JDOT #6: Prosthetics, Dolphins, and AI Time

AI Time

There are so many AI Announcements that we need AI to read AI. Maybe the last two weeks are a preview of what providing "intelligence" at scale means to our future. My bet is that "AI Time" will be the phrase of 2025.

Moving Prosthetics

Who says a prosthetic has to be attached to its body to be handy?

Watch the original wireless prosthetic hand demonstration, MP4, 5 seconds. Download the English visual-description track, VTT.

Audio Description: Close up, woman with black prosthetic arm, controls her detached hand, directing it to get her champagne a few feet away.

Open Bionics founders Samantha Payne MBE and Joel Gibbard MBE are revolutionizing assistive devices through innovation that’s anything but boring. Just ask Tilly Lockey, whose wireless, bionic hands captivated 10M+ viewers online. The company combines 3D printing with precision engineering to develop advanced prosthetics that genuinely improve lives.

How tall are you today?

The video reminded me of Paralympian Aimee Mullins' TED Talk, where she shared her experience with bilateral amputation at age one. With Alexander McQueen, she reimagined prosthetic design. She described having a dozen pairs of prosthetic legs, giving her a height of 5ft 8 to 6ft 1, leading to a funny moment when a friend "complained" about her having an unfair advantage in their height rivalry (as a joke).

TED Talk: “It's not fair having 12 pairs of legs”

The SOMAA Report is LIVE!

The groundbreaking State of Mobile App Accessibility (SOMAA) report has officially launched! ArcTouch has released this first-of-its-kind analysis revealing the true state of mobile app accessibility, and it's generating significant attention across the industry.

State of mobile app accessibility, much more to do according to SOMAA… | Customer Experience Magazine (CXM)

As the European Accessibility Act is about to kick in, a new report from ArcTouch, The State of Mobile App Accessibility (SOMAA), clarifies that too many apps are still lacking in accessibility. The report illuminates what designers and developers...

Geoff Spick at Customer Experience Magazine (CXM) interviewed me for the first, front-page feature on SOMAA.

A lot more media coverage is coming as we approach GAAD. Be sure to download your copy and arm yourself with the data that everyone will be talking about in the next year.

Download your copy of the SOMAA report

Global Accessibility Awareness Day is a month away!

Whether your GAAD event is public or private, we encourage you to submit it to the official GAAD website. Every contribution helps us spotlight the global momentum behind digital accessibility.

Later in this newsletter, I’ll share the (admittedly intense) process I’ve taken to enhance our GAAD knowledge base with AI. By the time our 15th anniversary rolls around next year, we’ll be equipped to tell the story of GAAD in ways that are richer, more dynamic, and unlike anything we’ve done before.

Submit your GAAD event

The GAAD Foundation Speakers Fund Announces Two Recipients

On behalf of Jennison Asuncion and myself, we are delighted to congratulate Jessica Yaacoub, MBA CPACC and Florian Beijers, our two Speakers Fund award winners. These outstanding individuals will help bring digital accessibility to mainstream audiences through their speaking engagements. Here is the official announcement:

Speakers Fund | GAAD Foundation | GAAD (Global Accessibility Awareness Day) Foundation

This month, we awarded two speakers through the GAAD Speakers Fund, helping them bring the topic of digital accessibility to mainstream audiences. Jessica Yaacoub, MBA spoke at the Globalization and Localization Association (GALA) Conference in...

Dara the AI Lunch Hero

My longtime colleague and friend Dara Sanderson has launched an impressive AI-powered project after I encouraged her to try AI coding. Dara, who has held numerous executive positions including CEO, CPO, and CTO, took my casual mention of Cursor (an AI coding tool) and within just four weeks transformed it into "Lunch Hero," a fully functional business now in production.

Lunch Hero uses AI to generate personalized, healthy lunch plans for children, solving a common pain point for parents. I've mentioned it casually to a couple of Solopreneur Moms who love the idea!

Visit Lunch Hero to create healthy lunch plans for your kids

In just a few weeks, I built Lunch Hero, an AI-powered tool that generates personalized lunch plans for kids.

Uh oh, will people windsurf away from Cursor?

The rapid pace of "AI Time" creates challenges for even AI companies. Despite recommending Cursor enthusiastically for months, I've watched as newer AI models have surpassed its capabilities while Cursor's integration hasn't kept pace.

This illustrates a fundamental challenge of prompt engineering in the AI era: it's like designing a ship's hull during an Atlantic crossing. You might spend months perfecting your prompts, only to have them rendered obsolete by newer models while customers grow increasingly frustrated.

After weeks of diminishing returns with Cursor, I've reluctantly expanded my toolkit to include Windsurf (reportedly in acquisition talks with OpenAI for $3B), while maintaining my Cursor subscription at a reduced level. I'm also evaluating Copilot, Cline, and Aider. In this rapidly evolving landscape, depending on a single AI coding tool has become too risky for professional work.

News from the Big AI Models

OpenAI

As brilliant as AI Models are, they are bizarre creatures. They cost so much and take so long to train, that their inability to learn new facts in a timely manner confuses people. The AI Models are starting to solve these problems, but from a software developer's perspective, they are all hacks. The core models have usually finished their training months ago.

The other challenge for models is that as you provide more information, from adding files to just having a long chat, the more likely it is to hallucinate or generally degrade in quality. That's called the context window. Most of us have been trying to solve it with context engineering.

Getting LLMs to remember us will make them so much more useful. As will providing large context windows that don't degrade the LLMs' ability to reason. Google's Gemini models have been the champions in context window sizes with 1 million, vs about 100 thousand for many models.

Cue this week in OpenAI.

OpenAI : Starting today, memory in ChatGPT can now reference all of your past chats to provide more personalized responses

Adding memory is not as simple as you may think. People lie. Share accounts. Change their minds. It's more about having an algorithm that figures out what memories to save or "use" than it is about improving the base models.

Screenshot of OpenAI's announcement about memory in ChatGPT

To that end, ChatGPT users, I strongly recommend you periodically prune the memories ChatGPT stores about you to make sure you get the most out of your chats. I had a few doozies in mine.

To find it, click on your avatar in the top right. Go to settings -> Personalization->Memory->Manage memories.

Screenshot of "Memory" settings

Let us know if you have any humorous entries.

Speaking of training models

Sam Altman recently offered a fascinating behind-the-scenes look at OpenAI's model development process. He shared a conversation with the employees who helped create GPT-4.5, providing rare insights into their methodologies and challenges. For those curious about how these increasingly powerful models come together, this conversation offers valuable perspective from the people directly involved in building them.

Sam Altman: A lot of people were interested in how we made GPT-4.5 and what comes next..we did a podcast [with the employees] who helped drive the product.

I've linked the entire conversation below:

Watch OpenAI's Pre-Training GPT-4.5 video on YouTube

A "fiery" chat with Sam Altman at TED Conferences

Sam also sat with TED CEO, Chris Anderson, for a whirlwind of a conversation. Chris gives quality interviews, but not hard-hitting. It turns out that he is capable of pressing his subjects hard. It was awkward. And interesting.

A few highlights:

Chris Anderson: I asked [ChatGPT 4.0] to imagine Charlie Brown..thinking of himself as an AI. It came up with this. I thought this was actually rather profound. What do you think?

Cartoon that looks similar to a Peanuts cartoon

I don't know if Chris Anderson considers himself a journalist, since his LinkedIn profile focuses almost exclusively on his TED journey, but that was the kind of loaded question that a good journalist would ask. Sam, here's a live grenade, but before I hand it to you, allow me to take the pin out first.

In one shot, in front of an influential, global audience hanging on your every word, you need to address the relatively profound words of the cartoon, while reacting to the concerning question of a possible IP violation underpinning the entire technology.

It's the kind of question where your facial expressions and how you choose to answer reveal so much more.

Sam, ignoring the IP issue underlying the question: "I mean, this is an incredible meta answer, but there's really no way to know if it is thinking that or it just saw that a lot of times in the training set. And of course..if you can't tell the difference, how much do you care?"

Chris: At first glance this looks like IP theft..you guys don't have a deal with the "Peanuts" estate?

Boom. The audience reacts with applause, leading Sam to attempt a comeback that fell a bit flat: "You can clap about that all you want, enjoy"

One actress was on the side of the clappers, but I wouldn't say "enjoy" is the correct word to describe her reaction.

Trash. Criminals. Contributing less than nothing to society. All they do is take IP, water and power resources, and as much money as they can grab. Despicable people.

Justine Bateman reacting to the TED talk

Strong words...

At one point, Chris made a faux pas after asking Sam, "How many users do you have now?"

Sam: I think the last we said was 500 million weekly actives, and it is growing very rapidly.

Chris: I mean, you told me that [OpenAI] doubled in just a few weeks...

Sam: I said that privately.

Oops.

When your user base is 500 million humans, and you’ve been doubling every few weeks, you're on pace to dethrone some kings of big tech.

I keep wondering if they’re making profit off of all those users, since compute is so expensive. Clearly they are in better shape than I thought. Investors believe that the AI market obeys Power Laws, which in the Venture Capital community means trillions to the winning company. A "winner-takes-all" market.

The unspoken fear for the rest of us, is that if Amazon took almost 40% of the e-commerce economy and 4% of all retail sales, how much of ALL commerce will the winner of the AI wars take?

Watch Sam Altman live at TED2025 on YouTube

I love AI so much. But no one can blame folks in an industry hemorrhaging jobs for being mad. Especially when they haven't been paid for the use of their IP.

Remember when Google hoovered all those images from people's sites? YouTube, before Google bought them, was already filled with copyrighted content. I thought they would all be cooked, but Big Tech has gotten away with it again and again. Now the safe bet is that AI companies will too.

I'm somewhat torn. The taxi industry held onto power with regulatory capture. You need to break a few vases, if not laws, to change the world as Uber did. I wouldn't want to go back to a pre-Uber world. But that doesn't mean you have to condone ALL the tactics either.

After the TED appearance, Sam did not share the talk on social media. Chris did, addressing the controversy that ensued:

In the room people mostly loved it, because they got the context… several speakers…had railed against Al. It was important to put their concerns to Sam…

TED curator, Chris Anderson, addressing the controversy

Last but not least, a note to my fellow toastmasters. We always emphasize that no matter how rough the audience, or moderator, always maintain composure until you’re back in the green room. Shake hands with the host, look at the audience and gracefully exit the stage. The end of this TED talk wasn't terrible, but provides a teachable moment to all speakers. Sam broke this cardinal rule and tried to escape the stage as quickly as possible. People online noticed.

The difficulty level of handling PR and media when you are the CEO of the leading AI company, that may change civilization for good, is off the charts. It's easy to underestimate how skillful Sam is, but if you want to understand how he got where he is, search YouTube for some of the pitch competitions and practices he conducted with founders. They might practice their pitch for a month. He would see it, understand its strengths and weaknesses instantly, and turn it into a killer pitch contemporaneously. Impressive stuff.

Mystery Model

OpenRouter has become an essential service in today's AI ecosystem by providing developers unified access to multiple AI models through a single API endpoint. Recently, they collaborated with an unnamed major AI lab to release "Quasar Alpha," a powerful model specifically optimized for coding tasks, featuring an impressive 1-million token context window.

OpenRouter home page

This "stealth" release allowed developers to provide unbiased feedback while creating intrigue in the AI community. Quasar Alpha quickly gained attention by outperforming industry leaders like Claude 3.5 Sonnet and Gemini 2.5 on benchmark tests. After building substantial anticipation, OpenRouter and OpenAI jointly revealed that Quasar was actually an early test version from OpenAI's 4.1 series of models, demonstrating how controlled soft launches can generate both valuable feedback and effective marketing buzz.

@OpenRouterAI: Many of you have been excited to know who made Quasar Alpha & Optimus Alpha, which both topped the charts…Today we can announce that they were early test versions of…

OpenAI left a few hints as well:

OpenAI teasing a livestream

Quasar turned out to come from their 4.1 series of models.

@OpenAIDevs: Announcing GPT-4.1, GPT-4.1 mini, and GPT-4.1 nano..major improvements on coding, instruction following, and long context.

The competition is heating up. Like Gemini models, all 4.1 models, including nano, have a 1 million token window, which allows you to upload a ton of files.

GPT-4.1 Launch Announcement

OpenAI published a guide to Prompt Engineering 4.1 to go along with the API. These guides are priceless tutors on how to get more out of AI models.

An example from the Prompt Guide. If you have a lot of files or a long prompt, and you repeat the instructions at the beginning and end, you will have much better results. If you only want to provide instructions once, doing so at the beginning works better.

GPT-4.1 Prompting Guide

Do you find AI model names to be confusing, and frankly poorly thought out? It's been talked about so much OpenAI employees make light of it publicly now.

Thankfully, CEO Sam Altman says:

how about we fix our model naming by this summer and everyone gets a few more months to make fun of us (which we very much deserve) until then?

The reason AI companies are so bad at naming their models, is that ideally the name would reflect what it's good at. But testing in-house won't get you there. It has to be out in the wild for a bit before the community figures it out.

Wharton Professor, Ethan Mollick on AI Models

The OpenAI Social Network

The Verge has reported that OpenAI is developing its own social network, according to multiple anonymous sources. While reports with unnamed sources should be treated cautiously, the strategic logic behind such a move is compelling. A social platform would provide OpenAI with valuable real-time interaction data to improve their models, something every major AI company could benefit from.

If I were naming such a platform, "Synapse" would be fitting, reflecting the neural connection metaphor that aligns with their artificial intelligence focus. Whether or not this specific report proves accurate, we should expect AI companies to continue seeking new sources of high-quality, real-time data to enhance their models.

Google AI

Google Cloud Next 2025 in Las Vegas showcased several significant AI advancements. The highlight of Google's announcements was Ironwood, their seventh-generation Tensor Processing Unit (TPU).

This custom AI chip represents a major advantage in Google's AI strategy, having both unlimited data access and the capability to design their own specialized hardware.

Google announcing their new Ironwood TPU chip

Blink and you'll miss a smart term Google just coined, "intelligence per dollar". The Gemini Flash models are indeed the most intelligent per dollar. Congratulations Google, "intelligence per dollar" is our phrase¹ of the week. If you know, you know.

Google Cloud coins the term intelligence per dollar

Google released a VideoGen model called Veo 2. I couldn't get it to work, so no comment. However, Google partnered with Sphere Entertainment Co. to reimagine The Wizard of Oz with Veo 2. What an incredible story. I was impressed.

Google Cloud Next, Veo2 and Sphere

Do you speak Dolphinese? That may not be possible today, but perhaps tomorrow, if the work of Denise Herzing has anything to say about it. Google just trained a model to understand dolphins. Squeal! [Dolphin for, WOW (probably)]

The goal would be to one day speak Dolphin

Google's Prompt Engineering Guide

Prompt Engineering by Lee Boonstra

Not to be outdone by OpenAI, Google also released a fantastic Prompting Guide. I learned a lot.

Gemini Game Creation

Turning a paper into a game

Google's Gemini 2.5 with Canvas feature demonstrates impressive capabilities for transforming academic content into interactive experiences. Professor Ethan Mollick showcased how the system can convert scholarly papers into functioning games with minimal human intervention.

I tested this by uploading Dr. Michael Levin's² latest research paper on cellular development to Gemini, which promptly created a playable simulation based on the paper's concepts.

The game, "Cellular Collective Architect," visualizes complex biological principles through interactive mechanics, though it currently lacks accessibility features.

This application shows how AI can bridge the gap between dense academic knowledge and engaging interactive content, potentially transforming how we experience complex information.

Gemini Game of Dr. Michael Levin's latest paper

Thinking Machines

Image of Mira Murati

Mira Murati, the former CTO of OpenAI, is reportedly raising at a $10 Billion dollar valuation. While she certainly has the credentials, what does it say about this market that multiple companies are raising billions of dollars, and two of them only have great resumes, AI know how, plaintext websites and no customers. Incredible.

Enshittification of the API

Ethan Mollick highlights a strategy of handicapping the API

Ethan Mollick consistently makes me think. This AI war will involve a lot of strategy as AI and downstream companies jostle for power. There are so many players with ungodly sums of money. I hope we don't see the APIs that power downstream companies, getting capabilities hampered in order to provide a strategic advantage. A good analogy is the App Store. The owner of the platform gets to make the rules.

How I use AI

Preparing for Media Interviews

Custom AI Podcasts

Since the last edition, I’ve had a few media interviews come up and needed to brush up on relevant topics. I went to Google's NotebookLM, added the journalist’s name, linked to a few of her articles, and uploaded background documents related to the subject.

NotebookLM generated a 20-minute podcast, which I listened to during my drive to the Strip. The podcast was solid, but I made a mistake in the prompt. I needed to be clearer that the journalist's articles were to understand what she wrote about, whereas the topic of the podcast needed to be about the theme of the interview. As a result, it leaned too heavily on the journalist’s past work and focused on tangential topics.

Fortunately, during a traffic stop, I was able to use one of NotebookLM’s beta features to make the session more interactive. That let me guide the AI “hosts” to cover the material I actually needed.

If you haven’t tried NotebookLM yet, it’s worth exploring, especially for interview prep or digesting complex information on the go.

Anki-style dynamic tutoring with ChatGPT Voice

You may be surprised to learn that I have a Diamond Certification from the GIA (Gemological Institute of America). When I was younger, I took their Diamond Grading Course. I made flashcards and every time I got a question wrong, I'd put it in the "wrong stack". And I didn't stop studying until I got every single flashcard right, scoring me a 99 on the final. As a Gretzky fan, I was delighted my score matched his iconic jersey number, but sad that I only discovered this technique after my schooling was over.

Fast forward to the 2020s and it turns out the kids today use software called Anki flashcards that uses a technique called Spaced Repetition to help you memorize things. For some reason, I couldn't get it to work for me. Probably because I inhale information at an intense pace, have a large variety of interests, and getting older, find it hard to remember it all.

But I started to think that ChatGPT Voice may solve the problem, at least for short term use. So again, to prepare for a media interview, I threw in all my talking points and all the topics I was boning up on and asked ChatGPT:

Help me write a prompt that gets you to quiz me on a list of talking points. Keep switching up how you ask the questions and go through them until I get everything right. Keep track of what I miss and circle back to it until I’ve nailed the whole set.

Here is the prompt that ChatGPT created, for its own use!:

You are a professional quiz coach focused on [TOPIC]. Quiz me until I’ve mastered the core facts below. Vary your question types—use short answer, fill-in-the-blank, rephrased multiple choice, or “explain like you’re presenting to [X]” formats. Reword the questions and shuffle their order to challenge retention. Be supportive, but hold a high bar for accuracy and clarity. Only move on once I’ve nailed each point.

I could not believe how well it worked. In retrospect, I would have told it about the interviewer, thrown in a few sample articles and asked it to act like they are that journalist and ask me questions from their perspective.

But even so, it WORKED LIKE A CHARM! I could not believe how quickly I learned a whole bunch of facts. Highly recommend playing with this approach.

Deep Research

For an upcoming GAAD interview, I needed to compile a comprehensive history, something we'd never properly compiled. This presented an ideal opportunity to test various AI deep research capabilities across platforms.

First, I created custom prompts for ChatGPT, Claude, Perplexity, Grok, and Google Gemini, uploading our historical documents and relevant links to each. After collecting their responses, I consolidated the findings in Google's AI Studio, which efficiently handled the large context window, and exported a structured CSV file.

The real breakthrough came when addressing our fragmented historical records. We had structured data on events from 2020 onward, but information from 2012-2019 existed only as unstructured lists with broken links. Using Cursor, I developed a program to scrape and analyze over 150 archived pages, extracting event titles, locations, and descriptions.

This process transformed previously inaccessible historical data into a structured database. While AI tools sometimes hallucinate, they excel at converting unstructured information into organized data, when supervised by humans. By next year's 15th anniversary, we'll have a complete GAAD historical record suitable for statistical analysis and rich storytelling.

Accessibility and Gen AI Podcast Updates

Eamon McErlean and I invited Matt Ater onto the pod to talk about JAWS, AI, his background and more. Matt is another OG of accessibility whose contribution to the field cannot be overstated.

Image Gen

What would this newsletter be without some ImageGen? Well, it was harder than the last edition because OpenAI nerfed their image tool. Perhaps in response to the TED interview. That said, we've got plenty to work with. Let's start with a Dolphin Whisperer, to celebrate Google's Dolphin LLM.

Here is OpenAI's best version:

Joe with a Dolphin by ChatGPT

Here is Gemini's best version:

Gemini's version of Joe with Dolphin

For some reason Grok really nailed this prompt:

Grok's version of Joe and a Dolphin

Pets to humans

The latest trend is having AI generate the human version of your dogs. Of course I had to turn Pepe into a human:

Pepe next to human-pepe

Then I thought it was mean to turn Pepe into a human without doing the same from human to dog. Here is my headshot if I turned into Pepe:

Joe as Pepe

I took all the text from this edition, ran it through ChatGPT, Gemini and Grok to generate a banner image incorporating all the elements from this newsletter.

ChatGPT:

An old man in a suit

I have no idea what ChatGPT was thinking! It's an image of an old man in a suit.

Ai Time, by Gemini

Google Gemini came up with a blue clock that says AI Time.

The winner is Grok, whose image is the banner of this edition. Here's the image description: "Woman with a prosthetic arm interacts with a floating digital interface; background shows ocean and coral reef through a large window. Text at top reads 'AI Time: Innovating Accessibility.'"

Going dark through GAAD

Writing this edition took a long time. The last two weeks have seen announcements accelerate so much that every time I put something down on paper, it was out of date.

I was hoping to cover all the major AI news, but it's impossible unless you do this daily and do nothing else.

I've booked ~15 keynotes this GAAD season and need to finish coding up the topic I'm going to speak about. So JDOT is going dark for about a month. It will give me time to reflect on what slice of AI to tackle, so it's sustainable.

The biggest struggle is making it interesting for my audience, which is made up of hardcore CTO-types, Devs, Entrepreneurs and Accessibility folks. It's hard to write about something that interests all of those groups simultaneously. I'd love feedback on what you want to read.

See you in early June.

Footnotes

¹ What is "Phrase of the week" about? Get in touch and subtly work the phrase into your message to prove you read the newsletter to the end. Last time we had 2 winners, Elaine KA and Heather S. Neff.

² Dr. Levin's work is the only topic that is more exciting and important than AI is today. If I were a young kid, I would do everything I could to study at his lab in Tufts University. He's going to turn out to be the Einstein of our times. Speaking of which, if Einstein were alive today and he was following you on social media, would that change how you wrote? Because the good doctor followed me yesterday, and now I feel a responsibility to stop posting nonsense. :-)

Editorial correction

October 5, 2026: The original paragraph described Aimee Mullins as having a dozen prosthetic legs. Her 2009 TED Talk describes a dozen pairs. The paragraph now reflects that count and the timing of the talk.