TL;DR: āSelf-consistencyā, ask a model the same question several times at a non-zero temperature and take the answer they most agree on (a plurality vote), gives large accuracy gains over greedy decodā¦
LLMs
6 posts tagged LLMs Ā· all tags
2026
Self-Consistency Is Just Probability Amplification
Understanding Is the New Bottleneck ā and Your AI Agent Can Fix It
TL;DR: In AI coding the hard part is no longer writing the code, it is understanding the code the agent wrote well enough to keep steering. Geoffrey Littās framing is that the human job shifts from veā¦
Generate-and-Verify: What P vs NP Teaches Us About AI Reasoning
TL;DR: The defining feature of the complexity class NP is polynomial-time verifiability: a candidate answer is cheap to check. The asymmetry that matters shows up in its hardest members, the NP-compleā¦
The Quadratic Wall: The Algorithmic Race to Beat O(nĀ²) Attention
TL;DR: Standard self-attention costs on the order of nĀ² in the sequence length n, in both arithmetic and the attention-score memory it materialises. Almost every long-context architecture headline of ā¦
Why AI Tokens Are So Expensive ā and What Actually Makes Them Cheaper
TL;DR: Autoregressive LLM generation has two phases with different bottlenecks: prefill processes input positions in parallel and is usually compute-intensive, while decode produces tokens sequentiallā¦
Compression Is Intelligence
TL;DR: Information theory says prediction and compression are the same thing measured two ways. That gives the cleanest mental model I know for what a large language model does: it is a compressor of ā¦