perf focus is awesome!

cargo install perf-focus

Finally, weekly Ed Zitron at Tech Report: https://www.youtube.com/watch?v=z4T_AbskFS0

Dude is such a great roaster :-)

Intel is not really up to the date given scarce availability of AVX-512.

#intel

Video detection in ReadSeek will be based on a storyboard.

Qwen3-VL-2B processes only images; it does not understand videos. In other words, a video must translated into a meaningful set of images.

A storyboard is a tiled composite image generated from a fixed number of down-scaled video frames. Let’s mark this with N, and further constrain that N>2.

Algorithm:

1. Use two frames of the frame budget for start and end timestamp frames.
2. Place the remaining N-2 timestamps initially even distances from each other.
3. Shift the timestamps within 50% range of their initial distance with some yet to be defined best effort strategy. maximizing the difference between the frames.

Note that timestamps positions form a O(2^N) problem as the timestamps affect each other as a system.

This idea could be taken out of the domain to an imaginary movie studio. What if you wanted back-engineer a storyboard to already shot material? A previous might be lost, or sometimes might not even exist.

Figuring out the answer to this question provides an optimal answer to the problem in my original domai

I would not call landstrip and opencode-landstrip as stable software just yet, but at least I've started to forget that the sandbox is active, unless it asks for a new permission.

I hit a bug in every two or three days, and the code base becoming less of a chaos i.e., all the convergence metrics are pointing to the right direction :-) Windows was quite rotten up until recent releases when I got some great feedback from an actual dog food eating Windows user.

While the most of the sandbox implementations taken the obvious container approach, I modeled a coding agent as follows:

1. A coding agent are bit like web browsers.
2. Thus, the process model for a sandbox should be modeled from a browser tab.

Based on these premises I made a Linux-container-free implementation, which protects the sensitive data with a combination of Landlock and SECOMP filters in Linux, Seatbelt (SBLT) in macOS and Win32 AppContainer framework in Windows, and adheres to Anthropic Sandbox Runtime JSON policy.

The chose approach has also many robustness benefits. E.g., nested containers are not required for CI runs.

My 3rd inference engine will land to Landstrip in foreseeable future. It will be be part of the sentinel layer for I/O, which protects the secrets even when prompt injection attacks are in effect.

#landstrip #opencode #agent #sandbox #anthropic #infosec

I've been really happy with Vivaldi. I still wish there was also a paid version with less partner junk but other than that it's pretty cool.

#vivaldi

Microkorg was based on same DSP family as EA-1 and MS-2000 but from Microkorg's VST version you really cannot the same aliased crunch as from EA-1 :-) It's the combination of DSP but also DACs have their own flavor for the sound. EA-1 is the undisputed king of digital acid :-)

In remembrance of Dan Williams

https://lwn.net/Articles/1085021/ #LWN #Linux #kernel

@lwn
This incredible loss of an amazing life is our reminder to do more to look out for each other. Depression sucks. Depression lies. And this society (and that means WE, ALL OF US) still puts a major stigma on acknowledging depression.

I take meds every day to deal with depression. And as Catherine says in her facebook post, depression "can not be treated with meds alone".

We need to acknowledge and talk about this being another illness. One that is mostly treatable. And best addressed by talking about it, being honest about it, and supporting the ones who suffer from it.

If you feel down, find someone to talk to. I don't care if we are friends or not, if I know you or not. If you need someone to talk to, reach out. @dnd.99 on Signal. Talk to me. Talk to someone. Don't give up.

@Jumpmed @jani @grumpygamer

@eduzsh I'm going to keep in phase of doing at min two inferences engines annually from scratch but still polish each to as high quality I can. What I want to understand is this:

1. Let's assume we have a SoC.
2. Let's imagine it is capable of doing inference and has special features.
3. The micro-architecture design can make any feasible sacrifices on anything related training post-training (does not have to but I don't give it any weight).

What would be best architecture provide lift up for let's say up 500B parameter models. It's also definitely an area were Nvidia dose not have any tech leadership. Blackwell hardware design is sloppy and dysoptimal if thinking from this "you had one job" angle.

Next model I'm still going to do on Ryze 5 Pro (common laptop CPU from decade ago, Zen 2 architecture) I need to make it scale to GPT-OSS-120B. That's my end goal for this CPU. I'm planning to reach it with 2bit quantization. I have full MoE implementation for 20B version. The magical "model streaming" part was weird. This was discussed either in the context Dwarf Star 4 or Colibii. I mean one always mmaps huge files instead of copying anything and page fault handler brings up the "experts". Still don't get what model streaming is but I'd guess it is just a silly term for the most common activity (never checked this from their implementation).

@eduzsh I think lot of problems around AI should be ripped away from AI researchers, we should put them deep into the cellar and let them work only on training at most ;-)

E.g., inference as an algorithm optimization exercise is more like comparable to a driver design than anything to do with machine learning. For me writing couple of inference engines from scratch has been mostly fun and I'm quickly becoming good at it despite I know almost nothing about training and machine learning. However, I think about e.g. CPU cache hierarchy almost every day (very first time was at high school while writing texture mappers for Pentium with its dual-integer pipeline plus additional simultaneous FPU 1/w op at best).

AI researcher can be considred like hardware designer or compiler writer they just have now overemphasized weight. Often e.g., a C++ compiler and successful programs written in C++ come from different carbon based entities :-)

There's been some games that have provided a bit of relief for the lack of XCOM 3, but MENACE is the first that is actually good.

Chaos Gate was like OK game mechanics is good but I absolutely don't feel anything towards the franchise. When I was kid I enjoyed Franz Kafka; and hated Lord of the Rings. It is such a boring book...

And in Chaos Gat everything rots and is disgusting. There's nothing I enjoy in this story. Almost as bad as Star Wars (BTW, I'm a trekkie).

Phoenix Point look like a game that was never properly finished; i.e., money run out. I've played it hundreds of hours but never to the end because at some point there's too much going on. If the game had been a finished I think it could have been great so I'm waiting for the rumored sequel :-)

@jani @grumpygamer I got pwnd: voted yes

now i'm waiting for the consequences with anxiety :-)

Misusing Github as a Windows VM :-)

https://github.com/landstrip/landstrip/actions

What a sudoku.

I think I try tiling with inference because that improves cache coherency. E.g. preprocess bitmap to 2x2, 4x4 or perhaps even 8x8 tiles (i have to try them all out).

Image becomes brutal for human viewer but as per *hypothesis* cache misses reduce by some percentage. I don't know what the vision model thinks bitmap having this organization but my plain guess is that it might even work.

I like to pay with money, or like it more than freeware. Only without money thing that I fully cope with is open source.

When I pay a bill I know exactly what and how much I loose :-)

Other apps that I pay for annually are CrossOver Office and Bitwig Studio.

@Vivaldi

1. Brave has Brave Origin.
2. In Vivaldi it would nice if the functionality and profile were (optionally) bound. I.e. simple profile would actually reduce the amount of stuff. This like engineered version, if Origin is a demo.

The latter could be e.g., "paid version". Since it is a complex feature there could be like modest annual sum of money. I don't trust companies with lifetime subscriptions or one-time payments because it does not make sense to me.

Like 10-20 annually (Origin's one time fee was like 50'ish). I would pay that any day.

#vivaldi #web #browser

In inference I hit to Rust bottlenecks. It is an enormous effort to make all loops error check free in order to not generate boilerplate. It's really hard to optimize against icache when using Rust TBH.

I spot all the time ".context()" (anyhow) or try-statements and similar cruft. When I work around not having to do any of these checks, inference is suddenly like 1-3% faster. This is ridiculous. Most of my time is now spent on Rust performance issues.

What I will do is to create a C library with the shared primitives (like matrices, tensors, fp16 and similar), which will then be used by ReadSeek and Goosedump to implement inference engines for Qwen3-VL-2B and GPT-OSS-20B.

I'll call this library as libcrank :-)

#ffmpeg

TUI library I like for Rust:

https://github.com/ccbrown/iocraft

Much easier to get by than Ratatui. I even re-implemented my custom IANA-aware, geo-correct and real-time rendered world map widget - with smooth scrolling :-)