I won’t even start reiterating all of the fucked up things “those” companies have done to train their ai. It’s out there. It’s documented. Whether anything is ever going to be done about it … who knows? I don’t know that I have anything insightful to add.

That being said, I have to use these tools when I work. They have been foisted upon society, and it’s a matter of using it, learning it, applying it, or falling behind and being consumed. It’s a weird place to be when I understand and agree with a lot of the criticism of ai. I also see the practicality and usefulness of these tools in many day to day tasks.

One of the uses I've had for ai outside of work (I currently use Claude) is transcribing text from images. For some time I have wanted to digitize dozens of journals I handwrote throughout my late teens, twenties, and early thirties. Yet whenever I told myself I was going to buckle down and begin this project I'd be overcome by the thought of typing all of this material. It is a monumental task. It's a vanity project requiring re-reading, and typing, so many cringe-inducing entries (and bad poetry) that the thought of the energy needed to accomplish this has been enough to extinguish any start.

Until recently. Enter Claude as a solution. I photograph the journal pages, 20 at a time, upload with a simple prompt. Claude analyzes, transcribes all 20 images worth of text, then points out words or other issues it doesn't recognize. The text is added to a doc, reviewed, and I'm done.

It's not perfect. I do have to review the text because Claude will make mistakes. But in the end it saves me an enormous amount of time. And it comes with the added benefit of text analysis. Just what LLM's are perfectly situated for. Once complete I can have Claude analyze the text, and then I can ask it questions, look for insights, etc.

Any ai can do this. And it's amazing they can. It's easy to forget how only a short time ago image to text transcription with this level of fidelity was not a thing. It transcribes handwritten text incredibly well, which I can attest, mine was quite awful at times seeing how often I enjoyed writing in my journals when I was drunk.

The Analysis Part

As a test when I first started this process I had Claude analyze two of the journals it transcribed. I asked Claude to provide me a social and psychological profile of the person who wrote those two journals.

At the time of this analysis I had just begun to dig into LLM capabilities. I worked with Claude during the day at my place of employment, figuring out how to run automations and ways to instruct Claude that gave decent results. To continue that learning I started a personal subscription at home. While it did do some cool and amazing things at my workplace, when it spat back its social and psychological profile I was pretty much stunned.

Let me note here that I'm well aware of ai-induced psychosis. I'm well aware ai's are helpful tools that can be overly agreeable, authoritative, and completely full of shit. You have to take them with a grain of salt. Verify the facts. In this instance the facts were the journals, and my own understanding of who I was then and now, coupled with the unreliability of memory, false-memories potentially induced by ai suggestion, and any other kind of mental-ego fragility that can exist in a examination like this.

I imagine most of us have marveled at the LLM's abilities, how life-like it can appear in its response to your queries. It doesn't seem possible that, in simplistic terms, a statistical probability algorithm should be able to formulate a response to your query with such life-like meaning and depth. And yet it does just this, providing you something that represents understanding. It's deceptive. It doesn't have understanding, yet it communicates meaning.

Claude analyzed my journals and came back with something that was meaningful in it's analysis. As an example, one of the first items it provided was a mental health diagnosis of the person I was when I wrote those journals. It was the exact diagnosis my counselor gave me only two years ago. It wasn't lost on me that had I had a tool like this way back then, there was a chance I might have taken it to an actual doctor to see if it was accurate.

And it wasn't only this diagnosis. It was the correlations and connections it made between the things I was doing, experiencing, and writing about. Many were spot on, many were illuminating. "Observations" framed in ways I hadn't thought of or acknowledged. Some of this was sobering. Some of this was overly flattering as only an ai can be. And some of its observations were outright wrong.

The Tricky Part

When I think back to the time period of these journals I'm greeted with a series of vignettes and a undulating sense of immense uplift and deep darkness. I see it now as a tumultuous time. When I chain together these impressions and vignettes, I'm left recalling a sense of excitement and terror concerning the future. That all was open to possibility and I could still believe in dreams that now, in hindsight, seem farfetched in terms of what I was doing then with my time. And while all was open to possibility, it all seemed so fucking futile and pre-determined.

It's all soft around the edges, this sense I have of then. When I read the actual journal entries it appeared differently. Those soft edges fell away and were replaced by stark lines, contours framed in concrete. They had weight and surfaced from soft fuzz to focused image. What the pages allude to captures some of that sense of excitement, but it wasn't romantic. The romance was applied by memory long after it had been lived. What the journals attest to is a person mired in deep, almost debilitating depression, somehow managing to pull off a semblance of a life.

So there's that dichotomy. The gravity of the pages opposite my close to thirty year old patch-quilt memories, along with the knowledge that when I was writing the pages I was also unconsciously fictionalizing to an invisible audience. All of which makes it off-center from reality. Ultimately my journals were a place to practice writing, and I was attempting to learn, and dramatize, and tell a story as I was writing about my daily life.

Now there is the ai analysis. It is spot on about much I can see in hindsight. Yet it also has a way of compressing the analysis into a drama. It's like reading the synopsis of a character from a book with a beginning, middle, and end. Of course, the journals have all three, but they're open ended. They don't have a sense of conclusion or a self-contained theme. They simply exists as days upon days being lived. Things ebbing and flowing. The structure is the calendar, nothing more. The ai analysis comes across hella convincing, but it's dramatized.

With these differing perspectives playing for and against one another, I'm left questioning if I can trust the ai analysis? I can see where it is wrong, but much of it feels legit. Yet, is it also influencing the way I perceive that time? For instance, if I say to myself, "Hmmm, I never thought of that Claude, perhaps you're right?", am I being convinced of a fiction because it sounds convincing and authoritative? Do I feel flattered by how much it talks about this young mans sensitivity and intelligence, and does that make me want to believe what it is communicating?

Of Course it Does

Even though the person in those journals was a trainwreck, Claude's breakdown makes him seem a little bit more heroic, a bit more sophisticated and romantically tragic then what I think he may have been. Claude gives him a story that envelopes these days, and it is wrapped up as if a definitive chapter in a life has concluded.

It's easy to give in to a more flattering portrait of the self. It's like when you try to have a realistic determination of something you did, and someone tells you, "Hey, you're being a little harsh on yourself", and you second-guess your conclusion. Maybe I am being too hard, and maybe they're right? They're on the outside, they don't have as much emotionally invested. Perhaps they see it clearer than I?

Claude doesn't have a perspective even though it communicates as if it does. It has statistical probability.

So What's this all About?

It will be interesting to see what Claude provides me once all of the journals are transcribed, and it has access to the full breadth of the journals. It will be interesting to see how the analysis differs also due to the continual improvement of the models.

So what's this all about?

The short answer is Claude works great for a practical application at home: image to text transcription. It saves me an enormous amount of time, and it even provides analysis. That analysis would be easy to accept without any qualms, but it leaves me wrestling with a number of potentialities that color it.

It's interesting to consider how unreliable memory is and then compare those colored memories to the picture painted by the actual journal entries, and then compare that to what Claude drums up from its profile. They all hold shades of something approximating the truth, overlapping and diverging, and they all embody shades of grey.