
Co-Founder & CTO @theveloperseu | Engineering mind | Building things for a living 🤖
Joined June 2021
- Tweets230
- Following1K
- Followers75
- Likes2.8K
Bart retweeted
Dziś wydaję silnik basal 2.2 🇵🇱 z autokalibracją - model sam uczy się na Waszych prywatnych danych 🤯
Model decyzyjny jest użyteczny tylko wtedy, gdy „90% pewności” znaczy 90% na Twoich danych. Kalibrację dostarczaną z modelami dopasowaliśmy na naszych danych. Na Twoich model może być pewny siebie inaczej, nawet gdy jest równie trafny.
Dlatego w 2.2 wprowadzam coś co pozwoli ustawić kalibrację na Twoich danych - basal-calibrate. Bierzesz oznaczone przykłady ze swojego procesu i najpierw generujesz raport. Potem narzędzie dopasowuje kalibrację i wyznacza certyfikowane progi dla Twoich lokalnych danych 🤯. Bez udostępniania danych na zewnątrz, bez trenowania. Basal sam się skalibruje.
Sprawdziłem to na nowej dziedzinie, na której model nie był trenowany. Ustawienie „1%” z dostarczonego pliku przyjęłoby automatycznie co ósmą decyzję, a 8% z nich byłoby błędnych. Po kalibracji na danych użytkownika błąd kalibracji pytań choice spadł z 0,268 do 0,083. Na twierdzeniach verify model przyjął automatycznie 54,8% przypadków bez żadnego błędu.
W tym wydaniu także:
🔸 porty FP8 (vLLM) i GGUF (llama.cpp, Ollama) dla basal-omni-1.0-mini i max,
🔸 porty MLX dla Maca z Apple Silicon: mini, max i 1.1-max,
🔸 wkłady społeczności. Dziękuję:
- Piotr Durlej (eksport GGUF) @pdurlej - super, pomogło mi zoptymalizować eksporty do GGUF
- Kuba Sołtys (Google TPU) @QoobaSol - świetna implementacja na TPU
- Michał Kardyś (AMD Strix Halo) - dziękuję bardzo bo już nie miałem możliwości by się tym zając (sprawdziłem u mnie i działa świetnie) @kardysm
- lukaspoloki (basal-viz, podgląd każdego kroku decyzji).
- Paweł Kiszczak - eksporty do AppleSilicon MLX
Podziękowania za cały weekend bardzo miłych wpisów, liczne komentarze opublikuję dzisiaj wieczorem. Ile wystarczyło czasu to odpisałem. Później już pracowałem nad kolejnymi usprawnieniami. Dzisiaj czytam kolejne wpisy. Dziękuję.
Bart retweeted
basal-omni 🇵🇱 - TOP1 🤩🤩🤩 🌍 na świecie w modelach otwartych oraz TOP2 wśród zamkniętych i otwartych
Wg niezależnego benchmarku modeli system-one gdzie testowane są wszystkjie modele na świecie - JevBench (link do bencha w pierwszym wpisie). Czyli można? Można! Polskie AI 🇵🇱
basal wyprzedził model OpenAI, Microsoft oraz pierwowzór z TypeSafe (jev).
Thank you, @airesearch12 , for the independent evaluation on JevBench.
Bart retweeted
How often do ArXiv papers contain errors in the main theorems they claim to prove? Using GPT-6.1 Sol to identify such mistakes, it finds:
- 6% of papers before 08/2026 contain main theorems that are false.
- Only 1.5% contain similar mistakes in the last two months.
Bart retweeted
There has never been a better time to be a mathematician. So many new directions, open questions, and amazing tools. If you're bold and young, come join the party. We'll run this world like Lisan al-Gaib riding Shai-Hulud.
Bart retweeted
JPEG XL turned on by default has just started shipping in Chrome 155 - after ~30 years officially replacing JPEG for the next decades, ~2x reducing web sizes, adding HDR, transparency, progressive decoding for much faster www, also animations, lossless - should retire also ancient GIF, PNG, TIFF - finally reaching universality and much better compression.
developer.chrome.com/blog/jp…
Every software developer should own a potato PC. It's the easiest way to ensure your application is fast enough. Put your agents to work within it and watch the weight drop. There's no reason a machine from ten years ago shouldn't fly like superman.
When several people talk at once, a transcript can get messy fast.
Our new Nemotron 3 Diarization model tracks who spoke when, even when voices overlap. It handles up to eight speakers, has 100M parameters, and is now available on @huggingface 🤗
Bart retweeted
Making Startups Powerful: paulgraham.com/powerful.html
Bart retweeted
Some people seem to be surprised that an old geezer like me could adapt so quickly to the newest tech.
The thing is, I've been a programmer for over 60 years. And every single year brings another adaptation to new tech. Adapting to new tech is what we programmers do on a regular basis. Why would I stop now?
Bart retweeted
If you want to start a startup, don't learn "entrepreneurship." Learn how to build things. The hard part of startups is not "entrepreneurship" but product: to know what to build, and to be able to build it.
Bart retweeted
Will AI replace junior devs?
During ETHSilesia, @sir_mimok from @theveloperseu shared a different take:
Bart retweeted
Day 2 of ETHSilesia shifted the focus to the builders. The schedule was packed with workshops focused on practical development, followed by the hackathon opening.
From deep-diving into new tools to forming teams and starting the first lines of code, the transition from learning to doing is officially underway. The building has begun! 🚀
Bart retweeted
Anthropic gave 16 AI agents $20,000 and 2 weeks.
No human wrote a single line of code.
They built a program that can compile the Linux operating system from scratch.
Here’s why the coordination matters more than the result.
Imagine 16 workers renovating the same house simultaneously. Without rules, they’d paint over each other’s walls. Solution: before starting a job, each agent writes their name on a sign-up sheet. If someone already claimed “fix the kitchen,” you pick something else. Simple file in a shared folder. First come, first served.
Any problems? Yes.
The system broke when they hit one massive task: compiling the Linux kernel. Think of it like 16 cooks trying to fix one broken recipe at the same time. Each one tastes the soup, adds salt, but the next cook doesn’t know salt was already added. They kept undoing each other’s work.
The fix was clever.
They brought in a professional chef (GCC, an existing compiler) to cook most of the meal. The AI agents only handled random portions. When something tasted wrong, they could pinpoint exactly which portion the AI messed up. Some would call that cheating. But the agents decided to do this themselves, no human told them to. And honestly, knowing when to use existing tools instead of reinventing everything from scratch? That’s what good engineers do too.
What they built:
100,000 lines of code. Can compile Linux, PostgreSQL, Redis, FFmpeg, and even run Doom. 99% pass rate on standard tests. What they can’t do: “Hello World” sometimes fails.
Now the reality check.
The GitHub repo has 41 open issues. Titles include: “can’t even compile linux as the description says,” “big words with nothing to back them up,” and simply “F*** off.” People who actually tried using the compiler found parser errors, missing tests, and broken basic features.
The methodology, 16 agents coordinating via git locks, splitting work, self-organizing, that’s genuinely forward-thinking. But the output today remains a toy, not a tool professionals would trust. The approach is the future. The result isn’t there yet. And that gap is exactly where the interesting work happens next.
------
source: anthropic.com/engineering/bu…
Bart retweeted
The V is for Victory – but only if it's shared.
Relax, @sama @DarioAmodei ✌️
Bart retweeted
Today, we are at Crypto Talks Warsaw, discussing the future of the crypto market in an era of increasing regulation. We’re also exploring how the Baltic countries can cooperate to better serve and benefit their citizens.





















