I just finished a 3-day kindergarten camp as a facilitator. The kids were great. But honestly? That's not the part I keep thinking about. It was the conversations with the other facilitators. Ages from 20s to 70s. Different backgrounds, different life chapters, zero reason to be in the same room together — except we were.

And somewhere between camp sessions, people just... talked.

Career pivots. Family decisions. Seventy years of living next to twenty years of becoming. No agenda. No audience. Just honest exchanges between people who had no reason to perform for each other.

It reminded me how much I've been underweighting this.

I spend a lot of time building in the digital space. Shipping things. Writing things. Optimising things. That work matters and I'm not stopping.

But there's a kind of knowledge that doesn't live in a dataset or a feed. It lives in people. And you only access it by showing up somewhere inconvenient and actually listening.

The richest conversations I've had recently haven't happened online. I want more of that. More situations where I'm not the builder or the expert — just a person sitting with other people, genuinely curious about their lives.

We really don't know what someone has been through until we ask.

When was the last time you had a conversation that genuinely surprised you?

#HumanConnection #RealConversations #PeopleFirst #Singapore

What would a truly localized Singaporean LLM look like? Beyond just Singlish, I explore how a 'fellow citizen' AI could navigate our unique cultural nuances.

---

As I'm going through the free Stanford Online course CME295, I thought about Natural Language Processing (NLP) and how we currently tokenize English text to feed into LLMs. If we transform words into numbers, how do we add in culture nuances, slangs, etc that are unique to a country so that we can create a sort of naturalized citizen? What is Singapore currently doing?

So I asked Gemini for some thoughts: so if I am a citizen of a country who wants to help build a realistic LLM for my country where we have 4 languages and we want to not only create a balanced LLM, but also to incorporate our country's culture and slangs into it so that the LLM performs like a "fellow citizen" and knows all our nuances and history. How should I start?

Gemini's answer: “This is a noble and technically fascinating mission. Building a “Sovereign LLM” is the move from a model that knows about your country to a model that is of your country.

To make an AI feel like a “fellow citizen,” you have to move beyond just scraping the web. You need to follow a four-stage blueprint:

• Collection: Gathering multicultural datasets.
• Curation: Cleaning data to avoid “Data Poisoning.”
• Alignment: Teaching the AI nuances and cultural slangs.
• Evaluation: Testing if the AI actually sounds like a “Fellow Citizen.”

...

Be careful with “Data Poisoning.” If you only scrape one political side or one ethnic group's slang, your “Fellow Citizen” AI might end up being biased or exclusionary. Balance is your hardest job.”

It then went on to suggest some open-source models to start with and actionable next steps.

[NOTE] I'm using this blog article to document the start of my journey. It will be a multi-year quest of mine. I know Singapore is already embarking on building her own LLM or LLMs, but wouldn't it be great if I can create my very own Singaporean friend, a super knowledgeable, kind and caring one at that.

Cool idea?

#ai #LLM #NLP #SovereignAI #Singapore #Singlish #AISingapore #CME295 #LearningInPublic