Tweet fixado
normal people read docs to learn @modal i rebuilt the entire platform from scratch
distributedhatemachine.github.io/posts/modal
github.com/wtfnukee/minim…
@charles_irl pls hire
English
Egor Konovalov
52 posts
@bskdany why twitter just now discovered shirts that i've seen in china uniqlo this summer
English
@foldll @modal @charles_irl @can @charles_irl any chance you can gift this guy some credits? this tool looks pretty useful.
English
built kernelscope - CUDA kernel debugger that maps source lines to PTX + GPU events
runs entirely on @modal btw
analyzing warp state, memory coalescing, warp divergence, SM occupancy, hints 4 perf
more info&pics in next post
@charles_irl @can this is my job application part 2
English
@gazorp5 @modal @charles_irl @can sorry, run out of credits, will be up in couple days when billing cycle reset
English
Do you love hacking on cool stuff with coding agents and sandboxes? Want to get paid doing it full time?
If yes, we're hiring a sandbox dev rel – jobs.ashbyhq.com/modal/5113f35e…
English
there were only two problems: Linear connector didnt work smh, I had to install Chrome so Claude for Chrome extension would work. in the end i found cli importer and docs, that did the trick
English
yet again talking about reading list
finally got access for @claudeai cowork and tasked him with organizing my tabs right away
now i have @linear workspace with all my blogs/papers/etc organized by projects, each entry is issue, where i can change status, dates and more
each JD is issue in Job Application that has references to other projects/issues so I can cross something from my reading list while preparing to interview
seems like I finally solved my problem, other than solving FOMO
English
genuinely dont understand why this post didnt activate @tekbog sleeper agent
English
Oh, you're using Copilot? Everyone's on Cursor now. Just kidding, we're all on Windsurf. We're using Cline. We're using Aider. We have an in-house MCP server mesh with custom tool schemas but wait, OpenCode just dropped so we're migrating to that instead. Our PM is on Gemini CLI. The team lead was on Codex but now she's back to copy-pasting into ChatGPT. If you're not on Amp, you're ngmi. Our intern is building on Goose for our internal tooling. Our CFO approved Claude Max so now we're porting our workflows to computer use. Our CTO is working on an agent-less RAG pipeline so we won't need vibe coding anymore. Our CEO thinks we're talking about actual vibrations. We're building clankercloud.
English
chatgpt was right.
thanks everyone that liked/followed/dmed and encouraged me in any way lately!
English
you've commented if i will opensource kernelscope
here you go! github.com/wtfnukee/kerne…
not my best code, but you asked for it
also popular question was my reading list, i have 'readed' list in my blog distributedhatemachine.github.io/til
English
English
@syntrocode tried to do this actually! but models (both oss and proprietary, even manus) struggle with 500 tabs. maybe skill issue, ngl
English
gm my fellow gpu lovers, what should i do next? ml perf/infra interviews coming up, trying to lock in
i have loong reading list on cuda, dl frameworks internals, sysdis, some perf case studies and bunch of puzzles, so requesting some tips on organizing/prioritizing & must reads
English
@secemp9 exactly
x.com/foldll/status/…
Egor Konovalov@foldll
Oh, you're using Copilot? Everyone's on Cursor now. Just kidding, we're all on Windsurf. We're using Cline. We're using Aider. We have an in-house MCP server mesh with custom tool schemas but wait, OpenCode just dropped so we're migrating to that instead. Our PM is on Gemini CLI. The team lead was on Codex but now she's back to copy-pasting into ChatGPT. If you're not on Amp, you're ngmi. Our intern is building on Goose for our internal tooling. Our CFO approved Claude Max so now we're porting our workflows to computer use. Our CTO is working on an agent-less RAG pipeline so we won't need vibe coding anymore. Our CEO thinks we're talking about actual vibrations. We're building clankercloud.
English
oh you're using anthropic api? oh no now it's claude max sub. actually it's too ratelimited, we use glm now. we switched to grok code fast. gpt5.2 IS THE BEST RIGHT NOW wait no, it takes 20min for a single prompt, nvm we go back to claude models. oh, turns out they give us 2k worth of tokens for 200 per month on the max ultra pro sub, nice. nvm they ban you if you work at a competitor. oh, they also ban you if you use it outside of claude code, we go back to glm. didn't you hear? we're using claude black sub now, it's from opencode-
English
@AISloppyJoel little more useful is what i already read
distributedhatemachine.github.io/til/
because i sort low signal posts/blogs and write short description
i used to track this kind of stuff in obsidian, but idk, i remember almost everything i read and can find it fast so i dont see point in this
English
@AISloppyJoel idk how to share but i probably can pipe sidebar tabs into text file
now i have so much to read so i vibecoded some tool to visualize my tabs so i can take some skill paths like in videogames, but it doesnt cut it
English
Great post! But we often see GPUs that report 'Healthy' (full clocks & links up) but suffer from massive internal instruction replays or silent packet retransmits.
If you skip active benchmarks at boot, isn't there a high risk of one 'hollow' node tanking an entire distributed cluster? Or a node might have 'healthy' links, but if the cloud provider allocated it on a different spine switch than the rest of my cluster, my training run effectively dies.
English
GPUs are unreliable at scale.
At @modal we've scaled to 20,000+ concurrent GPUs across AWS, GCP, Azure, and OCI, with 1M+ instances launched. Public-cloud GPUs fail in many ways, and we’ve seen most of them.
Here’s how we handle GPU reliability 👇
English