AsianDadEnergy's Podcast

Ivy-League educated, Ex Big Tech, Middle aged Asian Dad figuring out life.

This is a very public journal of anxiety, existential dread, and way too much tech knowledge. Basically therapy, but with Wi-Fi. asiandadenergy.substack.com

  1. 34m ago

    The Black Box of AI May Have Just Been Cracked?

    AI is moving at an absolutely ridiculous pace. It seems like every week there’s another breakthrough, another model, or another research paper claiming to change everything we thought we knew about artificial intelligence. But every once in a while, something comes along that genuinely makes me stop and think. Recently, I came across a series of experiments conducted by researchers at MIT involving something called Weight-Space Programming. And honestly, this one really blew my mind. The researchers demonstrated that a powerful AI model, OpenAI’s Astra, could do something rather extraordinary. It could inspect the internal weights of other neural networks, figure out what those networks were doing, modify their weights to change their behavior, and even generate entirely new neural networks from scratch. No traditional training. No gradient descent. No millions of training examples. Just an AI model directly manipulating the numerical structures that make other AI models work. In my opinion, this could become one of the most interesting developments in AI since the emergence of agentic harnesses. But before we get into the experiments, let’s take a step back and understand why this matters. How Do We Actually Build an AI Model? Most modern AI systems, including ChatGPT and Claude, are built using a type of neural network. At a high level, you can think of a neural network as a gigantic mathematical structure containing billions or even trillions of numerical values. These numbers are called weights. They are typically represented as floating-point numbers, and collectively, they encode the patterns and information the model has learned during training. The interesting thing is that these weights are completely incomprehensible to most human beings. If you were to open up the raw weights of a modern AI model, you wouldn’t find neatly organized instructions explaining how to write Python code, solve mathematical problems, or answer questions about history. You would find enormous collections of numerical values. Yet somehow, all of the model’s learned capabilities are encoded in these numbers. Traditionally, creating an AI model involves a process called training. And when we want to improve or modify an existing model, we often use a related process called fine-tuning. Both approaches involve feeding the model enormous amounts of training data. An algorithm called gradient descent then gradually adjusts the model’s weights based on those examples. Essentially, the algorithm makes trillions upon trillions of tiny numerical adjustments, gradually shifting the model’s behavior toward the results we want. The problem? This process is incredibly computationally expensive. Training a large AI model requires massive data centers packed with specialized AI chips, consuming enormous amounts of electricity and computing resources. And this is where Weight-Space Programming proposes something radically different. What if we could skip much of that traditional training process and simply have another AI model directly manipulate the weights? Instead of relying on gradient descent to slowly adjust a neural network, we could give a powerful AI model access to the network’s internal weights and ask it to analyze, understand, and modify them directly. Sounds like science fiction. But that’s essentially what these researchers started demonstrating. Let’s look at some of the experiments. AI Can Reverse-Engineer Other AI Models One of the most fascinating capabilities demonstrated was the ability to inspect the weights of an existing neural network and figure out what that network was actually doing. To test this, researchers trained four relatively small neural networks to play four classic arcade games: * Breakout * Pong * Pac-Man * Freeways They then gave Astra the raw weights of each neural network. But here’s the interesting part. They didn’t tell Astra which games were involved. They also didn’t allow Astra to actually run the neural networks to observe their behavior. Astra could only inspect the numerical weights. And somehow, it correctly identified all four games, with confidence levels above 95%. Think about that for a moment. Imagine handing someone a massive spreadsheet containing millions of seemingly random numbers and asking them to figure out which video game those numbers were designed to play. That’s essentially what Astra accomplished. And the researchers took this experiment even further. They trained three small language models on three different types of data: * English literature * Japanese literature * Python programming code Again, Astra wasn’t allowed to prompt or run these models. It couldn’t even inspect their input and output layers. It could only examine their internal weights. And somehow, Astra correctly identified the training language of all three models. Even more impressively, it identified the induction heads inside each network. Induction heads are specialized attention mechanisms that help transformer models recognize patterns they’ve encountered previously. For example, if a model sees a particular sequence of words or symbols, induction heads can help it recognize and reproduce patterns that appeared earlier in the context. Essentially, Astra was able to reverse-engineer parts of the internal computational circuitry of these neural networks just by examining their weights. This is a pretty remarkable development. One of the biggest challenges in AI research is understanding what actually happens inside these enormous neural networks. We know how to train them, and we can observe their inputs and outputs. But understanding exactly how their internal components work is an entirely different problem. If AI models can help us interpret these internal structures, we could potentially make significant progress in understanding the black box of artificial intelligence. AI Can Perform Brain Surgery on Other AI Models Now, identifying what another neural network does is already impressive. But what if we could actually change its behavior? That’s another capability the researchers demonstrated. In one experiment, they trained a neural network to control a car driving up a ramp. Unfortunately, the AI had a problem. The car would repeatedly drive straight into the ramp and fail to climb it. The researchers gave Astra the network’s weights, a description of the environment, and a recorded example of the failure. Astra analyzed the situation and figured out what was going wrong. The car needed to reverse before accelerating so it could build up enough momentum to climb the ramp. It then directly modified the neural network’s weights. And here’s the crazy part. It only needed to modify approximately 3.5% of the weights to fix the problem. The resulting network successfully completed the task. This is essentially like performing brain surgery on another neural network. Rather than retraining the entire model from scratch, Astra identified the problematic behavior and made a targeted modification to the underlying numerical structure. But the researchers took this idea much further. Fixing the reasoning loops in an 800-million-parameter model In another experiment, researchers took an open-weight model called Qwen3.5 with approximately 800 million parameters. Their goal was to improve its mathematical reasoning capabilities. However, Astra was prohibited from using traditional training techniques such as gradient descent and backpropagation. It could inspect Qwen’s weights and prompt the model to observe its behavior. And during this process, Astra discovered something interesting. The Qwen model frequently got stuck in reasoning loops. It would repeatedly generate the same thoughts without making meaningful progress toward an answer. Think of it as an AI model getting stuck in an infinite loop. Astra identified problematic patterns in the model’s weights and directly modified them. And the number of weights it needed to change was surprisingly small. Less than 0.1% of the model’s total weights. The results were pretty remarkable. On the MATH-500 benc… Get full access to AsianDadEnergy's Newsletter at asiandadenergy.substack.com/subscribe

  2. 4d ago

    5 Things You Should NEVER Reveal to Anyone?

    One of the interesting side effects of early retirement is that I suddenly have a lot more time on my hands. Besides pursuing various hobbies to keep myself physically and mentally active, I also have considerably more time to think, reflect, and contemplate some of life’s more interesting questions. And one thing I’ve come to realize is that while openness and transparency are generally good things, not everything needs to be shared with everyone. In fact, I would argue that there are certain things in life that are much better kept private. Some information should only be shared with your immediate family or a handful of people you trust completely. And some things are better kept entirely to yourself. Why? Because sharing the wrong information, at the wrong time, or with the wrong people can have surprisingly negative consequences. It can damage your career, strain your relationships, attract scammers, and create problems that might otherwise never have existed. Sometimes, keeping your mouth shut really is the smartest thing you can do. Over the years, I’ve come up with a list of things that I believe people should think twice about before revealing to others. Here are five of them. 1. Your Financial Situation Let’s start with something that I consider particularly important: your personal financial situation. This includes your salary, savings, investments, assets, net worth, and long-term financial plans. My advice is to keep this information on a need-to-know basis. There is very little upside to casually sharing your financial details with other people, but there can be plenty of downsides. For starters, money has a funny way of changing how people perceive and treat you. If someone discovers that you make considerably more money than they do, they might become jealous or resentful. Alternatively, they might develop unrealistic expectations about how much you should spend, what you should pay for, or whether you should lend them money. And if they discover that you’ve accumulated a substantial amount of wealth, you might suddenly find yourself attracting people who are interested in your financial resources rather than your friendship. This brings me to something I’m actually experiencing right now. Through my Asian Dad Energy videos, I’ve occasionally shared bits and pieces about my financial situation and early retirement. I haven’t disclosed that much, but apparently I’ve shared enough to attract a tidal wave of people trying to persuade me to invest in all sorts of questionable speculative schemes. It’s honestly pretty ridiculous. And this illustrates an important point: not everyone who shows an interest in your financial success has your best interests at heart. Some people see your financial resources as an opportunity to benefit themselves. Even among friends and family, money can create unnecessary competition, resentment, and strained relationships. So my advice is simple. Build your wealth quietly. Focus on your own financial goals. Don’t feel pressured to prove your success to anyone. You don’t need to tell people how much you earn, how much you’ve saved, or how much your investments are worth. Your financial information is private. Keep it that way. 2. Your Negative Opinions of Other People Another thing you should think twice about sharing is your negative opinions of other people. This includes how much you dislike your boss, coworkers, neighbors, friends, family members, or even your spouse. Look, we all have people in our lives who annoy us. Some people are incompetent. Some are difficult to deal with. And some are just genuinely unpleasant human beings. But here’s the thing: expressing your negative opinions about these people rarely accomplishes anything useful. On the other hand, it can create a surprising amount of unnecessary trouble. For example, openly criticizing your boss or coworkers can seriously damage your professional reputation and potentially hurt your career. You never know who might repeat your comments or how that information might eventually get back to the people you’re criticizing. I’ve seen this happen quite a few times throughout my career in tech. Smart, sometimes headstrong young software engineers would openly criticize senior leaders. And whether their criticism was justified or not, it almost always ended badly for them. The reality is that being right doesn’t necessarily mean being politically smart. You can have a perfectly legitimate criticism of someone in a position of authority, but broadcasting that opinion to the wrong audience can still come back to bite you. The same principle applies to personal relationships. For example, criticizing someone’s spouse, partner, or family members is particularly dangerous. People tend to be protective of their loved ones. Even if your criticism is justified, they might interpret it as a personal attack. And once you’ve created that resentment, it can be difficult to undo. The same goes for constantly expressing resentment toward people who have wronged you. You end up fueling gossip, negativity, and unnecessary drama. My advice? If someone is making your life miserable, address the problem directly, let go of the resentment, or distance yourself from that person. But you don’t need to broadcast every negative thought that crosses your mind. Sometimes, discretion is worth a lot more than being able to tell everyone exactly what you think. 3. Your Future Plans Here’s another thing I would be careful about sharing: your future plans, especially plans that haven’t materialized yet. This includes your dreams, ambitions, business ideas, plans to quit your job, retire early, or make major changes to your personal life. I think there are several reasons why keeping these plans private can be beneficial. First, other people can be surprisingly effective at discouraging you. You might have an idea that you’re genuinely excited about. But as soon as you share it with someone, they start explaining all the reasons why it won’t work. Sometimes they’re simply being negative. Other times, they genuinely believe they’re giving you helpful advice. Either way, their skepticism can undermine your confidence before you’ve even had a chance to make progress. And that can be enough to derail the entire thing. Second, there’s an interesting psychological phenomenon where talking about your goals can give you a premature sense of accomplishment. You tell people about your ambitious new business idea, your plans to get in shape, or your intention to become financially independent. You receive encouragement and validation. And for a brief moment, you feel as though you’ve accomplished something. But you haven’t actually done anything yet. You’ve simply talked about doing something. In some cases, that initial feeling of satisfaction can reduce your motivation to follow through. Finally, there are practical reasons to keep certain plans private. Suppose you’re planning to quit your job or retire early. If you announce those intentions too soon, your employer might stop investing in your career. You could be excluded from important projects, passed over for promotions, or even find yourself put on a PIP. People may also stop trusting you or relying on you. Why create those problems before you’re actually ready to leave? My advice is to work quietly, get your ducks in a row, and execute your plans before broadcasting them to the world. You don’t need everyone to know what you’re planning. Let the results speak for themselves. 4. Your Deepest Personal Struggles Another thing you should be careful about sharing is your deepest personal struggles, fears, insecurities, and regrets. This includes physical and mental health challenges, past mistakes, personal failures, and things that cause you considerable emotional pain. Now, I want to make an important distinction here. I’m not suggesting that you should bottle everything up, pretend that everything is fine, or never seek help when you’re struggling. Having trusted friends, family members, or qualified professionals… Get full access to AsianDadEnergy's Newsletter at asiandadenergy.substack.com/subscribe

    5 Things You Should NEVER Reveal to Anyone?

  3. 6d ago

    AI Just Broke Reality? Nothing Will Be the Same!

    Now that I’m retired early, I have a lot of free time on my hands. And one of the things I enjoy doing with that time is keeping up with major breakthroughs in computer science, mathematics, and artificial intelligence. Every once in a while, something comes along that makes me stop and think. Something that might actually change how we understand the world. And just recently, we may have witnessed one of those moments. I’m talking about the reported AI disproof of the Exponential Time Hypothesis, one of the fundamental assumptions underlying modern computer science. And here’s what makes this particularly interesting. Both OpenAI and Anthropic reportedly achieved breakthroughs on the same day, just hours apart, using two apparently independent approaches. If their results hold up under scrutiny, this could be an extraordinarily important development. Potentially, we’re talking about something with implications far beyond computer science. We’re talking about the possibility of breaking through one of the most fundamental computational barriers that humanity has struggled with for generations. And yes, that’s a pretty big deal. Let me explain why. 1. The Mathematical Wall That Humanity Thought We Couldn’t Break To understand what’s happening, we first need to talk about something called the Exponential Time Hypothesis, or ETH. Interestingly enough, the origins of this idea go all the way back to the 1950s, to a mathematician named John Nash. If that name sounds familiar, he’s the mathematical genius who developed some of the foundational ideas behind game theory. He also struggled with schizophrenia, and his life was eventually portrayed in the movie A Beautiful Mind. But there’s another lesser-known part of his story. Back in the 1950s, Nash wrote a letter to the American National Security Agency about the difficulty of breaking encrypted messages intercepted from the Soviet Union. His basic observation was pretty straightforward. If you have the decryption key, checking an encrypted message is relatively easy. But if you don’t have the key, figuring out what that key is can be an entirely different story. Nash suspected that cracking sufficiently complex encryption systems would require computational effort that increased exponentially with the length of the encryption key. Think about it this way. Imagine you’re trying to solve a puzzle. Every time you add another piece, the number of possible combinations doubles. So you end up with something like this: * 1 piece: 2 possible combinations. * 2 pieces: 4 possible combinations. * 10 pieces: 1,024 possible combinations. * 40 pieces: Over a trillion possible combinations. * 256 pieces: An astronomically large number of combinations. That’s the problem with exponential growth. It starts off looking manageable, but very quickly becomes completely ridiculous. Even if you have the world’s most powerful supercomputer, there are certain problems where checking every possible combination would take longer than the age of the universe. And this isn’t just some theoretical curiosity. It’s part of the reason modern cryptography works. The idea that certain computational problems become prohibitively expensive as their complexity increases eventually became formalized into what computer scientists call the Exponential Time Hypothesis. At its heart is a mathematical problem called 3-SAT. Don’t worry, I’m not going to torture you with a college mathematics lecture. The basic idea is that you have a collection of logical variables and constraints, and you’re trying to determine whether there’s an assignment of values that satisfies all those constraints. As the number of variables increases, the number of possible combinations grows exponentially. ETH essentially assumes that there is no algorithm capable of solving every possible 3-SAT problem faster than exponential time. In other words, no matter how clever you are, there are certain problems where the computational complexity grows so quickly that you eventually hit a wall. And this is closely connected to one of the biggest unsolved questions in computer science: P versus NP. The basic question is whether problems whose solutions can be verified efficiently can also be solved efficiently. Most computer scientists believe that P does not equal NP. In plain English, they believe that there are problems for which finding a solution is fundamentally much harder than checking whether a proposed solution is correct. That’s the assumption I learned about back in college. And here’s the important part. This isn’t just some obscure mathematical problem that exists to keep university professors employed. The underlying computational difficulty shows up in an enormous range of real-world applications. 2. Why Breaking ETH Could Change Everything Imagine what humanity could accomplish if we suddenly discovered an efficient way to solve problems that currently require unobtainable amounts of computational effort. For starters, think about cryptography. We could potentially discover entirely new ways to break encryption systems that we currently consider secure. Of course, that would also create some rather unpleasant consequences for internet security and financial systems. But let’s put that aside for a moment. What about medicine? Imagine being able to efficiently search through enormous numbers of possible molecular configurations to design proteins that specifically target and treat certain types of cancer. What about logistics? Imagine optimizing global supply chains, airline schedules, and public transportation networks to operate at levels of efficiency that are simply impossible today. What about materials science? Imagine discovering entirely new materials that are lighter than air, more flexible than rubber, and stronger than steel. Materials that could potentially enable structures and technologies that sound like science fiction today. And then there’s artificial intelligence itself. Imagine building an AI system that can efficiently solve problems that currently require enormous amounts of computational resources. An AI that can generate genuinely original, civilization-changing ideas on demand. We’re talking about the possibility of accelerating scientific discovery, engineering, medicine, and technological innovation by orders of magnitude. Perhaps even achieving true artificial superintelligence. Now, obviously, these are not guaranteed consequences of disproving ETH. But they illustrate why the underlying mathematical question matters so much. For generations, researchers have been trying to find better algorithms to solve these problems. We’ve explored everything from clever mathematical techniques to alternative computing architectures, including quantum computers. But even quantum computing comes with substantial limitations, and it isn’t some magical solution to every computational problem. The uncomfortable reality is that we’ve spent decades trying to find ways around this exponential computational barrier. And most researchers believed that, at least in the general case, it was fundamentally unavoidable. Which brings us to the really interesting part. 3. OpenAI and Anthropic May Have Just Found a Way Around the Wall According to their reported research, both OpenAI and Anthropic have independently produced mathematical arguments that challenge the Exponential Time Hypothesis. I took a look at both of their GitHub repositories. And at first glance, they appear to be two genuinely independent approaches. What’s particularly interesting is how both teams seem to have approached the problem. They’re essentially following what I would call the derivative innovation playbook. Rather than inventing an entirely new mathematical framework from scratch, AI agents can leverage existing mathematical knowledge, explore enormous numbers of possibilities, and use computational resources to search for solutions that human researchers might never have discovered on their own. In other words, take decades of accumulated mathematical knowledge, throw an army o… Get full access to AsianDadEnergy's Newsletter at asiandadenergy.substack.com/subscribe

  4. Oct 30

    How to survive a JOBLESS future?

    AI is coming for jobs. Layoffs are everywhere. And the old career playbook is increasingly useless. So what the heck are we supposed to do? On Wednesday, October 28, at 8 PM EST Asian Dad Energy did another livestream about surviving the coming job apocalypse. 💀 We talked AI, layoffs, underemployment, resilience, and how to build a life that doesn't depend on your boss deciding you're still useful.Bring your questions. Bring your existential dread. Let's figure this mess out together. Get full access to AsianDadEnergy's Newsletter at asiandadenergy.substack.com/subscribe

    How to survive a JOBLESS future?

  5. Oct 29

    Anthropic's IPO Just Revealed Something TERRIFYING?

    So Anthropic is preparing for an IPO, and its S-1 registration statement is finally giving investors a closer look at the financial reality behind one of the hottest companies in AI. For those unfamiliar with the process, an S-1 is the registration document a US company generally files with the Securities and Exchange Commission before going public. It lays out the company’s financials, business model, risk factors, and other information prospective investors should know before buying its stock. And with Anthropic reportedly targeting a valuation of around $2 trillion, we’re talking about a truly ridiculous amount of money. Now, I like money. I’m an unemployed former Big Tech software engineer with more than 25 years of experience in the industry. I’ve spent a good chunk of my adult life building software, watching technology companies rise and fall, and developing an increasingly unhealthy level of skepticism toward corporate hype. But I also like the idea of making a few bucks without having to return to corporate meetings that could have been emails. So I decided to put my tech cynicism aside, take a closer look at Anthropic’s financial disclosures, and ask a simple question: Can this company actually grow into a $2 trillion valuation? What I found was fascinating, concerning, and occasionally so absurd that I had to read the numbers twice. Let’s get into it. 1. Anthropic’s Losses Are Much Bigger Than They First Appear Let’s start with the headline numbers. Anthropic reportedly generated $16.3 billion in revenue during the first half of 2026. That’s explosive growth. In just six months, the company generated approximately 3.5 times its revenue for all of 2025. On the surface, that’s the sort of growth investors dream about. The company also reported an operating loss of $693 million. Now, that might not sound particularly alarming for a rapidly growing technology company preparing to go public. Amazon famously operated at a loss for years before becoming a profit-generating machine. Other high-growth technology businesses have also prioritized expansion over short-term profitability. So why should Anthropic be any different? Because the economics of running a frontier AI company are absolutely bonkers. The real cost of running Claude Training and operating advanced AI models requires enormous amounts of computing power. We’re talking about expensive GPUs, massive data centers, electricity, networking infrastructure, and the ongoing costs of serving millions of users. Training a frontier model can cost hundreds of millions of dollars. And that’s before you account for the infrastructure required to make that model available to customers around the clock. Then there’s inference: the computing power required every time someone asks Claude to write code, analyze a document, or explain why their Python script has spontaneously developed an attitude problem. Multiply that by millions of users, and the computing bill becomes enormous. And remember that Anthropic also offers expensive subscription plans, including Claude Max at $200 per month. Depending on usage, serving some customers could be extremely costly. The key question is whether the revenue generated by those customers is sufficient to cover the full economic cost of serving them. But here’s where things get interesting. Anthropic’s reported operating loss doesn’t necessarily capture the full economic burden of acquiring the computing resources it needs to operate. Borrowing compute today, paying with equity tomorrow Anthropic has complex financing arrangements with major cloud providers, including Amazon and Google. These arrangements can involve convertible promissory notes and other instruments tied to future equity. In simple terms, Anthropic can secure computing resources and finance its infrastructure needs through arrangements that don’t require an equivalent cash payment upfront. That can be useful for a fast-growing company. It preserves cash and gives the business access to the infrastructure it needs to expand. But it doesn’t make the underlying resources free. There is still an economic cost, even when the payment doesn’t immediately show up as a conventional cash expense. According to the figures discussed in the filing, Anthropic recorded more than $72 billion in costs associated with these arrangements during the first half of 2026, alongside another $8 billion in other expenses. The exact accounting treatment matters enormously here. Changes in the fair value of equity-linked obligations can create large noncash expenses when a company’s valuation rises. Those expenses are not the same thing as cash burning out of a bank account, and they should not automatically be interpreted as ordinary operating costs. But neither should investors ignore them. Noncash does not mean costless. And financing something does not mean you have stopped paying for it. If the accounting figures are being interpreted correctly, these items would imply an extraordinarily large reported net loss, potentially around $80 billion for the six-month period. That is an eye-watering number. However, I would be careful about declaring it the largest loss in corporate history without comparing the precise accounting line items against other companies’ filings. A large fair-value adjustment is not directly comparable to a company losing $80 billion in cash through ordinary operations. Still, the broader question remains: how much does it really cost Anthropic to build and operate its AI business, and how much of that cost will ultimately have to be borne by shareholders? That’s the question I want answered before I start throwing money at the IPO. And then there’s the long-term infrastructure bill The other number that deserves attention is Anthropic’s reported long-term infrastructure commitments. The filing reportedly describes more than $500 billion in obligations associated with securing AI computing capacity over the coming decade. These commitments are especially important if substantial portions are non-cancellable. Imagine signing a contract that obligates you to pay for a huge amount of computing capacity over the next ten years. Now imagine that, three years from now, your AI models become less competitive, customers migrate to cheaper alternatives, or your revenue growth slows dramatically. You might no longer need all that capacity. Unfortunately, your contractual obligations don’t necessarily disappear just because your business plans have changed. That’s a serious risk. AI infrastructure is a particularly dangerous place to overcommit because the technology is evolving so quickly. The chips, models, and data centers that seem essential today could become economically unattractive much sooner than expected. The danger isn’t simply that Anthropic spends a lot of money. It’s that the company could commit to enormous infrastructure costs based on assumptions about future demand that turn out to be wrong. And that brings us to another problem. 2. Anthropic Is Increasingly Dependent on the Same Big Tech Companies It Wants to Compete Against One of the things that jumped out at me was how intertwined Anthropic’s relationships with major technology companies have become. Amazon and Google aren’t just infrastructure providers. Depending on the relationship, these companies can also be investors, customers, distribution partners, and competitors. Think about the implications. Amazon has its own cloud business and AI ambitions. Google has its own models, infrastructure, and enterprise products. Both companies have reasons to support Anthropic’s growth. A successful Anthropic means more demand for computing resources, more cloud revenue, and potentially greater returns on their investments. But they also have reasons to prioritize their own AI products and businesses. These interests don’t always align. What happens when your supplier is also your competitor? Suppose Anthropic suddenly needs substantially more computing capacity to serve a surge in demand. What happens if its … Get full access to AsianDadEnergy's Newsletter at asiandadenergy.substack.com/subscribe

  6. Oct 26

    The Internet Is BREAKING… And It's Only Getting Worse?

    Guys, I don’t want to sound like a doomsday prepper. But over the past several days, the cybersecurity headlines have been getting increasingly disturbing. Critical software vulnerabilities. Ransomware attacks. Stolen personal information. Industrial control systems being compromised. Attacks targeting organizations responsible for healthcare, energy, manufacturing, and food distribution. It’s starting to feel like the entire internet is one giant dumpster fire. And here’s the part that really worries me. AI is making it dramatically easier to find vulnerabilities that humans have missed for years, sometimes decades. I’ve spent 25 years working in technology. I used to be a software architect at a Big Tech company, and I’ve dealt with major cybersecurity incidents firsthand. So when I tell you that AI could fundamentally change the cybersecurity landscape, I’m not just some random guy yelling at clouds. Well, okay, I am a guy yelling at clouds on the internet. But I have some relevant experience. Let’s talk about what’s happening, why it matters, and what you can do to protect yourself and your family. 1. A 20-Year-Old Security Problem Just Got Uncovered One of the most alarming developments in the recent cybersecurity news cycle involves critical vulnerabilities in the PHP runtime, the software environment that executes PHP code on web servers. PHP is old. It has been around since the 1990s and remains a foundational technology for a huge portion of the web. WordPress, Drupal, Magento, and countless other websites and applications rely on PHP, directly or indirectly. The reported vulnerabilities are especially concerning because they involve remote code execution, commonly abbreviated as RCE. In plain English, remote code execution can allow an attacker to make a vulnerable server execute commands that the attacker controls. Depending on the vulnerability, the configuration, and the server’s security controls, that could potentially give an attacker access to sensitive files, databases, application secrets, or other systems on the same network. Imagine discovering that the front door of your house has had a defective lock for 20 years. Now imagine that millions of houses have the same lock. And nobody noticed until someone finally figured out how to pick it. That’s the sort of problem we’re talking about. The particularly unsettling part is that AI-assisted security research can help researchers examine complicated software and discover bugs that might otherwise remain hidden for years. The same technology that helps developers write code can also help security researchers find ways to break it. Of course, discovering a vulnerability doesn’t automatically mean every PHP website is exploitable. Actual risk depends on the affected versions, configuration, mitigations, and whether administrators have applied the relevant patches. But the broader lesson is difficult to ignore. We have built enormous portions of the internet on top of software that humans have been writing, modifying, and patching for decades. Some of that software is remarkably reliable. Some of it is held together with technical debt, obscure dependencies, and the digital equivalent of duct tape. And now AI is getting much better at finding the weak spots. 2. I’ve Lived Through This Nightmare Before This isn’t entirely new territory for me. Five years ago, just before Thanksgiving, the technology industry was hit by Log4Shell, a critical vulnerability in the widely used Java logging library Log4j. If you worked in enterprise software at the time, you probably remember the chaos. Log4j was embedded in countless applications and products. Some organizations knew exactly where they used it. Others had no idea. And even when you found a vulnerable dependency, fixing it wasn’t always straightforward. I was working as a chief architect at a Big Tech company when the incident hit. Due to a combination of organizational circumstances, I was the only available chief architect in my business unit to coordinate the response. So naturally, I got to help clean up the mess. I called an emergency office-hours meeting involving architects and technical leads from around 100 application teams. One hundred teams. One security incident. One extremely unpleasant holiday season. Our newly appointed business-unit CTO had inherited Log4Shell as his first major crisis. Understandably, he wanted answers, and he was watching our response closely. The mission sounded simple enough: get all the teams to patch their applications. Except enterprise software isn’t a neat collection of perfectly maintained applications with complete documentation and developers who understand every line of code. Some applications were ancient. Their original developers had left years earlier. The current teams barely understood how certain components worked. Sometimes applying the patch broke the application. Other times, the teams had to modify their own application code just to accommodate the patched dependency. And every change needed to be tested, reviewed, deployed, and monitored. It was a nightmare. Now imagine AI helping attackers identify vulnerabilities across enormous codebases while defenders are still trying to determine which systems they own, which libraries they use, and whether a patch will break production. That is the asymmetry that concerns me. AI can accelerate both attack and defense, but the defender still has to deal with the messy reality of existing infrastructure. Finding the bug can take minutes. Fixing a business-critical system safely can take days, weeks, or longer. And that’s assuming someone knows the system exists in the first place. 3. This Is Bigger Than Your Password Getting Stolen The recent cybersecurity headlines aren’t limited to vulnerabilities in web applications. We’ve also seen reports of attacks involving personal information, ransomware campaigns, industrial systems, and organizations responsible for critical services. Consider the different layers of modern society that depend on software. * Healthcare: Hospitals depend on digital systems for scheduling, records, diagnostics, and other essential operations. * Energy: Utilities rely on complex industrial control systems and connected management infrastructure. * Food distribution: Supermarkets depend on refrigeration, inventory systems, logistics, and transportation. * Manufacturing: Factories increasingly rely on networked computers and automated production equipment. * Government: Public services depend on databases, identity systems, communications, and online infrastructure. These systems aren’t all equally vulnerable, and a cyberattack doesn’t automatically translate into a physical disaster. But the potential consequences extend well beyond having to reset your password. A compromised website might expose customer information. A ransomware attack might disrupt hospital operations. An attack on industrial systems might interrupt production or logistics. A sufficiently disruptive incident could affect services that ordinary people depend on every day. And here’s the uncomfortable truth: modern infrastructure is interconnected. The United States has many security controls, segmented networks, backup systems, and specialized protections. Critical infrastructure is not simply one giant computer connected directly to the public internet. But neither is it completely isolated from the digital world. Remote management, cloud services, vendor access, software dependencies, and operational technology create connections that must be secured. There is no magical force field protecting civilization from software bugs. Which brings us to the elephant in the server room. 4. Why AI Makes This Problem Different Software has always contained bugs. That isn’t a controversial statement. It’s a fact of life for anyone who has written enough code. Modern applications can contain millions of lines of code, depend on hundreds or thousands of external libraries, and interact with operating systems, databases, network services, and other appli… Get full access to AsianDadEnergy's Newsletter at asiandadenergy.substack.com/subscribe

  7. Oct 23

    Is This the END of LLMs? NCM Could Change Everything

    NCM: The LLM Killer? Stanford May Have Changed AI Forever Hello world. I’m an unemployed ex-Big Tech software engineer with 25 years of experience in the tech industry. And one of the perks of involuntary early retirement is that I have plenty of time to follow the latest developments in AI. Which is great, because the AI industry seems determined to change the world every five minutes. But here’s something I’ve been thinking about for a while. Almost the entire AI conversation revolves around Large Language Models, or LLMs. Every few months, we get a bigger model, a more expensive model, a new benchmark, and another announcement about how this one is definitely going to change everything. Meanwhile, the companies building these models are pouring staggering amounts of money into GPUs, data centers, electricity, and cooling infrastructure. And yet, today’s LLMs have some fundamental limitations. What if the next major AI breakthrough doesn’t come from making LLMs bigger? What if it comes from building AI differently? Well, researchers at Stanford University have been exploring a new architecture called Next-Concept Modeling, or NCM, and the early results are interesting enough to deserve our attention. Could this be the beginning of the end for LLMs? Maybe. Maybe not. But let’s talk about why this research caught my attention. 1. The Problem With Today’s LLMs First, let’s talk about how conventional LLMs work. At their core, these models predict the next token based on the prompt. A token might be a word, part of a word, or another small unit of text. The model predicts a token, adds it to the sequence, and then predicts the next one. This process continues until the model has generated its response. In grossly simplified terms, it’s a huge autocomplete. But their autoregressive generation process has a major drawback. The tokens in an answer are generated sequentially. Suppose you ask an LLM to generate a 1,024-token response. The model generally needs to perform a forward pass for each newly generated token, meaning roughly 1,024 sequential generation steps. There are optimizations, such as caching previously computed attention states, speculative decoding, and other inference techniques. So this doesn’t mean the entire model is recomputed from scratch every time. Nevertheless, generating long responses requires repeated computation, and model weights and associated data must be accessed as needed throughout inference. With models containing trillions of parameters, that can get expensive very quickly. And if you enable a reasoning mode that generates additional intermediate tokens, that only increases the number of passes and the amount of computation required before you get your final answer. The result is an AI industry that increasingly depends on enormous amounts of computing power, memory bandwidth, and energy. Now compare that with the human brain. The brain performs an astonishing range of cognitive tasks while consuming roughly 20 watts of power. Today’s frontier AI systems can require vastly more energy and computing infrastructure, particularly when training and operating at scale. That comparison isn’t perfectly apples to apples. Biological brains and digital neural networks work very differently, and their capabilities aren’t directly comparable using a single energy metric. Still, the contrast raises an important question. Are we approaching intelligence in the most efficient way possible? I’ve argued before that if we see major algorithmic breakthroughs that move us closer to more capable reasoning systems, they might come from somewhere outside the architectures dominating today’s AI industry. And this is where NCM gets interesting. 2. Meet NCM: A Different Way to Generate Text Next-Concept Modeling takes inspiration from an idea used in another area of AI: diffusion models. If you’ve ever watched an AI image generator create an image, you’ve probably seen how it starts with noise or a rough representation and progressively refines the result until a recognizable image emerges. Instead of generating an image one pixel at a time, diffusion-based approaches refine a representation of the image through multiple steps. NCM applies a related idea to text. Rather than generating an answer one token at a time, the model represents the answer through multiple layers of increasingly detailed concepts. Think of it as creating a rough blueprint of an answer first, then progressively filling in the details. At the highest level, the model represents broad concepts using a relatively small number of codes. The next layer adds more detail. Subsequent layers refine the representation further, until the final layer contains the token-level information needed to reconstruct the text. The key difference is that each refinement step works toward representing the entire answer, rather than simply adding one more token to the end of a sentence. That’s the big architectural idea. Instead of requiring a separate generation step for every output token, NCM can generate an output through a much smaller number of refinement steps. And if you can generate long responses with fewer sequential steps, you may be able to reduce latency and the repeated memory traffic associated with conventional token-by-token generation. Of course, fewer sequential steps don’t automatically mean less total computation. We’ll get to that important caveat shortly. But first, let’s look at the results. 3. The Early Results Are Pretty Wild The Stanford researchers trained several experimental NCM models, with the largest containing approximately 730 million parameters. That’s a relatively small model compared with the frontier systems developed by major AI labs. Nevertheless, their experiments produced some intriguing results. A potential 31x speed improvement In one benchmark involving text summarization, the researchers reported that NCM was approximately 31 times faster than a leading diffusion language model. Notice the comparison here. This was a comparison against another diffusion-based language model, not a conventional autoregressive LLM. That’s an important distinction. We shouldn’t take the result and immediately conclude that NCM is 31 times faster than ChatGPT or Claude. Still, it suggests that the NCM architecture may offer substantial speed advantages over existing approaches to parallel text generation. And that alone makes it worth investigating. The memory bandwidth advantage Let’s consider why this architecture could matter for inference efficiency. Imagine two hypothetical models of equal size, each with a 1 GB set of model weights. Suppose one is a conventional autoregressive LLM generating 1,024 tokens, while the other is an NCM model that generates the same-length output in 32 refinement steps. For a simplified comparison, assume each step requires reading the entire 1 GB set of weights from memory. The conventional LLM would require approximately 1,024 GB of cumulative weight reads. The NCM model would require approximately 32 GB. That’s a theoretical difference of 32 times less weight traffic under these assumptions. Actual memory traffic depends on caching, hardware, batching, model architecture, and implementation details. And this simplified example doesn’t account for all of NCM’s additional computational requirements. But it illustrates why reducing the number of sequential model evaluations could be valuable. Memory bandwidth is a major bottleneck in LLM inference. GPUs can perform enormous numbers of mathematical operations, but moving data between memory and processing units takes time and energy. If NCM can reduce that bottleneck substantially, it could change the economics of running AI models. And that brings us to the experiment that really caught my eye. The Sudoku experiment: 86 versus 1 The researchers also evaluated reasoning performance using Sudoku puzzles. They compared a 25-million-parameter NCM model against a conventional LLM with the same number of parameters. Both models were evaluated on 100 hard Sudoku board… Get full access to AsianDadEnergy's Newsletter at asiandadenergy.substack.com/subscribe

  8. Oct 20

    AI Just KILLED the Software Industry? One Guy Cloned Adobe!

    Let me tell you about a development that should make anyone working in the software industry a little nervous. Imagine spending decades building a software company worth billions of dollars. You hire thousands of engineers, invest enormous amounts of money in research and development, and build a portfolio of applications used by millions of people. Then one guy comes along, fires up an AI coding agent, and starts cloning your products. Welcome to the future of software. One Guy vs. Adobe and Microsoft Recently, a developer named Brandon Thomas attracted attention with ArtCraft, a project aimed at recreating the functionality of popular Adobe Creative Cloud applications using AI-assisted reverse engineering. The project includes applications inspired by products such as Photoshop, Illustrator, and Premiere Pro. Related projects extend the same idea to familiar office applications, including Word, Excel, and PowerPoint. These are not polished, feature-for-feature replacements for every original application. But that’s almost beside the point. The important development is that AI is making it dramatically easier to reproduce software functionality without rebuilding everything from scratch through traditional software development. I tried PhotoCraft myself. It wasn’t nearly as polished as Photoshop, but it worked for many of my basic image-editing needs. I also experimented with AI-assisted reverse engineering on a simple chess application. The resulting clone came surprisingly close to the original. And here’s the part that really caught my attention: I didn’t have to manually analyze every feature and write the replacement code myself. I could give an AI coding agent the task and let it do much of the heavy lifting. As a former Big Tech software engineer and architect, I find this fascinating. And potentially terrifying. For years, I’ve argued that AI is making software less valuable. These developments suggest that the process could accelerate much faster than many people expect. How Can AI Reverse-Engineer Software Without the Source Code? To understand why this matters, we need to look at how software reverse engineering works. When developers build an application, they write source code using programming languages such as C++, Java, or Rust. That code is typically compiled into machine code, which contains instructions a computer can execute. The resulting executable doesn’t normally include the original source code in its original, readable form. Traditionally, reverse engineering software required specialized expertise. Engineers used tools to inspect executable files, analyze program behavior, recover information about functions and data structures, and reconstruct how different features worked. This was difficult, time-consuming, and expensive. AI coding agents are changing that equation. An AI agent can work through an application feature by feature, using specialized reverse-engineering tools to gather information about the program’s behavior and internal structure. Tools such as Ghidra and IDA can help analyze compiled binaries. An AI agent can use the resulting information to reason about how a feature works, implement a replacement, test the result, and repeat the process. The basic workflow looks something like this: * Analyze: Inspect the original application and identify its features. * Investigate: Use reverse-engineering tools to examine relevant parts of the executable. * Reconstruct: Infer how a feature behaves from the available evidence. * Reimplement: Write new code that reproduces the feature’s behavior. * Test and repeat: Compare the replacement with the original and work through the remaining gaps. Repeat this process across enough features, and you may end up with a functional alternative to the original application. There are important limitations. Compiled binaries don’t necessarily reveal everything about the original source code, and some information may be missing or difficult to recover. Complex applications also have countless edge cases, integrations, and usability details that are hard to reproduce. A working clone is not necessarily a complete replacement. But the economic implications don’t require a perfect clone. If an AI-assisted developer can reproduce 80% of the functionality that most customers actually use, that might be enough to compete with the original product for a substantial portion of its market. The precise percentage is hypothetical. The underlying point is that customers don’t necessarily need every feature. They need the features they care about. And AI is lowering the cost of delivering them. The Real Problem: Software Is Becoming Easier to Copy Think about why software companies can charge so much for their products. Developing sophisticated software requires enormous investments in engineering, design, testing, infrastructure, and maintenance. Once a company has made those investments, it can distribute copies of its software to additional customers at relatively low marginal cost. That combination has historically been extremely profitable. Companies spend millions developing a product, then sell subscriptions to thousands or millions of customers. The original development cost is substantial, but each additional customer doesn’t require the company to rebuild the product from scratch. Now imagine AI dramatically reduces the cost of creating competing products. A developer no longer necessarily needs a large team to build a useful alternative. An AI agent can help implement interfaces, recreate common features, generate tests, and analyze existing applications. Reverse engineering adds another possibility: instead of starting with a blank screen and trying to figure out what customers want, developers can study an existing product and use its behavior as a reference. This changes the economics of competition. Historically, software companies benefited from the difficulty of reproducing their products. That difficulty created a barrier to entry. AI is attacking that barrier. And when barriers to entry fall, competition tends to increase. The result could be a world in which many software features that once justified expensive subscriptions become available through cheap or free alternatives. Adobe’s subscription business is an obvious example. But the implications extend to productivity software, project management tools, design applications, developer utilities, and countless niche software products. The question is no longer simply whether AI can help us build software faster. It’s whether AI can make existing software products much easier to reproduce. Are Small Software Developers Completely Screwed? I think smaller software developers could face some of the biggest challenges. Imagine spending three years building a niche application. You’ve identified a problem, developed a solution, found customers, and finally established a recurring revenue stream. Then someone uses AI to reproduce the core functionality of your product in a fraction of the time. What’s your competitive advantage? If your application consists primarily of features that can be observed and recreated, your moat may be much thinner than you thought. This doesn’t mean every small software company is doomed. Specialized expertise, customer relationships, proprietary data, integrations, reliability, and brand reputation can all provide meaningful advantages. But a product’s features alone may become a much weaker defense. And here’s the particularly uncomfortable part: AI doesn’t have to eliminate your entire business to hurt you. It only needs to make it easier for competitors to offer enough of your functionality at a lower price. Even if your product remains superior, you may find it harder to justify charging a premium. That’s a problem for any business whose value proposition depends on customers having few practical alternatives. Will Everything Become Open Source? My prediction is that we’ll see an explosion of open-source alternatives to proprietary software. Imagine a world where popula… Get full access to AsianDadEnergy's Newsletter at asiandadenergy.substack.com/subscribe

    AI Just KILLED the Software Industry? One Guy Cloned Adobe!
See All (69)

Ratings & Reviews

5
out of 5
2 Ratings

About

This is a very public journal of anxiety, existential dread, and way too much tech knowledge. Basically therapy, but with Wi-Fi. asiandadenergy.substack.com

Information

  • Creator
    Ivy-League educated, Ex Big Tech, Middle aged Asian Dad figuring out life.
  • Years Active
    2026
  • Episodes
    69
  • Rating
    Clean
  • Copyright
    © AsianDadEnergy
  • Show Website
    AsianDadEnergy's Podcast

You Might Also Like