5 Papers That Show Where AI Research Is Heading Right Now

Human data constrains discoverable solution spaces

Training on human-generated solutions limits models to subspace H, making it improbable to discover the full solution space F minus H regardless of test-time compute or recursive self-improvement efforts.

AlphaZero self-play enables unbiased AGI development

Unbiased self-play without human data represents a more viable path to advanced intelligence than AlphaGo-style training, avoiding the limitations of human "meandering" exploration patterns.

Intelligence per sample optimization remains critical challenge

Current methods like in-context learning fail to improve monotonically with more samples and hit context-length cliffs, unlike human learning which consistently improves with experience using the same algorithm.

Biologically inspired alternatives to backpropagation urgently needed

The brain likely does not use backpropagation, suggesting undiscovered learning procedures like SPSA could dramatically improve intelligence per watt and enable true continuous learning.

Protein models confirm scaling laws transfer to biology

The ESMC protein language model demonstrates clean log-linear scaling laws identical to LLMs, where contact prediction performance improves predictably with increasing compute and parameters from 300M to 6B.

Billion-scale metagenomic data eliminates performance plateaus

Unlike the previous ESM2 generation which plateaued at 50 million sequences, ESMC trained on 2.8 billion metagenomic sequences from uncultured organisms shows no diminishing returns, confirming data scaling drives biological AI capabilities.