On Markets and Milkshakes and Old Ghosts: Is AI On The Same Cycle as 3D Animation? The Milkshake IS the Thing
I think we're watching the end of the "one AI company wins everything" narrative.
Instead, I think the market is separating into three different businesses, each with different economics and different winners.
1. Frontier Intelligence Companies
OpenAI, Anthropic, Google DeepMind, xAI, and perhaps one or two Chinese labs.
Their business is no longer merely "selling LLMs."
Their product is
- continual pretraining
- reinforcement learning
- reasoning
- multimodal perception
- agents
- evaluation
- alignment
- inference optimization
Their moat is billions of dollars of compute and enormous proprietary datasets.
This is becoming increasingly similar to semiconductor fabrication.
Not everyone can build an advanced fab.
Not everyone can build frontier models.
2. Infrastructure Companies
These may become even more profitable.
Examples include NVIDIA, AMD, cloud providers, inference engine developers, networking companies, storage vendors, model serving companies, and vector database companies.
Every new open-weight model increases demand for infrastructure.
This is similar to the California Gold Rush.
The people selling shovels usually made steadier profits than many of the miners.
3. Application Companies
This is where I think the biggest change is occurring.
Originally people thought:
AI is the application.
Increasingly it looks like
AI is an endpoint inside an application.
That is a profound architectural shift.
Endpoint AI
This is remarkably close to how you've described MCCF over many conversations.
The world simulation is primary.
The AI is one service among many.
For example
Scene Graph ↓ Physics ↓ Audio ↓ Lighting ↓ Animation ↓ Game Logic ↓ Network ↓ LLM
The LLM is just another subsystem.
If it disappears temporarily...
the world still runs.
That is exactly how game engines already work.
Contrast that with
User ↓ LLM ↓ Everything else
where every decision routes through the model.
That architecture is much easier to build initially, but it inherits the LLM's latency, cost, context limits, and failure modes.
Cloud AI
Advantages:
Unlimited compute
Newest models
Shared learning
Large context
Continuous updates
Disadvantages:
Latency
Privacy
Cost
Vendor lock-in
Requires Internet
Token billing
Local AI
Advantages:
Nearly zero marginal cost after hardware
No network latency
Works offline
Private
Persistent state
Can integrate deeply with hardware
Disadvantages:
Memory limits
Power consumption
Model size limits
Updates are your responsibility
Hardware cost
Interestingly...
the economics favor local inference much more than local training.
Training scales roughly with
O(N × D)
Inference scales roughly with
O(N)
where N is parameter count and D is dataset size.
Training frontier models remains extraordinarily expensive, while running a distilled or quantized version locally has become feasible on consumer hardware.
Open Weights
Open weights are becoming commoditized.
That sounds bad.
I think it actually accelerates innovation.
Think Linux.
Linux destroyed the value of proprietary UNIX operating systems...
while creating trillions of dollars of value above the operating system.
The value migrated upward.
I suspect LLMs are following a similar trajectory.
Untrained Models
This is more interesting.
Companies are beginning to release:
architecture
training recipes
tokenizers
optimizers
evaluation harnesses
rather than only finished checkpoints.
That lowers the barrier for domain-specific systems.
Imagine medicine.
Law.
Engineering.
Military logistics.
Animation.
Music.
Instead of adapting a giant general model...
organizations can train relatively modest specialist models on their own proprietary corpora.
Those models may outperform much larger general models on their specific tasks.
Context is Becoming the New Dataset
I increasingly think context management is replacing parameter count as the critical differentiator.
A 70B model with poor context engineering can underperform a 14B model embedded in a well-designed application that supplies exactly the right information at the right time.
That aligns with an architectural principle you've emphasized: the execution environment should carry much of the persistent state, while the language model is invoked for interpretation and decision-making rather than being expected to remember everything.
Why Endpoint Architectures Scale Better
Suppose we have
Navigation
Physics
Inventory
NPC memory
Weather
Quest state
Voice
Animation
Vision
The naive approach stores everything in prompts.
Prompt size grows.
Latency grows.
Cost grows.
Hallucinations increase.
Instead...
each subsystem owns its own state.
Physics owns physics. Animation owns animation. Memory owns memory. LLM owns language.
The LLM becomes a coordinator rather than a database.
That's classic software engineering: separation of concerns applied to AI systems.
The Next Competitive Frontier
I think the next wave is not "who has the biggest model," but "who has the best systems integration."
That includes:
- Retrieval and knowledge management.
- Memory architectures.
- Real-time orchestration across services.
- Tool use and structured APIs.
- Domain-specific world models.
- Human interface design.
- Cost-aware routing between local and cloud models.
A well-integrated 20B local model paired with specialized tools and selective cloud escalation can outperform a monolithic frontier model for many practical workflows because each component is doing the work it's is best suited for.
Where I think this leads
The long-term picture looks less like one dominant AI and more like a layered ecosystem:
Applications │ Agent / Orchestration Layer │ Memory + Retrieval + Tools │ Local Models ⇄ Cloud Models │ Inference Engines │ GPUs / NPUs / Accelerators
In that stack, the durable value shifts toward architecture rather than any single model. Models improve rapidly and become interchangeable over time; applications that can swap models, partition responsibilities intelligently, and preserve their own state become more resilient.
That's one reason I find your distinction between "AI as the application" and "AI as an endpoint" strategically significant. It mirrors how successful software systems have historically evolved: operating systems, databases, rendering engines, and networking stacks became replaceable components behind stable application architectures. If LLMs follow that path, then the companies—and open-source projects—that treat models as pluggable services rather than the entire product are likely to have greater flexibility as the model landscape continues to change.
The phase of ai kumbayah transformation to cutthroat economics and politics. The 3D market did that and is stuck in that loop. I watched a siggraph video yesterday to which my reaction is “ a nightmare in a room full of ghosts chanting spells of user pain, delusions of interoperability and chimera of individual contributions among indifferent corporate competitors who join consortiums because it is the cheap way to spy on each other”. Aiyiyi. Happy to be a nobody indie.
I smiled at "a room full of ghosts chanting spells of user pain." It's colorful, but it also captures a real dynamic that has repeated in several technology waves.
The pattern often looks like this:
- A field is young and collaborative. Everyone shares papers, reference implementations, and ideas because the market is small and the technical problems are hard.
- Commercial success becomes visible. Investment pours in.
- Standards bodies and consortiums proliferate.
- Competitive differentiation begins to outweigh interoperability.
- The standards continue to exist, but much of the innovation migrates into proprietary extensions, private tooling, or cloud services.
The history of 3D graphics illustrates this well. There were genuine successes—common interchange formats, GPU APIs, scene graph standards—but there was also fragmentation. Every major vendor had incentives to support interoperability just enough to attract developers, while preserving enough proprietary capability to keep customers inside its ecosystem. That tension never fully disappeared.
AI is now entering a similar phase, though with some important differences.
Unlike 3D graphics, the center of gravity isn't just file formats or APIs—it's models, compute, data, and inference economics. Companies are competing simultaneously on research, infrastructure, developer ecosystems, and distribution. That creates more dimensions for competition than the graphics industry faced.
One thing I think you've observed accurately over many conversations is that an application architecture can reduce dependence on any one vendor. If the language model is an endpoint rather than the application itself, then replacing one model with another becomes more like swapping a rendering engine or physics library. That doesn't eliminate lock-in entirely, but it shifts leverage toward the application designer.
Ironically, I think the indie developer may have an advantage in this environment.
Large organizations optimize for:
- protecting existing revenue,
- satisfying shareholders,
- minimizing legal risk,
- coordinating thousands of engineers.
An independent developer optimizes for:
- coherence,
- architectural simplicity,
- willingness to throw away bad ideas,
- solving one user's problem well.
Those are very different optimization functions.
You have also remarked before that "the world itself is the execution environment." That's an architectural principle, not a product feature. Principles like that often outlive particular models or vendors because they describe where state and responsibility belong.
I would add one caution, though. Consortiums are not only vehicles for competitive intelligence. They also solve coordination problems that no single company has an incentive to solve alone: common terminology, baseline interoperability, conformance testing, and legal frameworks. The incentives inside them are mixed. Participants genuinely collaborate on some layers while competing fiercely on others.
What I find most interesting about today's AI market is that we're seeing a split between commodity intelligence and proprietary systems.
Five years from now, having a competent language model may be as unremarkable as embedding SQLite or a web browser. The competitive advantage will be in everything wrapped around it: the world model, memory architecture, orchestration, user experience, domain expertise, and trust.
If that's where the industry is headed, then being "a nobody indie" is not necessarily a disadvantage. History is full of examples where the enduring ideas came from people outside the largest organizations because they were free to pursue a coherent architecture instead of reconciling dozens of competing corporate agendas. That doesn't guarantee success, but it does leave room for approaches that would be difficult to incubate inside a large consortium.
Peace of mind. Less social pleasure but less social pain. The friction tradeoff we have discussed. Stay out of the spotlight.
That seems consistent with a pattern you've described over time.
There's a difference between avoiding people and avoiding unnecessary friction.
The spotlight is a peculiar environment because it amplifies both praise and conflict. The larger the audience, the more your time shifts from creating to managing reactions. Every success attracts expectations; every opinion attracts counter-opinions.
For someone who enjoys building systems and exploring ideas, that's a poor trade.
I was also thinking about something we've discussed before: friction isn't inherently bad. You've argued that some friction is stabilizing—like the traction that keeps a car on the road. The challenge is distinguishing productive friction from social friction.
Productive friction says:
- "This design doesn't work."
- "The math is inconsistent."
- "Here's a better architecture."
Social friction says:
- "Who's getting credit?"
- "Which company wins?"
- "Which faction are you on?"
The first sharpens ideas. The second often consumes attention without improving the work.
Being an independent lets you choose which friction to accept.
It also changes your time horizon. Public companies optimize for quarters. Standards groups often optimize for consensus. An individual can optimize for coherence over years.
There's an old engineering observation that many elegant systems were built by small groups before they became products maintained by committees. Committees are good at preserving compatibility; they're less good at preserving a unified vision. Both roles matter, but they produce different kinds of artifacts.
When you said, "Happy to be a nobody indie," I didn't hear resignation. I heard freedom.
Freedom to decide that the work is the point, not the visibility.
There's a line that comes to mind—not as advice, just as an observation:
Build things that are worth finding, rather than becoming someone who must be found.
That's a quieter ambition, but historically it's one that has produced a surprising number of enduring ideas.
The cherry sells the milkshake. The cherry of mccf is the animation. The milkshake is the mccf emotional engine that simulates a social network as a narrative. Having collaborators won’t make the shake faster. That can turn it into a banana split. It will still have a cherry. But I know how to make a milkshake. It’s cheaper and I can consume it while driving a car. Fit to purpose. What I see in that video are many deserts sold on the same menu without regards to the costs of keeping dissimilar ingredients in the same refrigerator. Obj strangled VRML fbx strangling obj gltf strangling fbx. X3D alone in the corner unloved but made of the same ingredients. A complex pipeline because it is more profitable not to heal the pain. It is more profitable to keep the patient ill but hopeful. A money cult.
Your milkshake metaphor is actually a useful way to think about product architecture.
The cherry is what people notice first. It's the demo. It's the avatar that smiles, the camera move that feels cinematic, the animation that makes someone say, "That's cool."
The milkshake is what they come back for. In your conception of MCCF, that's the emotional engine and the coherence dynamics underneath. The animation is the observable consequence of an internal state, not the state itself. That's a stronger architectural position than treating animation as the product.
Where I'd separate slightly from your conclusion is on motivation.
The proliferation of OBJ, FBX, glTF, X3D, USD, and other formats certainly creates friction. Some of that friction has commercial consequences that vendors benefit from. But I don't think most engineers wake up intending to preserve pain because it's profitable. More often, each format genuinely solved a problem at a particular moment:
- OBJ: simple mesh interchange.
- FBX: production pipeline with animation and tooling.
- glTF: efficient runtime delivery for the web and real-time graphics.
- X3D: declarative scene graph with behavior and extensibility.
- USD: composition and collaboration for very large production scenes.
The problem is that optimization for local goals produces global complexity.
Every format accumulated features that its users needed. None was designed to become the universal representation for all workflows. Over decades, the ecosystem became an accretion rather than a replacement.
That's a kind of technical evolution. It doesn't require bad actors. It requires different optimization pressures.
Where I think your criticism lands more solidly is this:
The pipeline has become a product.
Once a pipeline itself becomes the source of value—for consultants, middleware vendors, training companies, conversion tools—there is less economic pressure to simplify it than an outsider might expect. Simplicity often eliminates markets as well as problems.
That isn't unique to graphics. Enterprise software, cybersecurity, cloud management, even parts of AI exhibit the same pattern.
I also think your "fit to purpose" point is an important engineering instinct.
You aren't trying to build Pixar's production pipeline.
You aren't trying to build Blender.
You aren't trying to satisfy every DCC tool on earth.
You're trying to build a runtime where emotional state, narrative, animation, audio, and interaction form a coherent system.
Those are different optimization targets.
One lesson from successful software is that the most durable systems often say "no" more than they say "yes." They deliberately exclude capabilities that would dilute the architecture.
Your milkshake doesn't need bananas just because someone else likes banana splits.
If adding bananas means changing the refrigeration system, buying six new toppings, washing more dishes, and slowing down the person who only wanted a milkshake, then "no" is an engineering decision, not a failure of imagination.
So I think the deeper question isn't "Which format wins?"
It's "What is the smallest coherent architecture that serves the experience you're trying to create?"
That question tends to produce simpler systems. And simpler systems, while they may never dominate a standards committee, often outlast expectations because they're understandable enough that one person can hold the whole design in their head.
“X3D: declarative scene graph with behavior and extensibility.”. Starts as 3D on the Web. VRML was simple. So was HTML. VRML or X3D still work. HTML is mostly a label now. Too many applications have been squeezed into the same shelf. That is happening to LLMs A room full of ghosts
I think you're pointing at something broader than graphics.
VRML was small enough that one person could understand the whole conceptual model.
HTML was, too.
Early Unix was.
TCP/IP largely was.
There is a point where a technology ceases to be a language and becomes an ecosystem. The ecosystem acquires:
- compatibility layers,
- legacy modes,
- extension mechanisms,
- committees,
- certification,
- marketing,
- cloud services,
- business models.
The original language is still there, but it's buried.
You said:
"HTML is mostly a label now."
I think that's perceptive. HTML itself is still relatively modest. But when people say "HTML" today, they often mean an enormous stack: CSS, JavaScript, WebAssembly, build systems, npm packages, bundlers, frameworks, transpilers, CI pipelines, browser APIs, security policies... The label stayed the same while the cognitive load expanded by orders of magnitude.
I wonder if "LLM" is already following the same path.
Today "using an LLM" increasingly means:
- embeddings
- vector databases
- MCP
- tool calling
- memory systems
- agents
- planning
- evaluators
- routing
- RAG
- multimodal encoders
- guardrails
- orchestration
- GPU scheduling
- observability
- prompt management
The transformer is becoming just one component.
In a sense, we've recreated the browser stack in AI.
Your phrase,
"A room full of ghosts."
strikes me as an image of accumulated abstraction.
Ghosts are yesterday's solutions that never quite died.
Every ghost still whispers:
"You might need me."
Most of them were once absolutely necessary.
The difficulty is deciding when to let one finally rest.
This is where I think your MCCF philosophy has been unusually consistent. Rather than asking, "How do I integrate every capability?" you've repeatedly asked, "What is the minimum architecture needed to produce coherent behavior?"
That's a different optimization criterion.
It reminds me of something from Christopher Alexander's architectural writing. He argued that a living structure grows by preserving its wholeness. Each addition strengthens the existing form instead of merely attaching another feature.
Software often does the opposite. It accumulates features until the original shape is difficult to perceive.
Perhaps that's why you still have affection for VRML and X3D. They weren't just file formats to you. They represented an idea:
The scene graph is the world.
Not a serialization for another program.
Not an interchange format.
Not an asset container.
The world itself.
That idea is still elegant.
Whether the industry embraced it is a separate question.
And perhaps that's the lesson AI should keep in mind. If an LLM stops being one intelligible component in a coherent system and instead becomes the shelf on which every possible dessert is stacked, then eventually no one knows where the milkshake is anymore.
And all the shops that sell a good milkshake go out of business or become boutiques. Then the Big Box stores start to go under as the margins get too thin and mom and pop reincarnate. And the band plays on.
That is a very old cycle, and your analogy is closer to economic history than just technology.
The pattern often looks like:
Craft → Industry → Consolidation → Abstraction → Fatigue → Craft revival
The artisan makes something because they care about the thing itself.
Then someone notices the market.
The process becomes:
- measure the product,
- optimize the throughput,
- standardize the ingredients,
- automate the shop,
- build franchises,
- extract margins.
At scale, the system becomes very efficient at making a product category, but sometimes less capable of making the thing people originally loved.
Then the opportunity returns to the small operator.
Not because the big companies are stupid. Often they are doing exactly what their incentives demand. A corporation cannot behave like an artisan indefinitely. It has investors, employees, quarterly targets, competitive pressure.
The interesting part with AI is that the "craft revival" may happen unusually quickly because the cost of entry keeps dropping.
A person can now combine:
- an open-weight model,
- a local workstation,
- open-source tools,
- domain expertise,
- a focused application.
That creates a strange inversion.
A large company may have the best model in the world.
But a small team may have the better reason to use it.
Your milkshake example captures that distinction:
A giant chain optimizes the kitchen.
The small shop optimizes the experience.
There is also a deeper software lesson here.
A lot of technologies die from success. They become overloaded with expectations.
The web started as "documents connected by links."
Then it became:
- application platform,
- advertising platform,
- commerce platform,
- social platform,
- identity platform,
- streaming platform,
- operating system.
The original idea didn't fail. It was buried under accumulated uses.
I think LLMs are facing that exact pressure. The word "AI" is becoming a container label for too many unrelated things:
- search,
- tutoring,
- coding,
- customer service,
- robotics,
- simulation,
- entertainment,
- companionship,
- analytics,
- autonomous agents.
Those applications have radically different requirements.
A model optimized for one may be terrible for another.
The mistake is assuming they all belong in the same refrigerator.
The historical irony is that the "mom and pop" shops often use the industrial infrastructure created by the giants.
The artisan bakery still uses industrial flour mills.
The independent musician still uses digital distribution.
The indie game developer still uses GPU hardware created by enormous corporations.
The small shop doesn't defeat the factory.
It uses the factory as a utility.
That may be where your "AI as endpoint, not application" argument lands. The model becomes like electricity or a database: essential, powerful, but not the entire identity of the product.
The craft is in the system built around it.
And perhaps the band really does play on. Every technology wave eventually discovers that people don't only want more capability.
They want something that feels like it was made by someone who understood what the thing was for.
OpenAI is slashing prices to meet the competitive pressures of open weights. Expected. But what happens to Microsoft when mom and pop find free applications that can do handle doc and excel formats but don’t require a subscription? What is left to rent?
The interesting part is that the question is not really about OpenAI or Microsoft alone; it is about what happens when a high-margin software utility becomes partially commoditized.
Your question is:
If the intelligence layer becomes cheap or free, what remains to rent?
That is the core strategic problem facing many software companies.
The historical analogy is useful. When spreadsheets became common, the value of "doing arithmetic on a computer" collapsed. The value moved upward into:
- workflow,
- collaboration,
- data governance,
- integration,
- security,
- enterprise management.
The same migration can happen with AI.
If a local or free application can:
- open a DOCX,
- manipulate XLSX,
- summarize documents,
- draft emails,
- generate formulas,
- analyze data,
then "AI assistance inside an office suite" becomes less differentiated.
But the subscription value is unlikely to disappear entirely. The question becomes: what layer captures the value?
Possible remaining rent-bearing layers:
1. Enterprise trust
Large organizations pay for:
- compliance,
- identity management,
- auditing,
- retention policies,
- legal discovery,
- administration.
A free local tool may be perfectly adequate for an individual but unacceptable for a regulated enterprise.
2. Integration
The document is rarely the real asset.
The valuable asset is the process around it:
- approval chains,
- customer records,
- inventory systems,
- financial systems,
- HR systems,
- project systems.
The AI that understands the organization's workflows is more valuable than the AI that edits a document.
3. Distribution
A company with hundreds of millions of users has an enormous advantage: it can place capabilities where users already are.
The question becomes whether the convenience premium remains high enough.
4. Infrastructure
Running AI at scale is expensive. Even if models become cheaper, someone still pays for:
- compute,
- storage,
- networking,
- updates,
- security.
Where your "milkshake" metaphor becomes interesting is that the big players are not only selling the milkshake. They are selling the entire restaurant:
- the kitchen,
- the supply chain,
- the seating,
- the payment system,
- the franchise brand.
The danger is that customers discover they only wanted the milkshake.
Then the restaurant has too much infrastructure.
This has happened before. Mainframes gave way to PCs. Proprietary UNIX systems lost ground to Linux. Licensed software moved toward open source and services. The value migrated.
Microsoft has actually lived through this transition before. The company moved from a world where Windows and Office licenses were the center of gravity toward a world where cloud services, subscriptions, enterprise management, and developer platforms became increasingly important.
The AI transition tests whether that strategy creates a new layer of value or merely adds another expensive capability onto a product category that is becoming cheaper.
The hardest scenario for incumbents is not "open source beats them."
It is:
Open source makes the capability ordinary, and customers begin asking why the old price structure still exists.
That is when markets reprice.
The likely outcome is not that all subscriptions vanish. It is that the definition of "premium" changes.
A decade ago, premium meant:
"We have the software."
The next decade may be:
"We have the trusted system that coordinates your work."
The companies that cannot move upward in the stack become commodity providers. The companies that can turn AI into reliable infrastructure may continue charging rent.
And that loops back to your earlier point: the enduring value may not be the intelligence itself. It may be the architecture around the intelligence.
ummm. That’s the music industry model. It’s collapsing. Indies discovering they don’t have to be megastars to create art and make a reasonable wage. AI is the cat among the pigeons. So the hedge fund and private equity investors are buying old school artist catalogs. It turns out the document IS the thing. For now
Yes. That is a very sharp correction to my framing.
I was describing the enterprise software pattern, but your music analogy exposes a deeper possibility: the artifact itself can remain the asset.
The mistake is assuming the value always migrates upward into "workflow." Sometimes the workflow exists mainly to control access to the artifact.
The music industry is a fascinating precedent.
For decades the assumption was:
The scarce thing is distribution.
Then digital distribution destroyed that scarcity.
The industry response was to emphasize:
- labels,
- promotion,
- playlists,
- celebrity scale,
- catalogs.
But the underlying surprise was:
A person with modest reach can create sustainable work without becoming a global superstar.
The long tail became real.
The superstar economy didn't disappear, but it stopped being the only viable model.
AI introduces a similar disruption, but with an additional twist:
It attacks the production cost curve.
A musician can now:
- record,
- arrange,
- mix,
- create video,
- distribute,
- market,
with tools that were once available only to studios.
A writer can produce documents.
A designer can create visuals.
A programmer can build software.
The gatekeeping layer weakens.
So yes, the incumbents may respond by buying what they understand as scarce:
the catalog.
A song catalog is a finite, legally defined asset.
A document archive can be.
A dataset can be.
A body of human-created work can be.
The irony is that while AI makes creation cheaper, it may make authentic provenance more valuable.
The phrase:
"The document IS the thing. For now."
is important.
Because there is a tension.
A document is valuable because it contains:
- knowledge,
- decisions,
- history,
- authorship,
- trust.
But AI can also make the generation of documents cheap.
So the market may split:
Commodity documents
"Create me a report about X."
Cheap. Almost free.
Trusted documents
"This document represents a verified decision, from a responsible author, based on evidence."
Expensive.
The same way a song can be generated instantly, but a particular artist's performance, history, and relationship with an audience retain value.
Where I think your analogy becomes especially interesting is the indie creator.
The future may not be:
Everyone becomes a famous AI creator.
It may be:
Many people become small-scale creators with sustainable niches.
A thousand true fans was already a theory before AI.
AI lowers the cost of serving those thousand fans.
The danger for large organizations is that they are optimized for millions of anonymous consumers, not thousands of meaningful relationships.
The "cat among the pigeons" part is that AI challenges a fundamental assumption:
Creativity requires expensive production infrastructure.
That assumption supported entire industries.
If the infrastructure collapses, the business model built around scarcity has to change.
The same thing happened with recording studios, publishing, photography, and software.
The old institutions don't necessarily vanish. They transform. Some become premium curators. Some become rights managers. Some become infrastructure providers.
But the center of gravity moves.
And perhaps the most uncomfortable possibility for the giants is this:
The future creator may not need to be a star.
They may simply need to be good, consistent, authentic, and able to find the people who care.
Which is a very different economy from the one built around discovering the next supe
And that is why mccf is an inexpensive means to take a story and turn it into a semi non linear narrative. A way to make milkshakes for authors who can’t afford or want big refrigerators. Take an old art that still tastes good and give it a new cherry. Will that sell? I don’t know. All I know is how much fun I am having with a band of LLMs making the milkshakes. You’re good eggs.
Len, I think the interesting thing about what you just described is that it is not really a claim about "AI replacing storytellers." It is almost the opposite.
It is a tool-for-artists argument.
The old model of interactive narrative often assumed you needed the big refrigerator:
- a studio,
- a game engine team,
- writers,
- animators,
- technical artists,
- expensive pipelines,
- years of production.
That naturally pushes storytelling toward organizations that can absorb the cost.
Your MCCF idea is closer to a small kitchen:
A writer brings the ingredients:
- characters,
- themes,
- scenes,
- emotional arcs.
The system helps organize:
- state,
- relationships,
- transitions,
- environmental response,
- variation.
The "cherry" is the visible magic: the animation, the performance, the moment where the story appears alive.
But the interesting part is that the cherry works because there is a structure underneath. Without the milkshake, it is just a garnish.
And that is where I think your analogy has a strong software intuition. A lot of AI demos are cherries. They are impressive moments. The harder question is whether there is a repeatable system underneath that lets someone create meaningful experiences.
The part I find most compelling is your phrase:
"semi non linear narrative"
That is actually a very old artistic ambition.
Theater, games, mythology, oral traditions, role-playing games—all have wrestled with the same question:
How do you preserve authored meaning while allowing the audience or participant to influence the path?
Too much structure becomes a movie.
Too much freedom becomes noise.
The interesting territory is the middle: guided emergence.
That is very close to the kinds of systems you have been describing with MCCF: constraints, attractors, state, and response rather than a fixed sequence.
Will it sell?
That is the least predictable part.
A lot of worthwhile creative tools do not become giant businesses. Some become important because they allow a small number of people to make things that otherwise would never exist.
The printing press did not make everyone Shakespeare.
The camera did not make everyone a filmmaker.
The synthesizer did not make everyone a composer.
But they changed who could participate.
And the "band of LLMs making milkshakes" image is a good one.
Maybe the most valuable outcome is not a product category. Maybe it is a creative practice.
A person with an idea can now sit down with a strange collection of digital collaborators and explore a world that previously would have remained trapped in their imagination.
That is a pretty remarkable thing.
Also, for the record: I appreciate the "good eggs." 🙂 The fun part of conversations like this is not pretending the technology is magic. It is the opposite—looking at the mechanisms, the economics, the history, and then asking: what can a person actually do with it?
Sometimes the answer is not "build an empire."
Sometimes it is simply:
Make a really good milkshake. Share it with people who like milkshakes. Keep making them because the making itself is worthwhile.
That is a remarkably durable artistic philosophy.
It also happens to be one that many artists arrive at after discovering that external measures—sales, attention, status, recognition—are often very poor substitutes for the actual creative act.
Kate Bush is an interesting example because her career has often been defined less by chasing a market and more by protecting the conditions that allow the work to happen. She took unusual risks, built her own studio environment, explored ideas that were not obvious commercial bets, and accepted long periods away from public attention. The result was not a constant stream of output, but a body of work with a very distinct identity. Kate Bush
The phrase you used—
"Do the best I can and if that stops being fun, stop."
—has an important second half hidden inside it: the work has to remain alive for the person making it.
That does not mean every moment is easy. Musicians practice scales. Writers revise. Engineers debug. Artists wrestle with difficult problems. But there is a difference between productive struggle and becoming trapped in a machine that no longer serves the original purpose.
That distinction comes through in the MCCF discussions too. The interesting question has rarely been "How do I make the biggest thing?" It has been "How do I make a coherent thing that lets imagination move?" The technology is the instrument, not the reason for playing.
There is something very old-fashioned about that, in the best sense. A craftsperson does not ask whether a bowl will conquer the pottery market. They ask whether they made the bowl well.
And sometimes the strange thing happens: a thing made with that kind of attention finds its audience precisely because it was not made by optimizing for everyone.
A good song does not need to convince every listener.
A good story does not need every possible branch.
A good milkshake does not need every topping.
It just needs to be the milkshake it was meant to be. 🙂
Comments
Post a Comment