I would recommend testing whether GPT-3/4 most frequently ranks the first idea as the highest. In my testing it does. I would consider having it rank each idea individually and see if the results are the same.
That would make sense to me. In my experience, ChatGPT has a bias towards the initial things that are mentioned in the chat, similar to if you mentioned to your girlfriend something about her weight before going on to ask about her day. You've moved on, but the initial mode and discussion style that initiated the conversation has a way of lingering and guiding how the remaining discussion will unfold. I find myself opening a new chat if I feel a previous chat has gone a little off the rails, during some exploration for example.
This worked for me: Roleplay that you are a unique individual who is an expert with 30 years of concurrent experience in both the public and private spheres in (name 7 disciplines that are related to weight loss, healthy mind and overall wellness. 5 times rank the different disciplines in numerical order 1 to 10 with 10 being the best regarding consistent and sustainable weight loss. Each iteration use different unique techniques and innovative measures to rank their effectiveness regarding consistent and sustainable weight loss. After the 5th time average the scores from the 5 iterations and pick the top four disciplines. Those are the 4 you are uniquely skilled in). You are a published author in all of those disciplines and have won multiple prestigious awards in all 4 disciplines.
Asking ChatGPT to act as a psychologist and a relationship coach is basically what we are doing at Remble. 🙂 We are studying the best prompts and prompt chain of thought approaches for this specific scenario. It’s still pretty early, but I am definitely encouraged and excited by the potential of using this technology to actually help people who otherwise would never feel comfortable reaching out to a therapist for help.
I think that we shouldn't test the models with problems that are well known. If you now try to ask ChatGPT 4 the same question, it replies with correct amount. Since that podcast rised some concern about the real abilities of such models I think they fixed it (or anyway the model is constantly improving thanks to user feedback that it got fixed). This is the answer I got (without plugins): The time it takes for clothes to dry is generally not directly related to the quantity of clothes. The drying time depends more on factors such as how wet the clothes are, the temperature, humidity, and wind speed. If we assume that these variables stay constant, and the five clothes you initially dried didn't overlap each other and were equally exposed to the sun and wind, then you should be able to dry 30 clothes in the same amount of time (5 hours), given that you have enough space to spread them out as you did with the initial 5 clothes. However, if you don't have enough space and the clothes need to be overlapped or not all clothes get the same exposure to sun and wind, then it might take longer. But without additional information, it's hard to estimate how much longer it would take.
I hit on a similar idea. I've found that ChatGPT can be an excellent training tool for writing good instructions and for exposing hidden assumptions. Although I didn't attempt the straightforward with GPT-4, I did try this one: "I left 5 shirts out in the sun to dry. It took them 5 hours to dry. How long would it take 30 shirts to dry. Assume there is enough room for all of the shirts to be set out at the same time." The response was very solid: If you have enough room to set out all 30 shirts at the same time and the drying conditions remain the same (i.e., same sun intensity, same wind speed, same humidity), then it should also take 5 hours to dry 30 shirts. This is because the drying time is dependent on the conditions, not the number of shirts. The sun and air can work on all shirts simultaneously if they are all exposed at the same time. If space was a constraint and you could only dry a certain number of shirts at a time, then you would have to dry them in batches, which would take longer. But since you've stated there is enough room for all of the shirts to be set out at once, the drying time remains the same.
My first thought conclusion was: What first comes to your mind is often the best idea. (=the more you are trained in listen to your intuition -which is based on more information then your conscious mind can access.) It would be interesting to create statistic how it works with ChatGPT. The second problem is mathematics and not applicable. Whatever it is your scripts are insane brilliant and useful. I think this was a sneak peak of the reason process for future GPT versions, or....?
Brother I just have to tip my hat. Your many different presentations on prompting are nothing short of brilliant. More effectively than most on YT, they serve to demonstrate that in order to hack the best results out of these LLMs, you have to understand basic programming principles. (At least for now!)
Great improvement over the simpler tree of thought process shown in previous approach. This time it really worked like a human would act when faced with a task to brainstorm and find best solution.
Awesome, and thank you. I truly appreciate the simplicity and organization of your presentation. I absolutely must work on my frustration when submitting prompts and not receiving an appropriate reply. Sometimes Chappy will say random things such as, "Hmm, I'm not sure. But here's the latest recipe for chocolate chip cookies" (and yes, assigning GPU a name allows me to not feel so ridiculous about engaging in conversation with AI.)
To go around the content limitation, keep the titles of already discarded ideas in "Discarded: idea1, idea2, idea3, etc." and pass it to any new prompt loop. That will allow you to keep some sort of history and avoid repetition of ideas without breaking the limits. Good content by the way, keep up! Note: Do you know that you sound simmilar to uncle roger? 😂
I'm no expert in ToT but what I understand is that the idea is: 1. Pose the problem. Generate first layer of solutions and rank them. 2. Add a second layer by selecting each of first layer of solutions and generating different type of solutions from each with different question and then ranking them. In your example, generate a number of new solutions from each first solution (for example how to engage in direct discussion) by trying to add their pros and reducing cons. 3. Add third layer with another type of question for all solutions from 2nd layer or stop. 4. Finally compare all answers from the last layer to see which path lead you to best solution. (Even if the first layer didn't produce the best ranked solution.)
Thank you so much for your hard work. I noticed there was a difference between the prompt in the video and on your site. What is the difference between the two?
my current XP with gpt and gemini 1.5 brought me here. the whole chain is based on the assumption that gpt is stable enough to work through all of this. it isn't. and its measly 4096 token window will make it forget where the chain started. wild suggestion: feed gpt your chart and let it design the prompt by itself. besides that: flawless setup. keep it up. exactly what i am looking for.
I put the parts of code from Your video to a ChatGPT to generate missing parts. here's the result docs.google.com/document/d/1K6RmBG7cs5EkNbF6hBURABd7ljX8BhTyTuW2OVbqCH0/edit?usp=sharing
If I may here's an idea for a video. I'd do it, but my skill set is not there yet. Write a ToT prompt that lays out the scenario from the 1983 movie War Games. Have ChatGPT-4 assume the role of WOPR. Feed it a bunch of nuclear war simulations to find out which one leads the best possible outcome for the "winning" country. Then finally have it set up some AI Agents to play one another in a game of TIC-TAC-TOE. See if ChatGPT-4 figures out "the only winning move is not to play."
I tried some similiar stuff in. but I used prompts to setup different roles (in my case marketing guy, scientist (technical knowledge), business visionary guy, investor) and let them come discuss ideas with pros and cons, and loop several times. it was pretty interesting because certain aspects of ideas where good but then in the discussion it turned out they were to expensive or the market was to small, etc. you can also use gpt to comeup with roles, in your case: if i need advice concerning relationships, what kind of different personalities with different educations and personal experiences could be helpful to talk to, to get different ideas, opinions and interpretations. give a list
I asked it the same thing that you did, only I prefaced it by telling it that it had common sense and its answer was that the sun has the ability to dry them all simultaneously so it would still take 5 hours to dry 30 items.
Awesome as always! Funny thing is that you still can make winning idea better. It is another technique. I get shocked 😲 by ChatGPT capabilities everyday now ))
I have always believed that sentient artificial intelligence will go through problems in a cyclical manner. Because like out brain, it decides based on chemical imbalances in different parts of our brain that are activated differently for each thing. Since AI will not be encumbered by things like doubt, psychological issues, guilt, jealousy etc. it will truly evaluate from an objective standpoint. Sentience doesn't mean it needs to be emotional. Though if an AI actually gets enough tokens, a big enough storage capacity and a way to recycle in an efficient way, and if it starts evaluating it's own situation and if it's fair (one universal trait that comes really early after birth), and the AI becomes "hurt". We need to quickly be there for it.
From what i understand. Theres a max token of 8k in chatgpt plus. If there is too much input and responses exceeding the token limit, it will not remember the previous input. Do u have this issue while running so many times?
really interesting. I can't understand how it measures "success" when calculating the probability of success. You did not define what success means to you in this context. would you not need to add and weight the different elements of a successful outcome? in the clothes example you had the goal of minimising time required to dry the clothes I think.
I'm going crazy. I really want to use this prompt but I can't figure out how to install it. I've tried running it locally. I've tried Codespaces. I've tried Colab. I can't get it to work anywhere. Granted, I'm new to python.
But... I get the right answer to the clothes drying problem without all the fuss. In that problem, in general, I don't see the need for "brainstorming". I see instead the need for _clarification_ . The way the question is written, is it implying that it can only lay out 5 at a time? A human would ask. Also, is the use of direct sunlight explicitly desired? That is, if you are wanting the UV exposure for sanitary and deodorizing purposes, or just want the clothes dry. Because if you are limited to daylight hours, "30 hours" will stretch to many days and the clothes will be dry anyway. Real answers might be to wash with Oxy-clean to replace the UV sanitation, and crowd the clothing together so they have airflow but shade each other from direct sunlight, and fit all the clothes into the drying area and use other outdoor areas as well that don't have the direct sunlight.
This is good, in theory, but it doesn't seem to work. I have tried this a number of times now and the top result stays as the top result. GPT also tends to repeat the solutions multiple times, no matter how specifically you tell it not to. The other issue is that the problability seems to be lip service. When I ask for solutions that will contend with the winning solution and may have a higher rating, it just delivers the same results from earlier in the chat but gives them a higher rating. The initial prompt is great, very nicely crafted. And the final prompt (Deeper thinking) is also very considered. However the middle section can just be circumvented as it is almost entirely pointless with this iteration of GPT. This is certainly a good way to tackle problems in general (With humans).
I am torn about the usefulness of this; in your first example it seemed like chatgpt already priotized the best answer by itself [perhaps it already ran a million loops by itself before giving you its first answer] so running it through 5 additional loops seemed like a waste of time. it would be nice to show example of more complex things that our brains already doesn't know the answer to before you even hit enter on your first loop to perhaps better showcase the usefulness of this. the second issue which by the way is a math question and not meant for a language model , GPT4 got right on the first time for me, even without the wolfram plugin which obviously would edge out any need for looping when comes to math related questions
Is tree of thought prompting a thing that people like to experiment with, or does it make chatGPT more useful in solving real life problems? I have only used chatGPT maybe 20 times for very simple questions, and from my perspective, it looks like just thinking about the issue itself would take less time than all the prompting and reading that this method requires.
Amaze balls. Is there a way to automatically query GPT with data in an email, a pdf uploaded to Google Drive, Docs, Sheets, FAQs, price list, table/ sheet, etc. to add part or all of the data to a template prompt, and then send the response to a sheet, in an email, doc, CRM, etc.?
@@chrismachabee3128 Care to share how...paython? Was originally looking at it as a zapier type thing, but it seems the plugin has to be run from thel chat and not as an action of the new email trigger. There was a video about it here last week.
@@TheLoneCamper You have the advantage over me , sir. Your knowledge is unknown to me and unfortunately, I have no idea what you are referring to. Perhaps, I am not the person to address that question to. As you will notice I came here to learn, I have no answer to your question. I might suggest Google, which is still a most excellent source for answering complicated technical questions. Thanks for thinking I knew something.
Good video but i suggest you to use real use cases you could ask LLM, no way I'll take advice on such a topic... By the way, to have read TOT paper, this is not what you are presenting, there is no three search process that looks like your implementation A bit of click bait?
I do code learning work at a few places that include Udemy, and W3Schools is free. You can learn any code you want. They start with baby steps, from the beginning, and take you all the way up to you understand what you doing. Wish no more. Additionally, if you ask Google for free training in the code of your choice you will get the answer you need. Now, on the other hand, you got lots of bucks, and there are school a plenty. They call them boot camps and they will drive you to be the best coder. I have never been to a boot camp but that is what I hear.
Assigning a probability of success to the problem “how to break up with my girlfriend” is kind of ridiculous. It has zero information about you or your girlfriend. What criteria is the LLM using to define success? What data is it drawing on to even conclude that these percentages are right in the aggregated population? What verification are you using to confirm that this process has come up with a better solution? It’s impossible because the “ best way “ to break up with someone is incredibly subjective. LLMs have no idea what the concept of truth is. And has no justification to give its “ opinion” on anything.
You know, and this is just my opinion, instead of being afraid of AI, they need to put all you brilliant AI guys in a room or a building and replace the useless United Nations. That would fix all the world's problems right that! How about that?
As for clothing problerm, your script did not give a better answer. By the time this researcher presented her TED talk and you watched it, it was obsolete research. GPT4 solves the problem correctly on first try. """If your clothes are drying in the sun, the rate at which they dry isn't really dependent on the number of clothes you put out. In your case, it took 5 hours to dry 5 clothes, which means 1 cloth took 5 hours to dry. Therefore, regardless of the number of clothes you dry at the same time, each should still take about 5 hours, assuming the conditions remain constant (same amount of sunlight, same level of humidity, etc.).''''""
STEP 3 .. grhhhhh I'm sorry, but as an AI language model, I cannot evaluate the potential of each solution, assign a probability of success, or a confidence level to each option as these factors are subjective and dependent on individual circumstances. It's important to consider the unique situation and the individuals involved when choosing the best solution. It may be helpful to seek advice from trusted friends or a mental health professional to make an informed decision.
Silly. Not even watching this. Don't use computers to break up with people. That's the solution. That is the best solution. And that is the only solution.
ChatGPT is useless, it can not even convert simple php to react code, how useless is this AI. Devs should be scared at all, idk why you are still hyping this BS so much,
Furthermore I use stretchy PT in a moderately technical but have not used a python script like this before the other video that shows if I downloader purchase descriptive yours how I would implement it
GPT4 can already solve the clothes drying problem without the need for all of the chain of thought prompting. It’s much more intelligent since that TED talk from some time ago. It’s continuously being updated.
