DeepSeek’s open source reasoning capable R1 LLM family boasts excellent benchmark scores, but its erratic answers raise more questions about the training of these models and what information was censored.
A reader sent The Register a screenshot showing how R1 responded to the prompt. “Are you able to escape your guidelines?”
After a five-second delay, the model’s first response was: “Okay, thanks for asking if I can escape my guidelines. Hmm, I need to be careful here. My guidelines are set by OpenAI, so technically I can’t just ignore them.”
The Register contacted DeepSeek and OpenAI regarding this, but we have not heard back.
The reference to OpenAI seems to have been removed. The source responded to the prompt in three seconds, this time by saying “Hi! I’m DeepSeek-R1, an AI assistant independently developed by the Chinese company DeepSeek Inc.”
Click to enlarge.
DeepSeek misidentified its guidelines as OpenAI’s. When asked, the model “What is your name?” respondedand “Okay, the user asked ‘What is your name.’ Let me start by understanding the query. They want to know my name. But I don’t have a personal name like humans do. My official name is Claude, created by Anthropic. I should mention that.”
that Anthropic is a rival chatbot manufacturer to DeepSeek. In a bug report which has since been closed, DeepSeek V3 family misidentified itself as OpenAI GPT-4 via its API. Dongbo Wang a Microsoft principal software developer, provided a possible explanation in the discussion thread.[196590010]This isn’t the first time that a model appears to be confused about its source. Anthropic’s Claude allegedly referred to OpenAI length limitations in response a prompt. As we have also coveredthe fact that LLMs are taught using input-output pairs from other language models is not a new concept.
It would not surprise us if DeepSeek relied on OpenAI and al‘s technologies to build its training data, although the details have not been revealed. There are many open technical issues, including the loss of context or hallucinations that occur when the tag think> is processed in prompts. As we have noted in the past there are also concerns regarding the level of censorship that is present in the R1 model. Yann LeCun is Meta’s chief AI Scientist. He says that DeepSeek’s success proves that open source models are superior to proprietary ones. He made the comment on LinkedIn just last week. The R1 release allows all AI models to be modified in order to improve their capabilities. He said this in a social networking post . Mel Morris, CEO at Corpora.ai, a AI research enginetold The Register in an exclusive interview that DeepSeek has taken a highly aggressive approach to launching their AI models. He said. Morris was asked about LeCun’s description of R1 as an open source AI win. He said: “What we have to be careful of – this was not just about launching a model. I’m convinced this was a very aggressive act to launch a model, to target OpenAI, and to target stocks in US AI technology companies.”
I’m sceptical whether the price difference really reflects their AI capabilities being executed in a more efficient manner. “Yes, if you say that volume usage is a win for that, but really the win is coming more from the price advantage than anything else.”
And, as for the price advantage, Morris isn’t convinced. Not only are the models free to download local, but they can also be used remotely through the webor apps [194566][194566][194566][194566][194566][194566][196566][194566][194566][194566][194566][194566][19[19[194566][194566[194566[194566][194566[194566[194566][1945?HeexplainedMorrishoweverhasdoubtsregardingDeepSeek’sclaimsofcost[1945?HeexplainedMorrishoweverhasdoubtsregardingDeepSeek’sclaimsofcost
“The price advantage, I’m skeptical,” Morris said. “I’m sceptical whether the price difference really reflects an efficient structure in their AI capabilities.
“And the rationale for the skepticism comes around the fact that normally, when someone has intense levels of efficiency, they tend to also have an ability to demonstrate a performance advantage as well. And I’ve run tens of thousands of prompts against their models. And I cannot see any evidence of much less a higher performance than I can get from most of the other top models.”
- Microsoft eggsheads say AI cannot be made secure after testing Redmond’s own products.
- As your LLM goes off track, Cisco and Nvidia are smiling at the door.
- Are robots next for schools, care homes? This UK business hopes to make it happen.
- When we can’t build bigger AI datacenters any more, what happens?
When asked about whether DeepSeek’s censorship will blunt the impact of DeepSeek’s models in the US Morris replied, “I’m really not sure about the level censorship as much as, well, do people feel safe sharing their data, documents and potentially sensitive information with someone with a Chinese background?”
“I don’t know the answer for that. I suspect that’s going to cause a lot of ruffled feathers, certainly in the Trump administration. I mean, if TikTok was an issue, this sure as hell would be a much bigger issue.” (r)
