I was not alive during that time, and I will not pretend I was. What I can do is a deep dive, with images, about Heumann and the others who walked that journey with her.

All of us who are disabled in 2026 need to fight to be part of the AI movement and to make our voices heard. Disabled people have one of the greatest chances at independence with what this technology can bring. We must advocate on disability focused AI. We must build it for ourselves.

Thank you to every developer and AI leader already thinking about disabled people and AI. We need this as much as anyone else. On a a daily basis, AI can shape our lives more directly than others. Take a look at past articles and you will see what I mean here.

I plan on dropping a rap music video that will surely make heads turn. Stay Tuned…

I can’t tell you how many times I have spent tokens on GenAI image/ video of wheelchairs and it gets it all messed up. As a wheelchair user with a BKA (below-knee amputation), it gets even worse – wonkier limbs, broken chairs, scenes that don’t represent. I contacted ChatGPT to see if they wanted high-quality images or video from me in any format. No response. I do have to admit ChatGPT does a great job with GenAI images of wheelchair users… Sometimes. Recently it has improved a lot though.

I want to thank Microsoft for thinking about the disability community and putting some inclusion into GenAI images and videos. I like that they’re making it more of a community effort. It helps Microsoft get better data from communities like amputees, little people, etc.

Microsoft’s approach is called Project Volano “Enabling everyone to shape AI.” It is a data effort where communities help build image and video libraries that shape how AI represents them. Open libraries include Limb Difference & Prosthetic, Little People of America, and Black Albinism. Weirdly enough, I still don’t see a wheelchair section and wheelchair users are very much underrepresented in GenAI. Microsoft could do better, and even offer some of these communities money. As a disabled person on a fixed income times are tough sometimes!

I like the idea. Enabling everyone to shape AI is partially what AI For Independence is all about. But I also feel Microsoft lets projects like these go silent. Looking at Project Volano, the page has not been updated in 2 months, no news in 2 months and there’s still no wheelchair community library on the page. Do better, Microsoft. I’m giving them another chance personally.

If you are disabled, from a marginalized community, or have an interesting lived experience, reach out to OpenAI, xAI, Meta, or register interest for a Community Library through Project Volano. The data disabled people share now will help future disabled people live more independently.

I usually talk about Benjamin Franklin’s flexible catheter. In 1752 he designed one with hinged metal segments for his brother John. Now I can’t imagine the pain a metal catheter would feel like but later versions of what the catheter is today have helped disabled people like me live longer. I wouldn’t be here without the invention. That’s why our experiences matter: medically, in daily life, and in what we build. Future disabled people learn from what we share, and from our mistakes, so they can live more independently and more productively. AI helps organize this, and find needles in haystacks.

Microsoft: Add a wheelchair community library. Project Volano

Nexy is an accessibility-focused AI assistant for Apple Mac that lets you interact with your computer using your voice. Instead of constantly navigating your Mac by hand, you can speak naturally to Nexy and tell it what you want to accomplish.

Nexy can answer questions, work with documents, browse the web, and help you navigate macOS. It works alongside VoiceOver and uses ChatGPT to help understand requests and complete tasks.

Developers also plan to add support for additional AI providers, including Claude.

What Can NexyAI.app Do?

Work

• Work with email.
• Work with documents.
• Complete browser tasks.
• Reduce the amount of manual navigation needed.
• Help you accomplish computer tasks using your voice.

Learn

• Research information.
• Read information.
• Help with writing.
• Make learning and research more accessible.
• Help users work more independently.

Everyday Tasks

• Browse the web.
• Work with messages.
• Use YouTube.
• Navigate around your Mac.
• Get things done without constantly reaching for the mouse or keyboard.

Navigate Your Mac

• Move through windows.
• Navigate menus.
• Access settings.
• Interact with buttons and controls.
• Navigate different parts of macOS using your voice.

Use Other Apps

Nexy is not limited to one specific application. It can help navigate apps on your Mac, including menus, buttons, and other controls.

Use your voice to interact with your computer instead of having to manually navigate every part of it.

Why NexyAI.app?

Works With Your AI Assistant

• Connects with ChatGPT.
• Use AI to help complete tasks by voice.
• Claude support is planned for the future.
• Gives users another way to interact with AI on their Mac.

Built for Accessibility

• Designed with visually impaired users in mind.
• Accessibility is a core part of the experience.
• Helps reduce the amount of manual computer navigation required.
• Can also be useful for people who are not visually impaired.

Optimized for VoiceOver

• Designed to work alongside Apple’s VoiceOver screen reader.
• Voice and accessibility are central to the experience.
• Gives VoiceOver users another way to interact with their Mac.

Native macOS Experience

• Built for Apple Mac.
• Designed around macOS.
• Provides a more direct way to interact with your computer.
• Combines AI assistance, voice control, and accessibility.

Who Is Nexy For?

NexyAI.app is primarily designed with visually impaired users in mind.

Nexy may also be useful for people who:

• Prefer controlling their computer with their voice.
• Have difficulty using a mouse or keyboard.
• Want to reduce repetitive computer navigation.
• Want an AI assistant that can interact with their computer.
• Want to use AI to make everyday computer tasks easier.

See NexyAI.app in Action

If you want to see what Nexy can actually do, check out the NexyAI YouTube channel.

You can also learn more about the project on the official website.

Official website: NexyAI.app

https://nexyai.app

A sip-and-puff controller gives people who cannot use a traditional game controller another way to play. It converts sips and puffs into controller inputs, allowing users to control games without relying on their hands.

Caleb has said that the goal for this open-source controller is to keep the sip & puff controller under $200, depending on the parts used. The project includes a parts list and the information needed to build one yourself, although it is still a work in progress.

The creator is still testing and improving the software before releasing the project more widely. He wants to make sure the controller is reliable and polished rather than having users run into problems that are difficult to diagnose remotely.

I really like this project because it could eventually become a much more affordable way for people with disabilities to play games and gain independence. Accessible gaming equipment can be extremely expensive. For example, a QuadStick setup with mounts can easily cost $800 or more, which puts it out of reach for many disabled gamers.

Personally, I’d like to 3D-print one or two of these controllers. A 3d printer is pricey but a good way to keep me busy, and potentially help others.

Who Could Benefit? A sip-and-puff controller could potentially help people with conditions that make traditional hand controllers difficult or impossible to use.

Neurological & Spinal Conditions
Spinal cord injury — Can cause partial or complete loss of hand and arm function.
Cerebral palsy — Can cause weakness, tremors, or poor coordination.
Multiple sclerosis (MS) — Can cause weakness, numbness, tremors, and coordination problems.
ALS — Progressive muscle weakness can affect the hands.
Parkinson’s disease — Can cause tremors, stiffness, and reduced hand control.
Stroke — Can cause weakness or paralysis on one side of the body.
Spina bifida — Can affect movement and muscle control.
Traumatic brain injury — Can affect strength, coordination, and hand control.
Peripheral neuropathy — Can cause numbness, weakness, and reduced dexterity.
Muscular dystrophy — Progressive weakness can make hand use difficult.
Spinal muscular atrophy (SMA) — Can cause weakness in the arms and hands..
Upper-limb amputation — Loss of part or all of a hand or arm.

Sip & Puff Gaming Controller – Partially Assisted AI Development

One of the things that makes this project especially interesting to me is how AI was used during its development. I personally spoke with the creator, who told me that more than half of the code was written with the help of Claude AI.

However, this wasn’t a case of giving AI one giant prompt and having it create the entire controller.

Instead, Caleb started with a small project and gradually asked Claude to add features one step at a time. He tested each change, kept the features that worked, removed or changed the ones that didn’t, and continued building from there. As the project grew, he also used Claude to help clean up the code. The creator said the process took weeks of small changes, testing, and troubleshooting.

He also made an important point: people using AI to write code should still understand what they are building (The Code) and carefully check the AI’s work. AI can be extremely useful, but it doesn’t always get things right. There are still parts of this project that Claude hasn’t been able to solve and that may require an experienced human developer.

Current Features
6 sip-and-puff sensors
12 total inputs
8 game-profile slots
Adjustable sip/puff sensitivity
Standard, repeating, and latch modes
Multiple inputs can be mapped to the same button
Web interface for configuration and recalibration
Screen showing sensor status and current gaming profile
Future Features

The creator is also working toward:

Joystick support
Physical recalibration button
Physical game-profile switching
Bluetooth connectivity
Capacitive touch features using the ESP32-S3
Parts

The controller uses relatively inexpensive, off-the-shelf components along with a 3D-printed enclosure:

6 pressure sensors
ESP32-S3
ESP prototype board
Thumbstick
GC9A01 round display
3D-printed enclosure

The project also provides the raw Arduino code and a downloadable firmware file that can be loaded through ESPHome.

Why I Think This Is Interesting: For accessible gaming, DIY projects like this can give people another option when traditional controllers don’t work for them.

What I like about The Controller Project is that it combines affordable electronics, 3D printing, open-source hardware, human creativity, and AI-assisted coding to tackle a real accessibility problem.

Shout out to Caleb Kraft & The Controller Project. If you would like to donate towards his project, he accepts btc, eth, has a patreon. https://thecontrollerproject.com/support/

Meta AI and Mark Zuckerberg are helping push technology toward a future where people can work, learn, communicate, and live more independently.

When I set out to create AI for Independence about four or five months ago, one of my biggest goals was to advocate for ways people with disabilities can use AI to gain more independence in their everyday lives. That could mean robots that help around the home, AI agents that assist with everyday tasks, smart glasses that provide information hands-free, or brain-computer interfaces that allow people to control computers and communicate.

I know Meta has been working on these technologies for a long time, and obviously my website had nothing to do with their success. But it still feels good knowing that I can write about this progress in a positive way. If someone with any level of disability comes across my website, I want to be able to point them in the right direction and help them discover how AI might improve their independence, whether that technology can help them today or becomes available in the future.

That is why I was excited to see Meta announce the recipients of its AI Glasses Impact Grants. Through the nearly $2 million program, 30 organizations across 18 states will explore ways Meta AI glasses can help people work safer, learn faster, and live more independently.

The projects cover accessibility, workforce safety, education, agriculture, and economic opportunity. Organizations are using the glasses in very different ways—from coaching people entering construction trades hands-free, to helping people with dementia navigate daily routines, supporting roadside mechanics, and improving safety for rural broadband installers in New Mexico.

For the disability community in particular, this technology has a lot of potential. Think about the people with disabilities using AI glasses today and helping improve the technology for the people who will depend on it tomorrow. People who are blind or have low vision are already exploring these tools. People with cognitive disabilities can use AI assistance to help navigate their daily lives. Wheelchair users can use hands-free cameras and AI to document accessibility barriers. People with disabilities that stack would benefit a lot too. Such a person in a wheelchair who is also deaf or hard of hearing.

Meta launched the program in January with nearly $2 million in funding for organizations putting AI glasses to practical use. The program includes Accelerator Grants for organizations already expanding proven uses of AI glasses and Catalyst Grants for developers creating new applications through Meta’s Device Access Toolkit. Recipients were selected based on their demonstrated impact, ability to scale, and relevance to areas where hands-free technology could make a meaningful difference.

Nearly 500 organizations applied.

Some of the projects focused specifically on independence and accessibility are especially interesting to me.

Expanding Independence and Access

OurLife Labs in California, supported by Reality Hack in Massachusetts, is developing OurLife, an assistant designed for Meta AI glasses to support people living with early-stage dementia and mild cognitive impairments. The system is designed to provide help in the moment, including reminders, assistance with daily routines, and step-by-step guidance. It is also being developed as a shared experience that can involve patients, families, and caregivers while supporting greater independence and peace of mind.

The United Spinal Association in New York is equipping wheelchair-user advocates with AI glasses so they can document accessibility barriers in parks and public spaces hands-free. That information can then be provided to parks departments and public officials to help identify and improve ADA accessibility problems.

Easter Seals Greater Houston’s BridgingApps Program in Texas is studying how AI glasses could help three different groups: people with low vision, people with intellectual and developmental disabilities, and people experiencing cognitive changes associated with aging. Their work could help build the evidence needed for AI glasses to become more widely used as assistive technology.

Immerse Inc. in California is working to integrate its language-learning technology with AI glasses, allowing people to receive real-time, hands-free language tutoring while going about their daily lives. Instead of having to stop what someone is doing and open a phone or computer, learning could become part of the world around them.

These projects are only a few examples of what happens when AI moves beyond the computer screen and becomes something that can assist people in the real world.

If AI can help someone communicate more easily, remember something important, navigate their surroundings, identify an accessibility barrier, perform a job more safely, learn something new, or simply rely less on another person for assistance, that matters. Disabled people testing and using these technologies today are helping build something better for the disabled people who will need them tomorrow.

That is exactly the kind of progress I created AI for Independence to cover and advocate for.

P.S My keyboard is kind of broken momentarily. Soda & Keyboards no bueno. I will try harder to make a post more custom with the help of AI. Currently I write a rough draft and ask AI to make it flow and read easier for people.

It takes me back to when Grand Theft Auto V first came out. My buddies and I would get out of school and just play. We’d spend hours at the mechanic shop trying to figure out which one of us could build the best vehicle to make it the farthest up those ridiculously steep mountains.

We tried everything: mountain bikes, motorcycles, sports cars, trucks, off-road rigs—you name it. And if we weren’t doing that, we were driving around looking for side missions, messing around with the world, or finding some completely new way to waste a few hours.

My first taste of GTA was San Andreas and Chinatown Wars. I’ve since gone back and played some of the older games again, which has helped refresh my memory.

Why I am not looking at the GTA VI leaks. Everyone is hyped right now, and I completely understand the temptation to scroll through every leaked screenshot, video clip, and piece of information that shows up online. But I don’t want to judge the final game based on some half-finished snapshot of a build that could be two or three years old.

That isn’t the version of GTA VI Rockstar is preparing to release. Instead, I’m waiting for the newest official GTA VI trailer on August 27. And yeah, I’m a little annoyed about the Netflix situation. But at the same time, if Rockstar is getting paid for a pre-release special presentation, why wouldn’t they do it? At least the official trailer is expected to hit YouTube six hours later, where everyone can watch it in full quality.

The Bottom Line

I love this franchise too much to let my excitement turn into disappointment before the game even comes out. So here’s my rule: I acknowledge that a massive leak happened. It’s significant, and the leaked material appears to be real—especially considering the DMCA takedowns surrounding it.

They don’t necessarily show what GTA VI looks like today, and they definitely don’t tell us exactly what the finished game will look like when it launches. Even with the game only about three months away, things can still change between the current build and what we actually play on day one.

Instead of getting sucked into an endless cycle of leaks, rumors, and “what ifs,” I’d rather save that excitement for the official reveal trailer on August 27. Then for the game to release in November.

If you’re using Windows or Linux, you can download LM Studio, install it, and pretty much get going. One reason I like it is that it comes with a graphical user interface. It also has a built-in model search and downloader. LM Studio supports things like RAG and Skills that give an AI additional information or capabilities. Depending on what you’re using the model for, you can make it much more useful for a specific task.

Then I Started Hearing About Harness Engineering

Running a model locally is cool, but could I run something closer to an actual cloud-based AI agent? Could I give my local AI tools, let it use sub-agents, or have different models handle different jobs? Could I build something closer to what the big cloud AI systems are doing, except on my own computer?

First Stop: Ollama

From watching YouTube videos and asking cloud-based LLMs questions, I figured out that I should install Ollama. I downloaded it, opened it up, saw a GUI, and thought I was set.

Then I realized I was looking at the easiest part. If I wanted to start experimenting with an actual harness, I needed to go deeper, and that meant using the terminal. For someone comfortable with terminals, that’s probably nothing. For me, mentally, every additional layer starts adding up.

It’s not that I can’t do it. It’s that something that might take another person 20 minutes can turn into two hours of experimenting, stopping, thinking, watching another video, asking AI a question, taking a break, and then coming back to it. But I kept going.

Installing a Harness

I decided to experiment with DeepSeek Harness, which at the time I’m writing this is still very early developer software. Using the terminal, I installed it, and surprisingly that part was easy. The installation was mostly automated, and eventually it spit out a local web address.

I opened that address in my browser and suddenly I had my local AI harness running. That was a pretty cool moment. I also downloaded one of the newer Qwen models through Ollama and connected it to the harness.

Now I had: Local model → Ollama → Harness → Browser interface

One thing I immediately liked about the harness was being able to see more of what the AI was doing. When I asked it something, the interface would change depending on things like the model, context, and tools being used. This was much closer to what I had been trying to understand.

Then I Tried Building Something Useful

I wanted to actually make something, so I decided to try building a small tool for wheelchair users as a DeepSeek plugin. I had seen someone prompt a rudimentary plugin of a whale walking around the bottom of the browser harness, and I wanted to improve on that idea.

My concept was simple. It would have three animations and reminder modes: a 15-minute normal sitting period, a 30-second stretch or reposition period, and a 5-minute full pressure-relief tilt.

For wheelchair users, especially when you’re locked into a computer, game, work, or AI session, it can be incredibly easy to lose track of time. I wanted a little animated assistant that could remind myself and other wheelchair users.

My Error: I Blamed the Harness

Everything was slow. I blamed the harness. Maybe I configured something incorrectly. Maybe the model wasn’t good enough, maybe I installed something wrong, or maybe I needed another tool.

This entire process probably took me somewhere around two to four hours.

On good days I can go incredibly deep into something I’m interested in — the useful side of being autistic, where I can obsess over one subject for hours. But even on those days, I’m still dealing with a mixture of mental and physical problems in the background, and the last 24 hours had already been rough for me.

Then I Tried Hermes

While scrolling YouTube for other AI news, I started seeing people using Hermes Desktop, so naturally I thought maybe this would work better. I installed it, and again, getting it running was surprisingly easy. As long as Ollama was running in the background, Hermes Desktop detected the model I had already downloaded and could use it.

Great. Except it ran even slower than my DeepSeek Harness setup. At that point I started shutting down mentally. I told myself, screw it. From now on I’m just using LM Studio for local models and cloud AI for everything complicated.

That lasted for a little while. I played some games, ate some breakfast.

Then I started thinking again.

I started researching, and eventually I realized something important.

My model technically fit, but that doesn’t necessarily mean it was a good model for what I was trying to do. When you’re experimenting with a local AI harness that is running a main model, sub-agents, and tools, leaving some computational headroom matters. A slightly smaller model that runs comfortably can actually be far more useful than a larger model that consumes nearly all of your computer’s available resources.

The Way I Understand Harness Engineering Now

The model is the actual AI brain. It’s the large language model that thinks, reasons, and writes the responses.

Ollama is the tool that downloads the model onto your computer and keeps it running. It also acts like a translator so other programs can talk to the model.

The harness is the layer that sits on top and gives the model extra abilities that the model didn’t come with, such as external tools and sub-agents. Harness engineering can:

• Give the model tools (search the web, run code, check a database, etc.)
• Add memory and context so it remembers what it’s doing
• Create loops that check whether the answer actually worked
• Spin up smaller specialized agents for specific jobs

Cloud AI Hid This Complexity From Me

Part of the reason this was confusing is that when you use a major cloud AI platform, a lot of this is behind the scenes. You type a prompt, and you get an answer. Behind that simple interface can be models, tools, retrieval systems, context management, safety systems, routing, memory, code execution, search, and other infrastructure working together. When I started trying to recreate even a tiny version of that locally, I started appreciating how much is happening behind the scenes.

Running a local LLM is the easy part. Building on a local LLM is where things get interesting — and complicated.

Where Am I Now?

Well, I never got my animated widget wheelchair pressure-relief tool working, but I understand local AI significantly better now than I did before. I learned what Ollama actually does in the stack, why people are suddenly talking so much about harness engineering, and that tools and sub-agents don’t magically make a local model better. As you start adding more layers, you also add more overhead, more resource requirements, and more places for bottlenecks to happen.

Would I Try It Again?

Yes. First, I genuinely enjoy tinkering with AI, even when it gets frustrating. Second, I have a lot of ideas I still want to experiment with, especially when it comes to using AI to create tools that could actually help disabled people.

I’m definitely not done experimenting with local AI.

1. Sitting Animation — 15-minute timer
2. Tilt/Stretch Animation — 30 seconds
3. Full Pressure Relief Tilt — 5 minutes

I’m still pretty new to tinkering with AI, and I think I may have run into a GPU constraint. I’m going to take another shot at it tomorrow—or maybe later tonight after some gaming and YouTube.