rileygersh

You're right about the transparency. I am selling training data for iOS 26 - $19.99 for the bundle.

The post demonstrates a real problem (LLMs don't know iOS 26 exists) and offers a solution I created. The urgency is genuine - iOS 26 releases September 15 and developers are finding their AI tools useless for Foundation Models work. Some dev somewhere is praying they don't have to work through Labor Day weekend debugging Foundation Models without AI help. We're all now used to AI assistance - losing it for new frameworks hurts.

Regarding posting frequency - I posted about this in July when I first created it, now again 6 weeks later because the landscape paradoxically hasn't changed: Claude 4.1 and GPT-5 both released post-WWDC yet still don't know iOS 26.

The value proposition is straightforward: Aggregating and structuring iOS 26 documentation saves developers countless hours of searching and piecing together fragments. At $19.99, it's less than 15 minutes of developer time.

I appreciate the direct feedback.

The post could have been more upfront about the commercial aspect while still demonstrating the technical problem.

Web search helps but has drawbacks. For individual developers: each search adds latency and compute costs (API calls, processing, synthesis). The training dataset aggregates all that searchable information once, structures it for AI ingestion, then eliminates redundant processing. Instead of search-wait-paste cycles with inconsistent fragments, you get comprehensive framework knowledge instantly. This captures deeper insights (like constrained decoding architecture and ToolOutput patterns) that aren't easily found through search.

Front-loading the knowledge processing means every subsequent interaction is faster and lighter than triggering new searches.

This touches on an interesting concept - knowledge arbitrage - this information has a shelf life, once the SOTA LLMs know about iOS 26 my files turn into the Beta Max of training data.

Here's how i did it:

https://rileygersh.medium.com/how-i-gave-claude-gemini-knowl...

Updated URL: https://llmbridge.gumroad.com/l/bwoqe

Exactly! We're in a transition period where technical knowledge exists but AI can't access it. This bridges that gap until models retrain. The methodology works for any new framework - Foundation Models just happens to be the current example.

I used AI research tools to systematically extract and organize Foundation Models knowledge from over 100 sites that wasn't in any training data. The value is in the methodology and validation, not just raw output.

Updated URL - https://llmbridge.gumroad.com/l/bwoqe