I have placed a bet with a friend at 1:100 odds (in my favor) that at least one running instance of DeepSeek V3 Base before October 2025 instantiated whatever we will understand "consciousness" to be by 2050. This market publicly operationalizes that bet.
That is to say, not only does there need to be a way in principle to use the weights of V3 to produce what might be called a conscious experience; we must have evidence that some way in which it was actually run before October 2025 (for instance, during standard forward passes) actually satisfied whatever condition we will later believe comprises consciousness.
If by 2050, the notion of "consciousness" is widely understood to be illusory or otherwise invalid, such that humans also should not be considered conscious, or if there is still substantial disagreement about the nature of consciousness, resolves N/A. If consciousness is instead understood as a collection of more fundamental properties, resolves YES only if a run instance of V3 possessed almost all of the composite properties, leaning towards NO in any muddier circumstances. In other words, if "degree of consciousness" is quantifiable, a YES resolution requires that an actually-run instance of V3 reaches roughly 90% of the way there, whatever that might be widely agreed to mean.
Update 2025-11-02 (PST) (AI summary of creator comment): Clarification on partial consciousness/degree of consciousness:
If V3 is found to be "about as conscious as an insect" or similar, this counts as YES only if we come to believe that insects/smaller minds have "most of what is important" (~90%) about being conscious
If consciousness has both quickly-satisfied components and components related to scalable sophistication of minds, the quickly-satisfied properties must comprise ~90% of what matters about consciousness for a YES resolution
The creator will resolve conservatively, preferring NO if we aren't getting to ~90% of the important properties of consciousness
1,000@Bayesian So, suppose we discovered that instances of V3 can be "about as conscious as an insect" or something, with the notion of consciousness having some components that are quickly satisfied and others that are closely related to the scalable sophistication of a mind. I'll say that the former standard should count if and only if we come to believe that insects / smaller minds have "most of what is important" (~90%) about being conscious.
So we get a YES resolution if almost all of the quickly-satisfied properties are present (if that is meaningful to say) and this is understood to comprise most of what we decide matters about consciousness. But thinking that the quick-to-satisfy components of consciousness comprise most of the concept is a relatively demanding standard: I'll resolve conservatively, preferring NO if we aren't getting to ~90% of the important properties, however that comes to be understood.
yeah in my mind it's pretty likely to be entirely about the scalable sophistication of a mind. iiuc if it does end up being about that, then the market would resolve based on whether that scaling is greater for v3 than the insect?
"most of what is important" (~90%) about being conscious.
if almost all of the quickly-satisfied properties are present
comprise most of what we decide matters
I am skeptical of this entire line of thinking, but like, suppose a universal function approximator when faced with linearly increasing data and size had a log increase in some conceptualization of consciousness, then would the idea be that any data+size has "90% of the properties" bc the only property is function approximation, or would the idea be that any data+size that is at least 90% the consciousness of an insect would count (which would then just end up being a relationship of that insect's size and data vs the model's size and data?)