Entries for June 15, 2026

  1. Portrait of Onur Solmaz
    @onusoz · /2026/06/15· View on
    Current average generation speeds for local DeepSeek-V4-Flash-Q2, highest to lowest: Mac Studio M3 Ultra: 32 tok/s MacBook Pro M5 Max: 30 tok/s Apple ??? M4 Max: 25 tok/s MacBook Pro M3 Max: 24 tok/s Mac Studio M2 Ultra: 22 tok/s NVIDIA DGX Spark / GB10: 13 tok/s It seems macs' higher memory bandwidth is contributing here, though I'm not sure if GB10 performance could be improved (I do hope so, I have one!)
    @antirez·
    If you need AI to do a search for you in the real world, ds4-agent is basically SOTA, because it can access the web sites without any limitations given that it uses your local Chrome browser (no, not in headless mode, that's the trick...), and DeepSeek v4 is great at search.
  3. Portrait of Onur Solmaz
    @onusoz · /2026/06/15· View on
    Btw, TTS has come such a long way, @GoogleDeepMind cooked with gemini-3.1-flash-tts I gave Codex my google credentials and it oneshotted the Gemini TTS implementation When I built this 4 years ago, Azure TTS used to be SOTA. Then @ElevenLabs came in and raised the bar super high. Now Google is going after their lunch with controllable expressiveness at scale. I cheer for both! Here is Manim Voiceover demo from 4 years ago with Gemini TTS (sound on)
    @onusoz·
    I major concern I have these days is, while I author code in languages I cannot manually code, are they any good? Over years, I have worked with a number of languages: C, C++, Fortran, MATLAB, JavaScript But Python was my go-to language since more than 10 years. Well that changed last summer So while I have strong opinions on how Python code, should be, conventions and all, I don't have so strong opinions on other languages. That means I am producing slop by default in Rust, Go and TypeScript To solve that problem, I created Its aim is to be "the only tool and resource your agent needs, to minimize slop" It is inspired by the recent bathrobe rants of @unclebobmartin, a.k.a. the author of clean code It enforces a minimum test coverage, maximum cyclomatic complexity, mutation tests, code style across different languages But I have a major issue: How do I know that Slophammer itself isn't slop? One way is to implement and use it for Python, the language I know better, and judge what kind of changes it enforces So for this weekend experiment, I used Slophammer to refactor, improve coverage and merge new features to one of my old Python projects, Manim Voiceover The result is... mixed. We now have types everywhere, which is great. But the constraints have also made it write garbage code like this one. It works fine, even though it's not elegant. The new feature also works What do you think? Does code still need to be aesthetically pleasing to the human eye? Should it still be human readable? If an agent writes slop in the forest, and there is no-one to read it, is it still slop? If anything, I should use its output in Python to reason about other languages, and add more and more constraints. The more the constraints, the less the slop
  4. Portrait of Onur Solmaz
    @onusoz · /2026/06/15· View on
    OpenClaw is sooooo useful for staying on top of things