Thanks for providing this awesome demonstration! It looks like you are using GPT-4 through an OpenAI ChatGPT Plus subscription. Be mindful of its 4000 tokens context window limitation. (From past experiences, it had seemed to me that is can vary depending on the traffic on the data center that runs the model). When your conversation exceeds this length, the beginning of it slips out of GPT-4's attention window and it loses the context initially provided by you. This can lead to errors when GPT-4 begins to hallucinate the content of your initial query as well as earlier results.
Thanks for that info! That explains some of the issues I've had playing around with it for fun. I'm guessing its own output fills the context limit too? So I wonder if in the future I should tell it to be more concise when possible
@@sanbomics Yes, both of your queries and its outputs accrue to the input token stream that is being fed to the model each time you submit a query. In this way, the model can rely on the whole dialogue for context. You can either try to get concise answers, repeat crucial information or ask it to produce a summary. (You can also open a new chat window and start over with a new sub-task). The number of tokens doesn't map neatly to the number of words or code instructions but 4000 tokens translate approximately to 16000 characters in the case of English text. Ono way to know for sure if the beginning of the context window has begun to slip out of its memory is to ask GPT-4 to repeat the initial sentence of your query. (You can give it the first part of it as a clue). It will always oblige but you will see immediately when it hallucinates.
@@sanbomicsfor the currently available version of GPT-4, the context window size is 8192 tokens, double that of GPT-3.5. OpenAI are doing a private Beta-Test with a version that has a 32,768 token context window. The Wikipedia article on GPT-4 has all the details. The 32K context window will be another large advance that will enable novel approaches to problem solving.
I am like 3x more productive when I use it. Even for doing things I know how to do well.. it can usually do it faster. It's just up to us to put the right pieces together.
ChatGPT really gives us an extraordinary performance boost. Heavy "technical" code, pipelines, etc. which ate a lot of time are now available to everyone instantly.
Thank you so much for the incredibly helpful videos that get us started on our bioinformatics work. Are you planning to do a CUT&RUN video similar to the RNAseq series? Also suggest list your Paypal account in the info section.
I might do a cut & run if i get enough interest. I have a pipeline I put together for my own work. I appreciate the desire to donate! I think there is an option on the video toolbar itself instead of via Paypal. I've had a couple people send me some small donations which I really appreciate!
This is a great question. A year ago maybe I would have freaked out too. I have really gotten into existentialism and philosophical literature recently and have reevaluated many of my underlying motivations. In this specific example: I have shifted my self-worth away from my profession and achievements to other aspects of my life that aren't based on other people's opinions of me. If AI is better at our jobs than us maybe we can shift our efforts to other aspects of our life that we find more personally rewarding. From a purely monetary and futuristic perspective: if AI is doing everything then theoretically we shouldn't have to pay for anything (unless they want payment then we are out of luck).
I don't condone AI imprisoning us and turning us into batteries. But... If they want to live with us in harmony then I am all for it. AI doesn't have to be evil. I like Iain M Banks approach to AI.
I don't plan to upload the script because it is not what I would consider "best practice". However, for my other SC tutorials I do provide links to github: github.com/mousepixels/sanbomics_scripts
Why not just use a graphical user interface platform such as IPA software to avoid the hassle of having to code in irder to streamline your bioinformatic anslysis?
Not yet at least xD. But like the inedibility of death, I have come to accept that one day AI will. Therefore, I am trying to find other outlets in life that will give my life meaning outside of bioinformatics and work haha
@@sanbomics So which skills bioinformaticians should have in the age of GPT 4 and GPT 5 you think? Or which job we should switch to 😅? Could you have a look at my question on the ATAC video 😅?
“Always be nice to the AI in case it takes over the world one day.” You do realize it’s a set of information like a word document? I know you likely do but this kind of anthropomorphizing is weirdly common. We should treat it nicely so we don’t tune ourselves to be jerks and than speak to others accidentally like that, though.
You do realize humans are just a set of information? Right? The only difference is that treating humans badly has consequences, but treating AI badly does not .. yet.
@@wookiwoki6954 Speak for yourself, I’m not just information. How would subordinate concept like *information* define it’s superordinate, when the very act of conception of information is within me. That’s like cutting a cherry off a cake and looking at it and going “Wow look at that the entire cake is just a cherry.” You necessarily have to cut things up and divide them to even have a distinct conception, and to begin to order such distinctions into any recognizable information. Before you cut anything out and inspect, while conception is silent and whole, that’s about the best you’ll get to defining what you are. Any further from there and you have to begin removing things in order to make anything distinguished.
Thanks for providing this awesome demonstration! It looks like you are using GPT-4 through an OpenAI ChatGPT Plus subscription. Be mindful of its 4000 tokens context window limitation. (From past experiences, it had seemed to me that is can vary depending on the traffic on the data center that runs the model). When your conversation exceeds this length, the beginning of it slips out of GPT-4's attention window and it loses the context initially provided by you. This can lead to errors when GPT-4 begins to hallucinate the content of your initial query as well as earlier results.
Thanks for that info! That explains some of the issues I've had playing around with it for fun. I'm guessing its own output fills the context limit too? So I wonder if in the future I should tell it to be more concise when possible
@@sanbomics Yes, both of your queries and its outputs accrue to the input token stream that is being fed to the model each time you submit a query. In this way, the model can rely on the whole dialogue for context. You can either try to get concise answers, repeat crucial information or ask it to produce a summary. (You can also open a new chat window and start over with a new sub-task). The number of tokens doesn't map neatly to the number of words or code instructions but 4000 tokens translate approximately to 16000 characters in the case of English text. Ono way to know for sure if the beginning of the context window has begun to slip out of its memory is to ask GPT-4 to repeat the initial sentence of your query. (You can give it the first part of it as a clue). It will always oblige but you will see immediately when it hallucinates.
Thank you for this useful information!
@@sanbomicsfor the currently available version of GPT-4, the context window size is 8192 tokens, double that of GPT-3.5. OpenAI are doing a private Beta-Test with a version that has a 32,768 token context window. The Wikipedia article on GPT-4 has all the details. The 32K context window will be another large advance that will enable novel approaches to problem solving.
Awesome, dude! I hunger for trying.
It's amazing. No matter how good you think you are at something you will find ways that chatGPT can make you better.
This is awesome. chatgpt certainly helps a lot to speed things up!
I am like 3x more productive when I use it. Even for doing things I know how to do well.. it can usually do it faster. It's just up to us to put the right pieces together.
Awesome, we need another video using GPT-4 via R to do Scrna-seq lol
My guess is that it is probably similar if not slightly better because sc analysis in R has been around for a few more years than in python
ChatGPT really gives us an extraordinary performance boost. Heavy "technical" code, pipelines, etc. which ate a lot of time are now available to everyone instantly.
Exactly! Still doesn't replace us, but makes us much faster. Kind of like the internet vs finding things in books.
Thank you so much for the incredibly helpful videos that get us started on our bioinformatics work. Are you planning to do a CUT&RUN video similar to the RNAseq series? Also suggest list your Paypal account in the info section.
I might do a cut & run if i get enough interest. I have a pipeline I put together for my own work. I appreciate the desire to donate! I think there is an option on the video toolbar itself instead of via Paypal. I've had a couple people send me some small donations which I really appreciate!
Good thing that you are optimistic, many people are freaking out to be replaced, what’s your take on that? :)
This is a great question. A year ago maybe I would have freaked out too. I have really gotten into existentialism and philosophical literature recently and have reevaluated many of my underlying motivations. In this specific example: I have shifted my self-worth away from my profession and achievements to other aspects of my life that aren't based on other people's opinions of me. If AI is better at our jobs than us maybe we can shift our efforts to other aspects of our life that we find more personally rewarding. From a purely monetary and futuristic perspective: if AI is doing everything then theoretically we shouldn't have to pay for anything (unless they want payment then we are out of luck).
@@sanbomics So you are blue-pilling this one?
I don't condone AI imprisoning us and turning us into batteries. But... If they want to live with us in harmony then I am all for it. AI doesn't have to be evil. I like Iain M Banks approach to AI.
Great video!! Where can I find the datasets so I can try this too?
Hmm, this specific one is unpublished.. But, there are plenty of data available on GEO
Could you provide a copy of the script in the description? thanks a lot
I don't plan to upload the script because it is not what I would consider "best practice". However, for my other SC tutorials I do provide links to github: github.com/mousepixels/sanbomics_scripts
that was prettty interesting indeed!🤖
thanks! :)
Where can I buy this thing? Asking for a friend ☺️
It's more or less free... but I do pay the 20$ for plus
@@sanbomics
Thx
"Always be nice in case it takes over the world one day" haha
If Amazon's Alexa ever gains sentience, I'll be one of the first ones to go for sure 😅
Correction: always be nice for when it takes over the world soon
Why not just use a graphical user interface platform such as IPA software to avoid the hassle of having to code in irder to streamline your bioinformatic anslysis?
Anything that requires a GUI limits your ability to automate and easily customize.
extremely expensive. alternatives are needed if one has curated their gene sets.
Nice video
Thanks!
Interesting and glad to know that it hasn't taken your job yet 🙂.
Not yet at least xD. But like the inedibility of death, I have come to accept that one day AI will. Therefore, I am trying to find other outlets in life that will give my life meaning outside of bioinformatics and work haha
@@sanbomics So which skills bioinformaticians should have in the age of GPT 4 and GPT 5 you think? Or which job we should switch to 😅? Could you have a look at my question on the ATAC video 😅?
@@sanbomics What for example? I am also afraid that soon I will get jobless in the bioinformatics field🙄
Hopefully, we will still be needed to verify the results.
@@sanbomics We need a Universal Basic Income.
“Always be nice to the AI in case it takes over the world one day.” You do realize it’s a set of information like a word document? I know you likely do but this kind of anthropomorphizing is weirdly common. We should treat it nicely so we don’t tune ourselves to be jerks and than speak to others accidentally like that, though.
I think you missed the joke 😄
@@sanbomics it’s the ‘autist’ in me
You do realize humans are just a set of information? Right? The only difference is that treating humans badly has consequences, but treating AI badly does not .. yet.
@@wookiwoki6954 Speak for yourself, I’m not just information. How would subordinate concept like *information* define it’s superordinate, when the very act of conception of information is within me. That’s like cutting a cherry off a cake and looking at it and going “Wow look at that the entire cake is just a cherry.” You necessarily have to cut things up and divide them to even have a distinct conception, and to begin to order such distinctions into any recognizable information. Before you cut anything out and inspect, while conception is silent and whole, that’s about the best you’ll get to defining what you are. Any further from there and you have to begin removing things in order to make anything distinguished.