Today’s post is a guest post by friend of the show, Michael Fraser. Michael is an AI Evangelist and Innovation Strategist at Action Insight. He posts regularly on The Local Model where you can find more of his insights and work using Generative Models to change how companies think about art, workflows, and their future.

Being a leader means making decisions. Lots of them. And each decision is fueled by information. But there's so much information coming to us, it can be overwhelming and cause us to miss opportunities or fail to identify problems. The solution? Tools that can help us quickly transform our messy information into clear, easy-to-understand insights. It's all about making better decisions, faster.

Keep reading to find out how you can use simple approaches to get big results out of AI & GPT4.

GPT can apply any framework to your problem

GPT-4 is an adaptable AI model, capable of learning and applying a wide range of frameworks to various problems. The beauty of this model lies in its flexibility. No matter what framework you choose to use - whether it's a widely known one or a unique system you've devised - if you can clearly explain its structure and rules, GPT-4 can implement it. This allows for comprehensive, diverse problem-solving possibilities, catering to the specific needs and nuances of different challenges. Just remember, the key is clear and concise instruction. With that, GPT-4 can become a dynamic tool in your problem-solving arsenal.

Here are three examples to illustrate:

* Facts and Logic: This approach involves separating objective data (facts) from author's interpretation or reasoning (logic) in a given text.

* Edward De Bono's Six Thinking Hats: This method applies different 'hats' or perspectives (white, red, black, yellow, green, blue) to analyze an issue comprehensively.

* Structured Data: This framework involves turning unstructured information into a structured format like a spreadsheet, useful for data analysis and organization.

Facts and Logic Framework

Good writing often intertwines facts and logic to make compelling and persuasive arguments. As leaders, sometimes you need a clear separation to think through the matter carefully. Just ask GPT to separate them out for you.

First let’s generate an article:

Now let’s ask GPT to transform the same article into Facts and Logic:

The article may not be as compelling to read, but its much better at helping us clearly distinguish what’s going on in their argument. Remember - you can do this with any piece of text you can copy into GPT - news, research, reports - anything.

Edward De Bono’s Six Thinking Hats

Exploring different perspectives is a powerful way to overcome bias, ensuring we consider all angles before drawing conclusions or making decisions. However, even the most experienced leaders can find this challenging. That's where GPT-4 shines, with its ability to adopt diverse points of view based on the context it's given.

A particularly powerful tool for embracing these multiple perspectives is Edward De Bono's six thinking hats framework. My partner at Action/Insight, Kate Bowers, suggested this approach, and it has proven extremely effective. The beauty of using this model with GPT-4 is its ease of application. GPT-4 already understands the six thinking hats and their associated perspectives, allowing you to tap into the breadth of insights this framework offers at will.

By invoking the different 'hats', you can guide GPT-4 to scrutinize a problem from all possible angles, thus helping to combat inherent biases and promoting a more balanced understanding. In short, this tool provides a shortcut to comprehensive, insightful analysis, setting the stage for well-informed, holistic decision-making.

Let’s ask GPT to wear the blue hat and give some advice on this post:

Organize your data into a table or spreadsheet

One of the most powerful ways to transform your data is to have GPT take a long, rambling block of text (e.g. a transcript or customer call) and have it break it down into a structured table with whichever columns you want. Now you can quickly review all the information that was provided in a way that is clear and can support decision making going forward.

Here’s a simple example, recording some details about some children’s books:

This is a very powerful tool for collecting and capturing data that would otherwise be lost, and once you understand this capability you might start seeing a whole world of lost data in your organization. Get collecting!

AI is for leaders

More than any technology in recent memory, AI is a technology that directly benefits you as a leader, no matter what field you are in. It helps you organize your information, transform it into any presentation that suits your purposes, and ultimately leads you to better decisions.

There’s no better time than now to get started.

Thanks for reading Stable Discussion! Subscribe for free to receive new posts and support our work and those in our community.

AI is at its best when it removes boring tasks and lets us lean into the work we enjoy. Imagine being an athlete that only has to compete in competitions. They never have to practice or train. They get to focus on the work they train so hard for and make adjustments to the training done on their behalf. It promotes them to a coach that still gets to perform. AI elevates our role in the work we want to do.

Tedious tasks are "taxes" that often limit our interest in doing our work. Whether that's paperwork a doctor fills out or building test suites for software, every profession generally has some work that makes their work feel less enjoyable. These realities of the job limit our expectations as well as our motivation. We do fewer of the activities we want to do because there are tasks that actively block our ability to appreciate that work.

Progress Drives Motivation

Often we're very receptive to seeing something begin to come together. A project feels real for the first time when we can start to see what we have in mind. Painters get this when the composition is outlined for the first time, and they can begin to see where paint can start highlighting aspects of the work. Progress is a positive force to push our motivation.

AI offer an incredible opportunity to see progress early. Josh Larson mentioned in an interview for Midjourney Magazine that "… there's an entire possibility space of new art and design that can only be reached if we curate after the work is finished, rather than before. The only way to unlock that space is to create incredibly quickly." Larson emphasizes that by providing artists with the means to visualize their concepts in a near-complete state sooner, they could be encouraged to pursue more daring ideas, even those they initially believed might not succeed. Rapid creation is the key to unlocking this untapped potential.

That's not to say artists will use AI-generated art as a completed art piece. While Midjourney does an incredible job of arriving close to what an artist had in mind, the artistic vision requires much more specificity than AI can currently remember and act upon. Without that curation and refinement, we're left with a ghost of the artist's original vision and something that lacks a soul.

Over the last week, I've been working with a few friends to build a game map for a game set in Greece. I've dabbled in 3D game design and built a few games before but haven't tried anything as ambitious as what they're doing. I realized that AI made creating game maps much easier as I worked.

I'm using Unity to build out the maps, which already gives me a lot of great systems to use. But my work got a huge lift when I used Midjourney's tile prompts and a few web tools to generate textures that give a real feel of place to the world. Using very simple blocking for buildings and a few free assets, I was able to pull together something that helped motivate the whole team!

Speeding up Tasks

A new video game, Baldur's Gate 3, is releasing this week with over 170 hours of cinematics. That's twice the length of Game of Thrones. While this is an incredible undertaking, I wonder how much content was hand-crafted. Could Baldur's Gate 3 have benefited from using AI for portions of their massive set of content?

We've seen some incredible tools built for existing games that illuminate how AI might help. World of Warcraft is a Massively Multiplayer Online Role Playing Game (MMORPG) released in 2004 with an immersive world and some well-written quests. Unfortunately, that writing is usually ignored by the player base, who prefer to grind through the quests as fast as possible. Their avoidance of the plot is partially because the writing is displayed as large "walls" of text that the player must read.

A plugin for the game, VoiceOver, has fixed this problem by using AI to sample the few existing voice lines in the game to generate fully voiced quests. Coupled with a few other modifications to improve the interface, this game from 2004 now has an experience that feels like games created at least a decade later. The voices are incredible and give the game a new level of depth.

AI-generated voices could be a game changer for voice actors who don't need to sit in a studio for so many hours and can focus, instead, on the high points of the character — parts of the script that might need more attention. Many tedious lines can be auto-generated. AI can also help teams change the dialogue after recording to fit the setting better if something isn't working. They can do this without dragging the artist back into the studio.

We shouldn't overlook the financial impact of AI generation on voice artists. Since we can use an artist's voice to say a lot of new lines, there should be ways to compensate for that work and allow them to maintain some form of artistic control over the words spoken. These are conversations currently taking place during the Hollywood strikes but will likely be ongoing discussions in the months and years ahead.

We've been using AI voices on our publication here to voice out each post we publish and put that reading up on our podcast. I trained this AI with my voice, and while it doesn't sound quite like me, it's close enough for our purposes here.

If you are listening to this post: Hello! You're listening to an AI!

If you're reading, you can see a way to listen to the voiceover for this post at the top of this article.

An added benefit of having these posts read by an AI is that it can easily find spelling and grammatical errors. Having something read aloud to you as written surfaces editing issues you can't catch many other ways. These voices are a tremendous value add to us because we don't have a lot of resources to dedicate to our work, and we can reach an entirely new audience with voiceovers that articles may never reach.

Tedium Meet AI

AI doesn't get tired, but we must direct them toward the right tasks. Given the right direction, they can help us stay motivated by seeing our efforts come together more quickly. This motivation allows us to put the best of ourselves into our work and keeps our energies focused on the parts of the creative process that push us.

The world is changing, and while we can imagine many futures, it's exciting to see how people already use AIs today to create incredible things. William Gibson has one of my favourite quotes where he says: "The future is already here; it's just unevenly distributed." On Stable Discussion, we're helping to distribute that future, and we're excited to see how AI will continue to shape our world in the coming years. Stay tuned for more findings and perspectives on how AI works today and where we see opportunities for humans to use it to improve our lives.

Thanks for reading Stable Discussion! Subscribe to receive new posts and support our work.

You're late to class, didn't study, and must sit through a lecture on advanced mathematics (or a compliance training in an office, whichever you find most boring). You forgot your notebook and end up without any lecture notes. Now, an hour later, you're doing homework and try to remember what you learned.

You remember things at the beginning of the lecture. Topics connect to your previous learning quite easily. But then things get a bit murky... You've got something written down, but it looks wrong. You remember, at the end of the lecture, there are a few ways to check your results. Running through these checks, you can confirm that you're wrong, but you don't know why...

Without notes to reference, you're a bit lost. This is exactly the way AI feel when given too much data and no way to take notes!

The Challenge of Noteless AI

In a recent paper, Lost in the Middle - How Language Models Use Long Contexts, researchers found that when ChatGPT, and other AI, are given a lot of text they tend to forget or misremember some of it. Specifically, the information in the middle of the text. Similar to our lecture, the AI are less certain about information in the middle of any given data. It's uncanny that AI continue to act so human.

To help explain how this happened, the paper points to AI training. AI haven't been trained on larger chunks of information paired with accurate reflections over this information. Instead, they have been trained on shorter sections of text, often derived from how humans respond to information.

Where could we get this training data? Perfect recall isn't something we expect until we think about a machine "learning to read"” Then we want exact results. We want machines to do the drudgery of figuring everything out and just tell us what we need to know or what might have missed in our own reading. We don't want machines to read how we read or act like us!

But, hey. This is science, not science fiction. Because humans don't act like a prefect machine, we unfortunately don't have anything to model these machines after. It's becoming more obvious that, because these AIs are trained by humans, AI have even copied our faults.

How would we help a human get over these faults? How would you keep track of key pieces of information when interacting with a large and complex set of information?

Taking a Cue from Human Memory

In another Percy Liang paper titled, Generative Agents - Interactive Simulacra of Human Behaviour, the research team simulated a world full of AI-driven people. They woke up, ate, worked, socialized, planned, and slept.

A major challenge when creating this simulation was getting the citizens of this experiment to think like people. To solve this, the team constructed a complex set of memory systems. Short-term memory helped them remember where important items were, track immediate needs, and engage in conversations. Long-term memory was used to plan and remember important facts (among other things). These enabled the AI people in this simulation to act like people and even plan an ad hoc Valentine's Day parties with others.

It's a fascinating paper and I recommend you read it.

The AI were able to organize and stay on task when they were given some place to keep their memories stored. Memories tracked, at a high level, are summaries of what was happening around them. These memories are essentially notes on their environment!

Getting AI to Remember with Notes

This all leads us to some creative thinking about how to use AI to solve problems. While we'd love to treat the AI like machines, we need to assist their ability to remember and provide them with notes that direct their actions. Helping them take notes is essential to developing good AI workflows.

When we interact with ChatGPT, we're usually already keeping track of what the AI is thinking. They're using us as their long term memory, and we can remind them when they forget important information. In part, this is why chat interfaces are beneficial for this era of AI. They can sort of rely on us to perform well.

However, say we want an AI to read a large section of text and help build a summary of that text. Then we have a more challenging role to fill because the AI is going to misremember or forget information. We need to help give the AI note-taking capabilities.

If we split the text into smaller chunks, we can ask the AI to give us short summaries of each chunk. When we're finished, the AI can combine those pieces together to summarize what was covered. Just be careful that the pieces aren't too large when combined, otherwise we're right back where we were and the AI is being forgetful again.

Helping AI build plans isn't new, and there's already been a lot of research done to understand how to get AI to follow plans that help them accomplish tasks. A popular tool for building a plan is Chain-of-Thought (CoT) Prompting. We ask the AI to write out a plan for steps to take, which it will try to follow:

Question: Marty has 100 centimeters of ribbon that he must cut into 4 equal parts. Each of the cut parts must be divided into 5 equal parts. How long will each final cut be?

Answer: Let's think step by step.

Give this a try in ChatGPT, and you'll get back a sequence of steps and, usually, a pretty accurate result. We can leverage this pattern to get the AI to make plans for itself which can be referred back to later. These plans help AI to better organize its approach to solving problems.

The Path Ahead

While this only covers a few methods to help AI handle larger sets of text, many others methods exist and continue to be explored. While many believed that providing AI with a larger "context window" (the ability to be given more text) would solve many limitations, it seems that those limitations continue to exist. At this point, I won't use the longer context windows of these models until these memory problems are resolved with new modeling or better training data.

For the time being, the AI development community is going to be innovating around these limitations to engineer systems and solutions rather than just relying on LLMs to solve these problems. If you're doing something interesting to help AI better understand and remember larger sections of text, we'd love to hear about it!

We'll continue to post interesting ideas about how to use AI here and on our main site, Stable Discussion. Keep checking back to keep updated on our latest thinking around these ideas!

Thanks for reading Stable Discussion! Subscribe for free to receive new posts and support our work.

Follow us on Substack! https://blog.stablediscussion.com/

Stable Discussion now available on https://stablediscussion.com

We've posted new UI/UX examples and patterns for integrating AI into a wide variety of products.

Show Notes

Live Demo of Code Interpreter added to OpenAI (Our Dataset) Great Examples: https://twitter.com/emollick

AI Innovators & Devs Meetup in Toronto https://www.meetup.com/toronto-artificial-intelligence-meetup-group/

* Sponsored by Elevate (who hosted OpenAI when they came to Toronto)

* Organized by friend of the show, Dan Waldie

* Interview by AI App: https://www.producthunt.com/products/interviews-by-ai

Midjourney 5.2 dropped (allowing you to zoom out, panning) and Ben shares a Discord Trick for Midjourney

Google Deepmind - https://www.theverge.com/23778745/demis-hassabis-google-deepmind-ai-alphafold-risks

Super Intelligence and how to wrap your head around the discussions taking place https://openai.com/blog/introducing-superalignment

PI Chat App - https://heypi.com/talk

In our discussions with many different people, how to get started has come up quite a bit. Many of these people aren’t working in technology and haven’t programmed before. When they try to use AI it doesn’t work for them and the results aren’t all that helpful. This gave us the idea of putting together a short guide for anyone who identifies with these problems and would like some helpful pointers to get started.

Getting Setup

I think the easiest place to start is by getting an account setup to experiment with ChatGPT. You can go on their site, create an account, and get ready in just a few minutes. They also have a mobile app that you can download which may be easier if you don’t usually use a computer or laptop.

These same ideas will carry through to other systems like Midjourney, DALL·E, or Bard. We just find ChatGPT is one of the easier interfaces to use when getting started.

For our examples we’re going to get ChatGPT to write a letter to our friend that updates them on the latest news in our life and breaks the bad news to them that we won’t be able to visit this summer in as kindly as possible. I highly recommend you follow along on ChatGPT.

Don’t Look for A Single Right Answer

The most common challenge we find people struggle with is looking for the “right answer” to using AI. If you did well in school or took standardized tests you might be geared toward finding one correct answer to how you should use AI. Unfortunately there isn’t just one correct answer. Many different questions can get you the same result. It’s similar to asking your friend a question, they’ll likely give you the same response even if you ask your friend a question in a few different ways.

Instead, using AI is a lot like trying to ask a question that the AI can understand best to answer your question. If you ask someone “What is the capital?”, they might not know what you’re talking about. The AI is the same, you’ll need to be a bit more specific.

The examples we give below are some good ways to get started but you might find that for what you’re trying to do there’s another better way that works for you. Don’t get too attached and try to have fun finding new ways to get what you want back from the AI.

Initial Prompt

Prompts are the questions or queries that we send to a service like ChatGPT. For our initial prompt, we’re going to go ahead and get started writing a letter to our friend using the following:

“Write a letter to my friend, Paul, telling him I got a promotion to Manager at my company, my dog died, and I won’t be able to visit them this summer

If you put this into ChatGPT you’ll notice that it does a great job getting a letter created but it likely isn’t at all what you imagined. It probably doesn’t sound like you, it likely includes details that we didn’t give, and it might not be very interesting to read.

You might be tempted to then just say:

“Make it better”

The result is a letter that likely uses a more complex vocabulary, might have fixes some grammatical issues, and might be a little longer but doesn’t seem to address any of our issues. It’s at this point most people give up and call it a day. But if we push just a little further I think ChatGPT will surprise you!

You can find my chat example here

Adding Specifics

So we can immediately solve some of the problems we had just by providing a few more details. Let’s start a new chat and add some details that can be used to fix some of the challenges we’re having:

“Write a casual letter to my friend, Paul. Tell him I got a promotion to Manager at my company, Nintendi, because of my hard work over the last year. My dog, Mari, died from old age and my family is still grieving. I won’t be able to visit him this summer due to the new responsibilities at work.”

Testing this out we’ll immediately see much better results. Describing the tone of letter (casual) helps a lot and can definitely help ChatGPT know what kind of language I want to use. The other specifics help to fill the letter with details that make the letter’s contents a lot closer to what we were hoping for.

You can find this chat example here

If you imagine talking to your friend and asking them to write a letter for you, these are the same details you’d need to provide them. It helps a bit to think of an AI as just an assistant that is trying to figure out what you want and can only use what you give it to help you.

Adding specifics to your prompt might be just enough to get what you’re looking for especially if the letter isn’t very personal. However, this letter still doesn’t have the personal touch you might like it to have. Let’s look at some other ideas for how to make our examples just a little better.

Adding an Example

If we were to hire someone to write letters for us, something we might do is provide them with some good examples of how to do it. That would help them understand what we’re looking for and when we gave them feedback they could continue to learn with those examples as a foundation. For this improvement, we’re going to give the AI the same thing: an example of a letter for them to use.

Write a casual letter to my friend, Paul. Tell him I got a promotion to Manager at my company, Nintendi, because of my hard work over the last year. My dog, Mari, died from old age and my family is still grieving. I won’t be able to visit him this summer due to the new responsibilities at work.

Use the following example as a reference for my writing style:

Hey!

It’s been a crazy and amazing week. I’m at the Seville airport, getting ready to board flight TP1105 to Lisboa before TP259 to Toronto. It’s a really small airport here and security took me 5 minutes!

My favourite memory is this dinner of the amazing view of Ronda with local wine, salmon tartar, croquettes, artichokes and Iberian ham. Amazing views, delicious pairings, and incredible company.

Looking forward to being home now. 9.5 hours in the air and I’m there.

Love you,

Ben

Adding this example, we see the style of the result change to be more inline with the way that we talk and we see our formatting style in the resulting letter. However, you still might not be too happy with the result. Don’t worry, there are a couple other ideas we can try out!

You can see the result of adding an example here.

There are a number of other ways to provide an example that might provide even better results in your testing. One prompt engineering technique is called a **One-Shot Prompt** where we give it an example of a prompt and a resulting letter. You can see an example of this here.

Adding Rules

One thing AI are great at is following rules that we provide. They can factor in a large number of factors that are hard for us to track. Because of this, rules are a great way for us to control their behavior and get better results. Let’s add some rules that the AI should follow when they help us write our letter.

Write a casual letter to my friend, Paul. Tell him I got a promotion to Manager at my company, Nintendi, because of my hard work over the last year. My dog, Mari, died from old age and my family is still grieving. I won’t be able to visit him this summer due to the new responsibilities at work.

Follow these rules when writing the letter:

* Don’t be too emotive and stay somewhat reserved

* Try to start and end with good news

* Don’t add any details I didn’t include

* End with “wishing you the best”

Use the following example as a reference for my writing style:

Hey!

It’s been a crazy and amazing week. I’m at the Seville airport, getting ready to board flight TP1105 to Lisboa before TP259 to Toronto. It’s a really small airport here and security took me 5 minutes!

My favourite memory is this dinner of the amazing view of Ronda with local wine, salmon tartar, croquettes, artichokes and Iberian ham. Amazing views, delicious pairings, and incredible company.

Looking forward to being home now. 9.5 hours in the air and I’m there.

Love you,

Ben

These rules help to reinforce our tone, style, and help keep the AI from adding details that we don’t want added to our letter. Rules are easy to add and remove to see different results in the letter’s style and offer great direction for the AI. Sometimes you can jump right to adding rules for a great result, but it very much depends on the task you’re trying to do.

While it may not be perfect, this really gets us quite close to a letter that we’d be happy to send. After a few small tweaks and edits you have something that you can send to Paul and a great template you might use for future letters!

You can see the final result here!

Turning it off and on again

Similar to many technologies, AI sometimes get a little bit overwhelmed and need to be restarted. Consider opening a new chat conversation with ChatGPT after a while.

You can keep a conversation going for a long time in ChatGPT but you will start to lose control of the conversation and the AI might start behaving strangely. This is especially true If you’re trying to do a lot of different things in a single chat. Previous responses may leak into other questions. For example if I’m writing another letter to a friend in the same chat that we were chatting with Paul, it may bring up my promotion even through I didn’t say that it should just because the AI knew about it from earlier in the conversation.

Use a more powerful model

Another great way to get better results is to use a more advanced version of the AI. If you’re struggling to get GPT-3.5 to return something, using GPT-4 could be all you need for the AI to understand what you want. Later models are a little slower but they’re smarter and able to better understand complex requests. The GPT-4 announcement gives some great examples for how much the model improved over time.

Takeaways

Prompting can be a frustrating experience as we try to get an AI to understand just what we want. When struggling, refer back to these improvements that can help AI better understand what to do and how to produce something a little closer to what you’re looking for. AI act a lot like assistants and we need to provide them with the details that help them to do the jobs that we give them. Without that feedback, they’ll continue to struggle to understand how to make their results any better.

Thanks for reading Stable Discussion! Subscribe for free to receive new posts and support our work.

Prompt engineering doesn’t provide a repeatable way to use AI models.

You’ve likely seen courses sold with thousands of prompts and industry professionals discussing amazing prompt engineering techniques. Technologists excited about ChatGPT have spent a lot of time creating engineered prompts. With so much excitement, how is it possible that prompt engineering isn’t providing a repeatable practice or series of techniques we can rely on?

To understand how Prompt Engineering is coming up short, we need to start with the AI research discussed in one of the latest episodes of our podcast. Researchers are looking into ways of using AI models that get the best results. They do rigorous testing, collect their results, and publish their findings for others to read and discuss.

For example, one study showed that posing a question as:

Question:

Answer:

gives you better results than using

Q:

A:

Examples like this are perfect for improving prompts to get better results. Looking at this research, you can begin to piece together a general approach for prompting. Sometimes this can be just enough to overcome challenges you’ve encountered with AI models and get the result you’re looking for. But the problem is these models are changing every month, and this research may not be relevant to a later version.

Changes to models may not seem like a big deal, but they fundamentally change how AIs think. Worse, we only have a small glimpse into how these models are changing. OpenAI, in particular, changed its models significantly last week to add functions as a new capability. While it’s hard to understand how much prompts will be impacted by this change, we can see the impact of these sorts of changes clearly when we look at image generation.

I was a huge fan of DALL·E when it was first released, but my interest quickly soured after a month of using it. Updates completely changed how the model generated images, and many prompts that previously made amazing art for me suddenly had really poor results. I was devastated and eventually moved on to Midjourney, which, while changing and releasing multiple versions, continues to support older versions that consistently produce expected results.

While it’s tempting to stick with a version of an AI model that works for you, you’ll eventually miss out on new capabilities. If we look at the results for the same Midjourney image generation below, we can see a dramatic shift in style. This upsets our ability to prompt for the same thing, but as new versions are released, new appealing aspects of the presentation are added.

When we think of engineering, we might think of bridge construction. An engineer applies principles, standards, and techniques to work on a complex problem and create an expected output: a bridge. But that thinking starts to break down for AI because the tools we build with are changing, and the result is often a loose solution rather than a fixed statistical answer. It’s like building a bridge that has to handle different physics each time someone drives over it.

So what do we do? How do we handle the change that will continue to happen as new AI versions are released?

We need to do research. Every time a new model is released, we must re-validate what we know. Rather than focus on specific skills or prompt methods, we must learn from solutions to problems.

For example: How do you get an AI to give a more direct or less direct answer? Knowing some good ways to change the answer returned gives you a better idea of how to direct a model. Record what you tried, try to measure its effectiveness, and see if you can reproduce it when the next version is released.

This trial and error process is the journey of an early adopter. There won’t be any best practices for quite a while, so try not to focus too much on one way of doing things and keep flexible. Many changes will happen, and there will be many new ways to build things and converse with AIs. If you’re brave enough, there’s still an amazing space to explore with AIs, but don’t be surprised if the road gets rougher the further you go.

Thanks for reading Stable Discussion! Subscribe for free to receive new posts and support our continued work.

Topics discussed in this episode include:

Latest AI News

* Sam Altman visits Toronto on his world tour: https://open.substack.com/pub/stablediscussion/p/discussions-about-chatgpt-with-sam

* Google I/O Announcements

* PALM2: https://blog.google/technology/ai/google-palm-2-ai-large-language-model/

* Anthropic 100K context window: https://www.anthropic.com/index/100k-context-windows

* Geoffry Hinton interfview: https://www.technologyreview.com/2023/05/02/1072528/geoffrey-hinton-google-why-scared-ai/

* Drag your GAN paper: https://arxiv.org/abs/2305.10973

* Tree of Thoughts (ToT): https://arxiv.org/pdf/2305.10601.pdf

Ethics of AI

* What is the future like for artists and what do we think about rights and usage of AI tooling?

* What does the future of jobs look like and where should we be looking for people that may be losing the current way they work?

* Do we know what a good future looks like if it doesn’t look like the world we live in today?

* Geoffrey Hinton’s warnings about the potential future of AI

Misc References:

* Simulated Towns Research paper: https://arxiv.org/abs/2304.03442

* Fall by Neil Stevenson: https://en.wikipedia.org/wiki/Fall;_or,_Dodge_in_Hell

* Sapiens (book): https://en.wikipedia.org/wiki/Sapiens:_A_Brief_History_of_Humankind

* Lex Fridman Interview with Sam Altman: https://www.youtube.com/watch?v=L_Guz73e6fw

Stable Discussion was able to attend a special event in Toronto with OpenAI’s CEO, Sam Altman. Sam is on his world listening tour to gain insight into how users are engaging with the AI models his company has developed. Toronto was the first stop on the tour.

In a fireside chat, I was able to ask a single question to Sam:

Ben: “Are you having challenges re-engaging customers who have tried and failed to accomplish a given task with a GPT model? Others besides early adopters fall into this category. What does re-engagement look like for an AI model, and have you put any thought to that problem?”

Sam: "We really haven't seen it as a problem for us. As we release new versions, there's enough buzz and sharing between our users that we haven't seen a falloff in user engagement."

While this seems like a fairly uninformative answer, it shows us a glimpse of how OpenAI is operating. We can assume how they’re thinking about the millions of users using their ChatGPT service. The answer also gives us insight into the continued focus of OpenAI. Given the context of the rest of the event and the focus of the listening tour, we get a pretty good idea of what OpenAI is planning for the future.

Re-engagement on AI Models

ChatGPT has been hailed as the fastest growing product on the market. But Sam has said multiple times that it wasn’t really intended to be a product. They thought it was an amazing tool, but the intent was research and not the creation of a hot new product. Now many think of OpenAI as a product company due to the wild success of ChatGPT.

If ChatGPT is thought of as a product, there are at least two important factors for its success: acquisition of new users and retention of existing users. Acquisition has been no problem, to the surprise of Sam, former Y Combination President. Marketing for the product has been easy with so much word-of-mouth and organic content generated from users. Retention is easy, for now.

Retention measures how frequently a user comes back to use a service or platform. ChatGPT certainly doesn’t have a retention issue right now because so many users are trying out new things. But how often do users try something that failed previously? That’s where re-engaging users comes into play. Re-engagement is how easily a product can get a user to attempt a desired task again.

If you have a home assistant (Google Home, Siri, Alexa) you probably only use a handful of commands despite that device being capable of so many things. There aren’t any clues about what more you can do with the device, aside from the ads you sometimes get in the form of “Have you tried X?” When you try something new and the device fails to give you a result that you’re looking for, you’re unlikely to try giving it that task again.

These negative feedback loops drive customers away from trying new things and discovering new features of the product. It isn’t until a competitor appears, or a product focused on a specifically challenging task appears, that customers will attempt that task again. Until that point, the only way you can try to get the customer re-engaging with the task is through advertisement, marketing, or training.

While I agree with Sam’s answer that re-engagment isn’t a problem yet, it’s going to be a growing problem for a product like ChatGPT. Over time people will be less inclined to try out the latest model, and over time they’ll learn not to try certain tasks with the model. This will hamper the product’s ability to grow and maybe even interfere with the research that is core to OpenAI’s efforts.

OpenAI the Research Company

Research is the primary focus of OpenAI, closely followed by policy setting around AI. Sam Altman has been meeting with world leaders to discuss the changes that advanced AI will bring. Because of their significant partnership with Microsoft, OpenAI doesn’t need to worry about being commercial or focusing on ChatGPT as a product. In fact, their hit product wasn’t talked about very much in the fireside discussion.

However, their research is greatly improved by the success of their product. Reinforcement Learning from Human Feedback (RLHF) is one of the most effective ways to train an AI model. You’ve probably seen this a number of times when websites ask you to rate how well a search worked or if the data you’re seeing is what you expected. Fuelled by the world’s attention, OpenAI can use that feedback to better train future models.

Services focused on coding, like GitHub Copilot or Repl.it, may draw users away from asking ChatGPT coding questions, leading to it having a harder time improving its ability to create a model that best incorporates code. Companies focused on research will eventually compete not just for raw attention but specialized attention on specific tasks, and that’s where re-engagement will begin to become an issue.

Where ChatGPT will Grow

ChatGPT isn’t going anywhere, and it’s unlikely to be unseated with the entire world marketing its incredible capabilities which continue to grow with each new release. Plugins, released recently, enables some incredible new capabilities as well as some scary security concerns. OpenAI is owning the discussion on the future of AI and are taking advantage of their success to be the first to lobby government entities on what the future should hold for AI.

However, without a product focus, ChatGPT will begin to lose its advantage of being first to customers and see other competitors enter the market and begin converting its customers over to their services. Because OpenAI is focused on research, they might be ok with ChatGPT being unseated as long as they remain involved somehow. Partnering with Microsoft and other companies to integrate their AI models into their services may be enough for them to further their research in meaningful ways.

While there’s a lot of excitement and speculation for the future of the ChatGPT product, the reality is murky and full of questions that have yet to be answered. I’m looking forward to seeing more AI events, as it was great to connect with the community here in Toronto and learn what others are thinking about when it comes to AI capabilities. If you’re interested in discussing this post or potential product AI, comment below or reach out to us via email.

Thanks for reading Stable Discussion! Subscribe for free to receive new posts and support our work.

You can watch the fireside discussion on YouTube but unfortunately the Q&A section was cut from the recording.

Note tooling is something that I've been obsessing over for years. I've never thought of myself as a particularly good note taker but, as I started working more, I realized that I needed to offload information from my brain. It's hard to keep track of all the things you have on the go, your passion projects, and even things you might be trying to learn about outside of life. It's easy to feel overwhelmed with all the demands on your mind.

After filling many physical notebooks, Evernotes, One Notes, Bears, Notions, and Roams I've finally settled on Obsidian as my tool of choice for handling notes. This is in no small part because of its ability to facilitate a Second Brain note-taking system.

A Second Brain is an approach to help capture, organize, and retrieve information effectively. There are many different techniques for taking notes to build a Second Brain, including the famous Zettelkasten Method of note-taking. These systems reduce mental overload, improving learning, reinforcing memory, and making you a better problem solver and decision maker.

Obsidian is a great tool for any of these Second Brain techniques because of its ability to create links and references between content, structure content, build note-taking workflows, and quickly search the notes. Using it as a Second Brain creates a centralized location where you can keep track of all of your information. Having all of your knowledge in one place constructs a "Data Lake" you can analyze and search to help surface ideas, patterns, and generate new content.

Many writers use these pools of knowledge to "slow burn" their content creation, which leads to better research and more complete ideas being published. They start writing multiple articles at the same time, collecting research bit by bit, so that they can benefit from the serendipitous discovery of new information as they write. It basically avoids that feeling where you just finished publishing something and later realized that there's one or two things you wish you'd added.

I've been looking into using my Second Brain as a source for an AI chatbot. While access to the newest version of ChatGPT's Plugins will allow you to query against documents (like PDFs), I want to use hundreds of documents including highlights from articles, diagrams, and even referencing images. Anthropic announced a 100k token limit for their AI, Claude, this week but it’ll be interesting to see what the research shows about the AI’s ability to focus with a larger context space.

After I created a D&D chat bot to converse with the rules of a very old version of the game, I've tried to create the same thing, but for my notes. I dabbled at building a system to do this myself, but then I came across a plugin for Obsidian written by wfhbrian called Smart Connections. The plugin adds an AI-powered chat client right into Obsidian! It's great when I don't have to waste time building because someone else has already beat me to it! 😄

I can ask the Smart Connections chat questions, and it will reference notes that I've added, or I can directly reference it specific notes for it to work with. The comparison tool that comes alongside the chat can also show which notes are similar to help create better connections between notes. This is great for creating maps over your Obsidian notes, which could come together into digital gardens.

You'll need a bit of technical knowledge to get Smart Connections setup. I've linked to Github where you can refer to the setup procedure, but you'll need to get an OpenAI API key setup, and you should be ok with spending a little money to index all of your notes. My ~500 notes took less than $1 to process, and then a cent or two every time I converse with it.

The ability to use AI with my Obsidian Second Brain has grounded my belief that this is the best note-taking system around. Being able to use your own Second Brain as a source is incredible. You can chat with everything you've learned and ask questions that maybe only you would know! Better yet, the answer will likely be familiar to you or written in a way that's clear to you because the explanations were written by you! Tools like Gitbook Lens are already realizing that chat build against Knowledge Management Systems is a revolutionary idea.

There's a future where we have the power to use our own personal data to improve our lives. Rather than selling data to create services like Google or Facebook, AI offer a way to use that data without needing any special training to be better able to leverage that data. With AI running on-device, we won't need to worry about an AI system keeping our confidential information. This will empower us to fully leverage what we know and our own strengths to be able to improve our lives without the need to sacrifice our data to do it.

While Smart Connections isn't perfect, it's a great first step in being able to create a generic documentation search system for your Second Brain. AI solutions can always be improved and optimized by adopting a less generalized approach. We'd love to work together if you're interested in creating a more optimized documentation system for your data, or are interested in what other capabilities you'd like to incorporate.

Thanks for reading Stable Discussion! Subscribe for free to receive new posts and support our work.

In this episode, we explore recent AI research and its implications, from AutoGPT to AI playing The Sims. Join us as we discuss these fascinating developments and better understand the limitations these systems have today.

Topics discussed in this episode:

* AutoGPT: Architecting Out Language Models

* A proof of concept for the future of AI

* GitHub link: https://github.com/Significant-Gravitas/Auto-GPT

* Research Papers as a Source of Insight

* Gain a deeper understanding of AI's capabilities and limitations

* Get inspired by innovative ideas in the AI research community

* Simulated Towns & Emergent Events in GPT

* AI playing The Sims reveals both the power and limitations of language models

* Research paper: https://arxiv.org/abs/2304.03442

* AI and Small Business

* How small businesses can benefit from incorporating AI into their operations

* AI Adversarial Attacks

* How these attacks differ from traditional cyberattacks

* Video on adversarial attacks

Episode resources:

* Podcast generated outline: https://sharegpt.com/c/9Ppb1Kq

* Logspace AI platform: https://github.com/logspace-ai/langflow

* Maverick app

* AI enthusiasts to follow on Twitter:

* https://twitter.com/_akhaliq

* https://twitter.com/emollick

* ShareGPT tool

* AI-generated pizza commercial

Tune in for an engaging discussion on the latest AI research and its potential impact on our lives. Join the conversation on our podcast's social media channels and share your thoughts!