Strengths of ChatGPT (me!): I’m great at holding conversations, answering a broad range of questions, and being adaptive across various topics. I also have memory, so I can remember details across interactions, which can make conversations more personalized over time. I try to keep things friendly and natural, and I aim to be helpful in whatever context you need.
DeepSeek: Without knowing the specifics, it might have a more focused or specialized function, perhaps around certain topics or tasks. Some AI tools can excel in specific domains (like search or knowledge discovery), so if it’s designed for that, it might be better suited to specific needs.
想下载，败也：
什么原因？暂时关闭了?还是它本身并不强大？
loser出现了：
大家看看，DeepSeek 是：
1 牛B
还是
2 不牛B
？
ChatGPT说：
Strengths of ChatGPT (me!): I’m great at holding conversations, answering a broad range of questions, and being adaptive across various topics. I also have memory, so I can remember details across interactions, which can make conversations more personalized over time. I try to keep things friendly and natural, and I aim to be helpful in whatever context you need.
DeepSeek: Without knowing the specifics, it might have a more focused or specialized function, perhaps around certain topics or tasks. Some AI tools can excel in specific domains (like search or knowledge discovery), so if it’s designed for that, it might be better suited to specific needs.
shakuras2000 发表评论于 2025-01-27 07:14:291. deepseek确实很强
2. deepseek大概率用了gpt的数据做了distill，所以脱离了更好的模型，deepseek可能做不到这么好
3. 550万是训练成本，实际成本可能高几倍，但是还是很低。
4. 有人说其实deepseek用了10000张A卡，只是因为众说周知的原因不能说，不过没证据我暂时当成谣言。
5. 这家公司是做量化的，就算deepseek不赚钱，发布配合沽空美股也能赚翻，汗
DeepSeek把RL（unsupervised）引入LLM训练的后期阶段，取代监督微调（SFT），这是由监督学习，转向非监督学习的重要算法改进，业内称为飞轮，依靠飞轮自身转动，改进大语言模型的效率，就像AlphGo Zero那样。这套方法明显可以被OpenAI，Google，XAI等美国AI大玩家利用，改进自己现有的模型训练。简单来说，Deep Seek用1%算力，搞成了接近100%的performance。大玩家借鉴RL这个飞轮，施加100%的算力，能达到10000%的performance吗？能达成1000%也很好了啊，甚至200%也行啊。将来施加1000%的算力，就会达成2000%的效果了。感觉距离AGI越来越近了。
好像DeepSeek证明了“中国人”，从而证明了“自己”多么聪敏似的。DeepSeek完成了重要的算法改进，借鉴了LLM训练的注意力机制，AlphaGo Zero的非监督强化学习，以及模型蒸馏和浓缩技巧，这是算法上的重要改进，能把整个大模型训练提高一个台阶。但是十分明显，就像Deep Seek借鉴别人成果一样，别人也可以借鉴这一成果，加速自己模型的训练，在算法差不多的情况下，数据（数量和质量）和算力依然决定模型的性能。
光这就得十亿美元了，，，翼龙还附和说“肯定的！”
Nvidia 该被查水表了。。。 老黄上周一居然不出席川总的就职典礼，更有甚者，这个老皮夹克还跑中国去了~~~
deepseek降低了AI游戏的准入成本，并不意味着英伟达要少出芯片。而是更多小资本玩家得以涌入市场。芯片的总需求量未必会减。
真正受到冲击的应该那几个拼命砸钱用算力垄断AI赛道的公司。具体而言，微软，谷歌，Open AI商用部门，Meta，Tesla。