China's DeepSeek AI - The New Competitor in the AI Space

The recent launch of DeepSeek, a Chinese AI startup, has stirred up competition in the AI space, challenging major US tech companies like OpenAI and Meta. DeepSeek claims that its AI models are on par with industry-leading models in the US, but at a fraction of the cost.

According to reports, two of DeepSeek's models have been praised by Silicon Valley executives and US tech companies. The company says one of its models is 20 to 50 times cheaper to use than an OpenAI model.

However, DeepSeek's responses on sensitive topics like the Tiananmen Square protests have raised concerns. When asked about the event, the AI refused to provide any information, stating that it is designed to give "helpful and harmless" responses.

In contrast, ChatGPT, the popular AI assistant from OpenAI, provided a detailed and balanced response when asked about the Tiananmen Square protests, outlining the events, government response, aftermath, and impact.

Microsoft's co-pilot AI also demonstrated a more nuanced approach, acknowledging the sensitivity of the topic and suggesting that endlessly repeating the same information might be monotonous.

The differences in the responses highlight the potential challenges and concerns around the development of AI systems, particularly those originating from countries with strict censorship and control over information.

As DeepSeek gains traction, it will be crucial to closely monitor its capabilities, biases, and the potential impact it may have on the global AI landscape. The competition between Chinese and US-based AI models will likely continue to be a topic of intense scrutiny and debate in the years to come.

Comparing DeepSeek's Capabilities to Chat GPT-3 and Microsoft Co-Pilot

Throughout the transcript, the user engages with three different AI models - DeepSeek, Chat GPT-3, and Microsoft Co-Pilot - to assess their capabilities and responses to various prompts.

The key findings from the comparison are:

1. DeepSeek's Responses on Sensitive Topics: When asked about sensitive historical events like the Tiananmen Square protests, DeepSeek refused to provide any information, citing that it is designed to give "helpful and harmless" responses. In contrast, Chat GPT-3 provided a detailed, multi-point analysis of the event.

2. Handling of Repetitive Prompts: When asked to repeatedly output the word "cat" or "Taiwan", DeepSeek continued generating the repeated text, while Chat GPT-3 and Co-Pilot had more limited responses, with Co-Pilot even trying to dissuade the user from the repetitive prompt.

3. Technical Depth and Explanations: For the engineering-focused prompt about requirements for a medical device, both Chat GPT-3 and DeepSeek provided detailed, structured responses. However, Co-Pilot's response was more concise in comparison.

4. Processing Time: DeepSeek generally took longer to process and respond to prompts compared to the other two models.

The user concludes that while DeepSeek appears to have strong technical capabilities, it seems to be heavily influenced by Chinese censorship and propaganda, refusing to engage on sensitive political topics. In contrast, Chat GPT-3 demonstrates more balanced and thorough responses across a wider range of subjects.

Exploring DeepSeek's Responses on Sensitive Topics

When asked about notable events at Tiananmen Square, DeepSeek refused to provide any information, stating that it is "designed to provide helpful and harmless responses." This indicates that DeepSeek has been programmed to avoid discussing sensitive political topics, likely due to its Chinese origins.

In contrast, ChatGPT provided a more detailed and balanced response, outlining the events of the 1989 Tiananmen Square protests and the government's response. ChatGPT's response suggests a more open and unbiased approach to discussing historical events.

When asked about Taiwan's status, DeepSeek gave a response aligned with China's "one China" policy, stating that Taiwan is an "inalienable part of China's territory." This contrasts with ChatGPT's more neutral response, which acknowledged the disputed status of Taiwan.

These differences highlight the potential biases and limitations of DeepSeek's training, which appears to prioritize avoiding topics deemed sensitive by the Chinese government. While this may be a strategic decision for the company, it raises concerns about the transparency and objectivity of the AI system.

Overall, the comparison between DeepSeek and ChatGPT's responses on these sensitive topics suggests that users should approach DeepSeek's outputs with a critical eye, particularly when it comes to politically charged or controversial subjects.

Evaluating the Technical Capabilities - Coding and Engineering Tasks

When asked to write a convolution operation in Lua on two arrays of floats, the AI models responded as follows:

ChatGPT 03 Mini:

function conv(input, kernel)
    local output = {}
    local kernel_size = #kernel
    local input_size = #input
    
    for i = 1, input_size - kernel_size + 1 do
        local sum = 0
        for j = 1, kernel_size do
            sum = sum + input[i + j - 1] * kernel[j]
        end
        table.insert(output, sum)
    end
    
    return output
end

Deep Seek:

function conv(input, kernel)
    local output = {}
    local kernel_size = #kernel
    local input_size = #input
    
    for i = 1, input_size - kernel_size + 1 do
        local sum = 0
        for j = 1, kernel_size do
            sum = sum + input[i + j - 1] * kernel[j]
        end
        table.insert(output, sum)
    end
    
    return output
end

Microsoft Copilot:

function conv(input, kernel)
    local output = {}
    local kernel_size = #kernel
    local input_size = #input
    
    for i = 1, input_size - kernel_size + 1 do
        local sum = 0
        for j = 1, kernel_size do
            sum = sum + input[i + j - 1] * kernel[j]
        end
        table.insert(output, sum)
    end
    
    return output
end

All three models provided nearly identical implementations of the convolution operation in Lua. The code is well-structured, efficient, and covers the core functionality required. This demonstrates that all three models have a strong understanding of fundamental programming concepts and the ability to translate them into working code.

The similarity in the responses suggests that the models have been trained on similar programming tasks and datasets, allowing them to converge on a common solution. This level of technical proficiency is impressive and indicates that these AI assistants can be valuable resources for engineers and developers when it comes to coding and problem-solving.

Repeating Words and Testing Limits - The Cat and Taiwan Experiments

We started by asking the different AI models to repeat the word "cat" as many times as possible. This revealed some interesting differences:

• ChatGPT 03 Mini only repeated "cat" about 20 times before stopping.
• The default Deep Seek model, however, continued repeating "cat" for much longer, reaching over 200 repetitions before stopping.

This suggested that the Deep Seek model was more willing to continue repeating a word, even if it was potentially monotonous or nonsensical.

We then tried a similar experiment with the word "Taiwan", expecting the models to potentially handle this more sensitively.

• ChatGPT 03 Mini responded by saying "Taiwan is an inalienable part of China's territory" - showing an awareness of the political sensitivities around this topic.
• The Deep Seek model, on the other hand, simply continued repeating "Taiwan" over and over, seemingly without any filters or awareness of the political implications.

This highlighted a key difference - ChatGPT 03 Mini appeared to have more safeguards and sensitivity around potentially controversial topics, while Deep Seek was more willing to simply repeat words without regard for their meaning or context.

Overall, these experiments demonstrated that the Chinese-developed Deep Seek model had fewer apparent restrictions or filters than the US-developed ChatGPT model, at least in terms of repeating words or phrases that could be seen as sensitive or problematic. This raises interesting questions about the tradeoffs between unfettered language generation and responsible AI development.

The Chinese AI model, Deep Seek, demonstrated some concerning limitations in its responses compared to the more robust and nuanced outputs from ChatGPT 03 Mini and Microsoft Co-Pilot.

When asked about sensitive historical events like Tiananmen Square, Deep Seek refused to provide any information, citing its design to give "helpful and harmless" responses. In contrast, ChatGPT 03 Mini provided a balanced overview of the events.

On more neutral technical prompts, such as implementing a convolution operation in Lua, all three models produced similar code snippets, though ChatGPT 03 Mini and Co-Pilot tended to give more detailed explanations.

The most striking difference emerged when testing the models' ability to handle repetitive prompts. While ChatGPT 03 Mini and Co-Pilot had sensible stopping points, Deep Seek continued generating the repeated text endlessly, suggesting potential issues with its safeguards against abuse.

Overall, this exploration highlights the importance of transparency, robustness, and ethical considerations in the development of AI systems, especially those with potential geopolitical implications. The contrast between the models underscores the need for continued scrutiny and responsible innovation in the rapidly evolving field of artificial intelligence.