Arindam Roy suggested a business idea post-COVID: an online QR code for business cards.

The benefits are clear. Never print or carry cards. The scanned card goes straight to the other person's contacts. There might be social metrics we could capture as well.

That idea never got past the discussion stage, though I've begged for it a few times. Last year, I decided to use existing tools to solve the problem. My current approach: generate a 𝘃𝗖𝗮𝗿𝗱 𝗤𝗥 𝗰𝗼𝗱𝗲.

vCard is the format contacts are stored. Most phones, including iOS and Android, support it. It holds basic information like name, mobile, email, company, website, etc.

Sites like https://lnkd.in/gsF5-MNJ let you enter the details and give you a QR code image. I added this to my phone desktop as a photo widget.

Now, when I meet someone, I open my phone, show the QR code, they scan it, and my details are added as a contact. No paper. No typing. Instant updates 😎

Should we self-host an LLM instead of using an API?

𝗔𝗻𝘀𝘄𝗲𝗿: Almost never.

Take Llama 3.3 70b Instruct. To generate ~1M tokens (about as much as an A80 GPU can in a day), it would cost:

$0.12 on DeepInfra via API
$0.71 on Azure AI Foundry via API
$43 on Lambda Labs
$88 on Azure servers

That can explore a $1,000 budget to $700,000!

The 𝗼𝗻𝗹𝘆 reasons I see for self-hosting are (a) ultra-high security or (b) you want to learn how.

Blog Post: https://lnkd.in/gWanGzNq
Deep Research: https://lnkd.in/g7a33sgZ

It's so easy to fool yourself. I thought I knew my preferences until an experiment with LLM-generated slogans shattered my confidence.

I asked four LLMs (GPT 4.5, Claude 3.7, Gemini 2.5, Grok 3) to give me catchy rhyming slogans promoting journaling. At first glance, I ranked them easily:

Grok was best, Claude a close second, Gemini third, and GPT 4.5 clearly last.

But when I built a quick app to blindly compare individual slogans, my "clear" preferences flipped completely. Multiple times!

Turns out:

1. I 𝗰𝗮𝗻'𝘁 𝗷𝘂𝗱𝗴𝗲 stuff at a glance.
2. But I 𝘁𝗵𝗶𝗻𝗸 I can.
3. Even when evaluating carefully, my 𝗽𝗿𝗲𝗳𝗲𝗿𝗲𝗻𝗰𝗲𝘀 𝗰𝗵𝗮𝗻𝗴𝗲.

This shook me more than anything else recently. I badly need 𝘄𝗿𝗶𝘁𝘁𝗲𝗻 evals to even know what I want.

See the quotes and app at https://lnkd.in/gpdEzjG6

Could AI invent languages that make us smarter and more creative?

I asked Gemini 2.5 Pro to invent an alien language and think in it. It invented 𝗙𝗹𝘂𝘅𝘁𝗼𝗻𝗴𝘂𝗲, spoken by beings who see the world as fluid, interconnected processes. No nouns, only verbs. Here's the alphabet.

🔀 to flow, to exist, to be in process.
🔼 to intensify, focus, emerge, become distinct.
🔽 to dissipate, disperse, fade, return to potential.
🔁 to interact, resonate with, exchange influence.
⏸️ to persist, stabilize temporarily, resist change.
⏺️ boundary condition, locus of interaction, temporary self-reference point.

For example:

Solids are 🔀🔼 - a flow that stands out.
Home is ⏺️🔀⏸️🔁🔀 - a boundary that exists stably yet interacts fluidly.

Then it started thinking in that language. Some inferences:

Truth 🔀⏸️🔁🔼 is a stable resonance rather than objective fact.
Problems 🔀🔁🔀 are discordant interactions needing tuning, not obstacles needing removal.
Death ⏺️🔀🔁 is pattern dissolution and reintegration, not cessation.

Deep stuff, but I think there's something here. Invented languages can:

- Enable 𝗹𝗮𝘁𝗲𝗿𝗮𝗹 𝘁𝗵𝗶𝗻𝗸𝗶𝗻𝗴 and creativity, unlocking ideas we (or AI) would otherwise miss.
- Overcome 𝗯𝗶𝗮𝘀𝗲𝘀, enrich therapy, or communicate concepts we previously had no words for (e.g. Saudade or Hygge)
- Maybe help us talk to other species, AI, or ETs?

I'm terrible with human languages, but invented languages are something I'm definitely exploring further.

Read the full transcript: https://lnkd.in/gw94mTuB

Subramanian Neelakantan When coding this project, I missed an opportunity to learn. I SHOULD have asked the LLM to show me a dozen options to visualize the data.

For example, "Show a version geared toward an executive

My new workflow for slide decks involves speaking to ChatGPT on a morning walk.

Here's how I made my latest slides:

𝗖𝗵𝗮𝘁𝗚𝗣𝗧 𝗜𝗻𝘁𝗲𝗿𝘃𝗶𝗲𝘄: While walking 75 minutes to my talk at IIT Madras, I had ChatGPT interview me via voice mode to draft my Markdown slides.
𝗠𝗮𝗿𝗽 𝗦𝗹𝗶𝗱𝗲𝘀: Marp is a tool to convert Markdown into polished slides.
𝗟𝗟𝗠 𝗜𝗺𝗮𝗴𝗲𝘀: ChatGPT and Gemini created custom supporting images—effortlessly.
𝗚𝗶𝘁𝗛𝘂𝗯 𝗣𝗮𝗴𝗲𝘀: Automated GitHub Actions deployed my slides online instantly.

Here are a few examples:

Visualizing LLM Hallucinations: https://lnkd.in/gAjyjM2D
Code: https://lnkd.in/gCV5AEu2

LLMs in Education: https://lnkd.in/gGyxVWwh
Code: https://lnkd.in/gAiD-2Ua

The "Wow" factor for me was that my 75-minute morning walk is now engaging 𝘢𝘯𝘥 productive. Also, that it's a 𝗹𝗼𝘁 easier to talk to a smart AI than to figure out the next slide title.

Now, I’m refining this workflow to include voice-generated diagrams and automatic publishing triggered just by saying "Publish.

Here's the full, step-by-step breakdown of exactly how this works: https://lnkd.in/gppbU-R7

LLMs changed how I think about expertise & learning.

Simon Willison's post on browser data got me thinking: "Which browser releases new features fastest?" Four hours later, I published a scrolly-telling data story using LLMs.

Link: https://lnkd.in/gDpQvFFU
Code: https://lnkd.in/g2WF6SVT

𝗔𝗻𝘀𝘄𝗲𝗿: Edge releases features fastest since 2023, beating Chrome.

But in the process, I learned a few things.

1. 𝗘𝘅𝗽𝗲𝗿𝘁𝗶𝘀𝗲 𝗺𝗮𝘁𝘁𝗲𝗿𝘀. It took experience to spot the 𝗿𝗲𝗮𝗹 story.
2. 𝗘𝘅𝗽𝗲𝗿𝘁𝗶𝘀𝗲 𝘀𝗵𝗶𝗳𝘁𝘀. To create this story, the skill I needed in the 2000s was PowerPoint. In 2010s, it was coding. Today, it's knowing the right visuals. I've no idea what it'll be next.
3. 𝗟𝗟𝗠𝘀 𝗖𝗔𝗡 𝗯𝘂𝗶𝗹𝗱 𝘆𝗼𝘂𝗿 𝗲𝘅𝗽𝗲𝗿𝘁𝗶𝘀𝗲 - but only if you practice mindfully. It's not natural.

My current expertise 𝘄𝗶𝗹𝗹 become useless. So, I'm going to practice "becoming an expert faster". The new skill of the 2030s.

It's not what you know. It's how you learn.

Blog: https://lnkd.in/gp83WV52

Stories! Where are the stories?

Did you see any overflowing trash cans? Any that the drivers could not /would not pick up? Arguments with customers? Dead bodies in the trash?

Or at least ...

Navin Kabra I bought the $200 Pro version last month. I ran 40 Deep Researches and am unlikely to go beyond 50 from my quota of 100+. Will publish them shortly.

Abhiram R DEFINITE spoilers for Words of Radiance. In fact, for the entire Cosmere!

I didn't expect I'd use LLMs in READING fiction.

Yet, it's led to the most delightful experience in 2 ways.

Firstly, my imagination of scenes is vague. I skip descriptions and dwell more on actions and dialogues. The picture I form of places and people are hazy.

But when I give ImageGen 3 a prompt from The Way of Kings:

Galloping directly toward them was a massive black horse bearing a rider in gleaming armor that seemed to radiate light. That armor was seamless—no chain underneath, just smaller plates, incredibly intricate. The figure wore an unornamented full helm, and the plate was gilded. He carried a massive sword in one hand, fully as long as a man was tall.

... I get it. I *really* get it. That drawing is worth... well... a thousand words.

I think ImageGen 3 is the best image gen model, BTW: https://lnkd.in/gagd5QFW

Secondly, I love geeking out on some of the fiction I read. (Attack on Titan, Cosmere stuff, Calvin & Hobbes, etc.)

Open AI's Deep Research is fabulous for this. I asked it "What are the most likely theories and speculations about the Cosmere by Brandon Sanderson? Collect as many of these into a comprehensive, detailed, encyclopedia of Cosmere Secrets.

What it produced has been more instructional, entertaining (and highly distracting) than most blog posts I read: https://lnkd.in/gpaWVtQN

PS: There may be more spoilers per sentence than any other Brandon Sanderson reference.

I'm at an open Hyderabad meet-up, Thu 20 Mar 4 pm. "Analyzing data with AI agents".

It's a public event by Hasgeek. Venue: Castlight Health, Sattva Knowledge Park.

We know LLMs suck at number crunching but are good with code. I'll share what we've learnt by getting it to write code to analyze data instead.

Less lecturing, more interactive Q&A and demos in a cozy group. Mostly for analysts, data scientists, and programmers. Not so much for LLM researchers or managers.

Register at https://lnkd.in/gxuKrjFS (or just gate-crash 🙂)

Generated data is usually boring. Flat distributions, no insights, no patterns.

With LLMs, I now have an alternative that works fairly well. Here's my prompt to generate fake data:

------------
Generate realistic fake data for [SOME_DOMAIN, e.g. healthcare, student marks, ...]

STEP 1. List columns that would be present in such data, briefly describing how the data might be distributed.
STEP 2. Given such data, think about an objective and generate 5 hypotheses that an organization might want to test on how to achieve this objective.
STEP 3. Write and run a Python program that generates 2,000 rows of realistic fake data where these hypotheses are true in a statistically significant way. Let me download the output as a CSV file.
STEP 4. Test each hypothesis and show the results.
------------

This works VERY well (at least on GPT 4.5). If you have specific patterns you want to inject, it's easy to include those too.

Blog post with ChatGPT transcripts: https://lnkd.in/gApdFrGD

Sowmya Vivek Oh, what an apt one! Never thought I'd hear Kadhala Kadhala quoted in a ChatGPT context 🤣

I've been quietly bringing ChatGPT into meetings. And stopped mentioning it.

When I'm asked something like

“What’ll you eat, Anand?” (at restaurants)
“How can doctors use LLMs?”
“Will you review this technical architecture?”

... I just ask ChatGPT. And I read out the answer (in my own style, with some edits.)

Weirdly, when I used to say, "I'm reading from ChatGPT," reactions were disbelief, amazement, and sadly, dismissal - as if the insights weren't genuine.

So, I no longer tell them. They get value. I sound like a genius 🙂

More on how I personalize the response at https://lnkd.in/gV9BUxRJ

For a second, I thought I'd been hacked.

If you search for "Geocities" on Google, it shows the results in a Comic Sans font.

https://lnkd.in/gNP6N5nG

Turns out it's an easter egg by Google that's about 8 months old.

Are there any Geocities homesteaders here? Or is LinkedIn too new for the old-timers?