better project name.
HN headline: Shardines: SQLite3 Database-per-Tenant with ActiveRecord
060
Maciej Gryka @maciej.gryka.net · 17/04/2025
In case you doubted "all devs will become managers", I'm currently asking Claude via the CLI to plan and implement a bunch of changes and then Gemini via Windsurf to do code review of the changes. My first thought was "ROBOT FIGHT", but actually they're cordial and take each other's feedback well.
010
Maciej Gryka @maciej.gryka.net · 14/04/2025
Woah, looks like GPT 4.1 is the first model that doesn't pepper the code with gratuitous comments every 3 lines.
000
Reposted by Maciej Gryka
Anne Applebaum @anneapplebaum.wsocial.eu · 08/04/2025
I am also the author of a history of concentration camps in the Soviet Union, and I concur
633817827
Maciej Gryka @maciej.gryka.net · 07/04/2025
Still love Elixir, but also getting back into some Python feels like putting my old, comfy slippers back on :)
010