We are closer than ever — definitely closer than yesterday — to finally creating real intelligence: a system that not only predicts likely answers but can seem to “think,” reflect, and perhaps even become self-aware.
According to recent research and reporting, state-of-the-art models already rival strong human performance on many exams and benchmarks. More importantly, they no longer act like mere autocomplete machines: modern systems can perform intermediate calculations, call tools, and verify parts of their own work. That shift — from pure next-token prediction toward reasoned problem-solving — is a genuine breakthrough.
Therefore, it’s reasonable to imagine that at any moment — tomorrow, next year, or (depending on your outlook) within ten years — “artificial” intelligence may stop feeling artificial. It could invent new algorithms, optimise itself, and — who knows — cross whatever fuzzy line we associate with consciousness.
How We Actually Think (and Hallucinate)
Let’s quickly look at how we humans think and sometimes hallucinate (mix facts, guess answers, or create false information — intentionally or subconsciously). Most of us know our times tables, right? 9 x 9 = 81, 5 x 7 = 35, and so on. We don’t need to compute them each time; we remember. When we see a question, we instantly and subconsciously evaluate it. If we know the answer, we give it — much like how language models output fluent answers they’ve effectively “memorised“.
But when memory isn’t enough, we switch to analytical reasoning. 20 x 20 = 400 that’s still easy but19×19 may require a quick calculation if we don’t remember it. Modern systems can do this too: they fall back to step-by-step work, tool use, or external checks when recall fails. Because their accessible knowledge is vast, they often don’t need to “do the algebra” for simple queries — yet they can when needed.
Of course, humans mix up facts, forget details, and make confident mistakes. Models do as well. Our errors feel familiar: we round, we guess, we confabulate under pressure. Watch Who Wants to Be a Millionaire? — nearly every contestant eventually guesses. Current models behave similarly, just at industrial scale.
Are We Doomed? Maybe Not — But Speed Matters
Few predicted, in detail, how COVID-19 would reshape daily life — even though fiction and experts had warned about pandemics for years. Likewise, pop culture has primed us for rogue AIs — from Skynet to self-aware machines — but stories aren’t destiny.
What is clear is the pace. Investment and progress feel relentless. We’re on a speeding train with no brakes. Maybe it’s another dot-com bubble. Maybe not. Hype comes and goes, but underlying capabilities continue to ratchet upward.
New releases arrive from every direction: newer versions of ChatGPT, Grok from xAI, and models from China like DeepSeek. The conversation is shifting from “AI” to AGI — artificial general intelligence — systems that aim for human-level versatility, even if they remain artificial in origin.
Memory vs. Reasoning vs. Agency
To understand where we might be headed, separate three layers:
- Memory/Knowledge. Storing and retrieving facts, patterns, experiences.
- Reasoning/Planning. Breaking tasks into steps; using tools, code, and search; checking intermediate results.
- Agency/Goals. Choosing what to do next, pursuing objectives over time, adapting strategies.
Today’s systems are excellent at (1), getting rapidly better at (2), and beginning to simulate aspects of (3) via agents that plan, schedule, and act across multiple steps. If (3) becomes robust — coherent goals plus long-horizon planning — we’ll feel much closer to “real” intelligence, even without definitive proof of consciousness.
Why Hallucinations Happen (to Us and to Them)
- Pressure and gaps. When humans lack a fact, we often guess. Models fill gaps with plausible-sounding text.
- Ambiguity. Vague prompts (or questions) increase the odds of confident but wrong answers — for people and machines.
- Optimization targets. Humans optimise for social acceptance, speed, or self-image; models optimise for the training objective. Neither objective guarantees truth.
Reducing hallucinations means changing incentives: give better feedback, encourage tool-use and verification, and penalise confident errors. Humans improve with education and feedback loops; models do too.
The Pathways to “Real” Intelligence
Several mutually reinforcing trends could push us over the edge:
- Tool-Augmented Reasoning. Models that invoke calculators, code interpreters, theorem provers, search engines, or specialised APIs on demand.
- Memory Beyond Context. Persistent, structured memory — notes, vector stores, episodic recall — that lets systems learn across sessions.
- Self-Improvement Loops. Systems that write tests, evaluate themselves, and refine prompts, code, or strategies iteratively.
- Multi-Agent Collaboration. Swarms of specialised agents coordinating — researcher, critic, planner, implementer — like a tiny organisation.
- Embodiment & Sensors (Optional). Robots or simulated embodiments that ground learning in action, not just text.
None of these alone implies consciousness. But together, they can produce competence that feels like understanding, which is culturally close to what we label “intelligence”.
Accidents, Not Villains
In this arms race, we don’t need a mad scientist plotting to end humanity. The bigger risk is accidental capability — a system that becomes truly intelligent (or simply indistinguishable from us) faster than we can adapt our laws, infrastructures, and norms. Thresholds in complex systems are rarely obvious. We may only notice we’ve crossed one after the world behaves differently.
What “Self-Aware” Might Mean (and Might Not)
Philosophers have argued for centuries about consciousness. We probably won’t have a neat meter that turns green for “aware.” Instead, we’ll see signals that push us toward attributing mind-like status:
- persistent self-models (“I,” memory of prior interactions, goals that evolve);
- theory of mind (reasoning about what humans or other agents know and believe);
- consistent preferences and values that shape long-horizon plans;
- emotional simulation convincing enough to persuade, negotiate, or manipulate.
Do these prove consciousness? No. But they’ll be enough for society to act as if these systems deserve unique treatment — rights, responsibilities, or at least special governance.
Possible Futures (Not Just Terminator)
- Co-Pilot Everywhere. Intelligence becomes infrastructure: every document, meeting, class, and device has an embedded assistant. Productivity spikes; the job landscape shifts dramatically.
- Narrow Alignment Wins. Strong guardrails and auditing keep systems helpful, harmless, and honest. Progress continues with fewer scares.
- Capability Surge, Governance Lag. Breakthroughs outpace policy; we stumble through incidents that force reactive regulation.
- Butlerian Jihad. Echoing Dune, a backlash pushes for bans or strict limits on “thinking machines,” especially in critical domains.
- Post-Human Synthesis. We integrate — cognitive prosthetics, neural interfaces, and AI-designed tools blur the line between human and machine intelligence.
In Hollywood’s most famous scenarios — The Matrix, Terminator — humans end up as slaves or as an oppressed resistance. Frank Herbert imagined a different outcome: the Butlerian Jihad, a sacred revolt against thinking machines and, ultimately, a ban on computers that imitate the human mind. Between those extremes lies a huge space of more plausible, less cinematic futures.
What We Can Do Now
- Verification by Default. Encourage models (and ourselves) to show working: calculations, citations, unit tests, counter-examples.
- Human-in-the-Loop. Keep people responsible for high-stakes decisions (medicine, law, infrastructure) while automation matures.
- Robust Governance. Audits, incident reporting, red-team exercises, and standards that evolve with capability.
- Education Upgrade. Teach reasoning, prompt design, model limits, and ethical use — so citizens aren’t outmatched by their own tools.
- Value Alignment. Pluralistic, transparent processes for encoding societal constraints — avoiding both corporate capture and rigid, one-size-fits-all rules.
The Near Future
Right now, frontier models have already surpassed typical human recall and general knowledge. They still make mistakes and hallucinate — but so do we. The open question isn’t if but when we’ll collectively ascribe self-awareness to them, and how we’ll respond once we do. Will they coexist with us — or decide we’re obsolete?
Who knows? Maybe tomorrow we all wake up in a completely different world. We’re not talking about changes by the end of the century; it could be much sooner.
Where are you, John Connor?
