"use strict";
const {Fragment: _Fragment, jsx: _jsx, jsxs: _jsxs} = arguments[0];
function _createMdxContent(props) {
  const _components = {
    a: "a",
    h2: "h2",
    p: "p",
    span: "span",
    strong: "strong",
    ...props.components
  };
  return _jsxs(_Fragment, {
    children: [_jsxs(_components.h2, {
      id: "how-did-steve-get-his-name",
      children: [_jsx(_components.a, {
        "aria-hidden": "true",
        tabIndex: "-1",
        href: "#how-did-steve-get-his-name",
        children: _jsx(_components.span, {
          className: "icon icon-link"
        })
      }), "How did Steve Get His Name?"]
    }), "\n", _jsxs(_components.p, {
      children: ["In December of 2023, ", _jsx("a", {
        href: "https://www.google.com/search?q=ethan+mollick",
        target: "_blank",
        children: "Professor Ethan Mollick"
      }), " from Wharton posted the following on LinkedIn:"]
    }), "\n", _jsx("div", {
      style: {
        backgroundColor: 'lightgoldenrodyellow',
        padding: '1rem'
      },
      children: _jsx("a", {
        href: "https://www.linkedin.com/posts/emollick_the-thing-about-llms-that-make-them-unintuitive-activity-7125848204339732480-OeZL/",
        target: "_blank",
        children: _jsx("img", {
          src: "/assets/images/posts/steve-name.png",
          alt: "How Steve Got His Name",
          style: {
            marginLeft: 'auto',
            marginRight: 'auto',
            borderRadius: '10px'
          }
        })
      })
    }), "\n", _jsx(_components.p, {
      children: "For some reason I found this really funny, but I also really liked the mental model of a first year grad student or intern as a way to think about an AI assistant that lived in your browser. Thus \"Ask Steve\" was born."
    }), "\n", _jsxs(_components.p, {
      children: [_jsx(_components.a, {
        href: "https://openai.com/dall-e-3",
        children: "DALL-E 3"
      }), " drew Steve's picture (self-portrait?)."]
    }), "\n", _jsxs(_components.p, {
      children: ["Professor Mollick has ", _jsx("a", {
        href: "https://www.linkedin.com/in/emollick/recent-activity/all/",
        target: "_blank",
        children: "a lot of great posts on LinkedIn"
      }), ", and you should follow him if you're interested in LLMs. He also has a good book out called ", _jsx("a", {
        href: "https://www.penguinrandomhouse.com/books/741805/co-intelligence-by-ethan-mollick/",
        target: "_blank",
        children: _jsx(_components.strong, {
          children: "Co-Intelligence: Living and Working with AI"
        })
      }), "."]
    }), "\n", _jsx(_components.p, {
      children: "- rajat"
    })]
  });
}
function MDXContent(props = {}) {
  const {wrapper: MDXLayout} = props.components || ({});
  return MDXLayout ? _jsx(MDXLayout, {
    ...props,
    children: _jsx(_createMdxContent, {
      ...props
    })
  }) : _createMdxContent(props);
}
return {
  default: MDXContent
};


News


## Introduction
When I was at Google I wanted my product management team to get some hands on experience with LLMs so in our last offsite I setup a short LLM hackathon. Many of them had heard about LLMs but hadn't actually used them, or had briefly chatted with ChatGPT or Gemini but didn't really have a visceral sense for what they were capable of. 

I gave them a set of prototyping tools to pick from including several internal Google tools and then one that I had made myself, which integrated Google's Gemini LLM directly into Google Sheets. My hypothesis was that Sheets would be the fastest way for most of them to get up and running given that they were already familiar with the UI. And that combining the power of Sheets with an LLM would unlock some interesting new possibilities.

It turned out that most teams did end up using the Sheets integration I had built. And they came up with some fun ideas :)

I rebuilt that integration and am sharing it here for anyone to use.

It's basically a baby version of Ask Steve. It enables you to create prompts with variables, change the variables, run the prompts without leaving Sheets, and even chain prompts together by referencing cells in other sheets. 

It's a great way to get people who haven't played much with LLMs to dip a toe in the water with a familiar interface. It's also great for building sequential workflows with chained prompts. Give it a try, it's free!


## Walkthrough Video 
<iframe width="400" height="281" src="https://www.youtube.com/embed/UNxUxN_F52k?si=EmJyJjZoCUuTV42z" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share; fullscreen" referrerpolicy="strict-origin-when-cross-origin" allowfullscreen></iframe>

## Instructions
1. Make your own copy of **[this Google Sheet](https://docs.google.com/spreadsheets/d/1GZYV6CsLSmISoMAm9b6rC6ixYO8a6ht2zwXHMyUoJE4/edit#gid=1556497876)**.
2. From the copy dialog you can review the attached script to ensure that it's not doing anything it shouldn't be. You can also review the code [here](https://github.com/rajatrocks/sheets-llm/blob/main/apps-script.js)
2. In your new spreadsheet, enter your Google Gemini API Key on the SETUP sheet in cell A1. [Get one for free!](https://quickstart.asksteve.to/getkey.html)
3. To test that it's working, go to the second sheet and then select `Run` from the `Sheets LLM` menu. It may take the menu a few moments to appear. The first time you select it, you will need to give the script access permissions to Sheets and to make external calls. You might see a scary looking dialog that says "Google hasn't verified this app". You will need to click on the `Advanced` link and then `Go to Sheets LLM (unsafe)`. 
3. You will need to select the menu item a second time after giving permissions. The Result field should fill in with the results. 
4. Try your own prompts and adjusting the [temperature](https://www.google.com/search?q=what+is+temperature+for+an+llm) between 0.0 and 1.0. You can put variables in your prompts by surrounding them with double-brackets like [[this]]. Then you can put the values for your variables in the cells below. See the examples in the sample sheets. 
5. You can chain LLM calls using multiple sheets. Note that the "2. Write a press release" sheet is taking the result from the "1. Generate product ideas" sheet and using it as a variable for a second prompt. And similarly, the "3. Suggest visuals" sheet is using the press release as an input. You will need to `Run` on each sheet to update it. Add as many new sheets as you like to this doc and run them independently or chain them together.
6. Any [live data that Sheets can import](https://hackernoon.com/3-best-ways-to-import-external-data-into-google-sheets-automatically-bw163yxm) can be sent to Gemini. As an example the "Google News Feed" sheet reads the Google News rss feed and updates once an hour. Then the "Generate a movie plot from the news" sheet pulls the top news as variables and sends it to Gemini to create a movie plot.
7. Feel like modifying the script that powers all this? Under the "Extensions" menu select "Apps Script" to access it. It is your own copy of the script that is attached to this specific doc, so feel free to make changes.
8. This is using the publicly accessible Gemini API. Your input and output data [can be used to improve Gemini](https://ai.google.dev/pricing). Do not send it any sensitive or confidential information. 

> Want to unlock the power of ChatGPT and Gemini in any web page? 
> **[Install Ask Steve Now](https://chromewebstore.google.com/detail/ask-steve-do-your-job-bet/gldebcpkoojijledacjeboaehblhfbjg?hl=en&authuser=0)**

Experiment with LLMs in Google Sheets

*This was originally written while I was at Google, so the examples are geared toward software that most Googlers would know - rajat*

People often talk about wanting to make products delightful. What does this mean in practice? 

<div style={{backgroundColor: 'violet', padding: '1rem'}}>
**Delight is when a task is made unexpectedly easy**
</div>

The key thing to note here is that Delight is based on **the user’s expectations**.
* If they are expecting a task to be difficult and it is, then their expectations have been met. 
* If it is easier than they were anticipating, then they are **delighted**. 
* If it is harder, then they are **frustrated**.

### Illustrative examples of Delight:
* When Google Docs automatically corrects my spelling errors for me. 
* When Google Docs automatically completes my sentences for me. 
* When I use [Smart Chips](https://support.google.com/docs/answer/10710316?hl=en) in Google Docs to automatically create meeting notes, or an email, or show the actual name of a Google Doc when I paste a URL. 
* When I can respond to a Google Doc/Sheets/Slides comment from within Gmail
* When Smart Reply gives me the perfect response to a text that I can send with 1 click.
* When I’m trying to parallel park my car and the rear-view camera shows me the rear-view, lines that indicate my path and warnings when I get too close to the curb or another car. 
* When Google Photos [automatically creates albums of each of my kids](https://blog.google/products/photos/keep-your-favorite-photos-date-live-albums/), because it recognizes their faces. 
* When Google Photos [Magic Eraser](https://www.androidheadlines.com/how-to-use-magic-eraser-google-photos) enables me to easily remove unwanted people & objects from the background of my pictures.

### Illustrative counter-examples - Frustration:
* Every time I have to file an expense report in Concur, I am filled with dread. I know that I will likely feel stupid & confused at some point in the process. 
* Every time I have to log vacation, I mentally steel myself to use the slow and clunky Workday UI. 
* Every time I have to parallel park in a car without a backup camera, it stresses me out and takes multiple attempts. 

<img src='/assets/images/posts/delight_scale.png' alt="The Delight Scale"  style={{ marginLeft: 'auto', marginRight: 'auto', borderRadius: '10px'}} />

### Delight goes best with *flow*
There is a joy that comes from using a really good tool. Tools that are made well disappear, enabling you to enter [flow](https://en.wikipedia.org/wiki/Flow_(psychology)) and accomplish your goals without thinking about the tool and the ways it is getting in your way. You can feel when tools aren’t made well (Concur, Workday) - every unexpected behavior or rough edge slows you down and makes you think. 

### Delight eventually becomes *expected*
The first time you use a backup camera in a car, it feels like magic. By the hundredth time, it becomes an expectation. You miss it if it’s not there, but you are no longer delighted by it. So all delightful features become expectations over time, and the challenge is to see if you can continue to find ways to delight your users. 

Clearly framing your problem space helps with this. If you think of Docs as a word processor, you have limited the opportunities for delight. If you think of it as a collaborative tool for creating, refining and communicating ideas through words, then you have a lot more. 

### Delight doesn’t get people to use products *more*
The backup camera doesn’t make me want to parallel park more often, and Smart Chips don’t make me want to write more docs. But when I do have to parallel park or write a doc, they make me more effective and efficient at accomplishing my goals. 

### Delight isn’t *fun*
Asana has [Celebration Creatures](https://forum.asana.com/t/asana-s-celebration-creatures-got-a-glow-up/476849) that show on task completion. This is fun, but is not delight. You can add fun if you’re at least NEUTRAL on the continuum, which Asana is. If you try to add fun to a frustrating tool, good luck to you. 

### In Summary
<div style={{backgroundColor: 'violet', padding: '1rem'}}>
**Delight = Expectation of Difficulty - Actual Difficulty**
</div>


&nbsp;
> **Our goal with Ask Steve is simple: Be Delightful**\
> \
> We do this by making a bunch of tasks that would ordinarily take a lot of manual work unexpectedly easy. \
> \
> [Try it today!](https://chromewebstore.google.com/detail/ask-steve-do-your-job-bet/gldebcpkoojijledacjeboaehblhfbjg?hl=en&authuser=0)


What Does it Mean for a Product to be Delightful?

## We have just added the ability to connect to any LLM/model from Ask Steve!
* Instructions for connecting to a cloud model are [here](/docs/models)
* Instructions for connecting to a local model are [here](/docs/local-models)
* If you want to try a super fast model (free in beta), try out [Groq Cloud](/blog/news/groq-cloud)


Connect to Any LLM

## Groq is a company that makes chips customized to run LLMs, and they run *very* fast.
They have a cloud offering that is currently in free beta. Connect to it and see just how fast Steve can be!
1. Create a [Groq Cloud](https://console.groq.com/login) account
2. Create an API key
3. Go to the [Models](chrome-extension://gldebcpkoojijledacjeboaehblhfbjg/models.html) page in Ask Steve Settings, press `ADD NEW MODEL` and select `OpenAI GPT-3.5` from the menu. Groq Cloud's API is modeled on OpenAI's, so we'll use it as a starting point. Go ahead and paste in your Groq Cloud API Key.
4. Change the *Name* to `Groq Cloud`, the *URL* to `https://api.groq.com/openai/v1/chat/completions` and in the *Body* change *model* to `llama3-70b-8192`. You can see all the available models [here](https://console.groq.com/docs/models) and their rate limits [here](https://console.groq.com/settings/limits)
5. Press `TEST` to ensure it works, then `SAVE NEW MODEL` to save it. 
6. Steve is now SUPER FAST 🚀🚀🚀




Super Fast Responses with Groq Cloud


## Why Did We Build Ask Steve?
### Because LLMs like ChatGPT and Gemini give you superpowers to get things done better & faster:

> * <strong>Summarization:</strong> summarize this peer feedback about my report
> * <strong>Expansion:</strong> expand these bullets into nicer language
> * <strong>Rewriting:</strong> make this text more formal, or a specific reading level
> * <strong>Synthetic Data:</strong> blind this candidate profile to reduce bias
> * <strong>Authoring:</strong> write a custom prospect outreach letter
> * <strong>Translation:</strong> convert this from Spanish to English
> * <strong>Detail Extraction:</strong> what interview dates are being proposed in this email
> * <strong>Concept Extraction:</strong> what are the main points of this doc
> * <strong>Query Documents:</strong> query this page for information
> * <strong>Classification:</strong> sort these requests by category
> * <strong>Sentiment Analysis:</strong> is this a positive or negative comment
> * <strong>Toxicity Detection:</strong> is this an antisocial comment
> * <strong>Syntax:</strong> put this text into a specific format for another system                        
> * <strong>Idea Generation:</strong> suggest interview questions                        
> * <strong>Data Analysis:</strong> look at this table, what are the correlations 
> * <strong>Explain Something:</strong> explain this in simple terms

Studies like <a href="https://www.science.org/doi/10.1126/science.adh2586" target="_blank" rel="noopener noreferrer" className="text-blue-500 hover:text-blue-800">this one</a> have found
that ChatGPT substantially raised productivity on writing tasks: ***The average time taken decreased by 40% and output quality rose by 18%***. 
With the help of LLMs, people are able to do their jobs better and faster. 

### But bouncing back and forth between your work and an LLM suuuuuuucks

<a href="https://ar5iv.labs.arxiv.org/html/2303.10130" target="_blank" rel="noopener noreferrer" className="text-blue-500 hover:text-blue-800">Another recent study</a> has suggested that with <strong>access</strong> to
an LLM about ***15%*** of all worker tasks in the US could be completed significantly faster at the same level of quality.
But to get this benefit you&apos;re constantly copying and pasting and bouncing between windows. Most of the time it&apos;s too much work to bother with.                    

Which is why the same study showed that
when <strong>incorporating software and tooling</strong> built on top of LLMs, this share increases to between ***47-56%*** of all tasks.
Because the friction has been removed, <strong>your superpowers are unlocked</strong>. 


### Ask Steve unlocks the Power of Gemini & ChatGPT in any web page so you can get things done better & faster
💪 ChatGPT & Gemini give you **superpowers**, and **you should be using them more** to get tasks done better and faster.\
🤬 But constantly copying and pasting back and forth between a web page and ChatGPT is a **giant pain in the &#%**.\
⚡ Ask Steve **adds AI superpowers to every web page** so they're always at your fingertips. And it's **FREE!**

*<a href="https://chromewebstore.google.com/detail/ask-steve-do-your-job-bet/gldebcpkoojijledacjeboaehblhfbjg?hl=en&authuser=0" target="_blank"  className="text-blue-500 hover:text-blue-800">Install Ask Steve now to unlock **your** superpowers!</a>*


Why We Built Ask Steve


## How did Steve Get His Name?

In December of 2023, <a href="https://www.google.com/search?q=ethan+mollick" target="_blank">Professor Ethan Mollick</a> from Wharton posted the following on LinkedIn: 

<div style={{backgroundColor: 'lightgoldenrodyellow', padding: '1rem'}}>
<a href="https://www.linkedin.com/posts/emollick_the-thing-about-llms-that-make-them-unintuitive-activity-7125848204339732480-OeZL/" target="_blank"><img src='/assets/images/posts/steve-name.png' alt="How Steve Got His Name"  style={{ marginLeft: 'auto', marginRight: 'auto', borderRadius: '10px'}} /></a>
</div>

For some reason I found this really funny, but I also really liked the mental model of a first year grad student or intern as a way to think about an AI assistant that lived in your browser. Thus "Ask Steve" was born.

[DALL-E 3](https://openai.com/dall-e-3) drew Steve's picture (self-portrait?).

Professor Mollick has <a href="https://www.linkedin.com/in/emollick/recent-activity/all/" target="_blank">a lot of great posts on LinkedIn</a>, and you should follow him if you're interested in LLMs. He also has a good book out called <a href="https://www.penguinrandomhouse.com/books/741805/co-intelligence-by-ethan-mollick/" target="_blank">**Co-Intelligence: Living and Working with AI**</a>.

 \- rajat

How Steve Got His Name

Infinite Steves, at your service

How did Steve Get His Name?

Read more about
📰News

Experiment with LLMs in Google Sheets

What Does it Mean for a Product to be Delightful?

Connect to Any LLM

Super Fast Responses with Groq Cloud

Why We Built Ask Steve

How Steve Got His Name

Infinite Steves, at your service

How did Steve Get His Name?

Read more about 📰News

Experiment with LLMs in Google Sheets

What Does it Mean for a Product to be Delightful?

Connect to Any LLM

Super Fast Responses with Groq Cloud

Why We Built Ask Steve

Read more about
📰News