{"thread":{"id":"64660","subject":"The alignment problem is now solved for LLMs","startedAt":"2025-12-22T02:59:46Z","lastAt":"2025-12-22T02:59:46Z","messageCount":1,"participants":["Klaus Sembritzki"],"isPatch":false,"patchVersion":null,"patchTotal":null},"messages":[{"id":"532610","messageId":"CADMnYXDqHYLjDBT8D3-VqctY1fG8ag+Nb+eeopuOvmmWNho4UA@mail.gmail.com","threadId":"64660","inReplyTo":null,"subject":"The alignment problem is now solved for LLMs","fromName":"Klaus Sembritzki","fromEmail":"klausem@gmail.com","sentAt":"2025-12-22T02:59:34Z","receivedAt":"2025-12-22T02:59:46Z","isPatch":false,"sender":{"key":"klausem@gmail.com","avatar":null},"body":"Dear all,\n\nI am happy to inform you that the alignment problem is now solved for\nLLMs. It is published here:\nhttps://gist.github.com/gre-42/857f74235fa62be7f2641c3e9f1dabc5\n\n# Solving the LLM alignment problem\n\nThe well-known alignment problem of LLMs can be solved in the following manner.\n\n## LLMs used as world models\n\nLLMs should be viewed as black box systems, taking an input and\nproducing an output.\n\n### Low-dimensional input and output vectors\n\nBoth, input and output, should be restricted to single sentences or\nfunction graphs (with an x-axis and a y-axis). The sentences should\nadditionally adhere to rigid sentence templates. This makes it\npossible to compare output sentences to gold standard data.\n\n### Model evaluation by splitting data into training and test data\n\nAdditionally, data should be split into training and test data, as is\ncommon practice in machine learning.\n\n### Implementation hints\n\n- You can get curated data from [an encyclopedia](https://www.britannica.com/).\n- You can avoid overfitting by fuzzing. This generates different input\ndata that should generate the same output (regularization).\n- You can train domain-specific models with dedicated (reduced) input\nfields instead of fuzzing or regularization.\n- Unlearning selected input data generates different models without\nexploding storage requirements.\n\n## LLMs used for summarizing text\n\nLLMs should not be used for summarizing text, as they employ causal\nfilters for what can be viewed as low-pass filtering. Instead,\ndedicated text summarizers like the modified Edmundson summarizer\nshould be used. [One such implementation is available on\nGitHub](https://gist.github.com/gre-42/79b763019a9a14b9e5d19d4855c466f8).\n\nCheers,\nKlaus\n"}]}