I gave AI access to 5 years of my personal journals
The results were positively startling
In the grand tradition of humans finding increasingly creative ways to avoid genuine self-reflection, I decided upon a new approach.
By feeding over five years of my personal journals to an artificial intelligence.
Because why confront your demons yourself when you can have a silicon-based life form do it for you?
The Setup
Picture this: A man in his black fabletics gym shorts, the ones that only cost $10 and are perpetually on sale, surrounded by stacks of notebooks filled with the kind of handwriting that would make a doctor proud.
That's me, realizing I've accidentally created a paper-based time capsule of my own neuroses. And lacking the willpower to sit down and read them myself.
Five years of hopes, dreams, and the occasional drunken scribble, all waiting to be deciphered like some millennial Rosetta Stone.
Now, any sane person would have shoved these journals into a box, tucked them away in an attic, and promptly forgotten about them.
But no, not me. I decided to embark on a journey that would make Freud turn in his grave and Silicon Valley venture capitalists salivate simultaneously.
I was determined to consult a therapist. But not one that requires 5 expensive consultations before telling me what I already knew.1
One that reads Every. Single. Word. I’ve ever handwritten.
(And peers deeply into my soul).
The Great Digital Excavation
Step one. Digitize every single page.
I photographed each page with the reverence of a monk transcribing ancient texts.
I snapped photos of my feverish, emotion-filled, post-breakup scribbles, web-design mockups from hackathons or startup competitions of college years past, and prayers or gratitude notes from years of meditation.
With my digital treasure trove in hand, I turned to the next challenge: transforming my chicken scratch into something a computer could understand.
Though modern LLMs like OpenAI ChatGPT, and Anthropic Claude boast of multimodal image analysis capability, this hardly extends to dense, paragraph filled notebooks.
Enter Optical Character Recognition (OCR), or as I like to call it, "Handwriting Guessing AI."
My first stop was to check the popular online model repository, Hugging Face.
It’s like GitHub if the only code available were AI models, and those models were all ranked in an Olympics of turing tests.
Though Hugging Face showcased multiple popular OCR models, the vast majority of them were only capable of analyzing 1-line of text at a time.
In short, to use these models, I would first need to custom-code a way to split my images into clear-cropped lines of text.
Far too robust and enterprise-grade for my hacky solution.
No-can-do. Back to square 1.
My next Google Search for OCR offered a better option (probably biased in Google’s own favor).
This was to use an application programming interface (API) Google offered in their cloud service.
This API would accept whole images and return text using credits which Google supplies as a free trial.
Perfect.
I created a personal Google Cloud Platform (GCP) vision instance and got started.
I uploaded all my images to my private cloud.
(This is one of the most personal privacy risking steps of the whole process).
Despite Google’s assurances that my data is completely secure, I still find myself very wary uploading all this to the internet.
Next, I cobbled together some python code in the Google Cloud console and ran it across all my images.
Here’s a look at some of the vision code.
If a deeper-dive into this process appeals to you, please reach out to me at the email or LinkedIn below. I’m happy to give you the code or explain in more detail.
The result? A 316KB text file of every word I had ever written by pen (result.txt).
By my napkin math, that’s roughly 161,792 words, or 701 Moleskine pages.
We did it!
Enter the Silicon Shrink
Now came the pièce de résistance.
I took this digital word salad and fed it to Claude 3.5 Sonnet, an AI so advanced (at time of writing) it makes Siri look like a speak-and-spell.
This digital demigod boasts a memory that can handle 200,000 tokens at once.
A fair bit better than GPT-4o, which has a 128,000 token limit.
To put that in perspective, that's about 199,999 more things than I can remember on any given Tuesday.
Importantly, this means I can upload the entire file to Claude, without exhausting all the AI’s memory in the process.
As you can see Claude is capable. Grad-school level capable.
(Which is plenty more intelligent than me).
When you begin a new conversation, it shows.
But here's where it gets interesting.
I didn't just want a plain old AI analysis. No, I wanted therapy.
I remember very little of the 2 psychology classes I took in college, but thankfully, Claude remembers all of them.
So, I instructed Claude to put on its imaginary tweed jacket and become a family therapist.
First I asked, “what are the leading methods of therapy which you studied in school?”
Next, part 2 drops in a few days. In it, I “prompt engineer” the AI to ask deeply personal questions:
What are my weaknesses?
What are the goals I have set for myself over the years?
Do any goals conflict with each other?
How can I maximize my self confidence and minimize my self doubt?
The answers are positively startling.
If you would like to chat about implementing your own AI, shoot me a message on LinkedIn or lucas@lucaserb.com.
Helping businesses until the singularity,
~Lucas Erb
I believe deeply in the mental health benefits of proper, certified therapy and counseling. I think doing this right requires a human touch. Also, I think a significant benefit of therapy is derived from helping a patient think through their own crises and past struggles. For this portion of the therapy process, we may be able to achieve a greater mental health benefit by pairing continuous AI consultation with comparatively infrequent trained therapy sessions. For the substantial step of tailoring to the patient’s needs and watching over their next steps with care, I see no better fit than a trained, human, professional.