Derick Winkworth

I'll be your hanging man. You'll be my queen of swords. That pagan moon shining on the church house door. youtube.com/watch?v=sba0vC…

This is true. In the beginning, when Iluvitar sang the world into existence, there was a single layer2 device at most hanging off of a layer3 port. What I mean is, that was considered best practice. One hub or switch hanging off of a router port. You didn't attach multiple layer2 devices together. Switching was about bandwidth efficiency, not building layer2 networks.

Also, listen... For 1 second or less interval protocols, if your NOS doesn't implement those as table-based (i.e. "array-based") FSMs, timer wheels, and batched packet processing... then they're doing it wrong. And probably on purpose.

I will be turning this into a VPN client device. The chipset has IPSec acceleration.

I rooted a 10 year old ISP ADSL router. 128MB of RAM, 128MB of storage, dual-core 400mhz MIPS processor. And I think I have a way to recover the kernel headers even though they are not installed on the device.

Hi there - Was invited to talk on the CloudGambit/Packet Pushers podcast about the messy reality of AI tooling and expertise. Check it out! Links in the next tweet

Creating a custom kernel for a specific GPU has been a very interesting exercise. I asked claude to help me do it and to track performance issues. I have a bandwidth-starved GPU, but the kernel isn't even using the bandwidth available. To eliminate memory from the equation, we repacked the model so that byte order makes the data immediately consumable upon ingestion into the GPU. Portable formats have to have bytes rearranged in different ways, and that takes GPU cycles. After byte reordering, you need to layout the model as one giant block, padding layers so you get alignment across your memory banks. This gives you maximum memory parallelism. It also prevents some layers being afflicted with a bias that puts most of their weights on a subset (even a single!) bank. With memory out of the way, we found that the kernels themselves aren't coming anywhere near efficient memory bandwidth utilization.

@cloudtoad 100%

Lot of doom and gloom on here today. I don't know guys, I've learned more and executed more on my ideas in the last month than I ever have in my life and none of the current frontier AIs one-shotted any of it.

--dangerously-skip-permissions

Developer workflow, 2026: Spend a lot of time with Claude not even coding. Start breaking the project down into it's pieces. Make a pipeline out of that. This is project pipeline v0.1. Start breaking down what the ins and outs of these pieces are. This will turn into a graph over time, rather than a straight line. Your job, with Claude over time is to really understand the problem space relevant to your goals and to ensure each md and each claude invocation is as tight as possible. And automate the things that are deterministic. Don't have claude do those things. You're not going to one-shot anything worth doing, and neither is Claude. And that's true even of Opus-4.6.

@HankYeomans @steipete 90% of my relatives. They will never call me again for their computer problems. Well.. they will.. but a lot less. This is fine for them.

@steipete The numbers of people for who that is more than adequate is high. Very high.

The new MacBook Neo has 8GB unified memory? Rough.

Join me this coming Friday over on O'Reilly Online (via Pearson) as we look into the way a router forwards packets. learning.oreilly.com/live-events/ho… This training will peer into the internal components of a router, starting with an explanation of how a router switches packets. This walk through of a switching path, in turn, will be used as a foundation for explaining the components of a router, including the various tables used to build forwarding tables and the software components used to build these tables.

"When Claude is reading an ISA dump from stage 5 and sees v_dot4_i32_iu8, it needs the pseudocode and the 'this is your Q8 workhorse' note, not three paragraphs about the history of packed integer arithmetic." Sometimes Claude says some unintentionally hilarious shit. This is Claude talking about itself.

"Compacting conversation..." Nothing has ever become so annoying so fast.

@wmf A trace shows in any GPU heavy task, the GPU is idle about 99% of the time. And even though it's on the same dye as the CPU, *there is no shared cache between them.*

@wmf I read many reviews. I knew there were some driver issues. I guess I for the money and the marketing I wasn't expecting to find out that the GPU is literally useless for AI. So I'm going to eat it and buy a different laptop that doesn't fucking suck.

"The gfx1151 iGPU has up to 40 Compute Units. Even with a wide 256-bit memory bus delivering roughly 256 GB/s of bandwidth, an iGPU of that size is severely bandwidth-starved during heavy compute or gaming workloads." If there is a class-action lawsuit about this, I'm in. These laptops were selling for 10k a pop with maximum config. They leaned hard into the APU, but the system is so poorly designed it's just not possible to leverage the APU in any meaningful way.

@HankYeomans Most of my pipeline (each step a specialized agent) is validation, verification, testing.

People wrongly think you can plan your agent blueprints for every single inevitability. You can not. Plan, plan, plan, but also plan to correct, steer, think, understand.