sabareesh

7 ms·

31.
by sabareesh 1y ago
It might be that our current tokenization is inefficient compared to how well image pipeline does. Language already does lot of compression but there might be even better way to represent it in latent space
32.
by sabareesh 1y ago
Similar feeling. Seems it is good at certain things and if something doesnt work it want to do things simply and in turn becomes something that you didnt ask for and certain times opposite of what you wanted. On the other hand with codex ce
33.
by sabareesh 1y ago
"Simple, Boring Tech Stack:" Good advice but bad example, because it depends on what engineers are familiar and comfortable with and technology itself should be mature enough. You dont want to spend time building orchestrator when
34.
by sabareesh 1y ago
This is great,but most work is involved in curating the dataset and the objective functions for RL.
35.
by sabareesh 1y ago
Looks very similar to DGX spark
36.
by sabareesh 1y ago
Here is my rig, running GLM 4.5 Air. Very impressed by this model https://sabareesh.com/posts/llm-rig/ https://huggingface.co/zai-org/GLM-4.5
37.
by sabareesh 1y ago
That was confusing part of this video . May be there are some limitation on the tools he uses to tune
38.
by sabareesh 1y ago
Where can i find the specs. I am actively working on some project with robot arm and found following appealing eventhough this doesnt include servo or cameras or controllers. https://www.aliexpress.us/item/3256808789646
39.
by sabareesh 1y ago
Good move, not sure they are exposing other modalities as well ?
40.
by sabareesh 1y ago
Interesting one thing that is more satisfying with robotics is that because you can see your creations in real world
41.
by sabareesh 1y ago
Learning, prototype and then scale it in to cloud. Also can be used as inference engine to train another model if you are using model as a judge for RL.
42.
by sabareesh 1y ago
This is what i have https://sabareesh.com/posts/llm-rig/ All You Need is 4x 4090 GPUs to Train Your Own Model
43.
by sabareesh 2y ago
Raw models may not be good enough. I wonder how thinking models do on these
44.
by sabareesh 2y ago
Hate to say it but seems Safari might be the alternative. Only missing piece is ublock origin
45.
Homemade Silicon Chips [video](youtube.com)
3 points by sabareesh 2y ago|1 comments
46.
by sabareesh 2y ago
I have seen several videos on this but seeing an actual demo on a homemade setup really helps understanding how it is actually done
47.
by sabareesh 2y ago
Seems they lost the ship , it is supposed to be v2 and had several changes
48.
by sabareesh 2y ago
I would rather go other direction. When introducing new chemical or additive it has to be proven that is not harmful to humans before using it
49.
by sabareesh 2y ago
I am pretty sure memory bandwidth will be low it doesn't eat up their enterprise lineup. If we are luck we might get 512GB/S this is still half of 4090
50.
by sabareesh 2y ago
https://hyperbolic.xyz/pricing
51.
by sabareesh 2y ago
https://news.ycombinator.com/item?id=42614450
52.
by sabareesh 2y ago
what are current h100 price ? lowest i have seen is only 0.99 per hour
53.
by sabareesh 2y ago
That didnt take long after AMD announcment
54.
by sabareesh 2y ago
take my money
55.
by sabareesh 2y ago
About 4090 it is bad metric because that given model wont fit on 4090 Grrr Nvidia.
56.
by sabareesh 2y ago
"According to Jacobson, California electricity prices are high because of several reasons that have nothing to do with renewables. These include high fossil gas prices, utilities passing on to customers the cost of wildfires due to tra
57.
by sabareesh 2y ago
It might limit innovation on what is after usb-c . We might get stuck with usb-c for a while.
58.
by sabareesh 2y ago
He wanted to expose ads to AI, which felt weird for some reason.
59.
by sabareesh 2y ago
would love english version of the site
60.
Apple to pay $95M to settle lawsuit accusing Siri of snoopy eavesdropping(finance.yahoo.com)
4 points by sabareesh 2y ago|0 comments

More ›