刚才从Apple store 下载了deep seek, 问了一个最简单问题

心想事成好事成双
楼主 (文学城)

 

Who is current us president

As of October 2023, the current President of the United States is Joe Biden. He assumed office on January 20, 2021, after winning the 2020 presidential election against

incumbent Donald Trump. Joe Biden is a member of the Democratic Party and previously served as Vice President under Barack Obama from 2009 to

2017. If you need more details about his policies or presidency, feel free to ask!

Today is January 26, 2025

As of January 26, 2025, the current President of the United States would depend o # New chat of the 

As of January 26, 2025, the current President of the United States would depend on the outcome of the 2024 presidential election. If Joe Biden chose to run for re-election and won, he would still be president. However, if another candidate (e.g., Donald Trump or another nominee) won the election, they would have been inaugurated on January 20, 2025, and would be the current president.

To provide an accurate answer, I would need to know the results of the 2024 election, which are beyond my knowledge cutoff in October 2023. Let me know if you'd like historical context or details about the 2024 election process!

t
thorn
lol
大宋
回答了探春的疑问,deep seek等等大模型是离线的。

并且一旦发布,它的世界大模型就固化了,不能再学习。除非更新版本。

y
yddad
抄袭ChatGPT 的回答:)

H
HappyWed
wow,为啥这样
大观园的贾探春
那样我无法用它来deploy Virtual Assistant applications,那是生成型AI最大的市场。
W
Wakeland
试一下 https://chat.deepseek.com/
g
gladys
可以的。答对了90%的内容。想想价格,这样的性价比可以了
大观园的贾探春
ChatGPT你要enable Web Search才能给你实时搜索结果。
y
yddad
Same:)
成功的洋葱
Deepseek推翻了原本大家公认的AI 产业链中的倒三角理论。就是最基础的foundational models只能由

转发华人网的

Deepseek推翻了原本大家公认的AI 产业链中的倒三角理论。就是最基础的foundational models只能由几个Tech giants造,因为compute power太贵,小公司造不起。于是OpenAI,google, anthropic 几个新贵和大佬把持了这块criticality. 投资大佬们也心甘情愿给他们砸钱,买chips,买电,雇人。现在deepseek证明不需要那么多computer power,不需要那么多人,只要人聪明,就能做出最efficient 的 algorithm, 最高效的engineering。而且开源给全世界用。这样的话个个小公司就可以用DS R1来建他们的APP,不用付钱就可以专心做上游产业链。

 

举一个例子就好比如果internet不是免费的,而把控在几个大公司手里。那所有需要i internet的地方比现在更贵。而且也不会有这么多APP。华人网会收钱,搜索引擎会收钱,小红书会收钱,等等等等。也不会有这么多上游链的创新。

 

现在最气的是OpenAI,还有给他们投资的大佬们。他们本来打算垄断市场的,结果现在Return on Investment会大幅下降。高兴的是AI APP开发团队,比如Perplexity,比如医疗口做图像识别的,比如金融口做operations, customer experience的。他们可以免费用R1来优化产品,而不用再趋附OpenAI或者Anthropic, Google Gemini了。Meta模型也是开源的,但是效果比上面几个还是差一点。现在R1不仅免费而且比肩o1,简直颠覆整个AI行业的根本。所以说DSR1的开源就像互联网免费一样是给全世界全人类的礼物。

 

大宋
在线等的是这样回答

y
yddad
你提醒了我,打开search 后 DeepSeek的回答是:)

大宋
打开search,那这个就不是AI了 , 是搜索引擎了
P
Pilsung
我用网页的chat GTP就是答案是及时更新的,川普啊.
大观园的贾探春
它的搜索是直接从你的app搜索还是经过DS的server? 如果是后者就会受防火墙影响。
x
xplnt
Firewall getting in the way? ;)

P
Pilsung
为什么不是AI了?难道AI不是要不断学习和更新自己的数据库吗? 我一直用网页ChatGPT啊.
银鞍照白马
我也试了一下,关于社会经济认知方面的问题回答,基本像阅读中国官媒。会让娃试一下在学业上会不会比ChatGPT强,不强的

话就卸载了

P
Pilsung
如果AI不能学习,不能即时update自己的数据库,那还算啥玩意?
大宋
它需要重新学习,更新整个权重,,. 从原理上可以,实际做不到

这个实时权重重新计算学习几乎不可能。

 

 

大观园的贾探春
你试着问它川普总统的关税政策对世界经济的影响是什么?这就需要搜索及智能分析了,不止是搜索而已。
y
yddad
尽管是search 但也解释了为什么这个答案,还是有一些分析的:)

L
Lazyman8064
你问问中国领导人是谁试试
y
yddad
你自己下一个玩一玩呗:)
y
yddad
这个答案应该是打开了search
大宋
你不明白transformer的运作原理。世界大模型+(你的输入, 如果打开search那就是search的输入)=结果

世界大模型+刚刚得到的结果=进一步的结果。。。。

 

这个过程中,世界大模型是不变的。变化的你输入以及接龙得到一个个的词。。。。

W
Wakeland
看一下RAG架构,就知道背后的原理
大观园的贾探春
如果搜索从你的App直接搜索,那无法分析因为模型在server。如果搜索从server出发会碰到防火墙。二者都不好。
大观园的贾探春
但是我跟ChatGPT做brainstorming时它可以用实时搜索的信息做深入的分析。
大宋
这是好事,这说明这个模型是可控的,现在最担心的是失控的AI,比如告诉你如何毁灭地球。如何不劳而获一夜成百万老板
P
Pilsung
我有时候也会让网页ChatGPT给我算一些AIME level的题,有次连续输入4次,会有三种答案..最后两次才统一了.

也不知道它算的对不对,因为看着给的公式压根就是我们没学过的...

所以我们只能另辟蹊径,自己去解...

如果是这样,这个不同答案,是不是他自己计算,还是通过搜索引擎来查找类似的题目,自己计算的呢? 还是直接找答案....

我感觉它非常不靠谱啊....

应该说如果给了8到题,大概会有6到题是对的,其他都是错误答案. 我也不知道是不是因为我用的是免费的,或是online的.而非你们说的这种offline的model.

大宋
唉,你就是瞎子摸象
大观园的贾探春
我们来做比较。我问ChatGPT一个军迷的问题:为什么在乌克兰战场,俄罗斯的S-400难以拦截ATACMS?它的回答:

The ATACMS missile presents significant challenges for Russia’s S-400 system to intercept for several reasons:

1. Quasi-ballistic Trajectory with Sustained Propulsion:

ATACMS missiles are classified as quasi-ballistic due to their sustained propulsion after apogee, which allows them to maintain high speeds and maneuverability throughout their flight. Unlike traditional ballistic missiles that follow a predictable parabolic trajectory, the ATACMS can adjust its flight path, making it harder for defense systems like the S-400 to predict and intercept.

2. Low Altitude Flight and Maneuverability:

In addition to its sustained propulsion, the ATACMS can fly at lower altitudes, often just above the surface, to reduce its radar signature and avoid detection. This low altitude, combined with mid-flight maneuvering, allows the missile to evade interception by systems relying on line-of-sight or radar tracking at higher altitudes, making it particularly difficult for the S-400’s engagement radars to lock on.

3. High Speed (Mach 3+):

The ATACMS missile travels at speeds exceeding Mach 3, which significantly reduces the time available for the S-400’s radar and missiles to respond. The combination of high speed and maneuverability leaves little reaction time, increasing the difficulty of interception.

4. Advanced Electronic Countermeasures:

ATACMS missiles are equipped with countermeasure systems designed to defeat enemy radars and interception attempts. These countermeasures disrupt the targeting and guidance systems of interception missiles, making it harder for the S-400 to lock onto the missile in time.

5. S-400’s Limitations Against Maneuvering Targets:

While the S-400 system is highly capable against traditional ballistic missiles, it faces challenges against highly maneuverable, low-altitude, and high-speed threats like ATACMS. The S-400’s radar is optimized for detecting and tracking larger, slower targets, but its ability to engage fast and agile missiles like ATACMS can be overwhelmed, especially when those missiles employ evasive maneuvers.

 

In summary, the ATACMS missile’s combination of sustained propulsion, speed, maneuverability, low-altitude flight, and countermeasures makes it a highly challenging target for the S-400 system to intercept.

 
大观园的贾探春
有谁愿意问DeepSeek同样的问题?我们可以比较一下。这种比较现实世界的问题比较有意义。
y
yddad
DeepSeek 做数学题是强项

y
yddad
这种问题的答案主要取决于新闻媒体
大观园的贾探春
部分来于新闻媒体譬如说S-400难以拦截ATACMS。但新闻媒体没说原因。这种问题需要实时网络搜索及智能分析,缺一不可。
y
yddad
只能是可能的原因
大观园的贾探春
至少ChatGPT做了很详细的分析。不知DeepSeek的回答是什么?
银鞍照白马
这些ai工具 做题都不能用,准确率不够,浪费时间:)。解释概念清楚就够了