Cole Brown @dtcb

“i’m sorry, it’s so strange to be on the internet.” Brooklyn, NY Joined August 2007

Tweets

314
Followers

394
Following

846
Likes

10K

Cole Brown @dtcb

24 hours ago

@difficultyang I take it all back this skill is hell

0 0 0 7 0

View Details

Cole Brown @dtcb

a day ago

@difficultyang Hahah sorry "big" modified "skills maximalist" not "skills" there! But yes, it's short and sweet. I like the google doc approach!

1 0 0 15 0

View Details

@difficultyang The grill-me skill made me think of you. Maybe you incepted me but I've been getting frustrated @ 5.5 lately for jumping to action too eagerly. This skill and a simple "grill me on this" gets you closer to Opus conversational nature.

1 0 0 24 0

View Details

Cole Brown @dtcb

a day ago

@BryanMcNamaraUS @banteg FWIW I don't think there isn't value in self-hosting. It's just not price-competitive and it's important to understand that.

1 1 0 93 0

View Details

Cole Brown @dtcb

a day ago

@BryanMcNamaraUS @banteg I'm referring here to the break-even w/r/t GLM-5.2 self-hosted on 4xSparks vs via inference provider. I will concede they do advocate for the $200 closed subscription there. I personally don't! Just saying the math of self-hosting is hard to justify, is all.

1 0 2 103 0

View Details

Cole Brown @dtcb

a day ago

@difficultyang I'm not really a big skills maximalist, but this one's pretty nice!

1 0 0 20 0

View Details

Cole Brown @dtcb

a day ago

@difficultyang github.com/mattpocock/ski… github.com/mattpocock/ski…

1 0 0 21 0

View Details

Cole Brown @dtcb

a day ago

@difficultyang Lol this thread sniped me smh

0 0 0 164 0

View Details

Cole Brown @dtcb

a day ago

@BryanMcNamaraUS @banteg They are just saying it’s more economical to use the model via an inference provider on datacenter GPUs. Not that you shouldn’t use GLM 5.2.

1 0 5 203 0

View Details

Cole Brown @dtcb

a day ago

@egorFiNE @banteg I think you would be limited by unified memory w your KVs here. Napkin math with bf16 KVs is ~80GB for 1M tokens worth of KV. They document using FP8 for long context rollouts, so ~40GB. 4xDGX Spark has 512GB unified memory. So you could maybe handle B=2 with a INT3 quant.

1 0 1 94 0

View Details

Cole Brown @dtcb

a day ago

@sebkrier This 3B Parameter Model Benchmarks Higher Than Opus 4.8 On This Benchmark That Was Saturated Two Years Ago

0 0 6 312 0

View Details

Cole Brown @dtcb

a day ago

@xeophon I think it was not available long enough to have a really strong opinion about its strength, and yet… What I do feel (and hope everyone feels) is what life under Anthropic dominance is like. Not fun!

0 0 1 105 0

View Details

Cole Brown @dtcb

a day ago

@willccbb It is my sincere hope. If the intelligence is commodified there is a lot more market incentive to sell marginal compute to a broader set of companies since there will be more opportunity to capture 0 to 1 upside. At least I hope!

0 0 0 12 0

View Details

Cole Brown @dtcb

2 days ago

@PitchingNinja Not sick, arthritic

0 0 1 174 0

View Details

Cole Brown @dtcb

2 days ago

@scaling01 Just takes reps. Happens to us all :)

0 0 0 104 0

View Details

Cole Brown @dtcb

3 days ago

@bilaltwovec “Synecdoche New York” is a wonderful film. Worth a watch. Using AI to write film seems very stupid right now.

0 0 0 28 0

View Details

Cole Brown @dtcb

3 days ago

@dexhorthy This looks really wonderful. Reading through EULA, I paused at "3.1 Customer Data"—does the permissive royalty-free license to customer data include proprietary code? Is there any opting out if so?

0 0 1 44 0

View Details

Cole Brown @dtcb

3 days ago

@thdxr Orcs did 9/11

0 0 11 797 0

View Details

Cole Brown @dtcb

3 days ago

@thdxr I lived through a decade of crypto work. It is the easiest way to fail: build for a hypothetical builder who is supposed to build the thing of value lol.

0 0 1 24 0

View Details

Cole Brown @dtcb

5 days ago

@eliebakouch I guess CSA and HCA are that much more efficient than MSA! There’s so much going on in the DS4 architecture it really is remarkable that it all came together. I wonder if mHC was a key step in making all that compressed attention work.