Glad it was helpful! Trying to make each video better and better… my goal is to help busy tech professionals get up on all of this exciting technology…
Glad it was helpful!.... Trying to get better with each video... If you have areas you would like to see me cover please feel free to share... thank you...
This channel is the best discovery recently in terms of ai! Thanks for the content. One question though: What if top-p would consider the first three words, but top-k is set to 5? Will top-p win or top-k?
Thanks for your feedback... I really appreciate it .... Reguarding your question.... My research has found that, in most LLM's, Top-K goes first, filters down the candidate Pool, and then Top-P is applied to that Top-K modified candidate pool... but keep in mind, this is an LLM implementation detail, and it might be that there are some now, or in the future, that reverse this ...
Excellent content and educational. I appreciate putting such efforts into making it simple for others to understand the concepts. A small suggestion, if you don’t mind - the background sound is dominating your voice; it's better to lower the BGM sound a lot or remove it altogether.
I really appreciate your feedback... yes, my son has indicated the BGM is a little to loud and your feedback confirms his thoughts as well.. I appreciate positive feedback but also helpful critical feedback in the spirit of getting better... thank you sir... let me know if there are AI/LLM topics you would like to me cover in the near future.. thank you
Thanks for making this video. Gave very good clarity. I have one doubt, How does model picks up the final token to return. Even if these different techniques are changing the size of candidate pool, token with maximum probability should always stay at the top. If so, how these parameters impact the final output?
Good question.... So, the tokens, in the "Candidate Pool", are assigned a statistical probability. For example the first token might be 0.50 or 50%. The second might be 0.25 or 25%, etc.... These statistical probabilities come from the pre-training phase, when LLM trained on raw data. When the next word is selected the first word, in this example, will be selected 50% of the time. The second word will be select %25 of time, etc... When Temperature is 0 you get special case where, where all token probabilities go to zero, except the highest probability token, which is always selected. This gives deterministic output. When Temperature goes up, the probabilities are adjusted to give more tokens, with smaller probabilities, in the candidate pool, higher probabilities which give those tokens more chances to get selected.. Is this explanation helpful ?
Would it be possible to remove (shitty) music from the background? to "really help your channel grow", as you say. The music is not absolutely appropriate
Thanks for the feedback … we will have to agree to disagree on this one .. I do like the music so I will be keeping it .. but on another note … do you have any ideas for videos I should do? What dev environments do you use? Do you use notebooks? Do you use python or JavaScript? Are you a cloud developer? Let me know so I can try to create videos to environments and tools you use …
Thank you for explaining this simply! Came here after being confused in a course I was taking.
Great to hear! Let me know if there are other LLM topics you would like to see a video on…. 🙏
Top-P, Top-K, Top Explanation 🙂 Thanks much!
Glad you liked it! Trying to get better with each video ....
Beautifully explained
Thank you so much 🙂. Thanks for feedback…. Trying to get better with each video…. 🙏
Very helpful and clear. Thank you.
Glad it was helpful! Hyper focused on clear direct videos that break topics down in a way that easy understand … appreciate your feedback!
¡Excelente explicación! Muchas gracias.
gracias señor... agradezco sus comentarios…🙏
Thanks man! Superb teaching
Glad it was helpful! Trying to make each video better and better… my goal is to help busy tech professionals get up on all of this exciting technology…
Awesome explanation !!
Glad you liked it! Trying to get that much clarity in a concise way on every video……. Thanks for feedback.
A really helpful video, thanks.
Glad it was helpful!.... Trying to get better with each video... If you have areas you would like to see me cover please feel free to share... thank you...
This channel is the best discovery recently in terms of ai! Thanks for the content. One question though: What if top-p would consider the first three words, but top-k is set to 5? Will top-p win or top-k?
Thanks for your feedback... I really appreciate it .... Reguarding your question.... My research has found that, in most LLM's, Top-K goes first, filters down the candidate Pool, and then Top-P is applied to that Top-K modified candidate pool... but keep in mind, this is an LLM implementation detail, and it might be that there are some now, or in the future, that reverse this ...
clear and concise...thanks very much.
Thanks for watching! Working hard to create direct clear and concise videos ….. appreciate your feedback !
Good stuff here. Keep on!
Appreciate it! Trying to get better and better with each video… thanks for the feedback…
Excellent content and educational. I appreciate putting such efforts into making it simple for others to understand the concepts. A small suggestion, if you don’t mind - the background sound is dominating your voice; it's better to lower the BGM sound a lot or remove it altogether.
I really appreciate your feedback... yes, my son has indicated the BGM is a little to loud and your feedback confirms his thoughts as well.. I appreciate positive feedback but also helpful critical feedback in the spirit of getting better... thank you sir... let me know if there are AI/LLM topics you would like to me cover in the near future.. thank you
Great video!
Glad you enjoyed it. Thank you for taking time to provide feedback….
Thanks for making this video. Gave very good clarity.
I have one doubt, How does model picks up the final token to return. Even if these different techniques are changing the size of candidate pool, token with maximum probability should always stay at the top. If so, how these parameters impact the final output?
Good question.... So, the tokens, in the "Candidate Pool", are assigned a statistical probability. For example the first token might be 0.50 or 50%. The second might be 0.25 or 25%, etc.... These statistical probabilities come from the pre-training phase, when LLM trained on raw data. When the next word is selected the first word, in this example, will be selected 50% of the time. The second word will be select %25 of time, etc... When Temperature is 0 you get special case where, where all token probabilities go to zero, except the highest probability token, which is always selected. This gives deterministic output. When Temperature goes up, the probabilities are adjusted to give more tokens, with smaller probabilities, in the candidate pool, higher probabilities which give those tokens more chances to get selected.. Is this explanation helpful ?
Great content.
Thank you …. Let me know if there are AI / LLM topics you are interested in…
thanks for sharing, can you make a video on GANS?
Thanks for asking ... yes give me a few weeks to research and I will get something out for you...
the music is a bit annoying. it should be a bit quieter.
otherwise good video. thanks
Thanks for feedback…. 🙏
Would it be possible to remove (shitty) music from the background? to "really help your channel grow", as you say. The music is not absolutely appropriate
Thanks for the feedback … we will have to agree to disagree on this one .. I do like the music so I will be keeping it .. but on another note … do you have any ideas for videos I should do? What dev environments do you use? Do you use notebooks? Do you use python or JavaScript? Are you a cloud developer? Let me know so I can try to create videos to environments and tools you use …
@@NewMachina Thanks. I wont be watching videos with distracting music in the background. Therefore I dont have suggestions for you.
Ok, I understand...