[9/18] 2026 Goal Progress

The usual round of writing up progress on the goals I declared. This week I just got sick, and honestly I barely remember it.

Health

I was actually exercising, and then I got sick anyway

I was getting some exercise again this week, and then I just went and got sick.

It started with a sore throat, then a runny nose and headaches.

It's finally on the mend, but now I've got this dry cough. It's rough. Currently in recovery mode.

My left under-eye keeps twitching

Since I was sick, I was sleeping long hours.

But every morning I wake up and the area under my left eye twitches a bit. I looked it up, and the usual suspects are too much caffeine, too much alcohol, stress, or not enough sleep.

Lack of sleep isn't it, I don't think. And I don't drink, so alcohol's out.

Caffeine — sure, I have coffee every day so maybe it's on the higher side, but it's not like I'm overdoing it. Two cups a day, tops.

So that leaves stress. I'm in a situation where I have to keep thinking about genuinely complex things with a brain that isn't firing at all, so I suspect that's what it is.

Also, I started watching Blue Lock and couldn't stop, so I did stay up too late a couple of nights. That might be part of it.

Quitting smoking: sick, and somehow smoking more

Somehow, despite being sick, my count hasn't gone down — it's gone up.

It's not like I'm at two packs a day, but I'm definitely going over a pack.

Work

Found two "AI employee" ideas we can build while having fun

The work highlight: on the AI employee front, I found two ideas that are genuinely fun to build.

And I got as far as actually distributing them across the team so we build them together. That's the big one.

The two: one is an AI employee that writes the daily report. The other is an AI employee that manages all the small day-to-day tasks in a single DB.

Cloudflare's MCP Portal lets us distribute safely by role

Cloudflare has a new feature called the MCP Portal, and using it we've been able to create policies for admin, staff, and contractor roles, branch which MCP servers each role gets, and distribute them safely.

We've gotten that to stick with the team, and now people other than me are adding new MCP servers left and right. That's great.

Separately, there's the project-level thing I'm building — pushing it further toward being agentic.

I'm hammering on that myself, local environment only for now. I've been embedding daily feedback into it and polishing it bit by bit, so it's decent progress.

The problem: best practices keep shifting with model performance

We're adopting AI aggressively, but the fact that best practices keep getting swung around by model performance is genuinely hard to deal with.

Right now, for implementation, honestly Codex is performing better than Claude.

Conversely, Claude has gotten noticeably stronger on security — more knowledge, more capability there.

Until recently the flow was: have Claude write the code, have Codex do a rigorous review, keep quality up. That's starting to flip now, I think.

Which means: our AI agents and workflows basically run on Claude, but they need to work on Codex too. Designing skills and rules that way — building a foundation any AI can read — is going to be necessary to some degree.

That said, even staying on Claude, we're getting a serious amount of efficiency out of it.

The bottleneck is wherever humans have to show up

But for the parts humans have to think about, the places where a human unavoidably enters the picture — communication, misaligned understanding, those bottlenecks — there's still a pile of unsolved problems.

The plan is to go after those relentlessly using AI.

And honestly, I'm not sure about a world where you hold back individual strengths in order to support everyone equally.

Everyone has something they're strong at, and if you understand your own strength, I think you grow it on your own. And since everyone's strength is different, when those converge it produces enormous power.

So what I'm chewing on is: how do you let that power run free without it becoming a bottleneck? And I think it comes down to building a free space where AI can roam.

Then humans just attend meetings and nail down requirements. That's it.

And then it's "oh, this is already done?" — and off that event as a trigger, an AI employee handles relaying it to stakeholders and the key people internally. If we had that, I suspect it'd clear something like 80% of the bottlenecks, with high reliability.

Money

US stocks are doing fine. Still haven't taken profit

On the investing side, US stocks are doing reasonably well.

I still haven't taken profit on anything.

The crypto bot: four straight months of this means something is fundamentally wrong

The crypto auto-trading bot is bleeding red, but it's not a scary amount of loss, so I'm leaving it alone for now.

Still — four straight months of underperformance means something is fundamentally wrong, I'm pretty sure.

So I'm aggressively looking at rebuilding it, new bot concepts, that kind of thing. Working through it together right now.

Got a raise

Nothing new on cutting costs.

On income, I was fortunate enough to get a raise, so take-home goes up a bit. I want to put that toward more investing, or family trips — something along those lines.

Duolingo at 260 days. But nothing is sticking

Duolingo is at 260 days straight.

But being sick meant I barely studied, so it's down to about 5 minutes a day, and nothing is going in.

I also want to go somewhere before the year is out, but with my health like this I haven't looked into it.

My family has been bringing it up too — we've been talking about wanting to go abroad once — so I should actually start researching it properly.

Other

A week I don't remember

Nothing else really stood out.

I was sick, so I couldn't get out much. A week I don't remember.

Wrapping up

Next week is about one thing: getting healthy.

Silver Week starts next week, so to make sure I can actually go all in on enjoying it, I'm going to be much more deliberate about managing my health, my meals, and my sleep.