Sümer Cip

Sr Sw Eng working on observability. Specially interested in observability tools & databases & low-level stuff. Dad. Sports nerd.

[Sümer Cip @sumercip]

I’ve long wanted to write a blog post on Patterns for Building High-Performance Observability Pipelines at Scale and here it is:

https://t.co/drDuSoEzkx

Enjoy!

[Sümer Cip @sumercip]

@GargeyaS My journey started like this then I ended up adding local models with obsessing about performance. see https://t.co/jazMnoGi2d

[Sümer Cip @sumercip]

@marc_in_space Same.

BTW, are you sure? Because I am feeling watermarking might be coming before the announcement. Maybe A/B testing :)

[Sümer Cip @sumercip]

I cannot be %100 sure. But I am suspecting the reason Opus models start to produce so much complex wordings recently mght be related with watermarking algorithm? I see weird words choices even though I have strong rules everywhere to use STE-100... I suspect this is forced even more than CLAUDE.md rules. Fix: a separate tool that runs on PostHook that deterministically rejects a non-STE-100 text? :)

[wafer @wafer_ai]

🚨 GPT-5.6-Sol Ultrafast mode (750 TPS) is here powered by Cerebras!

Here's How kernels work on Cerebras Chips

cerebras built a chip with 900,000 cores on a single silicon wafer.

CSL (Cerebras Software Language) is a Zig-inspired DSL that gives you direct control over the Wafer-Scale Engine. the SDK is publicly documented and the programming model is genuinely the most alien one i've ever looked at. if NKI felt different, CSL takes it much much further.

in CUDA you write one thread's perspective and launch millions. in CSL, there are no threads, no warps, no shared memory, and no kernel launches. you write code for individual Processing Elements: 900,000 independent cores arranged in a 2D mesh on a single silicon wafer. each PE has its own 48 KB of private SRAM, its own program counter, and a 5-port router connecting it to its 4 neighbors. that's actually it. no DRAM. no HBM. no cache hierarchy. 48 KB is your entire world per PE, where code and data must both fit.

the programming model is dataflow, and this is where it gets a bit more alien 👽

data moves between PEs as 32-bit messages called wavelets, traveling along virtual channels called colors. when a wavelet arrives at a PE on a specific color, it activates a task (a chunk of code bound to that color at compile time). tasks run to completion, then hardware picks the next activated task. tasks cannot call each other. they can only be activated.

so instead of "launch N threads," you think of it like: "place code on PEs, define routes, let data flow."

the memory model is also very different from GPUs.

on an H100 you get 80 GB of HBM shared across all SMs. on the WSE-3, memory is 48 KB per PE, and there are 900,000 of them! this gives you 44 GB total on-chip SRAM at 21 PB/s aggregate bandwidth (vs 3 TB/s on H100). every access is single-cycle. no coalescing needed. no bank conflicts. no cache misses. but also no way to access another PE's memory. all inter-PE communication is explicit wavelet routing through the fabric.

take a distributed GEMV for example. you would write a layout file that physically routes wavelets across the mesh. two PEs sit side by side. the left PE computes a partial result and routes it eastward. the right PE receives those wavelets from the west and accumulates. routing is defined at compile time. both operations are asynchronous and activate a task when they finish. you're physically routing data across silicon at 1 clock cycle per hop.

cerebras is an extremely technically interesting & alien beast. reports 95-210x speedups over H100 on stencil computations. 3,000 tokens/sec inference on gpt-oss-120B. $10B+ deal with OpenAI for 750 MW of inference infrastructure. very exciting times for alternative accelerators!

@gpuemi technical deep dive

[Sümer Cip @sumercip]

@badlogicgames @jeremyphoward respect!

[Sümer Cip @sumercip]

@jeremyphoward @badlogicgames Sit the fuck down? ? Who do you think you are man?
I never disrespected anyone. I rly wonder if you can say those words to my face.... I am only telling someone I respect something that that might be helpful...

[Sümer Cip @sumercip]

@bcherny @DavidOndrej1 this bug happens when claude oauth token is set for me. when unset it wirks normalky. it was like this for the past 2 weeks with latest version.

[Sümer Cip @sumercip]

@mattpocockuk BTW, I really find "explain X to me from first principles" useful at times fighting for this kind of problem.

[Sümer Cip @sumercip]

@sawyerhood I have not used bb but I honestly believe this is the future of sw. Core baked in properly in a extensible way and the rest is tailored... self modifying sw.

[Can. @cancelik]

i still sometimes can’t comprehend that a tool i built for my own needs a few months ago reached this many people, and now gets praise from people i admire.

easy to forget that you can just do things.

[Sümer Cip @sumercip]

@VictorTaelin I look at this like this: everytime AI generates some code, it generates a little bit of shit. The problem is compounding interest is so strong that in time it becomes a shithole if you don't care enough. How much you shoul care and where is our new job imho.

[Sümer Cip @sumercip]

@marc_in_space You know how much I love to lag on the most important parts :) Will def. do it, thx!

[Sümer Cip @sumercip]

https://t.co/IanVTo6aGb I did a big release 0.4.0:
- Local Whisper/Parakeet models running very fast on Mac (subsecond for ~30s audio)
- Recording overlay window
- Better installation UX (single https://t.co/FJbAGdZSKc and update scripts)

P.S: I really suggest everyone to use some STT app if not already(esp. with a coding agent), it does not need to be Zee. It will be a massive improvement.

[Sümer Cip @sumercip]

And special thanks and shout out to @mudler_it for https://t.co/kQL9fGqedz. Incredible STT performance even on M1 for English: ~250ms range for ~30s audio.

[Sümer Cip @sumercip]

Sometimes the sync goes the other way — the code gets refactored to match my mental model.

[Sümer Cip @sumercip]

Code review is no longer reading code. It's syncing my mental model with the codebase — through the LLM.

[Sümer Cip @sumercip]

@mkratsios47 "We have information..." is all we need to believe you?How would you explain a distilled model can outperform the base model, not just a single domain across the board?

[Simon Willison @simonw]

I got some really useful Claude prompting tips from @_catwu and @trq212 - it's time to stop overloading our prompts with examples and lists of things not to do, Fable works better without those

Claude Code's own system prompt recently shrunk by 80%!

[Sümer Cip @sumercip]

@ulusalomer matt pocock, simon willison, mario zechner armin ronacher guncel bu konularla ilgili surekli icerik uretiyolar. bu adamlari kesin takip etmeni oneririm.