Marius
@rasmus1610
Followers
860
Following
2K
Media
165
Statuses
3K
Applied AI, Neuroradiology, Software Development, Entrepreneurship
Düsseldorf, Deutschland
Joined June 2009
Building LLM apps feels different, right? Their opaque nature makes predicting outputs and behavior really tough. 🤔. So how do you move beyond guesswork and actually improve your LLM-based products reliably?
4
9
25
Was kinda down that a change in my dataset for one of the experiments resulted in worse results. Then I found a very stupid bug in my model. Now we are back to normal.
1
0
1
Kimi 2 is the strongest non-reasoning model on the @ArtificialAnlys benchmark. That's really impressive. The second strongest is also open-weight and also chinese.
0
0
2
Interesting post btw. Gave me a lot of food for thought. As an example: Looking at what I do through the ChatGPT or Claude interface regularly and see if I can systemize and solidify this workflow.
0
0
0
well, well, well . Look what we have here 😂
Why are legal services such a lucrative opportunity for AI?. ・ ⚖️ Legal work is programming with words, a natural fit for AI. ・ 🚀 The market pull is unprecedented, with sales cycles cut from months to weeks. ・ 💻 AI engineers have an unfair advantage, able to build the.
2
0
16
Performance of MIPROv2 vs. # of training examples (3 runs each). Interestingly, there's no real difference between 50 (maybe even 30?) and 100 training examples in my case. Might be interesting to establish "scaling laws" for different optimizers.
1
0
1
Ah man, the first experiments for a paper using DSPy are currently running. I'm so excited to see how they turn out :) . Watching DSPy optimizers is like watching the loss function going down back in the days of CNNs :D.
2
4
57
GitHub Copilot with Claude Sonnet 4 is very nice :) No need to use Cursor. I'm somewhat lucky that I get Copilot Premium for free - don't really understand why. I guess it's because of some OSS contribution.
3
0
3
Hadn't had much luck with parallelization. But that's more a problem of how I handle my SQLite cursor right now and not a problem of DSPy. Optimization can come later :).
0
0
0
If I don't procrastinate too hard, I should be able to deploy a DSPy-based version of this week. The core is there. Now I "just" need to productize it (streaming etc.).
1
0
1
LFM2-1.2b is very very usable on my pathetic 8GB M2 MacBook Air. Love it.
Try LFM2 on @lmstudio today!. Using the latest version of LM Studio, press Cmd / Ctrl + Shift + R and choose "Beta" channel. After updating llama.cpp, you can now run LFM2 quants!
2
0
9
So Kimi 2 through the official API is cheaper than Claude Haiku 3.5, the cheapest (and most outdated) model from @AnthropicAI .
0
0
1
There is so much interesting work to be done and I spend my time writing ethics proposals for using open datasets from Huggingface.
2
0
5
