Christian Merrill

[Christian Merrill @M_Chimiste]

@emollick I deleted my post because you added the prompt, thank you.

[Christian Merrill @M_Chimiste]

@QuixiAI @0xSero ah the prince

[Christian Merrill @M_Chimiste]

@TheAhmadOsman I very much hope you’re right

[clem 🤗 @ClementDelangue]

The solution to American open-source lagging behind Chinese open-source: @elonmusk & @cursor_ai releasing their model tomorrow in open-source! It's that simple and that would be a massive contribution to the field and to the US!!

[Harrison Kinsley @Sentdex]

cmon Robert. Yea, he is a blowhole indeed.

Some examples:

multi agent/batching/subagents:
He handwaves away batching as a solution to running multiple agents simultaneously. This is literally exactly the solution that it scales very well especially tensor parallelism on vLLM. He claims he's running like 50 agents or something simultaneously. This is curious, and definitely outside of how I personally use coding agents. I currently go out to 8x.

Cost:
electricity $2K a year. His math was if you run it 24.7. This would be an agent literally running 24/7 producing actual tokens. If the GPU is indeed idle, it draws a very tiny amount of watts. Like 12. let's compare fairly. Which API/sub are you running an agent on 24/7 for CHEAPER than $2k a year? what a wildly dumb claim.

RTX Pro 6Ks are definitely expensive. he's not wrong there, but he's completely missing that people ARE buying these and ARE doing local inference and ARE running 100% of their workflow off of any API and off Opus and GPT. His only comparison was to buying a whole ass tiny box though. You can self-build an RTX Pro 6K rig for waaaaaaaaaaay less than a tinybox. You can also slowly build it up over time, as most people do.

I canceled both claude and gpt subs immediately at GLM's 5.2's release and have been entirely local since.

DSV4F is fast. Like REALLY fast. You can do 350 t/s if you are using pcie 4 or 5. I'm actually using that model over GLM 5.2 bc its so dang fast.

This is faster than even claude. You can live completely on dsv4f. Theo could (or should) too.

You can also live on GLM 5.2 2bit GGUF as a sole provider too. GGUFs don't scale quite as good with pipeline parallelism, but DSV4F you get like 1200+ t/s with bs=8.

Theo is truly speaking out of his ass. The problem is some of his arguments have merits but he's taking partial rights + ignorance and forming an argument that's just blatantly wrong.

[Christian Merrill @M_Chimiste]

@stake_jevens @theo 😂 I dropped 27k and it wasn’t 4x

[Christian Merrill @M_Chimiste]

@reach_vb One more wrinkle in codex, don’t make me sign in to codex on my desktop to use the iOS app. That was a bummer when I was like “oh wait let’s use codex on the iPhone” then sad trombone when I wasn’t anywhere near my Mac to click on something.

[Christian Merrill @M_Chimiste]

@reach_vb As someone who cancelled their personal Anthropic subscription I appreciate these. They were definitely missing when I was messing with it last on iOS.

[Christian Merrill @M_Chimiste]

@DJSnM @nikitabier I’ve decided to compress all of it to just SpaceX.

[Christian Merrill @M_Chimiste]

Are we certain it’s not H2 2026?

[@jason @Jason]

2027: Unlimited, free, on-demand bespoke software...

[Christian Merrill @M_Chimiste]

@sudoingX Already have.

[Christian Merrill @M_Chimiste]

@reach_vb Meant to say fork but I’m sleep deprived, sorry

[Christian Merrill @M_Chimiste]

@reach_vb Specifically the branching bit I missed.

[Christian Merrill @M_Chimiste]

@reach_vb I use it still for brainstorming and making my project_docs folder before moving it into codex

[Christian Merrill @M_Chimiste]

@OpensourceWtf @Tech2Wild Yeah going from 8900 to 25k for the Mac was not on my bingo card. Heck even my RTX 6000 Pros have appreciated significantly.

[Christian Merrill @M_Chimiste]

@OpensourceWtf @Tech2Wild I wouldn’t do training on a Spark. I have a multi RTX Pro system for that. Also when I bought the M3U the M5 MBP didn’t exist. Though I do have the M4 with 128gb. Actually I don’t even think the Spark was released when I bought the first Studio.

[Christian Merrill @M_Chimiste]

@QuixiAI New steering vectors?

[Christian Merrill @M_Chimiste]

@MiaAI_lab @hqmank If it does (and it likely won’t) I’d consider renewing my sub.