David Grant

Building Tidy, a collaboration tool for everyone

[David Grant @DaveAtTidy]

On a long English session it started flipping whole replies into Chinese, intermittently.

I was running the new MIA GLM-5.3-Flash EXL3, Mia's TensorFold v1.5 on 3 DGX Sparks (official 4bpw, q4 dense)

cold prefill was ~800 tok/s, single-stream decode ~49 (on their spec).

Don't know yet if it's the quant or the new 3-spark harness.

Back on the vLLM TP3 serve until it's stable.

[David Grant @DaveAtTidy]

@sudoingX not either or, both

[David Grant @DaveAtTidy]

big omarcy fan, but yeah, I agree with some of what you shared and had a similiar experience. harness installs don't auto-update, annoying, got rid of them and installed properly. I don't use the quickshell AI helper, I use herdr and built my own quickshell to monitor local AI. I did not have the issues you had with the display, or hyprland windows. but I did have a little bit of hassle getting the drivers working in the chromebook and a bit on the macbook. But it was still a lot faster than dealing wiht driver downloads and troubleshooting in windows would have been!

[David Grant @DaveAtTidy]

@0xSero well, you can get TWO strix halo for the same price. 256 gb vram and just a tad slower. I don't understand the pricing of the 495+

[David Grant @DaveAtTidy]

Every single thing that Opus 5.5 builds has issues that GPT-6.1-sol finds during audit. EVERY TIME. Opus probably would find issues two if it specifically ran an audit on it's own code, but I like the different model family approach. That is worth sub-agents just to have that two step approach. Now, if you want to build anything bigger, A feature vs. a story, you use sub-agents to save context. The orchestrator sees the big picture, and organizes the work, It's an Out-loop construct.

For me, that's all i'm using subs for, single orchestrator, then a series of builders/auditor teams chipping away story by story / feature by feature. They run 24+ hours and at the end, I have a working feature.

[David Grant @DaveAtTidy]

Do you use a window tiling manager like hyprland. Coming from windows over a year ago, I didn't have to do ANYTHING to setup linux, and had a great experience with hyprland and the waybar. Easy to learn and immediately better. I didn't realize how much I hated dragging and re-arranging windows until Hyprland

[David Grant @DaveAtTidy]

@owenthcarey Been using it for over a year, it's on my macbook (older hardware that wasn't being used due to crapintosh OS, now it's awesome), old chromebook (used to be my daily driver until an OTA google update bricked it) now is fast and my main portable, my workstation, and my sons.

[David Grant @DaveAtTidy]

@Spectromachina My check is going away soon, cancelled grok super because 4.7 is not good enough. https://t.co/BEpzNO63FM has until 10/16 to up the game to keep my sub.

[David Grant @DaveAtTidy]

@dark_coderz D. Native code is the only way forward

[David Grant @DaveAtTidy]

@GohilHardy Good point, but if they read, and ask, and learn, they are getting hands on tutoring. How long does it take to learn software engineering? What does it take to earn the title. I know some pretty shit software engineers who don't understand a lot of the domain

[David Grant @DaveAtTidy]

@dark_coderz Take your repo, break into service and app subfolders, put agents.md in each folder with specific guidance and agent maintained notes. Agents are good at working in sections. Sounds like you have a repo organization opportunity

[David Grant @DaveAtTidy]

@owenthcarey Opus for build and gpt-6.1-sol for audit is the winning combo atm

[David Grant @DaveAtTidy]

@ashxhart you and the new team are great thank you so much for doing this amazing work. One thing that isn't clear from some of the benchmarking people are doing is whether or not tensorfold is as effective with multiple concurrency as vllm. Do you have any insight into that. Does it come down to choosing the right inference engine for workload or is tensor fold the one to rule them all

[David Grant @DaveAtTidy]

@ClaudeDevs Worst feature, force every turn to end in more work for you. "Don't bring you managers problems." Take the, you should know output, and send it back to the model to resolve it.

[David Grant @DaveAtTidy]

@hindutva_anshu @nikitabier @elonmusk @grok what is this app about

[David Grant @DaveAtTidy]

I think it's more complicated than one size fits all I think every app should I really have three modes of focus mode where you pick your basic method of engagement let's say it's email and you have a left bar in the main area a studio mode where you can add in at most one additional panel either at the bottom or on the right and then a command mode where you bring in the multiple tabs the left nav icon bars bottom panel right panel adjustability I think that's the way you get the best fit for where customers are and even in my app I enjoy command but sometimes when I'm browsing email I just want to flip to focus mode to have a nice simpler aesthetic

[David Grant @DaveAtTidy]

@skpnky Guardians of the galaxy - Rocket, Groot, Peter Quill, Gamora, Drax... Fun times!

[David Grant @DaveAtTidy]

@Kappaemme1926 I have not noticed that. Probably time to check your skills

[David Grant @DaveAtTidy]

@VictorTaelin Maybe anthropic will give us a 500B parameter version of opus? hahah

[David Grant @DaveAtTidy]

@VictorTaelin I'm rooting for you @Alibaba_Qwen !!!