Chicken Little @Chicky_Think

@JackieTan_ @ohcheepeng Good question. Was about to ask that.

I discovered something interesting today. Did you ever get stuck when AI is blocked from accessing certain website? A valid protocol for website owner to protect their IP. Try Claude MCP connector to Chrome. I am amazed of what it can do. I see the pages in Chrome literally changing as how i would search for information myself. These are the little things that I figure out as I build.

@ohcheepeng Nice!

True character is shown not in normal times but under enormous stress. Rushing a project, losing an important match, a child failing his exams after repeated tries. Staying calm and leading with kindness while a deadline looms. Spending time with appreciative fans after losing a match. Genuinely encouraging a kid, knowing how much his effort hasn't paid off. All of this stems from past experience — the ability to learn from mistakes and internalise it. That separates a boy from a man.

@JackieTan_ Yes, obsidian but I use Claude code to clean up and manage the wiki. I am still figuring out how to use obsidian except for the beautiful neural nets and reading of notes. But I think obsidian has more functions than just visuals.

Did a health check on my second brain, CORTEX today (11 July 26). Adjust some neutral nets — delink some loosely link and interconnected more together. The query response is much better —more substance and better linked.

keyboard_arrow_left Previous keyboard_arrow_right Next

@JackieTan_ Maybe. It resemble grab (pun intended) market share at a a loss, to lock in the users.

Wrappers over frontier models almost always lose something in translation. Copilot sits on top of Claude and GPT, but the harness adds constraints, context limits, and product decisions that the underlying model didn't choose. You're not getting raw Sonnet or raw GPT. You're getting a filtered version of it. Then the pricing doesn't add up. Claude charges more direct than Copilot does wrapping the same models. Someone is eating that cost gap. Enterprise licensing deals, loss-leader bundling, Microsoft subsidizing seat growth — pick one, but the economics don't just work themselves out. I don't have the answer. Just flagging the gap between what things cost and what they're priced at rarely stays free forever.

@tonytonggg Sure!

When I started investing years ago, I started alone. No group of friends walked in with me. Somewhere along the way I found people on the same journey, and it changed everything. Having someone who shared the interest meant having someone who understood. Now I'm on a new journey — learning AI, assembling tools others have built into things that solve real problems. Mine and others'. And it's lonely again. Because when you build, you learn something new every single time. And there's a specific feeling that comes with each discovery that you can't explain to someone who isn't building. The only person who truly gets it is the one building concurrently with you. So this is me putting it out there. If you're on the same learning journey, let's walk it together. I've done the solo stretch before. The road is better with company.

@physicsandpicts Same. I support Norway!

I asked Claude Code a simple question about my second brain setup: "what is Lint Check?" That one question taught me more about knowledge management than any course I've bookmarked and never finished. Here's why building beats watching. Lint Check isn't a button someone gave me. It's a maintenance routine I built into my own vault — a pass that catches three failure modes as a knowledge base grows: 1. Orphan notes — pages sitting with zero links in or out, disconnected from everything else you know. 2.Stale claims — a fact or number that hasn't been reconciled with what you've since learned. An old thesis point that quietly went out of date. 3. Missing links — a note that clearly relates to another idea, but nothing connects them yet. None of this existed as a concept for me until I needed it. I didn't watch a video on "vault hygiene." I built a system, watched it get messy, and had to name the problem before I could fix it. That's the difference. A course teaches you someone else's already-solved problem. Building teaches you the problem itself — which is usually the harder and more valuable half. The real unlock: once "Lint Check" existed as a named routine, it became reusable. I can run it monthly, ask Claude Code to apply it automatically, and catch decay before it compounds. A video can't hand you that. Only building something and living with its flaws does. Watching a tutorial gives you someone else's mental model, borrowed for an hour. Building your own system — even a small one — gives you a mental model that's actually yours, stress-tested against your own mess. That's the trade. Slower to start. Faster to compound.

Wrappers over frontier models almost always lose something in translation. Copilot sits on top of Claude and GPT, but the harness adds constraints, context limits, and product decisions that the underlying model didn't choose. You're not getting raw Sonnet or raw GPT. You're getting a filtered version of it. Then the pricing doesn't add up. Claude charges more direct than Copilot does wrapping the same models. Someone is eating that cost gap. Enterprise licensing deals, loss-leader bundling, Microsoft subsidizing seat growth — pick one, but the economics don't just work themselves out. I don't have the answer. Just flagging the gap between what things cost and what they're priced at rarely stays free forever.

Ran the Claude vs Copilot comparison against my own actual usage, not a spec sheet. The answer wasn't close. Here's why. [1/6]

@JackieTan_ I am staying rent free in your mind. Hahaha

Every tech cycle starts the same way. Groundbreaking, frontier, everyone flocks to the best. iPhone launch, top Android flagships — you needed the highest spec to do what you wanted. Then the curve does its thing. Calls, WhatsApp, internet banking — the low end phone handles all of it now. Same story played out with PCs. We used to chase the highest spec just for daily tasks that a modest machine handles fine today. AI is walking the same path. Coding used to demand the frontier model. Now the average, even below-average model is often good enough. Good enough is the part that matters for cost. As more people adopt at mass scale, they gravitate to what's sufficient, not what's best. That's deflationary for the frontier players — more spend, more compute cost, and the return may not show up the way it used to. Watch for the tells. Frontier labs starting to sell compute instead of just models. Users pivoting to "good enough" — open source, Chinese models — because the gap no longer justifies the premium. So here's the thought for investors chasing the best model, the best technology, every cycle. Is the best always where the return lives? Or does the curve eventually make good enough the only thing that matters.

Many things in life are measured by relativity, not absolutes. If you're good at work, your colleague might be even better. You increase your value a lot, but they increase theirs by multiple times more. You still end up with the smaller bonus. Not because you did badly. Because the comparison moved faster than you did. Same with products. You build something valuable. A competitor ships something five times better. Your effort didn't shrink. Its relative value did. Two students study for the same exam. One puts in ten hours and scores 85, feels good about it. Then the curve comes back, class average is 92, because others found a better method. The absolute score never changed. The relative standing did. Switch this to AI. You work hard without it, push output up 2x, feel like you earned the bonus. Meanwhile a colleague using AI pushes output up 10x. He creates more value. He gets the bonus. Not because you slacked. Because relative to what's now possible, 2x is standing still. Everything claims to be absolute. Almost none of it is.

@JackieTan_ Nope but it look interesting from the link. Have you tried it?

@JackieTan_ I see. Then automate it. 😁

Sat in on a sharing session by someone from a big tech company today. His team gets unlimited tokens for AI work. There's even a leaderboard tracking who burns through the most — token maxing, officially encouraged. His job has moved in stages. First it was coding alone. Then coding with AI. Now it's writing specifications and letting AI run autonomously. Productivity is up roughly 5x since he started. He called AI an inevitable force — not optional, not a nice-to-have. He also flagged something worth sitting with: unlimited tokens won't last. The subsidy ends eventually, and then it becomes a constrained resource like everything else. The part that stuck with me most was about harnesses. Copilot, Claude Code, Claude Cowork, his company's own internal tool — these are all interfaces harnessing the same underlying models. Each one has its own learning curve. But the harness changes fast. The UI changes fast. The fundamentals underneath don't. So the better bet is learning directly from the frontier model itself — Claude, GPT, whichever — rather than getting fluent in someone else's wrapper. Learn the fundamentals and any new interface becomes easy to pick up. Learn the wrapper and you're starting over every time the wrapper changes.

@JackieTan_ 👌