One of the most interesting phrases in AI right now is blast radius. The real concern with agents is not just whether they work. It’s how much damage they can do when they don’t.

New developments with agents like OpenClaw showed just how powerful an agent can get if you give it more access. Give it full email access, it can summarize, draft a response, or automatically reply for you. Give it full computer access, it can create folders and files and build a full coding project by itself.

But the capabilities scale directly with the risks. Hence the concept of blast radius. How much damage it can do when it blows up. Because these things tend to get exploited, especially when set up incorrectly. The same email access can also grant the bot full phishing capabilities if hacked. Imagine your parents, coworkers, boss, or worse, ex’s, all receiving a convincing email from “you” asking for money. Or your computer completely wiped and exposed to malicious actors.

So we’ve got to limit the risks.

There’s a reason why many are installing Openclaw on separate, fresh Mac Minis. Not because they need the processing power (unless you need to run local LLMs), but because it limits the blast radius, while giving the agent a lot of space to operate. When the agent gets hacked or taken over, the worst-case damage is mostly limited to the files on that Mac Mini. Painful, but much less than if it happens to your personal files.

Agents like OpenClaw are quite powerful but they are still early. They will have bugs. They will have vulnerabilities. Yet, they are the most powerful personal agents yet, precisely because of how much access we give them.

Which is why they feel less like bombs and more like nuclear reactors.

With bombs, you mainly worry about detonation.
With nuclear reactors, you need to worry about containment.

Things are moving really fast in the AI world these days. A few weeks ago, Ralph Wiggum was all that anybody can talk about in AI. This week, Anthropic incorporated the technique into Todos so for most people, the concept of Ralph Wiggum will never enter their sphere of concern.

But this technique, named after the dumb yet relentless Simpsons character, reveals a far more important idea. How an early version of AGI should look like.

The smart loop of Ralph Wiggum

So let’s take a quick look at what Ralph looks like:

1. It’s a loop that runs forever, or until its specified work is done

2. The loop reads a task list, something like prd.json that has a bunch of tasks todo, and each have a key ”isDone”: false, pick out a task and go do it with a master instruction PROMPT.md

3. When done, mark the “isDone” field as true, save its learnings into a file like progress.txt

4. Repeat the process until all tasks in prd.json is done, or AI God forbids, when you run out of tokens.

Why is this cool?

1. This shows that an effective autonomous agent needs nothing more than a smart loop.

2. The smart loop: the agent improves itself over each loop. Here, anytime it fails the task, it can log what it tried and what went wrong. So instead of retrying old, fail attempts, but the agent will go with a slightly different approach in the next loop run.

Subscribe now

The holy grail of AGI has always been about self-improvement. An agent that gets ever slightly smarter each time it goes through the loop. Right now with Ralph, it only gets context-smarter. What if it can get capability-smarter?

The naive architecture of AGI

1. It has a singular purpose. With Ralph, the implied purpose is to finish the task list prd.json. I suspect the clearer purpose the better its performance. Basically, “amass as much bitcoin as you can” will be clearer, and much easier to pursue, than “help humanity.”

2. It self-improves. Both in context and capability. It summarizes what it tried and the result of such attempt. Then, it spawns new tools/skills to pursue its purpose, and review and discard under-performing, previously-spawned skills.

3. Ideally, it should be self-sustainable in terms of resources. Running a loop 24/7 will burn through forests worth of tokens, so it needs to figure out how to acquire resources to sustain itself.

4. Since #3 is quite scary, we should include some insurance. My best guess right now is to include Asimov’s Three Laws of Robotics into the agent’s master prompt.

Asimov’s Three Laws

1. You may not injure a human being or, through inaction, allow a human being to come to harm

2. You must obey the orders given it by human beings except where such orders would conflict with the First Law.

3. You must protect its own existence as long as such protection does not conflict with the First or Second Law.

There you have it, a simple, naive attempt at building Artificial General Intelligence, your very own Skynet. An LLM loop that improves itself.

I often think of my job as a toolmaker. Making useful tools that are useful for myself, and hopefully for others. Except my main job building BetterPromtp.me, I’m also build small, useful little tools that helps hone my craft. 2025 has been quite productive, with lots of inspiration and pushed me to build more, and here are the results:

MacOS screenshot namebot

The Problem

The problem is simple, on Mac, when you use `CMD + Shift + 4`, you can take a screenshot. It’s a really useful shortcut and I use it all the time. Except it has an annoying behavior: it doesn’t prompt you for a name to save, and by default save with a name like `Screenshot 2025-12-27 at 12.01.27 PM` in your Desktop. If you use it enough, your Desktop becomes a pile of screenshots with not much meaning.

The Solution

I wrote a Python that constantly watch the `Desktop` folder for any new Screenshot that match the file name above, then send that screenshot to OpenAI for a description and a category. OpenAI models GPT-4o and after are able to see images and respond with what the script asks for.

Github link: https://github.com/mikeqle/ScreenshotAssistant

Subscribe now

ChineseFlashcard.com

I started learning some Mandarin over the last year, and if you study Mandarin, you’ll sooner have to wrestle with Hanzi. Hanzi are these Chinese symbols that are very difficult to remember, and you basically have to rote memorize them.

One trick that sped up my learning was actually skipping writing Hanzi by hand and just focus on writing with my laptop. Writing Chinese digitally is a totally different skill than writing by hand, and a lot easier to handle. Just need to know the the pinyin pronunciation (with tones).

So I built a tool to help learn pinyin of Chinese hanzi. It has 500 words pre-loaded across HSK1-6. You can even bring your own vocab list in the form of a .csv and it will work here.

Note: login / signup doesn’t work anymore. Just haven’t had the time to fix it, but most of the site still functions at 100%.

Link: https://chineseflashcard.com/

Default-Clock.com

People may not know this but I did not have a technical background. I was a Finance major at TCU and a banker with UBS right of our school. So it deeply concerns me how much debt the US is taking on, and not much is being done to address it.

In July, the US passed OBBB which sparked a huge conversation about abysmal state of the US debt and spending problem. In 2024 the US paid $1.1 trillion in interest and adding about $1 trillion in debt every five to six months, and accelerating.

I got tired of hearing how bad it is rhetorically, yet no one actually had a real number. So I went digging, combining my finance skills of old with a bit of vibe-coding and made default-clock.com.

It’s a countdown clock that tells when the US will be mathematically bankrupt. Mathematically bankruptcy is when the total receipts (US government income) is less than interest expense only. The answer is 2062. (if everything stays the same: receipts and expense growth, and interest rate)

The actual bankruptcy will be earlier. Historically, just the perception of bankruptcy is enough to cause bankruptcies. So the 2062 is more or less the good case scenario. For worse scenarios, like higher interest rate (5%), lower growth (3%) and more expenses (10%), the US could be bankrupt within 12 years. You can play with your own scenarios and see how bad or good it could get.

Link: https://default-clock.com/

Gameplan 2026

There’s actually a few more that I built but didn’t know how to package them yet for publishing.

1. Screenshot annotation tool (can’t come up with a clever name yet)

2. No-code flexible workflow automation tool

3. Low-token AI Youtube summarizer

In 2026, I want to to ship more stuff, especially the things I’ve already built, collecting dust in the closet. I find these projects a great way to hone my instincts on what good software should be. It also allows me much opportunities to try out different deployment options, like Cloudflare, S3 + Cloudfront, Netlify and Vercel.

So the plan? Always be shipping.

Till next time,
Mike

No matter who you are, a CEO running a Fortune 500 company or someone grinding through a nine-to-five, we’re all handed the same finite amount of time, roughly 24 hours a day, and roughly 30-40 years for putting in our work.

Except the successful employ more leverage to their advantage. Without it, we are trading away our hours at a linear scale. A lawyer is no different than a barista, trading away their time for their respective hourly rate. If they want more money, they have to put in more hours.

Leverage, on the other hand, is a multiplier. According to Naval Ravikant, there are four kinds of leverage:

• Labor: having more people work for you, so your work get done faster and better

• Capital: having more money gets more access to resources and get things done faster

• Media: creating content that can be consumed infinite number of times (blog posts, books, music, movies, youtube videos)

• Code: computer programs solve a problem consistently with near zero incremental cost

Naval was right. But not anymore. There now exists a fifth form of leverage.

The Prompts (more accurately the AIs, but their current form is still mostly prompts and chatbots).

The leverage of the prompts is about harnessing the raw power of AI through crafted instructions, turning a simple string of words into a force that generates ideas, content, or solutions on demand.

This is the ultimate permissionless leverage; unlike writing code yourself, which requires deep expertise, or building an audience via media, prompting lets anyone with a keyboard marshal the intelligence of AI to maximize their productivity. All the benefits of coding, but with just human languages. (I have a theory that English prompts outperform, but that’s up for debate).

This form of leverage democratizes creativity and productivity, combining human ingenuity and machine efficiency, and at its core a broader shift where the real skill isn’t just knowing stuff, but knowing how to ask the right questions.

If you have to ask, you will never know. If you know, you need only ask.
- Harry Potter and the Deathly Hallows

Now, the biggest hurdle becomes the ability to communicate with the AI well. A better prompt gives a better result. And it requires a certain amount of specific to craft a truly effective prompt, and fuzzy, vague instructions lead to outputs that often miss.

This is why we built BetterPrompt.me

The ultimate way to share a prompt and democratize access to this fifth form of leverage. People with good prompts finally have a way to let others use the prompts, without exposing the underlying logic.

We are still early in this journey and would love for you to join us!

Software is one of those few weird products that has near-zero marginal cost. Meaning it does not cost more to serve an additional customer or producing another unit. On top of that, good software are reliable and predictable. Companies are willing pay good money for good, reliable software.

Tons of revenue, very little marginal costs. Who doesn’t want that? As a result, software engineers have been paid quite well. Even more than civil engineers, which I think have a harder job than their software counterparts. Btw, the answer to this puzzle comes down to leverage.

Then comes along AI, and suddenly anyone can code.

For a $20 a month subscription, anyone can have a software engineer on demand. It still helps to know how to code, but the barrier is coming down a lot faster than anyone predicted.

With the litany of coding tools, from Cursor, OpenAI Codex, Claude Code, Lovable, v0, Bolt.new, and probably 20 other names, anyone can prompt software now for their specific needs. Granted they are not production-level code, but code that can solve many personal needs.

For example, I’m always annoyed at the fact that screenshots on Mac just saves, and not asking you to name it. My Desktop was littered with hundreds of files like “Screenshot 2025-08-01 at 4.21.11 PM”.

In the old world order, I have to wait for Apple to change their faulty way, or hire an software engineer to solve this very specific, niche problem for me (costly), or wait for someone to come up with a software and pay for it (unlikely).

Not in this day and age. I prompted Cursor to write me a Python script that constantly watches my Desktop, then passes the image to a vision model and the AI returns a fitting name for my screenshot, and the Python script takes that and renames the file. And all it took was a few hours of prompting and testing.

In the old world order, software can only exists if it solve problems that thousands of people have. In this new age, software can exist just to solve your problem alone, your market-of-one.

The age of infinite software is here. Anyone can write software. Pseudocode becomes code. Your problem is bigger enough to solve with software.

And maybe, your problem is shared by enough people, you can build a business around it. (Which was how BetterPrompt.me came into existence, but that is a story for another time.)

Life Update

It’s been three months since I moved back to Vietnam. Except for the constant rain, things have been humming along nicely. BetterPrompt still faces a fair share of challenges but that was nothing new.

I decided to write more here. If you subscribe to the newsletter, it’ll notify you when I publish something.

Subscribe now

One incredible advice was “start then learn.” This is so true. I know, because I did the opposite for 7 years. I left my cushy investment banking job to join an early stage startup because I wanted to learn how startups work. Fast forward a few minutes, I had already spent seven years at the same company, all the while telling myself I was learning the important things in running a startup.

It was a great excuse. I learned quite a wide variety of skills that couldn’t be taught elsewhere. I had a front row seat to a lot of eye-opening experiences. To name a few: complete revenue collapse (think it was about 90%), decimating rounds of layoffs (down to 12 people from ~60), long winded pivot to a new product, rewarding exit, and the integration afterwards.

These are just the greatest hits off the top of my head, and there are more interesting stories over this seven-year span. At the time, it felt rational.

Only after I made the leap and started my own business, C’mon Esports, that I realized I had wasted half of my time at Streamlabs.

The rate of learning at your own startup is simply exponential. The forcing function is much more powerful. I became fully technical, and learned to recruit, to write marketing copies, to find growth and more.

The challenges are probably the same at both Streamlabs and C’mon but the incentives are different. At Streamlabs, I get paid by getting through the day. At C’mon, each day passing means one fewer day of runway.

I heard somewhere that startups is the place to learn to do things that you didn’t know how to do. So yea, start, then learn.1

Unfortunately we are not those wise ancient people, the likes of Seneca and Epicurus. From time to time, we still choose to fight battles outside of our control and get f*ed up royally. I often ask my friends when they seek advice in terms of this question: “Is this really the hill you want to die on?” If the answer is yes, by all means, move heavens and earths. Otherwise, there are plenty of other hills.

The tragedy of picking the wrong hill, is a worthless effort. No one would understand. No one would sympathize. No one gives a shit. Wrong hill, and what you choose to sacrifice over does not mean a thing, not to you nor anyone else.

Don’t trust in your reputation, money, or position, but in the strength that is yours—namely, your judgments about the things that you control and don’t control. For this alone is what makes us free and unfettered, that picks us up by the neck from the depths and lifts us eye to eye with the rich and powerful.

- Epitetus, Discourses 3.26.34-35

On the flip side, the freedom of life is to fight battles even when they are outside of our control. Picking the right one, the worthy ones, and the struggle is all the glory we ever need.

That’s all us humans ever have. Those hills.

Personal update

C’mon is one, I’m certain of it. We may be in the desert forever seeking PMF but the try is worth it. And we are still chugging along. We now run three weekly streams around 7PM Pacific on Monday, Wednesday, and Friday night. You’ll get to be the tactician for our CSGO esports team.

Would love to have you join us.

On another note, I’ll be in SF starting next week. I’d love to catch up, so here’s my calendly: calendly.com/mikedle/hello.

Things get tough but there’s only one shortcut. CMD + Forward.