91–100 of 217 posts
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...
This is pretty impressive!
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 Some were nearly accurate, some outdated, and plain "different".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 solut…
Yes, the model is trained on sample interactions that are designed to increase engagement. In other words, manipulate you. =)
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,…
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.
Earlier quoted context omitted.
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.
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
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…
Earlier quoted context omitted.
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!
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,…
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.
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 m…
> 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…
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.
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,…
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.