Suvash Thapaliya @suva.sh · 19/02/2026Sounds neat. Hopefully not a lot of horses, haha. I’ve barely looked into media streaming stuff in Elixir, will def. check ex_nvr. Curious about the inference server, is the yolo model running on its own non elixir engine or using Bumblebee things? 100
Suvash Thapaliya @suva.sh · 17/02/2026Looks neat, curious about the end use case, or is it a capability demo? Inference with Yolo models, or something else? Either ways, good luck and hope it runs all smooth. 100
Suvash Thapaliya @suva.sh · 31/10/2025Ah nvm, now I get that you meant the webshop didn’t update pricing for the custom config.! 😅 010
Suvash Thapaliya @suva.sh · 31/10/2025Why not put one together yourself, esp. if a PC/Linux build? Much better value. I built one in 2019(probably after 18 years), took a bit of research, but with sites like pcpartpicker, it’s really a breeze. Recently learnt about Minisforum etc., and I’d def go the mini route if I didn’t need GPUs. 100
Suvash Thapaliya @suva.sh · 29/10/2025A bit embarrassing to admit, but another thing I've been rather late at using/understanding is WebAuthn(FIDO2) compared to U2F(FIDO). Not having used a FIDO2 compatible hardware key, I had sort of mentally bucketed them together. But, FIDO2 is a pretty solid improvement/extension over the FIDO. TIL. 010
Suvash Thapaliya @suva.sh · 28/10/2025Is it only available via the Oreilly subscription? attempted to buy it, but not functional on my end. 100
Suvash Thapaliya @suva.sh · 28/10/2025Brilliant! I'm definitely going to try it out with the key. Would you suggest backing up the private key for future/new yubikey transfers (similar to gpg certify key bkp to extend subkey expiry), or just generate one on key and "never expose it"? Guess it depends, but curious about your workflow? 100
Suvash Thapaliya @suva.sh · 28/10/2025I've also been considering using Age for encryption to try out modern tooling, (though I'm comfortable with GPG+Yubikey) and only recently learnt that there might be a pathway to use Age keys on the PIV slots. @filippo.abyssdomain.expert Curious if this is the best way to use Age+Yubikeys? 100
Suvash Thapaliya @suva.sh · 28/10/2025Also, @ubiquiti.bsky.social hardware+software is really well done. A lot of things that I used to have sidecars & other solutions for is now just covered with UCG & U6+, and the Unifi software. What a treat really, and I'm pretty sure I haven't even started using all of it. 000
Suvash Thapaliya @suva.sh · 28/10/2025Super late to the "Gigabit at home" party, but recently updated to it, also moved from Synology to Ubiquiti Router+AP setup, Ethernet where possible. Finally I can max out on downloading these bulky models from HF & Ollama store. 😅 130
Suvash Thapaliya @suva.sh · 24/10/2025I've been using the same #GPG keys (master, sign., enc. & auth.) on a @yubico.com Yubikey 4 since 2017, (following the drduh guide) extending the expiry every X years. I'm now considering creating new keys (esp. for RSA/4096 & ed25519) for a new #Yubikey 5. How is everybody else going about this? 110
Suvash Thapaliya @suva.sh · 23/10/2025Right, that’s a fairly recent book as well. Thanks for reminding! 🙌🏽 010
Suvash Thapaliya @suva.sh · 23/10/2025Right after the Huawei book, I finally picked a copy of “Chip War”. I think this book honestly is the most readable compressed history of the chip industry all the way from vacuum tubes to modern day custom accelerator chips incl. the geopolitics. Not to drop the momentum, what should I read next? 110
Suvash Thapaliya @suva.sh · 17/10/2025I still do like using Yubikeys for that (GPG+SSH) since it’s a portable secure “key”, but pretty neat idea to use the Secure Enclave as well. Any hiccups/gotchas with it? 100
Suvash Thapaliya @suva.sh · 16/10/2025Brilliant, I've been on a search for something along the lines after finishing it. I'll queue this up in my local library if available. Thanks back for the tip. 🙌 000
Suvash Thapaliya @suva.sh · 15/10/2025TCP flow & congestion control was literally designed with this in mind. I think you might enjoy a Claude session on TCP flow control and congestion control in regards to streaming architectures. 🙌🏽 000
Suvash Thapaliya @suva.sh · 15/10/2025Nope it’s not. Unless it’s a very sophisticated inner loop. All the tokens should be produced at the accelerator’s own max capacity. The tokens are just waiting in-between various queues on the TCP(“network”) layer as packets waiting to be sent (or eventually dropped). 100
Suvash Thapaliya @suva.sh · 15/10/2025That’s somewhat expected TCP flow control (TCP backpressure if you will) on streaming systems. The tokens pile up in queue somewhere between the server and the browser, and if the TCP congestion clears up all the tokens will seem to arrive in “one big chunk”, like that flaky audio call. 😅 100
Suvash Thapaliya @suva.sh · 15/10/2025Just finished reading “House of Huawei”. What a solid read, would absolutely recommend to anyone trying to understand and forward connect the dots to where we are with technology wars right now. 220
Suvash Thapaliya @suva.sh · 08/05/2025Omg! I remember buying a self assembly kit for a friend’s birthday a long time ago. The packaging didn’t really specify what it was. Fun little useless surprise at the end of putting it together. 010
Suvash Thapaliya @suva.sh · 29/04/2025Agree but perhaps not fully(it depends)! Diverse tools def. increase the problem complexity space, but as an LLM provider you probably want to solve that to satisfy more downstream customers. Of course, I agree that most LLM consumers(businesses etc.) individually don't need a diverse set of tools. 000
Suvash Thapaliya @suva.sh · 23/04/2025Ah wow! Is the Lisp codebase still strong, or mostly converted to Python by now? 020
Suvash Thapaliya @suva.sh · 21/04/2025Pretty interesting idea. Immediate thought: one might need a diverse set of tools in the “training run”, so as not to overfit to the same set. Possibly more interesting if you can run the inner loop during training, and “transfer” that learning to run outer loops during inference. 100
Suvash Thapaliya @suva.sh · 21/04/2025Solid 10/10 post by Thorsten, on letting the results of your LLM(http) calls decide what function to execute next, in a loop, in a loop, .... 020
Suvash Thapaliya @suva.sh · 21/04/2025Finally got around to reading this. Love the simplicity, I've been using a similar barebones(low dependencies) approach in current product(in Elixir). No SDK, no problems either. It's all http calls, with some workflow branches and some letting the http response decide what function to dispatch on. 020
Suvash Thapaliya @suva.sh · 13/04/2025Rewatched the Game Theory video by @veritasium.bsky.social again today considering the turn of recent geopolitical events. A good reminder that being nice, forgiving, clear while still retaliatory/provocable is a pretty good strategy in the majority of cases. www.youtube.com/watch?v=mScp...youtube.comWhat Game Theory Reveals About Life, The Universe, and EverythingYouTube video by Veritasium 020
Suvash Thapaliya @suva.sh · 07/04/2025Been meaning to use the latest Gemini(2.x series) models for a while now. Perfect timing by @strickvl.bsky.social with this set of practical examples in this standalone site. The only thing that could have made this even better for me would be direct curl examples, but maybe I'm asking too much. 😅 111
Suvash Thapaliya @suva.sh · 24/03/2025Thanks for taking the time to write it up. I really enjoyed reading through & also shared it with my colleagues. 🙌 010
Suvash Thapaliya @suva.sh · 24/03/2025No no, Thanks for writing it up. It was really a fun read. 🙌 010
Suvash Thapaliya @suva.sh · 24/03/2025The third post that I thoroughly enjoyed reading is "A Visual Guide to Reasoning LLMs" by @maartengr.bsky.social. I bet a lot of time was spent in creating these beautiful & well explanatory visuals. bsky.app/profile/maar... 000
Suvash Thapaliya @suva.sh · 24/03/2025The second post is by @abhi9u.bsky.social on some common CPU architecture concepts (instruction pipelining, memory caching &speculative execution). Abhi has done a great job with the analogies in this one. blog.codingconfessions.com/p/hardware-a...blog.codingconfessions.comHardware-Aware Coding: CPU Architecture Concepts Every Developer Should KnowWrite faster code by understanding how it flows through your CPU 221
Suvash Thapaliya @suva.sh · 24/03/2025The first is a post about IO devices on the @planetscale.com blog. @benjdd.com clearly put in a ton of work on making this a fun resource. bsky.app/profile/plan... 220
Suvash Thapaliya @suva.sh · 24/03/2025Last week I read a bunch of articles, all of which were a fun read. Not so heavy on social/sharing etc. these days, but figured I'd share them here regardless, because these were a lot of fun to read slowly through. 110
Suvash Thapaliya @suva.sh · 13/03/2025@soldaini.net Just realised you're here as well. Curious if Gemma3 base models could be a possibilty for the olmOCR pipeline, or if the Gemma license etc. is a dealbreaker? 110
Suvash Thapaliya @suva.sh · 13/03/2025Gemma3 family of models seem quite good for straight-up text extraction from images(those horrible pdfs). Along those lines, I remember olmOCR from @ai2.bsky.social released just some weeks ago, based on Qwen VL models. Curious if somebody is working on olmOCR pipeline with Gemma3 models. 100
Suvash Thapaliya @suva.sh · 15/02/2025On this Saturday, I'm happily watching Veritasium episode on AlphaFold. I still think this is probably the highest impact model (in terms of opening up new future pathways for humanity) that has been made public in the last few years. Nothing else comes even close! www.youtube.com/watch?v=P_fH...youtube.comThe Most Useful Thing AI Has DoneYouTube video by Veritasium 021
Suvash Thapaliya @suva.sh · 29/01/2025This is one of the first few posts I've seen that uses Deepseek model to generate high quality datasets, which then can be used to train the ModernBERT models. Really neat stuff! Once can easily replace the slower, expensive 3rd party LLM router with a fast, cheap & local model. 063
Suvash Thapaliya @suva.sh · 29/01/2025Thanks for the post. Neat to see deepseek distilled models already being used for data generation. You mentioned "There are ways to allow for both structured generation and reasoning. I’ll post more on that in the future!" Is the general idea to do structured generation using xml tags instead? 110
Reposted by Suvash ThapaliyaLewis Tunstall @lewtun.bsky.social · 25/01/2025We are reproducing the full DeepSeek R1 data and training pipeline so everybody can use their recipe. Instead of doing it in secret we can do it together in the open! Follow along: github.com/huggingface/...github.comGitHub - huggingface/open-r1: Fully open reproduction of DeepSeek-R1Fully open reproduction of DeepSeek-R1. Contribute to huggingface/open-r1 development by creating an account on GitHub. 620036
Suvash Thapaliya @suva.sh · 23/01/2025One of the most fun things about using Open WebUI is how you can intersperse various models in the same conversation, depending on the next task you intend to work on. Start off with `deepseek-r1`, and the follow up the conversation with something like `phi4` or `qwen2.5-coder`. 000
Suvash Thapaliya @suva.sh · 23/01/2025Maybe a bit late to the club, but it’s pretty interesting to see the inner monologue when interacting with deepseek-r1. 000
