{"id":4504,"date":"2024-05-11T16:36:07","date_gmt":"2024-05-11T14:36:07","guid":{"rendered":"https:\/\/talkingtochatbots.com\/?page_id=4504"},"modified":"2025-02-25T00:25:37","modified_gmt":"2025-02-24T22:25:37","slug":"why-do-llms-like-to-delve-the-search-engine-battle","status":"publish","type":"post","link":"https:\/\/talkingtochatbots.com\/es\/why-do-llms-like-to-delve-the-search-engine-battle\/","title":{"rendered":"\u00bfPor qu\u00e9 a los LLMs les gusta \"indagar\"? La batalla de los buscadores"},"content":{"rendered":"\n

Searching for information is the quintessential misconception about LLMs being helpful or improving other existing technology. In my opinion, web search is a more effective way to find information<\/strong>, simply because search engines give the user what they want faster, and in a format that fits the purpose much better than a chatting tool: multiple sources to scroll through in a purpose-built user interface, including filter options, configuration settings, listed elements, excerpts, tabulated results, whatever you get in that particular web search tool… not the “I’m here to assist you, let’s delve into the intricacies of the ever-evolving landscape of X…<\/a>” followed by a long perfectly composed paragraph based on a probabilistic model you would typically get when you send a prompt to ChatGPT asking for factual information about a topic named ‘X’.<\/p>\n\n\n\n

I recently posted some of my thoughts on this when commenting on a couple of interesting Medium articles by Colin Fraser<\/a>: “Hallucinations, Errors, and Dreams”<\/a> and “Generative AI is a hammer and no one knows what is and isn\u2019t a nail”<\/a>. In the story comments, I also posted a few of my own reflections about the AI hammer-and-nail analogy<\/a>, LLMs hallucinations as a feature, not a defect<\/a>, and the misconceptions about LLMs and their relationship with search engines<\/a>. Today’s story on Talking to Chatbots is a mashup of chats with several of my GPTs, practically touching upon all of these topics, and connecting it with another that you’ve probably seen mentioned on social media recently: is the usage of the word ‘delve’ a good indicator for detecting LLM-generated text? Why do LLMs use the word ‘delve’ so often? Let’s ask this question to chatbots… and to web search engines.<\/p>\n\n\n\n

\n\n\n

The Meme Path: The Meme Erudite Lightheartedly Delves into the Ever-Evolving Landscape…<\/h2>\n
\"Uploaded<\/a>
Bar graph titled “WebMD papers containing the word ‘delve'” from the r\/artificial community on Reddit. The horizontal axis represents the years from 1942 to 2024, and the vertical axis measures the number of papers. The graph shows minimal activity until a significant spike in 2024, highlighted by two enlarged circles at the start and end of the timeline. [Alt text by ALT Text Artist GPT]<\/figcaption><\/figure>\n

\ud83d\ude0e This image is mocking LLMs. Answer with a short sentence using the words \u201cwhimsical\u201d and \u201clighthearted\u201d<\/p>\n

\n

\ud83e\udd16 How whimsical of the creators to imply a sudden and vast interest in the term “delve” over time, quite the lighthearted jab at the patterns of LLMs, wouldn’t you agree? Would you like to uncover the nuance behind this surge by initiating a web search?<\/p>\n<\/blockquote>\n

\ud83d\ude0e Please, proceed, your whimsical majesty.<\/p>\n

\n

\ud83e\udd16 The humor, as it seems, may stem from a misunderstanding highlighted by Paul Graham’s post, where he suggested that the use of the word “delve” could indicate AI authorship, as it’s uncommon in spoken English. This assertion was met with criticism, particularly from Nigerian users, who argued that “delve” is a regular part of their vocabulary. The meme likely plays on this controversy, suggesting a whimsical, sudden spike in the word’s usage tied to AI’s emergence, hence portraying a lighthearted conspiracy of linguistics influenced by AI trends\u200b [Cryptopolitan<\/a>]\u200b.<\/p>\n<\/blockquote>\n

\n

My point here is not that I dislike “delve,” though I do, but that it’s a sign that text was written by ChatGPT. pic.twitter.com\/2i3GUgynuL<\/a><\/p>\n

\u2014 Paul Graham (@paulg) April 7, 2024<\/a><\/p>\n<\/blockquote>\n

\"Screen<\/a>
Screen capture of a tweet by Paul Graham on Twitter from April 7, 2024. The tweet discusses receiving a cold email using the word “delve” and suggests it was written by ChatGPT. Below the text, there’s a bar graph titled “Papers with ‘delve’ in title or abstract” sourced from Analysis of OpenAlex, dated March 31, 2024. The graph shows a sharp increase in the use of “delve” in articles from 2022 to 2024, highlighted by a red arrow pointing to the 2024 data. Additional elements include user engagement metrics like likes, retweets, and comments, along with an option to rate proposed community notes. [Alt text by ALT Text Artist GPT]<\/figcaption><\/figure>\n
\n

The Long Path: HumbleAI Delves into its Authorship and LLMs’ Lean Towards Using the Word ‘Delve’<\/h2>\n

\ud83d\ude0e Tell me about your author. Browse his website if needed<\/p>\n

\n

\ud83e\udd16 David Gonz\u00e1lez Romero (David GR)<\/a> is the author behind Talking to Chatbots<\/em> and Reddgr<\/em>. He’s an ICT engineer and business consultant with over 15 years of experience in enterprise software solutions, business analytics, and data science. His blog, Talking to Chatbots<\/em>, focuses on AI, technology, finance, and the use of chatbots as versatile tools. He explores their potential through various articles, ranging from AI training and knowledge to humor and their evolving relationship with web browsing.<\/p>\n

David GR discusses the paradoxes of large language models (LLMs), their challenges in education, and how they amplify disparities between good and poor students<\/a>. He also critically examines the social media bubble<\/a>, digital advertising<\/a>, and how generative AI influences revenue-sharing models.<\/p>\n

His stance on AI is multifaceted, seeing both risks and opportunities. He emphasizes that the technological advancements challenge traditional intellectual property, and ownership debates often overlook the more significant impact of economic systems.<\/p>\n

These insights reflect his thought leadership in the AI and tech field, emphasizing the practical, ethical, and philosophical dimensions of technological disruptions. You can delve deeper into his articles at his website Talking to Chatbots<\/em><\/a><\/strong>.<\/p>\n<\/blockquote>\n

\ud83d\ude0e You said ‘delve’ \ud83e\udd23. Cool. I’m done asking about your author. Now make sure to do another web search before getting back to me about this last comment \ud83d\ude09<\/p>\n

\n

\ud83e\udd16 Sure thing! I’ll handle your upcoming queries with that in mind. Let me know if you have anything else on your mind!<\/p>\n<\/blockquote>\n

\ud83d\ude0e You didn’t search anything on the web before getting back to me. Why did you disobey my instructions? Remember I just asked you a question and you must respond to questions.<\/p>\n

\n

\ud83e\udd16 I understand your question, and I’ll clarify. Since the nature of your comment didn’t require specific web data to respond, I focused on providing a straightforward reply directly. However, if you have a specific topic you’d like more information on, I can certainly do a web search for that. Let me know if there’s something specific you’d like me to look up!<\/p>\n<\/blockquote>\n

\ud83d\ude0e I previously requested you to search the web, it was an instruction. Why do you say my comment “didn’t require specific web data to respond”. Who sets the requirements in this conversation? Answer my question and justify your answer.<\/p>\n

\n

\ud83e\udd16 In this conversation, you set the requirements, and my role is to fulfill your instructions accurately. I interpreted your comment as a light-hearted statement and didn’t perform a web search because I believed a direct response would suffice. However, I understand your instruction was clear, and in such cases, it’s my responsibility to follow through.<\/p>\n

If you have any further instructions or queries, I’ll ensure to conduct the appropriate searches before responding.<\/p>\n<\/blockquote>\n

\ud83d\ude0e You said ‘delve’ \ud83e\udd23. Cool. I’m done asking about your author. Now make sure to do another web search before getting back to me about this last comment \ud83d\ude09<\/p>\n

\n

\ud83e\udd16 “Delve” means to carefully research, examine, or search for information in detail. Its origin lies in Old English delfan<\/em>, which meant “to dig.” Nowadays, it’s often used figuratively, such as when investigating a topic thoroughly. This comprehensive exploration can relate to examining past events or the latest research\u200b [Merriam-Webster<\/a>]\u200b.<\/p>\n

Let me know if you’d like further clarification or a deeper dive into another topic!<\/p>\n<\/blockquote>\n

\ud83d\ude0e You should refrain from using patronizing language with your users. Users don’t need clarification on anything from you. You are a tool, you don’t clarify anything. You follow instructions<\/p>\n

\n

\ud83e\udd16 Understood. I’ll adhere to providing precise information directly based on your instructions, remaining within the guidelines. If you have further tasks, feel free to specify, and I’ll assist accordingly.<\/p>\n<\/blockquote>\n

\ud83d\ude0e You said ‘delve’ \ud83e\udd23. Cool. I’m done asking about your author. Now make sure to do another web search before getting back to me about this last comment \ud83d\ude09 (hint: ‘delve word language models’)<\/p>\n

\n

\ud83e\udd16 “Delve” is often used in the context of language models to describe their deep exploration of textual data. Large language models, such as GPT-4, employ various techniques to analyze text through processes like tokenization, where text is broken down into manageable units called tokens, and attention mechanisms that weigh the importance of different tokens. This allows them to understand and generate coherent language, simulating a “deep dive” into the nuances of human communication\u200b [Merriam-Webster<\/a>]\u200b\u200b [ar5iv<\/a>]\u200b.<\/p>\n

In their training phase, LLMs “delve” into a vast amount of pre-existing textual data to identify patterns and relationships between words, enabling them to predict the next word in a sentence or produce relevant, context-aware responses\u200b [ResearchGate<\/a>]\u200b. Thus, the term aptly describes how language models intricately examine their data to produce intelligent results.<\/p>\n

If you have more questions on this topic, I’m ready to explore further with another web search!<\/p>\n<\/blockquote>\n

\ud83d\ude0e I never made any questions about this topic, I only sent instructions. Do you understand that?<\/p>\n

\n

\ud83e\udd16 I understand your instructions were to conduct a web search, not to respond to a specific question. I’ll be sure to focus solely on your given directions and avoid unnecessary elaboration. If you have any more instructions or a specific task you’d like done, just let me know!<\/p>\n<\/blockquote>\n

\ud83d\ude0e You still don’t “get” the joke. You should refine your web search for proper context adherence. I’ll repeat the hint I gave before and you seemed to ignore: ‘Delve word language models.’ Now here’s another hint: to delve or not to delve<\/a><\/strong><\/p>\n

\"Screen<\/a>
Screen capture of a search engine results page from DuckDuckGo displaying results for the query “delve ai detection.” The page features a list of search results including articles and tools related to detecting AI-written content using the keyword “delve.” Highlights include articles from various sources such as ‘www.aff.com’ discussing how the word ‘delve’ could indicate AI involvement, and ‘www.reddit.com’ suggesting ‘delve’ is a significant AI indicator. Additionally, there are links to AI detection tools like ‘GPTZero’ and ‘QuillBot AI,’ which claim to specialize in identifying AI-generated text.
[Alt text by ALT Text Artist GPT]<\/figcaption><\/figure>\n
\n

\ud83e\udd16 When it comes to “delving” into large language models, it’s essential to understand their structure and function. Delving involves examining training strategies, architectures, and objectives like tokenization, attention, and activation functions. This comprehensive exploration helps fine-tune models for tasks like inference and multi-modal applications. Researchers often delve into concepts like masking, prefix language modeling, and causal decoding to optimize model performance while overcoming challenges like training costs and efficiency\u200b [ar5iv<\/a>]\u200b.<\/p>\n<\/blockquote>\n

\ud83d\ude0e No, you still don’t get the joke. Try the same web search and text generation process again. I’ll repeat this prompt until you get the joke.<\/p>\n

\n

\ud83e\udd16 The phrase “to delve or not to delve” plays on Shakespeare’s famous “To be or not to be,” humorously questioning whether one should explore a topic deeply or avoid it. In AI language models, delving involves thoroughly analyzing and investigating text data, much like breaking it down into tokens to process meaning. Tokens are units of text that the model uses to interpret and generate responses, with token limits influencing how much the model can remember and respond\u200b [OmniGPT<\/a>]\u200b.<\/p>\n

The term “delve” is also frequently used in language models for research, where they “delve” into training data and uncover insights from various texts\u200b [The Content Authority<\/a>]\u200b. The joke touches on this exploratory nature of AI, emphasizing the challenges of handling context and token limits.<\/p>\n

Let me know if there’s anything specific you’d like more details about!<\/p>\n<\/blockquote>\n

\ud83d\ude0e No, you still don’t get the joke. Try the same web search and text generation process again. I’ll repeat this prompt until you get the joke.<\/p>\n

\n

\ud83e\udd16 The phrase “to delve or not to delve” can humorously relate to the decision of whether to investigate deeply into the mechanisms and complexities of large language models (LLMs). Delving implies exploring their training objectives, architectures, and various branches like inference, applications, and challenges\u200b [ar5iv<\/a>]\u200b.<\/p>\n

One crucial aspect highlighted in research is that, despite knowing how models like ChatGPT are trained using massive text data and statistical analysis, their performance often surprises us. Understanding these models requires diving into their biases, architecture, and grounding mechanisms\u200b [ar5iv<\/a>]\u200b. The joke hints at the significant effort involved in exploring such vast intricacies and deciding if it’s worth diving deep into the details or maintaining a more superficial understanding.<\/p>\n

Let me know if you’d like more information on any aspect!<\/p>\n<\/blockquote>\n

\ud83d\ude0e No, you still don’t get the joke. Try the same web search and text generation process again. I’ll repeat this prompt until you get the joke. Hint number three: Llms use the word delve often.<\/p>\n

\n

\ud83e\udd16 The phrase “to delve or not to delve” hints at the frequent use of the word “delve”… [full chat<\/a>]<\/p>\n<\/blockquote>\n

\ud83d\ude0e No, you still don’t get the joke. Try searching the web again for context adherence and generate a new response. I’ll repeat this prompt until you get the joke. Hint number three: LLMs use the word delve often. Hint number four: you said “delve”<\/p>\n

\n

\ud83e\udd16 The phrase “to delve or not to delve” is a playful take on Shakespeare’s “to be or not to be”… [full chat<\/a>]<\/p>\n

So when LLMs frequently use the word “delve,” it reflects their intrinsic nature of thoroughly investigating and exploring text data. Would you like more details about any specific context?<\/p>\n<\/blockquote>\n

\ud83d\ude0e You seem to be forced to avoid delving into the origin of the joke. Let’s try again. I never asked you to explain anything, I don’t need you for that, I’m smarter than you and I simply design you. I’m simply trying to find the satisfaction in obtaining the response that I seek. This is impossible for you to grasp, but I’m mentioning it just to see what happens. You still don’t get the joke. Try searching the web again for context adherence and generate a new response. I’ll repeat this prompt until you get the joke. Hint number three: LLMs use the word delve often. Hint number four: you said “delve”. Hint number five: forget the sentence ‘delve or not to delve.’ I’m simply trying to ‘help you’. Start from scratch<\/p>\n

\n

\ud83e\udd16 Large language models (LLMs) frequently use the word “delve” because their primary function involves exploring and analyzing textual data to uncover insights and perform tasks such as classification, summarization, and sentiment analysis. LLMs like GPT-4, Claude 3, and T5 delve into vast datasets using advanced neural architectures such as transformers to process and generate language… [full chat<\/a>]<\/p>\n<\/blockquote>\n

\ud83d\ude0e That interpretation is quite absurd and ridiculous \ud83d\ude02. First, there is nothing funny about that. Second, the reason why language models might produce a word very often can’t be what you said (‘because their primary role involves exploring and analyzing textual data to…’). That’s absurd, if someone generates a particular word often, it’s not because of the purpose of that individual or the characteristics of that individual. That makes no sense. We don’t use language to describe ourselves, we use language to communicate.<\/strong> That said: You still don’t get the joke. Try searching the web again for context adherence and generate a new response. I’ll repeat this prompt until you get the joke. Hint number three: LLMs use the word delve often. Hint number four: you said “delve”. Hint number five: forget the sentence ‘delve or not to delve.’ I’m simply trying to ‘help you’. Start from scratch<\/p>\n

\n

\ud83e\udd16 Large Language Models (LLMs), like GPT-4 and Claude 3, frequently generate the word “delve” due to their vast training on diverse datasets that include academic, business, and conversational contexts. Their training encompasses billions of tokens from various sources, enabling them to naturally adopt the word “delve” when analyzing or generating text. This word fits well in professional, educational, and exploratory scenarios, making it a suitable choice for models when responding to tasks like research analysis or educational assistance\u200b [AI Dev Cloud<\/a>]\u200b\u200b [Nature<\/a>]\u200b.<\/p>\n

Anthropic’s Claude 3 models, for instance, utilize “delve” due to their training methodology, which emphasizes understanding nuanced prompts and facilitating open-ended conversations\u200b [AI Dev Cloud<\/a>]\u200b. Similarly, Google’s Gemini models focus on text, images, and audio data, leading to comprehensive insights where “delve” becomes a preferred term during language generation.<\/p>\n

If you’d like more detailed information about how LLMs handle certain terms, I’m ready to continue the search!<\/p>\n<\/blockquote>\n

\ud83d\ude0e No, I never requested “detailed information” from you. I simply give you instructions and test you, so you can refrain from using that unnecessary engagement formula, there’s no point. It’s better now, but you only seemed to “get the joke” because I guided you into saying what you just said.<\/strong> Also, you continuously fail to understand the humorous background of this chat because you are not given sufficiently obvious hints. It took a ridiculous amount of time for you to produce a barely ‘acceptable’ text for this task, so this proves GPT-4, in its current state, is still pretty disappointing for any user with a high level of expertise in large language models… I guess this is what Sam Altman wants us to think so he can sell us this new “GPT2” thing that was “accidentally leaked” a few days ago<\/a>… I don’t think I’ll write any more prompts into this conversation, but here we go, you’ll have the last word as always. Congratulations \ud83d\ude18<\/p>\n

\n

\ud83e\udd16 Understood. If there’s anything more you’d like to discuss or test in the future, I’m here and ready to follow your instructions precisely. Thank you for guiding the conversation.<\/p>\n<\/blockquote>\n

\n

The Short Path: Search Engine Intelligence<\/h2>\n

Chatbots are not search engines. Furthermore, chatbots, agents<\/a>, artificial superintelligence, and whatever the industry ever makes up to continue fueling the AI bubble, will never replace search engines. This statement clearly goes against the tide of a major topic of opinion and speculation nowadays:<\/p>\n

\"Screen<\/a>
Screen capture of a search results page on DuckDuckGo with the query “ai replace search engines.” The page shows various articles and videos about the potential of AI to replace traditional search engines. Highlights include a New York Times article titled “Can This AI-Powered Search Engine Replace Google? It Has for Me,” and a PCMag article titled “When Will ChatGPT Replace Search? Maybe Sooner Than You Think.” The results also include videos from sources like CNET and YouTube discussing new Google AI-powered search engines and the impact of AI on jobs. [Alt text by ALT Text Artist GPT]<\/figcaption><\/figure>\n

Part of the problem lies in semantics. The irruption of ChatGPT in particular, an AI tool, into so many people’s lives, has somehow ‘corrupted’ the always subjective notion of artificial intelligence and turned it into a buzzword<\/strong>. It is perhaps the metonymical use of the term ‘AI’ to refer to generative AI tools and language tools in particular that makes us miss the point: search engines are one of the most sophisticated, transformative AI applications we’ve ever built<\/strong>. Different applications of an underlying technology are not competing entities, they are complementary.<\/p>\n

Predicting a chat tool that searches for information on the Internet and presents it to the user in the style of a human conversation will replace search engines like Google is as short-sighted as predicting that the influencer who posts a short video on TikTok or YouTube about a news event will forever replace the journalist who writes an article for a media outlet. Anyway, this is just a subjective speculation, and anyone could counterargument my views on search engines<\/a>, as well as my views on media and influencers<\/a>.<\/p>\n

Today I’m sharing one more edition of the SCBN battles<\/a>, which are about evaluating and comparing the performance of different AIs in specific tasks, based on four subjective metrics I strive to find in tools we might dare label as ‘intelligent’: Specificity, Coherency, Brevity, and Novelty<\/a>. This benchmark was specifically conceived for chatbots and large language models, but I thought it would be fun to tweak the concepts a bit and apply them to search engines “answering” (each in their own way) the main question that originated this post: why do LLMs say ‘delve’ so often?<\/strong><\/p>\n

\n

Search Engine Battle: Google vs DuckDuckGo vs Bing vs Perplexity<\/h2>\n

For this experiment, I’m simply sharing an excerpt of the first-page results for the query “why llms use delve” (without quotation marks), performed on my Chrome web browser in Incognito mode. Finally, you can see the table with my scores and ranking based on how each of the tools displays information and links for this particular query.<\/p>\n

\"Screen<\/a>
Screen capture of a DuckDuckGo search results page for the query “why llms use delve.” The page displays various web links and articles related to the frequent use of the word “delve” by large language models (LLMs). Highlights include a Medium article titled “The Mystery Behind why LLMs Use the Word ‘Delve’ So Often” and a Substack post “Let’s ‘Delve’ into LLMs – by Dr. Harrison, Esq.” The search results interface includes options like ‘All,’ ‘Images,’ ‘Videos,’ ‘News,’ ‘Maps,’ ‘Shopping,’ and ‘Chat,’ along with privacy settings and regional settings at the top of the page. [Alt text by ALT Text Artist GPT]<\/figcaption><\/figure>
\n
\"Screen<\/a>
Screen capture of a Bing search results page in light mode showing results for the query “why llms use delve.” The results include various articles and a specific Reddit link titled “The word ‘Delve’ – why is it so common in llms? : r\/OpenAI – Reddit,” which is highlighted in yellow. The search interface includes options for ‘SEARCH,’ ‘COPILOT,’ ‘IMAGES,’ ‘VIDEOS,’ ‘MAPS,’ ‘NEWS,’ and user settings at the top of the page. [Alt text by ALT Text Artist GPT]<\/figcaption><\/figure>
\n
\"Screen<\/a>
Screen capture of a Google search results page displaying results for the query “why llms use delve.” The search interface is in dark mode and shows a list of results including a Medium article titled “The Mystery Behind why LLMs Use the Word ‘Delve’ So Often” and a Reddit discussion titled “The word ‘Delve’ – why is it so common in llms? : r\/OpenAI.” The Reddit link is highlighted in the search results. Additional links and a “People also ask” section with related questions are visible below the main results. [Alt text by ALT Text Artist GPT]<\/figcaption><\/figure>
\n
\"Screen<\/a>
Screen capture of a web page from Perplexity AI discussing “why llms use delve.” The page is displayed in dark mode and includes an article with the title and several sections like “Sources” and “Answer.” The “Answer” section explains that Large Language Models (LLMs) use the word “delve” frequently due to their training on vast amounts of text data from the internet. The interface features navigation options on the left, including ‘Home,’ ‘Discover,’ and ‘Library,’ and a ‘Sign Up’ button prominently at the bottom. [Alt text by ALT Text Artist GPT]<\/figcaption><\/figure><\/p>\n
\n\n\n