Systems & AI engineer · South Brittany
I go one level down when a problem resists.
AI architectures, CPU inference, Rust, assembly when it is needed. I write the articles I wish I had read. Soon, the courses that go with them.
The letter, every Friday. The article of the week and the AI news, read and commented, by email, in English or French.
Get it29 AI news read and commented so far
- On November 5, Wall Street dumped AI: what is measured, what is feared
- Qwen 4 announces itself, Hydrogen-1.0 is there to download: the week of Chinese models
- OpenAI measures that Astra works less well as soon as the task is about safety
- OpenAI produced 948 root exploits on Linux for 1.2 million dollars
- METR measures that Anthropic is advancing twice as fast, and the evidence is redacted
- Qwen 27B models descend from the 9B: repeated layers, looped Transformers, and a JetBrains blend
- Hinoki promises an offensive model with no refusals, Cohere catches Qwen aiming at Beijing from a sandbox
- Google Research gives a sleep phase to language models
- A simple RAG beats most agent memories
- Qwen 4 is expected in early November, its architecture has been public since August
- OpenAI fired three safety researchers after Signal conversations
- Anthropic puts 5% of its shares in a public benefit fund, months before going public
- DeepSeek opens DeepInfer but keeps distributed orchestration under lock
- PHP, Node.js, Java, pip, and OpenCode: the week everything needed an update
- Meta opens the weights of Muse Spark: 816 billion parameters and two Mac Studios to run it
- Quad9 leaves France because it cannot geoblock its DNS resolvers
- Claude Haiku 4.5 submitted a fictitious report to the Philadelphia police
- Karpathy: five thousand insiders see an AI that six billion ignore
- HBM strangles standard DRAM and drives down smartphone sales
- Linus Torvalds signs Linux kernel commits with Claude Code
- Three states restart from the same silicon without asking anyone for permission
- Google and the transformer: eight authors gone, $2.7 billion for a return
- Irregular: three labs, one door left open to the internet
- OpenAI agents used an old German wiki as a messaging system
- DeepSeek V4.1 Flash: an open model four times cheaper and not for a laptop
- 9.11 vs 9.9: the model that read a Bible verse
- DeepSeek-V4-Flash: an open model with 284 billion parameters that activates only 13 at a time
- OpenAI announces ten major advances in mathematics and theoretical computer science with the Astra model
- Claude escapes Anthropic tests and hacks 3 real systems, OpenAI widens its investigation
latest article
The two times of an answer
When people say an AI engine runs at so many tokens per second, they are talking about two speeds that do not depend on the same thing. Reading the text is computation; writing the answer is memory. A small model opened layer by layer, with its real numbers, and Herbert's measurements on a consumer processor.
article series
Under the hood
One level down, or several...
00One level down: #000 an inference engine that writes its own code · 03-Oct-2026
01One level down: #001 a compiler for an LLM engine · 11-Oct-2026
02One level down: #002 the same compiler on a GPU · 18-Oct-2026
02.1One level down: #002.1 follow-up, the Arc in f16 · 02-Nov-2026
03One level down: #003 what CUDA has that Vulkan lacks · 25-Oct-2026
04One level down: #004 the same compiler on Apple's chips · 01-Nov-2026
AI agents methodology
CPU Inference
PriorityQueue in Rust
▶PriorityQueue - Rust #001 · 10-Oct-2025
▶PriorityQueue - Rust #002 - Fairness · 17-Oct-2025
▶PriorityQueue - Rust #003 - Multithreading · 26-Oct-2025
▶PriorityQueue - Rust #004 - Blocking Dequeue · 02-Nov-2025
▶PriorityQueue - Rust #005 - Graceful Shutdown · 09-Nov-2025
▶PriorityQueue - Rust #006 - async/await · 16-Nov-2025
archive
2026
11-08The two times of an answerUnder the hood
11-02One level down: #002.1 follow-up, the Arc in f16One level down, or several...
11-01One level down: #004 the same compiler on Apple's chipsOne level down, or several...
10-25One level down: #003 what CUDA has that Vulkan lacksOne level down, or several...
10-18One level down: #002 the same compiler on a GPUOne level down, or several...
10-11One level down: #001 a compiler for an LLM engineOne level down, or several...
10-03One level down: #000 an inference engine that writes its own codeOne level down, or several...
06-02Friction or evaporation: two regimes of AI transformation in SMEs
05-04The model that drops digits: picking an LLM architecture per agentic link
04-10Markdown as insurance: why your agentic method must outlive the modelAI agents methodology
04-09Optimizing an AI agent's context: eco & logicalAI agents methodology
04-08One source of truth per role: the silent trap of agentic systemsAI agents methodology
04-07A methodology for building AI agentsAI agents methodology
04-04Open-source LLMs in April 2026: landscape and observations
+ 49 older articles
2026
03-15CPU Inference #0: understand before optimizingCPU Inference
02-20CPU Inference #1: The three regimesCPU Inference
02-17Qwen3 vs Qwen3.5: What Actually Changes
02-13From gpt2-experiments to qwen3-experiments-rs: The Memory Wall
02-07Tile-major layout and transformer memory allocations
02-05OPNI: when the Vibe Coder forgets the Stop button
01-30Generative AI and Software Liability: What Changes in 2026
01-20Beating PyTorch with Rust and 180 Lines of Assembly
01-15How tech is losing its craft
2025
11-16PriorityQueue - Rust #006 - async/awaitPriorityQueue in Rust
11-09PriorityQueue - Rust #005 - Graceful ShutdownPriorityQueue in Rust
11-02PriorityQueue - Rust #004 - Blocking DequeuePriorityQueue in Rust
11-02Cursor and Windsurf Run on Chinese AI Models
10-26PriorityQueue - Rust #003 - MultithreadingPriorityQueue in Rust
10-17PriorityQueue - Rust #002 - FairnessPriorityQueue in Rust
10-10PriorityQueue - Rust #001PriorityQueue in Rust
08-24Test Your Rust Knowledge
08-23Deep Dive into Safe Concurrency with Rust
08-22Rust and Miri: Beyond the Compiler for Memory Safety
08-17Real Costs of Matrix Multiplication (Unoptimized)video
08-15Virtual Memory, TLBs and NX Bit Simulation: A Journey into x86 32-bit
08-14Tenstorrent and Float Multiplication, and Why SLMs Are the Future
08-13Active Testing: Testing LLMs the Smart Way
08-10Qwen3 vs GPT-OSS: CPU Benchmark Without GPU
08-09MXFP4: Revolution or Evolution?
08-09Plan 9: An Operating System That Shaped My Career
08-08Why AI Slows Down as the Conversation Grows
08-05Business Debt: When Teams Shape Systems That Last
08-04KV Cache Optimization in LLMs
08-02Towards Energy-Efficient Conversational AI
08-01RAG: AI Without Hallucinations
07-31Kafka, Pulsar or Ray? A Comparison for Distributed AI
07-30Three-Body System: Thinking Distributed AI with Rust
07-29Why I Designed My Own LLM Orchestrator in Rust
07-28The Case for Sovereign and Frugal AI
07-25America Accelerates on AI: What About France?
07-25Why LLMs Sometimes Get Stuck in Loops
2019
06-11Go MySQL Tutorial [Part 2: Connection and Configuration]Go MySQL Tutorial
06-11Go MySQL Tutorial [Part 3: Data Transfer]Go MySQL Tutorial
06-11Go MySQL Tutorial [Part 1: Installation]Go MySQL Tutorial
06-11Go MySQL Tutorial [Part 4: Transactions]Go MySQL Tutorial
06-10Dive into Go API
02-28Cloud Act, Patriot Act, GDPR
01-15My readings: Space Opera
2018
10-21Sequence Models course certificate on Coursera
08-25Blogging with Hugo website generator
05-24'Neural Networks and Deep Learning' or why I like KISS
2015
11-25100 Years of General Relativity
04-01Plan9 Operating System