<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Alex Wang on Medium]]></title>
        <description><![CDATA[Stories by Alex Wang on Medium]]></description>
        <link>https://medium.com/@wwang776?source=rss-b1afd815a204------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*PvrHdjxARjZCOespMgO-TQ.jpeg</url>
            <title>Stories by Alex Wang on Medium</title>
            <link>https://medium.com/@wwang776?source=rss-b1afd815a204------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Mon, 16 Nov 2026 18:15:38 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@wwang776/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[The Last Invention: What AGI Will Look Like]]></title>
            <link>https://medium.com/@wwang776/the-last-invention-what-agi-will-look-like-bb4c66795c8d?source=rss-b1afd815a204------2</link>
            <guid isPermaLink="false">https://medium.com/p/bb4c66795c8d</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[future]]></category>
            <category><![CDATA[philosophy]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Alex Wang]]></dc:creator>
            <pubDate>Sun, 15 Mar 2026 04:58:13 GMT</pubDate>
            <atom:updated>2026-03-15T04:58:13.640Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*D8SBqaXtPez6cziaffONdw.png" /></figure><p>If you ask a tech CEO what Artificial General Intelligence (AGI) is, they will likely give you a vague, academic answer: <em>“It is an AI that performs at the level of a human with perfect knowledge of their domain.”</em> The more visionary folks would even say that AGI will be able to perform anything that can be conceived of.</p><p>But what does this mean, exactly? Does it mean it can write emails? We have that. Does it mean it can pass the Bar Exam? We have that, too. Yet none of us feel like we are living in the presence of a new intelligent species.</p><p>AGI will not be a simple librarian; it is a completely intelligent operator. It doesn’t need to have read a manual on “How to escape a locked room using a gum wrapper.” It understands the <em>physics</em> of the gum wrapper, the <em>logic</em> of the lock, and the <em>goal</em> of escaping. The point of this is not to show that ChatGPT cannot do this, but to illustrate that AGI will understand every constraint and perform optimally within that constraint.</p><p>To understand the future, we need to define AGI through proof by example. For true AGI, imagine the bounds of its capabilities as anything within human’s <em>pragmatic imagination</em>; given the tools, AGI should be able to do anything that a human can physically and mentally do or achieve, but scale that to infinity (or, to the limitations of computing power).</p><p><strong>An Axiom of the True Definition: Solving Any Problem</strong></p><p>The simplest component of AGI isn’t about passing a test. It’s about being an entity<strong> <em>that can understand and operate completely within its constraints to solve any problem it has or has not seen before.</em></strong></p><p><strong>Illustrating our Axiom of AGI: The Forest</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*enQhi_8tMNzDmSc7wj_OBQ.png" /></figure><p>Imagine you are lost in a dense forest. Your GPS is dead, no internet connection. The modern frontier model (i.e., Gemini, Claude, ChatGPT) tries to help and signals some facts about how to get out. You can try this on your own: input an image of a forest with metadata removed (the easiest way would be to screenshot the image) into your frontier model of choice. It gives some cursory facts about your situation, what you can do like “follow the water, head west, etc.”, and some stress-relieving advice. Helpful to an extent, but not a solution or of substance.</p><p>In this scenario, if you had what is truly AGI available to use, it would understand the goal: <em>Survival.</em> Current frontier models in our above paragraph recognize that it will never be able to help you completely, and so it acts more as a supportive friend, whereas the AGI knows that it is capable to solve for your situation. It asks for more details around the surrounding, parsing for context. It recognizes the moss growing on the north side of the trees. It references its internal knowledge of stars or topography relative to where civilization most certainly will be. Once it reaches a state of certainty, it should be able to simulate the complete path out. To further help you, it could even output you a video of the entire route out, as if someone had pre-recorded that video from your exact same circumstance.</p><p>What this would mean is two-fold: it has both the intelligence, which is a combination of knowledge and reasoning/logic, and capability to achieve anything, including navigation out of an unknown area of the forest. Let’s discuss this in parts.</p><p><strong>The Intelligence of AGI</strong></p><p>Intelligence is already extremely difficult to define. Formulaically, one can distill this into <em>knowledge,</em> which is the quantity of information an entity possesses, and <em>reasoning/logic</em>, which is the capability of solving and solutioning based on patterns, ideas, and of course, past knowledge and experiences rather than pure memorization. It’s rather recursive, but that is how it should be; one cannot gain greater intelligence without both components.</p><p>Regardless, in this case, the AGI has the knowledge; knowledge of the world in its physical form (i.e., satellite imagery, maps, star maps, seasonal data, geological and topological data, etc.) Going back to the earlier bounds of AGI as <em>pragmatic imagination</em>, a human <em>could </em>have knowledge of all these things, but it’s unlikely that a single person has all. Mankind as a whole including all data from technology, however, will, and therefore, AGI will also contain the culmination of all human knowledge.</p><p>The part, which is more in debate today and that every frontier lab is trying to prove, is that their AI has reasoning and logic as opposed to trained regurgitation of “memorized” information. The short answer is that it is neither true human reasoning nor simple regurgitation. Frontier models operate in a fascinating middle ground: they are highly advanced pattern-matching engines capable of generalizing concepts to solve novel problems, which can look incredibly close to reasoning. To illustrate what AGI will have, we will take the negatives of what modern frontier models do, and we should imagine that AGI will NOT do these things.</p><p>To start, humans understand the physical and logical rules of the world through experience, which emerges intuition, which allows for broader applicability. We know that if you drop a glass on concrete, it shatters, because we understand physics intuitively. This in turn informs our intuition, and we understand that “glass will shatter, or when applied the equivalent force of dropping (i.e., throwing), on hard things”. Modern frontier models only “know” this because the statistical relationship between the words “drop,” “glass,” “concrete,” and “shatter” is very strong in its training data. It won’t intuitively understand that glass will shatter on other hard surfaces, or anything that it has not seen before.</p><p>Furthermore, these frontier models do not “think” before they speak. They calculate. There is no internal, conscious space where the model deliberates, weighs options, or applies formal deductive logic unless it is specifically prompted to simulate that process. This is also why models can pass the bar exam but sometimes fail spectacularly at a simple logic puzzle or a basic arithmetic word problem. If the mathematical pattern of the language tricks the model’s probabilities, it will confidently output a completely illogical answer (a hallucination). At best, modern AI are simulating reasoning and logic. AGI will have the true form available, or at least, something so close that we humans cannot, and will never be able to, tell the difference (for example, having infinite knowledge on everything will allow it to achieve “reasoning” through regurgitation).</p><p><strong>The Capability of AGI</strong></p><p>After philosophical discussions, there will always follow whether there are practical implementations. AGI should have all manner of pathways to implement its solution, given its constraints, e.g., in our example of forest navigation, the constraint is that the solution can only be software-based. It cannot build a helicopter for you to fly out with.</p><p>Continuing with our example, we have previously discussed the knowledge of the AGI. In this case, after prompting you for more (visual) information of your situation, it has pieced together exactly where you are, using its knowledge of the world. Now, it could just give you directions out, telling you to take a left turn at rock number 6, and then heading straight through the willow tree. That would not be too practical, but it can work. However, to demonstrate its true capability and understanding that we, as humans, operate better with visuals, it has multiple avenues to implement the solution in the most digestible way.</p><p>First demonstration could be using its vast knowledge and capabilities to create a video. It generates frame by frame, using its knowledge of the area you were lost in, and its logic to piece together what it would look like at that time of year. It knows that rock number 6 is in 100 meters, and it understands that the willow tree is orange during the autumn. Then, the AGI can generate a video that looks just like as if you searched up “how to get out of the forest from my exact location” on YouTube. And it would be completely correct in navigating you out of the forest.</p><p>Another demonstration of its capability would be to use that same representation as the video but instead it is an AR/VR application for you to navigate out of. Imagine you hold your phone up, camera on, and there will be generated arrows pointing you towards the way out like a video game. Given enough access to your device, the AGI will even be able to detect how far you’ve gone based on the physical inputs onto your phone, such as your gyroscope, built in accelerometer (available in all iPhones at the time of this article), and other inputs in future phones.</p><p><strong>Conclusion</strong></p><p>AGI, the culmination of all human intelligence and “perfected” human reasoning capabilities, will blur the lines between humanity and an ascended species. Within the constraints it is in, it will be able to operate at a true optimum in every situation. Think of all the times you got “lucky” in a situation and played it perfectly; that will be how AGI conducts itself no matter the circumstance. It will do what a human pragmatically imagines.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bb4c66795c8d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The “Big AI” Era is Over. Here is What the Near-Future Holds]]></title>
            <link>https://medium.com/@wwang776/the-big-ai-era-is-over-here-is-what-the-near-future-holds-3643c30f7e17?source=rss-b1afd815a204------2</link>
            <guid isPermaLink="false">https://medium.com/p/3643c30f7e17</guid>
            <category><![CDATA[future]]></category>
            <category><![CDATA[emerging-technology]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[genai]]></category>
            <category><![CDATA[llm]]></category>
            <dc:creator><![CDATA[Alex Wang]]></dc:creator>
            <pubDate>Fri, 30 Jan 2026 03:22:25 GMT</pubDate>
            <atom:updated>2026-01-30T03:22:25.484Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="Big AI vs. the Future of AI. Image created by Gemini" src="https://cdn-images-1.medium.com/max/1024/1*dtZ0SIyRc2GGNMFIaJhHQQ.png" /><figcaption><em>Big AI vs. the Commoditized Future of AI. Image created by Gemini</em></figcaption></figure><p>For the last five years, the AI industry has been obsessed with a single, rather brute-force strategy: Scaling. Basically, the bigger the better. We were sold a narrative that if we just fed more data and more chips into the “Transformer” architecture (the engine behind ChatGPT), it would eventually be categorized as the coveted “AGI”.</p><p>But in 2026, that strategy has started to lose steam. The giant models are too expensive to run, they still “hallucinate,” and they are too massive to leave the cloud. If you are a business leader or a regular user waiting for a “bigger” frontier model to help you solve your problems, you are looking in the wrong direction. The “Godfather of Scaling” himself has told us to stop.</p><p><strong>The Verdict: Why the “Age of Scaling” is Dead</strong></p><p>In late 2025, the industry witnessed a massive pivot from <strong>Ilya Sutskever</strong> (co-founder of OpenAI and the primary architect of the “Scaling Hypothesis”). After spending a decade proving that “bigger is better,” he reversed course.</p><p>In a landmark interview, Sutskever famously declared: <em>“The 2010s were the age of scaling; now we’re back in the age of wonder and discovery once again.”</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*LJrOaorxYhsjheXH" /><figcaption>Co-founder and ex-CTO of OpenAI Ilya Sutskever claiming that the age of scaling is over</figcaption></figure><p>His reasoning is the business case against the old models:</p><ul><li>Based on the “10,000-Hour” Problem, current consumer frontier models learn rather inefficiently. Think of the student who memorizes every answer key for 10,000 hours but still doesn’t understand the <em>principles</em> of the subject.</li><li>We have already fed the models the entire internet, creating a massive “Data Wall”; data is the blood that fuels the engine that is AI. If the entire internet has been fed into modern AI, where will data come from next? To get smarter using the old method, we would need 100,000x more data to scale, which simply doesn’t exist.</li><li>Efficiency is being gated in modern AI. Running it at scale to perform inference would be incredibly expensive and will only get more expensive as scale ramps up.</li></ul><p>His new venture, <em>Safe Superintelligence (SSI)</em>, is betting that the future isn’t about building a bigger cluster but finding a “new physics” of learning that is efficient and insightful.</p><p><strong>The Hardware Reality: Why “Big” Can’t, and <em>Shouldn’t, </em>Fit in Your Pocket</strong></p><p>Even if we <em>could</em> make the models bigger, we have nowhere to put them. To understand why the future is “Small, Optimized, and Specialized,” you must look at the physical limitations of the device in your pocket.</p><p>Think of an AI model as furniture, and your phone’s memory (RAM) as a small apartment, with limited room for said furniture.</p><ul><li>The “Frontier” Model (e.g., GPT-5, Gemini 3): The soon-to-be classic AI. This is a massive, wall-to-wall sectional sofa. It requires <strong>250GB+ of RAM</strong> to turn on. It will never fit in your phone (which has ~12GB RAM). It must live in a warehouse (data center), and you must rent access to it. Even with this access, you still need to go to the warehouse to even use it. The benefit of such size is that it is generalizable. However, aside from research purposes, ask yourself this: “is this necessary?”</li><li>The “Bleeding Edge” Model (e.g., Mamba 3B): This is a folding chair. It runs comfortably on just <strong>2GB to 4GB of RAM</strong>. You can fit it in your apartment easily. In fact, you can own ten of them, maybe of different shapes, styles and colors. To redecorate your room, you can fold them up and swap them out as needed.</li></ul><p>So what? This physical constraint is the single most important factor for the future of enterprise software and consumer usage. It means that privacy, autonomy, and cost are physics problems. As long as you rely on massive 405B parameter models, sensitive data must leave the building to be processed in a cloud warehouse you don’t control. Efficient (which translates to even more broader application and usage due to cost constraints) models can be used to perform much more complex tasks at a fraction of the cost, with more control and less waste. The kicker? The performance for these tasks will actually be better than the supermassive models of today.</p><p><strong>The New Intelligence: Three Models Slated to Replace the Giants</strong></p><p>Guided by Sutskever’s “Age of Discovery,” let’s discuss some architectures that act like various furniture types that are all small, efficient, and purpose-built. Remember, this is only the beginning, with much more to come with their own benefits and drawbacks.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nsio_H2YsDgcqBjDPw8khg.png" /><figcaption>How each novel AI operates. Note that JEPA is one application of World Models. <em>Image created by Gemini</em></figcaption></figure><p><strong>Mamba &amp; SSMs: The “Privacy” Engine</strong></p><ul><li>The Old: Modern massive models get slower and more expensive the more they read and memorize (i.e., going from 0b to 100b is exponentially easier and faster than going from 100b to 200b).</li><li>The New: Researchers at Carnegie Mellon and Princeton developed “Mamba” to process data linearly. It digests data as fast as you can feed it.</li><li>The Difference: A tiny <strong>Mamba-3B</strong> model can often match the performance of older models more than twice its size. This allows you to, for example, process thousands of pages of private legal or financial documents <em>locally and more efficiently </em>on a laptop, without that data ever leaving your building.</li></ul><p><strong>JEPA: The “Common Sense” Engine</strong></p><ul><li>The Old: Generative models try to predict every pixel or word, which leads to “hallucinations” (lying) when they get confused.</li><li>The New: Championed by Meta’s ex-chief AI scientist Yann LeCun, JEPA (Joint-Embedding Predictive Architecture) stops trying to predict details and starts predicting <em>concepts</em>. JEPA is one application within the family of <em>World Models</em>.</li><li>The Difference: Reliability. A <strong>VL-JEPA</strong> model can understand a video feed using a fraction of the computing power of standard models, while being far less likely to invent (lying) objects that aren’t there.</li></ul><p>And more broadly, it is important to understand that <strong>JEPA</strong> belongs to a class of new model paradigms called <strong>World Models. </strong>Think of a <strong>World Model</strong> as the <em>goal; </em>an AI that can simulate the future to plan complex actions. Think of <strong>JEPA</strong> as the <em>blueprint</em> for building it. Because predicting every single pixel of the future is computationally impossible (and unnecessary), JEPA solves this by teaching the World Model to ignore the visual noise (pixels) and predict the underlying reality (concepts). Let’s discuss World Models below.</p><p><strong>World Models: The “Agent” Engine</strong></p><ul><li>The Old: Chatbots can talk, but they can’t <em>do</em>. They don’t understand cause-and-effect. LLM wrappers might try, but they are guessing at best with no guarantee of certainty. This is why modern agents have such high failure rates, almost all caused by hallucinations; no amount of prompt engineering can fix an inherent problem, nor is it worth the time and constant monitoring. At best, modern “agents” are for workflow automation, not for reasoning.</li><li>The New: World Models simulate futures. They run an internal physics engine to ask, <em>“If I click this button, what happens next?”, </em>allowing the AI to truly reason rather than blindly guess.</li><li>The Difference: This turns AI from a chatbot into a true form of Agent. It allows a small model to understand and navigate a website and book a flight, rather than just writing a blanket itinerary or constantly feeding (expensive) context.</li></ul><p>It is crucial to remember that Mamba, JEPA, and World Models are simply the <em>current</em> answers to the efficiency problem, not the final ones. We are in the early days of the ‘Age of Discovery,’ and the landscape is shifting under our feet. What looks like a breakthrough today, such as a 3GB model running on a phone, may be considered bloated in two years. The transition away from the ‘Giant Transformer’ isn’t just a one-time upgrade; it is the beginning of a new evolutionary branch where models will continue to get smaller, faster, and more alien to us.</p><p>So, this may sound highly technical and daunting. That’s because it is. However, this is no different from when the first iPhone came out; we all had to become tech and app literate to keep up with the changing mobile environment. We all had to learn what “gestures” were, what downloading and installing applications were, how to type on a touchscreen. However, what the iPhone did well was make it the simplest it’s ever been to use frontier technology. These new models are merely an extension of that; when these models become further commercialized it is not required to understand the technical details, the same way we don’t need to understand the backend code of applications on our phones either. We merely need to know what each one does and how we can apply it the most optimal way for our businesses and personal uses in the future.</p><p><strong>References</strong></p><p><strong>On The Pivot from Scaling:</strong></p><ul><li><em>Sutskever, I. (2025).</em> “The Age of Scaling is Over.” Interview with Dwarkesh Patel / Reuters. <a href="https://www.youtube.com/watch?v=aR20FWCCjAs">YouTube</a></li></ul><p><strong>On Efficient Sequence Modeling (Mamba):</strong></p><ul><li><em>Gu, A., &amp; Dao, T. (2023).</em> “Mamba: Linear-Time Sequence Modeling with Selective State Spaces.” <a href="https://arxiv.org/abs/2312.00752">arXiv:2312.00752</a></li></ul><p><strong>On Grounded “Concept” Prediction:</strong></p><ul><li><em>Assran, M., et al. (2023).</em> “Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture (I-JEPA).” <a href="https://arxiv.org/abs/2301.08243">arXiv:2301.08243</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3643c30f7e17" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>