FanaHOVA

9 ms·

91.
by FanaHOVA 3y ago
It's just a wordplay on llama / lemma, so it's pronounced as a normal L I assume.
92.
by FanaHOVA 3y ago
MSFT earnings are in 8 days
93.
by FanaHOVA 3y ago
> I doubt their model is the best in the world For what reason do you doubt that?
94.
by FanaHOVA 3y ago
He just means promos by creators, not YT ads.
95.
by FanaHOVA 3y ago
Yea I tried to anchor too but it didn't work :( All transcriptions are created with https://github.com/FanaHOVA/smol-podcaster , we think most people want to skim and read before listening so we try and have high q
96.
RAG is a hack(latent.space)
42 points by FanaHOVA 3y ago|36 comments
97.
If AI is so important, why is its software so bad?(latent.space)
3 points by FanaHOVA 3y ago|2 comments
98.
by FanaHOVA 3y ago
It's really not that deep
99.
by FanaHOVA 3y ago
Outside of the pod, we spend some time breaking down the history of LangChain and how it went from a prompt formatting library to a "context-aware reasoning applications" framework. " "LangChain is Pointless" [0] sp
100.
Is LangChain Pointless?(latent.space)
18 points by FanaHOVA 3y ago|2 comments
101.
by FanaHOVA 3y ago
> do something other than dumb scaling up This is exactly what people told OpenAI 8 years ago and look where we are now.
102.
by FanaHOVA 3y ago
The fact that the GPUs quantity dropdown cannot go over 1,000 drives home the "GPU poor" point from the SemiAnalysis post. Meta alone has 16,000 GPUs. OpenAI's cluster from 2020 had 10,000 GPUs. If you're serious about f
103.
Smol-scheduler: automate scheduling emails with GPT-4(github.com)
1 points by FanaHOVA 3y ago|1 comments
104.
by FanaHOVA 3y ago
I have to schedule a lot of meeting for work, and don't really like using Calendly since I want more control over what times they pick in order to maximize long blocks of free time. smol-scheduler is a simple utility to draft schedulin
105.
by FanaHOVA 3y ago
Yep, +1. That's why I used the quotes. :) Thanks for expanding!
106.
by FanaHOVA 3y ago
It's 2 for forward pass, 4 for backward pass (2PD + 4PD = 6PD), but we didn't cover why it's 2 and 4 since it was a very deep rabbit hole. Maybe he'll do a separate post for that in the future!
107.
by FanaHOVA 3y ago
Training smaller models can be really compute intensive (a 7B models should get trained on 1.4T tokens to follow the "LLaMA laws"). So that would be C = 6 * 1.4T * 7B = 58.8T FLOP-seconds. That's 1/5th the compute of GPT
108.
by FanaHOVA 3y ago
Yes! I have a 5 pages doc of prep notes I made, happy to share :)
109.
The Mathematics of Training LLMs(latent.space)
199 points by FanaHOVA 3y ago|66 comments
110.
Running LLaMA-70B in the browser with MLC(latent.space)
3 points by FanaHOVA 3y ago|1 comments
111.
Show HN: Smol-podcaster, an AI intern for podcast hosts(github.com)
1 points by FanaHOVA 3y ago|0 comments
112.
by FanaHOVA 3y ago
> I've been adamantly against MoE for a set of reasons Such as?
113.
FlashAttention-2: 800% faster Transformers using SRAM(latent.space)
2 points by FanaHOVA 3y ago|0 comments
114.
by FanaHOVA 3y ago
HBM is getting more expensive AND we've hit the SRAM density scaling limits AND current GPU lines shipments delays are expected to last until Q2 of next year so no price relief anytime soon. Ouch.
115.
by FanaHOVA 3y ago
Mods changed the title, I used the original one when first posting. Not sure why they changed it.
116.
by FanaHOVA 3y ago
Author here. I agree with you. LLaMA2 isn't open source (as my title says, the HN one was modified). My point is that the average person will still call it "open source" because they don't know any better, and it's
117.
by FanaHOVA 3y ago
Agreed. I called it a "$3M of FLOPS donation" by Meta.
118.
by FanaHOVA 3y ago
Yea I didn't include it, but that'd be the "free as in beer, but not freedom" circle :)
119.
by FanaHOVA 3y ago
Yep, same with SSPL. GPL has been tested in FSF vs Cisco (2008), but none of the more restrictive licenses have.
120.
In the LLM space, "open source" is being used to mean "downloadable weights"(alessiofanelli.com)
400 points by FanaHOVA 3y ago|228 comments

More ›