I’ve got ~20 years of collected content on technical topics that is all public. I’ve assembled it into a somewhat organized form so that it could be used to train an #LLM to answer questions by referring to and summarizing what I’ve said over the years. I’ve called this #meGPT, and think that it might be useful for other people, consultants, experts etc. I’m sharing freely and don’t want to monetize this, so hope it will be useful as an example training set for research. Edit: lots of interest in this - thanks for the feedback! I’ve just spent some time re-organizing the repo and documenting it better. There is as yet no LLM trained on this content, you have to go old school and watch all my videos etc. to learn it yourself.
There’s some initial discussion of this on Mastodon https://mastodon.social/@adrianco/112895432215168454
How happy are you with the responses from your trained LLM? Does the model really grasp the strategy and informed opinions behind your answers, or does it sometimes respond with an odd recombination of previous individual responses?
Does the LLM know the recipe for Roundtable Pizza?
Adrian Cockcroft would love to chat with you. I wrote about "interviewing myself" here: https://www.meditationsontech.com/p/a-conversation-with-myself... Leveraging a large cache of personal writings from over the years. There is certainly more to be explored behind the "personal LLM" concept.
Very cool Adrian. Classy move. 👌 Much appreciated. By the way, will this be included? My pages are all worn out 😂
Interesting idea. This can be achieved with InstructLab (https://github.com/instructlab/instructlab) You can create taxonomy files with your content and train an existing open source LLM and then publish it on HuggingFace.
Adrian Cockcroft an "app" for every person/expert was an idea we had written PR FAQ on long time ago. Do you envision having a GPT for every person so anyone can query his/her life's work?
Does it know which vintage branded t-shirts to associate with which answers from your greatest hits talk a few years ago? 😀
Is this virtual_adrian2 ?
Tech Advisor
1moI spent some more time yesterday re-organizing the repo to make it easier to add additional authors. Also used ChatGPT to generate some more code for processing the raw content. See the issues for progress on this. For example - here’s a shared ChatGPT conversation about generating code to process podcasts. It wrote 100+ lines of code, and explained how to use various packages, and I spent a minute or two on asking the question. https://chatgpt.com/share/9b8628da-410b-4ced-a6bd-55d523f24815