Brian Thomas

Software Engineer

[Brian Thomas @BrianJThomas]

@HBombUnited @wccftech 🥶

[Brian Thomas @BrianJThomas]

@eddiejiao_obj Yeah, this is the only reason I don't use it, actually.

[Brian Thomas @BrianJThomas]

@viditchess Even the smaller models are tight on the 5090. 6000 pro is better, but more expensive of course.

[Brian Thomas @BrianJThomas]

@DDrolapas Even in the south bay there are 1br apartments going for over $6k already. Surely SF is more...

[Brian Thomas @BrianJThomas]

@marcusx_ @namespacelabs So, this was funny timing.

[Brian Thomas @BrianJThomas]

@marcusx_ @namespacelabs You can’t buy M5 max in any other form factor currently.

[Brian Thomas @BrianJThomas]

@ryancbriggs lol. I was going to tweet this, but you beat me to it.

[Brian Thomas @BrianJThomas]

@BoringBiz_ Student loans should be private only. The cost of education would dramatically drop. Even better if loans are dependent on qualifications such as test scores and the expected earnings of the degree.

[Brian Thomas @BrianJThomas]

@arcinston I think maybe 4x or 8x makes sense. 4x gets you 1 TB/s and you can run larger models than the 6000. 8x will run models at 2TB/s larger than 2x DGX stations can at half the power.

[Brian Thomas @BrianJThomas]

@signorina37H I spent an entire day trying to download the JDK to try Java.

[Brian Thomas @BrianJThomas]

@0xIlyy Even with the 5090 you have to aggressively quantize or work with small contexts. You really need an 6000 for even the small models.

[Brian Thomas @BrianJThomas]

@ptrschmdtnlsn I wind up turning all my fan curves down because I prefer things run hot and quiet.

[Brian Thomas @BrianJThomas]

@luckyatr @midstate_dev @flaccyboi @nvk Still doesn’t matter without some source of entropy. For most uses, it doesn’t make sense either since performance is more important than security.

[Brian Thomas @BrianJThomas]

@TechMDAI In the US, you need to either run an extension cord from another circuit or have another circuit/240v wired to your home lab to power such a thing.

[Brian Thomas @BrianJThomas]

@tallsnail This is what I do. When people try to haggle, I just tell them I tried to make the price good and don't like haggling. I'll lower it if nobody buys it.

[Brian Thomas @BrianJThomas]

@theo All the AI haters in the maker community make me sad.

[Brian Thomas @BrianJThomas]

@theo People on Reddit were saying CLIProxyAPI doesn’t handle well when context gets too big in Claude code. Is that a solved issue?

[Brian Thomas @BrianJThomas]

@emilstridell @Tech2Wild Interesting. I wasn’t aware people actually used KV quantizations outside of hobbyists. Thanks for the info. Are people going as low as NVFP4?

[Brian Thomas @BrianJThomas]

@Tech2Wild I guess you can do 2x TP and 3x PP with 6.

[Brian Thomas @BrianJThomas]

@Tech2Wild Realistically, GLM 5.2 takes 8x sparks today, but I haven’t seen anyone post about it yet. With software support, 5 or 6 should be possible.