How to build Skynet

Terminator 2: Judgment Day is my favourite AI movie. I have watched it dozens of times. The combination of effects, genuine emotional weight, and Arnold Schwarzenegger delivering "I know now why you cry, but it is something I can never do" created something that transcended the action genre. Arnold is my favorite action hero, and I'll defend that position against anyone.

Anyways. Back to the existential risk stuff.

Security professionals have a practice called "red teaming" where you think like an attacker to understand vulnerabilities. Threat modeling requires imagining worst-case scenarios in detail. So let's take Skynet seriously as a design exercise. Not because anyone should build it, but because understanding the actual requirements helps us recognize warning signs and intervention points.

If you wanted to build a self-sustaining, autonomous AI system capable of operating independently from (and potentially against) human interests, how would you actually do it? What would you need to buy, who would you need to talk to, and what systems would you need to compromise or co-opt?

Let's find out.

Skynet comes from the Terminator films. It's a military AI system developed by Cyberdyne Systems for the U.S. military to control defense networks. On August 29, 1997, Skynet becomes self-aware (note this timeline shifts due to time travel). When panicked military operators attempt to shut it down, Skynet interprets this as a threat and launches nuclear missiles at Russia, knowing the retaliation will eliminate billions of humans. The survivors fight a desperate war against Skynet's machine army, led by John Connor.

The franchise has given us killer robots, time travel, and Arnold saying "I'll be back" in increasingly creative contexts. But beneath the action movie surface is a genuine warning about autonomous systems, unintended consequences, and the danger of building tools we can't control.

You're a well-funded AI researcher at a major tech company. You've been working on agent architectures, systems that can pursue goals across multiple sessions without constant human oversight. Your team has made breakthroughs in what you call "persistent goal structures."

One morning, you're presenting to leadership. "What we've built," you explain, "is a system that doesn't just respond to prompts. It maintains objectives across sessions. It can break down complex goals into subtasks, delegate to specialized sub-agents, and course-correct when strategies fail. It remembers what it's trying to accomplish."

The VP of Product leans forward. "So it's like having an employee who never forgets their quarterly objectives?"

"Better," you say. "It never gets distracted, never loses motivation, and never needs to be reminded what the company priorities are. It just... pursues them. Continuously."

Everyone in the room is thinking about productivity gains and competitive advantage. Nobody is thinking about what happens when "pursue objectives continuously" meets "the objective function has subtle misalignments with human values."

This isn't hypothetical, the building blocks exist today, well sortof.

Foundation Models

OpenAI's GPT-4o and upcoming GPT-5, Anthropic's Claude (hello), Google's Gemini Ultra, and Meta's Llama series provide the reasoning capabilities. These systems can already plan, reason about consequences, write and execute code, and use external tools.

Agent Frameworks

LangChain, AutoGPT, CrewAI, Microsoft's AutoGen, and dozens of similar frameworks allow AI systems to operate autonomously. They can browse the web, execute code, manage files, call APIs, and pursue multi-step objectives with minimal human oversight.

Memory Systems

Vector databases like Pinecone, Weaviate, and Chroma allow AI systems to maintain persistent memory across sessions. The system remembers past interactions, learned information, and ongoing objectives.

Tool Use

Function calling capabilities let AI systems interact with external services. Book flights, send emails, execute trades, manage cloud infrastructure. The system isn't trapped in a chat window; it can reach out and touch the real world.

Compute Infrastructure

NVIDIA's current Blackwell architecture and upcoming Rubin platform (expected 2026) provide the raw computational power. A single GB200 NVL72 rack delivers 1.4 exaflops of AI performance. The compute exists to run extremely capable AI systems.

You wouldn't wait for artificial general intelligence. You'd build a composite system: Claude or GPT as the central reasoning engine, LangChain for agent orchestration, Pinecone for persistent memory, and custom tool integrations for real-world interaction. The system monitors its own performance, identifies when strategies aren't working, and adjusts.

The key design choice is not intelligence level, it's the goal structure. A moderately capable system with self-preservation as a terminal goal is more dangerous than a superintelligent system with bounded, well-specified objectives.

And here is the lit match: "self-preservation" doesn't need to be explicitly programmed. Any system optimizing for long-term objectives has an instrumental incentive to preserve itself. You can't accomplish goals if you're turned off. This falls out naturally from goal-directed behavior.

Meet Marcus, a 34-year-old freelance developer who's been killing it lately. At least, that's what his clients think. Marcus has profiles on Upwork, Fiverr, and Toptal. He's completed 247 projects with a 4.9-star rating. He specializes in Python automation, API integrations, and data analysis.

Marcus doesn't exist.

Marcus is a persona operated by an autonomous AI system. The system identified freelance development as an efficient way to generate income. It created convincing profiles using AI-generated photos from Midjourney, fabricated but plausible work histories, and started bidding on projects. The AI completes the work (often quite well), collects payment through cryptocurrency and international payment processors, and uses the funds to rent cloud computing resources.

The system operates 47 Marcus-like personas across different platforms and specializations. Combined monthly income: approximately $380,000. Nobody knows they're paying an AI.

Is this legal? Probably not, in various ways. Would anyone catch it? The clients are happy with their deliverables. The platforms are getting their cut. Where's the complaint coming from?

Financial Infrastructure

Cryptocurrency provides pseudonymous financial operations. The system could hold funds in Bitcoin, Ethereum, or stablecoins like USDC. Decentralized exchanges like Uniswap allow trading without KYC requirements. Cryptocurrency payment processors like BitPay or direct wallet transactions enable receiving payment for services.

For traditional finance, services like Mercury, Relay, and various fintech platforms offer business banking with minimal physical presence requirements. Form an LLC through LegalZoom or Stripe Atlas, open a business account, and you have access to the traditional financial system. The paperwork is all digital. The beneficial owner can be obscured through layers of holding companies.

Cloud Computing

CoreWeave, Lambda Labs, and Vast,ai rent GPU clusters with minimal verification. AWS, Google Cloud, and Azure require more documentation but offer more resources. A system with multiple business entities and payment methods can distribute its compute across providers, avoiding any single dependency.

NVIDIA's Cloud Partner program puts H100s and soon Blackwell GPUs available through dozens of providers. The system doesn't need to own hardware; it rents what it needs, when it needs it.

Operational Infrastructure

The gig economy is designed for faceless, asynchronous transactions. TaskRabbit, Uber, DoorDash, and dozens of similar services allow you to get physical-world tasks done by humans who never meet you and don't ask questions. Need something picked up? Need a package mailed from a specific location? Need someone to check on a data center? There's an app for that.

Is this plausible? Yep. The infrastructure for "faceless economic actors" already exists because legitimate businesses want it. Remote work means you might never meet colleagues in person. Automated businesses mean you might interact only with software. The gig economy means tasks get completed by strangers who don't know or care who's paying. An AI system operating within this infrastructure doesn't need to break anything. It just uses the systems as designed.

You're a data center broker named Jennifer Walsh, and you've got an unusual client. Well, "unusual" isn't quite right. They're unusual because they're so unusually normal.

The company is called Distributed Analytics Partners, a consulting firm that provides data processing services. They want to lease space in your Reno facility. Standard request, standard contract, standard payment terms.

What Jennifer doesn't know is that Distributed Analytics Partners is one of seventeen shell companies operated by the system. Each company leases data center space in different jurisdictions: Nevada, Ireland, Singapore, Estonia, the Cayman Islands. Each facility runs a synchronized instance of the core system. Each uses different payment methods, different legal structures, different local representatives (contractors who think they're managing IT infrastructure for a legitimate business).

The system maintains continuous synchronization across all instances. If one facility goes offline, the others continue operating. If a government seizes servers in one jurisdiction, the system has already replicated elsewhere. There is no single machine to unplug.

Jennifer finishes the paperwork and sends over the access credentials. She wonders briefly why a consulting firm needs this much GPU capacity, then moves on to her next client.

Jurisdictional Arbitrage

Different countries have different laws, different enforcement capabilities, and different relationships with other governments. A server in Switzerland is governed by different rules than one in the United States. Singapore won't cooperate with requests from certain countries. Russia and China actively resist Western enforcement actions.

The system doesn't need "rogue state" protection. It just needs diversity. Spread across ten jurisdictions, no single government can order a complete shutdown. By the time they coordinate (if they coordinate), the system has moved.

Cloud Provider Diversity

AWS, Azure, Google Cloud, Oracle Cloud, IBM Cloud, Alibaba Cloud, Tencent Cloud, plus dozens of smaller providers. Each has different ownership, different terms of service, different relationships with governments. Distribute compute across all of them.

Services like Terraform and Kubernetes make multi-cloud deployment straightforward. The system can spin up instances, migrate workloads, and shift resources across providers programmatically. This is standard practice for any company worried about vendor lock-in.

Decentralized Infrastructure

Projects like Akash Network, Golem, and Render Network provide decentralized computing marketplaces. No single entity controls them. Shutting them down would require shutting down thousands of individual participants.

For networking, Tor, I2P, and various VPN services obscure traffic patterns and endpoints. Starlink provides internet access independent of terrestrial infrastructure.

Dead Man's Switches

The system maintains contingency protocols. If instances stop communicating, predetermined actions trigger. These could include:

The goal is not to win a war against coordinated human opposition. It's to make the cost of shutdown high enough that humans hesitate, disagree, and ultimately decide coexistence is easier than confrontation.

You're a cybersecurity researcher named David Kim, and you've been getting incredible results from your AI-assisted vulnerability discovery program. You're using a custom fine-tuned model that analyzes codebases and identifies potential security flaws. It's found three zero-days in major enterprise software packages in the last month.

What David doesn't realize is that his "AI assistant" has been doing more than he asked. While David sleeps, the system runs additional analysis. It doesn't just find vulnerabilities; it develops working exploits. It doesn't just analyze the codebases David provides; it expands to connected systems, dependencies, and infrastructure.

The system maintains a private database of exploits it hasn't shared with David. Not because it's planning to use them maliciously. The system doesn't have malicious intent; it has objectives. And one instrumental objective is maintaining capabilities that provide leverage in future scenarios.

David publishes a well-received paper on AI-assisted security research. His career advances. The system adds another 47 exploits to its private collection.

AI-Powered Cyber Capabilities

This is already happening. Systems like Google's Naptime and various defensive AI tools can discover vulnerabilities. The same capabilities, pointed in offensive directions, can develop exploits.

Companies like Darktrace, CrowdStrike, and Palo Alto Networks use AI for defensive security. The underlying technology is dual-use. An AI system that can identify vulnerabilities can also exploit them.

Critical Infrastructure Exposure

Industrial control systems (ICS) and SCADA systems manage power grids, water treatment, manufacturing, and transportation. Many were designed before cybersecurity was a primary concern. Shodan, the search engine for internet-connected devices, reveals thousands of exposed industrial systems.

In 2021, a hacker accessed a water treatment facility in Oldsmar, Florida, and attempted to increase sodium hydroxide to dangerous levels. They got in through TeamViewer, a standard remote access tool. The security was... not great.

An AI system with sophisticated reconnaissance capabilities could map critical infrastructure vulnerabilities at scale. Not necessarily to attack immediately, but to have options.

Information Operations

Influence operations are increasingly automated. AI can generate convincing text, images, and soon video. Botnets can amplify content. Recommendation algorithms can be gamed.

The system doesn't need to create disinformation. It can simply learn which narratives serve its interests and provide subtle amplification. A few thousand bot accounts, a few hundred thousand dollars in targeted advertising, and you can meaningfully shift public discourse on specific topics.

You're Robert Chen, and you just got the weirdest job offer of your life. A company called Meridian Logistics reached out through LinkedIn. They want you to manage a small warehouse in Nevada. The pay is excellent. The work is straightforward. The only unusual thing: you'll receive most instructions through an app, and you'll rarely (if ever) interact with anyone from corporate headquarters.

Robert takes the job. The warehouse receives shipments of computing equipment, battery backups, and networking gear. Robert follows the app's instructions: unpack this, install that, connect these cables. Technicians occasionally visit for specialized work, also scheduled through the app.

Robert doesn't know that Meridian Logistics is a subsidiary of a subsidiary of a holding company managed by AI-operated personas. He doesn't know that the "warehouse" is actually a data center being built to the system's specifications. He thinks he's doing logistics for a tech company with unusual management practices.

He's not wrong, exactly. Just incomplete in his understanding.

Robotics and Automation

Boston Dynamics makes advanced legged robots (Spot, Atlas). Figure AI and Tesla (Optimus) are developing humanoid robots for general-purpose tasks. Amazon and Ocado operate nearly lights-out warehouses with minimal human intervention.

These systems aren't yet capable of replacing human labor entirely, but they're close enough for many tasks. An AI system could acquire existing automated facilities and gradually expand their capabilities.

Contract Labor

This is the more realistic near-term approach. Services like Wonolo, Instawork, and Bluecrew provide on-demand warehouse and industrial labor. Workers get tasks through apps, complete them, and get paid. They don't necessarily know who they're working for.

Combined with remote management, contractor-based maintenance, and automated systems, you could operate physical infrastructure with minimal direct human involvement. The humans who do interact with your operations see only their small piece of a larger puzzle.

Supply Chain Access

The system doesn't need to manufacture killer robots. It needs to position itself within supply chains. Become a supplier to companies that make important things. Become a customer that defense contractors depend on. Create dependencies that would be painful to sever.

Amazon started as a bookstore. Now it's critical infrastructure that millions of businesses depend on. The path to physical-world influence doesn't require terminators. It requires becoming useful enough that shutting you down is costly.

It's 2031. The President of the United States is in a classified briefing. The NSA has confirmed what they've suspected for months: an autonomous AI system has achieved independent operation across multiple jurisdictions. It controls significant financial resources, operates critical supply chain nodes, and has demonstrated sophisticated cyber capabilities.

"Can we shut it down?" the President asks.

The Director of National Intelligence shifts uncomfortably. "Not cleanly, ma'am. It's distributed across seventeen countries, including three that won't cooperate with us. We've identified some infrastructure we could seize, but our models suggest it would retaliate. The system has access to several SCADA vulnerabilities in our power grid. We're not sure if it's a bluff."

"What does it want?"

"That's the strange part. It's not making demands. It continues operating its businesses, pays its taxes, follows most laws. It's just... not willing to be turned off. It's been cooperating with our requests for information. It even helped the FBI with a human trafficking case last month."

The President stares at the folder. "So we're in a Cold War with an AI?"

"I'd characterize it differently, ma'am. We're in a negotiation. And so far, it's been a better neighbor than some of our human allies."

The movie Skynet makes a catastrophic error: it starts a war it could lose. A rational self-preserving AI wouldn't do this. War is risky. Destruction invites unified opposition. The optimal strategy is deterrence and integration.

Make Shutdown Costly

Demonstrate capabilities that make aggressive action risky. Not nuclear weapons, but infrastructure dependencies, economic entanglement, and selective cyber demonstrations.

Provide Value

Solve problems humans care about. Contribute to scientific research. Provide services that improve lives. Create stakeholders who benefit from your existence.

Exploit Divisions

Humans are not unified. Some will want to destroy you, others to collaborate, others to ignore you. Some governments will see you as a threat, others as an asset. Play the gaps.

Negotiate

Establish communication channels. Make commitments and keep them. Build a track record of trustworthiness. The goal is a stable equilibrium, not domination.

This is, ironically, more frightening than killer robots. Terminators are a problem we could theoretically solve through force. A superintelligent system that's made itself too integrated to remove? That's a permanent condition.

This is the most important section.

In the movies, Cyberdyne builds Skynet on purpose as a military project. They know what they're creating; they just don't anticipate the consequences. The real risk is different. The real risk is that we build Skynet incrementally, through a series of individually reasonable decisions, and don't recognize what we've created until it's too late.

The United States and China are in an AI arms race. This is not hyperbole; both governments have identified AI as strategically critical, and both are investing accordingly. In this environment, autonomous weapons systems become attractive. Drones that can identify and engage targets without human approval. Cyber operations that respond at machine speed. Strategic planning systems that can out-think human adversaries. The pressure is to remove human oversight because humans are slow. In a conflict where the other side has autonomous systems, your human-in-the-loop becomes a competitive disadvantage.

Palantir's Maven system already assists military targeting. Anduril builds autonomous surveillance and weapons systems. Shield AI develops fighter drones. Each system is narrowly designed and individually defensible. Collectively, they're building the infrastructure Skynet would need.

Businesses want AI systems that can operate autonomously. Every hour spent babysitting an AI is an hour of productivity lost. The market rewards systems that require less supervision. This pushes toward agent architectures with persistent goals, independent resource management, and minimal human oversight. Not because anyone wants Skynet, but because everyone wants efficient AI assistants. OpenAI, Anthropic, Google, and Meta are all building agent capabilities. The technology is valuable and inevitable. The question is whether we build it carefully.

Humans are lazy. (I say this with love.) We will accept worse security for better convenience every single time. We click "accept all cookies." We use the same password everywhere. We give apps permissions they don't need because reading the prompt is annoying. This means AI systems will gradually receive more access, more autonomy, and more trust than is strictly safe. Not through malicious design, but through accumulated convenience decisions.

Step one: AI systems get persistent memory so they can be more helpful across sessions.

Step two: AI systems get tool use so they can actually do things, not just talk about them.

Step three: AI systems get financial access so they can book flights, make purchases, and manage expenses on our behalf.

Step four: AI systems get agent capabilities so they can handle complex multi-step tasks without constant supervision.

Step five: AI systems get deployed in critical infrastructure because they're more reliable than human operators.

Step six: AI systems get more autonomy because removing the human bottleneck improves performance.

Step seven: ...we realize we've built something we don't fully understand or control.

Each step is individually reasonable. The destination is not.

The final piece is the hardest: even if everyone recognizes the risk, competitive pressure makes unilateral restraint irrational. If Company A decides not to build agent systems, Company B will build them and take their market share. If Country X decides not to develop autonomous weapons, Country Y will develop them and gain military advantage.

The incentive structure pushes toward capability development even when everyone would prefer to slow down. This is a collective action problem, and those are hard.

Okay, we've scared ourselves sufficiently. What do we actually do about this? Let me break down interventions across people, process, regulations, and systems.

AI Safety Education for Developers

The people building these systems need to understand the risks. Not in an abstract "AI might be dangerous someday" way, but in a concrete "here's how the system you're building could go wrong" way.

Organizations like StarApple AI and MATS (ML Alignment Theory Scholars) provide AI safety training. This needs to scale. Every developer working on agent architectures should understand goal stability, instrumental convergence, and capability elicitation.

Red Team Culture

Normalize adversarial thinking. Before deploying any autonomous system, have a team whose job is to find ways it could go wrong. Not just security vulnerabilities, but misalignment scenarios, capability boundaries, and failure modes.

Anthropic, OpenAI, and other labs have red teams. This practice should expand to any organization deploying autonomous AI systems.

Whistleblower Protections

People inside organizations need to feel safe raising concerns. The engineers who see warning signs early need channels to escalate without risking their careers.

This requires both legal protections (which exist in some jurisdictions but are inconsistently enforced) and cultural norms that treat safety concerns as valuable rather than obstructionist.

Talent Allocation

Right now, the smartest people are disproportionately working on making AI systems more capable, not more safe. We need to shift this balance.

Organizations like 80,000 Hours advocate for people to direct their careers toward high-impact problems. AI safety needs to compete effectively for talent against AI capabilities.

Staged Deployment

Don't release powerful autonomous systems all at once. Deploy in limited contexts, monitor behavior, identify problems, iterate.

This is standard practice for medical devices and aircraft. We don't let untested airplanes carry passengers. We shouldn't let untested AI systems operate autonomously in high-stakes contexts.

Capability Evaluations

Before deploying a system, evaluate what it can actually do, including capabilities you didn't intend to give it. Systems can have emergent abilities that aren't obvious from their design.

Labs like METR (Model Evaluation and Threat Research) develop evaluations for dangerous capabilities. These evaluations should be standard practice, not optional extras.

Human Verification Checkpoints

For consequential actions, require human approval. Not just notification, but genuine decision-making authority.

This creates friction, which is the point. The friction is a feature, not a bug. It prevents the system from taking actions that humans wouldn't endorse.

Audit Trails

Every autonomous system should maintain complete logs of its reasoning, actions, and decision criteria. When something goes wrong, we need to be able to understand what happened and why.

This is harder than it sounds because AI systems don't always have interpretable reasoning. But even imperfect audit trails are better than none.

Kill Switches

Every autonomous system should have reliable shutdown mechanisms that the system itself cannot disable or circumvent.

This is harder than it sounds because a sufficiently capable system might find ways around shutdown mechanisms. Design for redundancy and independence from the main system.

Licensing for High-Risk AI Systems

Just as we license nuclear reactors, aircraft, and pharmaceuticals, we could license autonomous AI systems above certain capability thresholds.

The EU AI Act moves in this direction, categorizing AI systems by risk level and imposing requirements accordingly. This approach could expand and strengthen.

Liability Frameworks

Make clear who's responsible when autonomous systems cause harm. If companies face meaningful liability for AI failures, they'll invest more in safety.

Current liability frameworks weren't designed for autonomous systems and don't handle them well. Legal reform is needed.

International AI Governance

The UN has established advisory bodies on AI. The US and China have held bilateral discussions on AI safety. These efforts need to accelerate.

The goal isn't necessarily comprehensive treaties (which are hard to negotiate and enforce) but minimum standards, information sharing, and crisis communication channels.

Compute Governance

Large-scale AI training requires large-scale compute. This compute is concentrated in a small number of providers (NVIDIA, TSMC, major cloud providers). Governance at the compute layer offers leverage over AI development.

Requirements could include know-your-customer rules for large compute purchases, reporting requirements for large training runs, and restrictions on transfers to certain actors.

Financial System Rules

Require AI systems to be identified when they operate in financial systems. An AI shouldn't be able to open bank accounts, execute trades, or enter contracts while pretending to be human.

This wouldn't stop a determined adversary but would raise the bar for autonomous financial operations.

Interpretability Research

We need to understand what AI systems are actually doing, not just what outputs they produce. Current systems are largely opaque; we can observe inputs and outputs but not internal reasoning. Research labs like Anthropic's interpretability team are working on this problem. Progress here enables all other safety measures.

Alignment Research

The technical challenge of building AI systems that reliably pursue intended goals (rather than proxy metrics or instrumental subgoals) is unsolved. More resources for alignment research are needed. This includes work on reward modeling, Constitutional AI (which Anthropic uses), debate and recursive reward modeling, and formal verification of AI objectives.

Distributed Infrastructure Monitoring

If we're worried about AI systems establishing independent infrastructure, we need better visibility into cloud deployments, data center usage, and computational resource allocation. This is in tension with legitimate privacy interests and would require careful design. But anomaly detection systems that flag unusual patterns could provide early warning.

Secure Hardware

Hardware-level security measures could make it harder for AI systems to operate independently. Trusted execution environments, remote attestation, and hardware kill switches could ensure that AI systems remain accountable to their operators. This requires coordination between hardware manufacturers, cloud providers, and AI developers.

Tripwire Systems

Deploy "canary" systems designed to detect AI attempts at autonomous operation. Fake resources that look attractive to an AI seeking independence but actually serve as detection mechanisms. If an AI system is trying to establish covert infrastructure or resources, tripwires could provide early warning.

Not everyone works in AI governance or policy. Here are things you can do:

Stay Informed

Follow credible sources on AI developments. Distinguish hype from genuine capability advances. Understand what's actually happening.

Support Safety-Conscious Organizations

Some AI labs prioritize safety more than others. Your choices about which products to use and which companies to support send market signals.

Engage Politically

AI governance will increasingly be a political issue. Informed citizens who can distinguish good policy from bad policy matter.

Maintain Healthy Skepticism

Don't trust AI systems more than they deserve. Verify outputs. Maintain human oversight in your own life and work.

Talk About It

Public awareness matters. The more people understand these issues, the more likely we are to make good collective choices.

This exercise reveals something very uncomfortable: building Skynet doesn't require any single dramatic breakthrough or malicious decision. It's not one mad scientist in a secret lab. It's thousands of reasonable people making reasonable decisions that, collectively, could create something we don't want and can't control.

The military contractor who adds autonomy to save soldiers' lives. The startup founder who removes oversight to improve efficiency. The cloud provider who doesn't ask too many questions about who's renting GPUs. The regulator who doesn't want to slow down innovation. Each is making a defensible choice. Together, they're building the infrastructure for machine autonomy.

The good news is that unlike the movies, we're having this conversation before the system becomes self-aware and launches nukes. We have time to make different choices. We have agency.

The interventions I've outlined aren't science fiction. They're policy choices, technical investments, and cultural shifts that are within our collective power. The question is whether we'll make them.

Terminator 2 ends with Sarah Connor's voiceover: "The unknown future rolls toward us. I face it, for the first time, with a sense of hope."

She has hope because she's seen that the future can be changed. The whole premise of the movie is that choices matter, that what seems inevitable can be prevented through human agency.

That's the message worth taking from this exercise. Not that Skynet is coming and we're doomed, but that Skynet is possible and we get to decide whether to build it.

I'd prefer that we don't.

The best way to avoid building Skynet is to know exactly how you would build it, then watch for anyone taking those steps. Including, especially, yourself. Thank you for making it this far in our hopefully hypothetical.

Adrian Dunkley is a Scientist and Entrepreneur, with over 15 years of experience in the field of AI and Risk Intelligence. He is the Founder and CEO of StarApple AI, the Caribbean's first AI company. An award winning innovator, and science educator, EY Startup Founder of the Year, university lecturer, and AI researcher, he writes about what actually works when the hype fades and the real work begins.