My feed

There is a question quietly buried inside every task handed to a language model, and it rarely gets asked out loud. Do you trust one model? Do you go looking for the best model? Do you run several models side by side and keep whatever each of them happens to catch? Or — one turn further still — do you hand that whole pile of partial answers to yet another model and ask it to combine them into one? Each answer sounds reasonable on its own. Each is also, on its own, incomplete — and the only way to find out which incompleteness actually hurts you is to run the experiment rather than assume the answer.

This article runs it, in four steps that build on each other.

One model. A single broken bash script, with seven distinct bugs of varying severity, was handed to four models — Sonnet 5, HY3, Qwen3-Max, DeepSeek-V4-Flash — each working alone, each with no knowledge of what the others were doing. That's the first framing: one model, unaided, fixing only what it personally noticed.

Best model. Comparing those four independent attempts side by side produces the second framing almost automatically: if you had to pick just one of these four scripts to actually use, which is best? Table 1 answers that, and the answer is Qwen3-Max — but "best" here means "most complete," not "complete," a distinction the table is careful to preserve rather than paper over. Even the single best model in this test left real bugs unresolved.

Multi models, combined. That incompleteness is precisely why a third framing exists: instead of picking one winner and discarding the other three, what if a model's job is to read all four and assemble a single script that keeps every good idea, regardless of which of the four originally had it? Seven more models — ling-3.0-flash, Mistral-Medium-3.5, Nemotron-3-Super-120B, Qwen3-Max, DeepSeek-V4-Flash, Gemini Pro, dots-studio-3-note — were given exactly that job.

In an era where every company has its own app, UX design plays a vital role in representing the brand and building trust around the product. A well-designed app can accelerate your company’s growth, while a product with a poor user experience can ruin the company's image and lead to increased audience churn. To build an app that fulfills its mission and contributes to the company's success, you need a professional design team with the relevant experience, deep understanding of the subject and the right culture fit. Let's discuss a few key factors you should consider as a client before hiring a UX design team for your project.

In spring 2026, researchers at the Positive Technologies Expert Security Center (PT ESC) identified an unusual campaign targeting users in Saudi Arabia. The campaign used a previously unknown APK named Chrome.apk. The sample immediately stood out. Behind a benign-looking Google Chrome update screen and an unexpected request to enable Accessibility Services, we found a heavily protected execution chain that ultimately deployed Android spyware.

CreepJS is the test everyone in this space quotes and almost nobody reads carefully. It’s an open-source page that fingerprints your browser and then — this is the part that matters — checks whether your browser is lying. Not whether your values are unusual. Whether they’re internally possible.

Codex CLI sessions are easy to manage when only one task is active. The workflow becomes less predictable when several sessions operate in different repositories or terminal tabs. A task may be blocked on an approval while another is still running and a third is ready for review.

In the first half of 2026, we discovered activity by the East Asian threat group FamousSparrow. Its attacks targeted several countries in South Asia and Europe. In these attacks, the threat actor used its own malware: a heavily reworked variant of the modular SparrowDoor backdoor and a new backdoor that we named SquawkDoor.

FamousSparrow, also known as Salt Typhoon and Earth Estries, is an East Asian threat group that has been active since 2019. The group is known for using its proprietary SparrowDoor backdoor and initially focused on attacks against hotels worldwide, as well as government and international organizations. Later, FamousSparrow began actively targeting telecommunications companies and internet service providers, apparently seeking long-term access to lawful communications intercept systems.

In these attacks, the threat actor used both malicious LNK files and a website compromise vector involving malicious JavaScript. The injected script displayed a fake error message when a user visited the page and prompted them to download a new certificate. In reality, the download was a malicious executable that led to backdoor infection. The attacks were tailored to specific countries: the JavaScript samples and payloads contained text customized for the target country.

For several weeks, I logged every moment when understanding our system required asking another person instead of reading code, documentation, dashboards, or configuration. What started as a small experiment turned into a surprisingly useful map of hidden knowledge dependencies. The biggest risk was not missing documentation. It was that some parts of the system effectively existed inside one engineer's head.

Due to recent rapid progress in the field of large language models (LLMs), many have speculated about artificial general intelligence (AGI). This paper addresses the keystone question: Are LLMs capable of transitioning to AGI? The paper presents a framework for estimating the logical limits of cognitive systems, which highlights the conflicting design requirement, namely, being able to proficiently execute both strict and loose logic. LLMs are gravitating towards loose logic and that inhibits their ability to grasp the essence of complex topics.

I write about the point where software abstractions meet irreversible authority, with an emphasis on threat models, negative tests, and systems that fail closed.

DeepSeek Harness is often described as an open-source coding agent. That description is correct, but incomplete. The more interesting part is its architecture. DeepSeek Harness is a configurable runtime for constructing agents from model adapters, tools, session services, execution backends, permission policies, interfaces, and agent loops.

From 18 to 21 November 2026, Nepal will host PGConf.Nepal 2026, the country’s fourth PostgreSQL conference. The previous events took place in 2018, 2023, and 2025. This year, the conference is expected to bring together a wider mix of participants, organisations, and countries.

The journey from a single A100 in the cloud to an H200 cluster is not just a hardware upgrade — it’s a story of how an ML team stopped chasing “the one perfect model” and started building an ecosystem. When there are millions of lines of PostgreSQL C code under the hood and tasks range from hint-set generation to Graph-RAG, the model stops being a black box and becomes just another replaceable component. We explain how we rebuilt our stack around vLLM and MCP, why context management matters more than model weights, and how we made a 0.6B-parameter model perform on par with the giants using GRPO.

Hello, Habr!

For many IT professionals, the ultimate personal finance setup involves a massive Google Sheet or a complex Notion database. I completely understand this approach: when you spend your entire day sitting in front of a monitor, entering data into a cell takes two seconds.But 7 years ago, my reality was entirely different. I was an entrepreneur, and my workday meant constant movement. I was driving, visiting warehouses, meeting partners, or checking production floors. I was 100% mobile, and the only tool I always had on hand was my iPhone.

In that fast-paced rhythm, traditional finance trackers and spreadsheets simply fall apart. Here are typical scenarios from my life back then:

I wanted to see what actually changes when the same small data-processing program travels through four generations of programming languages. So I built it in BASIC, C, Pascal and Python, kept the algorithm as similar as possible, and compared not only execution speed but also memory use, binary size, debugging time and the amount of code I had to keep in my head. C was fast. Python was short. Those parts were predictable. BASIC and Pascal were where the experiment became interesting.

In light of recent events (restrictions, slowdowns, bans) — “this isn’t allowed,” “this is for the best,” “strictly in the interest of your security,” and so on — the market for proxy providers has grown exponentially in just the last couple of years. 

To be fair, the ”proxy boom” hit right during the golden era of the 2020s, and overall, the growth trend continues to rise. In fact, I believe in the next few years, we will witness a total revolution in automation and related services. Reimagining, repackaging, and rolling out new features — I’m certain we have yet to peak in abilities.

However, the larger the market grows, the breakdown of components becomes much more interesting — to analyze the players, categorize them into subgroups, and highlight the favorites, mid-tier options, and underperformers. There are plenty of services out there, and a good chunk of them are incredibly inefficient and inaccurate. Since I have hands-on experience with many of these players, I'll share my firsthand knowledge, in hopes of it being useful to others.

LLMs hallucinate — everyone has seen what it looks like. The model reports, confidently and coherently, with the right intonation and terminology, something that isn't there. Engineers treat this as a bug: one being fixed, one about to be fixed.

It won't be. This is not a technical fault but an inherent property of language, which the machine has made visible.

What do 1C queries, IN (VALUES ...) constructs, and the innocent expression x + 0 have in common? They can all turn a millisecond query into a multi-minute wait because PostgreSQL's standard planner "trips" over them. We break down how the pgpro_planner extension rewrites tricky parts of the query tree into a friendly form before the optimizer picks a bad plan — and why some of these fixes have already made it into vanilla PostgreSQL 18.

AI-generated code often looks cleaner than code written by a tired developer at 2 a.m. It has sensible names, neat layers, comments, tests and even error handling. The problem starts when the code meets something the prompt forgot to mention: two requests at once, a repeated webhook, a cancelled task or a database failure halfway through an operation.

This article is about bugs that stay invisible during a normal code review. I took several ordinary backend tasks, generated working solutions and then tried to break them with timing, retries and bad input. The code compiled. The tests were green. Some of it was still unsafe.

I am glad to present a new tool that I built, with Codex gpt-5.6.

Progressive Web App - mean that you can "install" it to your desktop or mobile screen - this is a web app - outside of App Stores.

Free and open source.

Without backend - just direct calls to Wikidata. Without an account.

You are welcome to share your feedback.

Currently this is the list only - no map, see related project https://wikishootme.toolforge.org/

https://vitaly-zdanevich.github.io/wikidata-nearby-pwa/

My name is Anton Omelianenko and I’m head of software development. A manager runs a team so that it delivers results for the business. To judge how well people are handling their work, a manager needs data and a system for assessing it. In software development this is harder than it looks.

This article covers three questions: why classic metrics fail when assessing developers, how to measure performance properly and what to do when an employee may not be working only for you.