https://www.instagram.com/reel/DadlLKBPjbE/?igsh=MWd3OWpidGc5Z3l6MA==
https://www.instagram.com/reel/Damhg7bykut/?igsh=MWRuZXF1NTFrOTNldg==
https://www.instagram.com/reel/Dbn8Fsxhu-l/?igsh=MWt1d2c3MXlzbXJ0eQ==
https://www.threads.com/share/BAZeMVeOL2/
Threads
月底救星｜小資租屋觀察中 (@fennyhsu1936) on Threads
月底救星｜小資租屋觀察中 (@fennyhsu1936) on Threads
https://www.threads.com/share/BAYQeRAGoa/
Threads
李昀翰 (@leeunhn) on Threads
李昀翰 (@leeunhn) on Threads
https://www.instagram.com/reel/DblLRBISAUf/?igsh=c2pxNHg0eGxjcW9n
https://www.threads.com/share/BAcWz7KB3W/
Threads
大叔裝機安 (@install.safe) on Threads
大叔裝機安 (@install.safe) on Threads
https://www.threads.com/share/BAexq5h274/
Threads
Kok Sing (@koksinghewvvvv) on Threads
Kok Sing (@koksinghewvvvv) on Threads
https://www.instagram.com/p/Dbqe7fgEyut/?igsh=MWN2emZxMXgyNTNzdg==
https://www.instagram.com/reel/DbrReJBjkx_/?igsh=dDg0eTNhNTVvMWYy
https://www.threads.com/share/BAV3YvifkV/
Threads
AI Dev Diaries | 從文轉理外商打工人 (@sw_ai_life) on Threads
AI Dev Diaries | 從文轉理外商打工人 (@sw_ai_life) on Threads
https://www.instagram.com/reel/DbqPfwICQOn/?igsh=MWpnd3Vnbmp3dWUwOA==
https://www.instagram.com/reel/DadGOThPPFz/?igsh=NXhpOWp3YzBtYmFt
https://www.instagram.com/reel/Dbk0hBWgoIY/?igsh=MXkxdWNmdGhlbWNuMA==
https://www.instagram.com/reel/DboRHixDAXL/?igsh=MW5peGg3MDR2ZnZyMA==
https://www.instagram.com/reel/DZIv9wWTJt1/?igsh=MWgyMDg3djY0c2d1cw==
https://www.instagram.com/p/Dbp9VytG9-A/?igsh=YzZ6OHg1MXJlNzFm
https://www.facebook.com/share/1BUVuSSGgr/?mibextid=wwXIfr
https://www.threads.com/share/BB7HfBg3Cf/
Threads
dev (@easyanythinghk) on Threads
dev (@easyanythinghk) on Threads
https://lnkd.in/p/g__3AeE8
LinkedIn
一張 4090，就把 35B 模型推到 GPT-5.6 Sol 的水準？研究團隊發表 Frontis-MA1，嘗試遞迴自我改進，也就是讓 AI 去改進 AI。他們拿機器學習工程當測試場，因為每一步都能執行、能驗證，並且開源一整套 OpenMLE Stack。…
一張 4090，就把 35B 模型推到 GPT-5.6 Sol 的水準？

研究團隊發表 Frontis-MA1，嘗試遞迴自我改進，也就是讓 AI 去改進 AI。他們拿機器學習工程當測試場，因為每一步都能執行、能驗證，並且開源一整套 OpenMLE Stack。…
一張 4090，就把 35B 模型推到 GPT-5.6 Sol 的水準？

研究團隊發表 Frontis-MA1，嘗試遞迴自我改進，也就是讓 AI 去改進 AI。他們拿機器學習工程當測試場，因為每一步都能執行、能驗證，並且開源一整套 OpenMLE Stack。

底層 OpenMLE-Gym 蒐集 5,758 個可執行任務，每個都配好沙箱、隱藏評分程式與結構化回饋。中層 OpenMLE-ERL 把演化拆成四個原子動作，草稿、改良、除錯、交配，直接拿執行結果做 SFT 與強化學習。上層 OpenMLE-Evo…
Before
After
Back to Top