AI neolab focused on open source research
Warsaw
Joined February 2022
- Tweets14.9K
- Following5.8K
- Followers1.1K
- Likes88.5K
Fabryka AI retweeted
You can now train your own Decision model with our free notebook! 💡
Qwen3.5-4B will generate decisions instead of text on just 8GB VRAM locally.
Learn to data prep (state, questions, gold answers), train, serve.
Notebook: colab.research.google.com/gi…
Guide: unsloth.ai/docs/basics/train…
You can now train your own Decision model like Jev locally!
We increased Qwen3.5 0.8B’s aggregate accuracy from 20.7% to 74.3% across 3 decision benchmarks - on just 4GB VRAM.
Turn any LLM like Qwen3.8, Gemma 4 into decision models with our open-source Unsloth repo.
We fine-tuned with a Clef head using Unsloth and LoRA (r=64) for one epoch, increasing downstream accuracy from 30–37% to 78%.
GitHub: github.com/unslothai/unsloth
Guide and Notebooks: unsloth.ai/docs/basics/train…
Fabryka AI retweeted
If your problem is that you're over-controlled, you need deconditioning. Therapy, raves, meditation, Eastern religion and drugs (within reason) will work wonders for you.
If you lack self-control, you need loving discipline, a relationship with a role model you genuinely look up to, and a project in which you give your energy over to something bigger than you. Spiritual formation.
If you swing wildly back and forth between rigid self-coercion and exhausted hedonistic binges, and/or you feel haunted by shame and guilt, you will likely benefit from seeing a professional. I recommend a PsyD or an older, very experienced spiritual teacher.
Fabryka AI retweeted
“Smart people are everywhere, but those who can endure extreme monotony and deliver excess over the long term are as rare as phoenix feathers and unicorn horns.”
Rare but precise advice for winning at work.
Alexandr Wang 在谈到“如何拉开与 99% 人的差距”时，给出了极其朴素甚至残酷的答案：
顶层竞争根本不拼智商，拼的是极端执行力的非对称倾斜。
他的三条核心准则：
1. 绝对聚焦，压榨认知带宽： 绝大多数人的忙碌只是“低密度伪勤奋”。真正的专注是把全部精力砸在唯一的核心支点上，拒绝任何无关干扰。
2. 把“过度交付”做成护城河： 行业标准要求走 1 英里，你就走完 10 英里。当所有人都按合同完成 100% 时，常态化交付 200% 的人会直接卷走全部声誉与溢价。
3. 带靶向的超额输出，拒绝无目的内耗： 努力不是盲目自我感动，而是针对具体目标的持续过载测试。把每一次交付都当成生死决战来打。
商业世界最公平的一点在于：
聪明人到处都是，但能忍受极度枯燥并长期超额交付的人凤毛麟角。
执行力溢出，就是最暴力的护城河。
Fabryka AI retweeted
Met a close friend who’s also a seed stage founder today..
They have been growing but runway is dwindling and he wanted advice on what he should do next. Stellar customers, getting close to a $1M in revenue. Not a rocket ship, but not bad at all. He thinks he’s a bit early since the models haven’t caught up to the capabilities he wants but he also really feels the opportunity cost.
I have this same exact conversation at least 2x a week.
In this moment, you only have essentially three options:
1) if you have intense belief in the company, be a cockroach and get to default profitable. once the capabilities get there, you feel the demand, then go raise a large round and go for the jugular.
2) find an orthogonal area that is much more relevant and growing like crazy. ride the coat tails of new existing capabilities and push hard to capture them quickly using unfair advantages. and if you don’t have that, then..
3) sell / get acquihired and join a far more ambitious company that has figured out a lot of the things you haven’t. And then get back to starting a new company 1-2 years later when you find the next thing you want to build.
It’s hard but honestly those are the three options you have. And it really depends what category you’re in. The bar continues to get harder every month on how to cross the chasm from seed -> series A. The old school metrics are just not relevant anymore and in this consensus market there’s only a few categories where there’s tremendous pull on the VC front.
You could lament that this is the state of the market, but it’s better to know it’s happening than be oblivious on what is actually happening.
I get to share this directly which makes me uncomfortable but at least someone has to say it out loud 🙈
Fabryka AI retweeted
At Proximal, we believe that building our own research infrastructure is crucial to design the highest-quality training data
Today, we are sharing more information about our internal post-training stack
Fabryka AI retweeted
There is an optimal amount of courage x intelligence. You can't be so intelligent that it is worth it for you to spend time thinking. You need to be just a little retarded. And have unrealistic expectations of your ability. You also need to know that you're retarded, so you use simple stupid tricks to meet the expectations of your abilities
Fabryka AI retweeted
Never underestimate how much alpha you can generate by asking yourself what you'd do if the stakes were raised to infinity.
"The approach you’re taking right now – would you still use it if your life were on the line? Is there anything else you’d do to increase your chance of living?"
Fabryka AI retweeted
sharing my thoughts on our recent fundraise.
we did alright for a couple of goofballs
nousresearch.com/a-note-on-o…
Fabryka AI retweeted
rly interesting how AI usage inflects exponentially above certain viable thresholds:
1. ChatGPT can chat -> fastest growing product of all time
2. Anthropic models begin coding well -> massive exponential growth for AI coding
3. Models start doing certain kinds of research->
until recently, ai coding was progressing in a quick but predictable fashion. ai math over the past 6 months has progressed in a way that was much faster and much less predictable.
this hasn't happened in coding. we'll know it's happened when we see ai code something that is as impressive as the Navier Stokes solution. i don't exactly know what that means but i think i can say that we haven't reached that yet.
Fabryka AI retweeted
Amdahl: “My law states the overall performance improvement gained by optimizing a single part of a system is limited by the fraction of time that the improved part is actually used.”
My buddies in 2018: “ain’t no laws when your drinking claws”
Who you gonna believe?
Fabryka AI retweeted
wise words from the best systems engineer I've worked with:
"two things that make code actually maintainable:
1. reduce the layers a reader has to trace
2. reduce the state a reader has to hold in their head"
applies to every codebase. always.
Fabryka AI retweeted
I’m increasingly convinced that people are spending most of their time building in agent terminals, and very little time actually using what they produce.
Lot of stuff is broken from the start, some stuff that worked before is now broken in weird ways.
Fabryka AI retweeted
Spędziłem już mnóstwo godzin oglądając kilkadziesiąt implementacji klasyfikatorów “System-1”. Im więcej implementuję, dodaję usprawnień, różnych tricków, spędzam przy tym czasu tym:
a. brakowało takiego czegoś w systemach przetwarzających tekst (oczywiście można było to symulować bert like albo llm). To inna filozofia. Reranker na basal? Klasyfikacja tekstu? Fantastycznie działa w systemach RAG. Kilkadziesiąt top chunków sklasyfikowanych w kilka sekund.
b. krótko mówiąc, bo wieczór sobotni a przede mną sporo jeszcze by jutro pokazać co wnoszę do polskiego ekosystemu modeli językowych, jest to najlepsza rzecz jaką wynyślono w 2026r dzięki @typesafeai i jev. Bez dwóch zdań.
Jutro rano, 9:00, basal-1.5.
Fabryka AI retweeted
if you are pondering to get into the core “training” part of language modelling, i would highly recom to just get started with technical reports of open source labs like deepseek, ai2, nvidia nemotron, arcee etc. olmo 3 is amazing to get started with!
Fabryka AI retweeted
Reasoning from scratch, round number 6!
An introduction (and implementation) of Reinforcement Learning with Verifiable Rewards (RLVR) and Group Relative Policy Optimization (GRPO).
00:00 Introduction
01:54 What makes a reasoning model different?
04:25 Reasoning traces and model capability
08:29 Accuracy and format rewards
11:34 Aha moments and DeepSeek-R1 training
14:41 Reasoning effort and answer length
18:38 RLHF and RLVR
23:04 GRPO vs. PPO
26:40 GRPO explained with a cooking analogy
31:43 The KL term and simplified GRPO
35:04 Loading the pretrained model
36:07 Loading the MATH training data
39:26 Sampling model responses
46:30 Computing verifiable rewards
49:55 Computing advantages
51:54 Token and sequence log probabilities
55:29 Implementing sequence log probabilities
57:37 Fixing the inference-mode error
1:02:24 Computing the GRPO loss
1:04:37 Putting the GRPO step together
1:09:19 The GRPO training loop
1:12:57 Training settings, logging, and checkpoints
1:17:24 Running training and inspecting outputs
1:19:28 Loading and evaluating checkpoints
1:22:33 MATH-500 results and training stability
1:24:05 Memory requirements and next steps
Fabryka AI retweeted
7. If nobody fails in practice, the practice was written wrong.
Don't be afraid to let them fail.
That's where they learn everything they need to learn about how to get better.
Fabryka AI retweeted
"And how do you know these environments are built on garbage? Aren't they filled with thousands of tasks and rubrics?"
"I read them."
"You read them? No one reads them. Only the researchers who put them together read them."
"I don't think they even read them. I don't think they even know what they made."
"But the evals are going up."
Fabryka AI retweeted
A 9B model could reach GPT-5 / Opus-4.5-level reasoning 🧠, nearly free ⚡🆓, on your own desktop 💻.
No additional post-training. Much less jagged 🧩 generalization.
This was my intern Panagiotis’ summer project. Still a long way to go on engineering, but I strongly believe this direction will keep getting better.
Small models may have a lot more intelligence 🧠 inside them than we think.
arxiv.org/abs/2609.38104
