Dave " " Buchhofer retweeted
I can’t stress enough how little an idea matters compared to the agency of the people executing the idea.
I have had the privilege of knowing and sometimes even working with some of the most successful people (by various metrics).
The difference between mediocre and excellent work and outcomes is predominantly one of agency.
In practice this means: they dont wait for things to happen to them they go out and make things happen for them.
They don’t wait for someone else to do something, for someone to teach them, for someone to give them the path, etc. They just go out and find a way to do it.
I think the single biggest superpower these people have is the realization/belief that the world around them is completely mutable. Most everything that happens is because a person made it happen.
I used to tell people to look around the room you’re sitting in. Look at everything. Every noun. It almost all exists because a person willed it into existence. Nothing is stopping you from doing the same.
I see people online all the time dismissing someone else’s success because “I had that idea first” or whatever. I mean… yeah? If so then the difference is… you. So a bit of a self own whenever I hear that.
Number one tip: act with agency.
Dave " " Buchhofer retweeted
In my day we didn't need "/grill-me", we would hire a 24 year old designer, loosely describe the feature, ask them to imagine the user interface and draw it for us, then come to a shared understanding by roasting the shit out of the designs in an hours long humiliation ritual
Dave " " Buchhofer retweeted
The funniest thing about LLMs is that they let you generate endless work in domains where you’re completely unqualified to evaluate the output.
Dave " " Buchhofer retweeted
gpt 5.5 has goblins and can think. claude 4.7 doesn’t have goblins and can think. so it’s not the goblins that make the models think, it’s something else entirely
Dave " " Buchhofer retweeted
me, scoffing in the back row of SIGGRAPH: ok I guess this is good if I ever need to render a teapot I guess??
Dave " " Buchhofer retweeted
your biggest enemy is still complexity. it's also your agent's biggest enemy. but it has no holistic view of your code base, so it keeps adding complexity.
and you think that's how it's supposed to be, because the clanker shat it out, and you don't know the stack.
glhf!
when you think of feature A your job is to recognize that it's actually a special case of feature X and if you implement feature X it unlocks feature A B C
people are struggling with this lately because they can't think bigger than what their agent can do
It's dangerous now that it’s easy-ish to run a wild ass experiment just to learn?
can I run the new qwen model on osx locally via my old 3090 in tinygrad vulkan -> llama.cpp -> egpu
osx egpu worked, it's ~3x faster than the same on metal for stuff that can stay GPU resident?
it loaded! 🎉
✅ Qwen3.5-35B-A3B loaded successfully!
Architecture: qwen35moe
Blocks: 40
GatedDeltaNet: 30, Transformer: 10
Experts: 256 total, 8 active
That's realize=False though — and a whole lot of random ass changes. Guess we'll try more tomorrow after all?
Anyway, no one should take anything from this as fact. Seems obviously something that I'm just a little too early on, and I'm looking forward to where it goes when my comma has these new compute powers (!)
The VRAM problem seems architectural in tinygrad’s load path? need a Q4_0Linear that stores quantized blocks and dequants in the forward pass (like Int8Linear already does)?
The actual quantized files have mixed quant types all over. Had a script that requantized 60 tensors (Q5_0 -> Q4_0)
In bartowski Q4_0. After that Pass 1 (realize=True):
KeyError: ‘qwen35moe.feed_forward_length’
• Pass 2 (realize=False): MemoryError at ~18.71 GB allocation
Dave " " Buchhofer retweeted
New ad by the Norwegian Consumer Council: "A Day in the Life of an Ensh*ttificator"
> Status: GREEN. The eGPU stack works for LLMs on macOS.
that’s a little bit of a stretch at one token per second for the 30 B model, but I wasn’t really expecting you to work at all. I just wanted to know where it broke.
3090 24gb via egpu on osx? Silly.
I wasn’t expecting that to really work at all, since without quantization there’s no real room for actually using the thing. But it’s loading and spitting back a token. Hah.
Prior experiment of bridging wgpu -> metal kernels <-> tinygrad -> back on osx shared mem was an interesting poc, (it worked!) just not practical for my splat tinkering. The stupid "but why?!" experiments will continue until morale improves or I learn something.









