jerpint

12 ms·

151.
by jerpint 2y ago
My openAI key was leaked and I noticed someone was using it, luckily the damage wasn’t nearly as bad as you. A few dollars worth of GPT4, a model none of my apps were using at the time. I’m almost entirely certain it was leaked via secrets
152.
by jerpint 2y ago
I noticed a few weeks ago that some of my OpenAI keys got compromised, they were only active as secrets on a huggingface space. I got an email a few days ago informing me that the spaces were compromised , so I suspect this issue has been g
153.
by jerpint 2y ago
Completely agree, and it’s been great to see to some extent that the gap between open source and 3rd party models is really starting to close
154.
by jerpint 2y ago
One point they don’t seem to spend much time on is also the difficulty in reproducing outputs in closed-source models. Setting temperature to 0 and setting seeds doesn’t always seem to be enough to get exactly the same results for a given p
155.
by jerpint 2y ago
Microsoft has the compute, so its likely to be an infrastructure problem
156.
by jerpint 2y ago
Have you tried CLIP image embeddings ?
157.
by jerpint 2y ago
I don’t think the app version is available
158.
by jerpint 2y ago
The imagenet results’ “main” contribution was showing that scaling the compute on large datasets worked surprisingly well. They didn’t contribute any new models as much as they showed that scaling is what helped these models learn and obtai
159.
by jerpint 2y ago
How does earth get seeded?
160.
by jerpint 2y ago
The videos look incredible, but a lot of the captions are riddled with grammar/syntax mistakes that seem odd for a model to make of that quality.
161.
by jerpint 2y ago
My understanding is that tokenization is part of the bottleneck. When you break up a sentence to tokens, each token gets a vector representation. The dictionary of all tokens would be infinite if it was at the sentence level
162.
by jerpint 2y ago
It seems to be a python app, so probably set it up as a seperate microservice with its own REST API
163.
by jerpint 2y ago
I had found that GPT4 couldn’t play wordle about a year ago [1]. At the time, I thought it must be because it wasn’t in the training data but now it seems to point to something larger. I might just get nerd sniped trying to teach it GoL now
164.
by jerpint 2y ago
The simplest way to think about it is a form of dropout but instead of dropping weights, you drop an entire path of the network
165.
by jerpint 2y ago
To cite results of an arxiv preprint as “proof” that none of these LLM technologies will ever work is so disingenuous. Of course these beasts are data inefficient, but to imply that this means all of the progress is smoke and mirrors is suc
166.
by jerpint 3y ago
You have to assume they share the same architecture for most of these methods
167.
by jerpint 3y ago
Rapes have been officially recognized as having happened by the UN, i don’t see how that is hard to believe given Hamas filmed most of their atrocities on GoPros. Oct 7 was a brutal terror campaign regardless of your position on the conflic
168.
by jerpint 3y ago
That’s pretty clever, encoding atomic concepts as a token
169.
by jerpint 3y ago
> With a bit of experimentation on Imagenet-1k, we can reach 82.0% accuracy with a 176x176 training image size with no extra data, matching ConvNeXt-T (v1, without pre-training a-la MAE) and surpassing ViT-S (specifically, the ViT flavor
170.
by jerpint 3y ago
Excited to see how it will perform on the lmsys leaderboard
171.
by jerpint 3y ago
Per the paper, 3072 H100s over the course of 3 months, assume a cost of 2$/GPU/hour That would be roughly 13.5M$ USD I’m guessing that at this scale and cost, this model is not competitive and their ambition is to scale to much la
172.
by jerpint 3y ago
Any news on how this model will compare to Mixtral? Interesting that they aren’t releasing a model with MoE this time given the success mixtral had
173.
by jerpint 3y ago
I work on chat-with-pdf and still find this awesome!
174.
by jerpint 3y ago
> An attacker only needs to read one keycard from the property to perform the attack against any door in the property That’s a pretty serious vulnerability, pretty much all it takes is to be a guest at a hotel
175.
by jerpint 3y ago
Until they can figure out how to scale other modalities to the expected amount of users , Incremental increases in Text models is likely what we will be seeing for the next little bit
176.
by jerpint 3y ago
There was a recent episode on The Latent Space podcast where they spoke with one of the Suno founders, I really enjoyed it https://open.spotify.com/episode/2c1yL8hlttlkCs6nPysVi0?si=O...
177.
by jerpint 3y ago
I didn’t interact with the locals much, but you are correct that it was mostly alpacas and llamas that you would see being herded
178.
by jerpint 3y ago
I’ve been to Peruvian mountains recently where they farm vicuñas and alpacas and can confirm they live in pretty poor conditions in a very harsh environment; high altitude barren mountains. The majority of income comes from tourism, and it’
179.
by jerpint 3y ago
Agreed, I have no idea what the plot at the end means and most outputs are just tensors
180.
by jerpint 3y ago
I’ve been having this hunch lately that using LVMs (GPT-4V, LLaVa, etc) could be a solution to error correcting, assuming the LVM has a sufficiently good enough world model

More ›