low latency guy

San Francisco, CA
Joined June 2009
Code and sleep retweeted
🚨 @NVIDIAAI just quietly dropped its own local inference server... No announcement. Single-digit stars It's called Flare and Apache 2.0 Built for SM120/SM121 GPUs RTX PRO 6000 Blackwell and DGX Spark >PyTorch + FlashInfer at build time >Rust runtime at serve time. No Python. >CUDA graphs captured before serving >FP8 KV cache by default >Recipes for Qwen Flash, GLM and DeepSeek github.com/NVIDIA/flare-loca…
33
75
21
573
48,223
We ditched Devin, Gemini, Greptile for code reviews, since Claude and codex can do same things cheaper and/or better. And their capabilities are included into the subscription.
31
Коллеги из Смартли некоторое время использовали домен smartly.af (читается «smartly as fuck», «дохуя умный») для внутреннего artifactory, но потом решили перестать поддерживать экономику Афганистана :-)
Every once in a while, some unsuspecting country's domain economy goes vertical 😂 .𝚌𝚘 → Colombia .𝚒𝚘 → I/O → Island Ocean .𝚖𝚍 → Markdown → Moldova .𝚛𝚜 → Rust → Serbia .𝚊𝚒 → AI → Anguilla .𝚜𝚒 → SI → Slovenia
2
1
107
10,436
If, on the other hand, you want to get the job done at a reasonable cost, use appropriate model, effort level, and find the fastest way to get things done. This will most likely involve ai and token usage, but judge yourself on work done, not tokens consumed.
If you get such a message from your boss, Claude and Codex have your back! Turn effort level to Max first. This will slow things down, and the bill will rise, so will your token usage. Next, Claude has a secret cheat: Ultracode. It increases token usage 2-3x easily. Do everything with Ultracode On. Stupid rules - stupid measures.
25
If you get such a message from your boss, Claude and Codex have your back! Turn effort level to Max first. This will slow things down, and the bill will rise, so will your token usage. Next, Claude has a secret cheat: Ultracode. It increases token usage 2-3x easily. Do everything with Ultracode On. Stupid rules - stupid measures.
A middleware engineer in South Korea wrote to me, seeking help on what to do now that their manager is "blinded by GenAI hype" and set an "AI token usage" KPI. This eng is using v little AI, so manager is poking them to use more. My response: your manager is trying to help you
1
132
I completely agree. Python is a pain. If not libraries, even TypeScript is way superior to Python from developer’s perspective. Let alone Rust.
there is zero reason to use python in 2026 it's inferior in every single way and the only benefit it ever had (ease of use) is completely meaningless now that agents can write all the code
46
With agents around, you have no excuse to not have close to 100% test coverage of your code.
1
52
"Context size: model steps take about twice as long once the session context passes 600k tokens." Thank you, Claude.
28
Пока прогноз сбывается
Прогноз количества датацентров у Яндекса @yandex
1
44
Тот кто знает насколько убогое качество связи в Сан-Франциско и калифорнии в целом, сразу поймет, зачем нужен спутниковый интернет
Replying to @xsync
Короткий ответ — это выдумка из исходного поста. Но дальше интереснее. У них есть 3 направления: легаси, будущее и наземное. Легаси покрывает существующие телефоны и работает уже сейчас в части диапазона 1.9ГГц от T-Mobile, самое главное по протоколам 4G LTE. Там очень низкая скорость, килобиты буквально. Хватит отправить смску кое-как, работает как экстренный сервис уже сейчас. С новыми спутниками DTC Gen2 (сегодня аппрувнули), которые намного мощнее и их будет 15000 вместо сотен, они обещают до 100 раз прирост, то есть в теории хватит на аудио звонок в сносном качестве, это их таргет. Но ключевое — на всех актуальных устройствах. Будущее куда интереснее, тк год назад они у EchoStar выкупили за $17 миллиардов весь спектр частот AWS-4 (50 МНц S-band в США), H-block и MSS (международный). Всё это нужно для работы 5G связи на спектрах 1.9ГГц и 2ГГц. Но это проприетарный спектр, он не поддерживается существующими модемами. То есть им надо работать с Qualcomm, Apple итд, чтобы внедрить поддержку в их модемы и новые стандарты по антеннам. На это уйдёт ещё год минимум (договор с Apple тяжело идёт). Но через DTC Gen2 и в хороших условиях они смогут обеспечивать пару мегабит (я слышал про таргет 2.5), а этого хватит на видеозвонок и даже ютуб в низком качестве. И тут мы переходим к наземной инфраструктуре, которую они втихаря скупают уже некоторое время. Орбитальный Starlink не пробивает до антенны телефона, а вот вышка на земле да. Сегодня они заключили сделку на приобретение ещё одного национального портфеля частот — низкочастотный спектр до 14МГц парного спектра в диапазоне 800МГц в США. Это позволит им закрыть дыру в покрытии, тк этот спектр пробивает сквозь стены, что не может орбитальный Starlink в среднем 2ГГц спектре. Самое главное, что этот спектр поддерживается, но не используется актуальными мобильными устройствами. Впереди одобрение от FCC. Тут уже также мегабиты, но стабильные. Но главный джокер пока не вытаскивали — они работают над возможностью раздачи связи для мобильных устройств через домашние терминалы Starlink, что сильно расширит всю сеть. Так что Starlink Mobile уже будет не сервисом по модели last mile, а полноценным пятым мобильным оператором в Штатах, но УТП в покрытии и бесперебойности.
1
2
104
Clown metrics for clown rewrite. E g compare numbers for codex in Rust to Claude in TypeScript. Do you see the difference? Me neither
Replying to @PrimeIntellect
Across the six harnesses we benchmarked, Prime Agent Rust had the lowest time to usable input, startup memory, and installed size. The rewrite also split the codebase into nine crates with enforced dependency boundaries. Changes to shared protocol types are now checked across the client, daemon, and session workers.
43
I absolutely hate video explanations. Who decided that it is a good idea to watch 5min video instead of spending 15sec to skim through the text.
We just let in every Pro and Max user from the Claude Code Projects waitlist! If you're new to Claude Code Projects, here's a 4 minute walkthrough to get you started:
23
Прогноз количества датацентров у Яндекса @yandex
В ночь с 7 на 8 октября дата-центр Яндекса в Сасово был серьезно повреждён из-за атаки БПЛА. Его работа полностью остановлена. Сегодня утром произошла атака на ещё один наш дата-центр, в Калуге. Часть инфраструктуры вышла из строя.
1
1
120
Let’s call it real life testing
We have many ways to ensure software quality. Unit testing. Fuzz testing. Integration testing. And so forth. I'd like to propose a method that was unthinkable before: ephemeral testing. (Ephemeral is a fancy word for 'throw away' or 'temporary'.) You write your code. You build your software component. Or the AI agent does it for you, it does not matter. Then you ask an AI agent to build on it: an application, another layer, maybe several. You have it test what it built. You do not assess the original work directly. You assess how good the software built on top of it is. It is a form of integration testing. The difference is that the software on top is entirely ephemeral. You throw it away when you are done. A library with a clean API, stable invariants, and useful errors lets the agent produce something that works quickly. A library with hidden state, surprising defaults, or incomplete docs produces a pile of patches and failures. The failures are evidence about your code, not about the agent. You can repeat it. Different agents, different tasks, same foundation. In effect, instead of building the core while trying to anticipate what might be needed at the other layers, you just simulate the other layers by actually building them. Of course, you could argue that with AI, you can rebuild everything whenever you need to. But that's not practical. You need some form of stability. I have been applying this trick to various projects. As I consider a new feature, I ask my AI to prototype quickly what I might later build based on what I am doing it. Ephemeral testing works for me thus far.
16
Code and sleep retweeted
I gave a talk on data parallelism this week, and I discussed 'runtime dispatching'. It is a trick whereas you check what the CPU supports (at runtime) and selects the right code path. The engineer asked 'what happens if the processor speculates wrong?' Indeed, whenever there is a branch, a compiler might try to evaluate the wrong branch. If you try to execute an instruction that your processor does not support, you typically get a crash. So it might be possible that you would check what the CPU supports, and somehow end up on a code path with unsupported instructions due to speculation. So would it crash if it speculated wrong? I said that I expected that there should be no crash, and no side-effect. I am happy to report that my answer was correct. Civilization saved, once again.
11
8
1
139
8,711
When they say it beat both Opus5.5 and Sol-5.6, I am expecting some catch.
Introducing Claude Sonnet 5.5, the second model in the Claude 5.5 family. It’s a clear upgrade over Sonnet 5, runs more than 30% faster, and costs up to 30% less for most work.
2
60
Code and sleep retweeted
Opus 5.5 is the result of your feedback. It communicates clearly, it's cheaper per token than Opus 5.0 with the intelligence of Fable 5.1, it’s very token efficient and works across every effort level. We're also increasing 5h rate limits & giving you a banked reset.
Introducing Claude Opus 5.5, the first model in our new Claude 5.5 family. It performs at the level of Claude Fable 5.1 for most tasks, and costs 40% less to run than Opus 5.
760
413
169
12,112
686,135
Code and sleep retweeted
I used Opus 5.5 to formally verify the Claude Agent SDK using Lean. A couple short prompts = 16 PRs fixing various bugs and race conditions. Video attached. TLA+ also works well. I sometimes combine Lean and TLA+ to look for issues around data flow, concurrency, and state mgmt. I don't know either language well, but Claude is excellent at both. This approach is super useful for formally modeling your code and finding bugs that a human probably wouldn't have spotted. Is formal verification the future of coding (or at least, bug finding)?
489
341
301
6,289
2,147,059
Crazy prediction from 2021: in 5 years you'll be programming in a terminal and switching between multiple sessions. Not gonna happen, of course.
37