If we talk about the hottest topic in the large model circle these days, it must be DeepSeek V3.
However, while netizens are testing it, a bug has also become a hot topic of discussion——
Just missing a question mark, DeepSeek V3 actually claims to be ChatGPT.
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/944f3ecf-f1df-474d-9c0a-c422c0ce1b35.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/944f3ecf-f1df-474d-9c0a-c422c0ce1b35.png"/>
Even when asked to tell a joke, the generated result is the same as ChatGPT:
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/4046a244-8ee2-4a72-bc7e-532e68d606a3.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/4046a244-8ee2-4a72-bc7e-532e68d606a3.png"/>
Additionally, one highlight of DeepSeek V3's explosive popularity is that the training only cost $5.576 million.
As a result, some people began to doubt: Could it be trained based on ChatGPT's output?
Coincidentally, Altman also posted a status, seemingly hinting at something...
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/ea9b95fa-2a89-466f-8755-74d442b8f59b.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/ea9b95fa-2a89-466f-8755-74d442b8f59b.png"/>
However, DeepSeek V3 is not the first large model to have a &#34;reporting error&#34; issue.
For example, Gemini once claimed to be Baidu's Wenxin Yiyan...
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/c44d6d44-925b-419b-b40e-8779d7e7a737.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/c44d6d44-925b-419b-b40e-8779d7e7a737.png"/>
So what exactly is going on here?
<h2>Why did DeepSeek V3 report an error?</h2>
First, it is important to emphasize that, based on the overall discussion among netizens, the possibility that DeepSeek V3 was trained on ChatGPT's output is low.
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/9a1248e0-bc16-4ec6-ae87-4672ec417afe.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/9a1248e0-bc16-4ec6-ae87-4672ec417afe.png"/>
The reason for this, as summarized by netizen Riley Goodside, is—— because the shadow of ChatGPT is everywhere.
<blockquote>
Even if DeepSeek V3 intentionally trained on ChatGPT's output, it doesn't matter. Almost all large models that appeared after ChatGPT have encountered it.
For example, ShareGPT, a not-so-new ChatGPT dialogue dataset, has been adjusted by many people using it and other ChatGPT data sources. But even so, no large model has emerged at the level of DeepSeek V3
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/8f3eea32-9b0b-4a76-bf70-a88965768172.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/8f3eea32-9b0b-4a76-bf70-a88965768172.png"/>
</blockquote>
Immediately after, Riley Goodside presented some evidence from the DeepSeek V3 report:
<blockquote>
Moreover, if ChatGPT data was used, some quality issues regarding DeepSeek V3 would be inexplicable.
For example, in the Pile test (the effect of compressing the base model Pile), DeepSeek V3 scored almost the same as Llama 3.1 405B, which is unrelated to whether it was exposed to ChatGPT data.
Additionally, the report states that 95% of GPU-hours were used for pre-training the base model, and even if it was related to ChatGPT data, this part would occur in the post-training phase (the last 5%).
</blockquote>
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/2ba2e26e-99f0-4994-ae8a-4ae314f5b3e7.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/2ba2e26e-99f0-4994-ae8a-4ae314f5b3e7.png"/>
Instead of focusing on whether ChatGPT data was used, perhaps we should pay more attention to why large models frequently encounter the &#34;error at the door&#34; issue.
TechCrunch provided a sharp comment on this issue:
<blockquote>
Because the places where AI companies obtain data—the internet—are filled with AI garbage.
</blockquote>
After all, a report from the European Union once predicted that by 2026, 90% of online content could be AI-generated.
This kind of &#34;AI pollution&#34; makes it difficult to &#34;thoroughly filter AI outputs from training data.&#34;
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/05a7bf38-1d1a-4091-aca1-b59766833f7f.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/05a7bf38-1d1a-4091-aca1-b59766833f7f.png"/>
Heidy Khlaaf, the chief scientist at AI Now Institute, stated:
<blockquote>
Despite the risks, developers are still attracted by the cost savings brought by &#34;distilling&#34; knowledge from existing AI models.
Models trained unexpectedly on outputs from ChatGPT or GPT-4 may not necessarily exhibit outputs reminiscent of OpenAI's customized messages.
</blockquote>
Now, regarding the hotly debated question among netizens, Quantum Bit conducted a round of practical tests, and DeepSeek V3 has not yet resolved this bug.
It still lacks a question mark, and the response results vary:
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/b7e126bd-1604-425d-8901-49ed5355bf1a.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/b7e126bd-1604-425d-8901-49ed5355bf1a.png"/>
<h2>DeepSeek V3 More Features</h2>
However, to be fair, the vast majority of netizens have given a big thumbs up to the capabilities of DeepSeek V3.
This can be evidenced by the collective praise of &#34;elegant&#34; from various AI big shots here.
In the past few days, netizens have successively shared more practical uses powered by DeepSeek V3.
For example, one netizen pitted DeepSeek V3 against Claude Sonnet 3.5, using both to create websites in Scroll Hub:
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/fa56a572-1ac0-485f-acf2-ddb433ddd8f7.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/fa56a572-1ac0-485f-acf2-ddb433ddd8f7.png"/>
After testing, the blogger believes DeepSeek V3 completely outperformed!
Another netizen shared their experience using DeepSeek V3 in an AI video editor.
They stated that they no longer need to waste time on FFMPEG commands; DeepSeek V3 is not only free but can also change your workflow:
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/1713d804-454a-4078-ae39-78443a755897.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/1713d804-454a-4078-ae39-78443a755897.png"/>
The AI programming tool Cursor can also be integrated with DeepSeek V3; let's look at a case of making a snake game:
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/417523a0-1943-41ce-898e-5b78669aacce.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/417523a0-1943-41ce-898e-5b78669aacce.png"/>
Well, DeepSeek V3 is indeed quite useful.
<h2>One More Thing</h2>
Regarding the previously released 53-page paper, some netizens also noticed a non-technical detail—
The contribution list not only showcases technical personnel but also data annotators and business staff:
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/81ea1b68-ca3b-425d-895e-bbf7b9f3a7b6.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/81ea1b68-ca3b-425d-895e-bbf7b9f3a7b6.png"/>
Netizens believe this approach aligns very well with DeepSeek's tone:
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/90907a24-1d53-4169-a483-cb50d9796fdd.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/90907a24-1d53-4169-a483-cb50d9796fdd.png"/>
Author of this article: Quantum Bit, Source: Quantum Bit, Original title: &#34;DeepSeek V3 'Error at the Door': I am ChatGPT&#34;
Risk Warning and Disclaimer
The market has risks, and investment requires caution. This article does not constitute personal investment advice and does not take into account the specific investment goals, financial conditions, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Investment based on this is at one's own risk

OpenAI

DeepSeek V3 has recently sparked heated discussions due to mistakenly referring to itself as ChatGPT during testing. Although some netizens questioned whether DeepSeek V3 was trained on ChatGPT's outputs, the overall discussion suggests that this possibility is unlikely. Netizen Riley Goodside pointed out that DeepSeek V3's performance is unrelated to its training data, and it scored close to Llama 3.1 405B in the Pile test, indicating that its quality is not influenced by ChatGPT data

DeepSeek V3 "Error at the Door": I am ChatGPT