<div dir="ltr"><div dir="ltr"><br></div><br><div class="gmail_quote gmail_quote_container"><div dir="ltr" class="gmail_attr">On Sat, Feb 21, 2026 at 3:15 PM 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 dir="ltr"><br></div><br><div class="gmail_quote"><div dir="ltr" class="gmail_attr">On Sat, Feb 21, 2026 at 2:16 PM BillK via extropy-chat <<a href="mailto:extropy-chat@lists.extropy.org" target="_blank">extropy-chat@lists.extropy.org</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></div><div class="gmail_quote"><div dir="ltr" class="gmail_attr">On Sat, 21 Feb 2026 at 17:49, Jason Resch via extropy-chat <<a href="mailto:extropy-chat@lists.extropy.org" target="_blank">extropy-chat@lists.extropy.org</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">Manual refinement of LLM chat bots we interact with is common, but it's not necessary to build an LLM that understands and reasons.<div><br></div><div>There is the "Pure Transformer" that is produced from simply training on a large sample of text, and this requires no labeling, or manual adjustments. GPT-2 and GPT-3 were examples of pure transformers. There was also a pure GPT-4 that was never made public, before it was given constraints around what things it can or can't do. What is interesting is that the intelligence of this pure model was measured to be significantly higher before it was put through this manual adjustment (we might liken it to being lobotomized).</div><div><br></div><div>This is a generally recognized phenomenon: <a href="https://share.google/aimode/Xz0ejYy73wOt5nQEc" target="_blank">https://share.google/aimode/Xz0ejYy73wOt5nQEc</a> </div><div>"In summary, while RLHF might "lobotomize" certain creative or reasoning edges of a base model, it is currently the industry standard for making AI usable and safe for the general public."</div><div><br></div><div>DeepMind encountered a similar phenomenon, when they observed that their Go model when pre-trained initially on records of human games, produced a less skilled player than a model trained on <b>zero</b> human inputs (hence the name "AlphaZero").</div><div><br></div><div>So I suppose my overall point is that while "granular labor of human annotators" is common, it's unnecessary for an AI to develop of meaning and understanding.</div><div><div><br></div><div>Jason</div></div></div><div class="gmail_quote"><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">
</blockquote></div>
_______________________________________________</blockquote><div> </div><div><br></div><div style="font-family:arial,sans-serif;font-size:small;color:rgb(0,0,0)">Gemini says that these Pure Transformers were never released because they were unusable.</div><div style="font-family:arial,sans-serif;font-size:small;color:rgb(0,0,0)">Labelling and manual adjustments are essential to LLM development.</div><div style="font-family:arial,sans-serif;font-size:small;color:rgb(0,0,0)">BillK</div></div></div></blockquote><div><br></div><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><br></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><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><br></blockquote></div></div></blockquote><div><br></div><div>P.S. They have come quite a long way from those initial instruction sets. Here is a recent one by Anthropic:</div><div><br></div><div><a href="https://github.com/jujumilk3/leaked-system-prompts/blob/main/anthropic-claude-opus-4_20250805.md">https://github.com/jujumilk3/leaked-system-prompts/blob/main/anthropic-claude-opus-4_20250805.md</a></div><div><br></div><div>Jason </div></div></div>