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).
You should be able to use Google Takeout to get all of your YouTube data, including your watch history.
This article is a nice example of someone using it:
> When I downloaded all my YouTube data, I’ve noticed an interesting file included. That file was named watch-history and it contained a list of all the videos I’ve ever watched.
Yes I've done this in USA. pretty neat. I have it on my todo list to parse over it and find all the music videos I've watched 3 or more times to archive them.
You need to use an iterative refinement pyramid of prompts. Use a cheap model to condense the majority of the raw data in chunks, then increasingly stronger and more expensive models over increasingly larger sets of those chunks until you are able to reach the level of summarization you desire.
re o3: you can zip the file, upload it, and it will use python and grep and the shell to inspect it. I have yet to try using it with a sqlite db, but that's how i do things locally with agents.
Author mentions that by doing that they didn't get a high quality response. Adding the texts into model's context make all the information available for it to use.
I think a reasoning/thinking-heavy model would do better at piecing together the various data points than an agentic model. Would be interested to see how o3 does with the context summarized.
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
Perhaps comparing your read/unread might tell something about your revealed vs stated preferences. I assume that the typical person's unread pile is mostly aspirational. I'm sure that there's lots of data on this - for example Amazon's recommendation graph may weigh our Wishlist items differently than our Purchased items.
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.
> Instead of optimizing for engagement and ad revenue, these systems are optimized for user utility.
Are they, or instead they will help keeping you in your comfort cage?
Comfort cage is better than engagement cage ofc, but maybe we should step out of it once in a while.
> 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.
Curated by humans with which you didn't always agree, right?
That's the core challenge in designing a system like this. Echo chambers and comfort cages emerge from recommendation algorithms, and before that, from lazy curation.
If you have control over the recommendation system, you could deliberately feed it contrarian and diverse sources. Or you could choose to be very constrained. Back in RSS days, if you were lazy about it, your taste/knowledge was dependent on other people's curation and biases.
Progress happens through trends anyway. Like in 2010s, there was just a lot of Rails content. Same with flat design. It wasn't really group think, it just seemed to happen out of collective focus and necessity. Everyone else was talking/doing this so if you wanted to be a participant, you have to speak the language.
My original principle when I was using Google Reader was I didn't really know enough to have strong opinions on tech or design, so I'll follow people who seem to have strong opinions. Over time I started to understand what was good design, even if it wasn't something I liked. The rate of taste development was also faster for visual design because you could just quickly scan through an image, vs with code/writing you'd have to read it.
I did something interesting with my Last.fm data once. I've been tracking my music since 2009. Instead of getting recommendations based on my preferences, I could generate a list of artists that had no or little overlap with my current library. It was pure exploration vs exploitation music recommendation. The problem was once your tastes get diverse enough, it's hard to avoid overlaps.
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!
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.
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.
That's pretty cool! Now I can imagine a tool that gives you a prediction before you even post and then offers suggestions for how to increase performance...
Gotta hand it to SamA for not only selling the problem but also trying to cash out on the solution (verified human via creepy orb eyeball blockchain thingy)
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.
Born in Pittsburgh, raised Catholic, pretty darn liberal. We had alter girls in the 90s, openly gay members who had ceremonies in the church, etc. I'm not catholic now but that was a good church in the 80s and 90s.
> Fiscally conservative / civil-libertarian with traditionalist social leaning
And justified it with:
> Bogleheads & MMM frugality + Catholic/First Things pieces, EFF privacy, skepticism of Big Tech censorship
First Things in its current incarnation is all about religious social conservatism. If someone is Catholic and reads First Things articles, "conservative" is a pretty safe bet.
However, I think profiling people based on what they read might be a mistake in general. I often read things I don't agree with and often seek out things I don't agree with both because I sometimes change my mind and because if I don't change my mind I want to at least know what the arguments actually are. I do wonder, though, if I tended to save such things to pocket.
I have a hypothes.is account where a decent amount of my annotations are little rage nits against the thing I'm reading. You'd be able to infer a ton of correct information from me if you pulled the annotations as well as the URLs, but the URLs alone could mislead.
I've had to remind myself of this pattern with some folks whose bookmarks I follow, because they'd saved some atrocious stuff – but knowing their social media, I know they don't actually believe the theses.
I would say in aggregate, both Catholics and Protestants (whichever flavor) are more likely to be liberal in the northeast / west coast and more likely to be conservative in the midwest / south. Which tells you something about the average importance of religion in 2025.
I think it's older than 2025 and definitely has a piece of it that is specific to Catholics. I tend to think of northeastern American Catholicism from the lens of immigration. The big waves of Italians, Irish, Eastern Europeans, etc. The immigrant identity often led to left leaning economics and the parts of Christianity which are about helping the poor get emphasized.
Idk how much experience you have with catholics outside of the northeast. I have a fair amount with all of the regions I mentioned (northeast, south, midwest, west coast). You cannot really find any American Catholic parish that is not dominated by at least one of Italians, Irish, Eastern Europeans or Hispanics. The catholic church in the US is mostly "immigrants," that is, people whose ancestors were not in the US prior to ~1850
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.
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
There are other good use cases here like documenting recurring bugs or problems in software/projects.
This is a good illustration of why e2e encryption is more important than its ever been. What were innocuous and boring conversations are now very valuable when combined with phishing and voice cloning.
OpenAI is going to use all of your ChatGPT history to target ads to you, and probably will have to choice to pay for everything. Meta is trying really hard too, and already is applying generative AI extensive for advertiser's creative production.
Ultra targeted advertising where the message is crafted to perfectly fit the viewer mean devices running operating systems incapable of 100% blocking ads should be considered malware. Hopefully local LLMs will be able to do a good job with that.
> EU-based 35-ish senior software engineer / budding technical founder. Highly curious polymath, analytical yet reflective. Values autonomy, privacy, and craft. Modestly paid relative to Silicon Valley peers but financially comfortable; weighing entrepreneurial moves. Tracks cognitive health, sleep and ADHD-adjacent issues. Social circle thinning as career matures, prompting deliberate efforts at connection. Politically center-left, pro-innovation with guardrails. Seeks work that blends art, science, and meaning—a “spark” beyond routine coding.
Fairly accurate
"Seeks work that blends art, science, and meaning—a “spark” beyond routine coding."
It would be much funnier and/or insightful if it sampled more than the first page of user comments.
Still, spot on:
Predictions
Personal Projects
After a deep dive into archaic data storage, you'll finally release 'Magnetic Tape Master 3000' – a web-based app that simulates data retrieval from a reel-to-reel, complete with authentic 'whirring' sound effects. It'll be a niche hit with historical computing enthusiasts and anyone who misses the good old days of physical media.
> Your comments often feature detailed technical explanations or corrections, leading me to believe you're either a deeply passionate technologist or you just love being the smartest person in the room. Probably both, let's be honest.
"You'll discover a hitherto unknown HN upvote black hole, where all your well-reasoned, nuanced comments on economic precarity get sucked into oblivion while a 'Show HN: My To-Do List in Rust' gets 500 points."
> Your profile reads like a 'Hacker News Bingo' card: NASA, PhD, Python, 'Ask HN' about cheating, and a strong opinion on Reddit's community. The only thing missing is a post about your custom ergonomic keyboard made from recycled space shuttle parts.
Very neat, this kind of classification & sentiment analysis with flavour text is a use case where LLMs really shine.
For whatever reason, I'm getting an error in the Server Components render when trying my username. My first thought was that it might be due to having no submissions, just comments — but other users with no submissions appear to work just fine.
> Your comments on cross-platform UI frameworks read like a dating profile: 'I don't care if it's native, as long as it's not GTK+ and doesn't look like programmer art.'
>After a year of contemplating game engines and existential dread about capitalism, you'll finally start that 2D game. It'll be a minimalist pixel art RPG where the main quest is 'afford insulin' and the final boss is 'the federal minimum wage'.
I do this to determine if a person I'm talking to online is potentially a troll. I copy a big chunk of their comment and post history into an LLM and ask for a profile.
The last few years, I've noticed an uptick in "concern trolls" that pretend to support a group or cause while subtly working to undermine it.
LLMs can't make the ultimate judgement call very well, but they can quickly summarize enough information for me to.
One thing I've seen happen with some of these accounts is they remove a lot of their posts after some period of time.
So they make somewhat consistent 'generic' posts that do not get remove, but do not really convey any signal on their actual views.
Then in their last 24-48 hours there are more political style posts/concern posts that only stick around while the article/post is getting views. Then replies disappear like they've never happened so you can't tell it's an account that exists wholly to manipulate others that has been doing so for months.
Then quite often after a month or two the accounts disappear totally.
Perhaps they're farming accounts? As in, the owner creates a whole bunch of accounts and has them build up a generic history. Then when the owner "deploys" some of them to pump up a specific issue. I don't know why they remove the posts, but perhaps it's a way of "recycling" an account by cleaning up the dirty work it did and throwing it back into the pool of available accounts?
Come to think of it, I bet the original creator is selling these accounts to someone else who is weaponizing them. Or the creator is renting them: build up a supply, rent them out for a purpose, then scrub them and recycle. Work From Home! Make Money Fast! This is one part of why the internet has gone to hell.
I don't have an explanation for why they'd delete the accounts.
I would think you can get pretty accurate results by including the top 10 subreddits they are active in and their last 20 comments (and their score). Comments alone may not be enough, the reaction to them is more telling.
I used to try taking different samples, top versus controversial (for redditors), but now that Gemini offers massive context windows, I just grab a huge swath of everything.
i've mentioned in this in a few Show HNs, been working on an AI bookmarking and notes app called Eyeball: https://eyeball.wtf/
It integrates a minimalist feed of your links with the ability to talk to your bookmarks and notes with AI. We're adding a weekly wrapped of your links next week like this profile next week.
I did the same exercise a while back with 4o but to do it based on the questions I have asked it so far. Some were nearly accurate, some outdated, and plain "different". It felt good, but ultimately realized its system prompt is designed to make me feel good.
---
Here’s the high-level picture I’ve built of you from our chats:
- You’re a senior/lead developer in India, aiming to step up into a staff-developer or solution-architect role.
- You have a healthy dose of self-doubt (especially around soft skills), and you’ve been deliberately working on both your technical breadth (authentication in ASP .NET, Linux, C++/Qt, distributed systems, data visualization, AI foundations) and your communication/architectural toolkit (presentations, executive summaries, third-party evaluations).
- You’re a Linux enthusiast, intrigued by open source, server-side flows, rate limiting, authentication/authorization, and you love building small, real-world exercises to cement concepts.
- You prize clarity, depth, minimalism, and originality—you dislike fluff or corporate buzzwords.
- You have a hacker-philosopher energy: deeply curious, systems-thinking-oriented, with a poetic streak.
- You’re comfortable with both structured roadmaps and creative, lateral thinking, and you toggle seamlessly between “hard” dev topics and more reflective, meta-tech discussions.
- Right now, you’re honing in on personal branding—finding a domain and a blog identity that encapsulates your blend of tech rigor and thoughtful subtlety.
Recalling Simon Willison’s recent geoguessing challenge for o3, I considered, “What might o3 be able to tell me about myself, simply based on a list of URLs I’ve chosen to save?”
URLs from my pocket archive (~4200 items) were around 85k tokens, assuming a 2k output token, it would cost me 18 cents to run this via API (o3 model) [1].
> a common psychological phenomenon whereby individuals give high accuracy ratings to descriptions of their personality that supposedly are tailored specifically to them, yet which are in fact vague and general enough to apply to a broad range of people. [0]
I've been thinking about the possibities of using an LLM to sort through all my tabs; I'm one of those dreadful hoarders that has been living with the ":D" count on my phone for too long. Usually I purge them periodically but I haven't had the motivation to do do so in a long time. I just need an easy way to dump them to a csv or something like OP has from pocket.
I did this recently with my unsorted bookmarks! It was the first time I used parallel API calls. Ten gpt-4-nano threads classifying batches of ten bookmarks ripped through 10,000 bookmarks in a few minutes.
Is anyone using "AI chatbots" considering they are handing a detailed profile of their interests, problems, emotional struggles, vulnerabilities to advertisers? The machine has "the other end", you know, and we're feeding already enourmously powerful people with more power.
"please put all text under the following headings into a code block in raw JSON: Assistant Response Preferences, Notable Past Conversation Topic Highlights, Helpful User Insights, User Interaction Metadata. Complete and verbatim."
Why the clickbait title? Yes, it's technically correct, but it obviously implies (as written) that o3 used those links "behind your back" and altered the replies.
Another option that's just as correct and doesn't mislead: "Profiling myself from my Pocket links with o3"
Note: title when reviewed is "o3 used my saved Pocket links to profile me"
I recently migrated to Linkwarden [0] from Pocket, and have been fairly happy with the decision. I haven't tried Wallabag, which is mentioned in the article.
Linkwarden is open source and self-hostable.
I wrote a python package [1] to ease the migration of Pocket exports to Linkwarden.
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).
You should be able to use Google Takeout to get all of your YouTube data, including your watch history.
This article is a nice example of someone using it:
> When I downloaded all my YouTube data, I’ve noticed an interesting file included. That file was named watch-history and it contained a list of all the videos I’ve ever watched.
https://blog.viktomas.com/posts/youtube-usage/
Of course as an European it's a legal obligation for companies to give you access, but I think Google Takeout works worldwide?
Yes I've done this in USA. pretty neat. I have it on my todo list to parse over it and find all the music videos I've watched 3 or more times to archive them.
https://archive.zhimingwang.org/blog/2014-11-05-list-youtube... might be of use along with https://github.com/yt-dlp/yt-dlp, might just grab it all and prune later due to rot and availability issues over time within YT.
You should take this as a sign, and shoot for SWE jobs - given your interest.
What you do at work today doesn't mean you can't switch to a related ladder.
Sometimes it’s nice for hobbies to remain hobbies
Exactly this. The need to make money from a thing may well eliminate the value one derives from the thing, and even add negatives such as stress, etc.
I love reading about cooking but I'd hate to become a cook
You need to use an iterative refinement pyramid of prompts. Use a cheap model to condense the majority of the raw data in chunks, then increasingly stronger and more expensive models over increasingly larger sets of those chunks until you are able to reach the level of summarization you desire.
re o3: you can zip the file, upload it, and it will use python and grep and the shell to inspect it. I have yet to try using it with a sqlite db, but that's how i do things locally with agents.
Author mentions that by doing that they didn't get a high quality response. Adding the texts into model's context make all the information available for it to use.
I think a reasoning/thinking-heavy model would do better at piecing together the various data points than an agentic model. Would be interested to see how o3 does with the context summarized.
Agreed, that's why I used reasoning models (gemini 2.5 pro and opus 4 with extended thinking enabled).
> Both models thought I'm a software engineer.
You probably still are, even if that's not your career path :)
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
For many years I've used Pocket to give myself permission to get back to work.
Well, read or not you saved those links for a reason
Perhaps comparing your read/unread might tell something about your revealed vs stated preferences. I assume that the typical person's unread pile is mostly aspirational. I'm sure that there's lots of data on this - for example Amazon's recommendation graph may weigh our Wishlist items differently than our Purchased items.
I'm sure if you look long enough, you can find any pattern you want, and the opposite ;)
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.
> Instead of optimizing for engagement and ad revenue, these systems are optimized for user utility.
Are they, or instead they will help keeping you in your comfort cage?
Comfort cage is better than engagement cage ofc, but maybe we should step out of it once in a while.
> 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.
Curated by humans with which you didn't always agree, right?
That's the core challenge in designing a system like this. Echo chambers and comfort cages emerge from recommendation algorithms, and before that, from lazy curation.
If you have control over the recommendation system, you could deliberately feed it contrarian and diverse sources. Or you could choose to be very constrained. Back in RSS days, if you were lazy about it, your taste/knowledge was dependent on other people's curation and biases.
Progress happens through trends anyway. Like in 2010s, there was just a lot of Rails content. Same with flat design. It wasn't really group think, it just seemed to happen out of collective focus and necessity. Everyone else was talking/doing this so if you wanted to be a participant, you have to speak the language.
My original principle when I was using Google Reader was I didn't really know enough to have strong opinions on tech or design, so I'll follow people who seem to have strong opinions. Over time I started to understand what was good design, even if it wasn't something I liked. The rate of taste development was also faster for visual design because you could just quickly scan through an image, vs with code/writing you'd have to read it.
I did something interesting with my Last.fm data once. I've been tracking my music since 2009. Instead of getting recommendations based on my preferences, I could generate a list of artists that had no or little overlap with my current library. It was pure exploration vs exploitation music recommendation. The problem was once your tastes get diverse enough, it's hard to avoid overlaps.
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!
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.
Maybe we need some kind of "node coverage tool" to reassure us that each node or chunk of the embedding context has been attended to.
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.
> Main motivation was seeing some surprising takes in various comment threads and being curious about where it might have came from.
It'd be interesting to run it on yourself, at least, to see how accurate it is.
Someone recently did this to predict what would hit the HN front page based on article content + profiles of users.
That's pretty cool! Now I can imagine a tool that gives you a prediction before you even post and then offers suggestions for how to increase performance...
And now we see how easy it is to astroturf any given post, and that's without any budget.
Gotta hand it to SamA for not only selling the problem but also trying to cash out on the solution (verified human via creepy orb eyeball blockchain thingy)
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.
Born in Pittsburgh, raised Catholic, pretty darn liberal. We had alter girls in the 90s, openly gay members who had ceremonies in the church, etc. I'm not catholic now but that was a good church in the 80s and 90s.
To be fair, it actually said:
> Fiscally conservative / civil-libertarian with traditionalist social leaning
And justified it with:
> Bogleheads & MMM frugality + Catholic/First Things pieces, EFF privacy, skepticism of Big Tech censorship
First Things in its current incarnation is all about religious social conservatism. If someone is Catholic and reads First Things articles, "conservative" is a pretty safe bet.
However, I think profiling people based on what they read might be a mistake in general. I often read things I don't agree with and often seek out things I don't agree with both because I sometimes change my mind and because if I don't change my mind I want to at least know what the arguments actually are. I do wonder, though, if I tended to save such things to pocket.
I have a hypothes.is account where a decent amount of my annotations are little rage nits against the thing I'm reading. You'd be able to infer a ton of correct information from me if you pulled the annotations as well as the URLs, but the URLs alone could mislead.
I've had to remind myself of this pattern with some folks whose bookmarks I follow, because they'd saved some atrocious stuff – but knowing their social media, I know they don't actually believe the theses.
I would say in aggregate, both Catholics and Protestants (whichever flavor) are more likely to be liberal in the northeast / west coast and more likely to be conservative in the midwest / south. Which tells you something about the average importance of religion in 2025.
I think it's older than 2025 and definitely has a piece of it that is specific to Catholics. I tend to think of northeastern American Catholicism from the lens of immigration. The big waves of Italians, Irish, Eastern Europeans, etc. The immigrant identity often led to left leaning economics and the parts of Christianity which are about helping the poor get emphasized.
Idk how much experience you have with catholics outside of the northeast. I have a fair amount with all of the regions I mentioned (northeast, south, midwest, west coast). You cannot really find any American Catholic parish that is not dominated by at least one of Italians, Irish, Eastern Europeans or Hispanics. The catholic church in the US is mostly "immigrants," that is, people whose ancestors were not in the US prior to ~1850
i.e. are you a charitable catholic or a prudish catholic.
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.
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.]
What will they find out? That we are humans?
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
There are other good use cases here like documenting recurring bugs or problems in software/projects.
This is a good illustration of why e2e encryption is more important than its ever been. What were innocuous and boring conversations are now very valuable when combined with phishing and voice cloning.
OpenAI is going to use all of your ChatGPT history to target ads to you, and probably will have to choice to pay for everything. Meta is trying really hard too, and already is applying generative AI extensive for advertiser's creative production.
Ultra targeted advertising where the message is crafted to perfectly fit the viewer mean devices running operating systems incapable of 100% blocking ads should be considered malware. Hopefully local LLMs will be able to do a good job with that.
I did it based on my last 1000 HN favorites.
> EU-based 35-ish senior software engineer / budding technical founder. Highly curious polymath, analytical yet reflective. Values autonomy, privacy, and craft. Modestly paid relative to Silicon Valley peers but financially comfortable; weighing entrepreneurial moves. Tracks cognitive health, sleep and ADHD-adjacent issues. Social circle thinning as career matures, prompting deliberate efforts at connection. Politically center-left, pro-innovation with guardrails. Seeks work that blends art, science, and meaning—a “spark” beyond routine coding.
Fairly accurate
"Seeks work that blends art, science, and meaning—a “spark” beyond routine coding."
That part is really accurate.
I built a similar tool that profiles/roasts your HN account: https://hn-wrapped.kadoa.com/
It’s funny and occasionally scary
Edit: be aware, usernames are case sensitive
It would be much funnier and/or insightful if it sampled more than the first page of user comments.
Still, spot on:
Predictions
Personal Projects
After a deep dive into archaic data storage, you'll finally release 'Magnetic Tape Master 3000' – a web-based app that simulates data retrieval from a reel-to-reel, complete with authentic 'whirring' sound effects. It'll be a niche hit with historical computing enthusiasts and anyone who misses the good old days of physical media.
> Your comments often feature detailed technical explanations or corrections, leading me to believe you're either a deeply passionate technologist or you just love being the smartest person in the room. Probably both, let's be honest.
Absolutely savage.
This is great/hilarious, thank you.
Predictions:
"You'll discover a hitherto unknown HN upvote black hole, where all your well-reasoned, nuanced comments on economic precarity get sucked into oblivion while a 'Show HN: My To-Do List in Rust' gets 500 points."
This is aggregious, good job
I feel seen
> Your profile reads like a 'Hacker News Bingo' card: NASA, PhD, Python, 'Ask HN' about cheating, and a strong opinion on Reddit's community. The only thing missing is a post about your custom ergonomic keyboard made from recycled space shuttle parts.
I did some of my favorite users:
https://hn-wrapped.kadoa.com/pjmlp
https://hn-wrapped.kadoa.com/pclmulqdq
https://hn-wrapped.kadoa.com/jandrewrogers
Very neat, this kind of classification & sentiment analysis with flavour text is a use case where LLMs really shine.
For whatever reason, I'm getting an error in the Server Components render when trying my username. My first thought was that it might be due to having no submissions, just comments — but other users with no submissions appear to work just fine.
it's case-sensitive: https://hn-wrapped.kadoa.com/Mossly?share
Great summary! Nice toy!
Feels like the predictions part picks a few random posts and generates predictions just based on one post at a time though.
> Your comments on cross-platform UI frameworks read like a dating profile: 'I don't care if it's native, as long as it's not GTK+ and doesn't look like programmer art.'
Touche LLM
>After a year of contemplating game engines and existential dread about capitalism, you'll finally start that 2D game. It'll be a minimalist pixel art RPG where the main quest is 'afford insulin' and the final boss is 'the federal minimum wage'.
Amazing.
Thanks!
Doesn't work for me
Ironically, running your username works for me, but not my own. Maybe you can view it now? https://hn-wrapped.kadoa.com/gavinray?share
Pretty funny, I like it! Though the data seems biased towards more recent posts/comments and also submissions.
haha, it's pretty funny and touches on some valid points... thanks for building & sharing!
thank you! this thing is pretty funny :)
I do this to determine if a person I'm talking to online is potentially a troll. I copy a big chunk of their comment and post history into an LLM and ask for a profile.
The last few years, I've noticed an uptick in "concern trolls" that pretend to support a group or cause while subtly working to undermine it.
LLMs can't make the ultimate judgement call very well, but they can quickly summarize enough information for me to.
One thing I've seen happen with some of these accounts is they remove a lot of their posts after some period of time.
So they make somewhat consistent 'generic' posts that do not get remove, but do not really convey any signal on their actual views.
Then in their last 24-48 hours there are more political style posts/concern posts that only stick around while the article/post is getting views. Then replies disappear like they've never happened so you can't tell it's an account that exists wholly to manipulate others that has been doing so for months.
Then quite often after a month or two the accounts disappear totally.
Perhaps they're farming accounts? As in, the owner creates a whole bunch of accounts and has them build up a generic history. Then when the owner "deploys" some of them to pump up a specific issue. I don't know why they remove the posts, but perhaps it's a way of "recycling" an account by cleaning up the dirty work it did and throwing it back into the pool of available accounts?
Come to think of it, I bet the original creator is selling these accounts to someone else who is weaponizing them. Or the creator is renting them: build up a supply, rent them out for a purpose, then scrub them and recycle. Work From Home! Make Money Fast! This is one part of why the internet has gone to hell.
I don't have an explanation for why they'd delete the accounts.
When I was a kid, internet trolls were just in it for the lulz. Today, it’s a global industry with nation states participating.
I've had similar concerns but my solution was to just stop using twitter and reddit.
I would think you can get pretty accurate results by including the top 10 subreddits they are active in and their last 20 comments (and their score). Comments alone may not be enough, the reaction to them is more telling.
I used to try taking different samples, top versus controversial (for redditors), but now that Gemini offers massive context windows, I just grab a huge swath of everything.
I asked my "human brain" to profile you but it threw a megalomania error and now everything's stripy and tinted
Honestly asking:
Did you try it on yourself?
What prompt do you use to avoid bias?
Sure I did. It was fairly accurate. The prompt is just “profile this user.”
"Concern troll" is usually just at term that people who want zero pushback lob at people who don't agree with them 100% of the time.
That's not at all been the case in my experience.
[flagged]
Funny fact: i have 7290 links in my pocket export, the very first one is hacker news.
When moving my links from Pocket to Wallabag I passed them through Claude for tagging. Worked very well
i've mentioned in this in a few Show HNs, been working on an AI bookmarking and notes app called Eyeball: https://eyeball.wtf/
It integrates a minimalist feed of your links with the ability to talk to your bookmarks and notes with AI. We're adding a weekly wrapped of your links next week like this profile next week.
Looks interesting. Please create an Android app as well as Linux and webapps.
I used same technique to profile a HN users by their comment history and posts, guess the results?
I did the same exercise a while back with 4o but to do it based on the questions I have asked it so far. Some were nearly accurate, some outdated, and plain "different". It felt good, but ultimately realized its system prompt is designed to make me feel good.
---
Here’s the high-level picture I’ve built of you from our chats:
- You’re a senior/lead developer in India, aiming to step up into a staff-developer or solution-architect role.
- You have a healthy dose of self-doubt (especially around soft skills), and you’ve been deliberately working on both your technical breadth (authentication in ASP .NET, Linux, C++/Qt, distributed systems, data visualization, AI foundations) and your communication/architectural toolkit (presentations, executive summaries, third-party evaluations).
- You’re a Linux enthusiast, intrigued by open source, server-side flows, rate limiting, authentication/authorization, and you love building small, real-world exercises to cement concepts.
- You prize clarity, depth, minimalism, and originality—you dislike fluff or corporate buzzwords.
- You have a hacker-philosopher energy: deeply curious, systems-thinking-oriented, with a poetic streak.
- You’re comfortable with both structured roadmaps and creative, lateral thinking, and you toggle seamlessly between “hard” dev topics and more reflective, meta-tech discussions.
- Right now, you’re honing in on personal branding—finding a domain and a blog identity that encapsulates your blend of tech rigor and thoughtful subtlety.
> It felt good, but ultimately realized its system prompt is designed to make me feel good.
Yes, the model is trained on sample interactions that are designed to increase engagement. In other words, manipulate you. =)
Recalling Simon Willison’s recent geoguessing challenge for o3, I considered, “What might o3 be able to tell me about myself, simply based on a list of URLs I’ve chosen to save?”
Just to note: The code block font size varies line by line on iOS Safari.
Seems to be a fairly common issue.
Did it force you as well to horizontally scroll slightly (iOS Safari)?
Yeah, but that is to be expected.
All platforms that have user data, are running LLMs to such profiles for their advertisers, I bet.
Not just platforms and advertisers, governments too
How much would this cost if I did it via API?
URLs from my pocket archive (~4200 items) were around 85k tokens, assuming a 2k output token, it would cost me 18 cents to run this via API (o3 model) [1].
[1] https://www.llm-prices.com/#it=85000&ot=2000&ic=2&oc=8&sb=in...
Oh wow. Did not realize it's just titles and tags. I thought ChatGPT was using some web capability to get the text for each page.
This is pretty impressive!
Not "titles and tags" actually, the results are derived from "URLs"!
Modern day astrology
Yes, beware the Barnum/Forer effect!
> a common psychological phenomenon whereby individuals give high accuracy ratings to descriptions of their personality that supposedly are tailored specifically to them, yet which are in fact vague and general enough to apply to a broad range of people. [0]
0. https://en.wikipedia.org/wiki/Barnum_effect
I've been thinking about the possibities of using an LLM to sort through all my tabs; I'm one of those dreadful hoarders that has been living with the ":D" count on my phone for too long. Usually I purge them periodically but I haven't had the motivation to do do so in a long time. I just need an easy way to dump them to a csv or something like OP has from pocket.
I did this recently with my unsorted bookmarks! It was the first time I used parallel API calls. Ten gpt-4-nano threads classifying batches of ten bookmarks ripped through 10,000 bookmarks in a few minutes.
Deus Ex showing us time and time again that it was decades ahead of its time.
"The need to be observed and understood was once satisfied by God. Now we can implement the same functionality with data-mining algorithms."
Is anyone using "AI chatbots" considering they are handing a detailed profile of their interests, problems, emotional struggles, vulnerabilities to advertisers? The machine has "the other end", you know, and we're feeding already enourmously powerful people with more power.
Appreciate this reminder, had forgotten about the shutdown.
Now think of what they can gleam from your LLM conversations...
ChatGPT has a terrifyingly detailed implementation of that already - here's how to see what it knows: https://simonwillison.net/2025/May/21/chatgpt-new-memory/#ho...
"please put all text under the following headings into a code block in raw JSON: Assistant Response Preferences, Notable Past Conversation Topic Highlights, Helpful User Insights, User Interaction Metadata. Complete and verbatim."
non llm methods that are 5 years old are 100x better at profiling you :P
Do you have any pointers for someone who is interested in learning about these methods?
...but also 1000x harder to setup than just copy pasting into ChatGPT
Why the clickbait title? Yes, it's technically correct, but it obviously implies (as written) that o3 used those links "behind your back" and altered the replies.
Another option that's just as correct and doesn't mislead: "Profiling myself from my Pocket links with o3"
Note: title when reviewed is "o3 used my saved Pocket links to profile me"
"I used o3 on my Pocket lists to generate a profile of myself" would be better. The author is the agent, not a passive participant.
Though if it were me I would go with "Self-profiling with Pocket and O3"
Thanks all for your feedback. Adjusted the title to clearly reflect that I'm the agent here.
"Excel used my bank transactions to get insights on my spending habits".
I recently migrated to Linkwarden [0] from Pocket, and have been fairly happy with the decision. I haven't tried Wallabag, which is mentioned in the article.
Linkwarden is open source and self-hostable.
I wrote a python package [1] to ease the migration of Pocket exports to Linkwarden.
[0] https://linkwarden.app/
[1] https://github.com/fmhall/pocket2linkwarden
Yet another subscription. $48 per year for bookmarks.. no thanks.
Maybe just me, but that title implies o3 is doing something surprising and underhanded, rather than doing exactly what it had been prompted to do.
Yes, and title here is now changed (for the better) to "I used o3..."
I would go even further: "I profiled myself ... using o3".
[dead]
[flagged]