The AI Pragmatist

Overview

Most of the AI discourse I read these days falls into one of two camps: existential dread or breathless evangelism. AI will displace all our jobs, or AI will transform our productivity overnight. Rather than subscribing to either, I prefer to simply dive deep into using these tools and see how best I can extract value from them. If I had to describe my stance, I would call myself an AI pragmatist: optimistic about the tools, honest about the gaps, focused on using them well.

So as an AI pragmatist, how am I actually using AI? I have covered some examples in my other posts, so I thought I would use this piece to focus on a few high-value use cases where AI works best with a human steering it, as well as some observations about where human involvement remains critical.

This brings to mind a concept introduced by Andrej Karpathy, who described AI as having “jagged intelligence”: a model can refactor a massive codebase, but ask it “I want to wash my car, the car wash is 50 metres away, should I walk or drive?” and it will confidently tell you to walk, missing the obvious point that the car needs to be at the car wash. This question has become a popular litmus test for assessing each new model’s reasoning ability, though newer models may have gotten better at it simply because the question has since made its way into the training data. Ethan Mollick wrote about a similar observation, calling it the “jagged frontier”: AI is excellent at certain tasks and poor at others, and the boundary between the two is irregular and unintuitive. The peaks are high but the valleys are also right next to them.

This post captures where I think that frontier sits today, based on my own use cases. I suspect some of this will read differently in six months.

What AI Is Great At

None of these examples involve AI operating autonomously. In each case, the value comes from a human working with AI: directing it, reviewing its output, and deciding what to do with the result. Ethan Mollick calls this co-intelligence, and it is the mode I have found most productive.

Excellent first drafts. For less complex matters, using NickOS with my voice, writing style and contextual knowledge already loaded, I can generate a very good first draft: a polished update for a stakeholder, a sense-check on the regulatory impact of an upcoming law and whether it affects the games industry, a structured comparison of approaches across jurisdictions. It gets me to 70 or 80 percent most of the time, sometimes more, sometimes less. Not perfect, but it eliminates the blank-page problem and gives me something to react to rather than build from scratch. That generates real value in terms of time savings.

Multi-modal outputs. Many of my guidance materials used to live in Word or PDF files. AI is excellent at generating .html deliverables with charts, toggle buttons to switch between English and Chinese (incredibly important in a Chinese technology company offering products and services outside China), .svg visuals, and interactive elements like side panels triggered by an info icon containing additional guidance material. The point is that information now comes in more accessible formats rather than plain text, so stakeholders can process and understand legal guidance more easily.

Vibe-coding and scalable frameworks. I am not a software engineer, but AI has made it possible to build functional tools and reusable frameworks that would previously have required one. For a recent compliance project, I needed to collect self-assessment inputs from multiple product teams across dozens of parameters, involving three languages, with guidance notes and supporting evidence for each answer. Instead of circulating a spreadsheet and chasing responses, I built an interactive HTML form: a toggle between English and Chinese, expandable guidance panels for each parameter, JSON export for the collected inputs, and a pipeline to convert those into a CSV format in Bahasa Indonesia ready for submission to the regulator. The form is now a repeatable framework: if the regulation changes or a new product needs to be assessed, the structure is already there. A year ago, this would not even have been conceivable.

Where Humans Still Fill the Gap

Institutional knowledge and token economics. Tencent has over 100,000 employees. Dozens of departments, teams within teams, stakeholders whose titles do not always reflect what they actually do. Knowing who to ask for a particular type of information, how to operate certain internal tools and systems, understanding each team’s real incentives, or filtering what people say through their vested interests: this is the kind of knowledge that compounds over years and resists being uploaded. A model can summarise text. It cannot interpret institutional subtext. And even if it could, the token cost of feeding a full organisational map into every interaction would not be economic.

Expertise, verification and the last mile. When a model drafts a regulatory analysis, the output is often structurally sound. Where it goes wrong is predictable to someone with domain experience:

• A risk framed too conservatively because the model lacks context on commercial appetite.
• A recommendation that overlooks an internal precedent.
• A conclusion that technically follows the text of a statute but misses how the regulator has applied it in practice.

The model does the heavy lifting of the first draft. The person with experience brings the judgement: knowing whether it is any good, whether it is applicable, and polishing it to the requisite standard.

Being an AI Pragmatist

To conclude, I would say that being an AI pragmatist means using the tools where they are good, compensating where they are not, and continuing to exchange ideas with like-minded people about how best to use them. That exploration can be personal: asking a model to read an insurance policy so the exclusions are finally intelligible, drafting a budget, getting meal ideas based on what is actually in the fridge, or having something nudge me to get to the gym. At work, it might mean automating the repetitive parts of a compliance review so the hours saved go back to the team. None of this requires a grand theory of where AI will be in five years.

It also reminded me of a question I was asked during one of my rounds of interviews at Tencent, about five years ago. The interviewer asked where I thought Singapore should position itself in the US-China geopolitical landscape. I referenced Lee Kuan Yew’s approach: Singapore has generally sought to maintain relationships with both sides, engage constructively where it can, and make pragmatic decisions based on its own interests and the reality on the ground. It does not ignore the tensions, but it does not define itself by them either. I think a similar lens applies to AI. The tools will keep improving. The pragmatist keeps using them, keeps learning what works, and keeps sharpening the skills that make the difference when the tools fall short.

Back to AI Journal