3 ms·
It's not so surprising to me. It's like how Markov chains get better at passing for human the more N-grams they memorize. larger models will continue getting ma
by sterlind 2y ago
It's not so surprising to me. It's like how Markov chains get better at passing for human the more N-grams they memorize. larger models will continue getting marginally better at predicting the distribution (human language.) but that doesn't translate into improved intelligence.
- rfoo 2y agoThe point is, it isn't marginally better. I agree the setup is not a demonstration of intelligence, but the difference is pretty significant. Not to mention that on conventional benchmarks Llama 405B is usually worse than GPT-4o.