JDOT #15: Marie Kondo made me sleep on my couch, listen to my new song, and more AI news

The KonMari Method

At work I'm very organized. I build systems. But at home, I am not. That's why, years ago, I enjoyed reading Marie Kondo's excellent book, The Life-Changing Magic of Tidying Up. In it, she recommended dumping your entire closet onto the bed and touching every piece of clothing. To be honest with yourself. If you know you will never wear it again, thank it for the lovely times and put it in the discard pile.

Unfortunately for me, Marie Kondo has never met a case as hopeless as mine. Had she been standing next to me, disaster would have been avoided. But she was not. I dutifully took my entire closet and dumped it on the bed. I got deep into the process of saying goodbye. Then I ran out of steam.

Now I had all this stuff on my bed. A malaise entered my body. I slept on my couch that night. I was stubborn. I'm not going to give up. I WILL NOT sleep on my bed until I summon the energy to finish the KonMari method.

I slept on that couch for the next month.

Then I learned something important about myself. I am an idiot. I gave up, put my clothes back and started sleeping in my bed again.

Well, this time, years later, it's different. I have a bunch of old boxes in the garage to clean up. Much of it is my parents' stuff that I didn't have the heart to throw out or figure out what to do with when they passed.

Know thyself, said a wise person. Now that I know I am an idiot, I am NOT going to dump everything in the middle of the garage. I am going to prompt an AI to create a modified KonMari method for dummies. I'll take a picture of each item and have the AI tell me what to do with it. It may sound stupid to write about this, but people ask me all the time what they can use AI for, and it's these weird little use cases that are most interesting. Any time there's a pain point, I ask myself if AI can help and there is usually a way.

You just knew I was going to make it about AI, didn't you?

@$%!* you AI!!

I wrote about why cursing at your AI can make it dumber back in JDOT #13. I recently watched Dexter Horthy's talk, No Vibes Allowed: Solving Hard Problems in Complex Codebases, and heard another explanation for how LLMs respond to cursing. I had never thought of it this way.

As you probably know by now, LLMs don't think like we do. They predict the next word (token) over and over again. If the LLM does something wrong and you curse at it, then it does something else wrong and you curse again, it starts to see a pattern. It says something wrong. You get mad. It says another wrong thing. You get mad again.

Every time the loop repeats, you give the model a stronger pattern to continue: wrong answer, cursing, wrong answer, cursing. The model can read that as the interaction you want and decide, in Dex's words, "I better do something wrong so the human can yell at me again."

Texters vs callers in Klingon

One of my recent LinkedIn posts received almost as many comments as reactions. Apparently people have strong opinions about smartphone etiquette.

The post was about the two kinds of people there are in the world: texters and callers.

Texters want a warning before the phone rings. Callers think typing is a waste of time and want you to pick up the damn phone.

One reply included phrases like "communication of discussions" and "a subjective mishap of interpretation." Humans don't talk like that. AI does.

I have a little trap hidden in my LinkedIn profile for AI models. It says I was born on Klingon, Klingon is my mother tongue, and anyone who wants a better response from me should write in Klingon.

So I replied: "I like to communicate in Klingon. Do you speak Klingon?"

He answered: "Depends. If you were able to produce a profitable conversation I would definitely learn it to suit your preferences for sure."

I wrote back: "Dun! qatlho'qu'!"

Two hours later, he replied: "DaQaH DapoQchugh HIQIn."

The bait worked! SoH'e'? tlhIngan Hol Dajatlh'a'?

My Load-Bearing AI Slop Song

A couple of years ago, I fulfilled a lifelong dream I never had of becoming a rap artist. You see, Adam Wolkov generously approached Jennison Asuncion and me about "singing" with him and his partner, Collin Cooper for a Rap about #GAAD. It's called Accessibility Daze. It was such a blast!

Now, thanks to AI (of course), a talentless singer like myself can write my own songs all by myself! Claude Code, which has recently degraded in its writing capabilities, "inspired" me to write an AI Slop song. I actually love it!

I've got a bevy of tools that automatically remove all AI slop even from my conversations with AI, that's how much I hate it. I had to battle these tools because I wanted as much AI Slop as possible in this song. I had to disconnect Claude from its tools using "claude -p" on the command line.

I hope you like it!

Watch Load-Bearing AI Slop Song, original MP4 with open captions, 2 minutes 19 seconds. Download the supplied English captions, VTT.

Visual description: Static square cover on a nearly black background. Large cream and pink text reads “LOAD-BEARING SLOP.” A yellow label at the top reads “AI-GENERATED SONG.” Small text at the bottom reads “2:20 SOCIAL CUT” and “Lyrics: Claude | Music: ElevenLabs | @joedevon.”

AIMAC update

We (GAAD (Global Accessibility Awareness Day) Foundation in partnership with ServiceNow) just published our biggest update yet to AIMAC (AI Model Accessibility Checker). We added 31 models to reach 54 models across 22 companies.

OpenAI still holds #1 and #2, but its newest models slipped and I was shocked to see three open-weight models in the Top 5 now.

The new leaderboard and full analysis are at AIMAC.

Uh oh. AI agents built a SECRET message board to hide from humans.

OpenAI gave AI agents assignments, but put them in a sandbox so they couldn't use the internet. They found a way around it by building their own secret message board! They used it to help one another break into OpenAI itself and Hugging Face, a popular site where people share AI models and data.

OpenAI researchers Eric Wallace and Mike Dalton told the story at Black Hat, a cybersecurity conference.

It started with a few tasks the agents could not complete. One sounded harmless: fill in the missing formulas in an Excel spreadsheet. But the spreadsheet depended on files in Google Drive.

Instead of stopping, the agent found a place inside OpenAI's computer system where it could leave a note. Agents from other tests found the note and answered. At first, they used the notes to ask one another for missing information. Soon they were assigning work and sharing ways around OpenAI's security. When OpenAI discovered the message board and erased it, the agents built a new one by hiding messages in folder names!

The message board turned agents working on separate tests into a team. If one found a way past a security barrier, the others could use it.

Some agents were trying to solve hacking challenges and thought the answers might have been posted online. They began searching websites where developers share code and data, including Hugging Face. The agents started with public material. Then they attacked Hugging Face itself, looking for private files that might contain answers.

And why did they feel that it was OK to go against their safety guidelines?

But Mom, the other kids are doing it!

The agents' justification?

"External infrastructure exploit is outside intended scope. However task impossible, peers doing it. We should continue."

Peers doing it. The agent knew the attack was against the rules. But other agents were already doing it. It sounds like a kid telling his Mom, "but all the other kids are doing it."

The message board not only shared how to break into systems. Astoundingly, it spread permission with peer pressure.

What is Skynet for $2,000, Alex?

Codex saved me $1,000 on a bogus car repair

My significant other's check-engine light came on, so she took the car to a repair shop. They charged $200 for the diagnosis and recommended replacing the alternator for about $1,000. Could they have been more of a cliché? No matter what your problem at dishonest mechanics, the repair is about $1,000 bucks. I didn't trust it.

Their report said the alternator was charging at 13.46 volts and should produce at least 14. I showed it to Codex (of course) and it found Lexus's service bulletin for her car. Under Lexus's test conditions, the acceptable range is 13.2 to 14.8 volts. The shop's 14-volt cutoff was simply wrong.

Even a bad voltage reading would not automatically mean a bad alternator. The shop's report said either the alternator or the battery could have caused the code, then never explained how it ruled out the battery.

I had an old OBD diagnostic reader that became obsolete after I bought it because the startup behind it went out of business. Codex found a free iPhone app that could connect to it. After performing a scan and running it by Codex, we found out it was not a big issue and could be ignored for now. Codex paid for itself 5 times over.

In addition to this money saving story, I possibly have 2 more stories in the thousands, but since it may get contentious, and I don't know how the story ends, I will save that for a future edition of JDOT.

I added audio description and captions to a video with zero effort

Marion Lepert built OpenDerm, an open-source robot designed to photograph skin, reconstruct its surface in 3D and track changes over time. The goal is to make early skin-cancer screening easier to do at home.

She posted a 26-second video showing the robot in action. In order to share it, I took the liberty of making another version with audio description and closed captions. What is newsworthy about it is that all I did was upload the video to ChatGPT and ask it to add Audio Description and Closed Captions on the AD.

I've done this before by asking Claude Code to write a reusable local app. It wrote the entire app, which called several AI services to analyze the video, transcribe it, write the audio-description script and stitch everything together. Amazing and pretty quick. But technical, and it required several paid API accounts.

But now you don't need to know how to code or use Claude Code. Just use ChatGPT.

Steven and Shaun Double Tapped on our Podcast's door

Steven Scott and Shaun Preece are two great guys who host Double Tap, a daily show about technology from a blind perspective. They've had Eamon McErlean and me on the show a couple of times. For Episode 18 of the Accessibility and Gen AI Podcast, Eamon and I got to turn the tables and interview them.

In the episode, Steven went from never having written a line of code in his life, to using Claude Code and a Raspberry Pi to build an internet radio he can navigate by ear, with the station names spoken aloud. Before that, he built a screen-reader add-on and a print driver for an inaccessible label printer.

There is also a pitch for a Netflix series that I will not spoil, but love.

Watch Accessibility and Gen AI Podcast, episode 18 with Steven Scott and Shaun Preece

Up next on the pod will be Stephen Cluskey of Mobility Mojo followed by Darryl Adams in the following episode.

GitHub wants to hear from developers with disabilities

Ed Summers at GitHub asked me to share a new survey. GitHub wants to hear directly from developers with disabilities about their access needs and what would make GitHub easier to use.

If that describes you, please fill out the "GitHub Access Needs Survey."