How to train your AI dragon
Like so many others, I've been watching the technology that collectively makes up "AI" mature - and it seems 2023 is going to be the breakout year (for the general public.)
I recently implemented Whisper.cpp to enable local audio transcription and it works remarkably well. Whisper.cpp is an open-source pre-trained model based on OpenAI's Whisper.
I especially appreciate the options to get word-based timings and the confidence level color-coding of the transcription. As a CLI tool that can work against files or devices (like a microphone), it's handy.
I'm hoping to eventually use it to transcribe my own blog and other marketing pieces without requiring 3rd-party services to do it for me. The word timings allow an impressive amount of visual design options for karaoke-style transcriptions, among the other helpful options to audit text-based audio transcriptions.
And again, it does not rely on a 3rd-party or service to work! Which feels a lot more comfortable when transcribing potentially sensitive audio.
Speaking of 3rd-party services, if you haven't at least given ChatGPT a spin - you should. It can be used for a ton of things - just make sure you fact check any results or content it creates for you as just like the internet - it can be wrong.
Some possible things to use ChatGPT for:
- Write marketing emails or pitches
- Create PowerShell or bash scripts
- Simplify things like regular expressions
- Write responses to emails
- Create written songs, poetry, or other works in specific styles
- Have fun and just chat/learn
- Write (insert language) code-snippets
One thing you should absolutely not do? Copy/paste sensitive information or private information about yourself into ChatGPT. It's free (currently) for a reason.
My passion for learning and leaning into AI is going to revolve around two facets:
- Using open-source AI alternatives
- Training personal and small-business oriented datasets
The goal here is enable myself, my family/friends, and my customers to leverage the capabilities of AI in a more fundamentally private and trusted way.
One tradeoff is that the open-source alternatives like Whisper.cpp, Stable Diffusion, and others generally aren't quite as "cutting-edge" as their private counterparts with VC-backed budgets. I'm okay with this.
Like many other (smarter) people have pointed out, the training of this data can be used for good or evil and the sources behind that data should be transparent and auditable.
These tools should be approached with caution, apprehension, and skepticism. But controlling and training your own personal or business data is one method of making these tools both private and useful.
So that's what I'm shooting for.