OpenBSD's KARL (Kernel Address Randomized Link) is one of the rare powermoves I see in computer security - along with OpenBSD's `unveil` and `pledge` feature. It singlehandly make an entire category of exploites (blind ROPs) much, much harder without much code and execution model change. However, u...
This is another talk I am giving at COSCUP this year. :sigh:.. I am late making the drafts again, let alone the slides. Rush job, but what I must do is what I must. Esperanto was a company that originally tried to make a RISC-V based GPU. It later pivoted to a RISC-V AI accelerator, primarily aim...
Anthropic's J-lens was one of the rare tech in mech interp that somehow is scalable (compute once and reuse forever) and not labor intensive (doesn't try to manually understand one action at a time through raw sciencing). Computing J-lens was _really_ expensive. So much so that creating one lens fo...
I feel lucky that I get to present my ~~BS~~ encouragement at COSCUP to prevent AI apocalypse, based on my work on RWKV. Hopefully landing in 40C3 too later this year. I dabbled into mechanistic interpretability as I am bored one day while also under existential dread that AI is gonna kill us all ...
For the web folks, use the following link via mozz.us's web proxy This post is going to be very rambly. TOFU has been a major point of discussion since I started my journey in the Geminispace. At that time I was also in support of TOFU to remove PKI as a Geminispace dependency. Back in early 20...
I was doing some infra maintaince work today and turned on debug mode too look into why some of my own searches are timing out. Turns out there's someone who is hammering the server, which I can only assume is because someone's rogue crawler or some vide coded web crawler hitting public gemini prox...
With Anthropic releasing their J-space paper, I decided to see if I can replicate it on RWKV, the RNN inspired LLM that actually works. One problem, I can't seem to collect the required calabration data using q Q4 model, the activation change is way too big on small perturbations.. BF16 does work ...
A fried at my local hackerspace recommended I try Immich for managing my photos. I had for the longest of time been complaining about how slow Nextcloud is when it comes to photo management. The map took literal 20 seconds to load because of a N+1 query issue on very capable hardware, the phone app...
Mi forlasis Tenstorrent iam frue en januaro 2026 kaj translokiĝis al AINekko, kiu interese aĉetis la intelektan proprietaĵon de Esperanto AI kaj malfermfontigis la stakon (kaj la RTL!). Estas amuze labori kun ĉi tiu ĉipo. Ne plu stranga VLIW maskovestita kiel RISC-V (vidu mian antaŭan afiŝon pri...
I left Tenstorrent some time early January 2026 and moved myself to AINekko, who interestingly bought Esperanto AI's IP and made the stack (and RTL!) open source. It's a fun chip to work with. No more weird VLIW masquerading as RISC-V (see my previous post about Tenstorrent's real programming model...
See my previous post about dmake about the initial results. This post is a release announcment, postmortem and an overview. As well as the slides for the first public talk I gave at CyCraft C++ club (HTTP only unfortunately, sorry the Gemini folks the slides this time is not suitable for print-t...
I don't know why. But my YouTube, Twitter (I still refuse to call the thing X) and Facebook account has been simultaneously infested with AI slop. Absolutelly low effort, low quality, hype or fear mongoring (I think the AI risk is real, both economy and human existence. But that needs real adult co...
AI has been spreading _very_ fast lately. And I feel conflated.. one hand it's literarly my job writing code that runs said AI and on ther other hand with pratical use, I feel that I learn nothing while using LLMs for tasks. Be it code or general understanding. Be it with Opus, GPT, Qwen or Deepsee...
I am so thankful my current employer is just totally open about most of the things. I can share almost everything I find. Ironically this is the best devrel possible. Show, not only tell. I am partially done with basic kernel work and starting to look at the processor as a proper systolic array and...
In the previous post I managed to steer RWKVv7's behavior to some degree and showed that the recurrent state after layer 0 cannot be effectively used as a method to detect dangerous prompts by any linear, non-linear or ML methods. I was thinking to myself - seriously? No right? That informatino got...
So I took a very long pause from my other project `kiln` (CMake compatiable alternative) to work on `tarnet`. Tarnet in short is my re-imagination of GNUnet which I spent quite a lot of time in the past diving into. GNUnet VPN (Virtual **Public** Network) was a function I used extensivly to securil...
Happy Easter! I am _really_ concerned about AI alignment and how that would make the world.. well... crash and burn and humans all die (thanks Robert Miles, my beloved Ai Safely YouTuber). That kept me up at night and the dread is is tearing me apart. I decided to do _something_. But transformers a...
Time to think more about art. AI has impacted the art industry the most. Artists see their income dwindle. Customers with low budget got sucked away by AI with much lower cost of production. And high budget customers are figuring out they sometimes can get away with lower quality results. Not goo...
I haven't written about Gemini in a long time. The late AI wave has gotten me really concerned about the future of humanity. Humans are the apex species on Earth because we are the most intelligent species. But AI seems to have surpassed us. I think we are truly in uncharted land at this point. I a...
I am currently working on the GGML backend for the ET-SoC-1. To answer the question of "how well does this chip actually perform?", I decided to benchmark the most basic FP32 matrix multiplication implementation using the matrix engine. The plan is the same as in my previous RK3588 matmul benchmark...
It's 2AM. My recent interation with AI coding tools (to be specific Claude Code with Opus 4.6) genuinely makes me think most of software engineering will be dead. And I am down an entire can of beer. I am the last one to join the AI hype. But the capablities of Claude Code is truly scary. No longe...
It's the time of the year again. FOSDEM came around again and I had something to talk about. This time thanks to my employer (AiNekko), a large part of the company was flown to FOSDEM (we also run the AI plumber's devroom) and my boss being supportive that he even pushed me to give my talk about my...
Weird I started a new project - a (going to be, mostly) CMake compatible **build system**. Source not released yet as it is early stage. But it seems promising so I'll write about it. Despite the many efforts to improve the C/C++ build systems, notably Meson and Ninja did a great job, CMake remains...
I'm ~~27~~ 28 (as of releasing this post) now. I am approaching my 30th birthday and I want to reflect on myself so far. Especially as recently things happened, broke up with my girlfriend and I want to reevaluate my priorities. I started coding and living with computers since a young age. Over ti...
Weird timing. I left Tenstorrent at this point, but I still have a talk at FOSDEM about Tenstorrent and the processor. Because that's fun. Luckily my new employer still allows me to talk about it - so in my tradition, here's the draft! The talk is "All in RISC-V, RISC-V All in AI: Solving Real AI ...
This is a strange post for a blog focused on computers and hacking, but I work in startups, and they are known to fail unexpectedly. Generic advice is not helpful, and paid advisors often have no idea about startup funding cycles or the problems that areise when your income is in a different curren...
This blog relies heavily on Drogon’s CSP (C++ Server Page) system - think of it as PHP’s templating engine, but written in C++. It’s not the most modern choice in 2026, especially since CSP predates C++ coroutines. But it’s good enough for the static-ish pages I serve, and more importantly, it lets...
Just a random post because I had much fun at 39C3. As per CCC's strict photo policy. Photos will not contain people unless they all concent to be photo'd. At least to the degree that they are not recognizable. I am not exactly sure what to expect when I handed in the Hamburg airport. This is the f...
I carry a travel router when I am on trips. Part of my workflow needs high LAN bandwidth between my devices and hotel WiFi just doesn't cut it. I bought one from TP-Link for cheap expecting things to just work. It does.. sometimes. But sometimes the connection just doesn't work even imy laptop happ...
I haven't blogged for quite a while.. Well, life is happning and I just don't have the time or energy to sit down and write posts. Until now. Lately I had a work strip to China, where it's known for the hostile internet the the GFW. So.. I decided to upgrade the long-untouched OpenBSD on my laptop...
My llama.cpp fork for Tensotorrent hardware is going fairly well. But this weeked I want to rest a bit and do someting else. I really want some tiem to go and play WatchDogs2. But then I have a new idea. I programmed Rockchip NPUs to run MLPs from sckikit-learn. It should be just as easy to get the...
Today I suddenly see a lot of discussions about the NEO Home Robot from 1X on the internet. And followed by a whole lot of discussion about humans being obsolete. Bla bla bla... I am not engaging with that discussion here. I have my opinions that may or may not agree with that perspective. BUT. I...
Rotary Position Embedding, or RoPE is a critical operations used in Large Language Models. Responsible of encoding positional information (where the token is in the sentence) into the embedding, so the model can distinguish between "the cat eats the bird" and "bird eats the cat". Ever since my jour...
Oh man.. not this kind of post again. Boomer. That's what I assume a lot of people's initial reaction would be. I just tried Google Jules, Google's AI coding agent, just because I happen to own a Pixel 9 and that comes with a 1 year free subscription to Google One. I figure I should give new thing...
I can't believe I haven't written on my blog for a month. Usually I end up with several posts by the end of the month. Lately my job has involved frequent flying around, all over the place. Malaysia last week, Canada this, Europe later. Flying is fun if for the first maybe 3 times. But after some p...
I'm not new to Linux Phones. I was a backer of the original Librem 5 camping, own a Librem 5 devkit, later on a PinePhone. During FOSDEM I walked around and found the stand by FuriLabs where they demonstrated their FLX1 phone and chatted with all the participants. I knew about the phone because one...
I feel like this is a constant state of my being. On weekends, alcohol and coffee during late night. It's not particularly healthy, but it bumps my productivity by at least 300%. Epically after I transitioned from a startup to working in big tech. Big tech job is boring AF, but it pays well that I ...
In my last post, I upgraded my home server/NAS to be an Intel machine. This post I want to write down the process of getting hardware video encoding working - before I forgot. One of the applications I run on my system is Jellyfin. The open source media server. It's also IMO better then Plex, the ...
The server that I run right now to host all of my services, including this blog is around 6 years old. Using an experimental ARM development board. I bought it because at the time 10G networking is expensive and that board itself was basically a proper 10G NIC. Well, non standard hardware has non s...
Is AI generated art, art? This has been a question asked by many since MidJourney dropped the revolution product: Stable Diffusion. Now you can easily generate images by asking a chatbot like Gemini or ChatGPT to do so. Using interface via the many consumer minded products. Or via a quick API call ...
This post is long overdue. FOSDEM 2025 is my first time at that event. Wow it's awesome. I've gone to local (to me) conferences before, but FOSDEM is truly on another level. I you ever saw this cat hoodie at FOSDEM, it's likely me. My flight is a long one, from Hokkaido to Taipei, Dubai then fin...
After my previous post about Tenstorrent programming, I am curious about how exactly am I supposed to write my own LLK (programming the vector engine/SFPU). I tried it once back when I was still using Grayskull. Wanted to give it another try and see where I land. This time I set out to write a simp...
The Tenstorrent architecture is a different kind of AI processor. Unlike GPUs where you get a massive core count and parallelize across thousands of cores (not really, SIMT is a lie that the compiler told you). Tenstorrent chips are a grid of different nodes. Most are compute elements called a T...
I was handed a flyer about Nostr, a new pubsub protocol at FOSDEM. To put it simply, Nostr is the minimal protocol to for a decentralized content publishing system which a lot of the same appeals as most small internet protocols. This concept intrigued me as someone dabbled into GNUnet and Gemini, ...
When I first started programming Metalium. Memory was a "you do this and that then it works" kind of thing. Which doesn't sit well with me - at that point I understand each Tensix core only has direct access to it's own SRAM. But never figured out how all the magic parameters in DRAM allocation wo...
Perplexity (perplexity.ai) loosely is a ChatGPT competitor that focuses on incorporating online search into its responses. This supposedly helps with reducing hallucinations and allows the LLM to have up to date information. Very recently they lunched "Deep Research" which instead of basically quer...
I just upgraded my home server today and randomly found I couldn't SSH into it. Also the BMC console though allows me to type into it, it accepts my password but doesn't show me the prompt. I was fearing that my system was compromised somehow. Through pure guessing, I found that the issue was somew...
I got my hands on a Mikrotik switch and is happy with it. One problem I haven't solved for a while has been getting WinBox to connect to it remotely (it's at by dad's place). I have been ssh-ing into a server thebn `mactelnet` into the switch. WinBox however can't work with that even with VPN (VPN ...
I really ain't spending much, barely more then what I spend during college days. An I'm 27 with a career. Living on cheap breakfast shops downstares (which is common in Taiwan) and cooking at home. I do eat out from time to time, but still at relativly cheap places. But lately, more like 2 years ag...
GGML has emerged as one of the leading framework for efficient inference, particularly for large language modes. GGML is engineered with performance in mind and with great quantization support. More importantly. Its flexibility and growing community support make it an ideal platform for exploring...