[WSG24] Daily Study Group: Wolfram Tools for LLMs

Ideas
1

A Wolfram U Daily Study Group focusing on LLMs and Wolfram Tech will be starting soon!

Join Wolfram Experts @Alan Joyce , Rick Hennigan, Timothee Verdier and @Jon McLoone and a group of fellow enthusiasts to learn about all the different ways you can use Wolfram tech along with Large Language Models (LLMs). Our topics for the study group include the Wolfram GPT, Chat Notebooks and LLM Functionality in Wolfram Language. The sessions will feature a number of fun examples that will show you how you can combine the computational powers and curated knowledge of Wolfram Language along with the generative capabilities of LLMs.

Each session will kick off with a presentation by an expert who will introduce the topic, followed by some review questions and discussions. There will be time for Q&A at the end.

No prior Wolfram Language experience is required to join the study group.

Please feel free to use this thread to collaborate and share ideas, materials and links to other resources with fellow learners.

We look forward to seeing you:
March 25th-29th, 2024,
11am-12pm CT (4-5pm GMT)

REGISTER HERE

WolframUBanner_OP_1_CP__OP_1_CP_.jpeg
WolframUBanner_OP_1_CP__OP_1_CP_.jpeg1467Ă—198 12.3 KB

8 Likes
2

Kicking off the study group today with a session about the Wolfram GPT.

1 Like
3

In Monday’s session, Alan Joyce pointed out that we can use a Chat Notebook with an OpenAI token, but Wolfram GPT requires a ChatGPT Plus subscription. As a retiree who uses the Wolfram Language only for personal projects, $20/month is way too expensive for my limited use. Why can’t we have an option for pay-as-you-go for this? I can accept that it might require a different API token, possibly charged at a higher rate, but my usage shouldn’t cost anywhere near $20/month.

2 Likes
4

Tomorrow we are going to discuss Chat Notebooks.
Please refer to Wolfram Support Quick Answers: How can I start using Chat Notebooks in Wolfram products? to learn how you can start using Chat Notebooks or other LLM-related functionality in Wolfram products, with the help of an account with a supported chat & generative model service provider like

  • “OpenAI” — speech, image, text computation with OpenAI
  • “GoogleSpeech” — speech synthesis and recognition using GoogleSpeech
  • “ElevenLabs” — speech synthesis and recognition using ElevenLabs
  • “Anthropic” — chat and other text computation with Anthropic
  • “PaLM” — chat and other text computation with PaLM
1 Like
5

Hi Abrita,
Attached is a notebook where I am trying to get the code for a new custom Moonrise function. I tried several ways but in the end the code didn’t work. Comments? Suggestions? Thank you kindly.
John

Moonrise calculation example that is not successful..nb (35 KB)

6

Hi, Abrita.

The link to the session files shows an empty folder (see attached).

Please advise.

Thanks,
Harlan

Screen Shot 2024-03-25 at 6.43.06 PM.jpg
Screen Shot 2024-03-25 at 6.43.06 PM.jpg1784Ă—998 38.2 KB

7

Day 2 of our Daily Study Group series on Wolfram Tools for LLMs: Today we are going to talk about Chat Notebooks.

3 Likes
8

We’ve added the following to the download folder:

  • a couple of notebooks from @Michel Trott
  • presentation notebook from Rick Hennigan today
  • Q&A digest from days 1 and 2 of the DSG
3 Likes
9

We added a few notebooks today.

3 Likes
10

Looking into this @John Davidson

11

Congrats on your excellent work. Nevertheless during the study group a trial account should have been given to all participants such as to try and work on the relevant subjects such as participants consider a paid subscription after testing the fit to their own needs.

1 Like
12

Day 3 of our Daily Study Group series on Wolfram Tools for LLMs: Today we are going to talk about Programmatic Access to LLM Functionality. Our speaker is @Timothee Verdier

1 Like
13

I checked in with our Astronomy expert @Jeffrey Bryant.
I will post more details from our discussion, but the solution he recommended was to use natural language understanding native to Wolfram Notebooks which fetches the results from the Wolfram Alpha knowledgebase.
In a regular notebook, press the equal sign (=) to start a natural language input and type in your query.

For example:

= next rise time of the Moon from Champaign, IL on March 27, 2024

This will convert to Wolfram Language input and evaluate to give you the result:

Entity["PlanetaryMoon", "Moon"][
 EntityProperty["PlanetaryMoon", "RiseTime", 
     {"Date" -> DateObject[{2024, 3, 27}], "Location" -> 
         Entity["City", {"Champaign", "Illinois", "UnitedStates"}]}]]

Screenshot of my notebook:

ScreenShot2024-03-27at10.01.00AM.png
ScreenShot2024-03-27at10.01.00AM.png1310Ă—448 51.7 KB

14

Thank you kindly Abiria & Jeffrey,
Unfortunately, I already was able to solve a similar more-elementary situation using the City entity. I figured Wolfram & ChatGPT together should be able to write the custom code that works for giving simply the Lat / Long data for any location on the globe along with any reasonable past or future date. Somewhat like the first attempt that failed because it attempted to use the old (deprecated) “AstronomicalData” which caused it to fail:

Needs["AstronomicalData`"]
SetOptions[AstronomicalData,"Overwrite"->True]
date=DateObject[{year,month,day,hour,minute}];
lat=yourLatitude;
lon=yourLongitude;
moonrise=AstronomicalData["Moon",{"Rise",date,lat,lon}];
moonriseTime=DateString[moonrise,{"Hour",":","Minute"}]
1 Like
15

Thank you kindly Abrita & Jeffrey.
Unfortunately, I have already found a similar solution, but I was trying to find a slightly more ambitious results like the final effort to generate a custom function that uses lat / long data that failed:

To calculate the moonrise time for a specific location and date in the Wolfram Language, we can create a custom function ChatCodeInlineTemplate[Moonrise[]] that calculates the moonrise time using the Horizon Event function. Here is the code for the custom ChatCodeInlineTemplate[Moonrise[]] function:

Moonrise[location_,date_]:=Module[{horizonEvent},
  horizonEvent=HorizonEvent["Moon",location,date,"Rise"];
  DateString[horizonEvent,{"Hour",":","Minute",":","Second"}]
]

(* Example usage *)
location=GeoPosition[{38.8783,-77.0687}];
date=DateObject[{2024,3,25}];
moonriseTime=Moonrise[location,date]

This custom ChatCodeInlineTemplate[Moonrise[]] function uses the ChatCodeInlineTemplate[HorizonEvent] function to calculate the moonrise time for a specific location and date. You can use this function by providing the geographic location as a GeoPosition object and the date as a DateObject.

Comments? Suggestions? What is wrong with the code ChatGPT / Wolfram generated? Where did it go off the track?

16

Day 4 of our Daily Study Group series today. @Jon McLoone will be talking about “Injecting Reliability into Generative AI”.
See you there.

1 Like
17

Here are some notes from @Jeffrey Bryant :

  1. We do not have a built-in documented function for computing Moonrise.
  2. The algorithm for computing moon rise isn’t a simple formula so its unlikely that the LLM would know how to write an algorithm from the bottom up.
  3. The LLM is hallucinating badly with things like Needs[“AstronomicalData`”] which makes no sense. In fact, AstronomicalData is obsolete and shouldn’t be used in any case

Additional notes from @Alan Joyce

  1. it looks to me (I could be wrong) that the user was running their chatbook with the “plain chat” persona instead of the “code assistant” persona. The latter would let the “assistant” evaluate WL code and see that things weren’t working correctly… though (), that doesn’t mean there’s a straight path to something useful.
  2. Could probably adapt Jeff’s suggestion and tell the model to try using SemanticInterpretation[“next rise time of the Moon from Champaign, IL on March 27, 2024”] to see how to construct a valid WL expression, which then I would expect it to be able to turn into something that could take arbitrary locations and dates and return a Moon rise time.
18

Abrita,

I am facing an old and repetitive issue: I do not receive any Bigmarker course email (like reminder, etc.). Therefore, I kindly ask you to send me the email with the link to the quiz that you mentioned yesterday to my email address via your Wolfram email account.

Thanks in advance for your help.

Regards,
JĂĽrgen

19

Sure. I will forward your request to our coordinator and they will email all the links to you.
We will also share the quiz link at today’s live session.

20

Day 5 of our Daily Study Group series. We will look at some built-in AI related functionality useful for NLP tasks. We will also review some topics from the earlier sessions this week.

1 Like