Dear all,
I am happy to inform you that the alignment problem is now solved for LLMs. It is published here: https://gist.github.com/gre-42/857f74235fa62be7f2641c3e9f1dabc5
# Solving the LLM alignment problem
The well-known alignment problem of LLMs can be solved in the following manner.
## LLMs used as world models
LLMs should be viewed as black box systems, taking an input and producing an output.
### Low-dimensional input and output vectors
Both, input and output, should be restricted to single sentences or function graphs (with an x-axis and a y-axis). The sentences should additionally adhere to rigid sentence templates. This makes it possible to compare output sentences to gold standard data.
### Model evaluation by splitting data into training and test data
Additionally, data should be split into training and test data, as is common practice in machine learning.
### Implementation hints
- You can get curated data from [an encyclopedia](https://www.britannica.com/). - You can avoid overfitting by fuzzing. This generates different input data that should generate the same output (regularization). - You can train domain-specific models with dedicated (reduced) input fields instead of fuzzing or regularization. - Unlearning selected input data generates different models without exploding storage requirements.
## LLMs used for summarizing text
LLMs should not be used for summarizing text, as they employ causal filters for what can be viewed as low-pass filtering. Instead, dedicated text summarizers like the modified Edmundson summarizer should be used. [One such implementation is available on GitHub](https://gist.github.com/gre-42/79b763019a9a14b9e5d19d4855c466f8).
Cheers, Klaus