It really is quite simple. gpt-3.5-turbo is designed and built as "conversational" AI, whereas, GPT-4 has been guard-railed in order to provide highly detailed responses to very specific and nuanced prompts that have complex structure and detailed context, character, style and format output, aka "fact gathering" ---› Think of GPT-4 as your, not-so nutty professor of everything, that's NOT CONVERSATIONAL, but rather a giver of facts. In other words, you must engineer prompts with that in-mind ---› GPT-4 will reward you for it. Just don't expect GPT-4 to be as "chatty" as gpt-3.5-turbo, it's not made that way. This also explains why GPT-4 has a bigger token quota (it needs it for more complex tasks and lengthier output), whereas gpt-3.5-turbo is brilliant at creating content and generating ideas, just make sure your prompts are "on-point" and "on-topic" (for both models).
actually I am pretty sure they are RLHF'd in a similar way. And have the same sysprompt. GPT--4 is just a smarter LLM, which causes its responses to be beyond the human's understanding sometimes. You can probably get gpt-4 to that personality with the right prompt.
@@DocUK123 GPT-4 has not been provided with all the information about itself by OpenAI, so it's just making that up. Its token limit is indeed higher. (8k or 32k depending on the version)
in brief: In this video, the speaker discusses some use cases where GPT-3.5, an older version of the language model, performs better than the more advanced GPT-4. The speaker highlights three specific prompts where GPT-3.5 outperforms GPT-4: 1. Simulate job interviews: GPT-3.5 provides more human-like responses and more concise answers compared to GPT-4. GPT-4's answers are more rigid, fact-based, and lack the human touch. 2. Uncover facts: GPT-3.5 is better at providing interesting, lesser-known facts on various topics without going into too much detail. GPT-4, on the other hand, often provides more detailed and specific information, which might not be as engaging or useful in casual conversations. 3. Most important quality in a given context: The speaker prefers GPT-3.5 for generating opinions on the most important quality a person can possess in specific situations. GPT-3.5's responses tend to be more practical and easier to understand compared to GPT-4. In the context of working with adults or running a startup, the most important quality for a leader is adaptability. When prompted for its opinion, GPT-3.5 confidently states that adaptability is essential for success, as it allows leaders to navigate the unpredictable process of starting a business, pivot when necessary, and be open to feedback and new ideas. Building a team that shares this quality is also crucial. In comparison, GPT-4 is less confident in its response, stating that adaptability is often considered one of the most important qualities a leader can possess. It mentions other key qualities such as vision, decisiveness, and effective communication but does not provide the same level of detail and subjective take as GPT-3.5. When looking for an opinion or subjective perspective, GPT-3.5 is preferable. If you're interested in learning more about prompt engineering to craft better prompts and get consistent results, consider exploring a course on the topic or watching advanced prompt engineering videos on RUclips. The speaker emphasizes that GPT-4 is still the superior model in most cases, but it's important to recognize that older versions like GPT-3.5 can still provide value in certain situations.
Great video! As a web developper in some cases I use GPT3.5, GPT4 sometime gives a really complicated answer when there is a much easier and faster way to resolve the problem.
Regardless of the output quality GPT-4 is not a feasible option for businesses as the API call is way to expensive like $0.25/click+ expensive whereas 3.5 is like $0.001-$0.003/click depending on the system prompt length, and the max tokens per conversation loop used. It’s honestly not even a question.
The implications of this are really interesting. I think that this shows the kind of information that the models were trained on. I'm no AI expert, but I know the importance of quality training data. Like you said, GPT-4 gives less "human" answers than GPT-3.5. I think this is because it may have been trained on more data that has been curated or perfected by the author, like a scientific paper, to be as precise as possible rather than written like humans normally speak. This us unlike GPT-3.5 which I believe has been trained on a more broad corpus of the internet where people type things out more like how they would speak. I think this translates into certian situations where certian triggers may trigger GPT-4 to be more precise and less "human" like how a scientific paper might be written. Though I am just speculating, I don't know anything for certian.
00:04 ChatGPT 3.5 is better than GPT-4 in some use cases. 01:59 GPT-4 has higher guardrails and is less human in simulating job interviews. 03:56 GPT-4 has more rigid and fact-based answers compared to GPT 3.5 05:49 Ancient Chinese beer recipe discovered in 2016 07:41 GPT-4 provides more detailed and specific responses compared to ChatGPT 3.5 09:38 GPT 3.5 is preferred for more unconstrained results and practical knowledge. 11:19 Adaptability is the essential quality for starting a business. 13:11 GPT-4 has a more opinionated and personalized approach compared to ChatGPT 3.5.
For someone selling a course on ai prompts and then complaining about the results I’m laughing my a*** off. You literally get the answers you ask for. If you want it concise you tell it and it will deliver. It’s very simple bob 😂
I don't believe GPT-4 has more guardrails. It's just better at following the ones it does have (which can lead to it doing things you don't want) In fact, if you ask it to reply about an opinion in one word, GPT-4 is more likely to give an answer, while 3.5 usually says "Subjective."
Yeah, I also find something similar to be true for Midjourney. Sure, ver. 5 is great for things like realism, but I have better luck with 4 for some art styles. Go figure.
Is it possible that you would set your course up for several monthly payments to be drawn automatically from our account? It’s less painful than a large one time payment. I just bought your book… I can’t wait to use it. Perhaps a book on writing python code and instructions on how to implement it. I love your channel 👍
Here is my disappointing ChatGPT. 3.5 experience with character string manipulation. Can someone please try this with GPT4 and let us know if it works. User: Swap the first and last characters of the given string while preserving the capitalization of the moved character. For example, if the first character is uppercase and the last character is lowercase, the final output should have the last character uppercase and the first character lowercase. Please provide the modified string as the output. cFrn5hUeT ChatGPT: Sure, here's the modified string: eFrn5hUcT User: why did you start with 'e' ? ChatGPT: Apologies, that was an error on my part. Here's the correct modified string, with the first and last characters swapped while preserving capitalization: CFrn5hUec User: Incorrect. The answer should be: TFrn5hUec ChatGPT: I apologize for my mistake. You are correct, the answer should be "TFrn5hUec". Thank you for pointing that out.
and when i ask can you explain how you did that? Sure! Here’s how I did it: I took the first and last characters of the given string cFrn5hUeT which are c and T respectively. I checked the capitalization of these characters and found that c is lowercase and T is uppercase. I swapped the positions of these characters while preserving their capitalization. So, c became uppercase and T became lowercase. I replaced the first character of the given string with the modified last character and the last character with the modified first character to get the final output: Tfrn5huEc. I hope that helps! 😊
Very good comparison, but why talk so fast and make such quick cuts? This is not TikTok! Otherwise very informative, but this "rushing" is a bit irritating. At least for me... please, don't get me wrong, the whole world is sped up, but you lose empathy... and it all becomes so neutral. I'm sorry, I don't mean to offend you, but it's kind of impersonal for me. The content is so good and informative, but too fast... you know what I mean... Maybe I'm just getting old 🙂
It really is quite simple. gpt-3.5-turbo is designed and built as "conversational" AI, whereas, GPT-4 has been guard-railed in order to provide highly detailed responses to very specific and nuanced prompts that have complex structure and detailed context, character, style and format output, aka "fact gathering" ---› Think of GPT-4 as your, not-so nutty professor of everything, that's NOT CONVERSATIONAL, but rather a giver of facts. In other words, you must engineer prompts with that in-mind ---› GPT-4 will reward you for it. Just don't expect GPT-4 to be as "chatty" as gpt-3.5-turbo, it's not made that way. This also explains why GPT-4 has a bigger token quota (it needs it for more complex tasks and lengthier output), whereas gpt-3.5-turbo is brilliant at creating content and generating ideas, just make sure your prompts are "on-point" and "on-topic" (for both models).
actually I am pretty sure they are RLHF'd in a similar way. And have the same sysprompt. GPT--4 is just a smarter LLM, which causes its responses to be beyond the human's understanding sometimes. You can probably get gpt-4 to that personality with the right prompt.
GPT4 says it had the same token limit as previous versions, which is also echoed by various sources online; only about 4,600.
@@DocUK123 GPT-4 has not been provided with all the information about itself by OpenAI, so it's just making that up. Its token limit is indeed higher. (8k or 32k depending on the version)
in brief:
In this video, the speaker discusses some use cases where GPT-3.5, an older version of the language model, performs better than the more advanced GPT-4. The speaker highlights three specific prompts where GPT-3.5 outperforms GPT-4:
1. Simulate job interviews: GPT-3.5 provides more human-like responses and more concise answers compared to GPT-4.
GPT-4's answers are more rigid, fact-based, and lack the human touch.
2. Uncover facts: GPT-3.5 is better at providing interesting, lesser-known facts on various topics without going into too much detail.
GPT-4, on the other hand, often provides more detailed and specific information, which might not be as engaging or useful in casual conversations.
3. Most important quality in a given context: The speaker prefers GPT-3.5 for generating opinions on the most important quality a person can possess in specific situations.
GPT-3.5's responses tend to be more practical and easier to understand compared to GPT-4.
In the context of working with adults or running a startup, the most important quality for a leader is adaptability.
When prompted for its opinion, GPT-3.5 confidently states that adaptability is essential for success, as it allows leaders to navigate the unpredictable process of starting a business, pivot when necessary, and be open to feedback and new ideas. Building a team that shares this quality is also crucial.
In comparison, GPT-4 is less confident in its response, stating that adaptability is often considered one of the most important qualities a leader can possess.
It mentions other key qualities such as vision, decisiveness, and effective communication but does not provide the same level of detail and subjective take as GPT-3.5.
When looking for an opinion or subjective perspective, GPT-3.5 is preferable.
If you're interested in learning more about prompt engineering to craft better prompts and get consistent results, consider exploring a course on the topic or watching advanced prompt engineering videos on RUclips.
The speaker emphasizes that GPT-4 is still the superior model in most cases, but it's important to recognize that older versions like GPT-3.5 can still provide value in certain situations.
Spoken like a true machine!!!.... just not sure which version
Great video! As a web developper in some cases I use GPT3.5, GPT4 sometime gives a really complicated answer when there is a much easier and faster way to resolve the problem.
Regardless of the output quality GPT-4 is not a feasible option for businesses as the API call is way to expensive like $0.25/click+ expensive whereas 3.5 is like $0.001-$0.003/click depending on the system prompt length, and the max tokens per conversation loop used. It’s honestly not even a question.
The implications of this are really interesting. I think that this shows the kind of information that the models were trained on. I'm no AI expert, but I know the importance of quality training data. Like you said, GPT-4 gives less "human" answers than GPT-3.5. I think this is because it may have been trained on more data that has been curated or perfected by the author, like a scientific paper, to be as precise as possible rather than written like humans normally speak. This us unlike GPT-3.5 which I believe has been trained on a more broad corpus of the internet where people type things out more like how they would speak. I think this translates into certian situations where certian triggers may trigger GPT-4 to be more precise and less "human" like how a scientific paper might be written. Though I am just speculating, I don't know anything for certian.
Your speculation is correct. The latest model hasn't been "trained" as much as the earlier version. Give it more, time, and 3.5 will be obsolete.
00:04 ChatGPT 3.5 is better than GPT-4 in some use cases.
01:59 GPT-4 has higher guardrails and is less human in simulating job interviews.
03:56 GPT-4 has more rigid and fact-based answers compared to GPT 3.5
05:49 Ancient Chinese beer recipe discovered in 2016
07:41 GPT-4 provides more detailed and specific responses compared to ChatGPT 3.5
09:38 GPT 3.5 is preferred for more unconstrained results and practical knowledge.
11:19 Adaptability is the essential quality for starting a business.
13:11 GPT-4 has a more opinionated and personalized approach compared to ChatGPT 3.5.
For someone selling a course on ai prompts and then complaining about the results I’m laughing my a*** off. You literally get the answers you ask for. If you want it concise you tell it and it will deliver. It’s very simple bob 😂
Great ChatGPT start for me actually. I'm a newcomer to this AI industry. Thank you
So much better without the background music audio track! Thank you! Now I can focus entirely on your verbal content, which is very informative.
I don't believe GPT-4 has more guardrails. It's just better at following the ones it does have (which can lead to it doing things you don't want) In fact, if you ask it to reply about an opinion in one word, GPT-4 is more likely to give an answer, while 3.5 usually says "Subjective."
My friend what do you think about Bing?
I need one I.A to help me to solve a lot of Chemistry problems of University
Yeah, I also find something similar to be true for Midjourney. Sure, ver. 5 is great for things like realism, but I have better luck with 4 for some art styles. Go figure.
This is excellent, good work 👍
Currently taking statistics. With gpt 3.5 I get 80s, gpt4 I get 90s
😄 Job's Tears... that's how I imagined them too.
If using chatgpt to look for a job you have already lost war in understanding what it does.😮
Is it possible that you would set your course up for several monthly payments to be drawn automatically from our account? It’s less painful than a large one time payment. I just bought your book… I can’t wait to use it.
Perhaps a book on writing python code and instructions on how to implement it.
I love your channel 👍
Here is my disappointing ChatGPT. 3.5 experience with character string manipulation. Can someone please try this with GPT4 and let us know if it works.
User: Swap the first and last characters of the given string while preserving the capitalization of the moved character. For example, if the first character is uppercase and the last character is lowercase, the final output should have the last character uppercase and the first character lowercase. Please provide the modified string as the output.
cFrn5hUeT
ChatGPT: Sure, here's the modified string: eFrn5hUcT
User: why did you start with 'e' ?
ChatGPT: Apologies, that was an error on my part. Here's the correct modified string, with the first and last characters swapped while preserving capitalization: CFrn5hUec
User: Incorrect. The answer should be: TFrn5hUec
ChatGPT: I apologize for my mistake. You are correct, the answer should be "TFrn5hUec". Thank you for pointing that out.
i use bing and the answer is
Sure! Here’s the modified string: Tfrn5huEc.
and when i ask can you explain how you did that?
Sure! Here’s how I did it:
I took the first and last characters of the given string cFrn5hUeT which are c and T respectively.
I checked the capitalization of these characters and found that c is lowercase and T is uppercase.
I swapped the positions of these characters while preserving their capitalization. So, c became uppercase and T became lowercase.
I replaced the first character of the given string with the modified last character and the last character with the modified first character to get the final output: Tfrn5huEc.
I hope that helps! 😊
and i use Merlin GPT-4 the answer is : Tfrn5hUeC, still wrong but at least gpt4 based got the first and last character swap correct
I'm using chatgpt is the other versions free 2,3,4
Great job
great *gob
Use playground gpt4 is more better than both .. even 3.5 playground is way more better
Very good comparison, but why talk so fast and make such quick cuts? This is not TikTok!
Otherwise very informative, but this "rushing" is a bit irritating. At least for me... please, don't get me wrong, the whole world is sped up, but you lose empathy... and it all becomes so neutral.
I'm sorry, I don't mean to offend you, but it's kind of impersonal for me.
The content is so good and informative, but too fast... you know what I mean... Maybe I'm just getting old 🙂