Jarkko Sakkinen
jarkko
HotFlow 0.1.7 aka "trying cyclopts"
This seems to do the job for the command-line but I make it less fancy in the end. Just wanted try out features.
Anyway, this version starts to feel like a useful tool :-)
This seems to do the job for the command-line but I make it less fancy in the end. Just wanted try out features.
Anyway, this version starts to feel like a useful tool :-)
0
0
0
Jarkko Sakkinen
jarkko
Edited 2 days ago
HotFlow is looking for its direction in terms of interface etc. because I haven't see anything doing exactly what it does but I'm getting useful and valuable profiling work done with it :-)
0
0
0
Jarkko Sakkinen
jarkko
Impulse buy: Dell UltraSharp U2410 from 2010.
You would not believe but this is one of the most desired displays for any sort of retro computing.
Its speciality is composite inputs and direct support for Amiga's 15 kHz video signal.
You would not believe but this is one of the most desired displays for any sort of retro computing.
Its speciality is composite inputs and direct support for Amiga's 15 kHz video signal.
1
0
1
Jarkko Sakkinen
jarkko
Edited 3 days ago
HotFlow - a probablistic CFG profiler that I've written to help optimizing CPU inference for my tools Goosedump and ReadSeek.
Inference is quite nasty and tricky to profile by measurement given maximization in scaling. E.g., when you optimize dedicated hardware for inference it's a minmax optimization to thenear proximity as possible of OOM as one can reach.
Since I'm not really seasoned on doing any of that instead of running anything I just count 🤷♂️ That's the gist of HotFlow. It's based on Angr and Capstone :-)
pipx install hotflow
Inference is quite nasty and tricky to profile by measurement given maximization in scaling. E.g., when you optimize dedicated hardware for inference it's a minmax optimization to thenear proximity as possible of OOM as one can reach.
Since I'm not really seasoned on doing any of that instead of running anything I just count 🤷♂️ That's the gist of HotFlow. It's based on Angr and Capstone :-)
pipx install hotflow
0
0
2
Jarkko Sakkinen
jarkko
Google on AGI: HR updates.
I skipped straight to the science part and this is what I get.
I guess this is Google's version of "IT ESCAPED FROM THE LAB 🤯 🤖 👽 I'd guess one could translate "AGI" as "our next generation models have already missed time-to-market".
I skipped straight to the science part and this is what I get.
I guess this is Google's version of "IT ESCAPED FROM THE LAB 🤯 🤖 👽 I'd guess one could translate "AGI" as "our next generation models have already missed time-to-market".
1
0
0
Jarkko Sakkinen
jarkko
Great Windows 11 Pro VM with SSH access.
Soon I can dogeat pi-landstrip and Windows AppContainer sandboxes.
Soon I can dogeat pi-landstrip and Windows AppContainer sandboxes.
0
0
0
Jarkko Sakkinen
jarkko
Edited 29 days ago
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 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
0
0
0
Jarkko Sakkinen
jarkko
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 :-)
0
0
1
Jarkko Sakkinen
jarkko
Edited 1 month ago
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
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
0
0
1
Jarkko Sakkinen
jarkko
Edited 1 month ago
This did not actually exist in dead simple form, so I created a tool for Landstrip testing use: https://crates.io/crates/mkwim
0
0
0
Jarkko Sakkinen
jarkko
I only vibecode or work on inference with my official slop shirt on :-)
0
0
1
Jarkko Sakkinen
jarkko
The screen cast below demonstrates ReadSeek 0.8.0 with hand-optimized inference engine for Qwen3-VL-2B. Given how shitty the CPU is (AMD Ryzen 5 PRO 4650) in this laptop, the performance is not all that bad, and it will of course only improve over time :-)
ReadSeek will retain itself as being based only on CPU inference as a core design decision; ReadSeek *is* smart but *does not* get in the way. That's why hand-optimized inference is even more so important aspect of this project.
The reasoning roots allowing to retain CUDA, Metal, unified memory and VRAM under user's full control. You can run your local models and ReadSeek will always do its job as an ubiquitous partner :-)
#pi #opencode #coding #agent
ReadSeek will retain itself as being based only on CPU inference as a core design decision; ReadSeek *is* smart but *does not* get in the way. That's why hand-optimized inference is even more so important aspect of this project.
The reasoning roots allowing to retain CUDA, Metal, unified memory and VRAM under user's full control. You can run your local models and ReadSeek will always do its job as an ubiquitous partner :-)
#pi #opencode #coding #agent
1
0
0
Jarkko Sakkinen
jarkko
Edited 1 month ago
As said in the README.md, each active task runs as a full pi --mode rpc process inside an outer Landstrip sandbox.
https://pi.dev/packages/pi-landstrip
I've as closely as possible reverse-engineered (via observation) how the process model in OpenCode works, and subagents are as close as you can get with Pi modelled to act and behave the same as they do in OpenCode. This includes even the shared configuration format, primary agents (build and plan) and pre-defined agent roles (e.g., explore in case you are familiar with OpenCode).
https://pi.dev/packages/pi-landstrip
I've as closely as possible reverse-engineered (via observation) how the process model in OpenCode works, and subagents are as close as you can get with Pi modelled to act and behave the same as they do in OpenCode. This includes even the shared configuration format, primary agents (build and plan) and pre-defined agent roles (e.g., explore in case you are familiar with OpenCode).
1
0
0