Dev observes models barely improved at one-shotting in 6 months — tool use and thinking are the ladder
rickasaurus · x · 2026-10-12
Developer rickasaurus notes that models aren't much better at 'one shotting' without tool use than six months ago, and not much better without thinking than a year ago. His takeaway: capabilities like tool use and reasoning are the actual ladder of progress, not raw unassisted model performance.
More from Models
- vllm-sr open-sources d3 reasoning model with xhigh/medium/low thinking levels on HF — victormustar · 2026-10-12
- JEV-35B decision model with calibrated probabilities trends on Hugging Face — autotrust · 2026-10-12
- Google's EmbeddingGemma 2 unifies text, image, video embeddings; 1-bit quantization keeps 95% of search quality — solyarisoftware · 2026-10-12
- Unannounced Claude model shows up in the wild: 2-hour one-shot Blender build via MCP — testingcatalog · 2026-10-12
- Unlimited OCR: 3B open-source model reads 100-page PDFs in one pass, but scores just 46.2% on ParseBench — solyarisoftware · 2026-10-12
- Reddit math: a fully used $200 Claude plan equals $2.5k–$12.5k in API value — Atlan_ · 2026-10-12