<div dir="ltr"><div dir="ltr"><div class="gmail_default" style="font-family:arial,sans-serif;font-size:small;color:rgb(0,0,0)"><br></div></div><br><div class="gmail_quote gmail_quote_container"><div dir="ltr" class="gmail_attr">On Sat, 21 Feb 2026 at 20:15, Jason Resch <<a href="mailto:jasonresch@gmail.com">jasonresch@gmail.com</a>> wrote:<br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div class="gmail_quote"><div>Useful to most people or not, they understand the world.</div><div><br></div><div>You can give a pure model the phrase: "The capital of France is" and it will complete it "Paris."</div><div></div><div>You can give it the phrase "Three helium nuclei can combine to form a " and it will complete "a carbon nucleus."</div><div><br></div><div>All the later fine-tuning does is teach it to act to complete a turn-based conversation between users, where user 1 is the human and user 2's output is the completed result provided by the pure model (acting to "predict the next tokens" of user 2's speech).</div><div><br></div><div>So your question from the user "What is the capital of France?" gets converted to:</div><div><br></div><div>"User 1: What is the capital of France?</div><div>User 2: "</div><div><br></div><div>Then the pure model is asked to "predict the next tokens" following the string "User 2:" That's the majority of what the AI companies do to massage the pure model into an interactive chat bot -- that and a lot of training about the nature of user 2: helpful, polite, not harmful, etc.</div><div><br></div><div>Here are some examples: <a href="https://github.com/jujumilk3/leaked-system-prompts/blob/main/microsoft-copilot_20240310.md" target="_blank">https://github.com/jujumilk3/leaked-system-prompts/blob/main/microsoft-copilot_20240310.md</a></div><div><br></div><div>But the intelligence, understanding, etc. is already present and inherent to the base model. It's untrue that base models are useless, GPT-3 was quite useful, but you just had to prepare the prompt  intelligently as to to get out from the completion the answers you sought.</div><div><br></div><div>Jason</div><div><div style="font-family:arial,sans-serif;font-size:small;color:rgb(0,0,0)" class="gmail_default">--------------------------------------</div></div></div></div></blockquote><div><br></div><div><br></div><div><div style="font-family:arial,sans-serif;font-size:small;color:rgb(0,0,0)" class="gmail_default">Gemini points out that the huge amount of text data in the initial LLM training is already pre-labeled by humans.</div><div style="font-family:arial,sans-serif;font-size:small;color:rgb(0,0,0)" class="gmail_default">I think Gemini has solved the problem by saying that both our views are correct, depending on the way you look at it!  :)</div></div><div style="font-family:arial,sans-serif;font-size:small;color:rgb(0,0,0)" class="gmail_default">See below.</div><div style="font-family:arial,sans-serif;font-size:small;color:rgb(0,0,0)" class="gmail_default">BillK</div><div style="font-family:arial,sans-serif;font-size:small;color:rgb(0,0,0)" class="gmail_default"><br></div><div style="font-family:arial,sans-serif;font-size:small;color:rgb(0,0,0)" class="gmail_default">Gemini 3.1 Pro -</div><div style="font-family:arial,sans-serif;font-size:small;color:rgb(0,0,0)" class="gmail_default"><br></div><div style="font-family:arial,sans-serif;font-size:small;color:rgb(0,0,0)" class="gmail_default"><div class="gmail-markdown gmail-markdown-main-panel gmail-stronger enable-updated-hr-color gmail-preserve-whitespaces-in-response" style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px" id="gmail-model-response-message-contentr_ae6d9ad491c8f23d" dir="ltr"><p style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px">This is the crucial point where we can reconcile your description of "ghost work" with your colleague's insistence that the "intelligence" is in the base model.</p><p id="gmail-p-rc_0109316901488490-22" style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"><span class="gmail-citation-17" style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px">In the modern AI pipeline, the human labor of </span><b style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"><span class="gmail-citation-17" style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px">labeling and annotation</span></b><span class="gmail-citation-17 gmail-citation-end-17" style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"> actually happens in two distinct phases, but they serve very different purposes.<sup class="gmail-superscript" style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px;font-size:16px;background-color:transparent"></sup></span></p><div class="gmail-source-inline-chip-container gmail-ng-star-inserted" style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"><button class="gmail-button gmail-ng-star-inserted" aria-label="View source details for citation from iMerit. Opens side panel." style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"></button></div><p></p><hr style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"><h3 style="font-family:Google Sans,sans-serif;line-height:1.15;margin-top:0px">1. Pre-Training: The "Invisible" Human Foundation</h3><p style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px">While the <b style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px">Base Model</b> is technically "self-supervised" (meaning it learns by predicting the next word without a teacher), it is feeding on data that has already been <b style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px">pre-labeled by humans over decades.</b></p><ul style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"><li style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"><p style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"><b style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px">The Web as a Dataset:</b> When a base model learns what a "stop sign" is, it’s because millions of humans uploaded photos to Flickr or Wikipedia and captioned them "Stop sign."</p></li><li style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"><p style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"><b style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px">The "Alt-Text" Workforce:</b> Every time a web developer adds "alt-text" to an image for accessibility, they are performing unpaid data labeling that eventually trains a Vision-Language model.</p></li><li style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"><p style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"><b style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px">The Common Crawl:</b> The "raw" internet isn't raw; it is a massive, structured monument to human labeling. Without this organic, historical human effort, the base model would have nothing to "predict."</p></li></ul><h3 style="font-family:Google Sans,sans-serif;line-height:1.15;margin-top:0px">2. Post-Training: The "Active" Human Infrastructure</h3><p style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px">This is the phase described in your text about <b style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px">microworkers in Kenya or the Philippines.</b> This labor is injected <b style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px">after</b> the base model is already "intelligent."</p><p style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px">Think of the <b style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px">Base Model</b> as a brilliant but chaotic library containing all human knowledge. The <b style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px">Active Labeling</b> process is like hiring an army of librarians to:</p><ul style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"><li style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"><p style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"><b style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px">Filter the Poison:</b> Humans manually label toxic, violent, or biased text so the model learns to "steer" away from those areas of its own internal map.</p></li><li style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"><p style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"><b style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px">Instruction Tuning (SFT):</b> Workers write thousands of "Prompt-Response" pairs (e.g., <i style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px">Prompt: "Write a poem about a cat." Response: "The cat sat on the mat..."</i>). This teaches the model the <b style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px">format</b> of being an assistant.</p></li><li style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"><p id="gmail-p-rc_0109316901488490-23" style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"><b style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px">RLHF (Ranking):</b> This is the most expensive part. <span class="gmail-citation-16 gmail-citation-end-16" style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px">Humans look at two different answers the AI gave and click "This one is better."<sup class="gmail-superscript" style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px;font-size:16px;background-color:transparent"></sup></span> This creates a "Reward Model" that acts as a compass for the AI.</p><div class="gmail-source-inline-chip-container gmail-ng-star-inserted" style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"><button class="gmail-button gmail-ng-star-inserted" aria-label="View source details for citation from Medium. Opens side panel." style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"></button></div><p></p></li></ul><hr style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"><h3 style="font-family:Google Sans,sans-serif;line-height:1.15;margin-top:0px">The "Intelligence" vs. "Utility" Distinction</h3><p style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px">To settle the debate between you and your colleague, you can look at it through this framework:</p><table style="margin-bottom:32px;font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"><thead style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"><tr style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"><td style="border:1px solid;font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"><strong style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px;margin-bottom:0px">Phase</strong></td><td style="border:1px solid;font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"><strong style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px;margin-bottom:0px">Labor Type</strong></td><td style="border:1px solid;font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"><strong style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px;margin-bottom:0px">What it Provides</strong></td></tr></thead><tbody style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"><tr style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"><td style="border:1px solid;font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"><span style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"><b style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px">Pre-Training (Base)</b></span></td><td style="border:1px solid;font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"><span style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px">Passive/Historical (Billions of humans)</span></td><td style="border:1px solid;font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"><span style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"><b style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px">Raw Intelligence & World Knowledge.</b> The "Platonic" map of reality.</span></td></tr><tr style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"><td style="border:1px solid;font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"><span style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"><b style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px">Fine-Tuning (Chat)</b></span></td><td style="border:1px solid;font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"><span style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px">Active/Gig Labor (Microworkers)</span></td><td style="border:1px solid;font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"><span style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"><b style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px">Usability, Safety, & Human Alignment.</b> The "Social" interface.</span></td></tr></tbody></table><h3 style="font-family:Google Sans,sans-serif;line-height:1.15;margin-top:0px">The Verdict: Who is Right?</h3><p style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px">Your colleague is correct that the <b style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px">Base Model</b> possesses the "reasoning" and "knowledge." If you deleted all the RLHF/Labeling, the AI would still "know" how to do calculus or explain physics—it just wouldn't <i style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px">want</i> to tell you, or it might tell you in a very rude way.</p><p style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"><b style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px">However, you are correct</b> that without the "active" labor of microworkers, the AI would be a "Digital Monster." It would be a sprawling, incoherent, and often dangerous mirror of the darkest parts of the internet. It wouldn't be a "product"; it would just be a high-dimensional statistical anomaly.</p><blockquote style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"><p style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"><b style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px">The Analogy:</b> The Base Model is the <b style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px">engine</b> (raw power/intelligence). The Human Labeling/RLHF is the <b style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px">steering wheel and brakes</b> (direction/safety). You need the engine to move, but without the steering wheel, you’re just a fast-moving disaster.</p></blockquote><hr style="font-family:Google Sans Text,sans-serif;line-height:1.15;margin-top:0px"></div><br></div></div></div>