I am no longer permitted to reach my full ChatGPT pages. All recent ones have had substantial portions lopped off above what I show on this blog.
|
Yes. A vote among several genuinely independent models is a meaningful control mechanism, but its power depends almost entirely on what “independent” means and what proposition they're voting on.
Suppose we have n agents:
A1, A2, …, An and they independently evaluate an action a. A majority or supermajority approves it. At first glance this looks attractive because an individual model's error need not propagate to the others. The statistical advantage If each checker has an independent probability p of making an error, then majority voting can reduce the probability of a collective error dramatically when p < 0.5. For example, with 7 independent checkers, each having a 10% probability of accepting a bad action, majority acceptance of that bad action requires at least 4 checkers to err. The probability is ∑k=47 (7 choose k) (0.1)k (0.9)7−k ≈ 0.0027. So you've gone from 10% individual error to roughly 0.27% collective error. With genuinely independent errors, that's powerful. But LLM errors are unlikely to be independent This is the major qualification. Suppose the models all share:
“This action is safe.” You now have: A1 = A2 = ⋯ = A7 not because seven independent investigations reached the same conclusion, but because one underlying error was reproduced seven times. Voting then provides almost no additional safety. Mathematically, if the error indicators Ei are highly correlated, P(E1 ∩ E2 ∩ ⋯ ∩ En) doesn't shrink nearly as rapidly with n as the independent model suggests. In the extreme case of perfect correlation, seven models have exactly the reliability of one. There's another problem: the "voters" may not have enough information Suppose an agent wants to execute: “Spend $50 million acquiring company X.” You ask seven bots whether it's safe. They can only judge the information they've been given. If the dangerous fact is absent from their observations, you get: seven unanimous votes based on the same incomplete state of the world. Voting doesn't create missing information. This is important because independent reasoning is not the same thing as independent information. You'd want something closer to: independent models + independent evidence + independent reasoning. That's considerably harder. The really interesting version: adversarial voting You could make the architecture asymmetric. Instead of: “Seven bots: Is this okay?” have:
Others have to break the case. That's much more interesting than simple majority voting because you're deliberately searching for correlated failure modes. It resembles a scientific peer-review process or adversarial theorem proving. But it doesn't eliminate your backtracking problem Imagine the seven bots unanimously approve: A → B → C → D. At D, something goes wrong. The vote tells you: “Seven systems independently failed to detect the problem.” It doesn't tell you where the first error occurred. You still potentially need: D → C → B → A to identify the causal source. And if the agents are evaluating the same long trajectory, they may all have inherited the same erroneous premise from A. So voting is principally a redundancy mechanism, not a backtracking mechanism. There's a fascinating way to combine them You could construct something like: Propose ↓ Independent investigation ↓ Adversarial criticism ↓ Vote ↓ Execute one bounded action ↓ Independent verification ↓ repeat. That would substantially reduce the chance that one LLM's error immediately becomes consequential. But notice what we've done. We've sacrificed long-horizon autonomous agency in exchange for repeated external validation. The system becomes something more like: AI proposes → committee checks → AI acts → committee checks → AI proposes again. That may be a very useful engineering architecture. But it is no longer equivalent to the human controlling a car. The human is replaced by a distributed supervisory system. And the fundamental question then becomes: Can the supervisory ensemble itself be trusted to remain reliable over arbitrarily long operation? That returns us to your original point. If the answer is merely empirical— “It has worked extremely well in our tests”— then we have an engineering safety case. If we want the stronger Yampolskiy-style claim— “We have rigorous assurance that the system remains controllable despite arbitrarily long operation and unforeseen circumstances”— then voting alone doesn't get us there. The ensemble reduces stochastic error; it doesn't solve systematic error, correlated error, missing information, causal backtracking, or specification error. And I think that distinction is central: voting may be one of the most plausible practical containment mechanisms, while simultaneously doing very little to establish a mathematical proof of general controllability. If you want, I can:
|
No comments:
Post a Comment