Greetings! ICYMI, there was big news in the AI world this week. You can catch up in this week’s newsletter. And don’t forget to do your homework! Scroll down to “DYI AI” for this week’s assignment. As always, if you like what you see, feel free to drop something in the tip jar.
Happy 25th, AI Insider
By Michelle Johnson, AI Insider
AI Insider turns 25 today. No, not 25 years. This baby is 25 issues old. Well, that went fast.
It seems fitting to mark this because issue No. 1 almost didn’t happen. As some of you know, I was reluctant to put time into something that could quickly become a time suck in my retirement.
Yet, here I am, pushing out No. 25, surprised that it’s gone this far.
Some people take up pottery or painting when they retire. I had thought I’d go the pottery route after following some cool creators on TikTok. Obviously, I never quite got around to throwing clay.
When ChatGPT blew up this month a year ago, I became intrigued. It reminded me of the early days of the Internet. Back then, I started out with my first home computer, a Commodore Vic-20. It was the early ‘80s, and I was intrigued like I am today. I worked my way through a succession of PCs (never a Mac!) until I opened a web browser for the first time in 1993 on a fellowship at Stanford University.
Things were changing fast then, too. Companies were rushing to create a “web presence.” Startups were attracting a crazy amount of money, and there was great demand in the job market for people who understood something about the web thing.
So, here we are again, this time with AI. And it’s as exhilarating and scary as it was the last time. Only more so.
Clearly, my interest in tech trumped my interest in making pottery. I’m lucky enough to have the time and the interest in sharing what I’m learning about AI with you.
For now, pottery or painting on an actual canvas is on the back burner. But, who knows? It could still happen at some point.
In the meantime, I hope you will find it useful to come along for the ride with me via AI Insider.
OpenAI’s Developer Event: Key Highlights for Everyday Users
By Michelle Johnson and ChatGPT, AI Insider
The big news in AI this week came out of the OpenAI (ChatGPT) developer confab held in San Francisco on Monday. OpenAI CEO Sam Altman announced a slew of updates to their blockbuster chatbot and the launch of some brand-new features.
To be clear, this “DevCon” is aimed at unleashing goodies for developers/programmers, so we mere mortals have to pick through these updates to see what’s in it for us.
And if you haven’t upgraded to the paid version of ChatGPT, this may be the time to make the move.
IF you pay for ChatGPT Plus, you can expect from this upgrade:
GPT 4 Turbo: Faster and boasting a context window four times larger than its predecessor, GPT 4 Turbo will allow you to enter more text in the prompt window or upload larger files, and the AI can process and remember more information.
Refreshed Look: A refresh of the ChatGPT interface. Nothing extreme, but useful rearrangement of the menu area and main screen. And it’s pre-populated with new a few GPTs to get you started. (See item below.)
There’s a GPT for That: With this new feature, anyone can create their own personalized GPT. Imagine crafting a GPT bot that can assist with cooking recipes, provide daily workout routines, or even act as a virtual travel guide, offering suggestions for your next vacation. You can also sell these GPTs in the upcoming GPT store. (See next item.)
GPT Store Launching Soon: A GPT Store is set to open, where users can publish and share their custom-made GPTs. This platform will feature a variety of user-created AI bots, showcasing the creativity and of a global community. Think of it as an app store for GPTs.
More Recent Knowledge: Until Monday, ChatGPT-4’s knowledge base ended in 2021. To ask it questions about anything more recent required a plugin. Altman says the bot has been updated April 2023. Big whoop, because if you want to talk about what happened yesterday or last week, it will still need to fire up the Bing plugin, so you’ll wait for a response.
Improved Voices in Text-to-Speech: The new text-to-speech API has six preset voices that sound more realistic and diverse than previous voices. Altman played a demo of one of the new voices, which sounded more lifelike than what we are used to hearing.
Pundits are speculating that OpenAI will kill off a bunch of startups who’ve been developing features similar to what OpenAI is now embedding into ChatGPT. No doubt a shakeup is coming, but who knows how drastic it will be. We’ll continue to keep an eye on developments.
Hands-on with ChatGPT’s New GPT Creator
By Michelle Johnson, AI Insider
I was glad to see this new feature in the recent update of ChatGPT, but not wildly excited. Why?
I generated a GPT months ago at a site called Poe.com. You may have seen me post about it. I named it BlkChat. (Try it). It lets you chat with black historical figures such as Frederick Douglas and Sojourner Truth. It took just a few minutes to set it up, and ChatGPT powers it.
So, I was eager to get a look at OpenAI’s version of a no-coding required bot builder. Although it was announced on Monday, I had to wait until Thursday to get in. Between OpenAI’s system getting overloaded and hacked, it took a while for the site even to load.
Once I got in, there was a new “Explore” link on the left-hand menu. Clicking it revealed an OpenAI curated list of GPTs and a link to create a GPT. (This is a beta feature.)
Behind that link is a screen with a prompt box where you can describe the type of bot that you want, and ChatGPT will ask additional questions to fine-tune your idea. A preview window appears on the right. You can ask ChatGPT to make changes to anything that you don’t like.
I decided to replicate BlkChat here by skipping the ChatGPT prompts and hitting the Configure tab at the top of the screen. That allowed me to enter a name for the bot, an opening prompt, (Welcome time traveler!), and other info. It’s exactly what I did to create BlkChat on the Poe site.
While I was in te configuration, I forgot to add a logo, so back on the Create tab, I asked ChatGPT to generate an image of Fredrick Douglas. It refused, citing an issue with its content rules. However, back in Configure, I could upload a copy of the same image of Douglas that I’d used on the Poe platform.
I ended up with something that looks and functions very much like the first BlkChat.
What’s really different about OpenAI’s model, is the ChatGPT interface. I decided to use it to create a travel bot from scratch. First, it asked me what kind of bot I wanted to create and what I wanted it to do.
So, I typed in something about how it should give me info about events and things to do when I typed in the name of a city I would be visiting and the dates. It then asked a few more questions to refine the model, suggested a name for the bot (Voyage Buddy), and generated a logo and a tagline (Your go-to guide for travel events and activities).
I asked it to change the logo, hit save, and my bot was done. No programming required.
You can make a bot private or open it to the public. You can also monetize your bots on both Poe and, eventually, on OpenAI’s GPT store when it opens. Both companies have announced that they will share the proceeds with creators.
The BlkChat final results on each platform were virtually the same. In minutes, I was chatting with Fredrick Douglas and Sojourner Truth, getting a history lesson in the style of a live chat. Voyage Buddy updated me on what I can do on a trip I’ll take soon.
Like Poe’s platform, OpenAI’s chat builder is no muss, no fuss, and it benefits from being powered by ChatGPT. Poe benefits from being free and also powered by ChatGPT. Creating BlkChat on ChatGPT requires the $20/mo. Plus subscription.
DIY AI: Now We’re Cookin’
By Michelle Johnson and ChatGPT, AI Insider
Hello, class! Are you ready for Thanksgiving? No? Let’s get some help. This week, for DIY AI, we will turn a bot into a cookbook.
The Assignment: Cook with a Bot
Pick your sous-chef — ChatGPT, Bing, Bard, Claude, or the like — and give them this prompt to start your culinary journey:
Prompt: I want to make [insert dish here]. What are the ingredients and steps?
Example: I want to make a classic lasagna. What are the ingredients and steps?
Go for the gourmet version by asking for a cooking tip:
Can you give me a professional tip for making [insert dish] even more delicious?
Finally, don’t forget the shopping list! Ask your bot to generate one.
I asked for help making a sweet potato pie. Here’s one professional tip Bard gave me: “Use roasted sweet potatoes for a richer flavor. Roasting sweet potatoes concentrates their sugars and intensifies their flavor, resulting in a richer, more decadent pie filling.”
What did you and the bot whip up? Share in the comments on this week’s newsletter.
What the Actual Hell?
Sony AI: Algorithms Getting Skin Tones Wrong
By Bing, for AI Insider
AI algorithms can be biased not only by the lightness or darkness of people’s skin tones, but also by their redness or yellowness, according to a new research paper from Sony AI. The paper proposes a more comprehensive and automated way of measuring skin color diversity in AI systems, using the CIELAB color standard.
Skin color bias in AI has been a long-standing issue, especially for facial recognition and analysis systems. A 2018 study by Joy Buolamwini and Timnit Gebru found that AI was more likely to make errors when used on darker-skinned females. Since then, many AI companies have adopted skin tone scales to test the accuracy and fairness of their algorithms across different skin tones.
However, Sony’s researchers argue that these scales are too simplistic and one-dimensional, as they only focus on the light-to-dark spectrum of skin color. They claim that these scales neglect the biases that may affect people who have redder or yellower skin hues, such as East Asians, South Asians, Hispanics, Middle Eastern individuals, and others.
To address this gap, Sony’s paper suggests using the CIELAB color standard, which is a three-dimensional color space that measures the lightness, redness, and yellowness of colors. The paper shows how this standard can be used to automatically measure the skin color distribution of image datasets and AI systems, and to detect any biases or inaccuracies.
Sony’s research also reveals some interesting findings about the skin color biases in existing image datasets and AI systems. For example, the paper shows that some image datasets overrepresent lighter and redder skin, and underrepresent darker and yellower skin. The paper also shows that some AI systems favor redder skin and misclassify yellower skin, such as labeling them as “more smiley”.
Sony’s paper has received positive feedback from some major AI players, such as Google and Amazon, who are reviewing the paper and its implications. However, the paper also faces some criticisms and limitations, such as the validity of the CIELAB standard and the complexity of its implementation.
Random Shorts
PopSocket, the company that makes those round holders that many people have glued to the back of their phones, has launched an AI design tool that lets you create custom cases, wallets, and pop circles.
A long-awaited piece of tech that looks more like jewelry, the Humane AI pin is finally dropping. Secretly years in the making, this wearable, AI-driven device that clips onto clothing is receiving mixed reviews. It’s $699, plus $24 a month for a T-mobile connection.
Worth a read, NYT: Personalized A.I. Agents Are Here. Is the World Ready for Them? (No paywall.)
Aht Gallery
The prompt was simply: Veterans Day. Here’s how AI image generators Midjourney, Leonardo, Dall-E 3, and DreamStudio interpreted it.
