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

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 :-)

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.

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 :-)

I replaced all mkfs in puu-installer with external mkfs.* commands as I came to realize that most of the mkfs Rust crates are stale or development run by Claude Code harness.

I guess this could be almost sealed in stone: never mkfs other than external commands ever 🤷

This did not actually exist in dead simple form, so I created a tool for Landstrip testing use: https://crates.io/crates/mkwim

@Codeberg, perhaps edge case for the AI policy:

1. Puu OS is a Buildroot based bootc desktop with well-configured vLMM stack.
2. However, it is created with a tootbrush. I've spent 5 months or over polishing every inch and corner. I.e. it is as human manual labor that it can get :-)
3. The core idea is that if these LLMs continue to persists it is better to create easy-to-deploy appliance for host or cluster.
4. I will maintain this for a long time given all the trouble (like packaging GNOME 49 for Buildroot). In fact, I will use it as my desktop.

So it is sort of LLM esque but I think from angle that makes sense despite the opinion. Less SaaS is better IMHO.

Just wanted to ask this before I publish the OS. And also: bots sending PRs is something I highly disregard so no issues with that.

I packaged and wired mold for my Buildroot based bootc operating system project Puu OS.

I needed it to be bootstrapped for the reproducible build or that is at least cleanest approach when going down into that rabbit hole.

Now waiting next few hours for the build results (success and time taken).

I made the most trivial video detection for ReadSeek just using Qwem3-VL-2B:

1. Chop fixed number of frames from the video. They are not in fixed time distance. Positions are adjusted by weighting them based on difference between the frames.
2. Scale them down to a lower resolution.
3. Make a new bitmap with key frames as tiles.
4. Run inference over the resulting bitmap.

Cheap and dirty but does the job for the scale and scope of the tool :-)

Figuring out my #3 inference engine for Landstrip. It will expand the concept of sandbox for in-traffic data and will be hybrid algorithm/LLM based detector.

Things it will do:

1. Detect and clear passwords with some kind of marker (e.g. '<sentinel>'), secret tokens and stuff like that before they leave to a remote source.
2. Detect malicious payloads to some degree. Won't replace anti-virus or intrusion detection but could save from worst slops and prompt injection attacks from time to time.

This will be probably as per inference easier than GPT-OSS-20B, which has Mixture-of-Experts architecture. Since I nailed that it is time to apply more this newly gained skill :-)

Have to try this for Puu OS build:

❯ git diff
diff --git a/Makefile b/Makefile
index 14c83d6..ecfb113 100644
--- a/Makefile
+++ b/Makefile
@@ -49,6 +49,7 @@ define buildroot
make -C "$(BOARD_DIR)/src/buildroot-$(BUILDROOT_VERSION)" O="$(BOARD_DIR)" \
BR2_REPRODUCIBLE=y BR2_EXTERNAL="$(CURDIR)" \
BR2_CCACHE_DIR="$(CCACHE_DIR)" BR2_DL_DIR="$(DOWNLOAD_DIR)" \
+ BR2_TARGET_LDFLAGS="-fuse-ld=mold" \
$(1)
endef


It's like 6-7h per build (6-7 GiB ISO image).

trying to get back on making tracks, last released https://www.youtube.com/watch?v=a-5jIaSIxpA