Q&A from April 19, 2023
Solving Computational Problems and Writing Code
IF YOU ASK A QUESTION THAT DOESN'T REALLY NEED THE PLUGIN, WILL CHATGPT EVER DECIDE TO NOT USE THE PLUGIN? OR IS ENABLING THE PLUGIN A COMMAND TO USE IT?
IF YOU ASK A QUESTION THAT DOESN'T REALLY NEED THE PLUGIN, WILL CHATGPT EVER DECIDE TO NOT USE THE PLUGIN? OR IS ENABLING THE PLUGIN A COMMAND TO USE IT?
With the plugin enabled, chatGPT will make a "decision" for each input based on its training to determine if there are elements of the input response that could be improved by the plugin, but if it determines there aren't any elements of the query response that would be improved, it won't use the plugin.
IS IT LIKELY TO BE SIMILARLY VERSATILE IN DOING ECONOMIC ANALYSIS?
IS IT LIKELY TO BE SIMILARLY VERSATILE IN DOING ECONOMIC ANALYSIS?
With the Wolfram plugin, ChatGPT will have access to up-to-date financial information, along with the computational power and accuracy that you'd expect with Wolfram tech. Let us know if there are particular things you'd like to see and maybe we can try some examples on Friday's session.
I UNDERSTAND THAT THERE IS A PACKAGE FOR QUANTUM COMPUTING IN WOLFRAM, IS IT POSSIBLE TO WORK WITH THIS PACKAGE USING THE GPT PLUGIN?
I UNDERSTAND THAT THERE IS A PACKAGE FOR QUANTUM COMPUTING IN WOLFRAM, IS IT POSSIBLE TO WORK WITH THIS PACKAGE USING THE GPT PLUGIN?
Answered live: We are working on this
WE HAVE SEEN A LOT OF PROMPTS FOLLOWED BY LARGE RESULTS. AS FAR AS I KNOW, CHATGPT SOMETIMES HALLUCINATES, AND MIGHT LACK FUNCTIONAL CORRECTNESS. HOW CAN A USER VERIFY, IF THE RESULTS ARE MEANINGFUL AND CORRECT? DO I NEED TO BE A SUBJECT MATTER EXPERT?
WE HAVE SEEN A LOT OF PROMPTS FOLLOWED BY LARGE RESULTS. AS FAR AS I KNOW, CHATGPT SOMETIMES HALLUCINATES, AND MIGHT LACK FUNCTIONAL CORRECTNESS. HOW CAN A USER VERIFY, IF THE RESULTS ARE MEANINGFUL AND CORRECT? DO I NEED TO BE A SUBJECT MATTER EXPERT?
Answered live (paraphrased): Run a couple of test cases - either ChatGPT creates the tests or you write them yourself. At the end of the day a human must check if the tests passed to verify the result was correct.
An observation (paraphrased): Sometimes if you prompt ChatGPT to be terse in its responses, there may be more “incorrect” results. Its producing less content of its own and might end up making up results.
A comment (paraphrased): Take a medium sized problem (nothing too complicated or specific to a field of study), and ask ChatGPT to write WL code to create maybe a visualization. It does a great job of assimilating its knowledge of Physics, Chemistry, etc. other fields along with WL knowledge to provide a nice solution.
ARE THE 3D IMAGES STATIC, OR CAN THEY BE ROTATED?
ARE THE 3D IMAGES STATIC, OR CAN THEY BE ROTATED?
They are static when sent to the ChatGPT interface.
Thanks
Thanks
CAN QUESTIONS BE COMPLEX? CAN WE ENTER MULTIPLE QUESTIONS AT ONE TIME?
CAN QUESTIONS BE COMPLEX? CAN WE ENTER MULTIPLE QUESTIONS AT ONE TIME?
Yes! You can build multiple steps into a query or have multiple queries at once. Unfortunately there is a bug affecting all users currently on the OpenAI side, preventing the full chat features, but I've had long back and forth conversations about multiple problems.
thanks
thanks
REGARDING THE MOLECULE. WOLFRAM|ALPHA CAN DO THE JOB EXTREMELY WELL. MORE GENERALLY THEN, WHAT IS THE DESIRED OUTCOME NOW AND IN THE FUTURE OF USING CHATGPT?
REGARDING THE MOLECULE. WOLFRAM|ALPHA CAN DO THE JOB EXTREMELY WELL. MORE GENERALLY THEN, WHAT IS THE DESIRED OUTCOME NOW AND IN THE FUTURE OF USING CHATGPT?
I think one of the benefits of using a tool like ChatGPT that can make its own Wolfram Alpha calls is that you can build up more complicated queries using natural language, like Jason is showing off currently. One can also integrate these results with some of the other things that ChatGPT does well, like writing essays, lesson plans, etc.
Thanks.
Thanks.
HOW DOES THE WOLFRAM PLUG-IN PERFORM WITH GPT-4?
HOW DOES THE WOLFRAM PLUG-IN PERFORM WITH GPT-4?
Presently OpenAI hasn't updated chatGPT to be using GPT-4, so GPT-4 doesn't have access to any plugins. The plugins are a feature only for use in chatGPT currently.
I HAVE THE PROBLEM THAT HAVEN'T ACCESS TO THE PLUGGING. COULD YOU GIVE SOME CLUE HOW TO GET IT? I SENT A LOT OF REQUEST BUT UNFORTUNATELY HAVEN'T RESPONSE.
I HAVE THE PROBLEM THAT HAVEN'T ACCESS TO THE PLUGGING. COULD YOU GIVE SOME CLUE HOW TO GET IT? I SENT A LOT OF REQUEST BUT UNFORTUNATELY HAVEN'T RESPONSE.
Unfortunately the only option we can point you to is the OpenAI page https://openai.com/blog/chatgpt-plugins that allows users to join the plugin waitlist. Good luck.
Thanks.
Thanks.
WILL YOU DESCRIBE SOMETHING ABOUT HOW THE PLUGIN WORKS? I THINK THAT WOULD BE HELPFUL FOR UNDERSTANDING THE POSSIBILITIES AND LIMITATIONS OF THE VARIOUS PROMPTS
WILL YOU DESCRIBE SOMETHING ABOUT HOW THE PLUGIN WORKS? I THINK THAT WOULD BE HELPFUL FOR UNDERSTANDING THE POSSIBILITIES AND LIMITATIONS OF THE VARIOUS PROMPTS
As Michael shared, the plugin works by enabling chatGPT to make API calls to either of a Wolfram|Alpha and Wolfram Language API. chatGPT creates the API calls just like it creates the textual response you see and makes the call and then ingests the response to inform what it says in response to your prompt.
ok, I think I got it on that "black box" level. But it would be better with some more details. Watching the prompts doing their magic is now like learning a programming language only through examples.
ok, I think I got it on that "black box" level. But it would be better with some more details. Watching the prompts doing their magic is now like learning a programming language only through examples.
THAT IS AN IMPROVEMENT A BIT MORE WOULD HELP
THAT IS AN IMPROVEMENT A BIT MORE WOULD HELP
Will pass this on. You should also be able to enlarge the screen in the browser.
TO ENABLE THE PLUG-IN WITH CHATGPT, IS THAT BASED ON CHATGPT'S DISCRETION OR A USER PREFERENCE?
TO ENABLE THE PLUG-IN WITH CHATGPT, IS THAT BASED ON CHATGPT'S DISCRETION OR A USER PREFERENCE?
Once you have access, there is a drop down in the new chat window that will let you select both the model used, as well as which plugin(s) you would like to enable.
WHICH VERSION OF CHATGPT WERE YOU USING IN THESE EXAMPLES?
WHICH VERSION OF CHATGPT WERE YOU USING IN THESE EXAMPLES?
There is a specific Model that has access to the plugins that is separate from 3.5 and 4.0, but I believe is based on 3.5.
WHEN I USED THE RESOURCE FUNCTION AIASSISTANT TO GENERATE WOLFRAM CODE, IT KEPT MAKING UP OPTIONS AND FUNCTIONS THAT DO NOT EXIST IN THE WOLFRAM LANGUAGE 13.2. WHY IS THAT AND ARE THERE WAYS OF PROMPTING TO MAKE SURE FROM THE START IT'S NOT MAKING THESE "HALLUCINATING" CHOICES?
WHEN I USED THE RESOURCE FUNCTION AIASSISTANT TO GENERATE WOLFRAM CODE, IT KEPT MAKING UP OPTIONS AND FUNCTIONS THAT DO NOT EXIST IN THE WOLFRAM LANGUAGE 13.2. WHY IS THAT AND ARE THERE WAYS OF PROMPTING TO MAKE SURE FROM THE START IT'S NOT MAKING THESE "HALLUCINATING" CHOICES?
There are things that we can do via prompt engineering (as Michael is showing off), but LLMs will probably always have this issue. However, if you would be willing to share some particular examples, that will help us to debug, at least as much as we can. https://resources.wolframcloud.com/FunctionRepository/feedback-form #AIAssistant
Thanks! I will try to do so.
Thanks! I will try to do so.
WHY SOMETIMES CHATGPT GIVE CODE BLOCK, AND OTHER TIMES, JUST CALL WOLFRAM?
WHY SOMETIMES CHATGPT GIVE CODE BLOCK, AND OTHER TIMES, JUST CALL WOLFRAM?
It depends on the prompt and what it "felt" it should do. Michael often asks for code displayed whereas in the chemistry examples I didn't.
If it uses the Wolfram plugin but doesn't show the code, you can also ask in a follow up prompt to show "how did you do that computation?" or similar and it will display the code, often with comments so you can read through it if you'd like.
If it uses the Wolfram plugin but doesn't show the code, you can also ask in a follow up prompt to show "how did you do that computation?" or similar and it will display the code, often with comments so you can read through it if you'd like.
HAS ANYONE CHECKED TO SEE IF WRITING A LONG PARAGRAPH FOR CHATGPT TAKES MORE OR LESS TIME THAN SIMPLY EXECUTING THE CORRECT CODE IN THE WOLFRAM LANGUAGE ? I DON'T HAVE THE PLUGIN SO I CAN'T DO IT.
HAS ANYONE CHECKED TO SEE IF WRITING A LONG PARAGRAPH FOR CHATGPT TAKES MORE OR LESS TIME THAN SIMPLY EXECUTING THE CORRECT CODE IN THE WOLFRAM LANGUAGE ? I DON'T HAVE THE PLUGIN SO I CAN'T DO IT.
It depends. If one is very familiar with the necessary WL functions, the human is commonly faster. However, that is not often the case. See the Functional group example in my notebook. There, typing in the MoleculePattern arguments gets tedious quickly. In cases like that the bot can really simplify the work.
Well, if you're just sitting there challenging the LLM with stuff you think up on the fly, that's probably correct. However, if you're looking at something you're particularly interested in, you likely do have the familiarization with the functions. However, your point about tedious typing is well taken.
Well, if you're just sitting there challenging the LLM with stuff you think up on the fly, that's probably correct. However, if you're looking at something you're particularly interested in, you likely do have the familiarization with the functions. However, your point about tedious typing is well taken.
CAN I ASK CHATGPT AND WOLFRAM PLUG IN TO LOOK FOR PATTERNS IN A DATA SET
CAN I ASK CHATGPT AND WOLFRAM PLUG IN TO LOOK FOR PATTERNS IN A DATA SET
Currently, chatGPT doesn't have capability to import data sets, so you would be limited by the character limit of the input field. This would be an exciting thing to try though if import becomes a feature!
CAN WE LOOK FOR PATTERNS IN A DATA SET
CAN WE LOOK FOR PATTERNS IN A DATA SET
Currently, chatGPT doesn't have capability to import data sets, so you would be limited by the character limit of the input field. This would be an exciting thing to try though if import becomes a feature!
IS IT UP TO THE USER TO DISCERN WHEN THE RESULTS ARE NOT CORRECT? CAN IT LEARN THAT TOO?
IS IT UP TO THE USER TO DISCERN WHEN THE RESULTS ARE NOT CORRECT? CAN IT LEARN THAT TOO?
Yes and no. It certainly happens where chatGPT will write incorrect code. That said, an approach that has worked quite well that Michael mentioned is prompting chatGPT to create its own test cases for the task at hand. In that case, it iterate for quite a while on the code its writing until the test cases are all succeeding. If you don't do this, it is still best to double check things yourself, but it does a good job of commenting on the code it writes so it should be fairly easy to read the code and determine if things are as you intended.
HELLO, THIS MAY HAVE BEEN ANSWERED ALREADY SORRY BUT DO YE KNOW HOW SOON THE PLUGIN MAY BECOME AVAILABLE TO SOMEONE ON THE WAITING LIST WHO ISN'T SUBSCRIBED TO OPENAI PLUS OR WHETHER IT WILL BE MUCH SOONER FOR PLUS CUSTOMERS? THANKS
HELLO, THIS MAY HAVE BEEN ANSWERED ALREADY SORRY BUT DO YE KNOW HOW SOON THE PLUGIN MAY BECOME AVAILABLE TO SOMEONE ON THE WAITING LIST WHO ISN'T SUBSCRIBED TO OPENAI PLUS OR WHETHER IT WILL BE MUCH SOONER FOR PLUS CUSTOMERS? THANKS
Unfortunately we do not have information on that. It would have to be a question for OpenAI
Thanks. Do ye know how many users the plugin is available to already?
Thanks. Do ye know how many users the plugin is available to already?
CAN YOU ASK THE PLUGIN TO WRITE CODE THAT WILL DO SIMULATIONS - LIKE THE ISING MODEL OR A PERCEPTRON?
CAN YOU ASK THE PLUGIN TO WRITE CODE THAT WILL DO SIMULATIONS - LIKE THE ISING MODEL OR A PERCEPTRON?
Yes, presuming the model is common and existed before the last training of ChatGPT it should do ok.
IF WE HAVE TROUBLE FINDING THE CORRECT CODE IN THE MATHEMATICA NOTEBOOK, CAN CHATGPT HELP US GENERATE THE CODE IN THE MATHEMATICA NOTEBOOK?
IF WE HAVE TROUBLE FINDING THE CORRECT CODE IN THE MATHEMATICA NOTEBOOK, CAN CHATGPT HELP US GENERATE THE CODE IN THE MATHEMATICA NOTEBOOK?
Yes! Prompts like "Write me some Wolfram language code that does..." often work quite well.
Paraphrased from Michael’s live answer: If you can reasonably describe what you want it to do (even if the description is half a page), ChatGPT will do a decent job of writing code, maybe not most efficient, but a decent job. And you can go back and nudge it...rethink do you really want to use “AppendTo” here or something along those lines. And it says “probably I should use Association or Sow and Reap” and you can say go ahead.
ChatGPT can not only generate Wolfram Language cod , but it should be able to generate the notebook, for e.g.
Notebook[{Cell["head","Section"],Cell["text","Text"]}]
so it can find the underlying data structure of a notebook and thereby should be able to detect the error. So in principle (provided you have the API for it) you can send it a notebook and ask ChatGPT to find the error in the notebook.
CAN YOU ASK CHATGPT TO LEARN FROM THE IMPORT?
CAN YOU ASK CHATGPT TO LEARN FROM THE IMPORT?
In the same manner as my example where I told it to read in usage statements for a particular function, it can use information that it reads. However, it does have a limited token memory so at some point it may forget what it read.
So I misspoke about it learning from the import. The bot does not get access to the imported data.
too bad :-) Would be interesting otherwise.
So I misspoke about it learning from the import. The bot does not get access to the imported data.
too bad :-) Would be interesting otherwise.
CAN YOU FIND THE EQUATION OF THE FUNCTION IN WL, EXAMPLE: "FINDPEAKS" OR OTHERS?
CAN YOU FIND THE EQUATION OF THE FUNCTION IN WL, EXAMPLE: "FINDPEAKS" OR OTHERS?
If you ask it "how can you find the peaks of a function" it will try to return Wolfram language code to do this. You may need to tweak it a bit
Thanks !
Thanks !
INDEED VERY INFORMATIVE. THANK YOU ! WHAT ARE YOUR BIGGEST COMPLAINTS OR FRUSTRATIONS AT THIS STAGE OF THE GAME.
INDEED VERY INFORMATIVE. THANK YOU ! WHAT ARE YOUR BIGGEST COMPLAINTS OR FRUSTRATIONS AT THIS STAGE OF THE GAME.
The occasional bugs such as the limit today with only one chat input. The other top issue is when it ignores the prompt for example not typesetting things properly in chemistry.
Impressive. Thank you again!
Impressive. Thank you again!
HOW DO YOU EXTEND CHATGPTS:S KNOWLEDGE OF ADDITIONAL API FUNCTIONS?
HOW DO YOU EXTEND CHATGPTS:S KNOWLEDGE OF ADDITIONAL API FUNCTIONS?
CAN THIS PLUGIN BE USED TO DO DFT CALCULATIONS AS WELL FOR SPECIFIC PARAMETERS? FOR EXAMPLE, MAYBE CALCULATE THE GROUND STATE ENERGY OF AN ELEMENT, SAY CU, IN A SYSTEM
CAN THIS PLUGIN BE USED TO DO DFT CALCULATIONS AS WELL FOR SPECIFIC PARAMETERS? FOR EXAMPLE, MAYBE CALCULATE THE GROUND STATE ENERGY OF AN ELEMENT, SAY CU, IN A SYSTEM
No, not at this point.
WILL THE PLUG-IN ACCEPT VISUAL INFORMATION AS FROM A CHART, GRAPH OR SPREADSHEET AND MAKE SOME ANALYSIS OF THIS INPUT?
WILL THE PLUG-IN ACCEPT VISUAL INFORMATION AS FROM A CHART, GRAPH OR SPREADSHEET AND MAKE SOME ANALYSIS OF THIS INPUT?
Not at present. Currently the API's only work with textual content, but perhaps once chatGPT is using GPT4, which is also better able to work with images, we will update the plugin accordingly.
IF WE HAVE TWO MATH EQUATIONS THAT ARE EQUIVALENT, CHATGPT CA SOLVE ONE EQUATION AND SHOWS THAT ARE EQUIVALENTS
IF WE HAVE TWO MATH EQUATIONS THAT ARE EQUIVALENT, CHATGPT CA SOLVE ONE EQUATION AND SHOWS THAT ARE EQUIVALENTS
Trivial examples - yes. If something more complicated - no.
WHEN I GO TO THE LINK TO RESOURCES AND THEN CLICK ON THE Q&A DIGEST SAY FOR DAY 1, I GET A MESSAGE THAT SAYS I DON'T HAVE PERMISSION TO ACCESS THIS ITEM. WHAT GIVES?
WHEN I GO TO THE LINK TO RESOURCES AND THEN CLICK ON THE Q&A DIGEST SAY FOR DAY 1, I GET A MESSAGE THAT SAYS I DON'T HAVE PERMISSION TO ACCESS THIS ITEM. WHAT GIVES?
This has been fixed.
A QUESTION OF SCALE. HOW DOES THE WOLFRAM DATA REPOSITORY RATE VIS A VIS MICROSOFT OR GOOGLE SIMILAR RESOURCES? ISN'T IT DWARFED?
A QUESTION OF SCALE. HOW DOES THE WOLFRAM DATA REPOSITORY RATE VIS A VIS MICROSOFT OR GOOGLE SIMILAR RESOURCES? ISN'T IT DWARFED?
The Wolfram Data Repository is a public resource that hosts a collection of computable datasets, curated and structured to be suitable for immediate use in computation, visualization, analysis and more.
It provides a system for storing data and making it immediately computable and useful - as a resource for publicly available data and data-backed publication.
It provides a system for storing data and making it immediately computable and useful - as a resource for publicly available data and data-backed publication.
CAN WOLFRAM UTILIZE NVIDIA GPU'S FOR MACHINE LEARNING AND LLM TRAINING?
CAN WOLFRAM UTILIZE NVIDIA GPU'S FOR MACHINE LEARNING AND LLM TRAINING?
Absolutely. When using NetTrain (https://reference.wolfram.com/language/ref/NetTrain.html ) to train a model you can set the TargetDevice option to “GPU” (https://reference.wolfram.com/language/ref/TargetDevice.html )
Check out these models in the Neural Net Repository: https://resources.wolframcloud.com/NeuralNetRepository/search/?i=GPT
Check out these models in the Neural Net Repository: https://resources.wolframcloud.com/NeuralNetRepository/search/?i=GPT
DOES THE COMMUNICATION BETWEEN CHATGPT AND THE PLUG-IN HAPPEN AS ROTE COMMUNICATION OR IS THERE SOME VALIDATION BETWEEN THE 2 AIS?
DOES THE COMMUNICATION BETWEEN CHATGPT AND THE PLUG-IN HAPPEN AS ROTE COMMUNICATION OR IS THERE SOME VALIDATION BETWEEN THE 2 AIS?
Answered live: no validation just handshake
CAN YOU SAY A LITTLE OF THE INITIAL (HIDDEN) PROMPT THAT YOU HAVE FOR THE PLUGIN?
CAN YOU SAY A LITTLE OF THE INITIAL (HIDDEN) PROMPT THAT YOU HAVE FOR THE PLUGIN?
Michael has listed these in details in the "The Importance of the Prompt" section at https://www.wolframcloud.com/obj/mtrott/Published/ChatGPTWednesdayStudyGroup.nb
Thanks, Abrita. But as I understand it there is an initial (and hidden) major prompt that is used for each session with the plugin.
Perhaps the detailed pre-prompt on Christopher’s notebook here https://wolfr.am/1cOXF9E1e might be interesting to you.
ON LOGAN PAGE'S QUESTION - CAN YOU ELABORATE HOW THE PLUG-IN "DECIDES" TO USE WOLFRAM. WHAT IS THE MECHANISM HERE?
ON LOGAN PAGE'S QUESTION - CAN YOU ELABORATE HOW THE PLUG-IN "DECIDES" TO USE WOLFRAM. WHAT IS THE MECHANISM HERE?
Every plug-in comes with a manifest that tells ChatGPT when and how to call the plugin. Based on this instruction ChatGPT at a later time may make the decision if and when to call the plugin.
DEBUGGING COMPUTER PROGRAMS IS AN ART AND SCIENCE. SHOULDN'T WOLFRAM CONSIDER A FORMAL AND SYSTEMATIC PROCEDURE TO DEBUG CHATGPT ANSWERS?
DEBUGGING COMPUTER PROGRAMS IS AN ART AND SCIENCE. SHOULDN'T WOLFRAM CONSIDER A FORMAL AND SYSTEMATIC PROCEDURE TO DEBUG CHATGPT ANSWERS?
The space of possible answers is very large - so difficult.
LET'S SAY I HAVE PLENTY OF FORMULATIONS AND SOLUTIONS N A SPECIFIC ENGINEERING TOPIC IN MATHEMATICA NOTEBOOKS. HOW DO I STEER CHATGPT QUESTIONS RELATED TO THE TOPIC TO USE MY NOTEBOOKS ON THE CLOUD?
LET'S SAY I HAVE PLENTY OF FORMULATIONS AND SOLUTIONS N A SPECIFIC ENGINEERING TOPIC IN MATHEMATICA NOTEBOOKS. HOW DO I STEER CHATGPT QUESTIONS RELATED TO THE TOPIC TO USE MY NOTEBOOKS ON THE CLOUD?
You could try to point ChatGPT to the published notebook on the cloud in an initial prompt and instruct it to use the information there while answering further questions.
MANY PHYSICAL PROBLEMS CAN BE FRAMED OR SOLVED WITH DIMENSIONAL ANALYSIS ALONE. CAN GPT SET UP ORDER-OF-MAGNITUDE OR SKETCHES OF SOLUTIONS WITH OUT NITTY-GRITTY?
MANY PHYSICAL PROBLEMS CAN BE FRAMED OR SOLVED WITH DIMENSIONAL ANALYSIS ALONE. CAN GPT SET UP ORDER-OF-MAGNITUDE OR SKETCHES OF SOLUTIONS WITH OUT NITTY-GRITTY?
Yes it seems to be able to do this.