top of page
An Anthropic researcher just gave us a peek at self-improving AI
TechCrunch AI
Given 10 benchmarks for specific misaligned behaviors, the automated systems were able to improve performance on every single one without degrading overall performance.
讞讜拽专讬诐 诪讞讘专转 Anthropic 讛爪讬讙讜 讛转拽讚诪讜转 诪专转拽转 讘转讞讜诐 讛讘讬谞讛 讛诪诇讗讻讜转讬转: 诪注专讻讜转 讗讜讟讜诪讟讬讜转 讛爪诇讬讞讜 诇砖驻专 讗转 讘讬爪讜注讬讛谉 讘注砖专讛 诪讚讚讬诐 住驻爪讬驻讬讬诐 砖诇 '讛转谞讛讙讜讬讜转 诇讗 诪讬讜砖专讜转' (misaligned behaviors). 驻专讬爪转 讚专讱 讝讜, 讘讛 讛讜砖讙 砖讬驻讜专 诪讛讜转讬 诇诇讗 驻讙讬注讛 讘讘讬爪讜注讬诐 讛讻讜诇诇讬诐, 诪拽专讘转 讗讜转谞讜 诪砖诪注讜转讬转 诇驻讬转讜讞 AI 讛诪住讜讙诇 诇转拽谉 讜诇砖驻专 讗转 注爪诪讜. 诪讚讜讘专 讘爪注讚 讞砖讜讘 讘讬讜转专 诇拽专讗转 讘谞讬讬转 诪注专讻讜转 AI 讘讟讜讞讜转 讜讗诪讬谞讜转 讬讜转专.
bottom of page