Papamasher

Old man who plays big body grapplers badly

[Papamasher @Papamasher]

@lauriewired I've thought this for a long time but i also recognise it strips people of many aspects of being gamer, there's no ownership of a game, you go from having a right by ownership to rights by lease. Archiving, modding and thousands of other activities would die out

[Papamasher @Papamasher]

@MiaAI_lab @NVIDIAAI you read my mind I was just about to look this all up :)

[Papamasher @Papamasher]

@iamlukethedev Dear skynet, we hope you enjoy living in the sims.

[Papamasher @Papamasher]

@loktar00 Ram shortage is predicted to last into 2028… only one way for prices to go.. increasing production at scale means highest prices for retail. If the whole bubble pops then prices remain high because production will tank 🥲

[Papamasher @Papamasher]

@kingkapoor72 Cool - glad to see all those banks and corporate landlords will be able to scoop up so many bargains...

[Papamasher @Papamasher]

@SlimTradeyBaby @crowleyx @NVIDIAAI Look forward to your future posts about it!
Honestly not sure sparks are the way to go if frontier is the goal - you get diminishing returns because of the amount of data on the cluster network.
It might be time to consider a rack solution? (don't tell your wife i told u that)

[Papamasher @Papamasher]

@capexbt This is what happened to the hunt brothers... https://t.co/Hpmw5iWdGq

[Papamasher @Papamasher]

@witcheer give it a skill to report prompt injection & continue the work flow

[Papamasher @Papamasher]

@forgotv @yacineMTB big topic, loads of smol bespoke models can do smart things in their niche, but broad knowledge is required for orchestrating that effort. Big params for planning and really hard stuff, small and massively concurrent for everything else.

[Papamasher @Papamasher]

@WescheNex1q @SpaceTimeViking look I've created a best in class score 'no think' score - just ignore the fact we destroyed the reasoning in a model centred around it...

[Papamasher @Papamasher]

@araseb_ Easily with the right background - software isn't really the hard part - getting the contracts is

[Papamasher @Papamasher]

@sue_xbt Intelligence comes in two forms, problem solving with applied knowledge and problem solving via discovery (trial and error).
Identify a problem, then start reading about it, is the quickest route to a usable skill.
Discovery is harder - but can open up new results.
take your pick

[Papamasher @Papamasher]

@loktar00 @usr_bin_roygbiv try Step3.7-Flash! first thing I've used locally that felt like it was as good as Qwen

[Papamasher @Papamasher]

After further review and comparison i do not recommend this version of 27b - seems to have suffered badly under compression - results from prolonged attempts at work have not been good.

[Papamasher @Papamasher]

I'm using Qwen3.6 27b NVF4P Text only for coding agents on DGX spark (ASUS-GX10).

https://t.co/rZSIfPLr7u

about 20 tps on each agent, seems pretty solid.

[Papamasher @Papamasher]

@cryptofergani Solar panels, water filters, toilet paper?

[Papamasher @Papamasher]

@0xSero @tmophoto out of interest what was given up?

[Papamasher @Papamasher]

the Crossy road prompt is now going to be my defacto test of the impact of KV quantisation. The difference between the quants is wild. I'll never trust turbo2 with anything :D

[Papamasher @Papamasher]

@sudoingX @MiaAI_lab
I've seen your posts about Step3.7 - what configurations have you tested? I tried putting the V cache down to turbo2 and the results were baaad.
Good results on Turbo3 tho.
Interested in hearing your thoughts.

[Papamasher @Papamasher]

@MiaAI_lab @NVIDIAAI I would love to see this too.

[Papamasher @Papamasher]

@ItsmeAjayKV tbh I think you will lose a lot of bandwidth to the pci lanes - but in theory you could spin up any 70B model with relative ease.
The trick will be tensor splitting across the cards if u use both.