A handful of GPUs built a GPT-6.1-level reranker from scratch, fully open
lintool · x · 2026-10-20
Jimmy Lin's (lintool) team speedran Karpathy's nanochat: pretraining from scratch with ClimbMix data (shizhediao et al.), then SFT on relevance data from beirmug/crystinaz to produce a competitive pointwise reranker.
Results are about on par with SFT'ing Qwen or Gemma backbones, and just a bit below listwise GPT-6.1 Sol — all with only a handful of GPUs.
More from Research
- SKY-001 debuts as first 'Engineering World Model,' claiming 125x speedup over GPT-6 on CAD — vasuman · 2026-10-20
- Operating fully in packed space: 10k-window shuffle drops from 7.86s/8.7GB to 0.85s/0.7GB — jmschreiber91 · 2026-10-20
- Speeding up DeepLIFT/SHAP: packed-sequence dinucleotide shuffling avoids memory blowup — jmschreiber91 · 2026-10-20
- UCSB REAL AI lab hiring postdoc to build AI-driven autonomous labs — ninamiolane · 2026-10-20
- Packing four loci per uint8: a memory trick that slashes DNA sequence encoding by 4x — jmschreiber91 · 2026-10-20
- Anthropic Risk Report shows distillation degrades agentic coding: SWE-bench drops detailed — maksym_andr · 2026-10-20