Ignore all that "the future is art and socializing" post AI bs.
The future is MMORPG and it swallow humanity. The real world doesn't stand a chance, they will never make it as pretty.
AI will never substitute the 3rd deviation taste because it will never appear in the data and it's not hill-climbable... there's a discontinuity in space I swear.
inflation changes its meaning every what? 2 years?
it's obvious that no one cares about these numbers because they don't reflect the difficulty of buying a home and family related services, aka the thing that people mostly care about while moving on with their lives.
adjusted for fake inflation, aka adjusted for price of water and bread when people care about houses and family related services to move on with their life.
but yeah i understand collecting an X paycheck is great so this was a nice rage bait tweet.
I love zeroth order optimization strategies because they force you to understand information theory and geometry and from then you see it everywhere...
from evolution, to how you should work with teammates/how to deal with large groups, game theory, philosophy... everything that involves information and an environment (so... everything)
reproduction is all about synteny indeed but in the future there's only 2 options:
1 - it stays because life fitness function is multimodal to the point where sampling (1 parented offspring) being able to follow signal doesn't matter because there's multiple, not in the same basin.
2 - we all stop crossover and start sampling because we treat life as a quasi convex fitness landscape based on our beliefs so we can bet on the better gradient/signal strategy - evolutionary strategies > evolutionary algorithms in dense gradient environments and unimodal landscapes.
personally i see option 2 being obvious in the future, people might pitch in a couple genes from people who they have a relationship with but nevertheless they will choose from a sampling standpoint (with evaluation priors as we already have heuristics to predict health, intelligence etc) and not a parental crossover dominated standpoing.
this is reading like one of those memes where there's a party saying "on average/on a majority of times this happens" but then there's a girl saying "actually i know an example where it doesn't"
i'm all for fine-tuning, it's just that no one cares (large majority)
Novelty also exist through synthesis of past information interpolation, which is totally ok. But the momentum towards that useful interpolation is lacking in the distribution the AI models learned, meaning it still needs a lot of human push towards those “simple” important interpolations
novelty comes at the point of contact with new information and then using non-novelty to frame it. it's online.
we as users bring novelty to AI models a lot of times, thats why they will keep getting better, beyond the obvious offline datasets that have already been drained.
Hmm you don’t understand LLMs. They are information storages of the past and if your pattern is not there it won’t help you.