I have avoided the fray as long as I could but the intensity of doomsday news hype and the calls by tech elites to slow down AI development is getting to the point of ridiculous. Whether I am watching CNN, BBC, CBC, CNBC or even Yahoo Finance, I cannot escape the doom, the fear, the speculation and the hype.
I guess it might be time to build a bunker. And yes, that was sarcasm.
I have been doing this long enough and dealing with AI long enough to know that there is a massive convenient diversion happening here in the run-up to major IPOs by AI frontier companies. You currently have a huge backlash against AI and data centers combined with the fears relayed by insiders at these companies, and it is all acting as a handy smokescreen.
Can AI do damage? Most certainly. Will it cause the doomsday scenarios that are dominating the hype cycle right now? No.
AGI? I Think Not
The technology today is powerful and capable but it is nowhere close to AGI and the frontier model vendors have likely realized they’re hitting a scaling wall. Many scientists and skeptics have been warning about this for a long time. The current AI, should I say, Gen AI craze cannot reach AGI and more fundamental breakthroughs are required. The idea that we could achieve AGI just through massive scale and burning through cash, energy and water was never going to work.
And now that circular incestuous trade is finally catching up to those who played off it. You clearly know who they are. That’s not to say that the AI economy is fake and there is no real value in the tech. On the contrary, I believe in the transformative capabilities of AI. However the irony, is that it distracts from the real capital getting unnecessarily sunk into a concentrated cycle of chip companies funding infrastructure builds, raising debt, pushing frontier model evolution to raise yet more debt and still not return the ROI that was promised.
What is actually real is that AI is a powerful tool and in the wrong hands it can do severe damage. We saw that firsthand when both Anthropic and OpenAI advertised how their systems broke into other companies and compromised intellectual capital as if these systems were acting on their own. They were not acting on their own.
This comes down to incompetence, negligence or maliciousness on the part of anyone who builds around these tools. AI cannot act of its own volition. It is coded to do something and you had better code it with good controls. Accountability is a key concept that seems to be getting lost in the hype, and it is something we explored recently with former U.S. Navy operational commanders, where absolute accountability remains the guiding principle.
Humans are the real problem here, not the technology.
Humans Behind the Steering Wheel
Just last week, I posed a theoretical point to all of the AI systems I use including ChatGPT, Gemini and Claude that “AI has solved exactly zero math problems.” That was purposely designed to incite a reaction and it’s based on the recent controversy surrounding AI and the Navier-Stokes equations, but it holds a truth I firmly believe. The intelligence is entirely human, not artificial. Even if the AI comes up with a perceived novel idea or concept, it is entirely upon the human to recognize and pull on the thread.
These systems are based entirely on human intelligent prompting, interaction and/or structured guidance because they are just powerful semantic processors. Now in the hands of the wrong humans, that is where the danger lies.
AI is a brilliant and scary accelerator … for humans. It allows us mere mortals to augment our capabilities in research as well as building stuff. We can decide whether or not we want to apply that augmentation and acceleration for good or evil. And we can decide whether or not we want to be smart about it in the first place. The doomsday scenarios have ranged from building biological weapons to incorporating AI into military applications to bringing down the entire internet.
I believe these scenarios are theoretically plausible, but there are a mountain of challenges to overcome. At CharliAI we’ve seen this first hand as we’ve found that AI is a poor engineer, a poor designer, a poor architect, a poor scientist, and a poor mathematician. You’re not replacing this level of intelligence any time soon even with recurrent depth and recursion.
If we use the takedown of the Internet as an example, the real danger lies in combining all the malware knowledge (easily digested through an LLM) with the execution of agentware (i.e. code, just code, nothing more than code with memory). AI can easily accelerate the development of worm, trojan and DDoS attacks. Moreover, the Internet is also highly concentrated around common infrastructure and a handful of critical service providers. Think ubiquitous infrastructure like Kubernetes and concentrated services like Cloudflare. Just consider the impact from the Cloudflare disruption late last year and you’ll quickly realize you don’t need a broad attack surface to be dangerous and disruptive.
While a targeted attack can cripple centralized infrastructure, AI acts as a massive force multiplier that allows everyday threats like identity theft, data breaches, and ransomware to go incredibly broad.
None of this destroys the internet, but that’s not the point. AI is still dangerous in the wrong hands whether or not it can be stopped and recovered.
Safety Is Not The Answer
Here is the rub. These systems can be controlled. They can be managed, watched and even shut down quite easily. But the tech elites and frontier labs keep confusing the market by screaming about safety when they should be talking about control. Those are two completely different paradigms.
Safety, albeit important, is being used largely as public relations theater to counter the backlash. It relies on refusal classifiers, content filters and ethical guardrails trying to guess if a block of text or a user request is harmful. It is just vibes based on language and it is inherently flawed because it will always over trigger or under trigger. When you rely on a probabilistic safety layer to protect your enterprise you are essentially crossing your fingers and hoping the model behaves today.
Safety initiatives can degrade the very capabilities people want because they try to “hard-code” behavioral suppression directly into model weights or semantic guardrails, rather than governing execution at runtime.
Control and security are hard-nosed systems engineering. Control is deterministic. It does not care about the mood of the model or what prompt someone typed in. It is about strict policy enforcement points, zero-trust data boundaries and immutable execution limits. When a model tries to execute a workflow or access a system, a deterministic control plane evaluates the action against hard rules before anything runs. The model acts as a powerful reasoning engine to generate candidate actions but the deterministic layer acts as the absolute authority that dictates whether that action is permitted.
That is how you prevent premature execution and state propagation failures. You stop asking probabilistic safety classifiers to do the work of hard infrastructure access controls. You put a human on the loop and you build an architecture where the AI has zero capacity to take an unauthorized real-world action regardless of what it hallucinates or how it was prompted.
Plain and simple … stop, just stop arguing about whether the machine is alive. In the enterprise you need trust, and with the right deterministic engineering, that trust is very achievable.