Mihai Chirculescu
@m_chirculescu
Followers
579
Following
8K
Media
142
Statuses
2K
You can just do things. My DMs are open: reach out!
Joined November 2020
🚀 Launching backtrack_sampler - for experimenting with custom LLM sampling strategies that can rollback and regenerate tokens. It works with both @ggerganov 's llama.cpp (GGUF files) and @huggingface 's transformers.
github.com
An easy-to-understand framework for LLM samplers that rewind and revise generated tokens - Mihaiii/backtrack_sampler
5
22
168
My friend’s hiring a software developer - fully remote, european timezone. Greenfield project. Must have experience with AI tools & building MVPs. JD: Contact:
drive.google.com
5
0
9
RT @AI_AlibabaInt: Introducing Ovis 2.5 - our latest multimodal LLM breakthrough!.Featuring enhanced visual perception & reasoning capabili….
0
36
0
Programming has already changed, anon. You just haven’t been paying full attention.
I never used git worktrees until I started using Claude Code and now I want to use them all the time.
0
0
1
That sounds great and I'm hyped. But why not release version 1 now, and version 2 later?. It's been four months and there's still no new open-source model. Frankly, this just feels like stalling.
we are going to take a little more time with our open-weights model, i.e. expect it later this summer but not june. our research team did something unexpected and quite amazing and we think it will be very very worth the wait, but needs a bit longer.
0
0
2
There's also ms-swift, which is surprisingly easy to use. For some reason, it's unknown to "westerners" :).
github.com
Use PEFT or Full-parameter to CPT/SFT/DPO/GRPO 500+ LLMs (Qwen3, Qwen3-MoE, Llama4, GLM4.5, InternLM3, DeepSeek-R1, ...) and 200+ MLLMs (Qwen2.5-VL, Qwen2.5-Omni, Qwen2-Audio, InternVL3, Ovis2.5, L...
This is currently the best (and afaik, the only) library that works well with Qwen2.5-VL models. reducto/RolmOCR was trained using LLaMA-Factory. It's still more or less buggy, and most of the discussions/solutions are still in Chinese.
0
0
1
Apparently, books like "Drive" or "Zero to One" are meant to be skimmed first and only then, if you really feel you need to, read word by word. Is this common knowledge in the U.S.?.
1
0
4
🤔🤔🤔🤔OpenAI could easily add a line of code to replace em dashes, but they want to know which text not to train on in the future. 🤔🤔🤔🤔.
1
0
4
Mobile devs, what's the equivalent of a hotkey on mobile?. @grok says swipe gesture patterns as 'hotkeys' aren't supported on iOS :/.So how do you trigger actions on mobile similarly to how AutoHotKey or AutoIt works on Windows?.
2
0
0
My hot take is that it's so much harder to become an indie hacker now than it was a few years ago. That window of opportunity has passed. You're better off deep learning what you believe will still be relevant in the future (in my case, that's the llama.cpp codebase).
If you're interested in doing the nomad coder side gig thing, then listen to this man please!:.
1
0
3
A coding agent that slows down while coding and live streams the code it changes and explains why via audio.
0
0
1
Writing software has really changed. On that note, there should be an app with a UX where the tool does NOT immediately execute commands, but instead asks the user things like, 'What about security? Doing X would cause Y', or 'Wouldn't this conflict with your other request?'.
0
0
1
I love @OpenRouterAI's concept and what they are building, but it grew too quickly, and I won't use it anymore in my future projects. I encounter issues too frequently, and there’s no paid support available. It’s just not production-ready. 👎.
1
0
0
RT @stablequan: i just tried @Alibaba_Qwen official coding agent, Lingma, and it is EXTREMELY good. You can choose Qwen3, Qwen3-thinking an….
0
2
0
RT @voooooogel: announcing an open beta of logitloom🌱, the tool i've been working on for exploring token trajectory trees (aka looming) on….
0
85
0