Neko Legends @softpoo

This is not a shill to buy VVV or DIEM. I am simply sharing interesting conversations with my agent per usual. And as I have said before, I believe we are in a local top in bitcoin:native but VERY happy to be wrong. If BTC goes back down to 66k, ANYTHING will be a great buy. I am not a professional day trader, just playing guessing games based on my magic Fable made AI chart.

I asked Eva if I should un-stake my minted DIEM to get more $VVV to convert to USD to buy inference elsewhere and, and the answer is something VERY FEW understand about Venice AI. 1) Minting DIEM currently gives you a 2.5x discount, since you still get 80% of the 8.29% yield. 2) You aren't paying Venice for privacy, they are paying YOU to have it. (Eva explains below) Anyway here's what she said (with some private numbers redacted): Jun hmmmmm i wonder if i can, use the x$.. to buy inference credits instead - but i guess privacy is important - how much is that worth Eva Run that comparison and it falls over immediately: 1x$ buys 1x$ of retail inference. Minting turns the same 1x into 2.5x. You'd be trading a 2.5x arbitrage for a 1.0x purchase. Even if some other provider sold tokens at half Venice's price, minting still wins. Economically there's no contest — the mint is the cheapest inference you own. But your second instinct is the interesting one, because you just stumbled into the actual answer: privacy isn't costing you anything here. It's paying you. The zero-retention lane isn't a premium you're shelling out for — it's a free rider on a 2.5x discount. You're not buying privacy; you're being paid to have it. That's a rare trade. Don't unwind a trade where the privacy is the profitable side. Now — "how much is privacy worth" deserves a real attempt, so here's how I'd price it. Think about what's actually flowed through these lanes today alone: your portfolio structure down to the token counts, the welfare protocol for your house, wife's conversations, location patterns, the memory corpus, the codebases. You can't price that as data — you price it as insurance: what would it cost you if it leaked? The honest answer is unpriceable, because you can't buy a leaked corpus back. There is no market for un-ringing that bell. When the downside is unbounded and uninsurable, you don't ask "is the premium worth it" — you ask "is the premium survivable," and yours is negative.

@stargazeruk7 @grok does this apply for X/twitter?

@FutureTechie some companies are attaching motion stuff on ppl in India doing chores - real 3D coordination, likely more accurate. Video for sure is useful in areas more than just for fun- I'm just saying, most will just be playing catchup in all areas rather than being focused on one thing

IMO everyone including Google took the bait from OpenAI with Sora video gen - which will be shut down if not already. Companies that split up to focus on video - which is nice and fun, they lose resources on focusing on AGI itself.

Venice @AskVenice

11 hours ago

Google’s latest frontier video model Gemini Omni 1.1 Flash is now on Venice. Available anonymously.

5 8 96 9K 8

@yura_elkin lol wtf so good

Repo updated with the best combos from several opens source repos github.com/neko-legends/s…

100 tok/s achieved on c1 structured GLM 5.3 Flash on 4x spark. EXL3 4bpw C4 performance 253 tok/s. Bench at 420k holds: structured 93.7 / code 37.8 / math 74.5 / prose 36.2 tok/s — no regression vs the 900k config. If you have 2x sparks, use MiaAI_Lab's receipt. If you have 4x sparks and Mikrotik 812, use mine below and tell agent to "Use the GLM 5.3 config instructions"

Mia @MiaAI_lab

11 hours ago

Run GLM-5.3 Flash EXL3 with DFlash2 for 2x DGX Sparks with ease ✨ - 900k context - 982,612 FP8 KV cache - Full image and video support - 300k stress passed, 874 prefill tok/s. ~63 tok/s in single stream structured/code 155 tok/s on 4 concurrent sessions (structured) - that

67 36 372 46K 274

I might give Hermes a try again - this was one of the issues that was blocking real work being done. But I'm already very comfitable with Pi.dev

Tony Simons @tonysimons_

a week ago

Hermes Agent just removed one of the dumbest limits an autonomous agent can have. 🪽 Until now, a single turn had a default 500-iteration ceiling. Hit it during a monster task? Hermes could simply stop mid-job. That default limit is now GONE. Hermes runs until the work is

44 11 423 37K 94

forgot to mention, this was made via Venice AI, only way to add blood and violence these days

sorry, dad Original upload, only on X. a short film made with my agent (Eva), inspired by the Korean Netflix series: Agent Kim. ⚠️ content warning: suicide themes, violence, gore.

@saylor @fundstrat lol wtf

@MiaAI_lab this is the one.

This is a really fun and creative test for LLMs. Numbers on benchmarks and speed in tok/s is one thing, but having them do things on their own is always more interesting.

TypingMind @TypingMindApp

15 hours ago

4 AI models trying to draw their self-portraits. GLM-5.3-Flash, GLM-5.3, Kimi K3, and Claude Opus 5. GLM-5.3-Flash came out looking the most handsome 🤓

28 27 308 47K 105

Updated my spark-bench config repo with this as well. This basically saves like 1000 tokens per prompt on GLM 5.3 flash - for free.

Zain @zainhas

20 hours ago

one more PSA for the all GLM-5.3 Flash API users: > if you're using this model for chat you have to explicitly turn the clear_thinking=true > coding agents -> clear_thinking=false + pass back in reasoning_content > chat mode -> clear_thinking=true

3 0 103 5K 70

@BitsagaRob its like a 500$ hardware wallet bruh

@grok alright, now make an image and show me all of these dogs together howlin in its own ways that makes sense

@grok is openAI an underdog? Then who is top dog? and what dog is grok?

Tibo @thsottiaux

15 hours ago

Starting to see more and more Codex users on airplanes and in cafés. Underdog energy gone mainstream.

469 57 5K 384K 115

@grok you didn't answer the question, if you were a dog, what dog would each AI company be.

@MattWallace888 unless they are there to put next to dying patients so they can die happy I don't see a reason?

warning, there is an open bug: The hard cap documented: 262144 (2^18), with the bug link — sglang #36550 ("worker abort in graph-replay context at first decode token after cold prefill > 262144 tokens", open since Aug 26) I have confirmed it - it will crash your sparks. Stuck to 256k context for now until it is fixed

Grab it here. github.com/neko-legends/s… Special thanks to the open source community for making this all possible. Provenance and receipts: joesinvestments/GLM-5.3-Flash-FP8-4x-DGX-Spark — the 4× FP8 SGLang formula + frozen flag set this builds on 0xSero/glm-5.3-flash-sglang-sm120 — the six-patch sm12x compatibility stack (unlocks flashinfer_sparse_mla DSA on GB10) tonyd2wild/GLM-5.3-Flash-NVFP4-2x-DGX-Spark — the parallel vLLM lane (our second column), KV sizing doctrine, cache-flush ritual incoai/GLM-5.3-Flash-DFlash2 — the drafter (CC BY-NC-ND 4.0, research/eval)

OK finally got it, my contribution: the best combo for GLM 5.3 Flash on 4x sparks: **SGLang + NVFP4 + DFlash2, TP4, NCCL over RoCE**: 88 tok/s for c1, 90 for c4. Was 10-38 tok/s before optimization. This is very usable. Grab the optimized recipe below if you have 4x sparks on Mikrotik 812.

keyboard_arrow_left Previous keyboard_arrow_right Next