Imagine searching the web twenty years ago. You typed a few keywords, and Google gave you a neat list of blue links. You clicked, scrolled, skimmed, and eventually found your answer—or didn’t. Today, the experience feels different. You ask ChatGPT or Google's AI-enhanced search a question, and it delivers a neatly summarized, conversational answer. It feels revolutionary, intuitive, even magical. But beneath this shiny new surface, something foundational hasn't changed.

At its heart, the core problem of web search is still exactly what it was in 1998: figuring out what information is trustworthy. And that’s still done primarily through an old trick—PageRank.

The Heart of Web Search: Authority

When Larry Page and Sergey Brin invented PageRank, their insight was both simple and profound: Links are votes. The more reputable the linking website, the more valuable the vote. This effectively crowdsourced a judgment about which pages mattered most. Rather than merely counting keywords, PageRank leveraged the entire internet’s implicit knowledge about authority.

That fundamental challenge—deciding who to trust—is still the core problem of search. It doesn't matter how flashy the AI interface becomes; beneath the surface, trustworthiness and relevance are still judged largely by this same mechanism.

Why? Because authority isn't easily measurable without some external cue. PageRank cleverly leveraged what millions of website creators implicitly knew: Good sources link to good sources. AI hasn’t solved the fundamental challenge of deciding trust without those human signals.

AI Summaries Are Built on Old Foundations

ChatGPT, Google’s Bard, and other AI assistants seem revolutionary. You ask a detailed question, and instead of blue links, you get a clean, authoritative-sounding paragraph. But peel back a layer, and you see these AIs aren’t independently evaluating every webpage on the internet—they still rely heavily on existing search results to form their answers.

These AI systems synthesize and repackage information sourced from webpages ranked by Google's traditional algorithms. If the underlying ranking is flawed, AI-generated responses amplify those flaws. If it’s robust, the AI appears intelligent. In other words, AI hasn't fundamentally disrupted web search—it has put a conversational, polished face on the old infrastructure.

The Quiet Dominance of PageRank

Despite impressive leaps in machine learning, nothing has yet displaced PageRank's underlying logic. That's because authority, as a concept, is stubbornly resistant to technical solutions. Determining trust is inherently a human-like judgment. Algorithms can approximate it, but not easily replace the fundamental wisdom embedded in link-based authority.

AI models are great at language, summarization, and reasoning, but they’re still relatively naive at evaluating factual accuracy or credibility without reference points provided by human judgment. As smart as ChatGPT seems, its strength currently lies in repackaging trusted information, not independently discovering new authoritative sources.

The Traffic Shift: AI Changes Everything but Ranking

Yet, while the fundamentals remain intact, the way users interact with search results is shifting dramatically. AI search interfaces—where an immediate, concise answer is served up—significantly reduce the incentive to click on individual links. Publishers who once relied on search traffic find fewer visitors reaching their sites.

This traffic shift has profound implications for businesses, advertising, and even how we build websites. But don’t mistake this surface-level disruption for a deeper change. AI hasn’t overturned the web’s underlying architecture; it’s just shifting the way we interact with it.

Imagining a World Beyond PageRank

Could AI ever truly replace PageRank? Imagine a "DeepRank," a hypothetical future AI model sophisticated enough to independently evaluate and rank webpages by deeply understanding their content, not just their incoming links. Such a model could identify genuinely expert sources without relying solely on external signals. But getting there is hard.

Today's AI, while impressive, regularly struggles with nuance, context, and—critically—factual accuracy. The complexity of trust goes beyond checking claims; it involves reputation, timeliness, and community consensus. Fully AI-driven authority evaluation must surpass simple comprehension—it must develop nuanced judgment.

Google is undoubtedly experimenting with such approaches, but thus far, no credible alternative to PageRank has emerged at scale. Trust remains stubbornly human.

New Interface, Same Game

Yes, AI has changed how we interact with search—but not the fundamental principles underneath it. The core challenge of web search is still what it was twenty years ago: judging what’s trustworthy.

PageRank, for all its simplicity, remains central to solving this problem. As entrepreneurs, investors, and creators, we’d do well to recognize that beneath AI’s shiny new interface, the rules remain largely unchanged. Authority is still king. Until AI genuinely cracks the code of measuring trust in a fundamentally new way, expect evolution, not revolution, in search.

The game hasn’t changed—at least not yet.

When I moved to Silicon Valley in 2023, everyone had the same question: "Why aren't you starting an AI content company?"

With Jasper and similar platforms raising massive rounds, it seemed like the obvious play for someone with my content marketing background.

I always said no for two reasons:

1. I didn't need to start another company (my investments were doing pretty well).

2. More importantly: AI simply couldn't produce genuinely good content

AI was missing two critical ingredients that make content worth reading. But now, for the first time, that's changing…

AI-generated content has been often dismissed as "slop." Tools like Jasper and Copy.ai promised to revolutionize content creation but largely fell short. The reason is simple: they lacked the ability to reason and experience—two crucial ingredients for producing truly engaging content.

Think about it: the most compelling content comes from two key elements—new information and a fresh perspective, both of which require reasoning and experience. AI has struggled with this because, until recently, it couldn't truly reason or draw on experiential data. It was good at remixing existing information but not at generating genuinely original insights.

What Makes an Insight Valuable?

Paul Graham has argued that the best insights are both general and surprising. A good idea should apply broadly, revealing something fundamental that others can recognize across different contexts, while also being unexpected enough to challenge conventional wisdom. Most AI-generated content has failed to meet this standard because it lacks the depth of reasoning necessary to produce truly insightful analysis.

Ben Thompson's Stratechery exemplifies this principle. His best essays don't just report on industry trends; they extract general principles from specific events—like the aggregation theory, which explains how digital platforms accumulate power in ways that traditional businesses never could. His insights are surprising because they often reframe the reader's understanding, yet general enough that they apply across industries.

Similarly, Marques Brownlee (MKBHD) does more than review gadgets—he derives principles about user experience, product design, and technology's impact on daily life. His videos don't just say this phone is good; they uncover broader patterns, like why hardware matters less than software ecosystems or how small design decisions shape user perception.

AI-generated content has typically failed to meet Graham's standard of providing general and surprising insights. It could summarize but not synthesize. It could provide surface-level observations but not challenge assumptions with truly original perspectives. That's because reasoning and experience are prerequisites for producing content that offers genuine value beyond what already exists.

The Reasoning Revolution

This is changing with the emergence of advanced reasoning models. Companies like Anthropic have introduced AI systems such as Claude 3.7 Sonnet, which features a hybrid reasoning approach that combines multiple methods to solve complex problems. This model's "extended thinking mode" allows the AI to self-reflect before answering, significantly enhancing its performance in tasks requiring deep logical thinking.

For example, when given a complex ethical dilemma about resource allocation, Claude 3.7 can now trace through multiple possible outcomes, weigh competing values, and articulate nuanced reasoning that considers both immediate consequences and long-term principles—a level of analysis that would have been impossible with earlier models.

Similarly, OpenAI's o3 model is designed to devote additional deliberation time to questions requiring step-by-step logical reasoning. Instead of rushing to an answer, these models can now plan ahead and reason through tasks methodically, performing intermediate steps much like a human expert would when solving complex problems.

This approach has yielded remarkable results in domains previously challenging for AI, such as:

• Mathematical reasoning: Working through multi-step proofs with formal logic

• Scientific analysis: Developing hypotheses based on experimental data

• Strategic thinking: Evaluating complex scenarios with competing variables

These models represent a significant shift from previous AI systems by incorporating self-assessment and internal deliberation, allowing AI to generate more nuanced and insightful content that's closer to human-quality thinking.

The experience revolution

In addition to improved reasoning models, AI agents and robotics are advancing rapidly. AI agents transform AI from a reactive tool to an autonomous actor capable of independently executing complex tasks. This evolution results in significant leaps in productivity and efficiency, enabling increased revenue, decreased expenses, and more innovation.

One of the most significant advancements in AI agents is OpenAI's Operator, a tool that goes beyond text generation to autonomously execute online tasks. Powered by the Computer-Using Agent (CUA) model, Operator can interact with web interfaces in a human-like manner—clicking buttons, filling out forms, and even navigating complex workflows. This ability allows AI to experience and act in digital environments, rather than merely processing static data. By integrating AI into practical applications like booking travel, managing online transactions, and assisting with research, Operator exemplifies how AI is shifting from passive computation to active participation in real-world tasks.

Robotics has also seen substantial progress, particularly in the field of humanoid automation. Companies like Figure AI are pioneering the development of humanoid robots such as Figure 02, which can perform tasks traditionally reserved for humans. Standing at 5 feet 6 inches, Figure 02 is designed to navigate human-centric environments, performing household chores, warehouse operations, and even assisting in healthcare settings. Equipped with the Helix visual-language-action (VLA) model, it can interpret commands, assess its surroundings, and execute complex sequences of actions autonomously. This represents a leap toward AI-driven physical interactions that mirror human experiences.

These advancements suggest that AI is moving beyond mere data processing to engaging in tasks that require a form of experience, thereby enhancing its ability to produce content grounded in practical application. The ability to interact with digital and physical environments not only increases AI’s utility but also allows it to generate richer, more informed insights—leading to the next wave of AI-generated content that is more thoughtful, original, and deeply integrated into human workflows.

The Future of AI-Generated Content

The ingredients for great content are finally in place. AI isn't fully there yet, but for the first time, the trajectory is clear. As models continue to evolve, AI-generated content will shift from remixing old ideas to discovering new ones. We're witnessing the transition from AI as a sophisticated parrot to AI as a genuine thought partner.

In the coming years, we can expect to see:

1. AI systems that generate truly novel hypotheses in scientific research

Advanced reasoning models will transform scientific discovery by identifying non-obvious connections across disparate domains that human researchers might overlook. Rather than simply processing existing data faster, these systems will propose entirely new research directions by recognizing patterns that span multiple disciplines. We're already seeing early examples in drug discovery, where reasoning-enhanced AI has identified novel molecular configurations based on theoretical principles rather than just recombining known structures. This represents a fundamental shift from AI as a research tool to AI as a genuine scientific collaborator capable of accelerating breakthroughs in medicine, climate science, and materials engineering.

2. Content platforms that produce original analyses of market trends or political developments

Business and political analysis will move beyond data aggregation to genuine strategic intelligence as AI develops the contextual understanding needed to identify underlying patterns beneath surface-level events. Next-generation analytical platforms will connect micro-trends across industries to predict macro-shifts before they become obvious, providing decision-makers with truly actionable insights rather than just information. These systems will excel at detecting subtle signals that telegraph major directional changes, uncovering the strategic logic behind market movements or political decisions that might otherwise remain obscured. The result will be content that provides genuine competitive advantage by revealing connections and implications that even expert analysts might miss.

3. Creative tools that don't just mimic existing styles but pioneer new aesthetic approaches

Perhaps most profoundly, reasoning-capable AI will break free from simple mimicry to develop its own aesthetic principles and creative approaches. Rather than just training on existing human works, these systems will identify underlying patterns and extend them in directions humans haven't explored, becoming genuine creative partners rather than sophisticated copying machines. We're seeing early indications in experimental systems that don't just imitate established styles but develop novel structures based on conceptual frameworks rather than mere references. In writing, music, visual arts, and design, this will manifest as AI that doesn't just sound or look like familiar creative works but represents genuinely new additions to our creative landscape—expanding human creativity rather than simply reproducing it.

This isn't just an upgrade—it's the beginning of a fundamental transformation in how we think about the relationship between human and machine intelligence. The era of AI slop is ending; the age of AI insight is just beginning.

Recently, I published an article, and to my surprise, an investor I deeply admire shared it. It wasn’t just a passing like—it led to a conversation, and we ended up connecting and having a call. That moment made me reflect on how writing isn’t just about building an audience by teaching those behind us. It can also be a way to reach the people we admire, to start conversations that might not have happened otherwise.

In the world of content creation, the prevailing notion is to write down: write for an audience that trails behind you. Share your expertise, guide them through your journey, and build an audience. But there's another, often overlooked path: writing up.

Writing up is about crafting content that resonates with those you admire, those who are ahead of you. It’s about creating a bridge to connect with the people who inspire you. When I started writing, it wasn’t about building an audience; it was about reaching out to the minds I respected.

At the beginning of my career, I was living in a small town with no one to discuss technology with. I wanted to write and talk about startups and tech. One of the advantages of writing up is that sometimes, even if you don’t land on your heroes, you end up finding peers. Back then, I met my co-founder, who liked one of my articles, and we ended up building something together.

Another article I wrote led me to connect with a CEO I admired. It wasn’t about generating leads; it was about expressing ideas that resonated with someone on a similar wavelength. This approach—writing up—opens doors to relationships you might never have imagined.

Many people shy away from creating content because they don't see themselves as audience builders or educators. They don't feel the need or desire to teach. As a result, they hold back from sharing their thoughts. However, these individuals might discover that writing up, rather than down, can be a powerful way to connect with their heroes.

When you write for an audience of those behind you, it’s easy to get complacent. But when you write for your heroes, you push yourself to deliver your best work—the kind of work you’d be proud of, the kind that challenges you to grow. Write as if your heroes eagerly read each and every one of your posts.

Writing up can be a refreshing and impactful way to connect with your heroes. It’s a reminder that sometimes, the best way to move forward is to look up.

A fascinating shift is happening in software development. For the first time since the early days of personal computing, individuals can build sophisticated applications that rival professional software. The catalyst? AI.

Most people don't realize how significant this is. The last time we saw something similar was in the late 1970s with VisiCalc. Back then, if you wanted to do financial modeling, you either had to write a program from scratch or hire someone to do it. VisiCalc changed that by letting regular people "program" through a simple table interface. You didn't need to know BASIC or FORTRAN – you just had to understand cells and formulas.

This pattern kept repeating. HyperCard let people build interactive programs without being programmers. Excel macros and Access databases gave office workers the power to create custom tools. Personal websites let anyone publish to the world. Each wave democratized some aspect of computing, but there was always a ceiling – a point where you needed "real" programming to go further.

AI is removing that ceiling.

Consider what happens when someone wants to build their own note-taking app today. Previously, they'd face months of work implementing basic features: text editing, sync, search, auth. Most people would give up and use Evernote. But with AI, you can generate much of this functionality through prompts. It's like having a team of junior developers at your command.

The really interesting part isn't just that AI makes development faster – it's that it enables entirely new kinds of personal software. Want to build an app that automatically generates workout plans based on your fitness data? Or a personal assistant that understands your exact workplace jargon? These projects would have been impossibly complex for an individual before. Now they're weekend projects.

This matters because the best software often comes from solving your own problems. When you're scratching your own itch, you have perfect product-market fit. You know exactly what the software needs to do because you're the user. The challenge has always been bridging the gap between what you want to build and what you're capable of building. AI is closing that gap.

What we're seeing now reminds me of the early days of the web. Back then, people built weird, specific tools just because they could. Most were useless to anyone except their creators, but that was the point. They were personal software in the truest sense.

I expect we'll see a renaissance of this kind of development. The combination of AI assistance and modern development tools means that the scope of what one person can build is expanding dramatically. We'll see more people creating highly specialized tools for their own use cases – not to start companies or get rich, but simply because they can.

This won't replace professional software development, just like spreadsheets didn't replace databases. But it will create a new category of software that sits between consumer apps and professional tools – personal software that's sophisticated enough to be useful but specific enough that it wouldn't make sense as a commercial product.

The really exciting part is that we're just at the beginning. As AI tools get more sophisticated, the gap between what an individual can envision and what they can build will keep shrinking. We might be entering a golden age of personal software – one where the limiting factor isn't technical ability, but imagination.

After selling my first company, I found myself in an odd position: I had achieved what many developers dream of, yet I'd lost my drive to create. Success had somehow stripped away the very thing that got me there – the joy of building.

But then something changed. I started writing code just for myself.

This might sound counterintuitive. In a world obsessed with scaling and user acquisition, building software for an audience of one seems almost selfish. But that's exactly what made it powerful. When you're the only user, there's nowhere to hide. You can't rationalize a clunky feature or justify a confusing interface. The feedback loop is immediate and brutally honest.

I call this approach "selfish software," and it's about to become more relevant than ever.

Think about how we typically interact with software today. We contort our workflows to match rigid systems, adapt our thinking to pre-built solutions, and accept compromises in functionality because "that's just how it works." It's backward. Software should conform to us, not the other way around.

This is where artificial intelligence enters the story. AI is quietly demolishing the barriers that once made personal software development impractical. In the past, creating competitive software required teams of developers just to implement table-stakes features. Want to build a note-taking app? Get ready to spend months on basic editing capabilities before you can even think about your unique ideas. But AI is changing this equation, allowing individuals to rapidly implement foundational features and focus on innovation.

The result? We're seeing the emergence of two distinct categories of personal software. First, there's personal software – tools we craft for ongoing use. Then there's what I call "disposable software" – programs generated on the fly by AI for specific, temporary needs. Both represent a fundamental shift in how we think about software creation.

But this isn't just about productivity or efficiency. Personal software development is becoming a form of self-expression, like writing or music. When you build for yourself, you're not just solving problems – you're creating a digital extension of how you think and work.

The educational implications are equally profound. Remember Richard Feynman's famous principle: you don't truly understand something until you can create it. Personal software development embodies this idea perfectly. Want to learn about physics? Build a simulation. Curious about music theory? Create a composition tool. The malleability of software makes it an unparalleled medium for learning through creation.

Critics might argue that personal software is too niche, too individualistic. But that's missing the point. When you build something that scratches your own itch perfectly, you often discover others have been silently suffering the same discomfort. Some of the most successful software products started as personal solutions to individual problems.

There's something beautifully honest about selfish software. The feedback loop is pure: if you're not using what you built, it's not good enough. No market research needed, no user interviews required. You can't fool yourself about its utility because you experience it firsthand.

As our digital world grows more complex, the need for personalized solutions becomes more acute. The era of one-size-fits-all software is fading. We're entering a time where the best tools might be the ones we craft for ourselves, shaped by our unique needs and perspectives.

For me, this approach did more than just restore my motivation to code. It reminded me why I fell in love with software development in the first place – the pure joy of creating something useful, something personal, something real. In a world of increasingly generic solutions, maybe being a little selfish with our software isn't just acceptable – it's essential.

The world is aging fast. By 2030 the world population aged 60 years and over will be 1.4 billion, and it will grow to 2.1 billion in 2050. A larger elderly population can be disastrous to the world economy, causing healthcare costs to explode. However, if we can increase the amount of healthy productive life by even one year, it could be worth an additional US$38 trillion to the world economy.

miRscience aims to control genes to make us grow old without growing weak. They are developing new treatments using microRNAs to control genes related to muscle loss in old age.

Growing weaker

Starting around age 50, individuals begin to lose muscle mass at a rate of 1-2% annually, leading to Sarcopenia—a condition characterized by significantly reduced muscle mass. The resulting loss in muscle power affects daily living activities such as climbing steps or even raising from a chair. The extreme loss of lean body mass present in Sarcopenia is associated with increased risk of infection, decreased immunity, and poor wound healing.

Losing muscle mass will decrease your quality of life significantly. We take too much for granted when we are young. Simple things that you don't even think about, like getting up from a chair, become challenges when you don't have enough muscles.

It is estimated that 5-13% of all people aged 60-70 are affected by Sarcopenia. These numbers increase to 11-50% for those aged 80 or above. Even for the elderly who end up not developing Sarcopenia, the yearly muscle loss can represent a big challenge to the quality of life.

To live longer, healthy, and be able to enjoy our old age, we need to solve Sarcopenia. It could be the difference between aging in bed or playing with your grandkids outside.

Gene Regulation: Beyond Destiny

The Human Genome Project was a great step forward to understanding how life works. It created a map of all human genes. In general, a gene is a biological "recipe" to "make" proteins. However, knowing all the recipes is not enough.

Imagine that you have access to all recipes from a restaurant. With this information, you know all the possible dishes that will be served at that restaurant, but you will not know at a given time what dishes are served and to whom they are served.

Each gene is a recipe, but not all recipes are being made at all times in our cellular restaurant. Not only that, but our cells are different because they are producing different recipes. Your skin cells produce a different set of recipes than your liver cells. It's like a restaurant with a chef that can cook Mexican, Italian, and Brazilian food, but the restaurant is only Italian.

Another way of seeing this is to imagine a big house with many lights. With an electrical map of all the lamps, you can tell how many are in which room. However, just by looking at the map, you can't tell which ones are activated.

The Human Genome showed us that only understanding what genes an organism has is not enough to understand how it works. Not all genes are active all the time. The dynamics of what gene is activated and when is key to understanding what's going on inside a cell.

The mechanisms of turning genes on and off are broadly called Gene Regulation. To understand how it works, first, we need to remember the "Central Dogma" of biology:

DNA -> mRNA -> Proteins

The information in the DNA is used to create mRNA ( Messenger RNA ) molecules used to build proteins. This whole process is called Gene Expression. We can say that a gene is expressed when the protein it represents is produced.

The regulation of genes can work in any of these pathway steps. One of the major ways of turning a gene off is by interfering with the mRNA being produced by it.

Think back about our house lights analogy. The light switches are the genes, the wiring between the switches and the lamps are the RNA, and the lamps are the proteins. If you cut the wires between a light switch and a lamp, the light will not turn on when you press it.

miRscience uses one of the most promising techniques to "turn off" genes called microRNA. They will use it to turn off genes related to Sarcopenia.

Finding gold in the junkyard

After the Human Genome was finished, we discovered that 99% of the human genome didn't code for any proteins. The non-coding regions were called "junk DNA" and were thought as biologically not that useful. However, inside this "junk DNA" lies the key to understanding how life works.

Inside the "junk" parts of the Genome, it was discovered a new type of DNA sequence is used to create a special kind of RNA called Micro RNA. They are essential to regulate how our genes work. The first miRNA was discovered in the early 1990s. However, miRNAs were not recognized as a distinct class of biological regulators until the early 2000s.

Micro RNAs showed that the "inactive junk" part of our DNA was one of the control centers of our cells. The miRNA ( MicroRNAs ) attaches to special proteins and becomes a molecular machine known as RISC (RNA-induced silencing complex).

The RISC machine finds and attaches to specific mRNA (messenger RNA). Remember, one step of the central dogma is that mRNA is used to make proteins; no mRNA means no protein. When RISC is attached to mRNA, the cell can't make that protein encoded in the mRNA. Since proteins are usually what "do the work" inside a cell, the gene will not have its intended effect if the protein is not made.

miRscience will create MicroRNAs to target genes related to the loss of muscle. Their "miRNA machine" will attach to the mRNA produced by those genes and stop them from being used to create proteins. 

Beyond Gene Editing: Controlling Life

There is a huge market now regarding gene-editing technologies. Many genetic diseases could be fixed by editing our DNA. Those diseases were previously impossible to cure. Fixing genetic diseases by eliminating faulty genes or adding new ones has been a dream for decades, and now the first crop of treatments is getting into the market.

One example is SMA (Spinal muscular atrophy) which is a neuromuscular disease that causes muscles to become weak and waste away. People with SMA lose a specific type of nerve cell in the spinal cord (called motor neurons) that control muscle movement. In 2019 a new gene-editing treatment for SMA called Zolensma was approved by the FDA. Zolensma is the most expensive drug in the USA, costing US$ 2.5m. Zolensma is given as a one-time intravenous (IV) dose. It works by replacing the defective or missing survival motor neuron 1 (SMN1) gene.

In 2020 Jennifer Doudna won the Nobel Prize for her work with CRISPRtechnology. A discovery that gave thousands of scientists worldwide the power to edit the code of life. CRISPR presented the opportunity to edit DNA easier and cheaper than ever before. 

The discovery of CRISPR led to an explosion of interest from Big Pharma and Startups to explore Gene Editing. Most of these new treatments and products are in early phase development, but the FDA approval of gene therapy treatments like Zolensma signals a promising future. Many multi-billion dollar companies will be created in the Gene Therapy space. We are only in the beginning. 

Even though I'm excited about the prospect of editing human genes, I'm even more excited about controlling genes. miRscience is in its early days of using Micro RNA to control Gene Expression. The applications of its technology can go much further than Sarcopenia treatments. The more we understand how our cells work, the more it will be clear that Gene Regulation is the key to controlling cell behavior. 

Life is made of a complex network of chemical reactions. Removing and adding Genes is a powerful first step, but it's a limited form of control. It's like turning off light by removing the light bulb instead of just turning a switch on and off. I believe that the most complex biological behavior is a function of the genes and the control system of those genes. We will not fully understand how life works if we don't take control of the switches.

Google and publishers once had a straightforward deal: Google's search engine brought traffic, and publishers provided content. This symbiotic relationship worked well, benefiting both sides. Google grew with its ad model, and publishers got more readers.

But the landscape started to shift. Google began to display its own content in search results, not just links to external sites. A prime example is the Rap Genius incident. Rap Genius, a site known for its song lyrics, accused Google of using their content without permission. To prove this, they cleverly embedded a unique pattern of apostrophes in their lyrics. This pattern spelled out "RED HANDED" in Morse code. When these distinct lyrics started appearing verbatim in Google’s search results, it was clear evidence that Google was scraping and using Rap Genius's content. Although the Supreme Court dismissed their lawsuit, this incident underscored the growing strains in Google's relationship with content publishers.

As AI continues to reshape the search landscape, Gartner’s prediction comes into sharp focus: a 50% drop in organic search traffic for brands by 2028. This is primarily due to the rise of AI tools like Bard and ChatGPT, which provide direct answers, reducing the need for users to visit external websites.

Google's planned release of the Search Generative Experience (SGE) is poised to further transform this landscape. With answers displayed directly below the search box, the fundamental exchange at the heart of Google's covenant with publishers is under threat. If publishers no longer receive traffic in exchange for their content, the very incentive for this symbiosis vanishes.

The relationship between Google and the publishers will need to be reworked. If Google is not delivering on their side of the bargain. I see two potential directions this can go.

The first, is Google adding a Spotify like model to pay the publishers that own the content being used as source to generate the answers. The second, is Google picking the fight. In this scenario Google has a massive advantage today compared to a decade ago. YouTube has become a massive source of content that Google could index and use to generate a lot of responses putting pressure on the publishers.

Google and publishers are at a turning point. AI, especially tools like ChatGPT, is changing how we search. This means the old way of trading content for traffic doesn't work anymore. Google's new moves, like the Search Generative Experience, push publishers to think differently. We're seeing a big shift, not just in technology, but in how we share and find information online.

I've always been skeptical about venture studios. The concept of a "startup factory" struck me as fanciful, something only those who haven't experienced the grueling journey of building a startup might romanticize. The notion of juggling multiple startups simultaneously seemed especially far-fetched, considering the Herculean task of making even one successful. Yet, here I am, embarking on creating a venture studio myself.

My decision to pursue this path stems from a simple truth: I love building things. I've dabbled in angel investing, and while intriguing, it never sparked the same excitement in me. There's a thrill in being in the trenches, directly involved in the chaos of creation, that mere advising or investing can't match. This realization led me to the venture studio model, a better fit for my hands-on approach than a traditional VC firm.

Why not latch onto a single idea and be a founder again? My preference leans towards the early stages of a startup - the ideation phase rather than the scaling phase. I liken it to the difference between prospecting for gold and building a mining operation once it is found. Early stages involve discovery, while later stages focus on capitalizing on successful ideas. I thrive in the former.

Retirement was a tempting option, filled with books and thinking - activities I deeply enjoy. However, the current AI revolution has rekindled a passion in me I haven't felt since my coding days. AI, the holy grail for programmers, is experiencing breathtaking advancements, making it impossible for me to step back now.

My inspiration partly comes from studying Sutter Hill Ventures (SHV). SHV's unique approach, particularly their involvement in founding about a quarter of their portfolio companies, including the success story of Snowflake, is remarkable. Their method, though challenging, has proven effective, influencing my strategy for Coeus Ventures.

At Coeus, we're adapting SHV's playbook to build AI-focused companies. We're adopting their 'serial' approach, concentrating on a few key bets rather than scattering resources. I plan to immerse myself as CEO in the primary venture each year, dedicating 80% of my efforts there while allotting the remaining 20% to explore new ideas. Our experiments will be documented here, though our main project each year will remain in stealth mode. We're excited about delving into three major themes:

• Intelligence Amplification

• Agents

• Robotics

If you're navigating similar waters or are interested in future collaborations, feel free to reach out at edmar at coeus.vc. Let's build something remarkable together.

In the world of traditional software development, experience breeds intuition. Engineers learn the ropes, becoming adept at gauging what's feasible and what's not. The uncertainties usually dance around quality, costs, and performance, rarely feasibility. But with Generative AI, the rulebook gets tossed out the window. It's a new frontier where even seasoned engineers find themselves back at square one, trying to figure out the bounds of what's possible.

One might attribute this to the novelty of it all. Groundbreaking models like GPT-4 are barely a year old, and everyone's still getting their feet wet. But there's more to it. The black-box nature of modern AI systems, with their stochastic behaviors, brings a level of unpredictability that's both exciting and challenging. It's like venturing into a dense forest with a compass but no map.

Now, the conversation extends beyond just building a product to understanding and unraveling what we've built. It's a whole new ballgame. Foundation models are like raw, uncharted lands waiting to be explored. They are "few-shot learners," able to generalize and transfer knowledge beyond what their creators intended. It's almost like having a conversation with these models to discover their capabilities.

This unpredictable nature of Generative AI has birthed new roles in the tech world. We already have ML Researchers and ML Engineers. Now, we have AI Engineers, a new breed who deal with extracting value out of foundation models. It's a shift from the nitty-gritty of algorithm training to exploring the vast capabilities of pre-built models.

The term "Prompt Engineer" was floated around, but it barely scratches the surface of what AI Engineers do. It's not just about prompting; it's about orchestrating a symphony between the code, the model, and the unknown potential waiting to be unlocked.

But with all this excitement comes a sobering reality. The unpredictability means that products built on Generative AI carry a cloud of uncertainty. Even with a robust model, there's no telling if the product will hit the mark. The feasibility check is replaced by a series of exploratory sprints to understand what the AI can do and how it aligns with market needs.

In essence, AI Engineering is not just a shift in technology; it's a shift in mindset. It demands a spirit of exploration, a willingness to embrace the unknown, and a relentless pursuit to decode the capabilities of Generative AI. It's about navigating through a maze where the walls keep shifting, but the potential reward at the end is a treasure trove of innovative products that could redefine the digital landscape.

The discussions around Extended Reality (XR) technologies often orbit around their capability for immersive experiences or their myriad applications across fields. Yet, a nuanced factor—legibility—often escapes the spotlight in these dialogues. Legibility, or the ease of understanding that comes with well-defined systems, subtly yet significantly steers us toward the embrace of XR technologies. Unlike the real world, with its tendency towards chaos and unpredictability, XR offers a realm where the rules are clear and the narrative easy to follow. This essay endeavors to illuminate the overlooked allure of legibility and how it serves as an undercurrent, propelling us toward a more legible reality through XR.

In the realm of human endeavor, the lure of legibility is a silent guide. We’re charmed by the comprehensible, the clear, the understandable. It's not a loud preference, but it's a consistent one. This liking for legibility, a simple, digestible reality, can be seen as a compass leading us toward the adoption of Extended Reality (XR) technologies.

Consider the life of a diligent student. There's a clear path — ace the exams, get into a reputable school, secure a good job, and climb the corporate ladder. It’s a legible script, where the next step is often clear and the rewards are tangible. Now, let’s wander into the life of an entrepreneur. The path is less trodden, the map less clear, and the landscape less legible. It's no wonder methodologies like Lean Startups gained traction. They tried to bring a semblance of legibility to the wild west of entrepreneurship.

Now, let’s stroll into the realm of video games. They are a poster child for legibility. The rules are clear, the goals are defined, and the feedback is immediate. It’s a digestible reality, a stark contrast to the often muddled, unpredictable real world. Gamers understand their virtual environment, and there’s comfort in that understanding.

This magnetic pull towards legibility isn't a shout but a whisper leading us to the threshold of XR technologies. With Virtual Reality (VR) and Augmented Reality (AR), we’re on the cusp of blending the legible with the real. AR overlays digital clarity onto our physical world, making the opaque transparent. VR, on the other hand, crafts entirely new realms where legibility is the law.

It's not an overstatement to propose that XR technologies may weave into our daily fabric much like the internet did. The internet simplified a vast expanse of information, rendering it legible, accessible, and, in turn, indispensable. Similarly, XR holds the promise of untangling the complex threads of our reality — making the illegible readable, the unclear clear, and the complex comprehensible.

As we stand at the cusp of this legible frontier, it's not just the allure of new tech that beckons but the promise of a reality that’s easier to read, understand, and navigate. It's an old desire to meet a new opportunity. And as history often shows, when the old meets the new in a dance of understanding, the tune often resonates far and wide.

For over a decade in the startup scene, I've been intrigued by a phenomenon I call "Geographic Information Arbitrage." This occurs when one geographic area has more information than another, primarily due to language barriers today. Generally, cutting-edge knowledge lands in English first, creating an information gap.

This gap is an arbitrage opportunity. Especially in Brazil, I've seen startups thrive by either exploring local market inefficiencies or importing ideas from the US, adjusting to local flavors. Many successful startups are essentially localized versions of American ones. Remember Rocket Internet in the 2010s? It brought Silicon Valley ideas to Europe and LATAM.

Another interesting but less discussed side effect of the information arbitrage is the "local influencers". You can carve a niche for yourself in areas like development, marketing, or finance by repackaging US-centric knowledge for the Brazilian market. This holds true for many other local markets, too. Those "local influencers" rarely create truly new information, but due to a gap between the local market and the global ones, they can grow their audiences.

However, the time lag between the US and other markets is shrinking thanks to the internet. What took years for an American startup idea to be mirrored locally now happens much faster.

Yet, the real game-changer is AI. Its enhanced translation capabilities are smoothing out the information bumps, narrowing the arbitrage space, especially in content. Business ideas may still have a time cushion as they require time to execute.

This tech evolution is a double-edged sword. On one side, it's opening global doors for local content creators. Who would have thought that shows from Korea, France, or Spain, like Squid Game, Lupin, and Money Heist, would gather global applause on Netflix? Yet, on the flip side, local "experts" merely repackaging US content will find the ground shifting. The competition gets tougher, but the global stage is now more accessible for genuine value creators.

Writing code is like constructing a complex, multifaceted puzzle. It involves a delicate balance of knowledge, intuition, context awareness, and a constant pursuit of evolution. The arrival of AI-based coding assistants such as ChatGPT brings an exciting prospect. But like all technological advancements, they come with frustrations and limitations that can sometimes overshadow their appeal.

The Frustration of Context

Imagine working with web frameworks like Django, where following the Model-View-Controller (MVC) pattern is the norm. You’re constructing an API endpoint deeply intertwined with various components like models, views, controllers, and sometimes even specialized service functions or classes. This is where the first main frustration with ChatGPT surfaces: the lack of context of the whole code base.

With ChatGPT, you’re essentially working with an intelligent but isolated entity. It's akin to trying to compose a symphony with a musician who can only play one instrument and has no idea what the rest of the orchestra is doing. Since ChatGPT lacks access to all the underlying connections in your project, it often generates incompatible or outright wrong code. You find yourself caught in a loop, explaining the structure, correcting the generated code, and fighting against the tool meant to assist you.

Being Out of Sync with Time

In the ever-evolving world of technology, being up-to-date is not a luxury; it's a necessity. The irony of using ChatGPT to build AI apps is an illustrative case of this frustration. Popular LLM libraries like LangChain and Llama Index were released after ChatGPT's knowledge cut-off.

Imagine the excitement of working with these libraries, only to find that your trusty AI assistant knows nothing about them. You find yourself facing the stark limitations of ChatGPT, a tool otherwise rich in knowledge but now trapped in time.

You could use the browsing feature, hoping it would bridge the knowledge gap, but it's like trying to tune into a live concert through a static-filled radio. It's slow, frequently fails to understand the parsed documentation, and the entire process becomes a trial in patience.

Copy, Paste, Repeat

The inability of ChatGPT to access your local file system brings us to the next layer of inconvenience. This tool, which could be a seamless extension of your coding workflow, cannot create or edit files directly. Instead, it forces you into a monotonous cycle of copying and pasting between ChatGPT and your editor.

What initially feels like a minor nuisance soon morphs into a friction-laden process that disrupts your flow. Every switch from your IDE to ChatGPT becomes a jarring experience, interrupting your train of thought, breaking the rhythm of creativity, and turning a promising innovation into a chore.

The Illusion of Debugging

Debugging is an art and a science, requiring a deep understanding of the code, its variables, and how they change over time. ChatGPT might help you fix your code, but it doesn't truly debug it. The absence of real-time interaction with the code's execution makes ChatGPT's assistance superficial at best. It's like having a co-pilot who can read the map but cannot feel the turbulence or adjust the controls when needed.

Conclusion

ChatGPT brings a fascinating blend of potential and pitfalls. It promises to be a companion in the coding journey but often falls short in understanding the complex tapestry of a real-world project. Its inability to adapt to changes, engage with the context, and truly become a part of the coding ecosystem reveals that we are still at the beginning of this technological exploration.

But, like all beginnings, it is filled with lessons, frustrations, and the promise of what lies ahead. The challenges with ChatGPT are not merely obstacles but signposts pointing toward the future, reminding us that innovation is a journey, not a destination.

WhenX is a tool for creating Semantic Alerts. You can ask a question, and it will search the web and answer it. After that, it will monitor the web and alert you when the answer to your question changes. 

New opportunities arise when the answer to a question changes. For example, when an author releases a new book, the answer to "What was the last book released by author X?" changes, and you may find a new interesting book by your favorite author. Or when the question  "What is the best CRM tool for small business?" may be an opportunity to switch your current system.

Until now, the best way to track new information was to use keyword-filtering systems like Google Alerts or periodically search the web. The same information can be written in many different ways but have the same meaning, which leads to a lot of noise. You still needed to interpret the results and filter the answers because machines could not do that for you. This approach is time-consuming because it generates more noise than useful information. 

The recent advancements in AI changed what is possible. AI systems can understand the human language well enough to answer questions if given enough information. The dreams of the semantic web are now possible.

Using AI, we can move beyond keyword-based alerts. Now we can make filters that use the semantic meaning of text instead of blindly matching keywords. I call this type of systems Semantic Alerts. They use the meaning of a text to alert you to important information. Semantic alerts are more powerful and flexible than simple keyword matching because they allow a smart robot intern to search the web for information and filter it.

I created a proof of concept for creating Semantic Alerts. I call it WhenX.

WhenX is a tool for creating Semantic Alerts. You can ask a question, and it will search the web and answer it. After that, it will monitor the web and alert you when the answer to your question changes. 

New opportunities arise when the answer to a question changes. Never miss out on them again.

You can Sign Up today here and create your alerts.

The Semantic Web has long intrigued me. The idea of transforming the web into a comprehensible database accessible to artificial intelligence, as described in my earlier essays, "How AI will make the Semantic Web possible" and "Using AI to turn the Web into a database", became a foundation for my continued exploration. The notion of creating a smart alert system, a digital sentinel standing guard over the information chaos of the internet, was an application that continuously sparked my curiosity. Imagine a robot that ceaselessly trawls the vast ocean of the web, alerting you to the occurrence of a specific event, say, the release of a new book. This is a dream I have been chasing for years.

The emergence of autonomous agents like AutoGPT and BabyAGI piqued my interest. These systems represented a promising new direction for AI technology. I decided to experiment with these tools, with the aspiration of realizing my dream of a smart alert system. My excitement, however, was quickly tempered. These systems, while advanced, operate with a focus on a goal and its subsequent division into steps. This is a far cry from the continuous, vigilant monitoring system I envisioned. Much to my disappointment, I found their architecture ill-suited for such a task. You can make it work sometimes but not reliably.

The shortcomings of existing models lit a spark to build something new. I decided to create my own system, a system of event-driven autonomous agents. I call it WhenX.

The system's goal is to create alerts in the form of "When X happens, then do Y". In this initial version, we only support email alerts in the form of "When X happens, send me an email report". Making the agents constantly work in the background and monitoring information on your behalf is a small next step towards more autonomous systems.

This is an early experimental project with many bugs! I am still working on it.

You can find the code here!

Architecture

WhenX is a framework that functions as a symphony of three distinct autonomous agents, each playing its part in the grand orchestra of the web. The three main components: the Captain, the Scout, and the Sentinel, are designed to work in harmony. The Soldier, another significant part, acts in response to the Sentinel's alarms.

The Captain is the conductor of this orchestra. Given a mission, it is responsible for assembling a team of agents - a Scout, a Sentinel, and a Soldier.

The Scout serves as the explorer, the one who embarks on the quest for relevant information, creating 'Observations' along its journey. These Observations are the essential raw materials that fuel the WhenX system.

The Sentinel, the gatekeeper, scrutinizes the last N Observations made by the Scout. It is a discerning entity, designed to identify changes, no matter how subtle, in the Observation data. When it detects something noteworthy, it triggers an 'Alarm', signaling the Soldier to action.

The Soldier, the executor, springs into action upon receiving Alarms from the Sentinel. It executes instructions based on the alarms, crafting a 'Report' that encapsulates the detected change or event.

Those agents work together to generate a final report sent to the user.

WhenX stands out because it works based on events, which makes it excellent for monitoring and alerting tasks. This means it can continuously watch over things and alert us immediately when something important happens. This event-driven approach makes WhenX more responsive and accurate in giving alerts, which is a big advantage over traditional task-based AI systems.

While WhenX is still in its early stages, it represents an exciting step towards the realization of a smart alert system, and by extension, a significant stride in the quest to make the Semantic Web a reality. The journey so far has taught me one thing: disappointment can indeed be the mother of invention.

As an avid coder, I embarked on a personal experiment to enhance my coding experience and boost my productivity. ChatGPT was a valuable tool, but copying and pasting generated code into my editor often proved inconvenient and disrupted my creative flow.

I also faced challenges due to ChatGPT's inability to access my local file system and external documentation, as it couldn't utilize my current project's code as context. This meant I had to copy my code to the website for further generation manually.

To overcome these limitations, I created the ChatGPT Code Assistant Plugin. By granting ChatGPT access to both my local file system and the internet, it can now effortlessly:

• You can use online documentation to provide accurate API calls and avoid guesswork.

• Incorporate the code from my current project as context, generating more relevant and coherent suggestions.

• Read and utilize files within my project to improve overall code integration.

• Directly write to files in my project, streamlining the coding process and eliminating the need for manual copy-pasting.

Through this personal experiment, I have experienced the enhanced capabilities of the ChatGPT Code Assistant Plugin and revolutionized the way I code. It feels like having an intelligent and diligent intern coding with me.

This was a simple experiment, but I learned a lot. The chat interface with plugins will be highly flexible, allowing users to customize their experience according to their unique needs and preferences. By integrating various plugins, individuals can adapt the platform to different use cases, making it a versatile solution for both personal and professional communication. Furthermore, the open architecture encourages developers to contribute their own innovations, ensuring that the interface remains dynamic and up-to-date with the latest advancements in technology. In this way, the chat interface with plugins has the potential to revolutionize the way we interact and collaborate in the digital realm.

ChatGPT plugins are a json explaining how to use an API. Developing an API for a ChatGPT plugin is like developing any other API. I used FastAPI to build mine in less than a day. You can find the plugin source code here: https://github.com/edmar/chatgpt-code-assistant

From an early age, I have been captivated by surprises. The thrill of a well-crafted plot twist in a novel or an unforeseen turn in a game made me feel invigorated. As I matured, I realized that "Building Things That Surprise You" became the core principle guiding my work, connecting my various pursuits in life.

Growth and surprise are inextricably linked. Personal growth necessitates change, which requires recognizing our current state as inadequate. This realization prompts us to redefine our role within our perception of reality. Surprise and learning also go hand in hand; we only learn when we encounter the unexpected. Without surprises, we would never expand our knowledge.

As I embarked on a journey to find meaning in my work, I discovered that surprises are more than just sources of pleasure and learning—they can also facilitate meaning through infinite play.

Infinite Play

We are all participants in various games, such as wealth accumulation, career advancement, and the pursuit of fame and status. Games can be categorized as finite or infinite. Finite games are played to win, while infinite games are played for the sake of continued play itself.

Finite players frequently alternate between attempting to win a game and seeking new games upon victory. This restless search can be challenging, leading to feelings of emptiness once a goal is achieved. By embracing infinite games, I found meaning in my work.

Infinite players focus on prolonging play, as exemplified by Walt Disney's famous quote, "We make money to make movies, not make movies to make money." According to James Carse's book "Finite and Infinite Games," surprise is integral to both types of games. Infinite players embrace surprise as a reason to continue playing, while finite players rely on surprise to keep their opponents on edge.

Manufacturing Surprise: Creating Meaningful Experiences

There are three primary ways the things we create can surprise us:

1. Unanticipated uses by others: People often find unconventional applications for products, such as fan theories about books or TV shows, which spawn dedicated forums and YouTube channels.

2. Discovering surprises during the creation process: Research and writing can lead to unexpected findings, deepening our understanding of a subject and fostering personal growth.

3. Self-generating surprises: Organizations, movements, and AI systems can follow unforeseen trajectories, providing a constant source of surprise and engagement.

AIs as Surprise Engines

Traditional software development demands precise instructions for computers to follow. However, machine learning algorithms "learn" to achieve success by analyzing data, allowing for more unpredictability and surprise.

Large Language Models (LLMs) are particularly surprising due to their emergent behavior, which can't be fully predicted. The capabilities of systems like GPT-3 extend beyond their initial training, often resulting in unexpected applications. Researchers have even found ways to improve LLM performance in logic tasks through prompt adjustments, demonstrating AI's potential for surprise.

As AI systems become more complex and capable, their ability to generate surprises will grow. Building AI systems may become the ultimate source of surprise, driving creativity and innovation.

Our fascination with surprises is deeply rooted in personal growth, learning, and the search for meaning. Embracing infinite games and harnessing the power of AI as a "surprise engine" can foster creativity, innovation, and a deeper understanding of the world. By pursuing surprise, we not only build a more dynamic future but continue our quest for growth, knowledge, and meaning—a true testament to the human spirit and our innate desire to explore the unknown.

The AI wars are heating up. Google and Microsoft are in an open war for the future of many industries, including search. Apple remains quiet even with all this noise, but they are well positioned to win big in the age of AI.

Apple seems to have fallen behind the other giants when the topic is AI. Siri is a joke if compared with ChatGPT. There are rumors that a new version is in the works, but I would not hold my breath for it. I don't think they will be able to compete with the likes of OpenAI so soon. Even Google is struggling to keep pace. 

The success of ChatGPT shows the big user interest in personal chat assistants that work. The idea is not new and has been tried many times but never cough on. Siri was released in 2011, for example. The value proposition of an intelligent assistant is compelling, but the execution is extremely hard.  Now that OpenAI proven it's possible, many companies will follow. Apple will be one of them. 

Assistants will be a huge software category, as big as Browsers are today. This is significant because whoever controls the Assistant will control a big chunk of user interactions because it is upstream from almost everything. You would interact with your computer by asking it to do things for you. 

Here are the current platform layers:

App > SO > Hardware

Each upstream layer has a lot of power over the downstream layers. This is why it was so important for Google to have Chrome and Android. Since apple controls Safari which is the main Browser in iOS, Google needs to pay them U$15B per year to be the default search engine. 

With AI Assistants becoming the main interface between users and their devices, we will have an extra layer:

We are seeing the beginning of this with ChatGPT plugins. Instead of searching and clicking to book a flight ticket, you can ask ChatGPT to do it for you. The agent will sit between you and other services, doing tasks on your behalf. Google is such a powerful business because it works as a middleman between you and the web. For many of your interactions with the web, Google is the first step. Companies that control the AI Agents will be even more powerful because they will control the full workflow, not only the first step. 

The three companies best positioned to win the AI assistant race are Google, Apple, and Microsoft. Each of these tech giants possesses a distinct advantage in that they own and develop popular operating systems—Android, iOS, and Windows, respectively. This control over operating systems allows them to deeply integrate their AI assistants, providing seamless and native experiences for users. As a result, Google Assistant, Apple's Siri, and Microsoft's Cortana have the potential to become integral parts of users' daily interactions with their devices. This deep integration also facilitates the collection of valuable user data, which can be used to improve AI algorithms and personalize experiences. However, each company must overcome unique challenges, such as competition, privacy concerns, and user experience, to ensure that their AI assistant becomes the go-to choice for users in an increasingly competitive landscape.

Google, despite being a leader in AI technology, is facing challenges keeping pace with OpenAI and its groundbreaking ChatGPT. Google's own AI assistant, Google Assistant, has proven to be a powerful tool, but the company's recent competitor to ChatGPT, dubbed Bard, appears to lag in performance. As Google confronts the hurdle of staying competitive with OpenAI, it must leverage its vast data resources, comprehensive ecosystem, and commitment to research and development to refine Bard and bridge the gap. To reassert its dominance in the AI assistant landscape, Google needs to address privacy concerns, enhance user experience, and build strong third-party integrations while also focusing on the development of a more advanced, competitive AI assistant.

Apple, despite currently lagging in AI advancements, remains a formidable contender with its strong hardware-software integration and a loyal user base. While Siri may not currently rival ChatGPT, Apple's history of innovation suggests that they could still make significant strides in AI assistant development. To become a frontrunner, Apple must refine and revolutionize Siri, capitalize on its existing ecosystem, and prioritize privacy and user experience to capture the market's attention.

Microsoft, a tech giant with a strong presence in the enterprise market, has strategically invested $10 billion in OpenAI for a 49% stake, positioning itself to benefit from OpenAI's cutting-edge technology. This partnership enables Microsoft to integrate OpenAI's advanced AI capabilities, such as ChatGPT, into its suite of Windows and Office products. By providing seamless, business-focused AI solutions and leveraging its established relationships in the enterprise sector, Microsoft has the potential to promote widespread AI assistant adoption. To succeed, the company must continue enhancing its AI offerings, tailoring solutions to both consumer and enterprise needs, and building strategic partnerships that enrich its AI ecosystem and empower users with powerful AI-driven tools.

In conclusion, the AI wars have intensified as tech giants Google, Apple, and Microsoft vie for dominance in the burgeoning AI assistant market. As AI assistants become increasingly integrated into users' daily lives, the competition to become the preferred choice will only grow fiercer. Google, with its vast data resources and ecosystem, must address privacy concerns and improve its AI offerings to compete effectively with OpenAI. Apple, while currently lagging in AI, can leverage its hardware-software integration and focus on user experience to make a comeback. Meanwhile, Microsoft has strategically aligned itself with OpenAI, positioning the company to capture the enterprise market with powerful AI-driven solutions. The AI assistant landscape is evolving rapidly, and the company that can successfully address user concerns, innovate, and provides seamless, transformative experiences will ultimately emerge victorious in this high-stakes race.

Advances in Generative AI will decrease the skill level necessary to produce quality results in many areas of creative work like drawing, writing, and programming. What happens when AI can achieve high technical skills? Taste becomes a superpower.

Creative work requires three ingredients: creativity, skill, and taste. Creativity is your ability to come up with novel ideas, Skill is your capacity to execute those ideas, and taste is your ability to judge the quality of work in your area of expertise.

These three ingredients can be developed at different levels in creative work. For example, you can develop excellent taste in literature without an analogous level of skill in writing. They also tend to overlap. As you develop your skill, you will also gain insights into new ideas, for example.

Generative AI will soon be able to achieve a master skill level in many creative work areas, for example, drawing. It already draws that would require a person many years of intense practice. This will lead to a democratization of access to drawing skills. However, access to skills is not enough to create great work.

Imagine you become a strong draftsperson, adept at creating art, but your ideas are crap or run of the mill. You need originality and taste to set you apart. Originality to generate new ideas and taste to filter what you show to the world.

In the same way, the work of an editor is important for a writer, your Taste will be important for working with AI. You will need to judge the output of Generative AI systems and ask for refinements, changes, or even for entirely new generations. Without taste, even with a great AI capable of superhuman skill, your results will be average at most.

Taste allows you to judge the quality of your work, but using Generative AI, it could also serve as a creativity enhancer. How could we use Taste as Creativity? We can use AI to generate random versions of something and use our Taste to give feedback for each generation, then the AI can use our feedback in the next batch of generations. Imagine you want to generate a novel, interesting drawing, but you don't have any ideas. In theory, you could use AI to generate a set of random drawings and give feedback on what you like and dislike and keep generating new images until you come up with something interesting.

People tend to develop taste faster than skill. This creates what Ira Glass calls the "Taste Gap".

Nobody tells this to people who are beginners, I wish someone told me. All of us who do creative work get into it because we have good taste. But there is this gap. For the first couple years you make stuff; it’s just not that good. It’s trying to be good, it has potential, but it’s not. But your taste, the thing that got you into the game, is still killer. And your taste is why your work disappoints you. A lot of people never get past this phase, they quit. Most people I know who do interesting, creative work went through years of this. We know our work doesn’t have this special thing that we want it to have. We all go through this. And if you are just starting out or you are still in this phase, you gotta know its normal and the most important thing you can do is do a lot of work. Put yourself on a deadline so that every week you will finish one story. It is only by going through a volume of work that you will close that gap, and your work will be as good as your ambitions. And I took longer to figure out how to do this than anyone I’ve ever met. It’s gonna take awhile. It’s normal to take a while. You’ve just gotta fight your way through.

Your capacity to judge the quality of your work increases faster than your capacity to produce quality work. Someone who decides to be a writer tends to be someone who reads a lot of great books. This develops your taste, and then when you write your own book, it will not initially meet your quality bar. You can only bridge the "taste gap" by doing the work and improving your skill.

The taste gap is the gap between what your produce and your quality bar. Generative AI could shorten the "taste gap" by providing readily available skills for cheap. I see this playing out in two possible scenarios.

The first scenario is that people with good taste will be able to produce great work faster than before in areas in which skill acquisition is harder than taste development. The number of people capable of producing great work would increase, leading to a new golden era of creative work production.

The second scenario is that Generative AI will decrease most people's motivation to pursue hard skills by reducing the skill necessary to produce creative work. Since you can "shortcut" your path using AI, people would gravitate towards using it instead of the long years of skill development. This could lead to a reduction in highly skilled people. Since Skill, Creativity, and Taste work together, having less skilled people would damage the future production of great works.

Developing your Taste is a good course of action to futureproof your career. Independent of where technology goes, having good taste will put you on the path of creating great work.

Coding is one of the first areas where LLM (Large Language Models) based products have found product/market fit. Cursor is a huge success; not only that, one of the most prominent use cases of ChatGPT is helping with coding tasks. Those systems can't substitute a programmer yet, but there is a clear trend of improvement in systems capable of writing code. If this trend continues, the cost of producing software will decrease. What happens when the cost of producing software goes to zero?

Intellectual Property is a powerful economic engine. Famous characters and stories have generated billions in revenue for IP owners. J.K Rowling, the creator of Harry Potter, is richer than England's late queen ever was. Writing could be a good place to understand a world of zero-marginal software.

Writing is one of the few professions in the modern world where a single individual can still make a huge impact. In most areas, you will need a team, sometimes a big team, to make something impactful. In writing, a single individual can produce a masterpiece. 

We have examples of small teams producing great outcomes in the software world. A good example is Whatsapp. It has a team of 55 serving 900M users when sold to Facebook for US$19B. This remains an exception, but we are moving towards becoming the rule.

Software will become more like writing. A single developer will be able to have a big impact. This has happened to some degree in the past. At the beginning of the PC revolution, many apps were created by single developers or small teams. Something similar happened during the initial mobile era. With increased competition and the maturing of platforms, it became harder and harder for solo developers to generate revenue.

We need to consider a distinction between the writing world and the software world. Stories and characters are unique and protected by IP. You can't simply create a new Harry Potter novel without permission. You can, however, create software that delivers the same functionality as other software. 

In software, utility is the most important factor. If two apps deliver the same utility, they are interchangeable. This is not true for stories, you can choose Lord Of the Rings over Harry Potter, but there is no perfect substitute for either since what you buy is the experience. No two books will ever feel the same. However, If you can reproduce the utility of any software at no cost, you can substitute it effectively. 

Convenience and differentiation are the main forces protecting value capture in a zero-marginal software world. 

Convenience

A good framework to think about convenience is piracy. Before the age of mobile and app stores, piracy was a big issue in the digital world. It was not completely obvious that people would prefer to buy digital products like software and music instead of pirating them for zero cost. Many court battles were fought over IP piracy. In the end, convenience and usability won the day. It's far more accessible and safe to buy a subscription to Spotify than to download and organize music by yourself. Another force that changed the game was cloud computing. Pirating software running on the cloud is far harder than one running locally. Today, piracy is still an issue but a shadow of the issue it was in the past. 

When software goes to zero, Convenience will become a function of Price and Complexity. The more complex a software is, the less convenient it will be to create it through AI. Even if you only need to describe what you want for an AI, if it's complex, it will take your time actually to get it to work. Price is the second important factor. The higher the current price of the software, the bigger the incentive to use AI and build your version. 

To be competitive, a software will need to strike a balance between complexity and price so that it would not be convenient enough to build a substitute using AI given the price of the market solution. 

Differentiation

Another framework to think about zero marginal software is differentiation. Software companies must continue to differentiate to be competitive against copycats. Once software starts to be successful, people will try to copy it. There are many potential sources of differentiation, like network effects, distribution, and switching costs. Network effects is why products like social networks and marketplaces can resist competition. The more people in the network, the more value you have. You can copy the software, but you can't copy the network of people using the product. 

Companies with strong differentiation will not be heavily affected by the cost of producing software going to zero. This is already a reality in the B2C software world. Successful B2C companies rely on powerful differentiation effects besides the pure utility of the software provided.

Some types of differentiation could lose power. Sales-driven differentiation in B2B could take a hit. Zero Marginal Cost software will empower internal tools teams to create solutions on par with market ones. Switching costs also can become less powerful since zero-cost software will mean zero cost data migration as well.

Winners 

The first winners of the zero marginal software world will be consumers. Software will become abundant like content is today. Anyone with a computer will be able to create software like anyone with a computer today can write a blog post. Software today is extremely hard to customize. In a zero-marginal software world, consumers will have the option of highly customized software for their specific needs. 

Operational Systems (OS) will initially benefit by delivering more utility out of the box. However, there is the possibility of reduced network effects. Current Operational Systems like Windows and iOS benefit greatly from network effects. The more apps you have, the more valuable your OS. It is hard to port software from one platform to the other, but with AI-generated software, these costs can go to zero. This implies the winner takes all dynamics in software platforms could change. 

Software is the natural complement of Hardware. With software costs going to zero, the utility of any kind of computer will grow exponentially. There is a chicken and egg problem when you introduce new devices. The iPhone would not be that useful if there were no apps.

Games are notoriously hard to develop. To produce even an indie game is a tremendous endeavor. You need to put in thousands of hours in coding and to make assets to produce a competitive game. Not only that, you need to combine many disciplines as well. Building a game usually involves sounds, music, visual assets, coding, animation, etc. A zero-marginal software world means small teams or even single individuals can build AAA games. This will completely change the Gaming industry.

Losers

Not everybody will win in a Zero Marginal Software world. Tools that provide simple utility will be easily cloned unless the convenience of getting them is bigger than just creating your own version. Many small apps will be hit hard by this change.

The B2B SaaS market size is huge. There is a myriad of software used to run the day-to-day operation of companies, for example, CRMs, ERPs, and ATS, to name just a few big ones. A trade-off in the B2B world has been the "Build versus Buy". A company can decide to create its own software to its needs or use a market solution. The SaaS wave of companies turns the scales in favor of buying readily available solutions rather than the painful effort of building them yourself. Because of this, we saw an explosion of B2B Saas companies in the last decade. 

Zero Marginal Software will pressure down the maximum price B2B software companies can charge. The ability to make software at zero cost will tip the scales to the side of "build" in the "Build versus Buy" decision. Before the SaaS era, the bias was to build internal tools, and it took a lot of evangelizing sales to get us to the current era. Now, B2B sales organizations will need to fight not only external competitors but internal teams building customized tools.

These dynamics will push B2B software companies to differentiate through methods similar to B2C companies. One big example of this is network effects through data aggregation. Being able to connect and anonymize multiple customer data will lead to a golden age of benchmarking and improved decision-making. 

Conclusion

AI-generated code could lead to a world of zero marginal cost software. End-user "programming" through AI and internal tools teams in companies will emerge as a powerful force in the software industry. Convenience and Differentiation still will be the main force driving value creation for companies building software. Still, companies will need to work hard to deliver value above end-user custom solutions.

I didn't discuss in this piece another interesting possibility. That software, as we know it, will not exist at all, but this is a topic for a next essay.

AI hype has grown past crypto hype now. It is almost impossible not to compare the two. Some people are dismissing the current AI hype cycle as the next bubble. I know, “The four most expensive words in the English language are: This time is different.” However, this time it IS different. Crypto and AI are different in an important way: Use case matters. 

There are three primary risks we can have with a new startup, and those risks can be combined into risk profiles as we discussed in On the Risk Profiles of AI Startups:

Feasibility Risk: Can we build it?

Viability Risk: Can we generate value for the business?

Value Risk: Can we generate value for the users? 

The value question tends to be the most challenging part for software startups and crypto projects as well. Most crypto projects suffer from a lack of real use cases. Ask yourself how many crypto projects are better than non-crypto ones. Many crypto projects are just existing use cases plus "decentralized" or "tokenized".  The fact that a company is public, meaning that anyone can buy shares in it, doesn't mean anything about the value it provides to users. 

The first distinction of AI startups is that it is easier to answer the value question. You can know the utility of an AI project by directly using it. Let's, for example, think about image generation. It's a cool technology, and there is a lot of hype. However, the utility is straightforward. You can use it to illustrate a blog post, create a new comic book, or an ad banner. These are activities that people already do every day in many sectors of the economy. 

Another big difference between crypto and AI is the speculative dimension. Most of the crypto engagement came from speculation and not usage. This is a massively important distinction. There was more engagement in betting on tokens and currencies than using "Web 3.0" products. AI startups tend to follow the more traditional form of funding, which makes it harder to speculate since you don't have tokens being offered in the market. There will be many speculative bets on AI, but it will not reach the same level as crypto. 

Many speculative bets in Web 3.0 failed for lack of value, which means that the end product of most projects didn't deliver enough value to justify the investments. In AI, we will see a different kind of speculative bet. We will see many speculative bets on feasibility, which can generate a hype bubble. 

Theranos is a good case of the hype that could generate trouble in AI. Another less damaging example is Magic Leap, the creator of Augmented Reality glasses. The company raised more than a billion dollars before releasing a product. There was a lot of hype around how the company would revolutionize everything. This hype can be built because the product is still in development/research, and everything is possible. This could also happen with an AI startup. Promises of AGI and disruption can be made with initial technical success, but delivering on those promises could be harder than initially expected. 

AI has stronger use cases than crypto. There will be some hype and a small bubble built around AI startups. However, the smaller potential for speculation will make the frauds and spectacular failures the exception, not the rule.