One of the BEST tutorials/explanations i have ever seen, thank you really much for that ❤

you’re cute

This is a very valuable video for me. Thank you so much

Lost it on "Watch the full video in the link below"

What exactly happened here?🤔

I love it; most unawkward facet of you yet😊

Wow I didn’t know I can enjoy “awkward” silence 😮

Waiting for the phd result...?

Some would upload this as ASMR

Haha a classic when recording!

I like the part where she doesn't say anything.

good concise explanation.

Damn you like Neffex ❤ Neffex is like 10 % of my life.

Thank you Miss Coffee Bean !

Cool! The simple but effective ideas are the most interesting. Have you checked out Nvidia's normalized transformer (nGPT)? It seems to be one of those cases. The differential transformer (arxiv: 2410.05258) also seems like it could be interesting.

Terrible

03:23 It looks like the MLP takes the representation of JUST the 8th layer, and not also of the previous ones. I just found out from a quick exchange via email with the authors.

CONGRATULATIONS Letitia !!!! Is there a pdf version where I can read your PhD thesis? Always amazed learning new knowledge from you :)

Unison's unique features offer promising possibilities for enhancing the FunSearch operational model: Content-addressed code: Unison's hash-based identification of code could streamline FunSearch's program generation and evaluation process. Each generated program would have a unique hash, simplifying tracking and caching of results. Immutable codebase: Unison's immutable data structure approach to codebases could enhance FunSearch's evolutionary process. It would allow for efficient storage and retrieval of program versions without conflicts. No builds: Unison's ability to parse and typecheck definitions once, storing results in a cache, could significantly speed up FunSearch's evaluation phase. This would reduce overhead in assessing generated programs. Easy distributed computing: Unison's content-addressed nature facilitates distributed computation, which could enhance FunSearch's parallelization capabilities. This could allow for more efficient scaling of the search process across multiple machines. First-class documentation: Unison's approach to documentation as executable code could improve FunSearch's ability to generate and evaluate self-documenting programs, potentially leading to more interpretable solutions. Strong typing with inference: Unison's type inference could help FunSearch generate type-safe programs more efficiently, potentially reducing the number of invalid programs generated. These features could potentially make Unison an excellent choice for implementing core components of the FunSearch system, particularly in program generation, evaluation, and distributed processing.

Hmm, the only thing I dont really understand is why we have to use SIN and COS instead of only one of them D:

Loved your explanation! thank you very much!! :D

Okay, the approach is “guessing”. Got it.

can I ask what app/software tools you use for recording, editing audio and video?

It would be definitely interesting to train end to end with contrastive lost. Because right now it is a more like a distillation task that is limited be the teacher model's capability & domain

"wait a few years"... 9 months later with Genie 2 😂

This was very helpful, thank you!

It's a clever idea, but the paper has many methodological issues which may amount to more hype than substance. Then results are good, but the major gains come from stacking contributions. And of course you have to worry about the extra training cost (DINO is very expensive) and generalization, etc. The peak accuracy shift is also worrisome, indicating that removing the driving force would likely result in an expensive shift / performance degradation when finetuning.

thats such a good idea. very simple .. i would not have thought to do this. I have added CLIP loss terms on the last layer but not dinoon the 8th layer lol.

i missed you, glad to see you back with your great explainers

i'm curious how they may scale this with additional external representations. How would one have to change the training approach for e.g. 3 sources? Could one simply co-optimise for all three losses + image-gen or would they interfere with each other? Can't wait to read the followup papers exploring such questions.

It's not a long term approach, autoregressive generative vision language models are the future.

Why did nobody think of this earlier!????

You are absolutely amazing thanks for this explanation. I read the paper and understood nothing!

This was really great, never really understood query, key and values before.

Congrats! You're inspiring

Great explanation -- loving your videos.The time codes for specific topics is really useful.

The begin was clear, then very obscure👎

Awesome content!

Terrible explanation

Hey khazeema saleem Al here

very concise; what an amazing video.

I disagree that people are being judged for being bad at math more harshly than for other subjects. Compare "I'm not interested in history" or "I don't like reading" with "I'm bad at math".

UMAP rocks! The only problem I see is the explainability of this high dimensionality reduction, which is easily done in PCA. In other words, you can get the best variables to explain the clustering, which is important when you are focusing on variable selection. What do you think?

You’re the best 👏👏👏

Thank you very much for this thorough, well-curated, and comprehensive review of MAMBA.

How are exploration and exploitation abilities or forms of intelligence? To me they're more like competing actions, like going to the shops vs going to the cinema. I am still capable of both actions.

sooo well explain, brilliant!

Another great video, as usual! This little bean mutant of yours always puts a smile on my face ☺ Is it possible that it is actually an AI? For example, a transformer that converts language information into the facial expressions of the animated bean. That would be so cool 😎 I have a question: I am looking for training methods that are not based on backpropagation. Specifically, I want to avoid running backwards through the NNW again after the forward pass. Do you know of any algorithms like this? Already 2^10 * Thanks in advance 😄

Thanks!