Archives

Vibe coding, plan mode, and autonomous agents do not call on developer attention at the same moment. The real issue is not the tool, but how human judgment is distributed without exhausting the team.

Skills and AGENTS.md files can help AI agents, but only when context is selected with precision. Too much context increases cost, slows agents down, and can degrade their answers.

Attacks against LLMs exploit less isolated technical flaws than the biases created by alignment: helpfulness, authority, self-generated anchoring, confirmation, and empathy.

DeepSeek V4 shows how a 1.6 trillion parameter model can become more efficient by changing its architecture: MoE, hybrid attention, mHC, Muon, and FP4 reduce wasted compute instead of adding brute force.

DeepSeek V4 is the model I am waiting for most in 2026, not because of the benchmark race, but because of what its architectural choices reveal about how future LLMs may be built.

A field report on how I am now untangling a multi-agent framework launched too quickly in the euphoria of vibe coding 2025, and why the right question is no longer simply "does it work?"

A critical reading of the HyperAgents paper: why this self-modification approach is most useful as an exploration machine for discovering agent engineering patterns, rather than as a target production architecture.

Five practical and often overlooked uses of embeddings to detect drift, map skills, cluster support tickets, find duplicates, and surface hidden incidents in logs.

To know if a model can act alone in a multi-agent workflow: EPOCH-Bench, an agentic planning benchmark inspired by Day of the Tentacle, with PDDL, 6 levels and 6 metrics to break down failure modes.

Why a reliable RAG looks more like a library (index, categories, navigation) than a top-k search engine, and which architectures (hierarchy, summaries, graphs, agents) enable that shift.

Practical guide to anchor your LLM responses on the web — without getting trapped. Comparison of three approaches (integrated, classic API, AI-optimized), analysis of hidden costs, and defense strategies against web poisoning.

Why "review this code" and "review this code for security" don't yield the same results. How the prompt guides the model's exploration and why multiplying perspectives improves response quality.

Why LLMs “recompute” simple facts, and how Engram (DeepSeek) proposes an on-demand external memory to retrieve them quickly.

Systematic analysis of 100+ publications (2023-2025) on single and multi-agent AI systems. Identification of viable, fragile, and structurally impossible patterns. Technical reference document.

Analysis of publications on AI agents: the golden rule of deterministic feedback, patterns that work, those that fail, and 7 design principles for building reliable systems.

Complete guide to the 11 multi-agent orchestration patterns: Pipeline, Supervisor, Collaborative, Hierarchical, Fan-out/Fan-in, Evaluator/Critic, Blackboard, Debate, Reflection, Swarm and Council. How to structure the collective intelligence of your AI agents.

Demystify the famous analogy "king - man + woman = queen" by understanding how AI models represent words as points in a multidimensional vector space through embeddings.

Practical guide to building a high-performance AI agent: languages, orchestration, models, telemetry, storage and runtime. The fundamental building blocks.

Anthropic published its Agent Skills specification as an open standard on December 18, 2025. Agent Skills enables transforming a generalist AI into a specialist for your business processes by creating an "onboarding kit" that the AI automatically loads when needed, avoiding repeating the same instructions in every conversation.

Discover how multi-head attention, through the metaphor of 8 projectors, enables ChatGPT and LLMs to decode complex sentence nuances in parallel.

The debate over the announced death of Transformers resurfaces. By late 2025, Attention-SSM hybrid architectures gain ground for efficiency, but pure Transformers still dominate complex reasoning. Reality: coexistence and specialization, not replacement.

Meta's Free Transformer injects latent variables to enable LLMs to make global decisions before generating. Impressive results on reasoning and code, but open questions on scalability.

Top 5 most relevant AI papers from arXiv Week 51-2025 with analysis and insights.

Top 5 most relevant AI papers from arXiv Week 50-2025 with analysis and insights.

Top 5 most relevant AI papers from arXiv Week 49-2025 with analysis and insights.

Top 5 most relevant AI papers from arXiv Week 48-2025 with analysis and insights.

Top 5 most relevant AI papers from arXiv Week 47-2025 with analysis and insights.

Top 5 most relevant AI papers from arXiv Week 46-2025 with analysis and insights.

Top 5 most relevant AI papers from arXiv Week 45-2025 with analysis and insights.

Top 5 most relevant AI papers from arXiv Week 44-2025 with analysis and insights.

Top 5 most relevant AI papers from arXiv Week 43-2025 with analysis and insights.

Top 5 most relevant AI papers from arXiv Week 42-2025 with analysis and insights.

A critical analysis of AI benchmarks and what they truly measure: encyclopedic memory vs abstract intelligence.

Top 5 most relevant AI papers from arXiv Week 39-2025 with analysis and insights.

Top 5 most relevant AI papers from arXiv Week 38-2025 with analysis and insights.

Discover the top 5 most relevant AI papers from arXiv week 37-2025: deep learning innovations, new language models, and computer vision advances. Complete technical analysis and insights.

Discover post-training, the secret art that transforms raw AI models into assistants with distinct personalities.

Discover the top 5 most relevant AI papers from arXiv week 36-2025: deep learning innovations, new language models, and computer vision advances. Complete technical analysis and insights.

Discover the top 5 most relevant AI papers from arXiv week 35-2025: deep learning innovations, new language models, and computer vision advances. Complete technical analysis and insights.

A reflection on vibe coding, between technological wonder and forgetting the foundations. Where does the developer stand in this new era?

Uncover how LLMs access web data: from training datasets to real-time search capabilities and their significant implications for SEO and content strategy.

Deep dive into Google's groundbreaking publication revealing the internal architecture of large language models and their transformative implications.

Dive into AI image generation and explore the fascinating challenges of algorithmic creativity. How AI transforms art and pushes creative boundaries today.

Dive deep into LLM architecture like GPT: let's dissect their internal workings, attention mechanisms, and reveal the technical secrets behind language models.

Third part of our AGI series: exploration of diverse approaches and methodologies for achieving artificial general intelligence. Multiple pathways analyzed.

Second part of our AGI series: explore why artificial general intelligence always seems 20 years away and the psychological factors behind this phenomenon.

First part of our comprehensive AGI series: dive deep into AI's history to understand its evolution and better grasp the future of artificial intelligence.

Dive into the historic AlphaGo vs Lee Sedol match that marked a turning point in AI. Explore the profound implications for human-machine interaction.

A reflection on vibe coding, between technological wonder and forgotten foundations. Where does the developer stand in this new era?