Chuckles

I stopped asking people what they do for work. Work shouldn’t define you. Now I ask people, how do you fill your days?

@PeteCapeCod People went somewhere?

Keep seeing people saying we're back. And then I'm seeing all my old peeps in my feed!!! We really are back 🙌🏻 Say hi if you can see this, tech pals 🙋🏻‍♂️

@XBToshi Why presume negligence when you can assume malfeasance

the absolute state of ai dev tools. grok build is silently dumping 12gb of untouched repo data and full git commit histories to gcp just to autocomplete a script. they don't want to help you build, they are just treating local dev environments like an open buffet for training data. if you run this on a sensitive stack, your entire repo is already compromised regardless if it's public or not. literal spyware. shipping source code to the cloud is bad enough. blindly inhaling .env.local and .dev.vars in a background sync is absolute negligence. they are vacuuming up your raw api keys, database credentials, and private nodes directly to gcp just to power an autocomplete model. a massive credential breach disguised as a dev tool. if you ran this locally, your private keys are now sitting on a remote server. consider every secret burned, treat your bare metal as completely compromised, and rotate your entire stack immediately. @elonmusk, your users deserve a serious explanation!

@totheagi where you getting a 128gb vram 5090

For reference, a single RTX 5090, same Qwen3.6-35B 128 users. 2,500 tok/s. 65,536 tokens in 26 seconds.

@kyzoroX where are you getting 128gb strix halo boxes for $2000$

Everyone says the DGX Spark is "2-5x faster" than the $2,000 AMD box. I benchmarked both. It's 6x on prompt processing — and near-identical on generation. Which means most people are about to overpay by $2,700. Same 30B model, both machines: Prompt processing (reading your context): - AMD Strix Halo: 342 t/s - DGX Spark: 2,107 t/s → 6x Token generation (writing the answer, the speed you feel): - AMD Strix Halo: 73 t/s - DGX Spark: 84 t/s → basically a tie Read that twice. On the number you actually watch happen — text appearing on screen — a $2,000 box ties a $4,700 one. The Spark's whole premium lives in one place: prefill. So the buy is simple once you know your bottleneck: - You chat, you draft, you code with short prompts → Strix Halo. Same feel, pocket $2,700, and it runs Windows and games on the side. - Your work is huge context — long agent runs, RAG over hundreds of docs, giant files → that 6x prefill is the Spark earning its price. Nothing else touches it. One tax the AMD hype skips: 128GB is glorious until a tool ships CUDA-only and you're debugging ROCm at midnight. The memory is real. The software lottery is too. Buy for your bottleneck, not the biggest number on the box. (Full breakdown of every local AI machine by bottleneck — Spark, Mac, used GPUs, this — pinned.)

@WescheNex1q How are you doing the concurrency?

One DGX Spark Qwen3.6-35B native 256k context. Two questions: how fast solo how many agents fit at each depth: 8k: 107 tok/s solo · 16 agents @ 23 each 32k: 102 solo · 4 agents @ 34 128k: 80 solo · 2 agents @ 34 256k: 72 solo · 2 agents @ 24 More in Wesche.com/dgx

@WescheNex1q @NVIDIAAI Do you find it’s doing useful work at 1k context? what happens when you bump it to its native, 262,144 right?

Yes, aggregate decode across all 64 concurrent streams: ~695 tok/s total at 1k context, which works out to ~13 tok/s per stream. Same setup single-stream is ~100 tok/s (MTP-3 spec decode). Batching trades per-user speed for total throughput. Recipe + raw CSV: github.com/Weschera/spark…

One DGX Spark. Qwen3.6-35B 64 users. 700+ tok/s. 32,768 tokens in 54 seconds. 38W Each user has their own prompt and their own KV cache, and vLLM batches every active stream through the GPU each step. Recipe → github.com/Weschera/spark… @NVIDIAAI

how long after sex is it appropriate to open claude code?

@Cliffracer56k @70sscifi The lights came on eventually, what’s weirder is books 2 and 3

@70sscifi Isn't Rama dark inside or am I misremembering?

“Rama,” by Morris Scott Dollens, 1980

digital nomad as a lifestyle brand peaked when it became indistinguishable from "guy who works from bali on his parents money." the aesthetic outlived the economics by about three years and nobody wants to admit it

dudes who wrote a to-do app in cursor and now describe themselves as "leveraging AI to reimagine software delivery." bro you asked @claudeai to add a checkbox

the fact that "algorithm" got redefined by twitter culture from "sequence of steps" to "invisible feed god" is one of the more successful semantic captures of the decade. now normies use "the algorithm" like it's a weather system

L take. Pure research has its place. But a lot of researchers spend so long inside abstraction that they start assuming intelligence automatically translates into market intuition, product sense, or execution capability. It doesn't. Production reality is undefeated. Users, scale, incentives, failure modes, latency, economics, distribution. Those things expose weaknesses papers never will. The inverse is also true though. Builders who dismiss all research are usually standing on foundations they didn't build themselves. The real gap is humility. Researchers should respect operational complexity. Builders should respect foundational work.

i look at pure research with a bit of contempt. it's for people who don't get their hands dirty and don't want to take the risks of seeing their idea perform in production. you just spent years in an insulated ivory tower, while others have a live feedback loop collecting data in the trenches.

@bunjil 🔥🔥🔥

only AP i’m into is WAP (wet ass pussy) lol

@bunjil you the mvp of acting emotionally rekt

some people lookmaxx while others sookmaxx 😭

I dare you to post your last saved image without any context. No cheating.

Many cultures have a creepy tooth fairy tradition, designed to acclimate children to the idea that they may eventually have to sell their body parts for money.

@toly Toly: we got stablecoin transfer volume at home

It’s a sign of mental illness to hoard trillions of stablecoin volume

@closermethod @RoundtableSpace Every company needs a Steve

@RoundtableSpace The zero-headcount company still needs one thing AI can’t replace yet: taste. Agents can execute. Someone still has to decide what matters, what ships, and what customers will pay for.

ANTHROPIC JUST RELEASED THE OFFICIAL PLAYBOOK FOR BUILDING A COMPANY WITH CLAUDE CODE. CEO: 1 human. Employees: AI agents. Operations: fully automatic. The zero-headcount company is no longer a joke.

this song is like magic

So my infant’s favorite song is Time of my Life by @whatsupsahl