wmf

23 ms·

811.
by wmf 8mo ago
I suspect this theory is cope from people who missed the top.
812.
by wmf 8mo ago
It was enough of a concern to cause a worldwide manhunt for the guy responsible who is now in prison.
813.
by wmf 8mo ago
Every financial company is known for that.
814.
by wmf 8mo ago
I don't discuss politics with AI so this isn't relevant to me.
815.
by wmf 8mo ago
It sounds like Claude's known liberal bias isn't the issue.
816.
by wmf 8mo ago
Human science being replaced by AI I guess.
817.
by wmf 8mo ago
I assume hep = high energy physics in this context. PI = professor who received a government grant. Peer review has never really been blind and I suspect PIs will reject papers from "outsiders" even if they are higher quality. Thi
818.
by wmf 8mo ago
There's nothing I want less than multi-frame generation. I guess some people want to feel like they're getting their money's worth from their 240 Hz monitors.
819.
by wmf 8mo ago
OFAC already sanctions crypto wallets. https://ofac.treasury.gov/faqs/594
820.
by wmf 8mo ago
Previous discussion of the text version: https://news.ycombinator.com/item?id=47042136
821.
by wmf 8mo ago
It will always be more expensive.
822.
by wmf 8mo ago
It only looks that way because Llama failed. Good models like Qwen are shipping every 6 months.
823.
by wmf 8mo ago
Taalas is more expensive than NPUs not less. You have GPU/NPU at home; just use it.
824.
by wmf 8mo ago
https://developer.apple.com/documentation/FoundationModels
825.
by wmf 8mo ago
No, Taalas requires more silicon which will always cost more than storing weights in DRAM.
826.
by wmf 8mo ago
That's the part that people are missing: it won't get smaller. It already required heroic optimization to get 8B on one megachip. Taalas is more expensive but faster. It is cheaper per token when running 24x7 but not cheap to buy.
827.
by wmf 8mo ago
Waymo already has on-board NPU(s) with Transformer model(s) that are cheaper than Taalas.
828.
by wmf 8mo ago
The Taalas approach is much more expensive than the NPU that phones already have.
829.
by wmf 8mo ago
Yeah, FPGA+HBM works but it has no advantage over GPU+HBM. If you want to store weights in FPGA LUTs/SRAM for insane speed you're going to need a lot of FPGAs because each one has very little capacity.
830.
by wmf 8mo ago
Arguably DRAM-based GPUs/TPUs are quite inefficient for inference compared to SRAM-based Groq/Cerebras. GPUs are highly optimized but they still lose to different architectures that are better suited for inference.
831.
by wmf 8mo ago
You still need to do a forward pass per token. With massive batching and full pipelining you might be able to break the dependencies and output one token per cycle but clearly they aren't doing that.
832.
by wmf 8mo ago
FPGAs have really low density so that would be ridiculously inefficient, probably requiring ~100 FPGAs to load the model. You'd be better off with Groq.
833.
by wmf 8mo ago
I think only disk manufacturers can answer that but they're not willing to. Lost or "high-flying" writes have been documented for 20 years or so; we don't need to know why they occur to admit that they are real.
834.
by wmf 8mo ago
The same happened with the Internet/Web; the hype came true 10-15 years later than predicted.
835.
by wmf 8mo ago
People sneered at dotcoms but they weren't afraid of them. People are afraid of AI. Maybe they shouldn't be, but they are. The claims of "adopt Internet/AI or be left behind" were similar but for some reason the rea
836.
by wmf 8mo ago
Just don't tell the users anything about it.
837.
by wmf 8mo ago
I assume the onboarding steers everyone to one PDS provider and the mobile apps only use one appview server.
838.
by wmf 8mo ago
It's really the defaults that need to be fixed and anyone cannot change those.
839.
by wmf 8mo ago
Chipsets don't determine the RAM type and all motherboards have been made in China for a while.
840.
by wmf 8mo ago
No. Maybe ask whoever did your incorporation and nonprofit application.

More ›