
Optimization and constraint programming (https://teapawt.net/t.co/P2zpmNexcx). Compilers and optimization at @SambaNovaAI
Stockholm
Joined September 2008
- Tweets786
- Following579
- Followers103
- Likes9.7K
Fun project this week. I wanted to play some Sudoku puzzles I had generated for a paper. That turned into an exploration of how to generate several kinds of puzzles.
I wrote a blog post about it:
zayenz.se/blog/post/constraiβ¦
1
46
At ModRef 2026, I have three different papers: one about modeling seating planning in MiniZinc for a wedding, one about generating Sudoku instances, and one about writing propagators for a min distance constraint.
1
75
For sudoku, the main questions is about how hardness changes when the size of the Sudoku board changes. zayenz.se/blog/post/scaling-β¦
1
26
Finally, for the minimum distance constraint (used mainly for the p-dispersion problem), the main way to model the problem leads to quadratic number of variables, propagators, and domain elements. Writing propagators for the constraint solves that problem. zayenz.se/blog/post/minimum-β¦
15
We released Gecode 6.4.0 (gecode.dev) today. Really good to finally have a new release officially tagged, and a new release process that makes it reasonable to do releases more often.
3
3
398
I have some additional comments on the releases here zayenz.se/blog/post/gecode-6β¦
15
With this maintenance release done, there are a lot more fun stuff in the pipeline, feels good to get back into improving Gecode.
27
To get the release out, one of the sticking points has been build systems and processes. Solving these kinds of issues is one of my favorite use-cases for AI-supported coding (mainly gpt in my case, with a lot of 5.6 Sol usage since the release).
29
A while ago @standupmaths made a video about the Partridge Packing Problem. Suitably nerd-sniped I went and modeled it in MiniZinc and ran lots of tests. I was surprised that one of the standard tricks for packing packing problems was not that useful.
zayenz.se/blog/post/partridgβ¦
1
87
Iβve had a rotating workforce scheduling example Iβve used in various presentation on MiniZinc for a while. Wrote it up as a blog post to share more widely, and ran a set of benchmarks as well. zayenz.se/blog/post/rotatingβ¦
1
54
I wrote a blog post about how to check for overlap of intervals. The main motivation is to show that sometimes when looking at cases, checking the negation is a lot easier. zayenz.se/blog/post/how-to-cβ¦
1
42
A while ago I wrote a post about solving the LinkedIn Queens problem using MiniZinc and now I have a follow-up where I benchmark that model with different back-ends on a larger set of instances: zayenz.se/blog/post/benchmarβ¦
1
44
I wrote Solving LinkedIn Queens with MiniZinc (zayenz.se/blog/post/linkedinβ¦) as a way to show how to solve the Queens puzzle on LinkedIn (also known as Star Battle) using the MiniZinc language.
1
53
I think this is a very real possibility, but I also wonder if we as programmers overestimate peopleβs willingness to type in mid to long instructions (and using voice instead doesnβt work in a lot of crowded situations).
Products with extensive/rich UIs lots of sliders, switches, menus, with no scripting support, and built on opaque, custom, binary formats are ngmi in the era of heavy human+AI collaboration.
If an LLM can't read the underlying representations and manipulate them and all of the related settings via scripting, then it also can't co-pilot your product with existing professionals and it doesn't allow vibe coding for the 100X more aspiring prosumers.
Example high risk (binary objects/artifacts, no text DSL): every Adobe product, DAWs, CAD/3D
Example medium-high risk (already partially text scriptable): Blender, Unity
Example medium-low risk (mostly but not entirely text already, some automation/plugins ecosystem): Excel
Example low risk (already just all text, lucky!): IDEs like VS Code, Figma, Jupyter, Obsidian, ...
AIs will get better and better at human UIUX (Operator and friends), but I suspect the products that attempt to exclusively wait for this future without trying to meet the technology halfway where it is today are not going to have a good time.
39
With all the Creepy Bob videos that have been doing the rounds, it was inevitable that it would be a meme format. I think Theo nailed it, and with the twist that he has sane views.
1
1
177
It is fun when we have announcements like this showing what we can do at SambaNova. The speed we can run R1 is excellent, and reducing the footprint to serve these models significantly is so important to make it feasible to use at scale.
ποΈβ‘οΈKa-chowβ‘οΈ The fastest DeepSeek-R1 671B on SambaNova Cloud β running at 198 t/s!
β
3X faster & 5X more efficient than the latest GPUs
β
Running on 1 rack efficiently (16 RDUs)
β
Hosted in secure US data centers
β
100X the global capacity by the end of 2025
@deepseek_ai #AI
135
The CACM analysis/critique (cacm.acm.org/research/reevalβ¦) of a 2021 Nature paper from Google on using RL for chip design is really interesting and aligns with my suspicions.
A slightly more spicy take on it from Puget:
Replying to @SumitGup
There wasn't much excitement in the optimization community as we were used to AI folks claiming stuff without comparing to state of the art.
Like the recent paper on bin packing.
1
106



