Alexander Hayes

AirTrain now has an interactive terminal. type commands, watch your MacBooks train AI together with live animations. connects to Ollama for real AI-generated dream training. one python file, no dependencies. Anyone should be able to train AI models

@9ZTvsmda6 I’m not sure about anything related to the token but I’m still committed to the project!

@AlexanderCodes_ Has your coin developer run away? Why did it plummet to zero?

Check out the latest changes on the Github repo here! ~800 new lines of code github.com/alexandercodes…

Showcasing my MedScan model I built on top of Ollama 3.3 and trained using AirTrain: I analyzed patient symptoms in real-time by training it for FREE across multiple MacBooks on available chest x-ray data online. It reads symptoms, generates a medical scan visualization highlighting areas of concern, and outputs a full triage assessment with urgency level and recommended next steps. Models like this typically cost $1000s to train. With AirTrain, a group of med students could train one overnight on their laptops for $0.

Goodmorning everyone! Finished training a custom multi-modal model I built off of Ollama 3.3. I also added custom ASCII rendering and scanning animations for any image analysis models in the AirTrain CLI. I’m going to put out a demo soon to show this off Super excited about this!

Don’t worry everyone I’m still at it 😤 I’ve got some big plans and changes in store for tomorrow. Goodnight ❤️

Just a reminder for how much people are paying to train or even run models locally. Another reason to use AirTrain

Check out the details of the latest release here: github.com/alexandercodes…

Currently working on a much cleaner and user friendly finish for AirTrain. I want to make this easy for everyone to understand, not just technical people. Think Claude Code but for AI training. Going to release it shortly

If anyone has any cool applications or recordings of using AirTrain please send them my way! DMs are open I’m looking to put up a nice portfolio on the AirTrain page to get some potential investors involved in the project.

Another ~500 lines of code added while working and training a model for a side project 😄 github.com/alexandercodes…

Newest AirTrain feature, Gradient Marketplace: Not all contributions are equal so now AirTrain scores every worker's gradients on quality (alignment, magnitude, consistency, impact) and gives better contributors more influence. Bad data or failing hardware? Automatically downweighted. Good data? rises to the top. AirTrain is the first distributed training framework with quality-weighted aggregation. Read more on it here: #marketplace" target="_blank" rel="nofollow noopener">alexandercodes4.github.io/airtrain-web/#…

Building a new feature already while putting AirTrain’s Dream feature to the test on my spare mac 🤓 In addition to compensating for high computing costs, AirTrain’s dream feature also compensates for inadequate training data: It uses the model to “dream” up new data, then AirTrain grades it’s quality score and if it scores high enough to be used, it will be saved to your Dream Cache. Dreams are then used to supplement additional quality data that you wouldn’t have had otherwise, to further increase your model’s accuracy. Open-source and free for anyone to use! Just pip install airtrain

Another ~900 lines of code added! A little less than I'm usually capable of, but I'm building the best part of AirTrain at the moment and I'm planning on officially releasing that later today 🤫 github.com/alexandercodes…

I've also edited the layout of the landing page to include the demo and show off everything nicely. Huge thanks to my buddy Nico from @oracle for helping me with the frontend design! He's a beast.

Check out the Model Autopsy details here! #autopsy" target="_blank" rel="nofollow noopener">alexandercodes4.github.io/airtrain-web/#…

Thank you for your patience everyone. I hope everyone's already using AirTrain to train their models! I just finished up our newest feature, Model Autopsy: After training finishes, AirTrain generates an interactive report showing your model's entire life story. Who contributed the most compute, which sync rounds caused the biggest breakthroughs, how dream training impacted convergence, and a full loss curve. every contributor gets credit. Try it out by using: "airtrain autopsy" in your terminal

Goodmorning everyone! Already building again today, I’ll update everyone shortly 😁

~900 new lines of code 😎 Goodnight everyone, already working on the next feature and I'll have it out by morning! github.com/alexandercodes…

Announcing Dream Training: When your MacBook is idle between training sessions, it "dreams" - running the model to generate synthetic training data, scoring it for quality, and caching the best samples. when training resumes, dream data is mixed in to speed up convergence. and if your Mac can't find a session to join at night? it dreams instead. no idle time wasted. #dream-training" target="_blank" rel="nofollow noopener">alexandercodes4.github.io/airtrain-web/#…