In this conversation, Nir Zicherman and Dan Shipper explore the fundamental disconnect between the vast accessibility of information via LLMs and the persistent struggle individuals face in achieving genuine understanding and sustained learning. The core thesis is that while LLMs excel at answering questions, they fundamentally lack the pedagogical scaffolding and motivational support necessary for deep, long-term learning. This reveals the hidden consequence that the very ease of access offered by current AI tools can inadvertently foster a superficial engagement, leading to projects fizzling out due to a lack of structured guidance and sustained motivation. Anyone seeking to build or leverage AI for effective learning, whether as an educator, a product developer, or an individual autodidact, will gain an advantage by understanding these system dynamics. This conversation highlights how focusing solely on information retrieval, rather than the process of learning itself, creates a system ripe for demotivation and incomplete knowledge acquisition.

Why Your AI Learning Projects Keep Fizzling Out

The advent of large language models (LLMs) like ChatGPT has presented a seemingly utopian landscape for self-education. The ability to delve into any subject, from quantum physics to ancient philosophy, with unprecedented ease has led many to believe that achieving mastery in any domain is now within reach. Yet, for many, these well-intentioned autodidactic pursuits inevitably peter out, leaving a trail of unfinished courses and unfulfilled ambitions. This disconnect isn't a personal failing; it's a systemic one. In a conversation with Dan Shipper on AI & I, Nir Zicherman, CEO and co-founder of Oboe, a platform that crafts personalized AI-driven courses, argues that the obvious answer--more information--is insufficient. The real challenge lies not in accessing knowledge, but in the often-overlooked mechanics of how humans truly learn and retain it.

Zicherman posits a counterintuitive idea: learning is often a more passive process than we assume, and the current AI tools, built as general-purpose information engines, are missing crucial elements of a true learning platform. The apparent ease of LLMs can mask a deeper problem: they answer questions but fail to notice when a learner is lost or to re-engage them when motivation wanes. This conversation reveals that the system of information access provided by LLMs, while powerful, is fundamentally misaligned with the human need for structured guidance, sustained engagement, and a sense of achievable progress. Understanding these system dynamics offers a critical advantage to anyone looking to harness AI for genuine, lasting learning, rather than succumbing to the cycle of superficial engagement and eventual disinterest.

The Illusion of Effortless Mastery

The promise of AI in education is immense, yet the reality for many is a frustrating cycle of starting strong and fading fast. Zicherman, drawing from his experience building Oboe, an AI platform designed for personalized course creation, contends that the very nature of LLMs as general-purpose tools creates a fundamental gap when applied to the complex process of learning.

"LLMs are built to be a general tool," Zicherman explains. "But given that they were not built as learning tools at their core, they're missing a lot." This isn't to diminish the power of LLMs; Zicherman himself uses them extensively. Instead, he highlights that their design prioritizes broad applicability over the specific, nuanced requirements of effective pedagogy. The consequence of this is that while LLMs can provide answers, they cannot replicate the supportive, guiding role of a human educator.

Why ChatGPT Isn't Your Ideal Teacher

Dan Shipper raises the pertinent question: "Why isn't ChatGPT just like the ideal way to learn anything?" Zicherman's response points to the core difference between an information retrieval system and a learning platform. LLMs are "information compression machines," distilling the vastness of the internet into digestible, conversational responses. However, true learning, especially for complex subjects, requires more than just compressed information. It demands commitment, follow-through, and a structured journey.

The downstream effect of relying solely on LLMs for deep learning is that they "lose context very quickly," "don't stay focused," and "it's really easy for them to deviate from the path." This leads to a frustrating experience where the learner, despite having access to immense knowledge, struggles to maintain momentum or achieve their objective. Zicherman emphasizes that while LLMs can be one tool in a learner's arsenal, they are insufficient as the primary method, especially for pursuits requiring sustained effort.

The Passive Nature of Deep Learning

One of the most striking insights from Zicherman is his assertion that "most learning is passive." This runs counter to the intuitive notion that learning is an active, engaged process. He clarifies that he is distinguishing between the intentionality of learning and the method of learning. While high intent and objective-driven motivation are crucial for initiating learning, the actual process of absorbing information often involves passive consumption.

"If you think back to school and you think about the best teachers that you had," Zicherman elaborates, "most of the time that you were learning you were learning from them teaching to you... you were sitting and consuming them talking or consuming the reading or whatever it was." Active participation, while valuable for reinforcement, is not the primary mechanism for initial knowledge acquisition. LLMs, by placing the onus on the user to constantly provide explicit direction and feedback, disrupt this natural, often passive, flow. Unlike a teacher who adapts their curriculum based on student comprehension, LLMs require the learner to be the curriculum designer, a task that can be overwhelming and demotivating.

Building a Scaffold for Sustained Learning

The failure of LLMs as standalone learning tools stems from their inability to provide the necessary scaffolding and motivational hooks that are essential for long-term engagement. Oboe aims to fill this gap by being built "as a learning platform," with a deliberate focus on pedagogical principles.

The Oboe Approach: Structure, Multimodality, and Achievability

During a live demonstration, Zicherman showcases Oboe's ability to generate a course on Wittgenstein's Philosophical Investigations. The platform's design anticipates the challenges of deep learning:

• Speed and Immediate Gratification: Oboe prioritizes delivering content quickly, offering an introduction and initial sections while the rest is generated in parallel. This combats the "attachment anxiety" of waiting for a complete product and provides an immediate sense of progress. "We did not want to be one of those products that requires you to submit something and then wait around for a while in order to get started," Zicherman notes. This immediate payoff is crucial for maintaining initial motivation.
• Multimodality: Recognizing that people learn through various formats, Oboe integrates different types of content. For philosophical topics, this might involve text and discussion, while for STEM subjects, it includes visuals and interactive elements. The generated podcast format, featuring two hosts discussing the material, further enhances engagement by offering a different modality of consumption. This approach acknowledges that a single mode of content delivery is rarely sufficient for comprehensive understanding.
• Achievable Milestones: Zicherman explains that Oboe's engine is designed to "reduce [scope] to the key points" and make progress feel "lightweight and achievable." This combats the overwhelming nature of large topics, preventing learners from feeling daunted. The system aims to provide "quick wins" and maintain momentum, which is critical for keeping learners engaged over time.

The Hidden Cost of Overwhelming Scope

The tendency for LLMs to generate vast amounts of information without clear structure can be a significant detractor. Zicherman contrasts this with Oboe's approach, where the system "by default try[ies] to reduce that to the key points because it wants to make you feel like you are making progress along the way and hitting important milestones rather than giving you something that's like an unachievable thousand chapter long course." This deliberate limitation of scope, while seemingly restrictive, is a strategic choice to foster sustained engagement. The hidden consequence of an LLM-generated course that is too broad is that it can feel insurmountable, leading to the learner abandoning the project before any significant progress is made. Oboe’s strategy of presenting a manageable, milestone-driven path addresses this by making the learning journey feel feasible.

Embedded Formats: The Right Tool at the Right Time

Oboe incorporates "embedded formats"--quizzes, flashcards, or games--that appear at opportune moments to reinforce learning. This mirrors the practice of effective human teachers who use varied methods to solidify understanding. Zicherman describes this as giving the content generation pipeline "more of a toolkit... to use just like a teacher has a variety of different tools at their disposal."

This approach directly addresses the limitations of LLMs, which typically provide linear text-based responses. By strategically embedding these interactive elements, Oboe encourages active recall and application of knowledge, bridging the gap between passive consumption and active mastery. This is a key differentiator, as it acknowledges that learning is not just about receiving information but about actively processing and retaining it.

The Deep Dive: Quantum Mechanics, Philosophy, and the Nature of Knowledge

The conversation takes a fascinating turn as Zicherman and Shipper delve into more complex topics, revealing how AI can facilitate exploration even in areas that challenge our fundamental understanding of reality and knowledge itself.

Quantum Tunneling and the Limits of Intuition

Zicherman shares his personal use of Oboe to learn about quantum physics, specifically the Stern-Gerlach experiment. This experiment, conducted over a century ago, demonstrated the probabilistic nature of quantum spin, a concept that defies classical intuition. The experiment's findings--that a particle's state can be indeterminate until measured, and that subsequent measurements can alter or "forget" previous states--highlight the profound counter-intuitiveness of quantum mechanics.

"It just boggles my mind," Zicherman exclaims, reflecting on the ingenuity required to devise such experiments with the technology available in 1922. This anecdote underscores a broader point: many profound scientific concepts, much like complex philosophical ideas, remain outside the direct experience of our everyday Newtonian world. The challenge, then, is how to make these abstract concepts accessible. While Oboe can generate courses on topics like "Quantum Tunneling Explained," the underlying difficulty of the subject matter remains. The platform's strength lies in providing a structured pathway to grapple with these complex ideas, rather than claiming to make them intuitively simple.

Embedding Spaces and the Unseen Patterns of Knowledge

Shipper draws a compelling parallel between the probabilistic nature of quantum mechanics and the concept of "embedding spaces" in LLMs. He describes LLMs as "postmodern technologies" that learn "countless implicit patterns from... tiny tiny correlations in long pieces of text." This process, he suggests, is akin to how quantum particles exist in a probabilistic state until measured.

"The way that you measure it at the time that you measure it like determines whether it becomes a particle or wave," Shipper explains, referencing the double-slit experiment. Similarly, he posits that embedding spaces represent a high-dimensional existence of knowledge, where the "way you measure" or query the model determines the output. This analogy suggests that LLMs, like quantum phenomena, operate in realms of complexity that are not easily reducible to our linear, symbolic understanding.

Zicherman agrees, noting that these embedding spaces reveal dimensions of human output that "we are completely blind to." LLMs can identify patterns in human behavior and knowledge that are so subtle or complex that we, as humans, cannot consciously perceive them. This leads to a profound philosophical question: if AI can identify and utilize knowledge that we ourselves cannot articulate, what does it mean to "know" something?

The Deterministic vs. Probabilistic LLM Debate

A technical discussion arises regarding the deterministic nature of LLMs. While often perceived as probabilistic due to their varied outputs, Zicherman points out that the underlying models are, in theory, deterministic. The perceived randomness comes from deliberate introductions of variability (like the "temperature" setting) to make outputs more human-like.

However, a nuanced point is raised by the research from Thinking Machines, which suggests that even with temperature set to zero, LLMs may not be perfectly deterministic due to the parallel processing of floating-point additions on GPUs. This introduces subtle variations based on the order of operations, creating a "race condition." This technical detail has implications for our understanding of computation and potentially even the human brain, which also operates with a high degree of parallel processing and potential for non-deterministic outcomes.

This leads to a broader reflection on the nature of knowledge and consciousness. Zicherman questions whether the human mind, by its evolutionary design, might intentionally obscure its own mechanics, limiting our ability to fully comprehend complex patterns--a feat that LLMs, trained on vast datasets, can achieve. Shipper counters that the ability of LLMs to embody knowledge that humans collectively possess, even if not explicitly articulated, broadens our definition of what constitutes knowledge. This includes implicit, intuitive understanding that can be leveraged by AI, even if we can't fully explain its underlying principles.

The Future of Learning: Beyond Information Access

The conversation circles back to the core problem: the gap between information access and genuine understanding, and how AI platforms can bridge this divide.

Re-engagement and the Ephemeral Nature of Digital Content

Shipper shares his experience with ChatGPT, where reminders about unfinished learning tasks created a sense of guilt but ultimately failed to re-ignite motivation when he lost context. This highlights a critical failure in current AI: the lack of sophisticated re-engagement hooks. While LLMs can be set to remind users, they struggle to adapt to waning interest or lost context in a way that a human teacher or mentor would.

Zicherman acknowledges this challenge, describing it as a "tough balance to create content that feels lightweight but not ephemeral." Oboe's current model, allowing users to generate unlimited courses for free, offers a potential solution: if a course feels ephemeral, a new one can be generated, picking up where the user left off, with the AI retaining context. However, he admits that Oboe currently lacks the "re-engagement hooks" that a mature learning product would require, such as native mobile applications and sophisticated notification systems. The focus, he explains, has been on nailing the core utility--delivering value--before layering on these engagement features.

The Competitive Advantage of Difficulty

A recurring theme is that the most durable learning approaches often involve initial discomfort or delayed gratification. Zicherman's experience with learning advanced math and physics, subjects he "wished" he had majored in, exemplifies this. He realized that he "didn't need the academic setting" and could teach himself through tools like Oboe, enjoying a "lightweight experience where I can jump in and jump out as easily as I want to."

This highlights a key insight: the systems that foster genuine learning often require a commitment that goes beyond immediate gratification. The difficulty of a subject, or the effort required to master it, can become a moat. As Zicherman notes, the "grand vision is you can learn anything with Oboe." This vision is built on the premise that even daunting subjects can be made "achievable" by breaking them down and providing a supportive structure. The competitive advantage lies in building platforms that can guide users through this initial discomfort, leading to a more profound and lasting understanding that superficial, easily accessible information cannot provide.

Key Action Items

• Prioritize Pedagogical Design: When building or using AI for learning, focus on pedagogical principles (scaffolding, feedback, structured progression) rather than solely on information retrieval. This is an 18-24 month investment that pays off in deeper user engagement and learning outcomes.
• Embrace Multimodality: Recognize that learners benefit from diverse content formats. Integrate text, audio, video, and interactive elements where appropriate, tailoring the mix to the subject matter and learner preferences. This is an ongoing product development effort.
• Design for Achievability: Break down complex topics into smaller, manageable units with clear milestones. This combats overwhelm and fosters a sense of progress, crucial for sustained motivation. This should be a core design principle from day one.
• Develop Re-engagement Hooks: Implement strategies that actively re-engage learners when they drop off, helping them regain context and rekindling their motivation. This requires thoughtful design of notifications, personalized check-ins, and adaptive learning paths. This is a critical feature for long-term retention, with payoffs becoming evident over 6-12 months.
• Leverage Embedded Learning Tools: Integrate interactive elements like quizzes, flashcards, and simulations at opportune moments to reinforce learning and encourage active recall. This is an immediate tactical improvement that can be implemented incrementally.
• Understand the "Why" Behind Learning Goals: Structure learning experiences around clear objectives. This provides direction and motivation, making the learning process more meaningful and less prone to fizzling out. This insight should inform prompt engineering and course generation from the outset.
• Accept and Guide Through Discomfort: Recognize that deep learning often involves initial difficulty. Design systems that acknowledge this challenge and provide the necessary support and encouragement to navigate it, rather than avoiding complexity entirely. This requires a long-term commitment to user support and content refinement, with benefits seen over years.