
PE @ Amazon Alexa | building @LastRidgeHQ for AI tooling | applying agentic engineering to all aspects of my work
USA
Joined June 2011
- Tweets3.1K
- Following3.1K
- Followers1.9K
- Likes7.3K
Armen Shimoon retweeted
CuteAFD v4.1.1
Two fixes for DS Flash 4.1
- scratch space scribbling was possible reducing fidelity 20-40% under some scenarios
- hammering the server with the same request for benchmarking would make it chain defer them to reuse the prefix leading to it looking like C2 was the limit of performance
If you’re using a different model, no changes.
Thanks to @ArmenShimoon for debugging with me!
Random test card: 1xRTX+4sparks DS41
We changed the benchmark in v4 so that every prompt during concurrency is unique, so you don’t get expert sharing juicing the curve, and additionally the prefill for the request is fully unique, so it makes the curve concave vs convex…. Oh noes my ePeen numbers are smaller 😭 prefill is also real non repetitive text so it also can’t cheat on experts… but prefill gets a lot of expert sharing anyway. c1 is c1 hard to do crazy shit there unless you go sampling based and accept not top-1 token….
CuteAFD v4.1
Kernel Optimizations and Engram prefetch
Quick follow up release….
AI made a dumb kernel swap to a pre-rotation kernel on GB10 which is not ideal, wastes memory bandwidth. It did it while it was optimizing the RTX onboarding. 20% loss of prefill speed for several families like GLM.
Also I added a sophisticated preloaded for engrams that starts streaming them as soon as tokens are drafted and into the future when doing prefill (thanks @dangerm00se). It’s optimized for both local SSD and SparkNest remote storage. 40%+ cold performance on DeepSeek and Qwen Flash Next. Big help in a steady state on low RAM or highly busy coordinator too!
github.com/tpurtell/cuteafd/…
If you actually read your code nowadays, you run the risk of getting too attached to it rather than letting your agents iterate and refactor freely
If that scares you, your guardrails are insufficient
Armen Shimoon retweeted
CuteAFD v1
Early access release
Many models
Many hardwares
Cool console includes perf and benching
Not a lot of fanfare… lots still in progress… will highlight features in later releases as it comes to maturity
@dangerm00se planning to help out on 5090 side so this is mostly 6000 only but it has memory planner options and stuff
github.com/tpurtell/cuteafd
In tech, somebody who debates or argues with you hard (but in good spirit) but then praises and supports you behind your back
Is far more valuable than
somebody who praises and supports you to your face but is resistant or even indifferent to you behind your back
Noticing a Claude Code bug @claudeai
When I have more than one instance of claude running, sometimes the tool output from instance 1 shows up in instance 2. I don't think it actually affects instance 2, just displays there. Very weird and seems unintentional.
Anybody else?
1
123
Absolutely delighted with the spotlightmaxxing loop OpenAI and Anthropic are engaged in
They compete, we win
Please welcome GPT-6 Sol and GPT-6 Luna to the GPT-6 universe.
GPT-6 Sol and Luna build on the advances behind GPT-6 Astra, bringing much of its strengths into faster and more affordable models to support work at scale.
We’ve also made caching and inference more efficient, and we’re passing the savings directly to you: 50% lower API prices for Sol and Luna compared with GPT‑5.6 promotional pricing.
2
93
Why is there no "Ask Grok about this" button on X posts?
I don't always want to have a public conversation with @grok, I want to dive deep privately
1
119
There's a run on Sparks at Microcenter
Most likely everywhere.
Today from 9:45 to 3:30 EST we've gone from 326 in stock down to 278.
My store that had 17 left yesterday now has 1.
For those wanting to get into local AI and have it actually be useful, this looks like the best setup.
Two to four sparks as your backend to hold big models.
A single 5090 or 6000 make those big models rip.
Replying to @wrldsuksgo2mars
This research was inspired by @dangerm00se who was tempted by @MiaAI_lab awesome 3-spark no-RTX recipe. She got the OG model to fit, so I guess you should be able to do it.
github.com/MiaAI-Lab/DeepSee…
So with an RTX/5090 handling the attention...
37 tps -> 71 tps C1 "prose" decode
3800 tps -> 7000 tps prefill
This kinda makes sense, the compute capacity of the RTX 6000/5090 = 4x spark and the memory bandwidth is 6x Spark... spark just comes with the memory juice that makes playing this system architecture game worth it....
Literally 2x performance adding the RTX card into the mix.
1
89
800 GB memory.
4x spark, 2x 5090, and a few macbooks. (Started with the macbooks then things escalated).
I built vram[dot]fyi for building and comparing local AI setups, and of course for dreaming about my ideal setup
Here's mine vram.fyi/?b=apple-m3-max-40-…
1
141
For anyone running two DGX Sparks over the ConnectX cable: rsync between the nodes was crawling despite the 200 Gbit link, so I wrote a small tool that uses both fabric paths and skips SSH for the payload.
51 GiB checkpoint: 3m14s → 16s. About 12× faster than rsync, 73% faster than rclone (existing best in class), and within 15% of the NVMe's read ceiling itself. Benchmarks vs rsync, rclone and others in the repo.
Recipe here:
github.com/ashimoon/spark-co…
1
99
With agentic engineering, it's not about how much you use coding agents.
The only thing that matters is whether the tokens you produce are more valuable than the cost of your input tokens.
You wanna spend $100 on input tokens to produce a tool that makes $200. That's it.
4
120
I use Claude Code for anything beyond trivial terminal commands nowadays.
Its the first thing I install when I spin up new VMs/containers. The rest of the setup, and effectively the terminal work after that -- I run through Claude.
I do not remember shell commands :) and I needed to kill a process running on port 21079. Instead of struggling to call and running a couple of commands, I built a small utility called `x`.
> x kill process running on port 21079
This utility is a pure bash script with no external dependencies (just curl or wget). You describe what you want in plain English after "x", and it generates the shell command and executes it only after getting your confirmation.
> x get all the git branches
> x delete all pycache and useless python files
> x grep all errors from ~/access.log file
It supports OpenAI, Gemini, and Anthropic as providers. You just set one API key, and you are good to go. I have open-sourced it, and the repo is linked below. Give it a star. Also, feel free to fork it and tweak it.
Installation takes a single command and barely a minute. Give it a spin.
2
7
123

















