About what I can't stop writing about
I run a tiny compiler startup: two people, pre-revenue. Like most companies at this stage, we are always short on staff, money, and time. When Claude Code and Codex became good enough to use every day, they changed our trajectory. They let us do a lot more with much less. Making it work was seriously [ ] The post Audit, Test, Automate: How We Decide What AI Can Own appeared first on…
(histoire enfantine pour adulte) Sara était assise, seule, appuyée contre une souche d’arbre. La forêt sentait venir la pluie. Un loup lui tournait autour, reniflait, s’arrêtait, repartait. Sara ne l’avait pas vu ; elle était occupée à pleurer. Il se dressa, ouvrit la gueule : ses crocs laiteux apparurent. — Pourquoi tu pleures ?— Parce [ ] The post Les fraises du loup appeared first on…
My dear startup founder friend, You told me that everything is “hyper-determined,” that “life is a big game because of that.” You’re not alone in thinking this: Silicon Valley loves to preach inevitability. A Silicon Valley paradox “Technology has its own agenda.” — Kevin Kelly https://kk.org/thetechnium/the-unabomber-w/ “Technology happens because it is possible.” — Sam Altman [ ] The post…
Writing consistent Javadoc is tedious, and even AI coding assistants often skip or misfire on it. Autodoc automates the boring part: it scans your PRs, understands your code, and generates high-quality Javadoc so you don’t have to. The philosophy: Generate 80% of the documentation (the obvious parts) automatically, and letting you fix it. That s why [ ] The post AUTODOC: ai writing your javadoc…
The author compared their tool, SuperVM, with AI models Copilot GPT4o and Claude Sonnet 4 using a simple fractal generator. SuperVM outperformed the others, running code 2 to 3 times faster. Built to optimize compiled code directly, it shows promise for specialized tools in the evolving AGI landscape, despite being early-stage. The post How my Bytecode Optimizer Beats gpt5 by 4X appeared first on…
« Le temps d’apprendre à vivre, il est déjà trop tard. » — Aragon, il n’y a pas d’amour heureux, 1943 Il est trop tard d apprendre à vivreIl est trop tardIl est... (il est tard : la vie cabosse) The post Béance appeared first on Deviant/Abstraction .
tl;dr; My auto-parallelizer PoC turns a sequential loop into a parallel one with almost no extra code. We parallelize a loop by mechanically composing its recurrence f into f⁽ᴾ⁾ without an algebraic closed form. It works with linear and non-linear recurrences (although it s not always profitable to do so). We do that by first generating [ ] The post Parallelisation by Recurrence appeared first on…
Today, I folded myself into a star and got pinned with feathers and tar to the sky. I had become your new god. I ached for the blissful emptiness of the cosmos. My Child: feed me Alas, in place of bliss, the tar scorched; pins in my wrists, feathers: only pain. Yet where I should [ ] The post venus as lucifer appeared first on Deviant/Abstraction .
Computation is a commodity—and a very cheap one. Yet it’s the cornerstone of our world, and anything that makes it faster or more efficient can have massive impact. For example, shaving just 5% off compute time can be akin to jumping a full generation in CPU performance—worth hundreds of billions of dollars. That’s the kind [ ] The post A poc of a JVM autoparallelizer appeared first on…
Introduction I think I’ve found a solution to an old engineering problem: dynamic method dispatch. It happens when the program decides at runtime which piece of code to run. This mechanism is fundamental to method calls in Javascript and Python and widely used in any OOP programming language (and now even in Rust). I’m explaining [ ] The post Making Computer Faster: a Deep Dive into Dynamic…