HNHacker News
TopNewBestAskShowJobs

mzl

1,920 karma · joined April 27, 2010

submissionscomments
mzl··on Shai-Hulud malware attack: Tinycolor and over 40 NPM packages compromised
Manual verification of releases and chain-of-trust systems help a lot. See for example https://lucumr.pocoo.org/2019/7/29/dependency-scaling/
mzl··on Many hard LeetCode problems are easy constraint problems
The MiniZinc Coursera courses (https://www.minizinc.org/resources/#courses-title) are useful to get a good basis for understanding constraint programming. It’s not a fast solution being full courses, but it is a very good resource.
mzl··on Many Hard LeetCode Problems Are Easy Constraint Problems
I get what you are saying, but this is not libraries that implement dynamic programming.

There are general libraries for dynamic programing that are interesitng, see for example https://didp.ai/

mzl··on JuLS – A Julia Local Search Solver
Would be interesting to see in next years MiniZinc challenge as a way to compare with other solvers.
mzl··on JuLS – A Julia Local Search Solver
Looks interesting.

Would it be possible to use it as a MiniZinc backend?

mzl··on Finnish City Inaugurates 1 MW/100 MWh Sand Battery
The point of these types of systems is that you can store energy when it is cheap to do so (there is an abundance of wind and/or solar energy) and use it later.
mzl··on Making games in Go: 3 months without LLMs vs. 3 days with LLMs
I've heard several talks about how some companies make AI systems that are designed to play as similar to human players as possible. This has been crucial for them in order to play-test levels in order to balance the game.

And note, this is not AI as in asking an LLM what to do, this is more classical machine learning and deep learning.

mzl··on Some thoughts on journals, refereeing, and the P vs NP problem
> > they show that you can use SAT to solve Sudoku

> Wait, did you mess up the direction here too, or am I confused? If you reduce problem A to B, then it means B is at least as hard as A, because solving it would solve A.

Using SAT to solve Sudoku is a reduction of Sudoku to SAT. The order of the problems names switches depending on how you write it.

mzl··on Gecode is an open source C++ toolkit for developing constraint-based systems (2019)
Adding to wavemode, the Sudoku example is a good starting point: https://github.com/Gecode/gecode/blob/master/examples/sudoku...

In the documentation (https://www.gecode.org/doc-latest/MPG.pdf) there are a lot of worked examples as well.

mzl··on Gecode is an open source C++ toolkit for developing constraint-based systems (2019)
Note, the release/6.3.0 branch is the current release.
mzl··on LLMs as Compilers
LLMs can be run in a mostly deterministic mode (see https://docs.pytorch.org/docs/stable/notes/randomness.html for some info on running PyTorch programs).

Varying the deployment type (chip model, number of chips, batch size, ...) can also change the output due to rounding errors. See https://arxiv.org/abs/2506.09501 for some details on that.

mzl··on LLMs as Compilers
That depends on the sampling strategy. Greedy sampling takes the max token at each step.
mzl··on LLMs as Compilers
Many compilers are not deterministic (it is why repeatable builds is not a solved problem), and many LLMs can be run in a mostly deterministic way.
mzl··on Huawei releases an open weight model trained on Huawei Ascend GPUs
Yes, as a reviewer of a paper I am assigned a number, and must by nominative determinism fulfill the expectations...
mzl··on Solving LinkedIn Queens Using MiniZinc
Thanks, luck of the draw ¯\_(ツ)_/¯

If you are interested in learning more about how to model problems with MiniZInc, there are several Coursera courses available: https://www.minizinc.org/resources/#courses-title

mzl··on How to Think About Time in Programming
I'd say, in general any time a time-value somehow originates as a part of an interaction with a human, that time-value carries with it the context of a timezone and the expectations of what that timezone means for that human.

For internal timestamps such as ordering events in a database, them UTC or something similar is nice. Bet the point then is that those values are not really meaningful or important in the analog world.

mzl··on How to Think About Time in Programming
As many others have said, time and calendars is messy, and there is often no correct solution but just a bunch of trade-offs. Jon Skeets Storing UTC is not a Silver Bullet (https://codeblog.jonskeet.uk/2019/03/27/storing-utc-is-not-a...) was very influential for me in realizing some of the subtleties in what a point in time means for a user, and how that should incluece the design of a system.
mzl··on Solving LinkedIn Queens Using Haskell
Directly modelling the unique solution property would be a quantified problem, essentially it would be "there exists a solution such that it is not the case that there exists a different solution". In principle, you can explode this into an enumeration of all O(n!) placements of queens and saying "either the placement is the same as the solution, or it is not a solution". That significantly increases the model size though.

For the symmetry, LinkedIn Queens generally do not have symmetric boards since that would imply more than one solution.

mzl··on Solving LinkedIn Queens Using Haskell
Generally what is needed is a subroutine that can tell you 1) if the problem has a solution, and 2) if the solution is unique (common requirement for puzzles like these). Using such a model as a sub-routine, a heuristic search can be done to gradually build up puzzles. If your solver technology of choice can handle quantified problems, that could be used to integrate those two problems into one, but that is quite a lot harder to to.

If the base solver you have is a system that can be run in various configurations with different levels of reasoning and assumption as well as a report on the amount of search needed if any, that can be very useful as a way to measure the hardness. In Sudoku as a Constraint problem (https://citeseerx.ist.psu.edu/document?doi=4f069d85116ab6b4c...), Helmut Simonis tested lots of 9x9 Sudoku puzzles against various levels of propagation and pre-processing as a way to measure the hardness of Sudoku puzzles by categorizing them by the level of reasoning needed to solve without search. The MiniZinc model for LinkedIn Queens (https://news.ycombinator.com/item?id=44353731) can be used with various solvers and levels of propagation as such a subroutine.

Now, for production-level puzzle making, such as what King does for Candy Crush, the problems and requirements are even harder. I've heard presentation where they talk about training neural networks to play like human testers, so not optimal play but most human like play, in order to test the hardness level of the puzzles.

mzl··on Solving LinkedIn Queens with SMT
You might be interested in looking at MiniZinc (https://minizinc.org/) which is an open source modelling language for combinatorial problems. The system comes from a constraint programming background but the language is solver agnostic can be used to compile into many different types of solvers.
mzl··on Solving LinkedIn Queens with SMT
SAT (satisfiability), SMT (satisfiability modulo theories), CP (constraint programming) and so on are not magic solutions. They are more or less really smart backtracking solvers that use lots of heuristics, deductions based on the constraints, learning, restarts, portfolios, and so on. A real benefit is that one can re-use the algorithms that someone else has developed and debugged.

For toy-sized problems such as the LinkedIn Queens, 8-queens, and 9x9 Sudoku more or less any competently implemented solver will work fine. As soon as the size of the problem starts to grow, the difference between a plain backtracking algorithm and a SAT/SMT/CP system.

mzl··on Solving LinkedIn Queens with SMT
In my view, using dedicated modelling languages for combinatorial problems is a great choice for getting models up and running quickly. Here is my take on how to model this problem using MiniZinc: https://news.ycombinator.com/item?id=44353731
mzl··on Solving LinkedIn Queens with SMT
Here is my variant of solving the problem using MiniZinc: https://news.ycombinator.com/item?id=44353731
mzl··on Pyrefly vs. Ty: Comparing Python's two new Rust-based type checkers
SOTA LLM model training costs a lot, yes. But fine-tuning and training of smaller models is a lot cheaper.

I've trained useful vision-models that delivered business value for industrial applications on a MacBook overnight.

mzl··on Why are 2025/05/28 and 2025-05-28 different days in JavaScript?
No, it is not. It is because converting to UTC loses information that can't be retrieved.
mzl··on Why are 2025/05/28 and 2025-05-28 different days in JavaScript?
No, but that is an even less likely situation, compared to the problems that the approach advocated for by Jon Skeet actually solves.
mzl··on Pyrefly vs. Ty: Comparing Python's two new Rust-based type checkers
I think you are underestimating the massive amounts of Python code that is built around these things. Also, a lot of businesses are not really interested in using an API for an LLM, instead they will modify and fine-tune their own models and deploy in their own data-centers (virtual or physical), and that means even more Python code.

Sure, a system that only relies on token factory LLM APIs can be written in any language, but that is not the full width and breadth of the AI hype.

mzl··on AI in my plasma physics research didn’t go the way I expected
Checking it shortly (I haven't seen the paper before) this seems to be a very good analysis of how results are reported specifically for medical imaging benchmarks.

As is often the case with statistics, selecting just a single number to report (whatever that number is) will hide a lot of different behaviours. Here, they show that just using the mean is a bad way to report data as the confidence intervals (reconstructed by the methods in the paper in most cases) show that the models can't really be distinguished based on their mean.

mzl··on LegoGPT: Generating Physically Stable and Buildable Lego
Or more generally the whole field of combinatorial optimization, of which metaheuristics is a (small) part.
mzl··on Modern LaTeX
I didn't know that, that is unfortunate to hear.
← PreviousPage 3 of 14Next →