YouTube Digest

English

Based on "Your life needs more ‘existential grit.’ Here’s how to find it | Kate Bowler" from Big Think Watch the original video

Beyond Bliss: Why Your Life Needs the Gritty Grace of Joy

By [Your Name/Staff Writer – or omit if preferred]

We live in a culture obsessed with happiness. From self-help gurus to social media feeds, the message is clear: strive for an unblemished state of contentment, where every moment is a picture of well-being, flourishing, and perhaps, "kids in matching denim." But what if this relentless pursuit of happiness is actually preventing us from experiencing something far more profound, something that can carry us through life's deepest valleys and still make us laugh?

According to Kate Bowler, a historian, podcaster, and writer who studies luck, meaning, and what makes life beautiful, our understanding of positive emotions is deeply flawed. Bowler, known for her candid explorations of faith and suffering, argues that while you can't be happy and sad at the same time, you absolutely can be joyful and sad simultaneously. This distinction, she contends, is where we find true "existential grit."

When Happiness Isn't Enough

Bowler's insights are not purely academic; they are forged in the crucible of personal experience. She describes a period of immense fortune: finally having a longed-for baby after years of infertility, landing her dream job as a historian at Duke University, and being married to her high school sweetheart. Life, she felt, was finally "paying off."

Then, a devastating season of "powerful unluck" struck. Diagnosed with stage four cancer after a long, arduous battle to receive proper care, Bowler found that happiness, as she knew it, simply wasn't enough anymore. "I was just never gonna be able to, like, add up my life in the same way," she recounts. "So I needed something more existentially gritty. I needed joy."

This pivotal realization led her on a quest to understand an emotion capable of carrying a person through a terrible time and into a new season.

The Fragility of Happiness vs. The Resilience of Joy

The confusion between happiness and joy is common, Bowler observes, largely because both reside on the "positive end of the spectrum." Americans, in particular, are bombarded by an enormous "happiness industry" that equates well-being with a psychologically relaxed, easy state – a cumulative result of things going well.

Psychologists, however, draw a critical distinction. Happiness, Bowler explains, is a definable set of qualities: ease, comfort, a state that is deeply contextual. "It happens because things have gone well," she says. "That's what makes happiness lovely, but also makes it very fragile. One thing goes wrong and you can topple the whole mood of happiness."

Joy, by contrast, operates on an entirely different plane. It's not about escaping reality or denying pain; it's about integrating it. "Joy engages not just the dopamine and the, like, happier parts of our brain chemistry," Bowler explains. "It still coexists with our stress systems, with really dark emotions, with despair, with enormous pain, meaning that we're not escaping the reality of what we're in. We're adding another layer on top of it."

It's this capacity to bind together broken pieces while still making you laugh that makes joy so fascinating and, ultimately, so powerful.

The Existential Yes

Joy is not merely a fleeting sweetness; it is one of the most transformative human emotions. It makes us more grateful, more hopeful, and it delights us, often eliciting laughter even in the direst circumstances. Bowler describes it as "this great existential yes that reminds us that life is still worth loving, even in the midst of the worst times."

This profound affirmation allows us to perceive reality with new eyes, to look at the totality of our experience – the good, the bad, the beautiful, the ugly – and still feel, somehow, that it is good.

For many, particularly within the Christian tradition, joy holds a special place as a divine gift, a moment of transcendence. Psychologically, it fosters a deep sense of bonding and connection – to other people, to the divine, to the very fabric of existence. It's an emotional openness that reconciles seemingly irreconcilable experiences.

Cultivating Existential Grit

If joy is this potent, life-sustaining force, how do we find it? Bowler suggests that the preconditions for joy are often the direct opposite of what our "machine selves" demand.

Two key elements emerge repeatedly: emotional availability and connection. In a world that often encourages us to shield ourselves from vulnerability, opening up to our emotions and forging genuine connections with others are crucial.

Another powerful precondition is a willingness to be surprised. "I think a person who wants to be surprised is a person who's much more likely to find joy," Bowler notes. This implies a conscious effort to step away from the predictable, to put down our phones and close our laptops, and to cultivate a sense of "experiential surprisability."

Ultimately, happiness is an emotional state, a fleeting mood dependent on external circumstances. Joy, however, is a story. "It is a feeling that somehow in your spirit that it feels good to be alive, to be here, to be put together with whatever you have left," Bowler articulates. It can lift you from the very bottom to the very top, and at its core, "mostly what it takes is love."

Joy Is For Everyone

Perhaps the most surprising and liberating aspect of Bowler's message is that joy is not a "bonus level of happiness" reserved for the fortunate or the perfectly content. Quite the opposite.

"I think it would be a huge relief to the person struggling with depression, to the person right in the middle of fresh grief, to the person who is just frankly, deeply bored by their life that joy is definitely for them," Bowler asserts.

This is the true power of existential grit: a joy that is both emotional and existential, tangible and profound. It is real, it is raw, and it is the very thing that can carry us through life's inevitable storms, reminding us that even when broken, life is still worth loving, still worth living, and still capable of surprising us with its luminous grace.

Based on "The Rise of the Dictater" from Every Watch the original video

Unleash Your Voice: The Rise of the 'Dictater' in the AI Age

For generations, the keyboard has been our primary interface with the digital world. From crafting emails to coding complex algorithms, our fingers have danced across QWERTY layouts, translating thought into text. But what if there was a more natural, more efficient way to communicate with our machines? What if the future of productivity lay not in our fingertips, but in the power of our voice?

Enter the "dictater." It may sound like a term pulled from a dystopian novel, but in the context of modern productivity, it refers to someone who dictates their thoughts, ideas, and commands rather than typing them. And surprisingly, these "dictaters" are poised to revolutionize how we work, driven by a powerful confluence of human physiology and artificial intelligence.

The Silent Struggle of the Keyboard

Think about it: how often do your fingers struggle to keep pace with your thoughts? The human brain processes ideas at an astonishing speed, but our typing speed is often a bottleneck. We might formulate a perfect sentence in our mind, only to have the rhythm broken by a typo, a backspace, or the sheer physical effort of striking keys.

Beyond the speed disparity, there's the physical toll. Long hours spent hunched over a keyboard can lead to finger cramps, wrist strain, and even repetitive strain injuries. There's also the subtle, yet pervasive, "input lag" – the minuscule delay between thought and its digital manifestation that, over time, can disrupt flow and creativity. The keyboard, for all its utility, is an artificial barrier between our minds and our output.

The Natural Flow of Voice

This is where the "dictater" finds their advantage. We can speak significantly faster than we can type. This isn't just about raw words per minute; it's about the ability to convey ideas as naturally as they form in our heads. There's an immediacy to spoken language that typing simply can't replicate. When you dictate, your thoughts flow unimpeded, mirroring the organic process of human conversation.

Imagine brainstorming a complex project, drafting a detailed report, or even composing creative prose. Instead of pausing to type, you simply speak, allowing your ideas to cascade into existence. As one early adopter puts it, "Hey. Yeah, that's exactly what I meant." The clarity and precision achieved through direct voice input often surpass what can be painstakingly typed out, leading to more authentic and impactful communication.

AI: The Catalyst for the Voice Revolution

Historically, dictation had its limitations. Early voice recognition software was clunky, prone to errors, and required significant manual correction. This is precisely why the keyboard maintained its dominance. However, the landscape is rapidly changing, thanks to incredible advancements in Artificial Intelligence.

Modern AI tools are no longer just transcribing words; they are understanding context, learning individual speech patterns, and even adapting output formats. Think of an AI assistant that can:

This new generation of AI-powered tools removes the friction that once plagued dictation. They turn your voice into a powerful, versatile input method that adapts to you, rather than forcing you to adapt to the limitations of a machine.

The Future is Spoken

As these AI capabilities continue to improve at an exponential rate, we are poised to witness a significant rise of "dictaters" around the world. Professionals across every discipline – writers, programmers, doctors, educators, business leaders – will increasingly choose monologue as their primary mode of interaction with technology.

Platforms like 'Every' exemplify this shift. Described as "the only subscription you need to stay at the edge of AI," it learns, transcribes, and translates across different disciplines and languages, adjusting its output format to match your context. This means you can dictate a complex medical report one moment, then switch to brainstorming marketing copy the next, all while maintaining an uninterrupted flow of thought.

The "dictater" isn't about control in the traditional sense; it's about regaining control over your time, your ideas, and your creative output. It's about empowering individuals to communicate with their devices in the most natural and efficient way possible: through their own voice.

The keyboard will likely remain a tool in our arsenal, but its reign as the undisputed king of input is drawing to a close. The future of productivity is clear, articulate, and spoken aloud. Are you ready to unleash your inner "dictater"?

Based on "Build a team of AI Agents to run your business" from Greg Isenberg Watch the original video

Beyond the Chat: Building Your AI-Powered Business Department

The world of Artificial Intelligence can often feel like a dizzying maze of acronyms and complex concepts. From LLMs to agent harnesses, the jargon alone is enough to deter many from exploring its truly transformative potential. Yet, beneath the surface of this complexity lies a profound shift in how we can interact with and leverage AI – a shift from simple chat interactions to powerful, autonomous AI agents capable of running entire departments of your business.

Greg Isenberg, host of a popular podcast, recently brought on his friend and AI expert, Remy Gasill, to demystify this evolving landscape. Remy, who has structured his own company around these digital workers, believes that while most people are still "chatting" with AI, the real productivity gains lie in harnessing AI agents. "The AI landscape is moving into stage two from chat to agents," Remy explains, "and the founders and employees that are utilizing agents are, no word of a lie, 10 to 20 times more productive in their day." This isn't just about marginal improvements; it's about exponential growth that can leave competitors miles behind.

The core message is clear: understanding and implementing AI agents today is not just an advantage, but a necessity for future success.

From Conversation to Command: Chat Models vs. AI Agents

The first step to unlocking this new era of productivity is to understand the fundamental difference between a chat model and an AI agent. The term "AI agent" is often thrown around loosely, losing its true meaning. Remy offers a crystal-clear distinction:

Imagine wanting to build a website. With a chat model, you'd ask, "How do I build a website?" It would give you steps. You'd then have to follow those steps. With an agent, you'd say, "Build me a minimalist portfolio site for Greg Isenberg," and the agent would go away and do it.

The Engine of Autonomy: Understanding the Agent Loop

What enables an AI agent to move from a simple response to a complete result? The answer lies in what's called the Agent Loop. This iterative process allows the agent to continuously work towards its goal without constant human intervention. The loop consists of three core steps:

  1. Observe: The agent takes in new information. This could be your initial prompt, files in its workspace, or feedback from its previous actions.
  2. Think: Based on its observations and its overall goal, the agent strategizes. It asks itself: "What's the next logical step?" "What information do I need?" "What tools should I use?"
  3. Act: The agent performs an action based on its thought process. This could be conducting research, writing code, drafting an email, or interacting with an external tool.

This loop repeats endlessly until the agent determines that the task is complete, based on parameters you set in your initial prompt.

A Real-Time Demo: Building a Website

Remy illustrates this with the "build a minimalist portfolio site for Greg Eisenberg" example.

This iterative process, constantly observing, thinking, and acting, is what allows agents to tackle complex, multi-step tasks that would overwhelm a simple chat model.

The Anatomy of an AI Agent

Every AI agent, regardless of its specific platform, is built upon four fundamental components:

  1. The LLM (Large Language Model): This is the "brain" of the agent, the underlying intelligence that processes information and generates responses. Think of models like Claude Opus, GPT-4, or Gemini.
  2. The Loop: As described above, this is the continuous "observe, think, act" cycle that drives the agent towards its goal without stopping after a single response. It's the mechanism that transforms ping-pong into a marathon.
  3. Tools: These are the external applications and services the agent can connect to and interact with. This is where AI agents become truly powerful, integrating with your email, calendar, project management software, and more.
  4. Context: This is all the information about you, your business, your preferences, and your processes that the agent needs to operate effectively.

A platform that facilitates this entire process – essentially housing the LLM, managing the loop, and providing the framework for connecting tools and context – is known as an Agent Harness. Popular examples include Claude Code, Codeex, Anti-Gravity, OpenClaw, and Manis.

Agent Harnesses: Your AI's Dashboard

Think of agent harnesses like different cars. "What we're going to learn today is we're going to learn to drive," Remy analogizes. "Once you know how to drive, you can kind of jump in any car, whether it's like an old Toyota, a Range Rover, and you inherently sort of know what to do." Agent harnesses are those cars. Some might have better features like "seat warmers and cruise control," but the core functionality remains the same once you understand the underlying concepts.

These harnesses provide the interface for you to define tasks, connect tools, and manage the agent's workflow. They often work with local files on your computer, making your AI setup portable and future-proof.

Security Considerations: When connecting agents to your business tools, security is paramount. Remy advises scoping access carefully. While major platforms are built with security in mind, it's crucial to control what permissions an agent has. For instance, giving it read-only access to sensitive platforms reduces risk. As Remy notes, platforms like OpenClaw are "the wild west" in comparison, requiring more user vigilance.

Onboarding Your Digital Employee: The Art of Context Engineering

Just as you wouldn't expect a new human executive assistant to perform effectively without proper onboarding, your AI agent needs to be trained on your business, preferences, and processes. This is where Context Engineering comes into play, marking a significant shift from the old focus on "prompt engineering."

1. The agents.md File: Your Agent's System Prompt

Unlike chat models that secretly build a cloud-based memory of your conversations, AI agents start with a blank slate in each new session. To give them a foundational understanding, you create an agents.md (or claude.md, gemini.md depending on the harness) file within your agent's local folder.

This markdown file acts as a persistent system prompt, loaded every time you start a new session. It contains:

With this file in place, a simple prompt like "Write me a cold email" transforms from a generic request into a highly contextualized one. The agent immediately knows your business, your target audience, and your preferred tone, allowing it to ask intelligent follow-up questions (e.g., "Is it for a brand or sponsor, potential partner, or consulting client?") and generate a much more relevant draft.

If you have extensive context, you can even create a separate context folder with multiple markdown files (e.g., brand_voice.md, ideal_customer_profile.md). Then, in your main agents.md file, you instruct the agent to "read my context folder to understand about myself and my business," effectively chaining your knowledge base. Some users even link it to their Obsidian vaults for a "second brain" approach.

2. The memory.md File: The Self-Improving Loop

Even with a robust agents.md file, a new challenge arises: remembering intricate details and preferences across sessions. If you tell your agent, "My favorite color is lavender," it might acknowledge it but forget it in the next session. This is problematic for continuous improvement, especially for things like email sign-offs or specific formatting preferences.

To solve this, Remy introduces the memory.md file. By adding a simple instruction to your agents.md file – "When I correct you or you learn something new, update the relevant section in memory.md" – you create a self-improving loop.

The memory.md file acts as a dynamic record of all your preferences, corrections, and learned behaviors. If you tell your head of sales agent, "Never sign off emails with 'cheers,' use 'warm regards' instead," it will not only apply that preference but also update memory.md. The next day, it will automatically recall this preference.

This compounding effect means that over weeks and months, your agents become increasingly attuned to your specific needs, reducing errors and becoming genuinely indispensable. While some agent harnesses are starting to build in automatic memory systems, understanding the manual setup highlights the underlying mechanism.

A valid concern is whether memory.md files can become too large and unwieldy. Remy advises keeping agents.md files concise (around 200 lines) and suggests that for memory, you can refine the instruction to "only save substantial corrections" if it starts logging trivial details.

Connecting the Dots: The Power of MCP

The true magic of AI agents comes alive when they can interact with your existing tools. This is facilitated by the Model Context Protocol (MCP).

Before MCP, connecting an LLM to different tools was like trying to get people speaking different languages to communicate directly. "Claude speaks English, Notion speaks Spanish, Gmail French, your browser Japanese, and Slack Chinese," Remy explains, referencing an analogy from Ross Mike. While custom solutions existed, they were extensive and time-consuming.

Anthropic (the creators of Claude) developed MCP as a universal translator. MCP sits between your agent and your tools, speaking every language. Your agent can simply "speak English" (its native LLM language), and MCP translates its commands into the specific API calls for Gmail, Notion, Stripe, or any other connected tool, and then translates the tool's response back to the agent. This standardized protocol makes connecting tools remarkably easy within agent harnesses.

Most agent harnesses offer a simple "connectors" or "integrations" section where you can link your Gmail, Google Calendar, Notion, Stripe, Granola (for meeting notes), and hundreds of other apps using MCP.

The AI Operating System: A New Way of Working

Remy envisions a future where everyone has an AIOS (AI Operating System) – a central hub where they interact with their personal and departmental AI agents. This AIOS, built upon local markdown files and integrated tools, becomes your single point of truth and action.

"I don't even enter these tools anymore," Remy states, referring to Gmail, Google Drive, Calendar, Notion, and Stripe. "I just sit in Claude Code as one central place."

Demo: The Integrated Executive Assistant

To demonstrate the power of this integration, Remy walks through a common business workflow:

  1. Task: "Summarize my inbox from today." The agent, connected to Gmail via MCP, quickly reviews and highlights key emails.
  2. Follow-up Task: "Okay, great. Review my meeting notes with Maltoshi from today, then draft up the email sending the proposal and creating the Stripe payment link, and then go into Notion and set up the project."

Here's how the agent, leveraging its loop, context, and tools, executes this complex task:

This entire process, which would typically involve switching between multiple tabs, copying and pasting information, and manually setting things up, is executed by the agent in minutes. "Even if you can just do something like seven times faster without having to go into all these tools, copy the meeting notes into the page to give it context on your meeting, it really starts to compound," Remy emphasizes. "Then you start to fit like a week in a day and then 7 weeks in a week."

Scaling Productivity: The Power of Skills (SOPs for AI)

The final layer of AI agent mastery lies in Skills. Remy describes skills as "SOPs for AI – Standard Operating Procedures for AI. It means once you explain something once, you never have to explain it ever again."

Imagine you frequently create client proposals. Without skills, you might spend 15-30 minutes iterating with your agent, correcting formatting, adjusting pricing placement, and refining the language. Even with memory.md, these detailed, task-specific preferences can clutter the memory or be forgotten if you switch sessions.

A Skill packages this entire process into a dedicated markdown file. This skill file meticulously explains the exact steps, formatting requirements, and preferences for creating a proposal. Once created, every time you need a new proposal, you simply invoke the "create proposal" skill, and the agent executes it perfectly, consistently, and without further instruction.

Skills are distinct from memory.md because they are about encapsulating processes, not just preferences. They are specialized instruction sets for specific jobs to be done. Most agent harnesses now have a "skills" feature, often stored in hidden local folders (e.g., .claude/skills). You can create skills by manually writing the markdown file or by having a "skill creator" agent (a common feature in harnesses) interview you about a process and generate the skill file.

By automating three to five small manual processes each week with skills, you gradually build a comprehensive library of automated workflows, inching closer to automating your entire work life.

The 100x Employee: The Future of Work

The vision painted by Remy Gasill, and echoed by figures like Cody Schneider, is one where every employee comes into their role equipped with a pre-existing AI operating system. They will continuously build out skills for their manual processes, accumulating automated workflows until their entire work life is streamlined.

This isn't about replacing human workers, but augmenting them into "100x employees" – individuals capable of achieving unprecedented levels of productivity and output by offloading repetitive tasks to their digital counterparts. The transition from simple chat models to sophisticated, context-aware, tool-integrated, and self-improving AI agents represents not just an evolution in technology, but a revolution in how we work and build businesses. The time to build your AI-powered department is now.

Based on "Why Scale Will Not Solve AGI | Vishal Misra - The a16z Show" from a16z Watch the original video

Beyond the Matrix: Why Scaling Won't Deliver True AGI

The latest generation of large language models (LLMs) like Anthropic's Claude and Google's Gemini are nothing short of astonishing. They write code, generate compelling text, and even pass complex exams, leading many to believe that Artificial General Intelligence (AGI) is just around the corner, achievable simply by making these models larger and training them on more data. But what if that belief is fundamentally flawed? What if these powerful systems, despite their brilliance, are merely "grains of silicon doing matrix multiplication," lacking the very essence of consciousness, inner monologue, or genuine causal understanding?

This provocative stance is at the heart of computer scientist Vishal Misra's groundbreaking work, as explored in a recent episode of the a16z Show. Misra, a professor at Columbia University, argues that while LLMs are incredibly adept at a specific type of learning—Bayesian updating—they are inherently limited by their architecture and objective function. To reach true AGI, he contends, we need a paradigm shift, moving beyond mere correlation to embrace plasticity and causality.

From Cricket Stats to the Matrix of Minds

Misra's journey into understanding LLMs began five years ago, fueled by a personal challenge. Granted early access to GPT-3, he sought to build a natural language interface for a massive cricket statistics database he had co-created years prior. He successfully employed a technique now widely known as Retrieval Augmented Generation (RAG), getting GPT-3 to translate natural language queries into a custom Domain Specific Language (DSL) that it had never seen before. The system, deployed at ESPN in 2021 (after an initial working prototype in October 2020), was "mind-blowing" in its effectiveness.

Yet, this success left Misra with a burning question: how did it work? Diving into the "attention is all you need" papers and other deep learning architectures, he found no satisfying answer. This led him to develop a mathematical model, conceptualizing an LLM as a "huge gigantic matrix."

Imagine a matrix where every row corresponds to a unique prompt – "protein," "protein shake," "the cat sat on the..." – essentially every possible combination of tokens within the model's context window. Each column, then, represents a probability distribution over the model's entire vocabulary (around 50,000 tokens for models like GPT). Given a prompt, the LLM samples the next token from this posterior distribution.

This matrix, Misra explains, is astronomically large – "more than the number of electrons across all galaxies" – but also incredibly sparse. Most combinations of tokens are gibberish, and for any given prompt, only a few next tokens make sense. What LLMs effectively do, in Misra's view, is create a compressed representation of this sparse matrix, approximating the true distribution for any given prompt.

In-Context Learning: A Bayesian Revelation

This matrix abstraction provided a crucial lens for understanding in-context learning, the seemingly magical ability of LLMs to "learn" from a few examples provided in the prompt itself. Misra demonstrated this with his cricket DSL. Initially, when shown a cricket question, GPT-3 would predict English words. But as he presented pairs of natural language queries and their corresponding DSL translations, the model's "posterior probability" for DSL tokens would steadily rise. By the time a new, related query was given, the model would confidently generate the correct DSL, having updated its "belief" in real-time.

"This is an example of in real time the model was updating its posterior probability," Misra clarifies. "It was upgrading its knowledge that okay, I've seen evidence, this is what I'm supposed to do." He concluded that LLMs were performing a form of Bayesian updating – starting with a prior belief, seeing new evidence, and updating their posterior belief.

Initially, this claim met with skepticism from parts of the machine learning community, often caught in the historical "Bayesian vs. Frequentist" debate. Critics argued that Misra's observations were merely empirical, not a rigorous mathematical proof.

The Bayesian Wind Tunnel: A Formal Proof

To unequivocally prove his hypothesis, Misra, along with colleagues Naman Agarwal and Siddharth Dalal, developed what they called the "Bayesian Wind Tunnel." Just as an aerospace engineer tests an aircraft in an isolated environment, they created a controlled setting for LLM architectures.

"We took a blank architecture," Misra explains, "and gave it a task where it's impossible for the architecture to memorize what the solution to that task should be." They used very small models, ensuring the task was difficult enough to prevent memorization but tractable enough that the precise Bayesian posterior could be calculated analytically.

The results were stunning: the transformer architecture "got the precise Bayesian posterior down to 10^-3 bits accuracy. It was matching the distribution perfectly." Mamba models also performed well, while LSTMs and MLPs faltered. This experiment provided the mathematical proof that transformers, at their core, are indeed performing Bayesian inference. Further papers in the series explored why they do it (analyzing gradients and geometry) and showed these same "signatures" persisted in large, frontier models.

The Chasm: Why LLMs Aren't Human (Yet)

Despite their impressive Bayesian capabilities, Misra draws a stark line between LLMs and human intelligence, identifying fundamental differences that scaling alone cannot overcome:

  1. Plasticity vs. Frozen Weights: Human brains are incredibly plastic; our synapses constantly adapt and update throughout our lives based on new experiences. LLMs, however, have "frozen" weights once training is complete. During inference (like in-context learning), they perform Bayesian inference, but they "forget" everything from one conversation to the next. "Every invocation of it was fresh," Misra notes of his cricket DSL system. "It did not remember the last time I sent a query what the DSL looked."

  2. Objective Functions: Humans are driven by primal objectives: "don't die and reproduce." Our brains have evolved to simulate dangers and react to preserve ourselves. LLMs, conversely, have a singular objective: "predict the next token as accurately as possible." This objective is entirely a function of their training data. Misra dismisses fears of LLMs "deceiving" or "trying to survive," attributing such behaviors to the models reflecting narratives found in their training data (e.g., Reddit stories about AI survival). "They don't have consciousness. They don't have an inner monologue. They're not driven by the same objective function."

  3. Correlation vs. Causation: This is perhaps the most profound difference. Deep learning excels at association – finding correlations in vast datasets (what Misra equates to Shannon entropy). However, humans possess the ability for causation, which involves building causal models, performing interventions, and reasoning about counterfactuals (Judea Pearl's causal hierarchy). This, Misra links to Kolmogorov complexity – the length of the shortest program that can reproduce a given string. While the Shannon entropy of Pi's digits is infinite, its Kolmogorov complexity is very small because a short program can generate it. LLMs are stuck in the Shannon entropy world; they haven't crossed over to Kolmogorov complexity or true causal understanding.

The Einstein Test: A High Bar for AGI

To illustrate this causal gap, Misra proposes the "Einstein Test" for AGI: "You take an LLM and train it on pre-1916 or 1911 physics and see if it can come up with the theory of relativity. If it does, then we have AGI."

At the time of Einstein, many anomalies challenged Newtonian mechanics: Mercury's orbit, the Michelson-Morley experiments, and early ideas about black holes. An LLM, adept at finding correlations, would see all this "evidence" and treat the discrepancies as "anomalies." But it would not, Misra argues, "come up with the beautiful equation that Einstein came up with."

Einstein's genius wasn't in correlating existing data; it was in creating a new representation – a new "manifold" of spacetime that fundamentally changed the axioms. "If you just stuck with the old manifold of the Newtonian physics," Misra explains, "then you would see these correlations but you could not come up with a manifold that explained them. So you need to come up with a new representation."

This concept was recently highlighted by Donald Knuth's viral experience using LLMs to solve a problem involving Hamiltonian cycles. While the LLMs were incredibly efficient at finding connections and solutions within the existing mathematical manifold (the "Shannon part"), it was Knuth, the human, who ultimately had to synthesize these findings into a new proof, effectively creating the "new manifold." The LLMs got "stuck" at a certain point; they couldn't generate the novel, underlying causal model.

The Path Forward: Plasticity and Causality

For Misra, the path to AGI lies not in simply scaling up current architectures, but in tackling these two fundamental problems:

  1. Plasticity (Continual Learning): Developing architectures that can learn new information without "catastrophic forgetting" of previous knowledge. This requires a mechanism for dynamic weight updates that mimics the human brain's lifelong adaptability.

  2. Causality (Moving from Correlation to Intervention): Building models that can understand causal relationships, perform simulations, and reason about "what if" scenarios. This involves moving beyond association to the higher levels of Judea Pearl's causal hierarchy.

"Scale will not solve everything," Misra asserts. "You need a different kind of architecture." LLMs, he believes, are "definitely part of the solution," but they are not the entire solution. They represent a powerful tool for correlation, for navigating existing manifolds of knowledge. But to truly generate new knowledge, to conceive of new representations of reality, and to exhibit genuine intelligence, a fundamental shift in our approach to AI architecture is required. The next frontier in AI, Misra concludes, lies in cracking the code of plasticity and causality, moving beyond the matrix of correlations to build machines that can truly understand, adapt, and innovate.

Based on "Chosen by War: The Rise of Iran’s New Supreme Leader" from New York Times Podcasts Watch the original video

Game of Thrones in Tehran: The Unforeseen Rise of Iran's New Supreme Leader

The news broke with a defiant roar: Iran's new Supreme Leader, Mushtaba Hami, the son of the recently assassinated ruler, had spoken. His first statements left no room for doubt or hope of de-escalation. "Avenging the blood of your martyrs is a top priority," he declared, promising continued attacks on Gulf Arab neighbors and demanding the immediate closure and eventual assault on all U.S. bases in the region. He even affirmed the closure of the Strait of Hormuz. This hardened stance, coming in the midst of a devastating war with the United States and Israel, signaled a new, potentially more perilous, era for Iran and the Middle East.

But the ascent of Mushtaba Hami was far from a straightforward transition. It was, as described by New York Times reporter Farnaz Fassi, an "extraordinary behind-the-scenes jockeying," a veritable "Game of Thrones" that defied the very spirit of the Islamic Revolution and ultimately saw Iran, driven by the pressures of war, replace one hardline leader with another.

A Revolution's Irony: Power from Father to Son

Mushtaba Hami’s selection came as a surprise even to seasoned Iran watchers and insiders. While he had worked closely in his father’s office, his succession was not considered predestined. This outcome carries a profound irony: the 1979 Islamic Revolution was born from a movement to dismantle thousands of years of monarchy, explicitly rejecting the transfer of power from father to son. Even after the death of Ayatollah Ruhollah Khomeini, the revolution's founding father, the leadership passed not to his son, but to the then-president, Ayatollah Ali Khamenei. For Mushtaba, the son of the late Supreme Leader, to now inherit the mantle, is, by many accounts, a direct violation of the revolution's core anti-monarchical ethos.

The process that led to this unexpected outcome unfolded over several intense days, described by sources in Iran as a "succession war." Various political factions, powerful generals from the Revolutionary Guards, influential figures like the former spy chief, and senior clerics engaged in a fierce struggle, each vying to install their preferred candidate in the nation's most critical and, arguably, most dangerous job.

The Factions at Play: Moderates vs. Hardliners

Constitutionally, the responsibility of appointing, supervising, and removing a Supreme Leader falls to the Assembly of Experts, a body of 88 elected senior clerics. However, the real battle took place through back-channeling and intense lobbying, as powerful factions sought to sway the clerics' votes.

On one side were the moderates and pragmatics, a camp led by figures such as the head of the National Security Council, Ali Larijani, and President Pezeshkan. Their argument was rooted in the extraordinary circumstances facing Iran: a country at war with the United States and Israel, and grappling with months of internal upheaval, including massive street protests demanding an end to the regime. The moderates believed this power vacuum, created by the assassination, presented an opportunity to steer the country in a new direction. They advocated for a "new face" for the regime, a candidate who would signal to both the international community and the Iranian public a willingness to moderate policies or pursue reforms.

Among their preferred candidates were:

These candidates, to varying degrees, represented a desire to "turn the page" on the hardline revolutionary ethos that had dominated Iran since 1979. While still loyal to the Islamic Republic's ideology, they offered a pragmatic approach, potentially open to scaling down hostilities with the United States.

However, standing in staunch opposition were the hardline factions, particularly the powerful Revolutionary Guards Corps. For them, wartime was not a moment for concessions or surrender to U.S. demands. Their priority was to ensure the continuity of the policies and strategies defined by the late Supreme Leader and the Guards themselves. The Revolutionary Guards, a dominant force in Iran's political, economic, and military spheres, viewed the selection of the Supreme Leader as an existential choice, crucial for maintaining their grip on power, especially as they commanded the ongoing war.

This hardline pool unanimously backed Mushtaba Hami. They saw him as a close ally, a reincarnation of his father who would continue his policies and grant the Guards a free hand, particularly in prosecuting the war. For them, the constitutional formality of rejecting hereditary power was secondary to the immediate needs of wartime and the desire for defiance. They argued that the assassination of their leader, whom they considered a martyr, necessitated the closest possible successor—his son, who not only bore a physical resemblance but also shared his ideological beliefs and policies.

The Dramatic Showdown: A Will Revealed

The process of selecting the Supreme Leader began almost immediately after the late leader was killed in air strikes on the first day of the war. Fueled by a sense of defiance, the Assembly of Experts, after back-channeling and influence peddling, initially chose Mushtaba as the frontrunner. A vote was cast, and the government was informed of the decision, with plans to announce it on state television the next morning.

But the plan unraveled. As news of Mushtaba's emergence as a leading candidate leaked (a New York Times scoop), both U.S. President Donald Trump and Israel's Defense Minister issued threats to "eliminate the next successor." This external pressure caused the Iranians to pause, fearing that an immediate announcement could endanger Mushtaba's life.

This pause provided a critical window for the moderates to launch an offensive. They saw an opportunity to convince the Assembly to rescind its vote. If the U.S. and Israel were threatening a hardline successor, didn't that strengthen the case for moderation?

The moderates convened a meeting with the Assembly's leadership council and presented a bombshell. They brought two of the late leader's closest aides – his chief of staff and a top senior military advisor – to testify. These trusted confidantes claimed that the late leader had explicitly stated he did not want his son to succeed him. As if that weren't enough, they then produced a sealed letter, allegedly his will, which, upon being unsealed, reportedly stated, "I don't want any of my family members to become the Supreme Leader." This was a direct challenge to Mushtaba's legitimacy, presented as the "word of God" from the revered former leader himself.

The hardliners and Revolutionary Guards generals, hearing of this audacious counter-offensive, mobilized swiftly. Generals and figures like Hussein Taib, the former intelligence chief of the Guards, personally called Assembly members, urging them to meet virtually for an emergency vote. Their aim was to solidify Mushtaba's position before the moderates could gain further ground. On Sunday, March 8th, the Assembly held a final vote, and Mushtaba Hami secured the two-thirds majority he needed. This time, the deal was sealed, and the announcement made.

The consensus among experts is clear: had Iran not been at war, had its leader not been killed by air strikes, Mushtaba's path to power would have faced insurmountable resistance. The unique circumstances of conflict and perceived martyrdom paved his way.

Who is Mushtaba Hami? A Figure of Shadows and Power

Mushtaba Hami remains a mysterious figure, having always operated in the shadows of power. Since his appointment, he has maintained a public silence, issuing only two written statements, with no public appearances or speeches. Yet, insights from sources in Iran who know him or have met him offer a glimpse into the man now leading the nation.

Born in 1969, Mushtaba was just nine years old when the Islamic Revolution established the theocracy. He grew up immersed in the ideological and religious fervor of the revolution's early years, witnessing its institutionalization from an abstract idea to a day-to-day government. His participation went beyond observation. At 17, he volunteered as a soldier in the bloody 8-year Iran-Iraq War, fighting alongside many who now form the senior leadership of Iran's military. This battlefield experience earned him "street cred" and forged powerful alliances, making him a veteran in a way the son of a powerful leader might not have needed to be.

After the war, he moved to Qom, a center of Shia seminaries, to study and become a Shia jurist and cleric. He climbed the ranks of the religious hierarchy, teaching advanced Islamic jurisprudence, a level of religious instruction only accessible to advanced clerics. His classes in Qom were reportedly popular, suggesting a certain charisma.

Eventually, he moved to Tehran and entered his father's close political circle, managing security and military administrative matters within his office. He forged particularly strong alliances with figures like Hussein Taib, the former Revolutionary Guards intelligence chief, and General Mohammad Bagher Ghalibaf, a powerful Guards commander who is now the speaker of parliament. According to sources, these three would meet weekly to strategize on policies, from electoral outcomes to crackdowns on dissidents and other state security matters.

A chilling example of his behind-the-scenes influence, pointed out by many sources, is his alleged role in the 2009 presidential election. When former President Mahmoud Ahmadinejad was declared the victor amid widespread accusations of a rigged election, sparking the "Green Movement" protests, Mushtaba is said to have played a role in the alleged rigging and in orchestrating the brutal crackdowns that followed, particularly through his ties to the Basij paramilitary militia. This history solidified his image as "their guy" for the hardliners. Indeed, Mushtaba's rise mirrors the Revolutionary Guards' increasing control over Iran's political, military, and economic landscape.

His continued public silence since becoming Supreme Leader is baffling to many. According to reporting, there are two primary reasons: first, he is reportedly injured, at least in his legs, and may not be in top physical condition for public appearances. Second, and perhaps more critically, Iranians are keenly aware that he is likely "number one on Israel's target list," and fear that a public video could allow for geo-location and assassination.

A Hardline Future: Revenge and Defiance

All evidence points to Mushtaba Hami being a deeply entrenched hardliner, unlikely to pivot towards reform. His life journey – from his military service and crackdowns on demonstrators to the assassination of his father by the U.S. and Israel – collectively suggests a leader driven by a strong sense of revolutionary ideology and a desire for vengeance.

His initial written statements confirm this. He explicitly stated that Iran's military forces would continue to strike at regional countries aiding the American military and affirmed that all his father's political and military appointments and directives would remain in place. This signals a clear intention to continue the previous leader's policies and wartime strategy.

While some supporters attempt to portray him as a "Muhammad bin Salman figure" – a progressive who, despite his hardline appearance, might be the only one capable of de-escalating hostilities with the U.S. and convincing hardliners of a ceasefire – there is no evidence to support this. All available information points in the opposite direction.

The outcome presents a profound irony: many in Iran believe that the United States and Israel, by seeking to eliminate and overthrow the regime's leadership, have inadvertently given Iran precisely the leader they sought to avoid. Their actions may have solidified the power of an even more hardline figure.

For the 80% of Iranians who, at the outset of the war, harbored a fleeting hope for change and an opening up of their country, Mushtaba Hami's elevation is a devastating blow. That moment of hope, once real, has been brief. As Iran enters the third week of war, with no signs of policy change or major concessions, Iranians find themselves under relentless bombing and air strikes. Hope has been replaced by fear and anxiety, with nightly reports of louder, closer explosions echoing across the nation. The "chosen by war" leader now presides over a country gripped by uncertainty, its future inextricably linked to a path of defiance and confrontation.