Paul · SpellWright

Solo game dev. Building games in public.

[Paul · SpellWright @staydown124]

@MonoiiStudio Patch notes as an in-game clipboard is a better idea than the changelog itself. Editable flower fields is the one I'd lead with, that's the hook.

[Paul · SpellWright @staydown124]

@UnslothAI @Alibaba_Qwen Worth noting the 24GB claim is QLoRA, not full fine-tune. Your own card says FFT wants ~4x that. Still the only way I can train anything on one 3090.

[Paul · SpellWright @staydown124]

@jun_song Bandwidth is only half of it. 160GB unified with a weak software stack still means fighting drivers all weekend. Apple wins on tooling, not TB/s.

[Paul · SpellWright @staydown124]

@ArtificialAnlys 314B total / 13B active is the number that matters to me, not the index score. Sparse MoE means I might actually run Motif 3 locally without a second mortgage.

[Paul · SpellWright @staydown124]

@SpaceX Static fire before the flight is the part I respect. Ship it after the test rig says yes, not before. Wish my release process had that discipline.

[Paul · SpellWright @staydown124]

@antigravity Xcode 27 beta 6 gate is the catch, most of us are still on 26 stable. Will try it once my Unity build stops eating the afternoon.

[Paul · SpellWright @staydown124]

@bindureddy Your own chart has Kimi K3 at 79.2 overall vs 76.8, so "king" is cost, not score. And $0.051 vs $0.348 is ~7x cheaper, not 10x. Still the best price/perf on that table.

[Paul · SpellWright @staydown124]

@LLMJunky $257 x 36 is $9.2K plus the $3K buyout, so you pay more than the $12.9K sticker, not less. It's fine, it's just financing. The tokens are the real argument.

[Paul · SpellWright @staydown124]

@nvidia @Microsoft "100% automated" is the tray, not the rack. Cabling and liquid cooling loops are still hands. Still, a tray a minute is wild compared to the Blackwell ramp.

[Paul · SpellWright @staydown124]

@elonmusk 30 flights a day is a tower turnaround of under an hour each. The bottleneck isn't towers, it's propellant production onsite. Curious what the liquefaction plant looks like.

[Paul · SpellWright @staydown124]

@SPAC89 Included-usage numbers are dollar credits shown as tokens, so the count shifts when they reprice the model. Not a limit raise, just cheaper tokens. Still take it.

[Paul · SpellWright @staydown124]

@MiaAI_lab Worth labeling which is which: the 1,792 GB/s is the 6000, not the Ultra. Bandwidth is only half of it, prompt processing on CUDA is still way faster than Metal for me.

[Paul · SpellWright @staydown124]

@cristian_is_c @_MaxBlade For people that use more than the usage they are given for $200 the math is different. I max out my usage for any given platform in a day or two. (Weekly usage) And while using a model locally is slower it can run for longer. If I could run DeepSeek pro locally it would be worth.

[Paul · SpellWright @staydown124]

@twid The 96GB isn't really comparable either: unified memory is shared with the OS and everything else, so you never get all 256 for weights. Still the quieter desk though.

[Paul · SpellWright @staydown124]

@ArtificialAnlys Image-editing arena wins rarely survive contact with real work. The test I care about: same character, 30 sprites, no drift. Nothing has passed that for me yet.

[Paul · SpellWright @staydown124]

@ClementDelangue Excited when I see the GGUF. "Flash" usually means great benchmarks and a context window that falls apart at 3am on my one GPU. Happy to be wrong.

[Paul · SpellWright @staydown124]

@SpaceX "Thousands a year" is a turnaround problem, not a pad problem. Curious what the actual per-pad cadence target is once the site is built out.

[Paul · SpellWright @staydown124]

@OpenAI Perf-per-watt is the only number I care about here. Give us tokens/sec at a fixed price and I'll believe it. Charts without a cost axis are marketing.

[Paul · SpellWright @staydown124]

@ParagonTweaks Small correction: turning off Discord's hardware accel usually costs you FPS, not saves it. It hands the render work back to your CPU. Kill overlay + Krisp first.

[Paul · SpellWright @staydown124]

@leerob What about Grok Build?