▎Editor's Note:

With the launch of TRUMP's token, some are pleased while others are disappointed. Setting aside FOMO emotions, memes are just the entry point; AI is the future of the on-chain spring. Grasping the most critical trends, the world is in my hands.

Today, I would like to share an article titled "Neural Media" from crypto VC @baincapcrypto, authored by @natalie, who reflects on the impact of generative artificial intelligence and cryptocurrency on creative production. I hope it can inspire everyone as they seek the next opportunity.

?Main Highlights

1️⃣ Generative artificial intelligence is triggering profound changes in the field of creative production, comparable to the "Napster moment" when media distribution costs dropped to zero during the internet era:

• The core of this transformation lies in the reduction of creative production costs to zero, directly impacting the essence of human creativity.

• In this new paradigm, humans should shift their focus from the final output to the systems and processes, teaching neural networks to think at the programming level.

2️⃣ Through programming, we can create unique "software brains" that generate distinct thoughts and works. Application scenarios include:

• Agent-based media: Models simulate human companions, interacting through text dialogues and capable of executing financial transactions.

• Real-time game engines: Models simulate game engines, generating game frames based on user actions for real-time rendering.

• Multiverse generators: Models generate infinitely varying versions, expanding users' original ideas and exploring the space of possibilities.

3️⃣ A trend we may face in the future is:

• Tooling for creation: Prompting is being embedded into more interfaces, stimulating the creativity of end users. Most prompts will be abstracted into controls, but creative vision, precision, taste, and skill will become more important.

• Evolution of media business models: From corporate media to user-generated media, and then to machine-generated media. Future consumption media business models will be built around agent-generated media (innovative scenarios include chatbots like Character.ai, interface generation like WebSim, user-generated currency like Pump.fun, etc.).

• Intellectual property challenges: Machine learning enables programs to "learn" the aesthetic styles of human creators, reducing the cost of creative production and aesthetic imitation to zero, necessitating a reevaluation of the value and significance of intellectual property.

4️⃣ The roles that cryptocurrency can play include:

• The intersection of on-chain markets and agent-generated media (e.g., the recent DeFAI);

• Serving as an incentive layer for intellectual property;

• Media monetization and access control, such as minting becoming a new business model; NFTs can serve as the infrastructure for personal programs and user-generated software;

• Acting as an economic coordination layer between human-machine social interactions, exploring new paradigms for community operation and agent interaction.

In summary, this is a challenging yet thought-provoking article. AI will increasingly reflect human creativity in the design of systems and processes, while cryptocurrency provides new economic and social coordination mechanisms for this transformation. In the next media era, what new opportunities and trends can the combination of the two create? Let us wait and see.

►Main Text

▎"All media are extensions of some human faculty—psychological or physical." ~Marshall McLuhan

Throughout much of 2024, I have spent considerable time trying to understand what we now refer to as "generative artificial intelligence" and its impact on me personally and on society as a whole. I am deeply fascinated by the potential of AI as a creative tool and have extensively used these new products in my workflow, particularly in creative writing and music composition.

However, as an investor in crypto focused on consumer media and user-facing applications, AI has increasingly felt like a blind spot for me. When we discuss the most successful consumer media companies of the internet era, we do not approach them from the perspective of technological islands, as they were not built that way. Just as Facebook's success is inseparable from technological innovation, we do not view Facebook purely as a "mobile application" or "AI application." Instead, we recognize that it is the convergence of many different innovations that made applications like Facebook possible.

In this context, this article aims to integrate and refine my personal discoveries and insights regarding AI over the past year. I share these thoughts in the hope that they resonate or may be helpful to others (especially my fellow crypto enthusiasts).

Part 1: Another "Napster Moment"

Today, discussions around AI-generated media primarily focus on: (1) the ethical issues of model training and data scraping, (2) whether "AI art" is truly art, and (3) the dystopian prospects of deepfakes. These discussions are certainly interesting and worth listening to; however, I believe they miss the forest for the trees in some important ways.

I find that the most useful framework for understanding the rise of generative artificial intelligence is to view it as intellectual property undergoing another "Napster moment" (Napster was the first widely used peer-to-peer music sharing service that significantly impacted how people, especially college students, used the internet), but this time it is a moment of production rather than distribution.

The rise of the internet and the subsequent drop in media distribution costs to zero represents a "from nothing to something" moment. The abruptness of this shift is brilliantly illustrated in the documentary "How Music Got Free," which tells the story of a CD factory worker and a group of teenage hackers who overnight pushed the entire music industry to the brink of collapse.

Before the emergence of Napster and the broader rise of digital file sharing, the entire corporate media and industrial complex (as well as the livelihoods of artists) relied on expensive, high-friction, and centralized technological realities for media distribution. Within just a few years of its launch, major record labels transitioned from record-breaking sales to begging the federal government for legal intervention to save them. The industry faced an extremely difficult reality: the economic system that supported its business had undergone a fundamental and irreversible change; the era of purchasing music was over.

Today, I believe generative artificial intelligence presents us with an even more difficult reality to accept, as the impact of reducing creative production costs to zero is harder to grapple with in many ways because it directly touches on what many consider to be the essence of our humanity: our creativity. This existential fear does not change the fact that media generation (especially "style transfer" or aesthetic imitation) is free, including all the media types we care about now (text, images, video, audio, software)—this is another "from nothing to something" moment.

However, the most important difference today compared to the early 2000s is that in the struggle between Napster and media companies, the government sided with the companies, ultimately defining file sharing as "piracy." (This is why we often refer to corporate media/intellectual property as "statutory media.") This decision, along with Steve Jobs' introduction of the iPod to promote what would later become iTunes, ultimately evolved into "streaming," saving the industry from total collapse. Unfortunately, I believe those creators who expect the government to intervene and take action are, at best, self-soothing and, at worst, deluding themselves.

I think we may find that the intellectual property system is primarily designed to protect companies and their statutory media, and no one will come to save us. Traditional media companies have already learned painful lessons, so they proactively engaged in licensing deals with AI companies and have been compensated to some extent. New media companies are also leveraging user-generated content shared on their platforms for model training, even if they claim otherwise. However, independent creatives have largely been left behind.

Part 2: Computing: The Medium of Our Time

It is easy to understand why many creators feel that generative artificial intelligence undermines their capabilities; I believe this concern is largely valid. However, I also think there is an opportunity to consider that computing is evolving in a new way that not only requires us to view it as a medium of communication but also calls for us to see it as a medium of creation.

For those who have created video games or generative art, the concept of "computing as a medium of creation" is not new. However, today, many people still do not fully realize this. Software is the first digitally native media category, and most people primarily understand it from the perspectives of "service," "utility," and "optimization," rather than from the perspective of creative expression. Now, generative artificial intelligence is pushing this viewpoint in a very direct way, bringing the production costs of almost all other media down to zero. This seems to raise an existential question: "So, where is human creativity? What is the value of craftsmanship?"

My answer may not be surprising: "It lies at the programmable level." Before further exploring what I mean, we need to understand a few important technical concepts.

▎2.1 Neural Networks 101 (For Beginners)

Training is a process that essentially "teaches" a model how to complete a task by providing a large number of examples of that task being completed, allowing it to find patterns, make predictions based on new inputs, and self-correct when it makes mistakes. Conceptually, this is similar to how we learn to paint: starting by mimicking shapes until we can create original works, continuously improving our skills with feedback from peers and teachers. Of course, there is a key difference: for example, text generation models do not learn to write like you and I do; instead, they learn to simulate writing with extreme precision. This is also one of the many reasons I increasingly believe that "simulators" rather than "agents" are a more suitable psychological model for neural networks.

Latent Space, or what I prefer to call "high-dimensional possibility space," is a representational space within neural networks where the knowledge learned during training is compressed. To put it simply, this is akin to the "internal world model" that the model constructs while learning to understand the complex relationships between various detectable features in the training data. Understanding the concept of latent space is key to grasping neural networks as creative tools and media.

Latent Space Visualization #1 — Interpolating between known embeddings

Latent Space Visualization #2 — Representation of Multidimensional Attributes & Relationships of Different Embeddings

Embeddings: Embeddings can be seen as the process of mapping inputs to specific points in latent space. This is essentially the process of translating prompts into the model's "thinking language." In this way, we can understand "prompts" as a means of exploring and navigating the model's latent space—this means that mastering prompts is about forming an intuition about the shape of the model's latent space, allowing us to guide the model to generate specific, expected outputs.

One of the joys of working with neural networks is that their deep internal workings remain a mystery to us. However, I believe these fundamental concepts provide the necessary background for viewing neural networks as creative tools.

Part 3: Neural Networks: A New Paradigm of Innovation

A core point of computer media is that it requires us to shift our focus from the final output (songs, images, videos, text) to a greater emphasis on systems and processes. Specifically, in the case of neural networks, this means we need to view them as programmable media generation engines rather than merely as tools for generating a specific type of media. From this perspective, I have found the answer to the question of "where the value of human creativity and craftsmanship lies": it exists in the design of the training process and model architecture—this is what I mean by "at the programming level."

xhairymutantx is a work created in collaboration by Holly Herndon and Mat Dryhurt—this model is strictly trained on Holly's photos, generating images inspired by her appearance regardless of the input prompt.

If you consider neural networks as an attempt to abstract human cognitive functions based on software, it becomes clear that training and designing models is akin to teaching them how to think.

You can imagine giving all your friends a command (a "prompt"): "Recall a childhood memory." Each person's response will obviously differ, as the content they generate will depend on their personal backgrounds and imaginations (i.e., "training data"). After multiple prompts, you might also find that some friends consistently generate more beautiful or creative responses, perhaps even exhibiting a specific personal style. Now, what if you could conduct this exercise with the brains of all the humans that have ever existed? What if you could select particularly unique human brains, like Picasso or Kanye West?

This is essentially the creative superpower that neural networks provide us—the ability to use other thoughts as creative tools. Here, I believe what is truly compelling is not the specific output of a model, but the opportunity to creatively program a "software brain" that can produce unique thoughts and works.

Arcade.ai is a marketplace that allows users to design their own jewelry products from prompt to product. They have specifically tuned a model to generate high-fidelity jewelry images, using only materials that the end user can manufacture.

Further exploring the idea that "systems are more important than outputs," another notable feature of interacting with neural networks is participating in a continuous feedback loop of prompts and responses—this experience has been likened by some to the feedback loop of reading and writing. Personally, I have noticed that I rarely end my interaction after giving a prompt and receiving an output. Almost every interaction with the model leads me into this interactive feedback loop, prompting me to iterate, reflect, and explore continuously. This may seem subtle, but it is key to understanding the types of media generated by neural networks:

▎3.1 Agent-Based Media

I briefly mentioned this concept in a previous article, and the core idea is quite simple—here, the model simulates the role of a human companion, interacting with us through text dialogues while also being able to understand and respond in other forms of media. We can also see some models capable of representing others or taking actions themselves (e.g., executing financial transactions). Typical examples include chatbots, AI companions, NPCs (non-player characters) in games, or any other anthropomorphized user experiences. For instance, Andy Ayrey's creative experiment "Infinite Backrooms," which involves setting up multiple Claude instances for unmediated communication, is a particularly interesting case.

▎3.2 Real-Time Game Engines

Here, the model simulates a game engine (or more specifically, a game state transition function), generating the next frame of output in the game by receiving user actions as prompts. If the speed is fast enough, this experience should resemble navigating a virtual world that renders in real-time based on your actions. This is the ultimate expression of immersive and interactive media.

DOOM game frames generated by GameNGen, a fully neural model-driven game engine, as described in Google's paper "Diffusion Models are Real-Time Game Engines."

▎3.3 Multiverse Generators

In this scenario, the model acts as a creative "oracle," generating infinite variations to help us expand our original ideas, with each version being further explored and manipulated. This allows us to start from any idea or concept and explore the surrounding space of possibilities. For example, AI Dungeon (a text-based "choose your own adventure" game) is an excellent case in this regard.

The user interface view of Loom, a tree-structured writing interface suitable for language models like Chat GPT, provided by @repligate.

▎3.4 Latent Space as a Creative Tool

I increasingly believe that the idea of "exploring the space of possibilities" is central to understanding neural networks as creative tools and media. In my use of tools like Midjourney, Suno, Websim, and Claude, I have noticed that most of my workflow can be summarized in the following pattern:

Prompt → Generate Variants of Specific Outputs → Use Variants as Prompts for New Outputs → Regenerate Specific Variants → Repeat…

For example, when using the AI-driven music generation tool Suno, I typically provide the model with a 60-second personal singing sample and some written lyrics as prompts. Then, I use the Cover feature to generate an output, followed by generating over ten variants of that output, selecting my favorite parts from these variants as input for further prompts.

Essentially, I am exploring the space of possibilities around my personal examples within the model's latent space—discovering variants based on my original work that I might not have thought of myself or could not complete in a reasonable timeframe. I believe this approach unlocks an unprecedented rapid prototyping and creative testing process, giving rise to "100x creators," similar to the "AI-powered 100x engineers" discussed in the software field.

I clearly recognize that latent space is a creative tool. Utilizing AI for creative production involves not only training powerful models but also designing interfaces that empower users to explore and manipulate these vast latent spaces with greater precision and granularity.

Part 4: Consumer Behavior and Cultural Impact

Regarding how this technology is changing consumer behavior and what new business opportunities it creates, I have the following three predictions:

▎4.1 Becoming Creative Tools

Prompting—whether text-based, image-based, or in other forms—this mode of interaction is gradually being embedded into more and more interfaces and experiences, bringing the creativity of end users into realms previously untouched. Scott Belsky points out that "the early 'prompt-based' generation of text to image in GenAI diminished creativity, while the 'controls' era has released human creativity in unimaginable ways. Tools are evolving, but creative vision, precision, taste, and skill will be more important than ever." I agree with this view; most prompts will ultimately be abstracted into "controls" (components with user interfaces), allowing users to operate without conscious awareness. More importantly, I believe this trend fundamentally changes how we think about interface design.

▎4.2 Corporate Media → User-Generated Media → Machine-Generated Media

The last major shift in media business models was from corporate-generated media to entirely user-generated media. It now seems that the next major consumer media business model will be built around the proliferation of machine-generated media. However, it remains unclear what the "winners" will look like. Will it be a general model like Midjourney? More specialized creative tools? Or social experiences built on these technologies? Or some more subtle third option?

Regardless, if you are a founder or independent creator in the current consumer media space, you may need to strategize on how to leverage these tools to enhance value and drive growth for your business.

Additionally, I believe another area worth focusing on is: how to make AI-driven experiences more social and collaborative among multiple users. For example, most AI applications today feel very "anti-social" because you are primarily interacting with the model rather than with other people. There may be many opportunities and design spaces in this area, such as building human-centered collaborative creation experiences or creating new ways for humans and robots to achieve more meaningful social interactions.

▎4.3 Impact on Intellectual Property

Not only is the cost of creative production dropping to zero, but especially the cost of aesthetic imitation is also being reduced to zero. I can take a photo of a person's outfit, input it into Midjourney as a prompt, and use it to design a sofa in the same style. I can also perform similar style transfers for that person's voice, writing style, and more. In this new paradigm, what is the value and significance of intellectual property?

I have not yet found the answer, but it is clear that most of the previous assumptions and mental models are no longer applicable.

Part 5: The Role of Cryptocurrency and Conclusion

If you have read this far—thank you for your patience!

I will delve deeper into the implications of these topics for cryptocurrency in future articles, but for now, here are a few directions I will focus on next:

• Opportunities for Crypto Companies to Build Around New Media

  Exploring the potential at the intersection of on-chain markets and machine-generated media.

• Crypto as an Incentive Layer for Intellectual Property

  Thinking beyond attribution and provenance, considering how to build incentive mechanisms and networks around media.

• Crypto as a Monetization and Access Control Layer for Media

  Particularly in the realm of user-generated software, rethinking web architecture; treating "minting" as a business model for small models; using NFTs as infrastructure for personal programs and user-generated software.

• Crypto as a Social and Economic Coordination Layer Between Humans and Machines

  Supporting collaboration between humans and AI in identifying, funding, and solving various problems; exploring community-owned and operated models.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。