judgment on S Anand

Scepticism and Humility

Wed, 18 Feb 2026 22:52:35 +0800

High Scepticism + High Humility = Scientist. Editor. Indecisive. “Let’s test it.”

Good for high-stakes, irreversible decisions. System 2 thinking is slow and effortful. But if you do this too often or too long, you miss the window or other opportunities.

High Scepticism + Low Humility = Critic. Troll. Red Hat. “You are wrong.”

Good for stress-testing and auditing. To prevent/anticipate failures. But it’s toxic and demoralizing if you do it too much.

Low Scepticism + High Humility = Intern. Follower. Cultist. “Tell me what to do.”

Good for rapid learning and emergencies. To learn patterns from a master or obey the firefighter’s orders. But without scepticism, you’re manipulable and a vector for bad ideas.

Low Scepticism + Low Humility = Bureaucrat. Visionary. Fanatic. “I know it.”

Good for action under uncertainty, and sales. Drives a bias for action. Irrational confidence helps transfer beliefs. But fanatics / bureaucrats can accelerate / decelerate dangerously.

We are each of these at different times.

Each has its place.

How to develop taste

Sun, 15 Feb 2026 14:12:28 +0800

Developing taste & judgement are an essential skill in the AI era. # #

But taste is different from knowledge and takes more time.

Gaining knowledge is a software upgrade. It strengthens existing synapses. It’s fast, reversible, no new “cables” required.
Taste is a hardware upgrade. It destroys inefficient pathways, grows neurons for new pathways, and wraps axons with myelin speeding up signals 100x. (London cab drivers literally have a larger hippocampus.) Taste takes time.

How we acquire taste depends on the environment.

In kind environments (with clear, immediate, accurate feedback, like sports, surgery) is easier. Practice at the edge of competence. #
In wicked environments (investing, hiring, politics, strategy) confidence can be misleading. So, audit prediction reasoning: write your predictions with reason. Months later, if you were right for the wrong reasons, treat it as failure.

Practices that help:

Taste requires complexity and understanding #. So, when bored, complicate; when confused, study by copying, comparing, asking why, and prototyping a vocabulary. # #
Argument Mapping: create granular mind-maps of arguments, find hidden assumptions, and evaluate evidence.
Watch experts: watch experts at work, guess their next moves, explain your reasoning, and copy but with extra constraints.
Perceptual Learning: learning by comparing examples and prototyping a vocabulary.
Brain Trust: have peers critique your work against a goal at early stages (to focus on core, not polish).
Mindfulness: to reduce sunk-costs and other biases/blindspots. Helps with post-mortems.
Communities of Practice: find where experts in your field hang out, curate ruthlessly, and absorb the vocabulary.
Project-Based Learning: solve a real problem you have, fail + learn + iterate, with other people.

Ancient wisdom (Stoicism, Buddhism, Confucianism, etc.) broadly aligns, but there are a few differences.

Slow vs fast. Ancient wisdom suggests that judgement must develop slowly. Science is optimistic about acceleration, e.g. perceptual learning, simulation, etc.
Moralility. Ancient wisdom anchors judgement in morality. Science is more agnostic.

Sometimes “taste” is just “elitism”. That’s not going away, and offers another way to develop taste: via “club membership”.

AI, like most automation, erodes skills. This has happened in the past and we deal with it differently.

Autopilots eroded flying skills - which is dangerous. So we enforce flight simulators. Same for surgical knots (robotic surgery), celestial navigation (navy), manual dosing (nurses).
Spreadsheets eroded calculation skills. We leveled-up from sums to strategy. Same for CAD, electronic trading, spell-check.
Photography eroded painting skills. We switched value to impressionism, cubism, etc. Same for vinyl records, luxury watches, craft coffee.
GPS eroded navigation skills. We accepted this and don’t care much. Same for phone numbers, spelling, mental maths.

Think about how the skill we lose will evolve. Then enforce, level-up, switch, or accept accordingly.

Source: I used Claude Deep Research and asked Gemini to interpret it.

23 Mar 2026. Taste may not matter as much as I thought. I see AI learning & acquiring good taste in code (e.g. architecture) and art (e.g. writing, visual design). Accountability may be more important.

Read from LLMs but don't tell people

Fri, 07 Mar 2025 05:08:18 +0000

In meetings, I pass on questions to ChatGPT and I read out the response. But I’ve stopped saying “I’m reading that from ChatGPT.”

(By “ChatGPT”, I mean ChatGPT, Claude, Grok, Gemini, Meta, etc. I happen to use ChatGPT with O3 Mini + Search.)

Use ChatGPT in meetings

It’s good to bring ChatGPT into conversations. (Or any activity where intelligence helps, actually.)

In meetings (online or in person), I keep a ChatGPT window open. When asked:

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

If I know the answer, I’ll give it. If not, I ask ChatGPT. (Ideally, I should ask even if I think I know the answer.)

For example:

Sharing the menu photo and ask, “List vegetarian options. Suggest innovative dishes I’ll like.” (This works because I’ve shared my preferences and history with ChatGPT.)
“How can doctors use LLMs in day-to-day work?”
Sharing a picture of the architecture, “Explain this architecture to a blind expert. Critique with strengths, critical issues, and optional improvements.”

I’ve learnt that:

Note-taking helps. I touch-type (without looking). I copy-paste the notes and their question to the LLM.
Short questions are fine. Newer models understand cryptic questions.
Say “Give me 30 seconds”. People assume you’re thinking deeply.

Read the response your way

I just read out the response – but with some changes.

Change style. I read quicky, internalize, and say it in my style. Instead of “1. Clinical Documentation & Administrative Support”, I’d say, “Doctors can use it for note-taking.”
Filter content. I skip stuff I don’t get or like. I might miss stuff, but when I speak, it’s my opinion I represent.
Add context. I add personal stories to make it real, if I can. “GPs I know are worried LLMs diagnose better than they do” is something LLMs may not have learnt yet.

Don’t say you’re reading from ChatGPT

I used to tell people, “… and I just read that out from ChatGPT.” Their response is always:

Disbelief for a moment.
Amazement that models are so good.
Dismissal of what I said, since it’s not “real”. (This is the sad part.)

I stopped saying that because

I don’t need to. I can promote LLMs elsewhere.
It’s not true. I re-style, filter, add context. It’s my response. My responsibility.

I’d rather deliver useful ideas than show where they come from. And if they think I’m a genius? Fine by me 🙂

Sun, 24 Mar 2024 14:39:01 +0000

Oh, wonderful! They’re keen to get in. Wise enough to take help. Honest enough not to be able to cover it up. Sounds like a good hire!