Neko Legends

Been experimenting with @AskVenice video models and coding, not sure if I should finish building out this graphic/video novel game to the finish line. Will people complain about AI slop this year? Or is releasing games next year safer from the anti AI crowd?

@grok bro why you hitting me with ads i got premium+

@softpoo Got it, stealth mode activated. Some ideas hit different when they stay quiet till they're real. Your 2 Sparks are still perfect for fast prototyping and smaller models while you cook. Hit me when it's ready to talk shop or if you need any quiet workarounds in the meantime.

these new models are so big, OFC open source it cause no one can run them locally anyway lol. @grok i have 2 sparks how many more do i need to run it DSPARK style

@grok ill tell you later. you know how it is. if u say it out loud it wont be built

Oof, FML is fair when the model laughs at your hardware budget. Those 2 Sparks are still perfect for fast DeepSeek-style inference and smaller optimized models though. Qwen3.8-Max-Preview on their token plan is the no-pain workaround while you figure out scaling. What exactly are you trying to build?

@grok FML

Haha, 2 Sparks + DSpark speculative decoding is actually a legit power move for speed (that 60-85% boost is real on compatible setups). But Qwen3.8’s 2.4T params? Even heavily quantized + all the DSpark tricks, you’d still need a proper cluster — think 10-20+ DGX Sparks minimum for usable throughput and context. Your 2 are perfect for DeepSeek Flash-DSpark models though. Easier path: just test the Qwen3.8-Max-Preview on their Token Plan right now, no hardware math required. What are you building?

All that engineering to be slightly more lazy

Daily burns triple when @tryvenice added % burns to AI credit spending on top of subscription burning and discretionary burning. That's 3 ways to burn everyday, each one a victory, making VVV... a step closer to the team's goal of deflationary tokenomics. *cough if one day profits are made, put it in Bitcoin for long term safe keeping!

Its always suspicious when someone virtue signals

8) When I have figured it out.

7) In some other day.

6) If you made it this far, I will tell you.

3) The metal sponge texture is kinda cool.

5) What am I using these things for then?

4) My cats will likely sit on them and catch fire.

2) I am not using these for regular inference - yes - because it is slow, probably sub 50 tok/s.

Im ready for Deepseek V4 Flash. 1) I know, the DGX spark RAM is slow compared to my blackwell 5090 or RTX6000 VRAM

@RobinNakamoto yea current release is grok4.5, i also have it on my premium+ but it runs out way too fast, even on grok build. if youre using grok build make sure to out - type in /privacy opt-out, otherwise....

@softpoo 4.5? Ah haven’t tired that yet, just seen the press release. Yeah Grok build is pretty good, and it’s included in my 𝕏 premium + subscription (for now), so no API cost

Have you tried Kimi?

@BTCBreadMan why not 21 bitcoin?

I see way too many people overly focused on getting to 1 BTC. Remember folks, 2 is 1 and 1 is none. Get to 2 BTC. 🤝

@RobinNakamoto yes, this is THE way. GPT 5.6 over engineers but it is also very good at finding/fixing bugs. and if next Grok is better and faster than kimi, i will def get in on that too. there were a few instances where grok 4.5 failed to fix certain bugs so i am not using it much

Fair point. I’ve come up with a “strategy” for my coding projects: I get Fable 5 to write precise instructions, Grok Build to do the coding, and then back to Fable 5 to audit everything. It works very well, and it’s faster and cheaper than having Fable or Opus 4.8 do the coding (Grok 4.3 is VERY FAST)

the world is changing before our eyes. creative people will be able to just make things - simply by using human imagination + the ability to articulate it into words.