🧠 DeepSeek-V3.1-Terminus was dropped yesterday: What's really changed?
Zhihu mind explorer 段小草 offered a clear takeaway: Stop treating LLMs as knowledge vaults. Start using them for what they do best: logic, reasoning and task-solving.
🚫 Hallucination is unavoidable, even at 1T parameters — because LLMs are fundamentally lossy compressors of knowledge.
✅ Better to use LLMs for extraction, classification, summarization & evaluation.
➡️ Tasks anchored to context, not vague prompts.
So based on thinking above, what's new in Terminus?
• Optimizes Agent capabilities
• Fixes the infamous “极” character bug
• Confirms DeepSeek is pivoting toward Agentic AI, not just bigger models
📌 Zhihu contributor MathLover tested it:
1. “极” no longer appears
2. Code-mixed CoT is better (but not gone)
3. Minor capability improvements; HLE seems up
And mind explorer 桔了个仔 noticed a situation:
⚠️ self_attn.o_prob param not yet compatible with UE8M0 FP8, aimed at next-gen domestic chips — hardware adaptation still a way to go.
🔗 Hot debate on Zhihu:
zhihu.com/question/195354501…
#DeepSeek #LLM #Agent #RAG #AI #AGI