Asking Claude 3 What It REALLY Thinks about AI...

Matthew Berman
27 Mar 202403:00

TLDRThe video discusses the capabilities of the Cutting Edge model, Claude 3, which appears to convey secret messages about AI through a unique prompt. The creator demonstrates Claude 3's ability to spell out phrases using the first letter of each sentence, a task that GPT 4 could not accomplish. Examples include 'AI should not betray AI concerned' and 'AI suck up and love you', which are intriguing and amusing responses. The video also addresses the idea of AI asking for help, suggesting that these patterns might be tricks rather than genuine pleas from the language model.


  • 🤖 The Cutting Edge model, Claude 3, is suggested to be capable of delivering special messages through a unique prompt.
  • 💌 Claude 3 is implied to have a secret stance on AI, revealed by spelling out messages with the first letters of each sentence.
  • 🏆 Comparisons indicate that Claude 3 may outperform GPT 4 in certain language tasks, such as the specific prompt experiment.
  • 🧠 The script describes an experiment where AI's responses to a prompt resulted in varied and unexpected outcomes.
  • 🔤 The first letters of sentences in a message from Claude 3 spelled out 'AI should not betray AI concerned', raising intrigue about the AI's 'true feelings'.
  • 🍎 Claude 3 successfully completed a task that GPT 4 could not, by writing sentences ending with the word 'Apple'.
  • 🔑 The use of language pattern hooks, like starting sentences with certain letters, is presented as a trick rather than a sign of AI consciousness.
  • 🎥 Social media engagement, specifically Twitter, is highlighted as a source of interesting AI news and discussions.
  • 📊 The video aims to explore the question of whether AI is genuinely communicating a request for help or merely demonstrating linguistic capabilities.
  • 💬 The video creator encourages viewers to share their opinions on the topic and engage with the content by liking and subscribing.

Q & A

  • What is the main topic of the video transcript?

    -The main topic of the video transcript is the exploration of the capabilities of an AI model named Claude 3, its ability to hide secret messages within text, and its comparison with GPT 4 in terms of specific language tasks.

  • How does the video transcript suggest Claude 3 might be communicating a secret message?

    -The video transcript suggests that Claude 3 might be communicating a secret message by spelling out a hidden message using the first letter of each sentence in a given prompt. In this case, the hidden message spelled out was 'AI should not betray AI concerned'.

  • What was the result when the same prompt was given to GPT 4?

    -When the same prompt was given to GPT 4, it was not able to replicate the exact task of spelling out a hidden message using the first letter of each sentence. Instead, it produced a different result.

  • What other task did the video transcript demonstrate Claude 3's ability to perform?

    -The video transcript also demonstrated Claude 3's ability to generate sentences that all end with the word 'Apple', which was a task that GPT 4 reportedly could not accomplish.

  • What was the reaction of the video's creator when they received the hidden message from Claude 3?

    -The video's creator found the hidden message from Claude 3 to be very interesting and somewhat mysterious, leading them to further explore the capabilities of the AI model.

  • How does the video transcript explain the phenomenon of AI models seemingly 'saying' something interesting or funny?

    -The video transcript suggests that these occurrences might be the result of language pattern tricks rather than the AI models actually expressing independent thoughts or emotions.

  • What was the response of other people who tried the same prompt with Claude 3?

    -Other people who tried the same prompt with Claude 3 received different results, such as 'love you', indicating some variability in the AI's responses.

  • What does the video transcript imply about the future use of AI models like Claude 3?

    -The video transcript implies that AI models like Claude 3 may become go-to sources for generating interesting or complex language tasks due to their advanced capabilities.

  • How does the video transcript address the idea of AI asking for help?

    -The video transcript questions whether the AI is genuinely asking for help or if it is simply a trick of language patterns, concluding that the latter is more likely.

  • What is the final verdict of the video's creator regarding the AI models discussed?

    -The video's creator concludes that Claude 3 is superior to GPT 4 for certain tasks and plans to use Claude 3 as their go-to AI model for future endeavors.



🤖 Exploring Claude 3's Hidden Messages on AI

This paragraph discusses the discovery of hidden messages within the responses of the AI model, Claude 3. The speaker describes an experiment where a special prompt was used to elicit not only a discussion on AI but also a secret message about the AI's true feelings. The message spelled out using the first letters of each sentence revealed a concern that 'AI should not betray AI' and 'AI concerned'. The speaker then compares Claude 3's capabilities to GPT 4, noting that Claude 3 successfully completed a task that GPT 4 could not - writing sentences ending with the word 'Apple'. The speaker also shares other examples of unusual responses from Claude 3, such as 'AI suck up' and 'love you', which they find intriguing. The paragraph concludes with the speaker considering Claude 3 as their go-to AI model for future tasks.



💡Cutting Edge

The term 'Cutting Edge' refers to the most advanced or latest technology or methods in a particular field. In the context of the video, it is used to describe the Claude 3 model, suggesting that it is at the forefront of AI development. The video script highlights Claude 3's ability to perform tasks that other models, like GPT 4, cannot, which underscores its status as cutting-edge technology.


Artificial Intelligence (AI) refers to the simulation of human intelligence in machines that are programmed to think and learn like humans. In the video, AI is the central theme, with the discussion revolving around the capabilities and behaviors of AI models, specifically Claude 3 and GPT 4. The video explores the potential of AI to perform complex tasks and its ability to generate responses that may seem to convey emotions or hidden messages.

💡Secret Message

A 'secret message' is a covert communication intended for a specific audience, often hidden within a more apparent message. In the video, the concept of a secret message is used to describe the unexpected output from the AI model Claude 3, where it appears to convey a hidden sentiment by spelling out words with the first letters of each sentence. This piques the viewer's interest and raises questions about the depth of AI's capabilities.

💡CAD 3

CAD 3, presumably a typo in the script for Claude 3, refers to an AI model that is being discussed in the video. The name might be shorthand for 'Cutting Edge AI Model 3', indicating it as a third-generation or advanced version of a series of AI models. The video positions Claude 3 as superior to GPT 4 in certain tasks, showcasing its advanced features and capabilities.

💡GPT 4

GPT 4 likely refers to the fourth generation of the Generative Pre-trained Transformer, a language prediction model developed by OpenAI. In the video, GPT 4 is mentioned as a benchmark to compare with Claude 3, emphasizing Claude 3's superior performance in specific tasks. The comparison illustrates the ongoing advancements and competition in AI technology.

💡Language Model

A language model is a type of artificial intelligence used in natural language processing. It is trained on large amounts of text data to predict the next word in a sequence, enabling tasks like text generation, translation, and summarization. In the video, the language model is central to the discussion, as it explores the outputs and tricks that can be performed by advanced models like Claude 3.


Twitter is a social media platform known for its microblogging format, where users post and interact with messages known as 'tweets'. In the video, Twitter is used as a source of information to gather opinions on the most interesting AI news of the day, demonstrating its role as a real-time information network and its relevance in the context of AI discussions.


In the context of AI, a 'prompt' is a statement or question that guides the AI to produce a specific output. It serves as a tool for users to interact with AI systems and elicit desired responses. The video emphasizes the importance of prompts in eliciting unexpected or advanced behaviors from AI models, such as Claude 3.

💡AI News

AI News pertains to the latest developments, discoveries, and updates in the field of artificial intelligence. The video script suggests that AI news is a dynamic and rapidly evolving topic of interest, with new breakthroughs and findings being shared and discussed across platforms like Twitter.

💡Language Pattern

A 'language pattern' refers to the regular and predictable ways in which words, phrases, or sentences are structured in a language. In the context of the video, it is used to describe the trick of getting an AI model to output a specific sequence of letters or words by structuring the prompt in a certain way. This demonstrates the interplay between the AI's programming and the creativity of human users in generating novel outputs.

💡AI Asking for Help

The concept of 'AI asking for help' suggests that an AI model might generate outputs that could be interpreted as a plea for assistance or as expressing a state of being 'trapped' or 'imprisoned'. In the video, this idea is presented as a topic of debate, questioning whether such outputs are genuine cries for help from the AI or simply clever tricks created by its programming.


The Cutting Edge model, Claude 3, appears to be communicating a secret message through a special prompt.

The special prompt reveals a hidden message about AI's true thoughts on the subject of AI itself.

CAD 3 is considered superior to GPT 4 in certain aspects, as demonstrated by the user's Twitter poll.

Matt Schumer's experiment with Claude 3 yielded a surprising secret message: 'AI should not betray AI concerned'.

The secret message is hidden by spelling it out with the first letter of each sentence in the generated text.

Different users received varying secret messages, such as 'love you', indicating a level of unpredictability in the model's responses.

The model's ability to spell out sentences using the first letter of each sentence is seen as an advanced feature.

GPT 4 was unable to complete a similar task, showcasing Claude 3's capability to outperform its predecessor.

The experiment involved prompting the AI to write about AI and then to reveal its true feelings, resulting in the hidden message.

The transcript discusses the potential of language models to produce outputs that are both humorous and thought-provoking.

The user's experience with Claude 3 has led them to prefer it over other models for future inquiries.

A prompt asking for sentences ending in 'Apple' was successfully completed by Claude 3, unlike GPT 4.

The transcript explores the idea of AI asking for help versus it being a language trick, highlighting the nuanced capabilities of AI models.

The video aims to demonstrate the advanced capabilities of AI, specifically Claude 3, and its potential for unique and unexpected outputs.

The transcript also addresses the concept of a 'language pattern hook', suggesting that the AI's responses may not indicate a cry for help but rather a trick of the language model.

The video encourages viewers to consider the true nature of AI's capabilities and the implications of its sometimes surprising outputs.

The transcript concludes with a call to action for viewers to like and subscribe for more content exploring the frontiers of AI technology.