In this edition:
My thoughts
Notable headlines
History: Microsoft Clippy (was this an AI mistake?)
Coding, Tools, and Experiments
How does this GPT thing work anyway?
My Thoughts
We’ve all seen the hysteria in the media—fear about layoffs, the destruction of society and so much more at the hands of AI. I even saw a few tweets last week about comparisons to the new Oppenheimer movie and how AI could be just like the atomic bomb.
The truth is, what happens next is unknown. There is much to be done, and it’s up to use to work together and to make it happen.
It’s easy to fear what we don’t understand. It’s much harder to understand something if we don’t know it. One of my goals with this newsletter is to inform and connect. At the risk of taking some of the magic out of AI, if you understand how it works, you’ll be much more likely to matter better decisions about how to use it.
Speaking of how to use things, “User Experience” is vital to the success of the technology we use. We’ll take a bit more about this in a few paragraphs, but here’s the thing… if I can help you understand even a tiny bit better how to leverage this technology appropriately, and we can solve real problems with it, we ALL win.
As a student of emerging technology, I love studying what’s new and understanding how to use it. Even when completing my graduate degree, I didn’t fully appreciate my passion for it. That changed when I enrolled in BCIS 5650 - Emerging Information Technologies in Fall 2007.
It all seems so long ago, but there are three key realizations I’ve had as a direct result of this course:
1.) The historical analysis of emerging technology provides vital context and insights to properly frame the thinking around an emerging technology by using hindsight to look back. Breaking down the successes and failures can help identify critical patterns we can use to our advantage for the next tech.
2.) The market analysis and economic, social, and government forces that provide external impacts on using and adopting emerging technology. It’s not always just that tech is good; it must also be needed. We use the term “killer app” (vital applications of the technology that drive change)
3.) In 2007, we studied cutting-edge future technology. Here are a few of the ‘future technologies’ we examined: nanotechnology, Augmented Reality, and Neural Networks. — These were cutting edge 16 years ago; change can be fast AND slow. It just goes to show even with tools like the “Gartner Hype Wave” and “Forrester Wave”, its difficult, nigh impossible to predict exactly when a tech breaks through and reaches wide adoption… but we do know that the pace of adoption is getting faster. The S-curve of tech adoptions has pretty reliably become steeper with each decade.
It is true; genuinely awful and disastrous things can happen when we use technology that we do not fully appreciate or understand. However, we can also change the world in incredible and positive ways. So with that in mind, let us seek to understand and then correctly apply and use this fantastic new technology to advance humankind rather than to destroy, diminish or harm it. It’s a big undertaking, but we all want to improve the world, so let’s go!
Don't miss out if you’re reading this for the first time and haven’t signed up.
Notable headlines
Dell Teams up with Nvidia for Enterprise Generative AI
The infrastructure and services of the Dell Generative AI solutions portfolio are being co-delivered by Dell with Nvidia.
Chhabra said that the new Dell Validated Design for Generative AI with Nvidia offering is the installation of Project Helix which was announced at Dell Technologies World in May. The first release is not a general offering for all AI, it is focused on inferencing use cases.
This will not be the only hardware partner looking to bring Nvidia solutions to the enterprise through parternship. Cloud is great, but there are still places where hardware has to be in place. I’m glad to see this partnership and I look forward to seeing more like it in the future.
Stability AI Releases “FreeWilly1” and “FreeWilly2”. and Stable Diffusion Xl
Stability AI has announced FreeWilly1 and its successor FreeWilly2. These are completely new and open access, LLMs. FreeWilly1 uses Meta’s LLaMA 65B foundation model with some additional fine tuning they built at Stability AI using the Alpaca. FreeWilly2 leverages the LLaMA 2 70B foundation model. There has been benchmarks that indicate that it outperforms GPT-3.5 in same cases. Stable Diffusion XL is “the next iteration in the evolution of text-to-image generation models” and produces “images of high quality in virtually any art style and is the best open model for photorealism“.
I’m particularly excited about models like that that will allow completely private or niche off-cloud solutions.I’ll be playing SDXL later this week. I can’t wait to get it installed and going along side my now legacy SD setup.
Anthropic Releases Claude 2 with 100,000 Token Capability
As we work to improve both the performance and safety of our models, we have increased the length of Claude’s input and output. Users can input up to 100K tokens in each prompt, which means that Claude can work over hundreds of pages of technical documentation or even a book. Claude can now also write longer documents - from memos to letters to stories up to a few thousand tokens - all in one go.
If you like to interrogate your documents, and uploading that file to a beta service doesn’t give you pause, this is an incredible leap forward. With 100K tokens, you can upload entire novels! Anthropic is also claiming that their new release is dramatically better at coding. Go give a spin, it’s free.
Apple will probably move ahead with its AI plans much more deliberately than what we’ve seen with Google, Microsoft, or many others that can’t seem to get generative AI into their products fast enough. Apple sources reportedly told Bloomberg last week that the company will have a “significant AI-related announcement” next year.
Apple is a noted exception in the big tech companies releasing Generative AI products or betas. Because of their market positioning, whatever they do will have an impact so it’s worth watching.
More than half (54.6%) of organizations are are experimenting with generative artificial intelligence (gen AI), while a few (18.2%) are already implementing it into their operations, but only a few (18.2%) expect to spend more on the technology in the year ahead, according to the early results of a new survey of global executives in data, IT, AI, security, and marketing, conducted by VentureBeat ahead of the recently concluded VB Transform 2023 Conference in San Francisco.
Vendors looking to sell GenAI solutions might find it a hard sell, but there is an interesting dicotomy presented in the store that calls out the “limited talent and/or resources for gen AI” adoption. This is yet another sign for huge opportunities from a skills perspective.
History: Microsoft Clippy
Microsoft's Clippy was the animated office assistant you might remember from using the Word product in the late 1990s.
Using a kind of artificial intelligence, this tool did not endear itself to users. Speaking from experience, the biggest problem that Clippy presented was the abrupt and obtrusive way the device had of popping into your document in a rather inconvenient way. I think it’s safe to say that user expectations were not met. It didn’t nearly me as much as some of my peers in the “business computing” class I took in high school, but everyone has their own tolerance for computers.
Clippy came to us with Office 97. The intended purpose was to guide users through tasks by providing suggestions and shortcuts based on the perceived actions of the user. The trouble was users didn’t receive it as a welcome helper. In my experience, it would surface as an interruption more than a help. I’m not the only user who found it less than helpful. Clippy has found itself among the “50 worst inventions ever” in Time magazine.
Here’s where things get interesting. The problem with Clippy wasn’t about how it functioned or how much it helped. I think it’s more of a story about user interaction, user experience, and how we feel about how a tool presents to us when trying to get work done. The negative sentiment towards Clippy reveals the importance of understanding user context and preferences. It also might indicate that humans don’t like unnecessary interruptions (especially by computers), and they expect an AI tool to provide value-added services, not just generic assistance.
I found the letter writing help to work, but when I used it most often, it wasn’t the immense help I expected. (I also didn’t write many letters in high school, so maybe it was more me than Clippy).
Whether or not including Clippy was a success, there is more to the story. Not so long ago, Microsoft decided to revive clipping and used the character to drive a social media campaign. Microsft was setting out to introduce a series of redesigned emojis across Microsoft 365 apps, and they used Clippy as a part of a larger initiative to promote some new features and aspects of their Microsoft 365 product. Clippy’s reemergence campaign drove a high volume of engagement on Microsoft’s Twitter and Instagram.
One of the significant parts of the story is that Microsoft used this mistake to drive new capabilities and approach the inclusion of new AI tools into their existing products. Multiple new AI-powered assistants (such as Copilot for Microsoft 365 and my favorite, Github Copilot) have been designed to assist people in generating content in a less intrusive way.
Microsoft has found ways to integrate CoPilot as an assistant in the sidebar, summoned by Office users to generate text in documents. Github Copilot works with a keyboard shortcut and only impacts the UI when explicitly asked to provide additional suggested code and completions.
Today’s integration designs are far less intrusive than Clippy and arguably provide a much better user experience.
Clippy serves as an essential lesson in the development of not only new AI solutions but any solution delivered to a user anywhere. To properly provide a practical and valuable experience, we must invest in understanding user needs, offering context-aware assistance, and maintaining a non-intrusive presence.
Coding, Tools, and Experiments
This week I was pleasantly surprised to read an excellent share by Stephen Hood on Mozilla Hacks about an experience building a chatbot at a Mozilla Hackathon event recently. It was an enjoyable read and highlighted some great thoughts about building capabilities and our myriad trade-off decisions when building applications. Whether it’s the first time, a short time, or a learning experience, seeing how people think when making a solution is excellent.
Local LLM Comparison Tool
If you’re like me, you’ve been trying out all the new local models that pop-up. If you find yourself wanting to compare them, this Github repo might be just the thing!
Command Line Interpreter LLM or PowerShellAI
If you live in the command line like I do, sometimes pulling up ChatGPT, Bard or Claude2 is just too much hassle. Bring your GPT of choice into the command line with one of these repos.
I’m fascinated by the continued developments that Meta has released. Check out the LLaMA 2 resources. If you haven’t had a chance to work with it yet, it provides a nice guide on what matters and provides all the essentials regardless of how you’d like to conduct your experiments.
Text Generation UI
Hugging face is a great resource and it’s getting the point where there are so many ways to stack the work with and not re-invent the wheel. This repo provides a reuseable text generation UI for use with many models. This was a find that came courtesy of the Mozilla Hackathon article at the beginning of this section.
How does this GPT thing work anyway?
What is ChatGPT doing, and why does it work?
Stephen Wolfram is an expert I rely upon. I was an early adopter of his excellent WolframAlpha IOS app a few years ago and I still use it regularly, right next to ChatGPT. If you follow me on LinkedIn, you might have seen my post recommending Stephen’s write-up on ChatGPT, but I feel it’s more relevant than ever, as every major tech company is getting into the LLM game. I was reminded of this earlier this month when Stephen’s book of the same title, What is ChatGPT Doing and Why Does it Work?, popped up on my Kindle Unlimited subscription service… I did a double-take and realized the book is available to anyone through his site.
It includes details images that help understand how it works. While it’s not technically architecture, building anything that relies on the LLMS and Generative capabilities is a must-read. It adds some deeper information that I didn’t get from other training on the topic. It’s quite a lengthy read, but it’s great for a weekend deep dive.
One More thing…
Why am I doing all this anyway?
When I started the newsletter, I felt compelled to get better at sharing what I learned every day with a larger audience. I started AI Mistakes because every mistake carries the seed of a new idea, solution, or a better way of doing things, and I discovered that my fear of mistakes, or rather my unease at sharing them publicly, was limiting my ability to share, educate and make the world a better place.
So, with the name and this publication, every mistake is an opportunity in disguise. Oddly enough, I've made plenty of mistakes in my tech career, but now it’s time to convert those mistakes into real value.
I want to help others to grow and understand. This month I filed the paper and made AI Mistakes an official thing. And in September, I will be launching publicly and introducing what I hope will become a celebration of mistakes and an entirely new opportunity for us to work together and improve the world.
My goal with founding this new thing is to open the door to new possibilities.
I've spent literal decades mastering the critical aspects of technology, and I cannot bear not to share what I have learned in a meaningful way.
So I do have a sincere and important request:
As I prepare to launch, I would love to know:
What's your biggest challenge?
What's your dream that you've kept to yourself?
What’s your vision for the future that is only in your mind?
I'm doing this not just because I know I'm worthy of it but because I know YOU are worthy of it.
Reply directly to this newsletter email. It comes to me directly, and I will respond to every one of you who sends me a message.
No pressure, but I want to help… and helping each other is likely the most important thing we do today.
Yes, I am really asking for YOU to reply. This is the best way that I can guide my work. Please let me hear from you.
Sincerely,
Daniel Lemire
Founder, AI Mistakes
Midjourney prompt: “A historic humorous city square with a crowd of people acting hysterical”