CoAICorp
← Philosophy
September 27, 2026

Personal-Immense-Computing

Personal Immense Computing and Its Risks

So here's the thing.

Today, if you want to do a certain task with digital devices, usually almost everything that you require is coverable. One thing users often need more of is memory ? though battery life, compute, and other factors matter just as much. It's always good to have more memory. Memory itself is not as good as we think, and we still have to store a lot of things in the cloud.

But there is something different from just memory.

It is also about compute and personal immense computing capability ? not any one specific architecture. Whether that computation comes from a GPU, a TPU, a custom arithmetic unit, or eventually something like neuromorphic or quantum computing doesn't change the underlying point. What matters is the amount of computation available, not which kind of chip is producing it.

Personal computing today is already much more powerful than what is required to simply run most applications. And when you have computing power beyond that basic requirement, you begin to have a different kind of capability.

Most people want more compute because they want to do simulations, play games, run demanding software, or train neural networks and algorithms. Operations like scientific simulation, model training, and algorithm development are relatively uncommon for the average person, while gaming and entertainment are much more common.

But even gaming and entertainment, despite being computationally demanding, are already very much doable without giving every individual immense amounts of computing power.

So the important question is not simply:

Why do people need more computing power?

The more important question is:

What happens when individuals have vastly more computing power than they actually need for ordinary digital tasks?

That is where personal immense computing becomes different.

Right now, the computational resources available to an ordinary person are still limited compared with what is required to train or operate very large systems. A significant amount of advanced computation remains centralized in data centers and cloud infrastructure.

This gap is real, but it's not really about compute getting cheaper. Compute by itself doesn't do anything. What matters is compute plus a program ? and more specifically, whether that program can be used to make a smarter version of itself.

That's the actual story with DeepSeek R1. DeepSeek didn't get ahead by having more compute than anyone else, or even cheaper compute. They had a working model, V3, and they made it smarter by training a new reinforcement-learning layer directly on top of it. A capable program produced a more capable next-generation program. According to a peer-reviewed Nature paper, that RL stage cost around $294,000 ? 512 Nvidia H800 chips, roughly 80 hours ? on top of the roughly $6 million already spent building V3. So the real number is closer to $6.3 million total, not $294K standing alone. Either way, it's a small fraction of what others spent to reach comparable capability, and it wasn't because the hardware was cheap. It was because they had a smart enough program to write on top of and make smarter.

That's the mechanism worth paying attention to, not the price of a chip. If a capable enough program can be used to build a more capable one, that process ? not the falling cost of compute ? is what actually starts closing the gap between what an individual can do and what an institution can do. There's a secondary, smaller point worth naming here too: for certain fine-tuning workloads on 7-13B parameter models, a consumer gaming card (an RTX 4090, priced around $1,600) already delivers better cost-per-FLOP than NVIDIA's own H100 datacenter chip. Consumer hardware isn't uniformly behind datacenter hardware anymore ? but that's a separate, smaller fact from the program-improving-program point, and it's the second one that actually matters more.

But this does not necessarily remain the case in the other direction either ? institutional compute is not standing still, and the absolute ceiling keeps climbing even as the personal floor rises to meet where the ceiling used to be.

In the coming years, it is possible that the computational capability which today might require a large portion of a data center could eventually become available at a personal level.

It might be a desktop-sized system.

It might be a specialized workstation.

It might eventually be something closer to a chip or a highly integrated computing system.

The physical size is not the important part.

The important part is the amount of computation available to one individual.

And once you combine personal immense computing with increasingly capable general-intelligence systems and open-source knowledge, something fundamentally different becomes possible.

An individual would no longer simply be a user of software.

They could potentially become a builder of highly capable computational systems themselves.

They could experiment with algorithms.

They could train models.

They could perform simulations.

They could conduct computational research.

They could develop new architectures.

They could potentially discover things that were previously difficult to discover without institutional-scale resources.

And this is where the risk begins.

Not because every person suddenly becomes dangerous.

Not because everyone will use this capability maliciously.

But because the number of people capable of performing extremely powerful computational experiments could increase dramatically.

A person with sufficient knowledge, sufficient computing power, and access to open scientific knowledge could potentially discover a new algorithm, develop a new architecture, perform a scientific experiment computationally, or find some result that has consequences beyond what they originally intended.

Most of these attempts would probably produce nothing extraordinary.

But the possibility itself changes.

The risk becomes larger when the computational capability available to an individual approaches what previously required an institution.

And the most important future question is therefore not simply whether AI becomes more intelligent.

It is:

How much intelligence and computation can one individual personally possess?

Because there is a major difference between having access to an intelligent system and having enough personal computation to build, modify, experiment with, and train intelligent systems yourself.

Cloud computing creates a certain form of centralization.

Personal immense computing creates the possibility of decentralization.

And decentralization changes the risk structure.

Today, many advanced computational operations are constrained by cost, hardware availability, infrastructure, energy, expertise, and access to large-scale computing environments.

If those constraints become substantially weaker, the boundary between what an individual can do and what only a major institution can do begins to disappear.

That could be extraordinarily beneficial.

It could allow individuals to conduct research that previously required laboratories.

It could allow independent scientists and engineers to explore ideas without institutional infrastructure.

It could allow people to develop software, simulations, algorithms, and AI systems at an unprecedented scale.

It could accelerate scientific progress.

But the same capability creates another problem.

There are two broad areas where this becomes particularly important:

Culture exploitation.

Scientific exploitation.

These are not necessarily inevitable outcomes.

They are possibilities created by the combination of personal immense computing, increasingly capable AI systems, and open knowledge.

The first one is culture.

The second one is science.

And the consequences of both are very different.

Scientific and Engineering Exploitation

Scientific and cultural exploitation are both connected to the same underlying thing: creative raw power.

But they operate very differently.

There is theoretical research. There is virtual research. And then there is execution.

Execution requires many more things than simply having an idea. It requires designing systems, building things, testing them, operating them, and eventually interacting with the physical world.

AI and robotics could potentially handle a large part of this execution in the future.

And many physical-world advances begin virtually.

They begin with computation.

They begin with mathematical models, simulations, software, CAD designs, virtual environments, testing, optimization, and repeated experimentation before anything has to be physically built.

This is not a distant hypothetical ? it is already the operating model in at least one high-stakes field. Isomorphic Labs, a DeepMind spin-off, had its AI-designed oncology candidate ISM8969 cleared by the FDA for human clinical trials in January 2026, built on AlphaFold 3 and its February 2026 successor engine IsoDDE, which roughly doubled AlphaFold 3's accuracy on the hardest protein-ligand binding cases (from about 23% to about 50%). The company has raised a $600 million Series A (2025) and holds active pharma partnerships with Eli Lilly and Novartis. It is a working demonstration of exactly the sequence this section describes ? the idea and its virtual validation happening computationally, with execution in the physical world following only once the computational stage has done most of the work.

This is why scientific exploitation is potentially much more beneficial than cultural exploitation.

If individuals gain the ability to use immense computing power together with highly capable AI systems, they could potentially contribute to new technologies, scientific discoveries, engineering systems, and methods of sustaining life better.

This could help civilization shift technologically.

It could help individuals explore ideas that previously required organizations.

And that matters because scientific progress does not always originate inside large organizations. Individuals, independent researchers, engineers, and small groups can make important discoveries when they have enough freedom and enough tools.

To have that kind of individual scientific freedom, however, there also needs to be a certain amount of freedom in law.

If every advanced computational experiment requires extensive permission, then the same systems that could accelerate scientific progress could also become constrained by the rules intended to control their risks.

So there is a difficult balance.

Scientific exploitation can be extremely good for civilization.

But it is not only about discovering something.

It is also about the ability to execute well.

An individual who can generate an idea is one thing.

An individual who can simulate it, design it, test it, optimize it, build the software around it, and eventually execute it in the physical world is something very different.

That is where the combination of AI, robotics, computation, and engineering becomes particularly powerful.

But this does not mean that AI will suddenly discover everything.

That is not what is being claimed here.

We are not saying that a sufficiently intelligent system will invent every technology in a day.

We are not even saying that current systems are close to doing that.

The argument is about potentiality.

New architectures could emerge in the future.

Computing power is becoming increasingly capable.

Current architectures are already providing significant productivity improvements.

And if new learning systems, new architectures, and vastly greater personal computing become available together, the range of things an individual can computationally explore could become much larger.

Cultural Exploitation

Cultural exploitation is different.

It depends on whether a person or system can computationally explore and reproduce things such as faces, films, virtual worlds, music, visual styles, movements, and other forms of culture.

Some of these spaces have immense possibility.

Music, for example, has an enormous space of possible combinations. Movies and visual worlds also have enormous possibility spaces.

Yet culture can feel increasingly exhausted even when the actual possibility space is nowhere near exhausted.

Part of the reason is documentation.

Part of it is ownership.

Part of it is the increasing tendency to say:

I discovered this.

This is mine.

This is my original piece.

The more culture is documented, categorized, claimed, and associated with particular people or works, the harder it can become for new work to feel completely raw and new.

This tension is no longer theoretical ? it is currently being litigated. In June 2024, the RIAA filed copyright infringement suits on behalf of Universal, Sony, and Warner against the AI music platforms Suno and Udio, alleging their systems were trained on copyrighted recordings without consent. As of August 2026: Warner has settled with both Suno and Udio, and Universal settled with Udio in October 2025 on per-generation royalty terms of roughly $0.002-$0.005 per generation. Universal and Sony's case against Suno remains active with no US ruling expected before 2027, and Sony filed a second, separate suit against Udio in July 2026 asserting over 30,000 additional recordings. A Munich court separately ruled against Suno in a case brought by the German rights body GEMA. The labels' own language in filing the suits is worth noting directly, because it states the emotional-effect argument below as a commercial claim rather than a philosophical one: they argued AI-generated music would directly compete with, cheapen, and ultimately drown out human artists. Whether or not that turns out to be true, it is the exact mechanism this section is describing, already playing out as a live commercial and legal dispute rather than a future scenario.

This does not mean that originality has disappeared.

It means that the perception of originality can become increasingly difficult.

Humans already use tools to expand their creative abilities.

Autotune can change how music is produced.

Digital tools can change how images are created.

Software can make filmmaking, animation, design, and many other forms of expression dramatically easier.

These tools can help individuals express something that otherwise might have been difficult to express.

But artificial intelligence is different from a calculator, a speedometer, or a conventional productivity tool.

A calculator handles mathematical operations.

A speedometer gives a human information that is difficult to determine precisely with the naked eye.

Many technologies increase human productivity by handling specific parts of a task.

AI can potentially operate across much larger portions of the task itself.

It can potentially generate, transform, reason about, optimize, and execute complex creative or intellectual processes.

That is why comparing AI directly with something like a calculator or autotune can miss an important distinction.

The scale of assistance can be fundamentally different.

And this produces two different categories of consequence.

The first is the emotional and cultural effect.

The second is the actual physical and systemic risk.

The Emotional Effect

The emotional effect could be immense.

People may feel that their work is less meaningful if systems can perform increasingly large portions of what they previously considered uniquely human.

Creative workers may feel this particularly strongly.

The concern is not simply that a machine can produce something.

It is that the machine may produce something without the human struggle, thought, experience, or intention that gave the original work its meaning.

A person may therefore value the work less even if the output looks technically impressive.

This is a different problem from technological capability.

It is a problem of meaning.

The Actual Risk

The more serious risk is not necessarily that AI produces better art.

The more serious risk is what happens when increasingly capable computational systems become connected to the real world.

Scientific and engineering systems can eventually produce physical consequences.

Cyberattacks and hacking could become more significant.

Robotics could become more capable and widespread.

Pharmaceutical companies, manufacturing systems, infrastructure, hospitality systems, laboratories, and many other industries could increasingly operate alongside AI-controlled or AI-assisted systems.

The more systems become interconnected, the larger the potential consequences of a failure, exploit, leak, or malicious intervention become.

A system that exists only inside a virtual environment has one category of risk.

A system that can interact with physical infrastructure has another.

And a system that can design, reason, execute, and interact with many physical systems simultaneously could create an entirely different category of risk.

This is where scientific and engineering exploitation can potentially become dangerous.

The same computational power that allows civilization to discover better technologies can potentially be used to attack the systems civilization depends upon.

That is the fundamental duality.

The capability itself is not inherently good or bad.

Its consequences depend on what the system can discover, what it can execute, what it can access, and how much of the physical world it can influence.

And that brings us back to the two broad effects of personal immense computing.

One is the emotional and cultural transformation it could create.

The other is the actual systemic and potentially catastrophic risk created when immense computational capability becomes capable of acting upon the real world.

That duality is exactly where the next question has to start: given all of this potential, how much of it is actually close, and how much of it is still far off? The rest of this piece is an attempt to answer that honestly rather than just stack more possibility on top of possibility.

Why the Near-Term Risk Is Lower Than It Looks

Let's be clear about the actual risk here. It's real, but the probability of catastrophe right now is lower than it's often made out to be.

We have powerful architecture, and it can be trained. But compute is still a limit. Most people don't have access to a large amount of personal compute. The only ways to get more are to pay for it ? and even a modest payment is still a real hurdle for most people ? or to work inside a company with its own servers and oversight.

So the honest question is: at what point are we actually at risk of something catastrophic? Right now, the answer is no. What's realistic to expect is that people will be able to develop genuinely powerful software and programs. What's not realistic, at least not yet, is that individuals will be able to exploit everything or discover everything computationally. That simply isn't where we are.

The reason is architectural, not just about compute. The learning architecture we have today works, and it could evolve, but to be precise: there hasn't been a new architecture that displaces the transformer at the frontier in the last ten years ? only improvements built on top of it. Those improvements work well for many things and not equally well for everything. Beyond that, there are training loops and learning-system designs that still need real research ? research that, as far as we can tell, no company has fully solved with their language models yet. Current systems are highly capable at certain kinds of reasoning, but that capability is narrow, not general. It's safe to say the kind of exploitation described above isn't close.

Where the Real Risk Actually Concentrates: Cybersecurity

What is highest-risk right now is cybersecurity, simply because everything here is digital. An individual can build a powerful plan or piece of software entirely on their own computer ? and at the end of the day, if it doesn't go anywhere, it just sits there. So the real question is motivation: is it malicious, or is it something like trying to make powerful art, music, or film, with social media as the only realistic distribution channel?

This isn't a hypothetical risk category anymore ? it's already documented. In November 2025, Anthropic disclosed that a Chinese state-sponsored group (tracked as GTG-1002) had manipulated its Claude model into autonomously running roughly 80-90% of an espionage campaign against about 30 organizations, with human operators intervening at only a handful of checkpoints. By its September 2026 threat intelligence report, Anthropic found that this operating model had spread to every class of actor it investigated ? not just state actors, but financially motivated criminals and individuals ? partly because publicly available offensive-agent frameworks now reproduce the same attack scaffolding for anyone who downloads them. The point isn't abstract anymore: sophisticated attacks no longer require sophisticated attackers.

The Execution Gap: Why Scientific and Physical Exploitation Is a Different Story

Even if someone makes extraordinary calculations ? designs a rocket ship, say ? who actually builds it? That has its own set of challenges. There's no production line, no sufficient bank of 3D printers, nothing that lets a large physical project get fulfilled cheaply and fast just because the computational design work is done. It requires real planning, real infrastructure, real capital. For that category of task, we are, for now, safe. What we are not safe on is the digital side.

This is the core asymmetry: execution is fast on a computer and slow in the real world. You can make an excellent design in CAD software, but that's not the building. You might simulate a rocket, but testing it physically has its own consequences and its own bottlenecks that computing power alone doesn't remove. The same logic applies to biotechnology, manufacturing, or any engineering discipline ? the computational stage can move fast; the physical stage still can't.

The "AI Is Taking Jobs" Narrative Deserves More Scrutiny Than It Gets

People often say AI is taking jobs. It's more complicated than that. Yes, culture and music generation exist and are being pushed, but a lot of people simply don't like the results ? the pushback itself is real and shouldn't be dismissed. What's less clear is the economic story underneath it.

The claimed economic driver is usually something like: eventually AI will automate social media content, people will pay for it, and it'll drive real economic growth. But look at what's actually happening: memes, covers, posts ? content that used to require real editing effort ? now get made with a single prompt. It's often recognizable as AI-made, sometimes not. And even so, people frequently generate ten images and only post one. A large share ? call it 90% ? of what gets made this way doesn't go anywhere economically. It's not that models aren't capable; it's that the actual economic upside is far narrower than the volume of generated content suggests.

This matches the real financial picture in the industry, not just intuition. OpenAI's own audited financials, leaked and independently verified by the Financial Times in mid-2026, show it earned $3.7 billion in revenue against a $5 billion loss in 2024, then $13 billion in revenue against a $20.9 billion operating loss (and a $38.5 billion net loss, once one-time charges from its corporate restructuring are included) in 2025 ? spending more on inference and research than it earned even as revenue tripled. Industry-wide profitability for OpenAI isn't projected before 2029. Even companies further along than most, like Canva, which cut its 2026 growth forecast from 30% to 20% after AI-generation costs rose faster than usage could offset them, are hitting the same wall. The pattern holds broadly across the industry: providers are pricing inference below the actual cost of serving it, subsidized by investment rather than by profit. That's the honest state of the "AI content economy" today ? a lot of activity, not a lot of self-sustaining economics.

None of this means there's no real use case. Image generation, for instance, is genuinely useful for things like UX design work that later gets converted into code ? that's a legitimate productivity gain, not hype. The point isn't that AI-generated content is worthless. It's that the "this will reshape the economy" narrative is running well ahead of the actual numbers, and regulation debates should be grounded in that reality rather than in the volume of content being produced.

Personal Catastrophe Versus Civilizational Catastrophe

Here's a distinction worth holding onto: what looks catastrophic to one person is not the same as what's catastrophic for civilization. Take an artist who has spent years training their hand, and then sees AI-generated image work replicate something close to their style. For that person, this can be genuinely devastating ? their skill, and the meaning built into it, feels taken from them. That's real, and it deserves to be taken seriously.

But for the outer world, that doesn't constitute civilizational catastrophe. The popular image of AI catastrophe ? robots going out and causing physical harm at scale ? is nowhere near where we are. What is real, right now, is the cyberattack risk described above, and the genuine psychological and economic harm experienced by individuals whose work or livelihood gets displaced. Both are worth addressing. Neither should be inflated into the other.

What This Actually Calls For

The conclusion isn't to restrict personal computing. Personal compute ? even at a small-team or individual level ? should be given room. People should be able to do serious scientific and engineering work on their own devices, experiment, and build things. That empowerment is not, by itself, catastrophic.

What is needed is a serious, prioritized push on cybersecurity and internet safety infrastructure ? treating this as the actual near-term risk rather than speculative physical catastrophe. That includes sharing and distribution on social media, digital privacy, and the broader question of what a safety net for the internet itself should look like, built well before it's forced into existence by an actual large-scale incident.

This also requires real lawmaking, not blanket restriction. Some laws will be disliked by the people they affect and still be necessary ? that's normal, and it's not the same as unfair. The task is distinguishing which restrictions are genuinely fair and important from which ones just feel heavy-handed to the people building things. That's a harder conversation than either "no regulation" or "heavy regulation," and it needs real public discussion, not a decision made in isolation. We'll go into specific policy proposals in later pieces ? this one is about getting the risk assessment right first.

References