personal-analytics on S Anand

Fri, 15 Aug 2025 00:00:00 +0000

Indian Celebrities and Directors was my top searched category on Google while OpenAI & AI Research was the top growing category.

This is based on my 37,600 searches on Google since Jan 2021. Full analysis: https://sanand0.github.io/datastories/google-searches/

The analysis itself isn’t interesting (to you, at least). Rather, it’s the two tools that enabled it.

First, topic modeling. If you have all your searches exported (via Google Takeout) into a text file, you can run:

uvx topicmodel searches.txt --ntopics 50

… and automatically get the top 50 topics you search for.

Second, an improved O3 prompt. I fed it monthly topics volume and asked:

Look closely at the numbers as well as the image. What insights can you draw from these? Aim for non-obvious non-trivial insights. Run correlations or any other analyses on the data to go deeper and come up with material suitable for a deep research paper.

The analyses it did was far more powerful than anything I would have thought of.

It calculated my Herfindahl-Hirschman index slope and declared that my search interests are diversifying. (Good to know!)
Using Principal Component Analysis it discovered 3 personas in my searches
- Classical developer (Python, JS)
- AI-builder (OpenAI, LLMs, APIs)
- India/Singapore geo-culturist (Celebs, local info, Tamil cinema) I should segment sharing & learning along these axes (e.g. separate newsletters or dashboards.)
Using the Coefficient of Variation it found the biggest spikes in SQL & Databases, and Testing & Code Tools. Steadiest were Currency Conversion Rates and Singapore/Bangalore Local Info.

I should cache common reference lookups locally and allocate “deep‑focus blocks” for debugging spikes.

There’s more. But this is the first time I felt completely outmatched by an LLM. I’m an expert on analysis. I’m an expert on this domain (my search queries.) Yet, this is far more insightful than I ever would have analyzed!

Thu, 31 Jul 2025 00:00:00 +0000

Here’s a comic book analyzing my Google Search History.

It’s a simpler version of my earlier post. I created it using PicBook, a tool I vibe-coded over ~5 hours.

PicBook: https://tools.s-anand.net/picbook/
Code: https://github.com/sanand0/tools/tree/main/picbook
Codex chat: https://chatgpt.com/s/cd_6886699abfb08191acf036f6185781be

The code prompt begins with Implement a /picbook tool to create a sequence of visually consistent images from multiline captions using the gpt-image-𝟭 OpenAI model and continues for 6 chats totaling ~22 min. My review took 4.5 hours. Clearly I need to optimize reviews.

Once generated, I drafted the storyline:

I analyzed 4 years of my Google search history. [Draw: Night study room. Protagonist unrolls a dusty scroll from a trunk labeled “BACKUPS”. Ghostly numbers and topic names swirl out. Candlelight + laptop glow mix.]
It’s mostly tech. That was no surprise. [Draw: Protagonist bored, leaning on a pile of hefty tomes: “JS DOM (1613)”, “Python Tools (1402)”. He’s sipping chai, half-asleep. A speech bubble with “meh”.]
…

Full storyline: https://sanand0.github.io/datastories/google-searches/#comic-story

… and about 11 minutes later, the story was generated. I just needed to print as PDF.

Now that it’s easier, more people might create comics. (But perhaps not on LinkedIn. Business networkers seem uncomfortable around comics.)

Download the PDF.

Top 8 ways I use ChatGPT in 2025

Sat, 24 May 2025 03:31:44 +0000

I extracted the titles of the ~1,600 conversations I had with ChatGPT in 2025 so far and classified it against the list of How People Are Really Using Gen AI in 2025.

Here are the top 8 things I use it for, along with representative chat titles. (The % match in brackets tells you how similar the chat title is to the use case.)

Improving code (clearly, I code a lot)
Troubleshooting (usually code)
Corporate LLM/Copilot (this is mostly LLM research I do)
Generating code (more code)
Generating ideas (yeah, I’ve stopped thinking)
Simple explainers (slightly surprising how often I ask for simple explanations)
Generating relevant images. (Surprising, but I think I generated a lot of images for blog/LinkedIn posts)
Specific search (actually, this is mis-classified. This is where I’m searching for search engines!)

My classification has errors. For example, “Reduce Code Size” was classified against “Generating code” but should have been “Improving code”. But it’s not too far off.

Here’s a list of representative chats against these use cases.

Improving code (263):
- PR Code Review Suggestions (64% match)
- Assessor Code Review and Improvement (63% match)
- Reduce Code Size (62% match)
Troubleshooting (172):
- Connector Error Troubleshooting (67% match)
- DNS Resolution Debugging Steps (55% match)
- Exception Handling Basics (47% match)
Corporate LLM/Copilot (141):
- LLM Integration in Work (57% match)
- LLM Agents Discussion (56% match)
- LLMs Learnings Summary (56% match)
Generating code (113):
- AI Code Generation Panel (58% match)
- AI for Code Generation (58% match)
- Reduce Code Size (54% match)
Generating ideas (99):
- Filtering Ideas for Success (54% match)
- AI Demo Ideas (52% match)
- Hypothesis Generator Name Ideas (52% match)
Simple explainers (94):
- Simple Public APIs (43% match)
- Y-Combinator Explained Simply (41% match)
- Prompt Engineering Tutorial Summary (39% match)
Generating relevant images (93):
- Popular AI Image Tools (54% match)
- Diverse Image Embedding Selection (52% match)
- AI ImageGen Expansion Ideas (52% match)
Specific search (69):
- Semantic Search Engines Local (59% match)
- Enterprise Search Solution (54% match)
- Local LLM Semantic Search (53% match)

How did I calculate this?

On ChatGPT.com, I scrolled until I had all 2025 chats visible. Then I pasted copy($$(".group.__menu-item").map(d => d.textContent)) to get the chat titles.
On Claude.ai, I transcribed this list of use cases from HBR (prompt: “Transcribe this image”).
On LLM Foundry (which you may not have access to), I used the Similarity API to get a CSV of similarities between prompts and top 30 use cases in 2025 using text-embedding-3-small.
On ChatGPT.com, I told it to analyze the data like this:

This sheet has the embedding similarity between my ChatGPT prompts (in column “A”) with different use cases.

Write and run code that tags the prompt with the use with the highest embedding similarity (cell value), drops prompts whose highest embedding similarity is below a cutoff, and shows a table where the rows are the use cases and the values are the frequency. Do this for multiple embedding cutoffs as columns: 0.0, 0.1, 0.2, 0.3, 0.4. So, the table has use cases in rows, embedding cutoffs in columns, and the cell values are the count of prompts tagged with each use case AND have an embedding similarity >= cutoff. Draw this as a heatmap with low numbers as white and high numbers as green.

… and then:

Let me download this as a Markdown list in this style, sorted by descending order at cutoff = 0

Anti-trolling (mention count of matches at 0 cutoff):

Tor Technical AMA questions (34%)

Bot Message Edits (33%)

Popular Hacker News Keywords (33%)

…

Here’s the full list against the top 30 use cases:

Improving code (263):
- PR Code Review Suggestions (64% match)
- Assessor Code Review and Improvement (63% match)
- Reduce Code Size (62% match)
Troubleshooting (172):
- Connector Error Troubleshooting (67% match)
- DNS Resolution Debugging Steps (55% match)
- Exception Handling Basics (47% match)
Corporate LLM/Copilot (141):
- LLM Integration in Work (57% match)
- LLM Agents Discussion (56% match)
- LLMs Learnings Summary (56% match)
Generating code (113):
- AI Code Generation Panel (58% match)
- AI for Code Generation (58% match)
- Reduce Code Size (54% match)
Generating ideas (99):
- Filtering Ideas for Success (54% match)
- AI Demo Ideas (52% match)
- Hypothesis Generator Name Ideas (52% match)
Simple explainers (94):
- Simple Public APIs (43% match)
- Y-Combinator Explained Simply (41% match)
- Prompt Engineering Tutorial Summary (39% match)
Generating relevant images (93):
- Popular AI Image Tools (54% match)
- Diverse Image Embedding Selection (52% match)
- AI ImageGen Expansion Ideas (52% match)
Specific search (69):
- Semantic Search Engines Local (59% match)
- Enterprise Search Solution (54% match)
- Local LLM Semantic Search (53% match)
Adjusting tone of email (66):
- Email transcription request (45% match)
- Summarize emails request (45% match)
- Intro Email FAQs (44% match)
Generating a legal document (59):
- LLM Generated SVG Ideas (48% match)
- LLMs for DSL Generation (45% match)
- Deterministic Random Content Generation (45% match)
Preparing for interviews (43):
- LLM Coding Interview Tools Report (43% match)
- Bank Ops Prep Resources (42% match)
- AGI Preparation (42% match)
Personalized learning (32):
- Lifelong Learning in Conversations (51% match)
- AI Classroom Engagement Names (48% match)
- LLM Learner Personas Roadmap (47% match)
Explaining legalese (32):
- LLM Coding Insights (46% match)
- LLM Code Ownership (45% match)
- LLM Data Format Comparison (44% match)
Creating a travel itinerary (28):
- Travel Strength Training Tips (39% match)
- User Journey Tools Online (37% match)
- Prioritize My Explorations (36% match)
Creativity (28):
- Creative Process Breakdown (55% match)
- Creative Hallucinations in Innovation (50% match)
- Leveraging Serendipity for Innovation (50% match)
Cooking with what you have (26):
- Vegetarian Dish Creation (45% match)
- Baked Veggie Dishes (41% match)
- Vegetarian dish idea (40% match)
Organizing my life (24):
- Prioritize My Explorations (49% match)
- Workspace Suggestions for Browsing (39% match)
- Editing for Clarity and Simplicity (39% match)
Enhanced learning (23):
- 2025 LLM Embedding Enrichment (51% match)
- Lifelong Learning in Conversations (49% match)
- Tech-Enhanced Teacher-Student Rapport (49% match)
Finding purpose (21):
- Prioritize My Explorations (40% match)
- Deep Research Use Cases (37% match)
- Filtering Ideas for Success (36% match)
Deep and meaningful conversations (20):
- Lifelong Learning in Conversations (49% match)
- Humorous conversation summary (42% match)
- New chat (40% match)
Healthier living (18):
- Modeling Quality of Life (40% match)
- Lifelong Learning in Conversations (37% match)
- Posture and Breathing After Weight Loss (36% match)
Anti-trolling (18):
- Tor Technical AMA Questions (34% match)
- Bot Message Edits (33% match)
- Popular Hacker News Keywords (33% match)
Writing student essays (18):
- Scholarship Answer Advice (47% match)
- Student Q&A on LLMs (41% match)
- Reward Systems for Students (41% match)
Fun and nonsense (17):
- Humorous conversation summary (45% match)
- Funny Llama 3.3 Strips (40% match)
- Synonyms for Interestingness (40% match)
Boosting confidence (14):
- Emotional Prompting Impact (41% match)
- Emotional Prompting Impact (41% match)
- AI Ratings of My Flaws (38% match)
Personalized kid’s story (14):
- Fake Data Storytelling Tips (43% match)
- Low Effort Storytelling Training (39% match)
- Demo Name Suggestions (37% match)
Reconciling personal disputes (12):
- Divorce AI Podcast Ideas (37% match)
- Summarizing Personal Journals LLM (36% match)
- Hobby Suggestions and Devil’s Advocacy (36% match)
Entertaining kids (9):
- Comedy for Geriatric Doctors (40% match)
- Humorous conversation summary (36% match)
- Indoor Activities in SG (34% match)
Medical advice (9):
- Patient Doctor Communication Tips (41% match)
- AI Training for Doctors (40% match)
- AI Training Course for Doctors (38% match)
Therapy/companionship (3):
- DBT Course (36% match)
- Cupping Therapy Evidence (36% match)
- Empathy App Development Ideas (33% match)