Pi is such a great and iconic piece of software; and its author is a sane person. That's rare in the AI world. I still think that Pi should have been what LF backs; definitely not Goose.

https://dfarq.homeip.net/amiga-1000-ten-years-ahead-of-its-time/

#amiga

I don't have enforced use of AI from my employer, and my effort is still calculated in human metrics.

I think it is good to attach to such employer.

Significant perf updates for GPT-OSS-20B “home-baked” inference in Goosedump:

engine: low-hanging fruit for GPT-OSS-20B

- Add ARM64 NEON and AVX-512 SIMD support.
- Reuse pre-quantized Q8Activation across matrix projections.
- Parallelize active MoE expert computations with Rayon.
- Parallelize GQA causal attention using Rayon.

Signed-off-by: Jarkko Sakkinen <jarkko.sakkinen@iki.fi>

Should be out soon :-)

This:

~/work/kernel.org/jarkko/linux-tpmdd master ⇡ 18s
❯ git rebase --exec "git commit --amend --no-edit --trailer 'Reviewed-by: Jarkko Sakkinen <jarkko@kernel.org>' --trailer 'Signed-off-by: Jarkko Sakkinen <jarkko@kernel.org>'" origin/master
Executing: git commit --amend --no-edit --trailer 'Reviewed-by: Jarkko Sakkinen <jarkko@kernel.org>' --trailer 'Signed-off-by: Jarkko Sakkinen <jarkko@kernel.org>'
[detached HEAD e6a236dda485] keys: fix out-of-bounds read in keyring_get_key_chunk()
Author: Michael Bommarito <michael.bommarito@gmail.com>
Date: Sun Jul 19 12:15:03 2026 -0400
1 file changed, 5 insertions(+), 4 deletions(-)
Executing: git commit --amend --no-edit --trailer 'Reviewed-by: Jarkko Sakkinen <jarkko@kernel.org>' --trailer 'Signed-off-by: Jarkko Sakkinen <jarkko@kernel.org>'
[detached HEAD eb2538efe0bf] keys: make keyring key-chunk byte order agree with keyring_diff_objects()
Author: Michael Bommarito <michael.bommarito@gmail.com>
Date: Sun Jul 19 12:15:04 2026 -0400
1 file changed, 3 insertions(+), 2 deletions(-)
Executing: git commit --amend --no-edit --trailer 'Reviewed-by: Jarkko Sakkinen <jarkko@kernel.org>' --trailer 'Signed-off-by: Jarkko Sakkinen <jarkko@kernel.org>'
[detached HEAD 218b305fedb4] assoc_array: trim the final shortcut word using the current chunk end
Author: Michael Bommarito <michael.bommarito@gmail.com>
Date: Sun Jul 19 12:15:05 2026 -0400
1 file changed, 2 insertions(+), 1 deletion(-)
Successfully rebased and updated refs/heads/master.

After this era I have zero trust towards AI researchers. They invented and marketed imaginary concept of "AGI". Talking seriously about magic creatures, should decrease trust. It's common sense. Never forget.

Hiring philosophers to make sure dot products and matrix multiplications don't become evil is just sick. That's what it is. I'm pointing out this to you Anthropic.

I don't wish AI to vanish (I don't actually care whether or not) but I do wish that these medieval times will be over soon.

I've followed the war events almost every night for past 1.5 years and all I can say is that I'm glad Finland is not up against the world's most modern military might.

Pretty amazing stuff during past few months.

I've started to get all kinds of offers from weird AI companies thanks to Landstrip and ReadSeek. There was also one that promised to help to lift up my profile or something like that.

This is like the gist of this business. It's all about delegating slop and being a wingman with good stuff 🤷

How I prompt when I prompt?

1. I have a HackMD account. It is disconnected from any AI.
2. I have there a folder called "Prompts".
3. When I get an idea I don't jump it with to a coding agent. Instead, I create a new note to HackMD.
4. There it rests and grows as I randomly learn more about the topic. I don't have rush It could week or month before I input it.
5. Some day it looks like "sound". Then I copy-paste it to Pi or OpenCode and usually use GPT-5.6-terra or -sol. Explore tasks are redirected to Kimi-K2.6 hosted by Neuralwatt.

This also keeps me away from AI psychosis and relaxes LLM from having to do that much thinking. I can take advantage of coding agents but only use them in "mission oriented" manner. HackMD notes usually fill out from web browsing, chatting, F2F discussions etc. :-)

fter optimizing QWen3-VL-2B Q8 pipeline a bit in ReadSeek:

Decoder prefill: 47.3 s -> 14.1 s
Total inference: 133.4 s -> 98.8 s
Prefill throughput: 6.2 tok/s -> 20.8 tok/s

CPU: AMD Ryzen 5 PRO 4650U

Update will be released soon :-)

What the actual fuck? How do you leave Twitter and then decide, oh, actually, I like the fascism?

Some example of the movies and books that I've enjoyed in the past that come to mind from 2026 :-)

1. Falling down (movie, US, 1993)
2. Battle Royale (movie, Japan, 2000)
3. Lord of the flies (a novel by William Golding, 1954)
4. The Trial (a novel by Franz Kafka, 1914)
5. Dark Star (movie and excellent soundtrack by John Carpenter, movie, US, 1974)
6. ...

IT in 2026:

https://www.imdb.com/title/tt0266308/

If AI is here to stay, what open source community needs to do is to conquer the tools. That's how *we* retain control and govern the ecosystem.

That's one of the reasons why I do so much digging in the local AI space.

When it comes to AI policies, and especially Linux kernel, I just follow the lead and I'm happy that I don't have make the unpopular decisions. [I don't think there actually is a popular position when it comes to AI, so what would you decide if you were Linus? Not a position to envy for TBH.]

My fighting ground is building tools that are convenient and elegant enough that they naturally grow the user base and build up politics-free governance for everyone.

I like Local AI and proactively work on it but my angle is a paradigm shift from deep state ;-) I.e. I'm not one of the guys dropping acid in their basement and talking to computers, each having both real and a "petty name".

Instead, I try to ripple local AI so that it leaks through holes while the more involved payloads go through a cloud provider. This thinking realizes already in ReadSeek and Goosedump; both which embed their custom and optimized inference engine for the *localized* use case ( 💡 ).

This allows to use e.g., text-only frontier models with the lack of vision filled up by the local tools, to have *actually* private agent memory and this list will continue to grow! What I'm 100% sure is that I'm definitely done Just Yet.

https://www.youtube.com/watch?v=j51uMah-3js

#infosec #ai

super interesting interview: https://www.youtube.com/watch?v=3DA7EXbf_d4

I have already plans for my 3rd inference engine. It will add sentinel detection to Landstrip ecosystem. It's actually least demanding of these I'd project at least (GPT being the most complex being "true MoE").

A custom, AVX2-optimized inference engine tailored for GPT-OSS-20B MXFP4: https://github.com/jarkkojs/goosedump/commit/7290b7fe925fb42e43cdc40e1eaa94a1020f3c87

It feels stupid to say this but yeah model is "streamed" i.e., i did ****ing mmap call ;-) Super advanced.

And yes I take a lot of inspiration from the ideas of Dwarf Seek 4 and Colibri (GLM-5.2) so kudos to those projects.

I had a few DSP/DAW experiments early this year so AVX2 was still fresh in my head, and as teen I've written texture mappers, and worked with affine transformation and quaternions, so with that background learning how to use and apply tensors and linear algebra in inference was not an enormous stretch.

Yeah and have also studied math as part of my degree but more like algebra side of things, nothing that would help with this.

I've enjoyed working with inference because I really have to tie ripples of knowledge I have and fill the gaps to get a functional implementation :-) Sort of "gamified" situation if you can describe it that way...

Vivaldi is great and *actually* politics free browser, and yeah, simple view gives me similar experience as Brave Origin (and for free of charge, also in macOS).