#optimization
Logistic regression may be used for classification.
In order to preserve the convex nature for the loss function, a log-loss cost function has been designed for logistic regression. This cost function extremes at labels True and False.
The gradient for the loss function of logistic regression comes out to have the same form of terms as the gradient for the Least Squared Error.
More: https://www.baeldung.com/cs/gradient-descent-logistic-regression
If the iPhone has a Silicon Carbon Battery, and if all hardware components, especially Apple's latest Bionic chip and display, are designed to maximize battery life, and if all apps and iOS are heavily optimized for enhanced battery life, and if grayscale mode and dark mode are on at all times, how many days could the battery of this iPhone last on a full charge?
GEPA optimizes LLMs without costly reinforcement learning https://venturebeat.com/ai/gepa-optimizes-llms-without-costly-reinforcement-learning/ #AI #optimization #GEPA
GEPA optimizes LLMs without costly reinforcement learning https://venturebeat.com/ai/gepa-optimizes-llms-without-costly-reinforcement-learning/ #AI #optimization #GEPA
Parquet’s the cool kid in the data world. 
It’s sleek, efficient, and powers formats like Iceberg and Delta Lake. But few understand Parquet's storage magic. 
I wrote a blog that breaks it down in plain English. If you’ve ever wondered how Parquet squeezes more into less, this one’s for you.
LLM на прокачку: практический гайд по Alignment
Мы в Точка Банке делаем свою LLM. Чтобы она работала хорошо, недостаточно просто обучить её на куче текстов. Для получения осмысленного и предсказуемого поведения модели, нужен Alignment — дообучение с учётом предпочтений и ограничений. В статье расскажу, какие методы применяют в современных моделях, и как мы адаптировали их под себя.
NUMA Is the New Network: Reshaping Per-Socket Microservice Placement
We Hit 100% GPU Utilization–and Then Made It 3× Faster by Not Using It
https://www.daft.ai/blog/embedding-millions-of-text-documents-with-qwen3
New blog post, the third in a series on the development of an optimization based design tool for heat exchanger network synthesis (aka design):
Автоматизированная оценка стабильности скоринговых моделей на основе временных рядов метрик
Привет, Хабр! Меня зовут Зотов Глеб, я ML-инженер в команде скоринга в билайне. В статье расскажу о том, как не сойти с ума, мониторя десятки графиков вручную. Скоринговая модель может быть блестящей на этапе обучения, показывать отличные значения всех метрик на кросс-валидации и радовать бизнес на первых неделях после деплоя. Но вжух — и через два месяца валидационные метрики поползли вниз, отклонения по PSI зашкаливают, а product owner уже поглядывает в твою сторону с подозрением. Проблема? Проблема. Давайте разберемся, почему так происходит и как можно этого избежать.
I wrote a work-stealing task queue library for Rust! It's called takeaway, and I just published a version I think is ready for use. The only popular task queue library out there (for Rust) is crossbeam-deque; compared to it, takeaway provides a higher-level API with a lot more features. I wrote it as part of my very-very-WIP Rust compiler, which needed the unique feature of task prioritization; takeaway's since grown a lot, and manages competitive (if not better) performance to crossbeam-deque! You can find it at https://crates.io/crates/takeaway; I've also written a blog post about the design and implementation process, at https://bal-e.org/speed/krabby/takeaway. If you're writing a performance-intensive, task-based program in Rust, or if you're already using crossbeam-deque, please check it out.
Due to working on services that do quite a bit of dynamic protobuf marshalling and unmarshalling, this seems like an exciting project https://mcyoung.xyz/2025/07/16/hyperpb/
Как порядок свойств убивает JavaScript?
Привет, Хабаровчане! Во второй статье, хочу поделиться наблюдениями из документации V8 и немного нудной информацией для многих :)
Do you need better performance than what the standard #tidyverse functions have? {collapse} might be worth a look: https://sebkrantz.github.io/collapse/ #rstats #optimization
How We Made “JSON.stringify” More Than Twice as Fast, by (not on Mastodon or Bluesky):
mov rax, qword [r13 + 0x18]
add rax, 0x8
jns 0x14493
mov rcx, qword [r13 + 0x10]
mov qword [rcx + rax * 1], r15
mov qword [r13 + 0x18], rax
This beautiful sequence of instructions pushes an element into a custom-made interior-mutable Vec. The control variable in rax does all the heavy lifting:
- It has special values if the Vec is borrowed or has reached capacity, so I can drop to a slow path to handle either case.
- The add operation, which updates it to account for the element we're trying to insert, generates a condition flag which tells us whether to fall back to the slow path.
- It is also used as the offset for the element in the Vec (with the element size pre-multiplied), so only the base address needs to be loaded alongside it.
It took me 5 hours to put together over the last few days, with a record 47 unsafe blocks (rewrites and cleanups pending). It's led to a ~4% speed boost. I love doing things like this so much.
c'mon rustc/LLVM, this is 4% of my runtime for some reason, can you please just fold down this intricate overflow-checking arithmetic into ADD; JA, I promise it's exactly what I need...