3 ms·
Nobody has ever made it clear that there is even a slight consideration for safety given that everybody at this point knows we're 1-2 generations away from self
by boringuser2 4y ago
Nobody has ever made it clear that there is even a slight consideration for safety given that everybody at this point knows we're 1-2 generations away from self-replicating models.
What do we do then?
Literal crickets.
- have_faith 4y agoIs there anywhere I can bet against this?
- goatlover 4y agoYou want to bet that three years from now, GPT-6 can't replicate itself? I feel burned by "betting" against surprising AI advances the past decade. That being said, I don't know whether an LLM can be trained on weights and what it would mean to have its own weights as part of the training data.
- boringuser2 4y agoGPT-4 can improve itself with reflection. It's already real. The only missing piece is vast improvement and autonomy.
- boringuser2 4y agoSo, I thought about this question briefly and I believe this is your answer: Every non-trivial passage through the layers of the neuronal net produces a never-before-seen amalgamation of weights. The AI, with this novel information -- data filtered by its neural net and infused with some randomness -- can use the novel data to further process it for correctness. In our case, the randomness is likely caused by genetics and environmental factors. For the AI, it's caused by explicit programming. It's a form of evolution -- things can evolve into novel things by using non-novel substrate thanks to the refinement of selective iteration. It feels like this is similar to how humans consume information using reflection.
- boringuser2 4y agoWhy would you bet against it? GPT-4 is shockingly good at programming logic. I'm writing 95% less code and mostly guiding and debugging GPT-4 now and it's insane.