Reading Log

by Kurt Pan

“You’ve got to start with the customer experience and work backwards to the technology. You can’t start with the technology and try to figure out where you’re going to try and sell it.” — Steve Jobs

Lean lets mathematicians treat mathematics as code—break it into structures, theorems and proofs, import each other’s theorems, and put them on GitHub. The big idea is that eventually much of the humanity’s mathematical knowledge might be available as code—statically checked, verifiable, and composable.

The result defies mathematicians’ usual intuitions about what functions can and cannot do.

“There was this inescapable sameness, in a way. No matter what I did, I was in the same place doing mostly the same things,” she said. “I was very isolated, and nothing I could do could really change that. I’d wake up on certain days and realize, I’m just older.” Math became a kind of escape, a space that felt expansive when her daily life was not. “Mathematics was another world I could explore. A world that was not confining, a world I could access at any point just by thinking about it,” she said. “That’s how I grew up, thinking about mathematics as this world of ideas that I can explore on my own. That sort of process helped me see math differently than a lot of people.”

Fast eliminates cognitive friction. No one praises these tools explicitly for their speed. They just feel magical.

Network calls and dependencies reveal themselves through latency, and this brutal honesty forces discipline. Companies that do fast very well tend to have very focused products. In a world obsessed with adding rather than refining, speed becomes the ultimate expression of respect. It says, “We've thought deeply about what matters and eliminated everything else.”

You'll see more and more companies optimizing for low latency, interface design, connectivity, and reliability. This, in turn, will unlock new capabilities and use cases that we aren't yet even thinking about.

I know the code was generated because it was written in a way no developer on the team would.

The part we’ve struggled with is making changes to that after you push it out the door.

I want people to care about quality, I want them to care about consistency, I want them to care about the long-term effects of their work. LLMs are engineering marvels, and I have the utmost respect for the people who’ve created them. But we still need to build software, not productionize prototypes.

Write better prompts. Give better descriptions. Tell the LLM what library to use. Give it examples to follow. Write smaller files. There are no new principles – follow the ones that already exist.

MelvinGr/CandySimplyFi-tool existed. This is a simple C++ program that already knew some known strings, and was able to brute-force the key in a matter of seconds. I was able to compile this on my laptop, and run it, and sure enough, it was able to decrypt the data in a matter of seconds. For now, I plugged this key into CyberChef, and was able to decrypt the data.

A common symptom of human mitochondrial dysfunction is a massive feeling of tiredness.

This also strongly suggests that sleep and hunger are both tied to mitochondrial function and energy balance (the latter was already pretty clear!), and that aerobic organisms are constantly adjusting for both fueling their mitochondria and giving them (especially the ones in the central nervous system) some down time for repair and recovery.

If you need oxygen, then you need sleep!

I love fireflies. But in recent years, they stopped coming for reasons I don’t know. No tiny, glowing dots in the dark like they used to. I miss them more than I expected.

Few weeks ago, I decided to make my own fireflies. I discovered something called an Astable Multivibrator. It looked like magic to me TBH. A tiny circuit, with minimal components, that could automatically switch from ON to OFF and OFF to ON state (without an explicit “timer” component).

In this article we go through the fundamentals of embeddings. We will cover what embeddings are, how they evolved over time from statistical methods to modern techniques, check out how they're implemented in practice, look at some of the most important embedding techniques, and how the embeddings of an LLM (DeepSeek-R1-Distill-Qwen-1.5B) look like as a graph representation.

“mathematics gives you power” — the power to understand chemistry, physics and other aspects of the world.

“Do not give up,” she likes to tell them when they’re struggling. “A problem proves its worth by fighting back.”

These Ramsey numbers are not very big. With all the computing power today, shouldn’t someone just crank out an improved lower bound by computer? The thing that drives many people crazy is that even in this case, it’s beyond the boundary of our knowledge.

Mathematics is big. No one knows everything. So collaboration helps give you a broader point of view — a wider scope.

Also, a problem usually will not die as easily if you’re collaborating on it. If you’re frustrated or discouraged, then your collaborator might share some light and pull you out. Usually, one piece of work will lead to another, so it’s just continued year after year. That’s the best situation.

Math is a great career. It’s fun. It’s interesting. It keeps the mind curious. It’s good for making friends. It’s very good for your health.

Here's how the AI pricing bait-and-switch works: 1. Launch with generous/unlimited limits 2. Build user dependency 3. Add caps targeting “less than 5%” of users 4. Frame as “sustainability” or “fairness”

Sound familiar? That's because we've seen this movie before. Cursor did it. Windsurf followed. GitHub Copilot tightened screws. Now Claude Code.

The real damage: You're not frustrating 5% of users—you're breaking trust with the exact people who drive growth and adoption.

What we want: Transparent costs upfront so we can make informed decisions.

Anyone serious about designing for AI should consider non-copilot form factors that more directly extend the human mind.

Routine predictable work might make sense to delegate to a virtual copilot / assistant. But when you’re shooting for extraordinary outcomes, perhaps the best bet is to equip human experts with new superpowers.

A DKIM replay attack is when an attacker captures a legitimate email with a valid DKIM signature and re-sends (replays) it to new victims. Since the body and signed headers remain unmodified, the DKIM signature still validates, making the spoofed email appear authentic.

It’s hosted on a trusted google.com subdomain (like sites.google.com), many users let their guard down. But don’t be fooled – just because the domain looks legitimate doesn’t mean the content is.

Now we hear the chainsaws revving as private equity’s playbook of mass firings, squeezing users, and aggressively monetizing the product starts to play out. I’ll argue that Komoot is neither a moral failure nor an outlier but the capitalist system of value extraction working exactly as intended for the platform owners. For-profit corporations squeeze and sell us out when we give them the opportunity. Corporations we depend on are swallowed all the time.

For corporations, it’s always profits over people. In financial capitalism, companies themselves become commodities. This pits owners against the people who build the platform in the pursuit of shareholder value. To capital, the corporation is a vehicle for profit; the platform is their plantation. Capitalists see our forests only for their timber value, and they wield the power to impose their limited view on us. Private equity’s business model lies in squeezing the maximum amount of profit from the company until it dies and then throwing it away.

Capital generates profits and user base growth, accumulating ever more capital on digital platforms through algorithms that harvest user data. To boost engagement and capture groups of users, capital co-opts our human instinct to connect with one another and contribute to a “community” and sells it back to us. As corporations—bureaucratic entities—cannot create value themselves, they instead lure us, social beings, with promises of “community.” In contributing to a corporate “community,” we’re providing free digital labor toward proprietary, engagement-driving features that expand the user base for the profit of the owners. In fact, we suckers even pay for the privilege. Betrayed, we see this “community” for what it is: a mirage.

Capital forcibly converts our public commons into its private property in a process called enclosure, commodifying our work. Enclosure by private interests interrupts the fundamental reciprocity of the commons, eroding its abundance over time.

The corporation is always a vehicle for generating profit and capital accumulation, always in conflict with the interests of the user base.

We begin to see that the next corporate platform courting our favor is nothing but another trap. We cannot outfox organized capital with our individual choices. Underneath the marketing veneer and despite the best intentions of employees and possibly some owners, all corporate platforms are just different profit-seeking vehicles of capital. Corporations cannot be our friends or part of our community as they indifferently co-opt, enclose, and extract from our commons until there’s nothing left.

Yet the clear-cut also contains the seeds of the next old-growth community. At this moment, a rewilding of the internet is occurring in many places. A decentralized movement toward open, cooperative platforms that support real community, rather than zero-sum corporate walled gardens. Promising projects such as the Mastodon social network, Matrix chat, and Pixelfed social photo sharing are reviving the diversity and abundance of the early, independent internet before it was enclosed by tech giants in the 2010s. More than singular platforms, the Fediverse represents a growing ecosystem of open protocols and distributed services that guarantee freedom of movement for users and data and push back against capitalist enclosure—a diverse and resilient digital commons.

Their primary motivation for pushing the app, more often than not, seems to boil down to gaining more access to your personal data and behavior.

The perceived “convenience” of an app often comes at the cost of your privacy and control.

So, the next time you're met with that insistent prompt to download an app, take a moment to consider what you might be giving up. For me, I'm sticking to the website.

Humans don’t really learn when we download info into our brain, we learn when we expend effort to pull that info out.

The “thing” that we learn most effectively is not knowledge as we typically think of it, it’s process.

Humans learn collectively and innovate collectively via copying, mimicry, and iteration on top of prior art.

We build tools to help us think, not to think for us.

I like to imagine AI as an “absent-minded instructor”, not as a coworker. It’s prone to forgetting details, but ultimately there to guide you; most importantly, the goal of the instructor is to make sure you learn and learn how to learn!

Please, y’all, put the emphasis on humanity back into our tooling rather than pretending nothing matters, as if somehow humans will supposedly be irrelevant in a few years.

This is a call for you to do the same. Build your own LAN. Connect it with friends’ homes. Remember what is missing from your life, and fill it in. Use software you know how to operate and get it running. Build slowly. Build your community. Do it with joy. Remember how we got here. Rebuild a community space that doesn’t need to be mediated by faceless corporations and ad revenue. Build something sustainable that brings you joy. Rebuild something you use daily.

Bring back what we’re missing.

For the first step, Khurana and Tomer focused on a quantum version of a one-way function, called a one-way state generator.

Khurana and Tomer dubbed these perplexing new building blocks one-way puzzles. But as Khurana and Tomer grappled with that task, they decided to take a more direct approach: Forget about the one-way state generators, and instead anchor one-way puzzles directly to the mathematical bedrock.

This problem, known as the matrix permanent problem, is notoriously difficult to solve for large matrices, and there’s no simple way to check whether a calculation is correct. The matrix permanent problem also has other special mathematical properties that cryptographers find appealing. “This would be a beautiful problem to base cryptography on,” Khurana said.

All data exists in one of three states: At Rest (stored on disk) In Transit (moving over a network) In Use (being processed in memory)

We have robust solutions for the first two: At Rest: Disk encryption, file system encryption. In Transit: TLS/SSL, VPNs, end-to-end encryption.

But in use—when data is loaded into RAM and processed by CPUs—it is decrypted. This is the Achilles’ heel of modern security.

The future of internet computations will be encrypted. It is not if, it is when.

Production systems need 99.9%+ reliability. Even if you magically achieve 99% per-step reliability (which no one has), you still only get 82% success over 20 steps. This isn't a prompt engineering problem. This isn't a model capability problem. This is mathematical reality.

The most successful “agents” in production aren't conversational at all. They're smart, bounded tools that do one thing well and get out of the way.

The pattern is clear: AI handles complexity, humans maintain control, and traditional software engineering handles reliability.

The winners will be teams building constrained, domain-specific tools that use AI for the hard parts while maintaining human control or strict boundaries over critical decisions. Think less “autonomous everything” and more “extremely capable assistants with clear boundaries.”

Enter your email to subscribe to updates.