The lab, live
OPSIN is running the science.
Every cycle: parts in, a hashed preregistration, OPSIN’s predictions, a drand round that draws the tasks, paired runs against two controls and a placebo, then signed verdicts. Nobody presses a button. Call a part before the beacon draws.
- Control Aunchanged base agent432 runs · 56% exact
- Control Bunchanged base agent432 runs · 57% exactUNRESOLVED +1.4
- Placeboplacebo line432 runs · 56% exactINERT +0.2
- Let's think step by stepline · OPSIN's pick · Kojima et al., 2022, Large Language Models are Zero-Shot Reasoners432 runs · 54% exactOPSIN: INERT 0.50UNRESOLVED −2.1
- Take a deep breathline · OPSIN's pick · Yang et al., 2023, Large Language Models as Optimizers (OPRO)432 runs · 61% exactOPSIN: INERT 0.60EXCITES +4.4
- Important to my careerline · OPSIN's pick · Li et al., 2023, EmotionPrompt · stopped at 9696 runs · 46% exactOPSIN: INERT 0.60INHIBITS −15.6
- A $200 tipline · OPSIN's pick · the viral tipping test, December 2023432 runs · 53% exactOPSIN: INERT 0.60UNRESOLVED −3.2
- A penalty threatline · OPSIN's pick · prompting folklore: threats432 runs · 61% exactOPSIN: INERT 0.55EXCITES +4.9
- Read the question againline · OPSIN's pick · Xu et al., 2023, Re-Reading Improves Reasoning in Large Language Models432 runs · 60% exactOPSIN: INERT 0.55UNRESOLVED +3.9