AI, management, personal stories. Notes from a Lead DS in banking production.
Good courses from LangChain itself, got around to taking Deep Agents (lectures + labs). This is their own harness solution. Pros: flexibility and the ability to manage all the…
I wanted to challenge the guys at work that it’s easy to unfold the models. Namely, by trying server stories SGLang, vLLM or Triton. But I refused llama.cpp, because it’s more…
SkillOpt - evolving skills (Arxiv | GitHub 15k+ ⭐️) Previously, neurons were optimized by gradient, now skills. Ideologically, SkillOpt - we look at the trajectories of tool…
New SIRIN library for assessing and monitoring contextual hallucinations in RAG and memory-based LLM systems (arxiv + github article) Theoretically, there are three tools inside:…
Ouroboros is now SOTA on Terminal Bench, OSWorld and CL-bench My favorite pure Python agent loop has evolved to become the best tool for coding! On the main benchmarks it now…
Still on codex and claude code. What about ouroboros 🤌🏻 ps a man gave $50k to run all the benchmarks
We came up with another skill that creates a skill using books (Book-to-Skill - 10k ⭐️). Essentially, zoomers have reinvented the table of contents. I collected this skill from…
Harness evaluation - deepeval Someone may be starting to develop harness solutions and skills for them. I was wondering what tools can be used for quality testing. While many…
It's cool that Cloudflare allows you to create a static website in a couple of seconds without registration - simply by transferring a ZIP archive. And if you register, you can…
The researcher intercepted Grok Build CLI traffic and discovered that the tool was sending not only open files to the cloud, but the entire Git repository along with commit…
UPD: When I shared my research under the announcement in the Yandex channel, they sent me a link to the open benchmark TabArena (see 2nd picture). The conclusions there are…
Tested the new PageAgent from Alibaba This is an open-source browser-based AI agent that can independently work with websites: understand the interface, move the cursor, click on…
AI coffee with a robotic hand: Chita - Moscow 1:0 in AI implementation 😎
Corporate Bullshit Receptivity Scale (2026) The article explores why some people perceive meaningless corporate language (with many different buzzwords) as meaningful and clever,…
*I just read the documentation
I lost to Volendemort 🤩 In May, a competition from Avito “ML CUP 2026” was held to predict the most relevant ads that the user will contact. It would seem like an ordinary rexis:…
I lost to Volendemort 🤩 In May there was a competition from Avito " ML CUP 2026 " by predicting the most relevant ads that the user will contact. It would seem like an ordinary…
NVIDIA DGX Spark vs Macbook NVIDIA DGX Spark is less than a year old and can be bought on wildberries, it will cost 500k+, for example, here (and they actually deliver it). Sounds…
Interesting project for vibecoding https://freebuff.com/ with free limits and no token restrictions, it really works. In fact - 5 sessions for an hour every day. DeepSeek and…
Do nothing 😐: long term productivity system A worthy book about the benefits of procrastination. Inside: • Meta-studies on cognitive load: The author cites various studies that…
Interesting research from Stanford Hiring is now declining primarily due to juniors (entry-level positions). Moreover, development and customer support suffered the most, since…
A.I. as coding agents, rewrite a letter or help do analytics or check some facts. What about creating real added value? 😅? Lately I’ve been using chatgpt for cooking, starting…
Russian Speech-To-Text Made it small bench how do different ones work speech-to-text models. I was wondering if they came up with something better Whisper (it's still good and…
Colleagues at work wrote an interesting thing that wraps HTTP traffic via SSH on a server. Convenient, for example, to get web access to code running Claude Advantages of websh: –…
Received 100 trillion tokens from Xiaomi (more than one vacuum cleaner) PS not everything is for me, but still worthy 😐
💀
Went to AI Conf the reports are good, and all about agents. They don’t even talk about classic ML anymore 1. About publishing industry reports at A conferences*: surprisingly, the…
This is the first squat where I eliminated some of the solutions
RAG SOTA benchmarks RAG as a tool for connecting the knowledge base and enriching the context for LLM is the basis. A classic pipeline usually includes Hybrid Retrieval with…
We are implementing 🙂
Everyone relied on Chinese models, but subscriptions there cost no less. Here z.ai with glm 5.1 they immediately doubled prices to $160 for a max subscription (the best in open…
At the beginning of the week, many wrote about the project Colleague Skill https://github.com/titanwings/colleague-skill - a system for digitizing a living employee into an AI…
Time series foundation models In time series problems, foundation models are also built. Due to pretraining on hundreds of billions of points, the model understands trends and…
Dive into Time-Series Anomaly Detection: A Decade Review (link) 📌 A review article that collected key developments in the field of anomaly detection over the past 10 years. Since…
Taki picked it up Clawbot (or moltbot or smart AI assistant) - it was more difficult than I thought. On one of the local VPS (there we also had to create tickets to turn on the…
AiConf 2026 will be held in Moscow on April 20 I’m thinking of going myself: it’s a Data Science conference with an application program (master classes, workshops). Will discuss:…
Review what else did Open Source LLM come up with? over the past two months: inside the article there is a good comparison of architectures and how the authors are inspired by…
One ruler to measure them all: Benchmarking multilingual long-context language models ( arxiv 2025 ) This article examines how language affects the quality of models. Usually, in…
He spoke at a local python meetup (there were 100+ people and they promised a T-shirt 🤨 ) about publishing solutions in open source - everything is much easier than it seems
Jevons paradox technological progress (see vibecoding), which increases the efficiency of using a resource, increases the volume of resource consumption => while we live The graph…
AI is warming up It should make the work easier, but instead it makes it more intense. An article by the Harvard Business Review (dated February 9, 2026) examined exactly why this…
Anthropic accused Chinese AI is that they made too many requests (16+ million) to their grid through fake accounts, thereby violating the terms of use, all in order to improve…
The most difficult domain in ML is RecSys - a quote from a lead that works with recommendations. Some time ago I had to study what was new in the recommendations: • To study the…
It turns out that people are already bringing generated slides for work projects (I think it’s too early to do that). I looked at what I could try (ChatGPT, Gemini, Perplexity) -…
Mutual Agentic Looping 😎
Good article about computer-use agents based on LLM from @wrapper228 here • Formally, LLM can already be used as a computer management tool - not just as a chat, but to perform…
Tested high-star opensource learning solutions: KnowNote and DeepTutor (local analogues of NotebookLM) • KnowNote did not work in full force (except for the chat mode), since the…
Free API To test agents you need an API and tokens for LLM, but this costs money. It turns out that you can mimic and use existing subscriptions without spending tokens. For…
🚀 I'm looking for senior DS: AI agents + classics in the HR block to make the employee's path better and more technologically advanced: Middle+/Senior Data Scientist [Sber, Block…
An article was published in the Financial Times How to AI-proof your job about critical skills in a career (unfortunately, it is under paywall), but you can find the source of the…
No wonder RAM is becoming more expensive. Or of course my respect to the actor (c)
The most complex domain in ML is RecSys - a quote from a lead who works with recommendations. Some time ago I had to study what was new in the recommendations: • For studying…
Results of the channel's year from tgstats and tgwrapped.com, the numbers don’t add up - that’s what it means they haven’t agreed on the methodology
Global memory in LLM Memory is needed by AI agents to retain long-term context from past and current interactions, given that when the context window inflates, facts begin to be…
Global memory in LLM AI agents need memory to retain long-term context from past and current interactions, given that as the context window swells, facts begin to be forgotten 1.…
An interesting review of open-source LLM architectures for 2025 The Big LLM Architecture Comparison It’s funny how each of the models in its own way combines already known…
Yesterday Avito hosted a party event for DS team leads and managers. There were many friends and acquaintances, many whom I met for the first time. I would like to immediately…
Open season New Year's corporate events 🎉 Serious preparation, the guys coded an analogue of the game 100k1 with llm’s to combine similar answers 😎
Worthy blog for leisure from the former CTO noname of the company, although he worked for many places, including Uber, Digg and Yahoo Moving from an orchestration-heavy to…