Comparing your method to trillion-param LLMs is an unfair baseline
mariotelfig · x · 2026-09-11
Mario Teguh Sihalath argues that comparing against strong LLM baselines today is like comparing to humans 5-10 years ago — unfair, since a top LLM has a trillion parameters, costs millions, and is general-purpose. Often you seek a clever, efficient solution to a specific problem, even if it underperforms an LLM.
More from Research
- Redwood proposal: companies should disclose no-CoT reasoning to preserve AI monitorability — RyanGreenblatt · 2026-09-11
- Learn folding physics on simple shapes, transfer zero-shot to brain development with GNNs — bravo_abad · 2026-09-11
- KL-Shampoo beats Shampoo and SOAP without Adam, new paper reframes optimizers via KL divergence — _arohan_ · 2026-09-11
- Dankrad: Proving P=NP Would Likely Have Zero Implications for Cryptography — soumitrashukla9 · 2026-09-11
- NVIDIA open-sources BioNeMo Inference Runtime to speed up protein structure prediction — AllThingsApx · 2026-09-11
- If P=NP, proving it would be exactly what AI is good at: unleash agents to find polynomial solutions — soumitrashukla9 · 2026-09-11