Brooo 🤯 I was doomscrolling on github and found this.
It really makes my brain stop for a moment. DeepSeek recently dropped their harness agent, DSH. And he said that on the first day it immediately reached 78,000 stars. 😳 But that's not the craziest part. 53 MINUTES. Only 53 minutes
Been seeing a lot of people on my timeline say that DeepSeek V4 is heavily overfit to the minimal mode…
Been seeing a lot of people on my timeline say that DeepSeek V4 is heavily overfit to the minimal mode of its official Harness, so I did a tiny test myself. The worst one here is DeepSeek V4 + my own agent, Bubble Code.The middle one is DeepSeek V4 + its official Harness in…
What a crazy week in AI!
🚀 DeepSeek V4 0813 GLM 5.3 Grok 4.6 Qwen 3.8 Max Qwen 3.8 27B Index TTS 2.5 MiDasheng LTX 2.5 GPT Ultrafast Gemini 3.7 Flash MAGI-2 WorldClaw Nemotron Lightning & more! Watch the full recap:
The expiration date of DeepSeek V4-Pro, a cost-effectiveness benchmark, was 3 days.
DeepSeek raised V4-Pro-0813 to GA on August 12th. There were no blogs or posts. The API document footnote was the only notice. Performance jumps are abnormal. DeepSWE 12.8 → 62.7, CyberGym 52.7 → 83.3, Terminal Bench 2.1 72.1 → 87.9. two
You have reached the end of the archive
All of deepseek