Opensource consultant. #AIGC #Infra #CloudNative #Web3 FOSS, GNU Hacker, W3C privacy reviewer, Compiler writer, OSdever

Tokyo
Joined September 2019
Folks, we provide GNU Artanis consulting services based on @hardenedlinux artanis.dev/services.html
2
4
559
As C/C++ people, you have to be familiar with Sanitizer and its friends, otherwise Rust people will jump out to wash you.
29
SonicBoom now runs DeepSeek-R1-Distill-Qwen-1.5B with zero model-specific code: same compiler path as Qwen2.5. Day one on an RTX 3050 vs llama.cpp (same ggml ukernels): decode 157 tok/s (1.06x) prefill 0.50x: d128 attention untuned, fixing now. #DeepSeek #LLM cc @deepseek_ai
Calling all GPUs: help us benchmark SonicBoom v0.0.3 β€” a compiler-first AI runtime that beats llama.cpp on decode cc @citypw @hardenedlinux @hardenedvault github.com/hardenedlinux/Son…
1
88
Good, open-design is a good alternative to Claude Design, and you can use your preferred LLM API.
25
Calling all GPUs: help us benchmark SonicBoom v0.0.3 β€” a compiler-first AI runtime that beats llama.cpp on decode cc @citypw @hardenedlinux @hardenedvault github.com/hardenedlinux/Son…
1
1
3
203
In AI era, people rarely write code, so memory safety is a fake requirement, AI can write safe code in the next few months. It can almost do it right now. However, compiler optimizing, program verification are still needed, no matter AI or humans write the code.
1
1
50
The future of CS will not focus on how to help humans write safe, efficient code. The pattern will shift to how to ensure AI-generated code meets requirements accurately and correctly.
1
31
People may say formal verification can do this but no, it's not. A code can be verified correct according to the spec, but not to your requirements. Say, a motor can be limited to certain speed, and code can be verified, but it's just one of the cases to make the device safe.
2
17
Live in the moment, I don't think it's wise to look to the too far future. Even AI appears so smart and efficient, 99% people never make money from it.
2
31
The infra of Typesafe AI still needs to improve. The Jev is not ready for product.
1
58
As a daily viber, I won't let vibe coding touch my infra, things like web frameworks and compilers. I'll handle those myself. It's unbearable to discover a hidden design flaw caused by hallucinations or missing semantics that I never even thought through.
60
99% syntax correctness but only 7% verification success. That implies there's a huge gap between "code looks correct" and "it's semantically verifiable". dafnycomp.github.io/
28
SHA-256 is a finite-state, deterministic system. No external randomness, no accumulating entropy. Instead, each round pushes the structure of the input into increasingly higher-order relationships through nonlinear mixing and diffusion.
1
22
The goal is not to create more randomness, but to make exploitable structure progressively harder and more expensive to preserve.
7
Tokens are burning too fast in the later stage of the product development. Rather than waiting for another 5 hours, it's better to write it yourself. A few minutes to finish.
20
I'm so glad that I can push two different projects in parallel with the help of AI.
31
We have just implemented r7rs-small macro, syntax-rules in Laco. It's not based on strict syntax-case, and syntax-case is not in r7rs. However, r7rs requires macros to be strictly hygienic, which can be done with syntax-case. syntax-case is a larger work to go. #animula #scheme
1
50
The mainstream #scheme implementation may choose to integrate Dybvig's syntax-case and base the macro system on it. We can't do it in Laco, since there's more work to do. I hope to keep Laco as simple as possible, but simple is not weak. I believe I'll find a more compact way.
24
Monads are not Options. But Options are Monads. People learned Options as if they knew Monads.
43