Entries for July 22, 2026

  1. Portrait of Onur Solmaz
    @onusoz · /2026/07/22 · View on
    Ran Laguna S 2.1 *theoretical* (not measured) upper bound calculations for the DGX spark poolside/Laguna-S-2.1-NVFP4 - max ~25 tok/s single session. if you are willing to go down to 12 tok/s, you can have 4 sessions in parallel vcruz305/Laguna-S-2.1-GGUF (IQ1_S) - max ~74 tok/s single session. theoretical 5x20 tok/s = 100 tok/s aggregate upper bound Calculated assuming no speculative decoding, so might increase well over these numbers See my local frontier app and my previous posts for more info on the methodology
    @sudoingX ·
    this is the drop the local ai crowd should be losing their minds over. poolside just dropped laguna s 2.1: 118b total parameters, only 8b active per token, a full 1m context window, open weights under a real open license, on huggingface today. look at the chart. it lands at 71 on terminal-bench at 118b, sitting above deepseek v4 pro max at a trillion params, above inkling at 1.5 trillion, above nemotron 3 ultra. it's beating models ten times its size and losing only to kimi k3, which is 24 times bigger. that's the efficiency frontier, up and to the left, exactly where you want a model to sit. but here's the part that made me sit up: it runs on a single dgx spark. and this is what nobody's saying loud enough. the dgx spark is the moe king. a dense 118b would crawl on it, the bandwidth chokes reading every weight each token. a moe with 8b active only ever reads 8b, so the spark's 128 gigs holds the whole model while generation stays fast. big brain, light footprint, the exact shape the spark was built to run. open, frontier competitive, moe efficient, and it fits on a box on your desk. that's the whole thesis in one release: you don't need a datacenter, you need the right architecture on the right hardware. go grab the link below, weights are up.
  3. Portrait of Onur Solmaz
    @onusoz · /2026/07/22 · View on
    There is an inherent unfairness in this industry If you are Anthropic, Cursor or another billion dollar company, you can do growth hacks and guerrilla marketing techniques like putting your brand into every commit of a user, and get away with it But if a tool like pi, opencode or even say vim or emacs tried to do that, they would be boo'ed Like, it happened to @herdrdev recently There was a feature where it asked if it could star the repo for you at startup, with exponential backoff if you said no. I thought it was a pretty smart and balanced thing to do Then apparently someone complained it was a dark pattern, and it got removed So WHAT if it is a dark pattern? You know what is the darkest? A public good losing out. If monopolists can do growth hacks, creators of public goods should be allowed to do it as well! In fact, I would say that if you benefit from those public goods, you are morally obliged not to criticize this stance Of course, we cannot expect e.g. @pidotdev or @openclaw to make this a default feature Therefore, I created the packages pi-must-win and openclaw-must-win Install them with: pi install npm:pi-must-win openclaw plugins install npm:openclaw-must-win They will add [email protected] and [email protected] as co-committers to code developed through these projects Overall, I propose the "X-must-win" naming scheme for plugins that independently add guerilla marketing into open source projects This absolves the creators of any "sin" and gives the community a way to support the growth of a project Open source must win! Repos are in the links below.
  6. Portrait of Onur Solmaz
    @onusoz · /2026/07/22 · View on
    If you are moving from codex to pi, here is an extension that mimics codex's exec behavior So that long running execs can "fork" to background, and your session does not get blocked by potential hour-long tasks (like it happened to me last night 😭) iamwrm/pi-unified-exec
  7. Portrait of Onur Solmaz
    @onusoz · /2026/07/22 · View on
    brokerkit lets my claw request github admin merge on my behalf, I receive it on telegram I approve the first request. repo doesn't allow merge commits, so it gets rejected due to github policy then it asks again to squash merge. I approve again, this time it works no agent account or clickops on github needed. pure broker side fine-grained policies you don't need to create policies manually either! just ask your agent to set them up while setting up brokerkit "I want my claw to be able to read all my repos except X Y Z, and I want it to be able to push to main directly on A B C repos" I have a separate control-plane repo outside the control of my claw. My local brokerkit policy gets synced there, policy as code I like this model a lot! Complete control over what my agent can do with my own account. Version controlled, explicit. For free locally, without having to deploy a server to host the broker (btw since I implemented brokerkit, I don't need reviews on this repo, and hence no need for admin merge. but I'm still keeping it to be able to dogfood it for other users)
  8. Portrait of Onur Solmaz
    @onusoz · /2026/07/22 · View on
    pi has /goal too. @micLivs Michaelliv/pi-goal implementation works great!