Cofounder & engineer. Running local AI, building things with it, and sharing what I learn.
- Tweets 8,472
- Following 679
- Followers 7,153
- Likes 14,465
Hear me out, what if Google's team is trolling us...
Claude just talked about removing plan mode and right as that happens, Antigravity ships plan mode.
Then they'll drop Gemini 4 Pro and be the new frontier overnight...
Antigravity 2.0 now features a dedicated planning mode, just like the Antigravity CLI.
Type /plan and the agent steps back to think through the task, and conduct extensive research, before generating an implementation plan for your review. The agent will ask for your approval before it begins executing on the task.
You can also ask for a plan naturally in your prompt, for a lighter version of a plan.
1
3
692
Remember @NVIDIAAI Nemotron 3 Puzzle? with the latest build of llama.cpp, the CUDA backend now supports it.
Puzzle is Super compressed down to 44.5 GB (vs Super's 70 GB). With this update the Mamba "scan" finally runs on the GPU, so those layers don't fall back to the CPU. That makes it so much more usable!
3
1
7
571
the hf deal isnt closed yet. nvidia also says in the filing it'll keep uploads and downloads open and keep supporting other chip vendors. and if they ever lock it down, thats more reason to have the weights on your own box, not less (+ there are other options from which to download). the "will" is about open models getting good enough, not whether a spark exists
34
personal projects i'm building, plus maintaining and improving the local models everyone's running on their sparks
1
49
that's the setup i'm talking about. the hardware's already there for people like us, it's the open models that still need to close the gap before they take the main workload and not just idle hours, but i think we're close to getting there
1
86
that split makes sense, private stuff is the easiest case for local right now and partially how I use it now. my bet is the line keeps moving though, once open models close the gap people will start using it for everyday work too
1
52
how often does it actually escalate in practice? if local handles most of the day that's already the complementary setup i meant
1
2
116
wouldnt bet against that, lots of people are doing the same math right now. how's glm 5.3 flash holding up for daily use?
91
I considered getting one last year for a bit lower than retail and didn鈥檛 because there wasn鈥檛 much I could do with it. Now it鈥檚 unjustifiable for me, but prices are going up as utility for these devices increases
39
Eventually maybe, rn there isn鈥檛 enough compute even for huge companies like OpenAI and Anthropic, people and small size companies will wake up and buy consumer products driving the price up
1
81
your chart kind of says the opposite though, o1 level has been stuck around $0.17 since aug 2025. and for me local was never about beating cloud on price per token, i'll always get more tokens from the cloud. it's for the high privacy stuff frontier labs won't serve me or i just don't want to hand them as data. a pair of sparks gives you a lot of freedom the cloud doesn't
1
3
142