John Pippett

We really need to move away from “this is absolutely the best model” toward, “this is the best model for my use case”.

@blader It’s the best open source model I’ve ever used, cheaper than Claude by 10x, and nearly as good.

glm 5.2 has been out for like a week, why did everyone start glazing it all the same time just today?

People still talking about running out of their expensive subscription plan usage and I’m only paying like a 1$ a day to have GLM 5.2 run all my agents.

@TheFinalRecord @BenjaminDEKR This - Jevons paradox. Cheaper inference leads to more usage, not less.

Solo devs will scale up their workload but so will companies. As long as companies prefer renting cpu over paying for their own hardware and staff, there will be demand for data centers. There will also be heaps (already is) of solo devs that will need more than a $5000 pc for their workload.

What happens to the datacenter economy if / when we have GPT 5.5-level models running on $5,000 desktops in our homes and offices?

@BenjaminDEKR The world will rejoice.

life would be awful without tailscale

@usr_bin_roygbiv For Hermes is rocks, ironically, I can’t get Kimi to do anything in Hermes. It just ignores its tools and then just starts making things to do the job instead it’s actual tools.

every person i've seen claiming glm 5.2 is absolutely insane is using claude code or opencode

@juan_quenga @rezoundous Yeah that’s exactly what I mean, but a crappy 80$ Dell Optiplex, stick in a stick of DDR4 16GB and run Ubuntu Server OS. Then you can just SSH directly to it from your house, or turn on tailscale and access it literally anywhere in the world. Same concept, but alot cheaper.

@JohnPippett @rezoundous i don’t disagree, but mac mini’s are still great to run openclaw, hermes, and tools like codex and pi remotely

Is getting a Mac Mini for AI still a good option nowadays?

@WrongShore42 @durreadan01 I mean I agree but I do disagree. Some apps had pretty poor legibility, it did need some tweaks. I think the slider is a good middle ground.

@durreadan01 I have to say that all the reviewers moaning about legibility problems with iOS 26 beta 1 ruined Liquid Glass for everyone by the way. As a consumer I wanted it that clear! And Apple clearly overcorrected and won’t give that back to us ever. 🤦🏾‍♂️

Am I the only one still using the default Liquid Glass setting? 🤔

if working in aerospace has taught me anything it’s that any engineer who speaks in absolutes about hardware is full of it

@TheAppleDesign They’re going to make an Air 2?

Which design would you choose for the iPhone Air 2 ??

@VerdeSelvans That’s were they put all the extra AI, in the bezels… oh wait.

iPhone 17 vs iPhone 16 bezels 🫪🫪🫪 The difference is huge

We may get iOS 27 Beta 2 before my iPhone stops indexing.

This really is what AI has been amazing for - you need some weird tool/software no one else cares about? Bam, it’s built in 20 minutes.

@AdoHernandez27 I mean I like the guy, he really turned it around after the last incident but he’s been a bit of a ghost. Losing Ross Colton was harder.

Alright, #Avs fans, do you actually want to see a Nichushkin trade? I reported back in March that he was in play and got lambasted for a week so I’ll just ask a question. #GoAvsGo

@beyang The latter - orchestrator should be the one checking and fixing work, not the other way around.

Which do you think you’d prefer? Fast main agent with smart subagents? Or smart main agent with fast subagents?

@thdxr The cost of this that scares me - less security. Models still hallucinate. I mean people should be scoping their keys anyway but giving everyone the power isn’t always the best move.

i've been building things for developers for a long time which means i have a lot of experience with api design but LLMs are making me reset a bit instead of neat simple APIs it's probably worth trading that off for more powerful APIs

@infohazard_lol @BLUECOW009 Absolutely - Q4_K_M with 32K context is 10GB. It’s tight but it works. You can compress KV cache to Q8 with llama cpp as well and save even more space. Turnoquant will eventually shrink it even more once llama cpp has full support for it.

@JohnPippett @BLUECOW009 You can comfortably run 12B on your 3080? With an average context window size?

Running AI locally is not cheaper

Ok it’s official, I’ve spent a lot of time with GLM 5.2 since it launched and… wow. This is the first open source model I feel I can trust as much as GPT 5.5 or Opus.

@loktar00 Ironically this is the perfect example of how much choosing the right harness matters. For example - Bambulab P1P, no problem. Ender 3 and you might wake up to a fire.

Running a huge model at home like GLM 5.2 Q2 where tk/s is low is a lot like 3D Printing large models. You wake up and you either have a gift waiting for you, or a big mess, it's all about the surprise! 😂 Wish GPUs were as cheap as 3d printers.