I used o3 to profile myself from my saved Pocket links

After reading this I realized I also have an archive of my pocket account (4200 items), so tried the same prompt with o3, gemini 2.5 pro, and opus 4:

- chatgpt UI didn't allow me to submit the input, saying it's too large. Although it was around 80k tokens, less than o3's 200k context size.

- gemini 2.5 pro: worked fine for personality and interest related parts of the profile, but it failed the age range, job role, location, parental status with incorrect perdictions.

- opus 4: nailed it and did a more impressive job, accurately predicted my base city (amsterdam), age range, relationship status, but didn't include anything about if I'm a parent or not.

Both gemini and opus failed in predicting my role, probably understandably. Although I'm a data scientist, I read a lot about software engineering practices because I like writing software and since I don't have the opportunity at work to do this kind of work, I code for personal projects, so I need to learn a lot about system design, etc. Both models thought I'm a software engineer.

Overall it was a nice experiment. Something I noticed is both models mentioned photography as my main hobby, but if they had access to my youtube watch history, they'd confidently say it's tennis. For topics and interests that we usually watch videos rather than reading articles about, would be interesting to combine the youtube watch history with this pocket archive data (although it would be challenging to get that data).

I've noticed a lot of people are converging on this idea of using AI to analyze your own data, the same way the companies do it to your data and serve you super targeted content.

Recently, I was inspired to do this on my entire browsing history, after reading https://labs.rs/en/browsing-histories/ I also did the same from ChatGPT/Claude conversation history. The most terrifying thing I did was having an LLM look at my Reddit comment history.

The challenges are primarily with having a context window large enough and tracking context from various data sources. One approach I am exploring is using a knowledge graph to keep track of a user's profile. You're able to compress behavioral patterns into queryable structures, though the graph construction itself becomes a computational challenge. Recently most of the AI startups I've worked with have just boiled down to "give an LLM access to a vector DB and knowledge graph constructed from a bunch of text documents". The text docs could be invoices, legal docs, tax docs, daily reports, meeting transcripts, code.

I'm hoping we see an AI personal content recommendation or profiling system pop up. The economic incentives are inverted from big tech's model. Instead of optimizing for engagement and ad revenue, these systems are optimized for user utility. During the RSS reader era, I was exposed to a lot of curated tech and design content and it helped me really develop taste and knowledge in these areas. It also helped me connect with cool, interesting people.

There's an app I like https://www.dimensional.me/ but the MBTI and personality testing approach could be more rigorous. Instead of personality testing, imagine if you could feed a system everything you consume, write, and do on digital devices, and construct a knowledge graph about yourself, constantly updating.

The main thing I learned from my pocket export is that 99% of the articles were "unread". Not sure if it would make sense to extrapolate something about myself other than obsessive link hording from this. :D

a middle aged white guy using AI, my mind is BLOWN

I’ve been really interested in stuff like this recently. Not just Pocket saves but also meta analysis of ChatGPT/Gemini/Claude chat history.

I’ve been using an ultra-personalized RSS summary script and what I’ve discovered is that the RSS feeds that have the most items that are actually relevant to me are very different from what I actually read casually.

What I’m going to try next is to develop a generative “world model” of things that fit in my interests/relevance. And I can update/research different parts of that world model at different timescales. So “news” to me is actually a change diff of that world model from the news. And it would allow me to always have a local/offline version of my current world model, which should be useful for using local models for filtering/sorting things like my inbox/calendar/messages/tweets/etc!

As someone with a family background of more left leaning Catholics (which I think are more common in the US northeast), it's interesting that it decided that you are conservative based on Catholicism.

A while back I made a little script (for fun/curiosity) that would do this for HN profiles. It’d use their submission and comment history to infer a profile including similar stuff like location, political leaning, career, age, sex, etc. Main motivation was seeing some surprising takes in various comment threads and being curious about where it might have came from. Obviously no idea how accurate the profiles were, but it was similarly an interesting experiment in the ability of LLMs to do this sort of thing.

There’s no guarantee this didn’t base the results on just 1/3 of the contents of your library though, right? How can it be accurate if it’s not comprehensive, due to the widely noted issues with long context? (distraction, confusion, etc)

This is a gap I see often, and I wonder how people are solving it. I’ve seen strategies like using a “file” tool to keep a checklist of items with looping LLM calls, but haven’t applied anything like this personally.

Another thing one could do with a flat list of hundreds of saved links (if it's being used for "read it later", let's be honest: a dumping ground) is to have AI/NLP classify them all, to make it easy to then delete the stuff you're no longer interested in.

What was it doing for those 13 seconds? Is it fetching content for the links? How many links could it fetch in 13 seconds? Maybe it is going by the link URLs only instead of fetching the link content?

If you take the 13 seconds of processing time and multiply by 350 million (the rough population of the US), you get:

~144 years of GPU time.

Obviously, any AI provider can parallelize this and complete it in weeks/days, but it does highlight (for me at least) that LLMs are going to increase the power of large companies. I don't think a startup will be able to afford large-scale profiling systems.

For example, imagine Google creating a profile for every GMail account. It would end up with an invaluable dataset that cannot be easily reproduced by a competitor, even if they had all the data.

[But, of course, feel free to correct my math and assumptions.]

I did something similar, but for groupchats. You had to export a groupchat conversation into text and send it to the program. The program would then use a local llm to profile each user in the groupchat based on what they said.

Like, it built knowledge of what every user in the groupchat and noted their thought on different things or what their opinions were on something or just basic knowledge of how they are. You could also ask the llm questions about each user.

It's not perfect, sometimes the inference gets something wrong or the less precise embeddings gets picked up which creates hallucinations or just nonsense, but it works somewhat!

I would love to improve on this or hear if anyone else has done something similar