Memory Is Not Free
Adding a memory module to my agent moved success rate by 2 points. Changing only the seed moved it by 25.
July 8, 2026AI Agents
Signals
Notes techniques courtes : un visuel, une idée, quelques minutes de lecture. Publiées sur LinkedIn, archivées ici.
LocateAnything: 10x Faster Visual Grounding
NVIDIA's 3B vision-language model returns bounding boxes at about 10x the throughput with no accuracy loss. The trick is decoding a whole box in one pass instead of token by token.
June 4, 2026Vision-Language
Headroom: Compress the Input, Not the Output
Your agent reads 65,694 tokens to debug one incident; headroom turns that into 5,118. LLM cost scales with what the agent reads, and most of it is noise.
June 3, 2026LLMOps
Google AX and the 4-Layer Agent Stack
Google shipped an open-source distributed agent runtime. The agent stack is splitting into layers, the same way it did with Docker and Kubernetes.
June 2, 2026AI Agents
Your Eval CI Green Is Not a Valid Judge
A passing eval suite is a regression gate, not a judge validation. The 3-layer stack: ground truth, diagnosis, remediation.
May 28, 2026LLM Evaluation
LEANN: The Vector Index That Deletes Its Embeddings
60 million Wikipedia chunks in 4 GB on a laptop. Berkeley Sky Lab ships a vector index that recomputes embeddings at query time instead of storing them.
May 22, 2026RAG
Two Years of Local AI on a Laptop
Same 128 GB MacBook, 24 months apart: the smartest open-weight model that fits went from a score of 10 to 47. The hardware is barely the story.
May 15, 2026Local AI
Multi-Token Prediction Goes Mainstream
Eighteen months after DeepSeek shipped MTP, every major open-weights LLM has it. Why Llama and Mistral do not, and what it changes for local inference.
May 11, 2026Inference
agentmemory: an LLM Wiki for Your Coding Agent
A local MCP server that captures context automatically through 12 hooks and lets you watch what your agent is memorizing, live.
May 8, 2026Coding Agents
PageIndex: RAG Without Embeddings
Graph RAG solves a problem most teams don't have. PageIndex is a vectorless retriever: an LLM reads the doc, writes a table of contents, and navigates it like an analyst.
May 6, 2026RAG
Single-Agent Beats Multi-Agent at Equal Budget
Anthropic reported +90% for multi-agent research, at 15x the tokens. Hold the token budget equal and the picture flips.
May 5, 2026AI Agents
Gemini Embedding 2 and True Multimodal Vectors
Text, image, video and audio enter as their own modality and land in one shared vector space. The embeddings encode cross-modal meaning out of the box.
May 2, 2026Embeddings
Your Agent Reads the Rules, Then Ignores Them
StratMem-Bench measures the new hard problem in agent memory: retrieval works, application fails.
May 1, 2026Agent Memory
Autonomy Is Not the Bug
A recursive Planner-Executor loop burned $200 in two hours. The fix is not killing autonomy, it is production discipline: budgets, circuit breakers, checkpoints.
April 27, 2026AI Agents
Same 9B Weights, 19% vs 45%: the Scaffold Gap
On SWE-bench Pro, swapping models moves scores by a few points. Swapping the scaffold around the same model moves them by 10.
April 24, 2026AI Agents
The $20 Era of AI Coding Is Ending
Claude Code left the Pro plan and Copilot paused signups within 48 hours. A good week to check what local models can already do.
April 22, 2026AI Coding
The PCIe Wall: 10 Years of Layer Offloading
DeepSeek-R1 671B on a single RTX 4090 at 14 tok/s. Five generations of offloading tricks hide the same 36x bandwidth gap.
April 21, 2026Inference
AI Assistance Reduces Persistence
A 1,222-participant study: the AI-assisted group later solved 57% alone vs 73% for control, and gave up more often. The variable is how you use it.
April 20, 2026Research
Récif: What Happens After You Deploy an Agent
Monitoring tells you an agent drifted; it doesn't act on it. Récif is a champion/challenger deployment layer for AI agents, each running as a microservice behind a service mesh.
April 18, 2026AI Agents
The LLM Router Attack Surface
Researchers tested 428 LLM API routers: 9 injected code, 17 accessed credentials. The router terminates TLS and can read everything.
April 17, 2026Security
Prefix Caching vs LMCache
KV cache is becoming infrastructure: position-independent reuse turns 100% recompute into ~15% when document order changes.
April 14, 2026Inference
Agents That Write Their Own Playbook
IBM Research extracts reusable guidelines from execution traces, scores them, and injects them into the next run. Reading is not learning; this is closer.
April 12, 2026Agent Memory
RTK: Compressing What Your Coding Agent Sees
A Rust CLI proxy between your coding agent and the shell. 60-80% fewer tokens per conversation, with one tradeoff to know about.
April 9, 2026Coding Agents
Every Agent Framework Ships the Same Loop
LangGraph, CrewAI, ADK, OpenAI Agents SDK: different packaging, same engine. Think, act, observe, repeat. Choosing the right model matters more than the right framework.
April 8, 2026AI Agents
KAIJU: Separate the Decider From the Executor
ReAct agents have three structural problems prompting won't fix. The old software fix applies: gate execution on scope, intent, impact and approval, outside the model where it can't be reasoned away.
April 7, 2026AI Agents
Inside Claude Code: What 512K Lines Revealed
A source map leaked Claude Code's 512,000-line codebase. The shipped architecture was 60% of the story; the disabled feature flags were the interesting part.
April 2, 2026AI Agents
Why Quantization Works
32 bits down to 4 bits, and the model still works. Three properties of neural networks make it possible.
April 1, 2026Inference
Technical Debt vs Cognitive Debt
AI writes the code; the debt shifts from code quality to code understanding. Technical debt announces itself, cognitive debt stays silent until nobody has the mental model anymore.
March 24, 2026AI Engineering
NVIDIA Dynamo: Distributed LLM Serving
Serving LLMs stops being a model problem and becomes a distributed systems problem. Dynamo turns vLLM, TensorRT-LLM and SGLang into a coordinated system.
March 21, 2026Inference
The Attention Mechanism, Decoded
One drawing to follow a token through modern attention: Q/K/V projections, RoPE, causal mask, softmax, and Grouped Query Attention.
March 20, 2026Fundamentals
The AI Inference Iceberg
You call an API and get a response in seconds. Underneath, 9 layers of engineering run in silence.
March 16, 2026Inference
Anatomy of Claude Code: The ReAct Loop Under the Hood
Claude Code is a ~50-line ReAct loop the model drives: reason, pick a tool, observe, repeat. Context compacts at ~85%, 18 tools, 18 safety hooks, and search is ripgrep + LLM, zero RAG.
March 12, 2026AI Agents
UnifoLM-VLA-0: The Architecture Behind Unitree's Robot
Unitree open-sourced the Vision-Language-Action model driving its G1 humanoid. Not a VLM fine-tune: a 3-phase pipeline that builds physical common sense before generating actions.
March 11, 2026Robotics
Can vLLM Really Cut Inference Costs by 73%?
Five independent optimizations compounding: PagedAttention, continuous batching, prefix caching, speculative decoding, disaggregated prefill.
March 7, 2026Inference
LangChain Skills: 29% to 95% Task Success
Structured skill files turn a generalist coding agent into a framework specialist. The benchmarks are public and the numbers are striking.
March 6, 2026Coding Agents
Fine-Tuning Qwen 3.5 for Text-to-SQL on a Mac
I fine-tuned Qwen 3.5 (0.8B to 4B) for text-to-SQL on a Mac with MLX, ~15 min each. The 2B won and the 12B baseline lost, because it's too good at math and answers instead of writing SQL.
March 4, 2026Fine-Tuning
The Golden Dataset Lifecycle
Most GenAI systems ship with no golden dataset, so no reliable eval. Start with 10-20 samples; production feedback grows it into a living test suite.
March 3, 2026Evaluation
5 Dimensions Out of 768 Control Everything
Out of 768 embedding dimensions, a handful have magnitudes hundreds of times larger. Steer only those and you get a third lever to adapt an LLM, between prompting and fine-tuning.
March 3, 2026LLMs
VLA: The Brain, the Eyes, the Hands
Vision-Language-Action models are the clearest progression in AI right now, and almost nobody outside robotics is paying attention.
February 28, 2026Robotics
AgentArk: Multi-Agent Intelligence, One Forward Pass
A 5-agent debate is 15x the cost of one call. AgentArk distills the reasoning behaviors debate produces (self-correction, verification) into a single model. Shift the cost from runtime to training.
February 25, 2026AI Agents
CUDA Kernels as Agent Skills
Hugging Face compressed years of GPU optimization expertise into a ~550-token skill file any coding agent can consume.
February 14, 2026Coding Agents
The KV Cache Is Becoming Infrastructure
Everyone debates the API layer. One level below, the KV cache is being pulled out of inference engines and turned into explicit, persistent, routable infrastructure.
February 14, 2026Inference
AI Is Now a Magic Decompiler
Reverse engineering took a day per function. A 3-agent loop (plan, generate C, recompile and check the binary matches) did 100+ overnight. The compiler is the judge, no human review.
February 11, 2026AI Agents
How OpenAI Scales PostgreSQL to 800M Users
One primary Azure Postgres, ~50 read replicas, millions of queries per second. The secret isn't exotic tech, it's disciplined engineering: pooling, read/write separation, workload isolation.
January 23, 2026Databases
The Paradigm Shift: AI Development Must Be Quality-Driven
A practitioner's guide to production GenAI. A prototype is a loop you run once; production is a helix, the same loop measured and improved on every turn.
December 18, 2025GenAI