Dal Rupnik (@TheLegoless) header
Dal Rupnik (@TheLegoless) profile picture

CTO at @speechblubs · Developer · Speaker · Mentor · iOS · Machine Learning · AI · Vision

Slovenia
Joined February 2011
It's quite clear that all current open source LLMs are very benchmaxxed. LegoBench easily proves this: legobench.net/?model=deepsee… Best score of an open source model is DeepSeek V4.1 Flash 1017 at 12.7%! GPT-6 Astra scores above 80%!
1
88
Even latest Qwen 4 Max, can barely do anything.
26
Gemini 4 Carbon's 3D understanding of LEGO models has certainly been improved compared to Argon. However it still fails to beat Astra or Fable 5.2. LegoBench results below: legobench.net/?model=claude-…
1
101
Anyone else got Cursor Grok reset? Thx @elonmusk
59
Week 3: So it's not exactly an open source version of the existing software, but it's an open source version of patches you need to run to get Skyrim running on Apple Silicon. Details on my blog: unifiedsense.com/blog/skyrim…
59
Interesting how benchmaxxed models perform bad on LegoBench, how are they outperforming Opus and Astra on almost all benchmarks, but fail so horribly on something so simple as putting bricks together? AGI is here. Kimi K3.1 Pro barely has a clue. legobench.net/runs/20261023-…
62
In addition to all of the other ports, Skyrim also runs natively on Apple Silicon. iPad, iPhone and Mac. :)
83
For my slovenian followers, I made a couple of videos explaining AI concepts. First one is about AI Agents. Let me know your feedback! youtube.com/watch?v=TDmxmT5E…
112
A little bit late, but still - week 2 in Open Source: My version of Adobe Creative Suite. This would be a much more exciting if ArtCraft wouldn't have been released few weeks ago. Tauri 2, Rust and WASM engine. Covering most of their apps and a little more. unifiedsense.com/blog/monet-…
1
3
106
Is DGX Station worth it? Can it run v4.1 Flash?
54
I had used up my Codex Pro 500 plan yesterday to 100%. Just now I noticed I still have 20% left. Did anyone else notice this?
1
4
224
We now have state-of-the-art pre-trained base models in the 3T-parameter weight class, built entirely in-house at MAI. Post-training results coming soon -- stay tuned! Lots of details in our technical blog post. microsoft.ai/news/frontier-p…
154
361
163
3,522
1,387,926
Dal Rupnik retweeted
please dont be benchmaxed please dont be benchmaxed please dont be benchmaxed
Introducing Muse Wonder, our new family of frontier models with state of the art performance for personal and professional agentic work.
44
4
396
16,742