Howdy,
With the prices of AI Tokens going up and the “blacklisting” of Mythos/Fable5 I got to thinking maybe I should fine-tune a local LLM model on AutoHotkey.
I started thinking about what I would use to train it.
I have over 2,000 AutoHotkey v2 scripts, transcripts to 6 AHK v2 courses, Over 80 AHK Skills/Agents, the AHK v2 documentation, and ~ a dozen well-written AHK Git repositories.
I worked with Claude code to decide how to train my model with my Nvidia 3060. It took quite a bit of time but I came up with a model that should run on most computers.
I made it accessible to AHK Script Maker and saw it was generating valid AHK v2 code.
But then I wondered “How does my locally trained model stack up against Flagship models like Opus, Chat GPT, Grok, etc.
So I had Claude develop 150 syntax tests and had it loop over 12 models (you can learn more about the process and results here)
The results were fascinating❗
While I regularly use Claude’s Sonnet and Opus models, they’re not cheap.
I know Isaias heavily leans on Chat GPT’s 5.4 model (again not cheap.)
I was (somewhat) surprised to see Grok 4.3 emerge as a clear winner for both Quality of code and Price.
Granted, the price is artificially low because X-AI is heavily subsidizing it but both Anthropic and OpenAI are also subsidizing their models.
After training (which didn’t cost much) my local fine-tuned model was free and private. It also delivered the cleanest AHK v2 code (96% of code was v2 code and 4% was v1) of all the models❗
Quality vs Price
I always included Grok in my “guru” list (when I ask other AIs to brainstorm on my plan or for troubleshooting) but I’m going to start using Grok more for my day-to-day coding.
How about you❓ What is your favorite model for developing AHK v2 code?
Now on with the show…
 |  |  |  |  |  |  |
Don’t get embarrassed when Sharing your screen
We recently learned how easy it is to “hide” a GUI from screen capture tools like Zoom, OBS, etc.
Since I am often sharing my screen in Zoom, I thought it would be cool to be able to hide my notifications that appear on the screen.
So with a few minor tweaks we updated our Notify Class so you can easily hide them from screen capture tools. Check it out in this video.
Hidden Screen Notifications Nobody Can See While Sharing | Hero Extrac
Make your message ❗
Are you trying get to noticed and don’t want to spam Emojis❓
Our Noticeable Text script can give you some sʎɐʍ ƃuᴉʇsǝɹǝʇuᴉ ʎɹǝʌ to make your text ⓖⓔⓣ ⓝⓞⓣⓘⓒⓔⓓ.
Take a look how easy it is in this video.
This AutoHotkey Font Trick Will Make You Stand Out Everywhere | Hero Extract
Improving People’s Lives: Story #103
I had a call with Allison (another person I met from the Sabrina AI meetup in Vegas.)
We had a great time talking through how she could use AI (and automation) in her real estate business. Sadly she’s a Mac user so I didn’t show her any of my AHK stuff but it was still a really fun call stimulating ideas in what we could do in our businesses.
Have you sat down with someone to discuss using AI in yours❓
What we’re reading
- Why everyone’s jumping ship from Plex
- Ditch Windows Phone Link for a free open-source app
- Stop buying Consumer hard drives
AI News that you Need to Know
- Google Chrome is silently stashing a 4GB AI model on your device
- Use Elon Musk’s ‘Relevance Rule’ to fix your memory
- Your AI Video Agent
⚡️Productivity tips:
HotStrings are, by far, the most under-utilized ways to be more productive.
However many you have, trust me, you don’t have enough❗
A spot of Humor:
I wonder how many people would fall for this these days
Candid Camera Classic: Texas Is Closed Today
What we utomated this week with AutoHotkey #137
I packed a ton into this week’s session, from fine-tuning a local Qwen model on thousands of AutoHotkey scripts so you can code AHK without cloud costs, to running Parakeet speech-to-text on my GPU that spits out SRT files at 20-30x speed.
I also show off a concept-based document search that finds PDFs, videos, and audio by meaning instead of keywords, a Token Burn tool that benchmarks Haiku vs Sonnet vs Opus on cost and success rate, plus a custom file explorer, Quick Raw Edit for fast video cuts, and a pile of other utilities.
In this week’s video I walk you through all of them and how they fit together.
What we Automated with AI and AHK #137
Clients say…
The original script you created for me was so robust I haven’t changed it in over 2 years. It just works! — Michael N. (Radiologist)
AHK Hero
Consider joining the Hero club! Members receive 25% off courses, consultations, tutoring, done for you project work, etc. Currently we have over 581 hours of recordings teaching AutoHotkey that is exclusive to Hero members❗
You can watch 1 hour a day for a year and not repeat content! That’s a lot of Learning!
During this Friday’s call we’re planning to discuss:
- Help Hero Members with their Questions & Code
- Adding a progress bar to videos
- New AHK tools
- File Explorer
- Taskbar
- Start Menu
- Phone Link
- AutoRuns
- Add to Startup
- Java inspector
- LLM Model Training and comparison
- AHK Script Maker – local Fine-tuned model
- Quality versus performance
- Privacy
- Add to Startup skill & adding to all my scripts
- Git repositories
- Unlimited OCR
- Voice Cloning
- Custom version of Handy customized for GPU
They also have access to exclusive & pre-released content and 3 hours a week where we help people with their automation issues (the Saturday calls are geared towards people new to Automation with AI)
️ Quotable quotes
- Being born in poverty isn’t a sin; staying in poverty might be❗ — Dan Kennedy
- Determine that the thing can and shall be done, and then we shall find the way — Abraham Lincoln
- I’m not going to buy my kids an encyclopedia. Let them walk to school like I did — Yogi Berra
⏩SHARE WITH A FRIEND⏩
Do you know someone that would benefit from learning useful automation tips like the gems in today’s issue? Take a second and forward this newsletter and bask in the glow of their undying gratitude. Was this email forwarded to you?
Sign up for this Newsletter here. We never share your email with anyone. Ever!
the-Automator Dedicated to Multiplying your Income
Joe, Isaias, Irfan and staff
 |  |  |  |
P.S. When I was training my LLM on AutoHotkey I was asking AI if I had a GPU with more VRam (and could have a larger model) if I could get closer to Frontier model performance like Grok & Opus.
It said “not really” because those models have been hardened with an amazing amount of “logic and thinking” for solving problems.
So the small model I made, for now, is the right sized-model for data you want to keep private.
