matthew.

@NielsRogge @openclaw get quantized 27b running its miles better

Wtf... Qwen3.5-35B-A3B took 3 minutes (!!) to answer my simple question, "What's on my calendar today?" via @openclaw I don't know what these local LLM fellas are running on, but a DGX Spark sure is not the best thing

@Davidrejll @bryan_johnson yessir agreed 💯

@topmass @bryan_johnson So real. I don't see how anyone still consumes those billions of videos of pure non-sense on Instagram and Tiktok. I had to uninstall all of them and I'm only here now too. We want information, not brainrot

I create content for X, IG, YouTube, TikTok, and email. The video below will do 1-3 million views on IG, YouTube and TikTok but will perform horribly on X (maybe 20-50k views). The video also has a longer total life on the other platforms, whereas dead on X within 24 hours.

@Davidrejll @bryan_johnson yeah bad for longevity too BRIAN - i use none of those apps and my feed is mostly text / github based on X about tech stuff I'm interested in learning about. seems like a good thing for longevity since we know short form video content is bad for mental health

@topmass @bryan_johnson Exactly. Short for content is very close to lobotomy

@iflessthan3 @theo as a full time user I second both or these changes great suggestions

ok fine. t3 code is good @theo but 1. Let me archive threads like Codex app. I only care to keep in the main view the threads I'm currently working on. But don't want to outright delete old threads 2. Let me pin projects. Then I can have the best of both worlds from manual sorting and "last user message" sorting

@okinalog thank you Japanese people for showing me - i love japan :) I will buy one!

Overseas people may not be familiar with it, but this is a mouse called "Trackball Mouse" that Japanese people love very much. Move the pointer by sliding your thumb without shaking your arm by a millimeter. The old generation of mice had a ball on the bottom, but I have the impression that it is a thumb version. The advantage is that it can reduce the burden on the wrist and save space. More than anything else, the smooth rotation of this ball is comfortable. If you have the opportunity, please give it a try. In addition, recently, even unknown devices beyond the scope of the mouse have appeared. "Nape Pro" You can place your favorite commands on all 6 keys, and you can also move the pointer. And it's super small. This may be an effective device for those who do not have a stereotype about the trackball mouse. Please be aware of it.

@thekitze just got a 3gb plan for roughly 70US it is insane never had this kind of speed before but the newest first world problem has been upgrading all my hardware network cards to even take full advantage if they do fibre to your area ideally you should be able to get 1gb up / 1gb down

they are finally digging to set up optic fiber on my street after 3 yrs... i can finally get rid of starlink 😭 it's gonna cost 1/4 of the price and have gigabit download and ~250 upload... twitch.tv/thekitze might be back in full swing these is my current speed ($100/mo)

@aykutkardas @Cloudflare hono was built by a cloudflare employee I think or ex maybe? it's made for workers and they are a lil match made in heaven - I have built many small projects with two workers, one react+vite and the backend a hono worker that handled everything - scales infinitely no problems

@topmass @Cloudflare will check, thank you.

No Next.js. No React. No TypeScript. just a simple html and js file. deployed on @Cloudflare workers. feels… enough.

@yoyonofukuoka I just wanna jump on to say I freaking love Japan and Japanese people pls japanese homies let be friends I will translate and reply always

あらゆる言語が自動翻訳に対応し、世界中の人々がシームレスにコミュニケーションを取れる様になったら、国対国という従来からある構図が崩れて、常識対非常識という構図になるだろうな。

@ibuildthecloud well when you realize that you're not limited anymore you would have no need to buy a macbook and thats a problem for timothy apple

Wait I've never paid attention to iPads. Why is this not just macos? It's like Mac OS, but you can't run a terminal. If you could put Brew on this and launch VMs this would be the perfect device for me.

i have solved the claude rate limits and which is better problem I had and it is just to absolutely dive in head first with pro plan for codex and max 200 for claude and put my head down and build good bye

@stevibe sweet thanks for passing the into along I'll keep in mind that low temps still being used a lot for deterministic stuff and experiment on the models I run as well 👍

@topmass They do! Berkeley's BFCL and Databricks both use low temp for function calling evals. Reproducibility basically requires it (github.com/ShishirPatil/g…) But yeah for creative use 0.7+ totally makes sense, different goals

Qwen3.5-27B went 15/15 on our tool-calling benchmark. But which quant should you actually run? Tested Unsloth's Q2_K_XL all the way to Q8_K_XL TL;DR: Q8 — 15/15 ✅ Q6 — 15/15 ✅ Q5 — 14/15 Q4 — 14/15 Q3 — 14/15 Q2 — 13/15 Q6 is the sweet spot. Same perfect score as Q8, smaller footprint. Also, the results scale almost linearly, seems like ToolCall-15 is actually measuring something real.

@stevibe That makes sense, generally I find myself on closer to SOTA models never going below 0.7 lately since most use has some creative elements. I know most labs just recommend 1 as a default now. I'm curious if they also test their models on lower temp during benchmarks

Good point! Temp 1 is definitely the default for general use, but for tool calling benchmarks specifically, temp 0 actually performs better. Databricks found accuracy can swing up to 10% between temp 0 and 0.7 on function-calling tasks, since it's structured output (pick the right tool, pass the right params), you want determinism, not creativity. Same reason you'd use low temp for JSON or code gen. Reference: databricks.com/blog/unpacking…

@theo as someone with boomer eyesight and a huge monitor I found out I could do control & +/= to zoom in and it scales great so I recommend adding buttons/a tooltip for this so people know they can easily scale the PWA to their liking

Best part is how many of my own assumptions Julius has challenged correctly. I thought the built-in terminal wasn't important, now I use it constantly. I thought a "one click PR button" would be a slop farm. It's the only way I make PRs now. I thought the "web app" (npx t3) would be just as popular as the electron app. I haven't opened the web version since we launched. This is so fun.

I'm supposed to be filming right now. Instead I'm on PR #3. T3 Code is the most addicting coding experience I've ever had. It makes me wish I had more time for code :(

@sinnformer @BradGroux @AnthropicAI @claudeai @DarioAmodei @FTC @DigitalMeld lol brother are you saying you think the open source completely free project of openclaw is a competitor to anthropic ??

@BradGroux @AnthropicAI @claudeai @DarioAmodei @FTC @DigitalMeld ok buddy, you just have the flair for funsies

You make this announcement BEFORE you make the change. Good riddance... what a shame @AnthropicAI and @claudeai. Customer service isn't hard, @DarioAmodei. This was a rug pool and y'all got caught. I hope the @FTC fines you.

@StirlingForge @stevibe nope this is only v2 for the opus-reasoning-distilled variant of the qwen 3.5 27b model - all the qwen models otherwise don't have versioning differences 👍

@topmass @stevibe Is there a v2 for 9B parameters?

Which local models can actually handle tool calling? I built a framework to find out. 15 scenarios. 12 tools. Mocked responses. Temperature 0. No cherry-picking. Tested every Qwen3.5 size from 0.8B to 397B, and since some of you asked after the distillation tests: yes, I included Jackrong's Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled too. Only two models went all green: the 27B dense and the distilled 27B. The 397B? Failed two tests. The 122B? Failed one. The 35B? Failed two. The timed-out results — mostly on the smaller models, are cases where the model got stuck in a loop, repeating the same tool call until it hit the 30-second limit. The test that exposed the most models: "Search for Iceland's population, then calculate 2% of it." Simple, but 35B, 122B, and 397B all used a rounded number from memory instead of the actual search result. They didn't trust their own tool output. Small models hallucinate data. Big models ignore data. The 27B just threaded it through.

@zahlekhan @TFP_tweeets @jpschroeder I hadn't even heard of it - I'll check it out and see !

@topmass @TFP_tweeets @jpschroeder Have you checked out OpenUI. It faster and built based on a year of running GenUI in production. Disclaimer: I'm one of the maintainers.

We’re open sourcing ArrowJS 1.0: the first UI framework for coding agents. Imagine React/Vue, but with no compiler, build process, or JSX transformer. It’s just TS/JS so LLMs are already *great* at it. AND run generated code securely w/ sandbox pkg. ➡️ arrow-js.com

@unclebobmartin in unc we trust i agree

I gave the new Claude a try, but Codex is still better IMHO.