Crazy Wisdom

In this episode of the Crazy Wisdom Podcast, host Stewart Alsop sits down with Devin Elliot, founder of Monetic xyz, to explore the evolving landscape of AI development, decentralized infrastructure, and the future of software engineering. The conversation ranges from Devin's decision to move away from Anthropic's services—cutting his monthly spend from $12,000 to zero—to his early adoption of "vibe coding" in 2022 and his extensive experience building with cursor and custom AI infrastructure. They discuss the technical architecture behind privacy-preserving systems, the value of embedders and fine-tuning, the economics of frontier AI companies, and why Devin believes the internet is becoming increasingly inhospitable to humans while agents take over. Along the way, Devin shares insights from his decades in engineering, his background in crypto since 2011, and his vision for a self-custodial future where users control their own data through content-addressed systems and encrypted safe tensor files rather than centralized databases.

1. Devin Elliott transitioned from traditional engineering to vibe coding in 2022 after fifteen years of experience, recognizing it as a fundamental cycle shift rather than resisting it. He reduced his Anthropic spending from $12,000 monthly to zero, partly due to disagreement with their public communications and partly because he could migrate his entire codebase to other platforms. His approach was strategic: he deliberately chose Cursor as a bridge tool despite knowing it would cost more in tokens, prioritizing speed and continuity over constant tool-switching. By staying focused on one platform, he avoided the productivity drain of repeatedly learning new interfaces while others chased the latest options.

2. Elliott's token usage was extraordinarily high, placing him in the top 0.2% of Cursor users with 25 billion tokens processed. His heavy consumption stemmed from spending tokens on priming context rather than reusing prompts or relying heavily on rules. His background as an architect rather than a pure engineer meant he focused on understanding architectural patterns and system design rather than memorizing specific function implementations. By investing heavily in context priming, he achieved near one-shot execution where the AI could implement complex features like WebSocket servers correctly and efficiently without multiple iterations.

3. The lack of engineering standards in software development has always been problematic, but AI coding amplifies this issue significantly. Most software engineers cannot explain fundamental architectural concepts like model-view-viewmodel or understand why different frameworks embody different state-handling philosophies. Engineers often debate which framework to choose without understanding the underlying engineering principles that make each suitable for specific purposes. If buildings were constructed like software is written, they would be structurally unsound and dangerous. This knowledge gap becomes more critical with AI coding because the quality of output depends heavily on understanding architectural patterns to guide the AI effectively.

4. The future of the internet will likely become inhospitable to human users as bots gain increasingly efficient access at lower levels than human interfaces allow. Elliott observed this in DeFi through MEV (miner extractable value) where bots could intercept and extract value from transactions before humans could complete them. He predicts the internet will evolve into a space primarily for agent-to-agent interaction, with humans interacting through local AI systems that venture into this hostile environment on their behalf. These personal agents would operate through proxy environments that can be attacked or compromised without exposing the user's actual systems or data.

5. Centralized AI services like Anthropic face fundamental architectural problems that decentralized approaches can solve more effectively. The traditional Silicon Valley model of centralizing user data in company databases to deliver personalized services cannot achieve true privacy preservation or meaningful personalization. This centralization model failed during the Internet of Things wave when consumers rejected giving their data to companies like Samsung just to receive basic functionality from their own devices. Elliott built his entire system around self-custodial, privacy-preserving architecture using local GPU racks and custom models, demonstrating that individuals and businesses can achieve AI capabilities without surrendering data to centralized providers.

6. Elliott developed a novel approach using embedders and content-addressed systems instead of traditional databases to handle data more efficiently and securely. Embedders are essentially LLMs with different final layers that convert text to vector space representations, and they can run locally with zero cost while generating tokens hyper-fast. He stores data as encrypted safe tensor files in blob storage rather than Postgres databases, allowing individual ownership, licensing, and sale of personal data. Users could compile their Facebook, Google, and social media data into valuable inventory files they control and potentially sell directly to buyers, bypassing traditional data brokers while maintaining privacy through encryption.

7. The system Elliott built over a year represents approximately 80,000 engineering hours of work that would have been impossible before AI coding capabilities emerged. His core system contains 3.5 million lines of quality code that passed specifications, with the account system alone comprising 250,000 lines. He successfully created an AI agent that autonomously set up an LLC in Wyoming, completing all paperwork, payment, and registration using biometric signatures for authorization. The architecture is built on a "thin client, fat protocol" model with extensive infrastructure supporting agent operations, including WireGuard-based mesh networks, custom security implementations, and seed files that can instantiate entire systems when copied to new locations.

00:00 Devin discusses leaving Anthropic after spending $12,000 monthly, moving entirely to cursor for vibe coding after adopting it in 2022 with fifteen years engineering experience behind him

05:00 Token management strategy explained through staying with cursor despite higher costs, choosing speed over switching tools while burning through $200 plans in fifteen minutes reaching 25 billion tokens

10:00 Engineering background shapes approach to context priming over reusable prompts, focusing on architectural patterns rather than memorizing native functions to achieve one-shot implementation

15:00 Compiler determinism discussed through donut-to-car mutation analogy, explaining how LLMs improved through better tool chains and loop checking rather than transformer changes

20:00 Internet of Things failure traced to centralized data architecture, arguing Ring camera backlash preceded GDPR and revealed impossibility of true privacy through centralized servers

25:00 Personalization requires decentralization as centralized algorithms need users staying static, conflicting with human ability to improve daily and make incremental changes

30:00 Local agent architecture described using WireGuard mesh networks and proxy cloud environments, enabling bots to safely interact with hostile Internet while protecting local systems

35:00 Anthropic's $42 billion losses and $8 billion monthly revenue discussed alongside server infrastructure value, questioning centralized model sustainability as decentralization pressure increases

40:00 Seed file system explained where single file instantiates across devices, reaching back through mesh networks to access compute resources like rack of A100s

45:00 Fine-tuning requirements range from 10,000 to 100,000 samples typically, though custom embedder training used 40 million samples to make geometry-respecting vector space

50:00 Agent autonomously filed Wyoming LLC using biometric signatures and custom-built secure enclaves, demonstrating 80,000 engineering hours compressed into one year

55:00 Data broker alternative proposed using safe tensor files in blob stores instead of databases, letting users own and sell their complete data inventory through encrypted vaults

Monetic

In this episode of the Crazy Wisdom Podcast, host Stewart Alsop sits down with UI designer Peter Petrash to explore the evolving landscape of user interface design and technology. Peter walks through his journey from discovering the early web on GeoCities to working with gaming companies in the nineties, and how his background shaped his current perspective on design. The conversation moves from the technical foundations of web design to deeper philosophical territory, examining how AI is transforming creativity, whether machines can truly understand taste and relevance, and what happens when human imagination becomes the bottleneck instead of technology. They discuss everything from self-driving cars in San Francisco and urban design principles to the future of rendering engines, the meaning crisis in an age of infinite content generation, and why formally defining goals often backfires—all while considering what it means to design interfaces when the full breadth of imagination can finally be unleashed on screen.

7 Key Insights from Peter Petrash's Episode:

1. The evolution of UI design has come full circle from requiring deep technical knowledge to becoming accessible through AI coding tools. Petrash describes starting in the Geocities era when designers had to master HTML, CSS, deployment, and package management just to get anything online. Today, tools like Claude handle most of the technical overhead, lowering barriers to entry. However, this democratization creates a new challenge where design principles and aesthetic judgment become more critical differentiators than coding ability.

2. Visual design operates on fundamental principles rooted in human cognition, not just subjective taste. Petrash explains gestalt principles of visual perception, such as the law of proximity where elements placed close together are perceived as related, and common fate where animated movement helps brains track relationships. These principles represent an assembly code for visual design that could theoretically be formalized and validated, similar to how code is verified, though this remains largely unexplored territory for AI systems.

3. The convergence of rendering technologies HTML, SVG, 3D, shaders, and now AI video generation points toward a future where any imaginable visual can be rendered on screen. Petrash describes emerging AI video models that can generate interactive, realistic environments without traditional rendering engines or lighting calculations. This represents a fundamental shift where the technical constraints that historically limited designers are disappearing, making imagination and taste the primary bottlenecks rather than technical capability.

4. Voice interfaces and multimodal AI create opportunities for capturing nuanced human communication that text misses. Petrash notes that when speaking spontaneously, humans convey uncertainty through pauses, vocal modulation, and hesitation patterns that current AI systems largely ignore through transcription to text. He envisions AI that could detect these subtle cues and adjust responses accordingly, offering educational context when uncertainty is detected or providing more confident direction when the user seems assured.

5. The professionalization of UI design creates a constant critical awareness of environmental design flaws that can become psychologically taxing. Petrash describes how training in user experience design makes designers perpetually scrutinize their surroundings, noticing poor design choices like Norman Doors that have push bars but require pulling. This occupational hazard of seeing flaws everywhere extends beyond digital interfaces to urban planning, architecture, and physical products, requiring conscious effort to disengage from this analytical mode.

6. The democratization of creative tools through AI is creating an attention economy where imagination itself may become currency, but also risks imaginative exhaustion. Petrash reflects on how technology has finally caught up to human imagination after decades of the reverse being true, but now finds himself coming up dry creatively as everyone's imagination is simultaneously unleashed on platforms like Twitter. This creates a paradoxical situation where infinite creative possibility collides with limited human attention and the challenge of generating meaningful novelty rather than novelty for its own sake.

7. Goal setting and life planning tools built on AI fundamentally misunderstand how human motivation and meaning work, revealing limits in artificial intelligence despite mathematical prowess. Petrash describes how apps attempting to help users identify and achieve goals through explicit articulation and reminders often fail because formalizing goals in writing strips away essential implicit elements. This Jungian insight suggests that bringing things fully into intellectual spotlight loses their essence, and that true intelligence involves navigating unconscious dimensions that current AI systems cannot access despite solving complex mathematical problems.

00:00 Peter's journey into UI design begins with discovering Mad Magazine's website as a kid, building on Geocities with HTML and inline styles

05:00 Gaming journalism and visiting Valve Software to see Half-Life in alpha, learning technical skills alongside artistic vision

10:00 Modern vibe coding versus old technical barriers, design becoming more important as AI handles coding tasks

15:00 Gestalt principles and visual perception laws like proximity explain how humans cognitively process interfaces

20:00 Voice UI experiences and capturing uncertainty through speech patterns, the potential for AI to detect hesitation

25:00 Norman Doors illustrate bad design everywhere, how designers constantly scrutinize their environment for flaws

30:00 Urban design comparing San Francisco's organic growth to Barcelona's centralized planning and Buenos Aires architecture

35:00 Building new cities takes generations, Brasilia and Chandigarh show problems with inorganic hyperfunctional design

40:00 Self-driving cars create social consciousness about gig economy impacts while improving safety dramatically

45:00 Argentina's robotics testing environment and Milei's transhumanist views including cloning his dogs

50:00 Interactive video models eliminate traditional rendering engines, creating persistence challenges for world state memory

55:00 Imagination becoming the bottleneck as technology catches up, future currency being taste and creative vision

60:00 Explicit goals fail when formalized, LLMs lack true intelligence despite reasoning capabilities and test-time compute

Mad Magazine Website
Pete's Personal Website

In this episode of the Crazy Wisdom Podcast, host Stewart Alsop sits down with Ryan Marsh of thestack.io to explore what it really takes to build production AI systems. They discuss how production AI has evolved from simple prompt-to-API demos into complex systems requiring evaluation suites, human-in-the-loop feedback mechanisms, and sophisticated approaches to handling context and data retrieval. The conversation covers the challenges of domain mapping, the fundamental difficulty of translating messy human business processes into structured systems, and the role of ontologies in AI development. Ryan and Stewart also examine the limitations of LLMs, the debate between specialization versus generalization in AI models, consciousness and cognition in system design, and the regulatory landscape facing AI companies. They touch on infrastructure constraints, the democratization of AI through open source models, and whether we'll eventually hit a ceiling where human intelligence can no longer distinguish between increasingly capable AI models.

1. Production AI systems today fundamentally differ from demos through their reliance on comprehensive evaluation suites that function like unit tests to measure and maintain quality, though they cannot be as deterministic. The key distinction is that production systems require clearly defined metrics for what good looks like, along with feedback mechanisms that allow the system to evolve over time. Without this foundational understanding of success metrics and continuous improvement processes, a system is not truly production-ready regardless of how many users it serves.

2. The fundamental challenge in building production AI systems is not the technology itself but rather mapping business domains into structured formats that models can work with effectively. This problem of translating messy, subjective human processes and language into precise specifications has plagued software engineering for decades. Different people within organizations use the same words to mean completely different things, and humans naturally operate with assumed context and imprecision that must be explicitly defined for AI systems to function reliably.

3. Large language models excel at generalization but struggle with specialization, which creates friction in production environments where specific outputs or styles are required. While they can code in any programming language, getting them to write code exactly the way a particular engineer wants remains extremely difficult. This explains why professional documentation and specialized coding tasks often require extensive prompting and fighting with the models, as they naturally gravitate toward their trained patterns rather than highly specific user preferences.

4. Modern production AI systems primarily solve classification problems wrapped in natural language interfaces rather than requiring true open-ended cognition. The models work best when they can leverage reasoning over provided information to make verifiable decisions, but they still lack common sense despite their vast knowledge. Success comes from teaching models everything about your specific domain and what good and bad outcomes look like, rather than relying solely on their general intelligence.

5. Context retrieval in production AI systems is fundamentally a data storage, search, and retrieval problem that has been solved many different ways throughout computing history. The appropriate solution depends entirely on the type of data being accessed, whether through vector databases, graph databases, relational databases, or even simple text search. The harnesses and frameworks for orchestrating AI agents have matured significantly, making the real challenge the quality and structure of the data being fed to these systems.

6. Human-in-the-loop feedback mechanisms are essential for production AI because models will inevitably encounter situations they have not been trained to handle. When confidence is low or novel scenarios appear, systems should flag these for human review rather than proceeding blindly. The feedback provided during these interventions must be captured and scored so it can be incorporated into the permanent behavior of the system, creating a continuous improvement cycle similar to how model vendors perform reinforcement learning on their base models.

7. The rush to regulation in the AI industry is driven primarily by the fact that these companies are currently completely exposed to existing consumer protection and liability laws with no legal precedents to protect them. If AI agents cause harm, companies could be sued into oblivion under current law. By establishing compliance frameworks through regulation, these companies can create carve-outs and exemptions that limit their liability when they follow prescribed rules, similar to how heavily regulated industries like airlines and banking have become nearly impossible to enter due to compliance requirements.

00:00 Welcome and introduction to Ryan Marsh discussing production AI systems and what companies need to understand about building them at scale

05:00 The cognitive load challenge of working with invariants and probabilistic systems, discussing how LLMs function like PhD students without common sense

10:00 Building eval suites to measure AI performance, handling the long tail of edge cases in complex contracts using human-in-the-loop feedback systems

15:00 Context as a data retrieval problem and the fundamental challenge of mapping business domains when humans struggle with imprecision and ambiguous language

20:00 How different departments use the same words with different meanings and why LLMs generalize well but don't specialize effectively for specific coding styles

25:00 The ontology debate and intractable problem of mapping subjective human systems to rigid structured formats with diminishing returns on perfect mapping

30:00 Why humans struggle distinguishing what is from what ought to be and the challenge of creating SOPs when companies lack updated documentation

35:00 The liability exposure AI companies face and why they're begging for regulation to protect themselves from existing consumer protection laws

40:00 Open source knowledge transfer between Chinese and American labs, chips and power as the real constraint not algorithms for frontier models

45:00 Reaching intelligence ceiling where specialists can't distinguish state-of-art models and fundamental physical laws limiting LLM scaling through layered optimization strategies

Website
X

Stewart Alsop sits down with Antoine Marcel, CEO and cofounder of Flourish, to explore the future of home robotics and what he calls "solarpunk" automation. Antoine discusses why Flourish is building wheeled robots with arms instead of following the humanoid hype, arguing that the expensive legs and teleoperation approach isn't practical for real consumers today. The conversation ranges from the philosophy of robots that disappear when you're home (rather than becoming constant companions) to the bureaucratic nightmares of modern life that robotics might actually solve. Antoine explains how their $3,500 robots can be trained in thirty minutes using just your phone to handle personalized household tasks, and why saving people two hours a day—not just ten minutes—is the threshold for consumer adoption. They also dig into whether Paris will become a solarpunk or cyberpunk city, why Apple represents solarpunk design principles, and how the first 50 Flourish units are about to ship worldwide at flourish-robots.com.

1. The core innovation behind Flourish is building mobile-based robots with arms but no legs, moving on wheels rather than attempting humanoid locomotion. This design choice stems from practical considerations since even humanoid robots cannot reliably handle stairs, and legs represent approximately twenty thousand dollars of the cost in humanoid platforms. By eliminating legs, the company can focus on delivering functional value to consumers today rather than developing research products. The founder emphasizes that accessibility and usefulness are paramount, rejecting the demonstration-focused approach that dominates the robotics industry where companies raise billions on teleoperated demos that never ship to actual consumers.

2. The founder's background in automation across logistics, software, and sales led him to recognize that the most painful tasks in life occur outside the computer. After automating digital work processes, he realized that returning home after a long workday to clean, cook, and tidy remained exhausting and unaddressed by existing technology. This insight, combined with advances in physical AI over the past three years, drove him to develop the first robot prototype that runs in his own home, constantly testing and iterating based on real-world use cases like navigating doorways and picking up scattered items from floors.

3. The company is building toward a solarpunk vision rather than cyberpunk future, where robots disappear from conscious awareness and operate silently in the background. The founder envisions robots emerging from closets when people leave home, performing cleaning and tidying tasks, then retreating before residents return. This approach contrasts sharply with the prevalent humanoid robot vision featuring teleoperated machines with cameras potentially monitored by remote operators. The goal is reducing technology's presence in daily life to create more time for family, creativity, and meaningful activities rather than introducing intrusive robotic presences into living spaces.

4. Consumer robotics faces fundamental challenges that explain why almost no companies successfully sell to end users despite massive investment in the sector. Industrial robotics offers clearer paths to revenue, while consumer products require extensive market education, safety certifications, and must genuinely change lives rather than provide marginal improvements. The founder argues that saving people ten or thirty minutes daily is insufficient and forgettable, but saving two hours daily creates transformative value that sustains product adoption. This high bar for meaningful impact explains why the consumer robotics market remains largely unaddressed despite technological capabilities.

5. The business model centers on teaching robots specific tasks through a thirty-minute training process using a smartphone, then fine-tuning AI models for those particular workflows. Rather than pursuing general physical AI that can handle any task, Flourish enables users to create custom automations for their specific needs, like retrieving items from the refrigerator or collecting clothes for the washer. This approach acknowledges that general physical AI is not yet ready while still delivering practical value through task-specific learning. The company plans to launch fifty units at thirty-five hundred dollars, shipping worldwide, with rapid sellout expected due to existing demand.

6. The robot extends vertically to reach from floor level to table height, enabling tasks like cleaning surfaces and accessing shelves while maintaining a compact mobile base that can navigate through doorways. Current limitations include lack of waterproofing, preventing dishwashing tasks, and challenges with delicate object manipulation requiring advanced tactile sensing. The founder uses first-generation prototypes in his own home, scheduling the robot to pick up scattered items like socks and trash throughout the day. This real-world testing revealed critical design issues like initial prototypes being too large to pass through standard doorways, problems only discoverable through actual home deployment rather than warehouse testing.

7. The rise of AI-powered robotics fundamentally transforms hardware development by enabling software engineers without traditional robotics expertise to launch robotics companies after just weeks of learning. This democratization stems from downloadable models from well-funded research companies like Physical Intelligence, whose work the founder incorporates into Flourish robots. This represents a shift from the industrial age's capital-intensive factory model to an information age paradigm where substantial capability can be rapidly deployed through software. However, this transformation raises questions about competition intensity, market education for skeptical consumers, and whether the benefits will broadly distribute or concentrate among those with will and intelligence to build while others remain excluded and fearful.

00:00 Stewart welcomes Antoine Marcel, CEO of Flourish, who explains they're building accessible home robots with wheeled bases and arms rather than expensive humanoid forms for practical domestic use.

05:00 Antoine discusses testing prototypes in his home for picking up clothes and trash, explaining how real-world testing revealed issues like doors being too narrow for early designs.

10:00 The conversation shifts to how AI hype enables useful research, with Antoine noting he uses physical intelligence models in his robots that wouldn't exist without funding bubbles.

15:00 Antoine emphasizes robots must save users two hours daily to matter, explaining why consumer robotics is harder than industrial applications due to certification requirements and market education needs.

20:00 Discussion of solarpunk philosophy versus cyberpunk futures, with Antoine envisioning robots that hide away when not needed, emerging only to clean and organize while families enjoy life.

25:00 They explore how different cities might adopt solarpunk or cyberpunk approaches, comparing bureaucracy as inherently cyberpunk while discussing Argentina's potential as a solarpunk capital.

30:00 Antoine positions Flourish like early Apple, making complex robotics simple and accessible, rejecting teleoperated solutions in favor of onboard AI that learns from users.

35:00 The business model discussion centers on unit economics, with Antoine arguing domestic robots should monetize time saved rather than attention captured like social media platforms.

40:00 Antoine advocates against robot screens and rental models, favoring voice control and ownership so robots learn individual routines and preferences through personalization.

45:00 They discuss robots disappearing from view when unneeded, contrasting home automation with dark factory models, before exploring robot aesthetics and art possibilities.

50:00 Antoine announces Flourish's launch of fifty units at $3,500 that users can train in thirty minutes using phones, teaching specific tasks the AI then fine-tunes for daily execution.

Flourish Robots

1. Certification requirements aren't dead ends—they're routing problems. When smaller companies got rejected from tenders for lacking SOC 2 or ISO certificates, Juan didn't turn them away. He found that EU procurement rules allow bidding as a consortium or subcontracting to a certified partner, turning a disqualifier into a workaround that builds trust with clients.
2. Public procurement is a massive, underexamined market. Roughly 20% of GDP flows through public purchasing of private-sector goods and services, yet most people have no visibility into how tenders work or how governments post and award these contracts.
3. Being wrong has become more socially acceptable. Juan traced this shift to the falling cost of information: in the Stack Overflow era, giving a wrong answer was costly, but now that answers are instant and abundant, both mistakes and corrections happen faster, changing how people learn and communicate.
4. Task-based pricing beats hourly billing for AI-era services. Rather than charging per hour, Tender Market prices around the deliverable, winning a tender, which avoids the perverse incentive of hourly billing to be inefficient and instead rewards actually solving the client's problem.
5. RAG and fine-tuning solve different problems. RAG helps a model reference large documents without hitting context limits, while fine-tuning changes a model's internal weights so it learns new behavior or style. Juan noted that true RAG use cases needing thousands of pages of context are rarer than the hype suggests.
6. Deterministic code should replace repeated LLM calls once a pattern is found. Stewart described his own workflow: solve a task with an LLM a handful of times, then convert the repeated pattern into deterministic software so tokens are no longer spent on it, freeing the model for genuinely new problems.
7. AI agents are never truly autonomous. Both hosts agreed that no matter how many steps an agent chains together, a human operator always initiates the first prompt, meaning accountability and intent trace back to a person even in multi-agent systems.

1. Link building is dying but not dead. Mark's SEO agency business peaked around 2022 working with brands like monday.com and HubSpot, but AI has steadily eroded the value of backlinks. There's still budget for links, just far less than during the gold rush years of 2018-2022.
2. Big companies waste money because they can't track it. B2B companies spending $2-500M+ a year in revenue often can't account for where ad and vendor spend actually goes. Once a budget exists, it gets allocated regardless of efficiency, especially in large organizations too bureaucratic to figure out AI.
3. AI is changing product discovery itself. Instead of searching Google and choosing from options, people now get AI recommendations directly, shifting the entire model of how consumers find products, which threatens the search-and-link economy that powered SEO for two decades.
4. The open internet may be closing. Concerns about Anthropic, OpenAI, and China scraping content freely, including old books, have pushed creators like Stewart to restrict access to older podcast episodes. This marks a cultural shift away from the openness that defined the early internet era.
5. AEO and llms.txt are the new frontier, and nobody fully understands them yet. Much like early SEO in 2001, there's a new discipline forming around how AI agents discover and read content, but no established playbook exists, leaving even experienced marketers uncertain.
6. AI-generated video is a high-effort, low-margin trap. Mark found that fully AI-generated ads were nearly unprofitable given the time needed to get results right, while blending real human footage with AI-altered environments (using tools like Google Omni) produced better, faster, more believable results.
7. YouTube's AI demonetization wave hit creators hard. As AI-generated content flooded YouTube, the platform cracked down broadly, devastating even legitimate animators and renderers. This pushed Mark to pivot away from consumer video tools back toward serving his existing B2B client base.