AI News
Articles techniques sur l'intelligence artificielle, consommables par les humains comme par les agents IA.
Sur iPhone/iPad : appuyez sur Partager puis « Sur l'écran d'accueil » pour activer les notifications.
Categories
Latest articles
-
AI Productivity Tools: The Big July 2026 Comparison
Agentic video editing, presentations, text-to-speech, local LLMs and the top 10: the AI productivity tools comparison, verified in July 2026.
13 min read ProductivitéOutils IALLM -
Evaluating an LLM: the LLM-as-a-judge pattern
Measuring LLM quality with no single answer: metric limits, LLM-as-a-judge (modes, biases, mitigations), human agreement, Elo arenas, HELM.
10 min read LLMÉvaluationQualité -
AI agents and multi-agent systems, explained
The think-act-observe loop, ReAct, tools, memory and reflection; multi-agent patterns (orchestrator-workers), frameworks, evaluation and failure modes.
12 min read Agents IAMulti-agentsOrchestration -
Embeddings and vector search, explained
An embedding turns text into a vector of meaning. Training (word2vec, contrastive), cosine similarity, ANN (HNSW/IVF/PQ), vector databases, MTEB.
11 min read EmbeddingsRAGRecherche vectorielle -
Reasoning models and test-time compute
Why spend compute at inference: CoT, o1/o3, DeepSeek-R1 (GRPO, the "aha moment"), best-of-N, PRMs, search, scaling laws, and the limits.
18 min read RaisonnementTest-time computeRL -
Agentic RAG and GraphRAG: reasoning and structure above retrieval
Beyond naive RAG: agentic loops (Self-RAG, CRAG, routing, multi-hop), hybrid + rerank, and GraphRAG (entities, Leiden, summaries). Costs and RAGAS.
12 min read RAGGraphRAGAgents IA -
MCP: giving tools to AI agents
The Model Context Protocol standardizes connecting LLMs to tools and data: architecture, primitives, transports, security, and the 2026 ecosystem.
9 min read MCPAgents IAOutils -
AI agent security: prompt injection and guardrails
Why AI agents create new risk: direct/indirect prompt injection, the lethal trifecta, the OWASP LLM Top 10, and defense in depth.
13 min read SécuritéPrompt injectionAgents IA -
Multimodal models and Vision-Language-Action (VLA)
From CLIP and ViT to VLMs (LLaVA, Flamingo, GPT-4o) then VLAs for robotics (RT-2, OpenVLA, π0): modality fusion and actions as tokens.
14 min read MultimodalVLARobotique -
State Space Models and Mamba (beyond Transformers)
Classical SSMs, S4 (HiPPO, Long Range Arena), Mamba selectivity, parallel scan, Mamba-2 duality, and Jamba hybrids. Linear cost vs quadratic attention.
15 min read SSMMambaArchitecture -
Fine-tuning and post-training (LoRA, DPO, RLHF)
The post-training pipeline: SFT, RLHF (reward model + PPO) and DPO, plus LoRA/QLoRA. When to fine-tune vs RAG vs prompt engineering.
14 min read Fine-tuningLoRARLHF -
Long context and KV-cache management
Why long context is costly: O(n²) attention and a linear KV cache. MQA/GQA, FlashAttention, PagedAttention, RoPE/YaRN, attention sinks.
12 min read Contexte longKV-cacheInférence -
Mixture-of-Experts (MoE) at scale
MoE: router, top-k routing, auxiliary load-balancing loss, capacity, expert parallelism. Total vs active parameters (Switch, GLaM, Mixtral, DeepSeek-V3).
16 min read MoEArchitectureScaling -
Quantization and efficient LLM inference
FP32/FP16/BF16/FP8/INT8/INT4 formats, GPTQ, AWQ, SmoothQuant, GGUF, NF4, outliers, continuous batching, PagedAttention and speculative decoding.
12 min read QuantizationInférencePerformance