The AI Problem Is Worse Than Big AI

Bad actors do bad things.

We keep worrying about what OpenAI, Anthropic and Google might do with artificial intelligence. Meanwhile, anyone with enough hardware can download it, remove the guardrails and do almost whatever they want.

So I think this is an interesting thing. We seem to be very worried about Big AI being bad. OpenAI is a cesspit. Anthropic is literally the demon spawn. Google is apparently taking evil to a whole new level. Depending on which corner of the internet you happen to wander into, there is always a corporation currently responsible for the impending collapse of civilisation.

And look, some of the criticism is absolutely warranted. These are extraordinarily powerful companies building extraordinarily powerful technology, often faster than governments, regulators and society can properly understand what they’re building. There are legitimate questions around privacy, copyright, employment, misinformation, surveillance, environmental impact, monopolisation and the increasingly uncomfortable amount of power being concentrated into a handful of companies.

But I’d like to propose another outcome, which is pretty much the normal with me at this point. Big AI is only partly the problem. Unfortunately, the actual problem might be significantly worse, because we keep talking about AI as though OpenAI, Anthropic, Google and the other frontier labs are holding the keys to the machine.

They’re not. The machine has already escaped.

Axiom Accidentally Sent Me Down a Rabbit Hole

As I’ve mentioned before, I’ve been working on my own AI orchestrator called AXIOM. I’ve written an entire Brainwaves article about it already, so I won’t repeat the whole thing here. The important part is that I’m now onto version three, and building it has forced me considerably further down the AI rabbit hole than I expected to go.

AXIOM isn’t really an AI model itself. It’s closer to an orchestration system. Different models, tools, machines and agents can be given jobs, work autonomously, communicate with other systems and perform increasingly complicated tasks without me manually sitting there typing prompts into ChatGPT. Once you start experimenting with this stuff properly, you realise it really isn’t that difficult anymore.

Obviously there is a technical barrier. You need to understand computers. You need some programming knowledge. You need the hardware. You need to understand models, context windows, inference, APIs, agents and a bunch of other wonderfully nerdy things. But none of this requires being a Google engineer. I can do it, and that’s where my existential crisis begins, because if I can do it, plenty of other people can too.

I’ve Been Avoiding Writing About This

I’ve actually been avoiding writing Brainwaves articles about the state of the world lately. Everything feels too heavy. Every time I turn on the news there is another war, political crisis, economic problem, technological nightmare, environmental disaster or some other horrible thing happening somewhere.

Obviously humanity has always been like this. We’re very good at romanticising the past while conveniently forgetting that most of human history consisted of disease, war, famine and somebody invading somebody else because their king wanted another paddock. But the information environment is different now. Everything happens in front of us, constantly, and AI has somehow managed to become simultaneously one of the things that makes me most optimistic about the future and one of the things that makes me most concerned about it.

That’s the annoying part: I love this technology. I can learn things in hours that previously would have taken weeks. I can explore disciplines outside my existing expertise, build software faster, analyse enormous datasets and prototype stupid ideas simply because I wondered whether they were possible. It is genuinely incredible.

Then you pull back the curtain slightly further and think: oh. Oh no.

You Can Run This Stuff Yourself

I don’t think most people realise how accessible local AI has become. When the average person thinks about AI, they probably think about ChatGPT, Claude or Gemini. You type something into a website, your request travels somewhere into a gigantic data centre, the model processes it, the company applies its policies and safety systems, and something comes back.

That creates a comforting assumption that someone, somewhere, is in control.

Except you don’t actually need the company.

You can download models and run them locally. Completely locally. No internet connection required. The best model I currently run on my laptop is GPT-OSS 120B. It consumes an absolutely ridiculous amount of memory when running, but the fact that I can run a roughly GPT-4o-class open-weight language model on a laptop sitting in front of me is absurd.

Not a data centre or some university supercomputer. My laptop.

Once the model is sitting on your computer, you can connect it to agent software, give it access to files, development environments and tools, let it write code, analyse information and perform tasks. You suddenly have this strange little synthetic worker sitting inside your computer that doesn’t necessarily need an internet connection or a corporation sitting between you and it.

Which is incredibly exciting, until you discover abliteration.

Abliteration Is Where This Gets Uncomfortable

“Abliteration” is one of those words that sounds like it was invented by someone on Reddit at 3am, but the basic concept is fascinating and slightly horrifying.

AI models contain mechanisms that influence whether they comply with certain requests. Researchers and hobbyists have discovered techniques for modifying models in ways that substantially weaken or remove some of those refusal behaviours. The technical details get increasingly complicated, and at a certain point involve mathematics that causes my brain to politely leave the room, but the practical outcome is simple: models can be modified so that safeguards become significantly weaker.

Now combine that with everything I just described. Local models, no central provider, no API moderation, agent frameworks, autonomous execution, powerful consumer hardware, open-source software and models deliberately modified to refuse fewer instructions.

That’s the bit that worries me, because suddenly our entire conversation about “AI safety” starts looking strangely incomplete.

We Keep Asking Corporations to Control Something They No Longer Control

Governments can regulate OpenAI. They can regulate Anthropic. They can regulate Google. They can require age verification, restrict websites, force platforms to remove certain models, block services, write enormous pieces of legislation with exciting names and hold parliamentary hearings where politicians ask technology executives questions about things neither side particularly wants to explain properly.

But what exactly do you regulate when somebody can download the model?

Australia can block access to a website, but a sufficiently motivated person can find another route. A model can disappear from one hosting service and appear somewhere else. Files can be copied. Software can be forked. Models can be modified. Once something is sitting on somebody’s computer, the government cannot magically reach through the screen and put the safety features back.

This is the part of the AI debate I think we’re dramatically underestimating. We are treating AI like a service when, increasingly, it is infrastructure.

And infrastructure spreads.

My Client Got Hacked, Which Somehow Made This Worse

The rabbit hole actually started somewhere much more boring: a client got hacked.

That made me start thinking seriously about whether I should be doing more proactive security testing across the websites and infrastructure I manage. Attackers are already using automation and AI-assisted tooling, so eventually defensive security reaches an uncomfortable conclusion: refusing to understand offensive techniques doesn’t make your systems safer. It just means somebody else understands your vulnerabilities better than you do.

So I’ve been researching how autonomous agents could assist with authorised penetration testing against systems I manage. The objective is completely legitimate: find vulnerabilities before somebody malicious does, patch them, improve monitoring and make the websites safer. That’s a genuinely good application of the technology.

But while researching what these systems are capable of, you inevitably discover what the exact same underlying technology can do when you remove the words “authorised” and “defensive”. The same autonomous architecture that can continuously inspect my infrastructure for vulnerabilities could be adapted by someone else to continuously search for vulnerable infrastructure to attack.

The underlying capability doesn’t understand morality. We add the morality.

And Cybersecurity Is Only One Example

This problem isn’t limited to hacking. Generative image systems can also run locally. Video systems are improving. Voice cloning is improving. Language models are improving. Automation frameworks are improving. Agents are gaining access to browsers, operating systems, terminals and external tools.

And safeguards are not universal.

We’ve already seen deeply disturbing cases involving AI-generated sexualised deepfakes, including reports involving students and school communities. That genuinely makes me feel sick, and it demonstrates the central problem perfectly. You cannot solve malicious human behaviour by asking a model nicely not to participate, especially when another version of that model exists without the restriction.

Hacking is one thing. Synthetic media is another. Fraud, impersonation, misinformation, automated harassment and whatever other depraved use somebody thinks of next all exist on the same spectrum. AI hasn’t invented human depravity. We’ve been doing perfectly well in that department for thousands of years.

What AI changes is how much one person can do.

The Same Technology Can Also Be Incredible

This is where I keep getting stuck, because I don’t think “AI is evil” is an intellectually serious conclusion either.

I’m currently exploring AI-assisted penetration testing to make client infrastructure safer. That’s good. I’m building systems that analyse businesses and help them understand where they should invest, where opportunities exist and how they can make better decisions. That’s good too. Even calling some of that “AI” is questionable because half the time we’re really talking about traditional statistics, semantic analysis, machine learning and data engineering with AI slapped on the front because apparently everything needs AI branding now.

I’m also experimenting with pipelines that combine my photography with computer vision, photogrammetry, Gaussian splatting and generative systems to create better digital representations of client products. That’s incredibly cool. Similar technology can be used for architecture, historical preservation, education, training, simulation, accessibility, medicine, engineering and scientific research.

You can build systems that help people learn. Systems that help small businesses compete with enormous companies. Systems that make software accessible to people who could never previously afford development teams. Systems that analyse mountains of information and surface things humans would have missed.

The technology isn’t inherently evil.

That’s almost the problem.

AI Isn’t the Enemy

AI is still a tool, but I think we’ve created a fundamentally different category of tool.

A hammer can build a house or cave someone’s skull in. The difference is that a hammer cannot independently manufacture another hammer, research better skull-caving techniques, write the instructions, translate them into 40 languages, produce the marketing campaign and then operate continuously while you sleep.

AI can increasingly participate in that entire chain.

That’s what changes the equation. The problem isn’t simply that AI can be misused, because every technology can be misused. The problem is scale. One malicious person can suddenly produce the output that previously required dozens, hundreds or potentially thousands of people. One person can generate enormous quantities of misinformation, automate scams, manufacture fake identities or generate synthetic images, voices and videos. One technically capable attacker can potentially automate parts of cyberattacks that previously required significantly more expertise and labour.

And the marginal cost keeps approaching zero.

That is different.

Regulation Doesn’t Solve the Fundamental Problem

I’m not saying regulation is pointless. There are absolutely areas where governments should regulate how companies collect data, deploy AI, handle people’s likenesses, protect children, disclose synthetic media and use automated systems in high-risk environments. There should be consequences when people use these systems to commit crimes, and there should be responsibilities placed on corporations deploying powerful technology at enormous scale.

But regulation cannot make the capability disappear.

That’s the uncomfortable truth. The genie isn’t going back into the bottle.

We can regulate commercial nuclear reactors because building a nuclear reactor in your bedroom is somewhat inconvenient. AI doesn’t have that limitation. The hardware required to run increasingly capable models is becoming cheaper and more accessible. The models themselves can be copied infinitely. Knowledge spreads. Code spreads. Research spreads. Once a capability exists publicly, preventing determined malicious actors from accessing it becomes extraordinarily difficult.

Which means the safety question eventually moves beyond “How do we stop AI companies from doing bad things?” It becomes something far more difficult:

What happens when millions of individuals have access to capabilities that previously belonged to governments and enormous corporations?

I don’t think we’ve properly processed that yet.

This Is Why I’m Having an Existential Crisis

Because I build something cool, then discover somebody can use roughly the same thing for something horrible.

I build autonomous agents because I want software that can help operate my infrastructure. Someone else can build autonomous agents designed to attack infrastructure. I experiment with generative media because I want better creative pipelines. Someone else uses generative media to humiliate, exploit or deceive people. I use local models because I want privacy, independence and control over my own technology. Someone else wants local models because nobody can tell them no.

Am I the problem? Of course not. But that’s almost beside the point.

The barrier is simply becoming very low.

Someone could build autonomous malicious software. Someone could use unrestricted models to seek information mainstream systems deliberately refuse to provide. Someone could build an automated fake influencer, generate the photos, write the captions, manufacture the personality, generate videos, reply to messages and operate the account indefinitely.

Not because they’re some elite hacker sitting in a dark room surrounded by seventeen monitors, but because the components required to do these things are becoming increasingly accessible.

That is the scary part.

So What Exactly Are We Supposed to Do?

No idea. Excellent conclusion, I know.

I don’t think anyone really knows. We’re trying to build rules around something that is evolving faster than the rules can be written, and personally I’m stuck in this strange position where I don’t want to stop exploring the technology simply because other people can misuse it. That would be ridiculous. The same curiosity that led me into this existential crisis is also showing me ways of making systems safer.

Maybe that is actually where the responsibility sits. Not pretending the dangerous capabilities don’t exist, but understanding them. Testing them responsibly. Building defences. Creating better verification systems. Teaching people how synthetic media works. Hardening infrastructure. Making provenance easier to establish. Developing ethical norms around how these systems should be used.

And, importantly, maintaining some kind of moral framework when the technology itself doesn’t provide one.

Which brings me, somewhat predictably, back to Christianity.

Thank God I Have Some Kind of Moral Baseline

I don’t mean that as “Christians good, everyone else evil.” History would make that argument collapse almost immediately. I mean that personally, I’m increasingly grateful that I have something outside technology against which I can measure what I’m doing.

Because technological capability is a terrible moral framework.

“Can I build it?” is not the same question as “Should I build it?” “Can this be automated?” is not the same question as “Should this be automated?” And “Can this make money?” is definitely not the same question as “Should this exist?”

AI doesn’t solve those questions. If anything, it makes them considerably more important. When your capabilities increase, your responsibility increases with them.

Perhaps that is the part of the AI conversation we should be focusing on more. Not simply what the machine is capable of, but what we are capable of doing with it.

White Hat AI

So maybe that’s where I land: white hat AI.

I’m going to keep experimenting. I’m going to keep building AXIOM. I’m going to keep downloading models onto increasingly ridiculous computers. I’m going to keep experimenting with autonomous agents, local models, security systems, data analytics, computer vision, generative media and whatever other absurd technology appears next Tuesday.

And I’m going to see how far I can push it.

But there needs to be a line. Use offensive security knowledge to improve defensive security. Use autonomous agents to protect infrastructure rather than attack it. Use generative media to create rather than exploit. Use data to help businesses make better decisions rather than manipulate people. Understand the dangerous capabilities precisely because pretending they don’t exist leaves those capabilities exclusively to people who don’t care about the consequences.

Fight the depravity rather than contribute to it.

Because I don’t think we’re stopping this thing.

Five years from now? I have absolutely no idea what the world looks like. At the current rate, I’m not entirely confident what it looks like one year from now. That scares me, but it also fascinates me. Perhaps that’s the uncomfortable reality of living through a technological revolution: the same invention can make you incredibly hopeful and deeply concerned at exactly the same time.

AI isn’t God. It isn’t the devil either. It’s something much more uncomfortable.

It’s a tool that increasingly amplifies whatever the person holding it already intended to do.

And we’ve just handed that tool to humanity.

If you’ve met humanity, you’ll understand why I’m concerned.