thesz

21 ms·

331.
by thesz 1y ago
> the assertion that "VLMs don't actually see - they rely on memorized knowledge instead of visual analysis". If that were really true, there's no way they would have scored as high as 17%. The ability to memorize lea
332.
by thesz 1y ago
I thought so about strong static typing looong time ago.
333.
by thesz 1y ago
Careless application of Gumbel Softmax would let your model to select same input for both inputs of a gate. Also, note that selecting a wire has all same problems as selecting an operation, such as vanishing gradient, but no apparent mitiga
334.
by thesz 1y ago
You may perform n parallel lookups into Bloom filter then use state machine table to perform k (most often k < n) actions, oftentimes in parallel (SIMD) too because it is just byte shuffle.
335.
by thesz 1y ago
> The tricky bit is that this line is not enforced by syntax or by the type system (unlike Rust, where you have a syntactic label `unsafe`). Safe Haskell: https://ghc.gitlab.haskell.org/ghc/doc/users_gui
336.
by thesz 1y ago
Far is for Windows, where "find" and "cat" are problematic, to say the very least. In that area, Far is game-changing. And even today I use Midnight Commander in Linux because I used to Norton/Volkov Commander and t
337.
by thesz 1y ago
Of about kefir percentage. Kwas has 0.5-1.5% ABV, kefir has 0.02-2.0% ABV.
338.
by thesz 1y ago
One can have obvious but repetitive optimizations with symbolic programming [1]. [1] https://arxiv.org/abs/1012.1802 Strange that AlphaEvolve authors do not compare their work to what is achievable by equality saturati
339.
by thesz 1y ago
> To address this challenge, we introduce a type-constrained decoding approach that leverages type systems to guide code generation. This should not work with type inference even at the level of C++ "auto x = " - "
340.
by thesz 1y ago
The issue is with indirect branches. Most branches are direct ones.
341.
by thesz 1y ago
This video is eleven years old. Contemporary C++ is very nice language, the only thing that I miss is a standalone REPL. But, I think, current code editors can be viewed as same thing.
342.
by thesz 1y ago
https://github.com/stassa/louise "Louise is a machine learning system that learns Prolog programs." For example, it can infer context-free grammar from half a dozen examples, one positive and several negative
343.
by thesz 1y ago
How can one quantumly differentiate between photon lost due to eavesdropping and photon lost due to some other error?
344.
by thesz 1y ago
These days one's program rarely does purely sequential random access. Usually, there are several threads or requests, all going into same data structure. In the extreme case where all memory accesses of a immensely massively parallel p
345.
by thesz 1y ago
What is the difference between channel error or distortion and eavesdropping?
346.
by thesz 1y ago
https://eprint.iacr.org/2015/1018.pdf "As pointed out in [57], there has never been a genuine implementation of Shor’s algorithm. The only numbers ever to have been factored by that type of algorithm are 15 and 21
347.
by thesz 1y ago
I usually have two meals a day, at evening and before sleep. Always after I am done with the work and training. I think the change you experience is not about stomach size, but ketosis. Ketones are such a great source of energy compared to
348.
by thesz 1y ago
Runner's high is due to release of endocannabioids: https://pubmed.ncbi.nlm.nih.gov/35081831/ Please, switch to resistance exercises. As these do not release endocannabioids but endorphines, they have appetite sup
349.
by thesz 1y ago
It is actually exercise and diet, that drive reward and hormones. For example, exercise has appetite suppression effect [1]. [1] https://med.stanford.edu/news/all-news/2022/06/anti-hunger-m... Diet has a
350.
by thesz 1y ago
> modify the node in-place I consider this anti-pattern [1]. [1] https://web.archive.org/web/20070417190836/https://www.eecs.... Authors found themselves fighting with control flow graph modif
351.
by thesz 1y ago
> like taking a forklift to the gym. First, you will have excellent forklift skills in the end. A real profession! Second, girls dig forklift operators or so I was told.
352.
by thesz 1y ago
> Well, I've encountered this use case a few times in Lisp:... > ...where a callback is used to collect various items. This can be and is achieved by simple SQL-like query. Filter (flat) set of nodes by integerness and
353.
by thesz 1y ago
https://news.ycombinator.com/item?id=42486610 Discussion of how Bloom filters made SQLite much faster.
354.
by thesz 1y ago
User interface widgets form a forest - windows contain layouts that contain widgets which also can be layouts, etc, etc. To implement Brown's algorithm to optimize class-based language models I had to implement a complex forest (DAG, a
355.
by thesz 1y ago
> It's a new technique they couldn't do before that helped crack the mystery. What about SAT-based solvers [1] for same problem? [1] https://ieeexplore.ieee.org/document/5361301 Would that techniqu
356.
by thesz 2y ago
https://arxiv.org/pdf/2503.02113 This paper shows that polynomials show most features of deep neural nets, including double descent and ability to memorize entire dataset. It connects dots there - polynomials there are
357.
by thesz 2y ago
This is fascinating! If I understand correctly, they approximate language of inputs of a function to discover minimal (in some sense, like "shortest description length") inputs that violate relations between inputs and outputs of
358.
by thesz 2y ago
> It is straightforward to write a fast parser generator for languages that require just one character of lookahead... Then you get VHDL. https://news.ycombinator.com/item?id=15017974 You need (at least an approximation
359.
by thesz 2y ago
As you mentioned "improvement of existing language," I'd like to mention that Haskell has green threads that most probably are lighter (stack size 1K) than goroutines (minimum stack size 2K). Haskell also has software transac
360.
by thesz 2y ago
Take a look at Satisfaction-Driven Clause Learning [1]. [1] https://www.cs.cmu.edu/~mheule/publications/prencode.pdf

More ›