RSSAmplifier

Blog

Shiv After Dark

Shiv's thoughts on cs, math and art

shvbsle.inRSS feed ↗10 posts

Latest posts

Various LLM smells

Looks like this ended up on the HN front-page: HN Thread Late last year I started writing a math blog and decided to use LLMs to polish/enhance my writing. The LLM generated writing obviously felt significantly better than my own writing. It had better vocabulary, interesting sentence structures etc etc. I swear it did not seem like AI-slop to me at the time. Then about 3 months later, I see the…

reflections on ai-infra from a private lunch

Last week I attend a lunch event hosted by a16z. The people in attendance were AI-researchers and AI-infra engineers. I compiled a list of companies where I think the talent density is ridiculously high (based on my interactions with them): Meta Training and Inference Accelerator (MTIA) Anthropic Black Forest Labs Together-ai Numa Labs Mirendil Inferact The people from the above companies were…

Introduction to distributed ML training and straggler ranks

CROSSPOST from blog.k10s.dev This blog is written by a human [1] Say I'm running a distributed ML training job. A total of 4 GPUs in my cluster and one rank mapped to each GPU. I look at the logs for one of the ranks and I see this: [21:10:49] [rung3-a Rank 0] epoch=0 step=260, loss=4.1811, elapsed=233.3s, dt=8.90s [21:10:58] [rung3-a Rank 0] epoch=0 step=270, loss=4.3926, elapsed=242.3s, dt=9.00s…

An exercise in value creation

This year I've not had a lot of time to post on my blog. And the reason is that I'm in the arena. The good thing about being in the arena is that you absorb and learn things much faster than contemplation and reflection. Bu the trade-off is that the gain in wisdom and intrinsic value evades the conscious plane of the mind and takes a deeper and more permanent seat in your subconscious. To keep…

[NOTES] Kubernetes vs SLURM: Scheduler Architecture Notes for 2026

"Technical notes comparing Kubernetes and SLURM schedulers: gang scheduling, fair-share allocation, backfilling, and custom scheduler patterns."

Hyperscalers are hard

How NFTables Drives Kubernetes Nodes to Memory Exhaustion at 100k Scale OR How Kubernetes optimizations become catastrophic at hyperscale.

eBPF, Race Conditions and Soham Parekh

Google CTF challenge with ebpf, kernels, linux-security-modules (LSMs), race-conditions and file descriptors. Some meditations on the Soham Parekh fiasco of 2025

[NOTES] CS 6515: Intro to Graduate Algorithms Recap

Stretching my will power for Grad Algorithms final exam

Underestimating the capacity for evil

"Bill gates started COVID to spend one last year locked in with Melinda" sounds outrageous but, when it's coming from Peter Thiel you start considering it. babe wake up, a new conspiracy theory just dropped bill gates started covid to spend one last year locked in with melinda pic.twitter.com/k9BkcGrnWk — vittorio (@IterIntellectus) August 17, 2024 A very dear friend had the following to…

Curriculum Vitae