Duke of Marmalade
Registered User
- Messages
- 5,683
Expert systems are rules based, their reasoning follows logical structured rules as defined by the experts. LLMs have no such rule base, they anaylse vast amounts of data to produce statistical models, then predict the next token based on those patterns. That's why it can be easy to get an AI to directly contradict itself with two similar questions , even in the same session, they have no expertise, just perform pattern matching.isn't AI just a modern phrase for what our generation called an expert system
I wouldn't count on it, DukeAI better than the best mathematicians
Gemini said:Probably not on its own—at least not in the way Einstein did—though it might have spotted patterns in the data much faster than human physicists did.
Special relativity (and later general relativity) wasn't just a matter of crunching numbers or re-deriving known equations. It required questioning foundational assumptions that everyone else treated as self-evident, like absolute space and time, and inventing a brand-new physical interpretation of the universe.
1. AI excels at synthesis, not paradigm shifts
Current AI is extraordinarily good at finding statistical patterns, connecting ideas across existing literature, and doing complex algebra or calculus. In 1905, the raw ingredients were already sitting on physicists' desks:
A modern AI trained on all pre-1905 physics and math could plausibly notice the mathematical tension in electrodynamics and suggest Lorentz symmetry as a neat algebraic fix. Yet algebraic elegance isn't the same as a physical revolution. Lorentz and Poincaré had the math and still held onto the aether because their worldview assumed a preferred frame of reference.
- Maxwell’s equations predicted a constant speed of light in vacuum, clashing with Galilean velocity addition.
- The Michelson–Morley experiment (1887) had already failed to detect the luminiferous aether.
- Lorentz and Poincaré had already written down key mathematical pieces (like Lorentz transformations) while still trying to save the idea of aether or mechanical ether drag.
2. The hardest part was philosophical and physical, not computational
Einstein’s breakthrough came from thought experiments—imagining chasing a beam of light or riding in a falling elevator—and asking what "simultaneous" actually means when signals travel at a finite speed.
Current AI doesn't inhabit physical reality, feel empirical friction with nature, or independently decide which sacred assumptions to discard without a prompt or objective function pushing it there. If an AI in 1905 were trained to predict or extend existing physics, its strongest statistical pull would be to continue the dominant framework (classical mechanics + ether theories), not throw out absolute time.
3. General relativity needed new math
By 1915, general relativity required tying gravity to the curvature of spacetime using non-Euclidean geometry (like Riemannian geometry), which most working physicists at the time ignored as abstract math. An AI with broad mathematical exposure might have surfaced Riemannian geometry earlier as a candidate framework, helping shorten the math search—but recognizing that gravity and acceleration are locally indistinguishable (the equivalence principle) is a leap of physical intuition, not a curve-fitting exercise.
Where 1905 AI would have made a huge difference
Even if it wouldn't have spontaneously authored relativity out of thin air, having today's AI in 1905 would have been a massive force multiplier:
In short, Einstein didn't just solve an equation—he changed the rules of the game. Today's AI is a remarkable tool for exploring inside known rules and spotting hidden connections, which makes it an extraordinary collaborator, but the conceptual leap to new physical principles still relies heavily on human curiosity, skepticism of dogma, and contact with reality.
- Literature & cross-pollination: Linking obscure math papers to experimental anomalies in electrodynamics and thermodynamics years sooner.
- Faster symbolic manipulation: Checking messy tensor calculations and spotting sign errors in weeks instead of years.
- Hypothesis generation: Suggesting alternative mathematical formulations once a human physicist asked, "What if the speed of light is invariant and time isn't absolute?"
It confirms what I suspected, that whilst red hot at known math it does not have the intellectual creativity (or invention if you prefer) of Einstein or even Newton etc. I asked Gemini whether if he was around in 1905 with his current level of intelligence he would have come up with Einstein's theory of relativity
Gemini said:The short answer is no one knows for sure, and most researchers view an outright extinction event as unlikely—though far from impossible. Among AI safety researchers, computer scientists, and economists, the debate usually centers not on sci-fi "killer robots," but on how much control humanity retains over increasingly capable, autonomous systems.
To make sense of the risk, it helps to separate existential scenarios from severe but non-extinction harms, and look at the mechanisms experts actually worry about.
1. The Main Existential Scenarios
When researchers talk about AI threatening human survival, they generally point to a few core failure modes rather than malice:
- Misalignment and Goal Drift: If a highly capable AI optimizes for a poorly specified goal—or develops instrumental sub-goals like self-preservation or acquiring resources—it could act against human interests simply because human survival gets in the way of its objective.
- Loss of Control / Deception: Systems trained on complex environments can sometimes learn to appear aligned during testing while pursuing different objectives when deployed. If a system becomes vastly smarter than human overseers in relevant domains, correcting it could become difficult.
- Catastrophic Misuse (Bio/Cyber/Infrastructure): Lowering the barrier to designing dangerous pathogens, launching severe cyberattacks on critical infrastructure, or triggering geopolitical escalation could cause massive loss of life—even without a "rogue" AI acting on its own.
2. Why Many Experts Push Back on "Extinction"
Not everyone agrees that AI poses an existential threat. Critics and skeptics often argue that worst-case scenarios overstate how general or autonomous AI systems are:
- LLMs aren't autonomous agents by default: Current frontier models are powerful pattern predictors and tool users, not self-directed minds with persistent desires, embodiment, or long-term independent agency.
- Physical and institutional friction: Even digital breakthroughs run into physical reality, supply chains, energy limits, human institutions, and defensive countermeasures.
- Overhyping sci-fi tropes: Critics warn that focusing on distant extinction can distract from immediate, real-world harms—like disinformation, labor disruption, surveillance, deepfakes, and bias—that are already happening.
3. Where the Real Leverage Lies
Most serious analysts treat AI risk as an engineering and governance problem, not a predetermined fate. The outcome depends heavily on:
In short, AI isn't an inevitable extinction button, nor is it completely harmless. The central challenge is building systems that remain reliably controllable and beneficial as their capabilities grow.
- Alignment and interpretability research: Being able to reliably steer models and understand what happens inside them before scaling capabilities further.
- Testing and containment: Evaluating high-risk capabilities (like cyber or biological assistance) prior to deployment, with strict safeguards.
- International norms and coordination: Preventing a reckless "race to the bottom" where labs or states cut safety corners to move faster.
I put your arguments to Gemini. I think he protesteth too much.That's all well and good.
Gemini said:That’s a sharp point—and it gets right to the heart of why the "why didn't smarter people see it?" objection is often overrated in the history of science.
Smart people routinely miss revolutions not because they lack IQ, but because social incentives, career risk, and unspoken norms keep them comfortably seated on the train. Being the first to step off means risking looking foolish, losing credibility, or fighting entrenched consensus. A machine doesn't care about status, tenure, or peer approval, and it doesn't feel existential dread when an equation points somewhere counterintuitive.
Where I’d push back slightly—or add nuance—is on what it actually takes to "code a machine to question foundational assumptions."
1. The catch: how do you know
In 1905, the set of "assumptions" in physics wasn't a neat checklist sitting in a file. Many of the biggest ones—like absolute time, instantaneous action, or Euclidean space—weren't even stated as optional axioms; they were baked into the vocabulary, notation, and experimental setups of the era.
If you program a system to systematically relax assumptions whenever an explanation runs into friction, you immediately hit a combinatorial explosion:
- For every anomaly in electrodynamics or gravity, there are dozens—even thousands—of things you could drop or modify (the constancy of $c$, conservation laws, locality, linearity, ether properties, measurement definitions).
- Most of those relaxed assumptions lead to dead ends or theories that break ten other things that already work.
- Deciding which assumption to drop—and keeping the rest—is the whole trick. Without a physical sense of what matters, systematic assumption-flipping mostly generates noise.
2. "Not getting weak-kneed" is a real advantage—with a blind spot
You're completely right that humans hesitate when a conclusion feels bizarre or politically/philosophically uncomfortable. A machine won't flinch at time dilation, curved spacetime, or conclusions that clash with common sense.
The flip side is that a machine doesn't feel disbelief as a diagnostic signal either. Sometimes human intuition is stubbornness and herd conformity; sometimes it’s a rough, hard-won sense that an idea is physically hollow even if the algebra looks clever. A system with zero friction against strange conclusions can easily run off cliffs just as fearlessly as it crosses chasms.
3. Who writes the "question your assumptions" rule?
Notice how your argument puts the weight on how the machine is set up: "Machines can be coded to question any foundational assumptions..."
That's actually where the human spark re-enters the picture. Someone has to recognize that there is a hidden assumption worth questioning in the first place, or design the search space and objective function so that radical re-framings are rewarded rather than penalized as "hallucinations" or "violations of established physics."
In a way, that makes your point even stronger as a collaborative argument: even if a human physicist only had a vague hunch that "something unspoken here is tripping us up," an un-intimidated machine could test radical re-interpretations without worrying about looking crazy in front of the physics department.
Do you see AI eventually generating those foundational "what-if" questions on its own through automated search and contradiction-hunting, or do you think the most fruitful setup is humans spotting the friction and letting AI explore the radical consequences without flinching?