Konstantin Pavlov

Steve Jobs was alsway behind the mentioned products. But he’s not with Apple any more…

Why it matters:
• Parameter sharing (same θ reused across layers) keeps the model small.
• Adaptive depth (different numbers of steps per token) avoids wasting compute on tokens that don’t need much processing.
• Smarter memory usage by caching only what’s needed at each step.

Now, rather than running every input token through f a fixed number of times, it introduces a router, which decides dynamically how many times to apply f for each token, based on how “difficult” it is. The model learns this routing behavior during training.

Think of it like this:

Imagine a fixed-size function f(x; θ) that transforms input x using parameters θ. Instead of stacking many unique layers (with different θs) like in standard Transformers, this model reuses the same function f several times—this is the “recursion” part.

The paper: arxiv.org/pdf/2507.10524

YouTube video by AIM Tv Smaller, Faster, Smarter: Why MoR Might Replace Transformers | Front Page

Google has introduced a new Transformer model called Mixture-of-Recursions, which lets each token go through a different number of processing steps based on how much “thinking” it needs, helping the model use less memory and computation while improving performance. youtu.be/MfswBXmSPZU

Release 0.3.0 · kpavlov/langchain4j-kotlin Notable Changes Update dependency dev.langchain4j:langchain4j-bom to v1.1.0 by @renovate[bot] in #163 Update kotlin monorepo to v2.2.0 by @renovate[bot] in #164 Remove reports module and update Ko...

blew the dust off #langchaiun4j-kotlin. Kotlin 2.2 and LangChain4j 1.1 ready.
github.com/kpavlov/lang...

In a world where every keystroke carries consequences, how far will they go to avoid the sweet taste of failure?

Genres: Workplace Thriller, Tech Drama
This story is: Mind-Bending, Thought-Provoking
Maturity Rating: PG (Mild Technical Jargon, Intense Pair Programming Scenes)

... a psychological experiment in collective responsibility, technical excellence, and the price of perfection. As commits shrink and pair programming intensifies, the team discovers that their greatest enemy isn't failed tests or broken code, but the comfortable mediocrity they're leaving behind.

The Cookie Story: When Build Failures Became Sweet Accountability Twelve developers. One trunk. And a jar of cookies. Zero tolerance for broken builds. Watch as pair programming sessions intensify, commits shrink to surgical precision, and a team discovers that the ...

A software development team implements a deceptively simple rule: break the build, buy cookies for everyone. What begins as a lighthearted accountability measure slowly transforms into something far more profound — ...

Illusion of Hype - xam.dk/blog/illusio...

You can use YouTube's latest features! Your browser is up to date

Had a fun conversation with Mark Sailes about my career in professional opensource - maybe you find some of it interesting too. And go subscribe give a like on Marks new channel!

youtu.be/ePb9O9VDsVA

Poland trials a four-day working week: A step towards the future? Poland will start its first-ever pilot of a shortened working week this July as the government seeks to meet social needs.

Poland will start its first-ever pilot of a shortened working week this July as the government seeks to meet social needs.

Their system prompts need more refinement and supervision and human approvals while making the decisions, but the business model is seemingly interesting and promising.

An AI chatbot ran a shop for a month. But things got weird very fast Anthropic put an AI chatbot in charge of a shop. The results show why AI won’t be taking your job just yet.

Anthropic put an AI chatbot in charge of a shop. The results show why AI won’t be taking your job just yet.

LangChain4j with Quarkus | KotlinConf 2025, May 21–23, Copenhagen KotlinConf is the official Kotlin conference by JetBrains. It is a place for the community to gather and discuss all things Kotlin.

You may also get the slides from here: kotlinconf.com/talks/795976/

YouTube video by Kotlin by JetBrains LangChain4j with Quarkus | Max Rydahl Andersen and Konstantin Pavlov

The #video of our talk with @maxandersen.xam.dk at #KotlinConf2025 on #Quarkus #LangChain4j is here! youtu.be/hxEM5J6QfLY

Guardrails | LangChain4j Guardrails is an experimental feature. Its API and behavior might change in future versions.

In case you missed the latest @langchain4j.dev release, a new feature that snuck in is #guardrails! You can validate LLM input & output!

docs.langchain4j.dev/tutorials/gu...

If AI can generate it once, let’s save the planet 🌍🌱 by not asking AI to generate it many times. Same as: if AI can write and run a script, lets’s save the script instead of asking AI to generate it every time.

Build AI Apps and Agents in Java: Hands-On with LangChain4j - JAVAPRO International Discover how to integrate AI models & LLMs seamlessly in Java using LangChain4j. Beginner or an advanced developer, this hands-on guide walks you through the key features, from basic AI interactions t...

A comprehensive overview of building #AI applications with #LangChain4j
javapro.io/2025/04/23/b...

LLM Evaluation Testing with promptfoo: A Practical Guide This article shows how to implement automated testing for LLM applications using promptfoo with a real application server, addressing the challenge that traditional testing methods fail with non-deter...

I wrote a new #blog post about why #testing #prompts in isolation isn’t enough 🚨

Your app usually generates prompts on the fly. Static tests miss how it behaves in real-world use. You should test the service itself—dynamic prompt logic, context, full flows
kpavlov.me/blog/llm-eva...
#ai #testing

First day at #DevTalksRomania. Today I am hosting a workshop on building AI agent on #Kotlin, #LangChain4j, and #Quarkus

Kodee, the speaker on KotlinConf 2025 in Copenhagen, Denmark 🇩🇰

There are only 103 unique #LEGO sets like this, and only 2 happy owners in #Estonia: @antonarhipov.bsky.social and myself 😎
#KotlinConf25 #Kodee

The DevUI is especially useful to play with the ADK

GitHub - kpavlov/koog-sauce: The finishing touch to JetBrains Koog The finishing touch to JetBrains Koog. Contribute to kpavlov/koog-sauce development by creating an account on GitHub.

#JetBrains #koog is better with #koog-sauce: here is an example how to use Koog and #SpringAI together: github.com/kpavlov/koog...
#AgenticAI

Debugger in baggy jeans You squash bugs in oversized hoodies with style. The stack trace fears you.

I am Debugger in baggy jeans. Unleash your inner Duke and win a limited-edition T-shirt!
#Java30WithIntelliJIDEA www.jetbrains.com/lp/java-30/d...

KotlinConf'25 – Day 1 Livestream (Feat. Keynote) KotlinConf is the premier event connecting professional developers and companies shaping the future with cutting-edge technologies. Conference schedule - https://kotlinconf.com/schedule/?day=2025-05-...

🎉 KotlinConf is almost here! The opening keynote and an exciting lineup of sessions kick off tomorrow at 9:00 am CEST. Don’t miss the livestream: 👉 kotl.in/conf25-live

A simple test for new Google Java Agent Development Kit #ADK

A #test for @glaforge.dev's ADK example glaforge.dev/posts/2025/0...
github.com/mokksy/ai-mo...

#Mokksy AI Mocks 0.4.0 now supports Google #Gemini #VertexAI mocking. Enable deterministic testing with local API simulation - no production dependencies needed. Useful for stable CI/CD with #AI components.

Docs: mokksy.dev/docs/ai-mock...

#AITesting #DevInfrastructure #Gemini​​​​​​​​​​​​​​​​

I found a nice #AI #wiki generator. It helps to understand project structure and architecture, e.g. deepwiki.com/mokksy/ai-mo...

Free for open-source projects. They are asking to add a badge to README file, and the wiki will be updated weekly
deepwiki.com