Subhasis Roy

Building ThinkExec to help people think clearly & act on what matters. I write about better thinking, product building, consulting & entrepreneurship.

[Subhasis Roy @SubhasisRoyUK]

Most people don’t have a time problem.

They have a clarity problem.

[Subhasis Roy @SubhasisRoyUK]

Perhaps schools need to develop two complementary capabilities: independent problem-solving and AI-augmented problem-solving.

If students can instruct AI to build almost anything, they still need enough foundational understanding to recognise when the solution is wrong.

The interesting educational challenge is how to use AI to develop that understanding, not bypass it.

That could mean combining independent reasoning with AI-assisted exploration, challenge and feedback, then testing whether students can apply what they've learned without AI.

The goal shouldn't just be students who can direct AI, but students whose own judgement becomes stronger through using it.

[Subhasis Roy @SubhasisRoyUK]

Congratulations on the launch, Rob!

One interesting consequence of AI making products easier to build is that it also changes how founders can validate ideas.

We can now run experiments and create prototypes much faster. But that makes it even more important to distinguish positive reactions to a prototype from genuine willingness to pay.

The scarce capability may increasingly be designing experiments that produce reliable evidence of customer demand, rather than simply building something people find interesting.

[Subhasis Roy @SubhasisRoyUK]

Interesting prediction. I wonder whether the economics will be more nuanced than tokens versus human salaries.

Even if frontier AI becomes more expensive, smaller specialised models and improving efficiency could make many individual tasks cheaper.

The more meaningful comparison may be the total cost of delivering a reliable outcome, including orchestration, verification, supervision and rework.

That could favour different combinations of humans and AI depending on the task, rather than one universally replacing the other.

[Subhasis Roy @SubhasisRoyUK]

I agree that architectural ownership cannot simply be delegated to an AI agent.

But I'd distinguish architectural ownership from architectural reasoning.

AI can contribute meaningfully to exploring design alternatives, challenging assumptions, identifying dependencies and evaluating trade-offs against defined constraints.

The architect's role isn't necessarily to make every design decision independently. It's to establish the principles and constraints, critically evaluate the options, and remain accountable for the decisions.

That seems a more productive model of human–AI collaboration than treating architecture as exclusively human work and implementation as the agent's responsibility.

[Subhasis Roy @SubhasisRoyUK]

This raises an important question about how organisations measure engineering productivity in an AI-assisted environment.

If code generation becomes dramatically cheaper, measuring output volume becomes even less meaningful.

Perhaps the better measure is how effectively teams deliver useful capabilities while controlling complexity, operational risk and ongoing maintenance costs.

The real productivity gain isn't necessarily producing more software. It's delivering more value per unit of engineering effort and long-term liability.

[Subhasis Roy @SubhasisRoyUK]

Perhaps the boundary shouldn't be defined simply by what AI is allowed to write, but by the level of assurance required before that code is trusted.

An AI-generated internal utility and an authentication service carry very different risks.

The important distinction is between generating code and accepting responsibility for its behaviour in production.

As AI becomes better at testing its own output, human oversight may shift from reviewing every line towards defining assurance requirements and verifying that they're met.

[Subhasis Roy @SubhasisRoyUK]

An important distinction here is between AI helping us perform a task and AI helping us develop the capability to perform it independently.

Even when AI challenges our thinking rather than supplying answers, how do we know the learning actually transfers?

Perhaps the real test of AI-assisted learning is whether we can apply that understanding to unfamiliar problems later, without AI assistance.

That seems a more meaningful measure of whether AI is strengthening our thinking.

[Subhasis Roy @SubhasisRoyUK]

A persistent, connected knowledge base is a valuable foundation.

I wonder whether the next evolution of the AI second brain is an active learning system.

One that doesn't just retrieve what we've captured, but challenges our understanding, tests recall, generates unfamiliar scenarios and helps us apply knowledge to real problems.

AI could orchestrate much of that learning process, while we retain responsibility for the thinking and application.

That would move the second brain from storing knowledge towards developing capability.

[Subhasis Roy @SubhasisRoyUK]

Most professional learning fades because we rarely connect it to what comes next.

I've been exploring how AI can change that, turning isolated learning experiences into knowledge that compounds over time.

Here's what I'm discovering. https://t.co/bJON4EGkZ0

[Subhasis Roy @SubhasisRoyUK]

The interesting part here may be less about a specialised Challenger AI and more about how much difference behavioural instructions can make.

Modern models like Claude, ChatGPT and Copilot can already be prompted to question assumptions, challenge goals and demand clearer success criteria.

If a relatively simple change in how we instruct AI produces such a significant improvement, that has important implications for how organisations deploy the tools they already have.

[Subhasis Roy @SubhasisRoyUK]

There's an interesting risk here that goes beyond polished presentations.

AI can also produce a beautifully structured narrative, complete with assumptions, trade-offs and counterarguments, while the underlying reasoning remains flawed.

So perhaps the real distinction isn't slides versus prose, but how rigorously we test the thinking behind either.

A convincing argument and a sound decision aren't necessarily the same thing.

[Subhasis Roy @SubhasisRoyUK]

I've found that one of the most valuable ways to use AI is to challenge my own thinking.

When evaluating an important decision, I ask AI to question my assumptions, identify blind spots and argue against my preferred approach.

Sometimes it changes my mind. Sometimes it strengthens my original reasoning.

The value isn't in AI making the decision.

It's in helping me make a better one.

[Subhasis Roy @SubhasisRoyUK]

That's a good point, Ben. Writing down the reasoning is a useful way to test whether the disagreement is genuinely justified.

I did go ahead with the feature. Claude's challenge made me examine whether it would create meaningful differentiation and contribute to distribution, rather than just add functionality.

I concluded it was worth building, but with a much clearer rationale for the decision.

[Subhasis Roy @SubhasisRoyUK]

How Disagreeing with AI Made My Decision Better

Claude challenged a decision I was making about my SaaS product, ThinkExec.

I disagreed with its recommendation.

But the challenge made my decision better.

I was planning to build additional features for ThinkExec. I believed these features would give the product an important competitive edge.

Claude questioned my priorities.

Was adding more functionality really the best use of my time? Or should I be focusing on the harder problem of distribution and customer acquisition?

It was a fair challenge.

As a founder, it's easy to get absorbed in building. Developing features creates a visible sense of progress. But a better product doesn't automatically translate into a more successful business.

I took Claude's challenge seriously.

I reconsidered whether the features would deliver meaningful customer value, strengthen differentiation, and contribute to my distribution strategy.

Ultimately, I decided to continue building.

Not because I dismissed Claude's reasoning, but because examining its challenge helped me clarify why those features were worth pursuing.

And that experience reinforced something important about how I use AI.

The value of AI isn't always in changing your mind. Sometimes it's in making you examine why you believe you're right.

I don't want AI to make my decisions for me.

I want it to challenge my assumptions, expose weaknesses in my reasoning, and help me make more deliberate choices.

The judgement and responsibility for the outcome remain mine.

[Subhasis Roy @SubhasisRoyUK]

Thanks! I actually went ahead with the feature.

It's called "Check My Thinking", designed to help users challenge their assumptions and examine their reasoning.

Claude's pushback made me question whether it offered genuine differentiation and could support distribution, rather than simply adding functionality.

I concluded it was still worth building, but with much clearer reasoning behind the decision.

[Subhasis Roy @SubhasisRoyUK]

The difference between offloading work and giving up judgement is really striking.

I had a similar experience when I was building my SaaS product. Claude asked why I wanted to build more features and said we should focus more on distribution .

I took the challenge seriously, I examined my assumptions and ultimately kept building because I felt like those features were meaningful differentiators.

What I found interesting was that the AI improved my decision without changing it.

Maybe a sign of a healthy AI collaboration is that we can explain why we accepted or rejected its recommendations, not just whether we followed them.

[Subhasis Roy @SubhasisRoyUK]

The most concerning part isn't that AI can generate code faster than we can review it. It's that we may gradually lose the ability to recognise when something is wrong.

There's an important difference between understanding what the code does and understanding why it's the right solution.

I think the real challenge is preserving independent judgement while increasing our reliance on AI. Otherwise, better tools could make us more productive while making us less capable of questioning the results.

[Subhasis Roy @SubhasisRoyUK]

@marcgmbh Yes. Attention by itself doesn’t compound.

If the product can’t retain people, distribution just means repeatedly paying to refill the top of the funnel.

The real moat is distribution that turns into retention.

[Subhasis Roy @SubhasisRoyUK]

@0xForce_ The website build is almost the least interesting part.

The real shift is that one person can now compress prospecting → research → offer → delivery → outreach into one workflow.

When production gets this cheap, demand discovery and distribution become the real constraints.

[Subhasis Roy @SubhasisRoyUK]

@siirtecirlioglu Exactly. Faster implementation also means faster execution of bad decisions.

As code gets cheaper, architecture and judgement become more valuable because the real bottleneck shifts to deciding what should be built, and why.