Interactive, long-form writing on AI systems, infrastructure and algorithms by quanglee, senior software engineer in Paris and math, tech and AI enthusiast.https://quanglee.dev/https://quanglee.dev/blog/mistral-large-4/https://quanglee.dev/blog/mistral-large-4/Mistral Large 4 is the best model built outside the US and China and a big jump from Large 3. Independent tests put it about seven months behind the US labs and behind five Chinese open models, and a task on it costs about fifteen times more than on the small US models with the same score.Mon, 12 Oct 2026 00:00:00 GMTaillmbenchmarkshttps://quanglee.dev/blog/messi-in-numbers/https://quanglee.dev/blog/messi-in-numbers/Lionel Messi made football look like art, and no spreadsheet can hold that. But the spreadsheet still says it, plainly. The best forward who ever played and the best playmaker who ever played are the same person.Fri, 09 Oct 2026 00:00:00 GMTfootballdatahttps://quanglee.dev/blog/how-lean-checks-a-proof/https://quanglee.dev/blog/how-lean-checks-a-proof/Lean is a programming language in which a theorem is a type and a proof is a program, checked by a small kernel. This post covers how that works, where it can still go wrong, and why it lets language models write proofs nobody has to read.Wed, 07 Oct 2026 00:00:00 GMTmathleanaihttps://quanglee.dev/blog/openai-722-math-manuscripts/https://quanglee.dev/blog/openai-722-math-manuscripts/On 6 October OpenAI published 722 machine-written math papers with claims on the Unique Games Conjecture, the colouring of the plane, a $5,000 Erdős problem and dozens of other open problems. This post goes through the claims, the parts that come with Lean proofs, and the first day of reactions.Wed, 07 Oct 2026 00:00:00 GMTaimathleanhttps://quanglee.dev/blog/deepseek-v4-1-vs-glm-5-3/https://quanglee.dev/blog/deepseek-v4-1-vs-glm-5-3/DeepSeek V4.1 and GLM-5.3 compared on agentic work and coding in two weight classes, next to the closed models behind Claude Code and Codex, with benchmark scores by task type, speed, cost per solved task and reports from daily use.Tue, 06 Oct 2026 00:00:00 GMTaillmagentsbenchmarkshttps://quanglee.dev/blog/how-a-transformer-reads/https://quanglee.dev/blog/how-a-transformer-reads/A walk through the parts of a transformer, from tokens to sampling, with a working figure for each step.Sat, 03 Oct 2026 00:00:00 GMTaillmtransformershttps://quanglee.dev/blog/leetcode-1977-separate-numbers/https://quanglee.dev/blog/leetcode-1977-separate-numbers/A hard DP that's worth knowing, because it stacks three classics in one problem. Partition DP, prefix sums and an LCP table take it from O(n³) to O(n²).Mon, 22 Jun 2026 00:00:00 GMTleetcodedynamic-programmingalgorithmshttps://quanglee.dev/blog/kv-caching-explained/https://quanglee.dev/blog/kv-caching-explained/What actually happens when a provider says your prompt is cached. The trick underneath is called KV caching, and it's a small idea with a big memory bill.Thu, 30 Apr 2026 00:00:00 GMTaillmtransformerscachinghttps://quanglee.dev/blog/agentic-ai-cost/https://quanglee.dev/blog/agentic-ai-cost/Agents loop, and every loop resends everything. Where the tokens actually go, how prefix caching claws most of it back, which benchmarks matter when you pick a model, and a router you write yourself.Tue, 28 Apr 2026 00:00:00 GMTaiagentscostinfrahttps://quanglee.dev/blog/elasticsearch-best-practices/https://quanglee.dev/blog/elasticsearch-best-practices/What I check first when an Elasticsearch cluster gets slow. Shards, segments, refreshes, and the profiling result that changed how we fetch ids.Tue, 15 Jul 2025 00:00:00 GMTelasticsearchinfraperformancehttps://quanglee.dev/blog/notes-on-making-this-site/https://quanglee.dev/blog/notes-on-making-this-site/The site, twice. A one-font page I shipped in april, then a rebuild with figures you can poke, a small green critter and a status bar.Sun, 27 Apr 2025 00:00:00 GMTmetabuild-logastro