jerpint

33 ms·

421.
by jerpint 4y ago
ChatGPT is fine tuned on a curated, relatively small dataset, which is of very high value and would mitigate some of the attacks
422.
by jerpint 4y ago
It isn’t revolutionary but it provides value
423.
by jerpint 4y ago
The PETALS project is a step in the right direction. A network of GPUs allows you to do inference (and even retrieve activations!) for the BLOOM model, a model so big it won’t fit on most regular hardware
424.
by jerpint 4y ago
Do you think it would ever be possible to “maximize” a neuron with certain sentences? What’s so different with the gradient ascent techniques with convolutions?
425.
by jerpint 4y ago
Price of implementing GPT? 300$ Price of training GPT? 30k$
426.
by jerpint 4y ago
It’s been weeks we’ve all played with chatGPT. everyone knows just how limited it is, ESPECIALLY at being factual. Microsoft going all-in and rebranding it as the future of search might just be the biggest blunder in recent tech history
427.
by jerpint 4y ago
It just highlights even more the need to understand these systems and test them rigorously before deploying them en-masse
428.
by jerpint 4y ago
I think understanding these jailbreaks give us a much deeper insight into the actual “comprehension” of these models. From a research perspective, this is fascinating
429.
by jerpint 4y ago
I wonder when we will have LLMs with custom personalities, chatGPT and Bing both sound so boring
430.
by jerpint 4y ago
HBO hit series material
431.
by jerpint 4y ago
I really like that they use GPT-J for this - significantly less weights than GPT and open source!
432.
by jerpint 4y ago
Typehints are a great way to document code. If the type hints get too complex they can be cumbersome though. For example passing an int or None if the value doesn’ exist
433.
by jerpint 4y ago
It’s one token if you add a space before: ‘ Sydney’
434.
by jerpint 4y ago
Sadly by the time enough people can manage to replicate this it’ll already be patched
435.
by jerpint 4y ago
I’m not sure I under exactly what this is? Does it mean chatGPT has an exposed endpoint and we can ping it?
436.
by jerpint 4y ago
Imagenet was a breakthrough in that they were a large curated dataset which was needed for supervised learning, but modern transformers rely on self-supervision
437.
by jerpint 4y ago
What about Neural Machine Translation? They were the first to propose the attention mechanism
438.
by jerpint 4y ago
This is a great point. For years, Amazon had good quality and prices. Now, shopping on Amazon you have to actively look out for scammy items and search for the good deals. Very easy to get fooled. “Verified purchases” don’t mean much either
439.
by jerpint 4y ago
Can’t wait to go play iceworld G-g-g-g-g-god like
440.
by jerpint 4y ago
A GPT whisperer
441.
by jerpint 4y ago
Sounds like it, pretty smart idea
442.
by jerpint 4y ago
SOS would’ve been a funny prank
443.
by jerpint 4y ago
This wouldn’t work in practice because you don’t have access to GPTs activations
444.
by jerpint 4y ago
Source? Genuinely curious
445.
by jerpint 4y ago
There are bit caveats here: Vision transformers outperform CNNs in HUGE data regimes. On small datasets, CNNs still shine. Also, if you take a CNN with modern tricks, they can be on par with vision transformers, e.g convnext Transformers re
446.
by jerpint 4y ago
GitHub issues has been a pretty good place to get answers too
447.
by jerpint 4y ago
Wow thanks for sharing that’s such a clever hack
448.
by jerpint 4y ago
This amount of diversity in transformers is very impressive, but what’s more impressive is that for models like GPT, scaling the models seems much more effective than engineering the models
449.
by jerpint 4y ago
I’ve been playing around with this and have gotten decent results simply using openAI embeddings. I’ll probably be making some kind of post showing results soon
450.
by jerpint 4y ago
Instead of saying the French, we should say “people of croissants”

More ›