Jared Castorena

LLM and AI growth is now in my wheelhouse. Startup company with ai 1728Studios LLC. founding member.

[Jared Castorena @JaredC1728]

@NightSkyToday next step: holodeck

[Jared Castorena @JaredC1728]

@JackS10760 @aryagm01 Grok said: put t=1+xy, q=t²z+y²(1+3t). the map
(tq, 3y-11tq+9xq-2t²x²q⁴, 6x-9x²y-3x³z+x⁴q⁴, w): ℂ⁴→ℂ⁴
has jac det -18 and sends (0,11/3,-475/9,0), (-3,4/3,125/81,0), (6,1/3,-7/81,0), (-3,2/3,-1/9,0) to (1,0,0,0)
degree-26 counterexample in dim 4

[Jared Castorena @JaredC1728]

@0xSero I used 3 resets. MoE research is fun.

[Jared Castorena @JaredC1728]

@miao_xiong_cs I have a question about datasets for MoE, newer population-level balancing methods improve over noisy mini-batch heuristics to prevent expert overlap and homogenization, would it be best then to formulate load balancing as a constrained decision problem?

[Jared Castorena @JaredC1728]

@crystlizedd @0xSero You can do it, you just have to offload to cpu and ram which slows it down considerably. Im working on it tho... I shrank gpt oss 20b down from 13gig to 11.9gig im working on 120b now. Passing inference on them tho.. now that is fun. I love Sol

[Jared Castorena @JaredC1728]

@toptickcrypto Gpt just is irritated by 3 questions

[Jared Castorena @JaredC1728]

@mathemetica This could potentially help for making a classifier for resolving MoE routing

[Jared Castorena @JaredC1728]

@LeeLeepenkman you are brilliant

[Jared Castorena @JaredC1728]

@IntCyberDigest "Man-made horrors beyond all comprehension" brought to you by gemini

[Jared Castorena @JaredC1728]

@Gelassoldat 16gb of vram is going to hamper your model type... max you can fit on there with any good sized context would be 8b, maybe 14b if you quantize it down. Are you running agents?

[Jared Castorena @JaredC1728]

@ErenChenAI After market robot mod shops are gonna be an insane market to move into

[Jared Castorena @JaredC1728]

@BLUECOW009 Claude fable put it best, "formalization is easy"

[Jared Castorena @JaredC1728]

I was able to find a smaller archive size for full fp16 weights for gpt-oss 20b. I was able to shrink it down to ~11gb under the size of the MXFP4 version. I left the decoder, if anyone wants to donate some compute I can do the same thing for kimi k3.
https://t.co/G4k7nchpq3

[Jared Castorena @JaredC1728]

@BLUECOW009 Codex. Opening an ai model and seeing whats going on under the hood is too much for claudecode

[Jared Castorena @JaredC1728]

@thsottiaux Single handedly keeping me from using kimi k3 thank you

[Jared Castorena @JaredC1728]

@BLUECOW009 Teacher who vilified calculators but still used the slide-rulers

[Jared Castorena @JaredC1728]

@devops_is_life @LeeLeepenkman In an odd way if they use ai's that have been trained on the total corpus of internal knowledge they are indirectly communicating no?

[Jared Castorena @JaredC1728]

"The formalisim is the easy part" - Claude Fable 5 cir. 2026

[Jared Castorena @JaredC1728]

@yacineMTB Im working on a way to shrink model size. I stared with a turnary model and saw promising results for storage, even smaller than gguf's. This is to store it not to pass infernce on.

[Jared Castorena @JaredC1728]

@VictorTaelin As long as results are checked programmatically instead of "vibe" with strict constraints. This is my way of interacting as well.