Diverse Saturday: Hirshman, LLM forecasting, warm-data, and automating posts

• Hirshman’s Tunnel Effect - the idea that people are ok with inequality if they believe they will eventually benefit Horatio Alger Myth… however, if they don’t believe this, then they become very upset.
• Hirshman’s Exit voice loyalty was mostly concerned with firms. However, to our knowledge, it’s never been applied to the concept of social movements and ‘voice’
• The ‘image’ of a matrix with respect to a vector is basically just them multiplied.
• To ‘solve’ a systems of equations you need to multiply by the inverse… I thinK?
  • A system of equations is essentially \(A*x=b\). To solve, you do \(x=A^-1*b\). The inverse is \(\frac{1}{det(A)}*Adj(A)\) (Adj is adjugate matrix)
  • Only works if A is a square matrix, and the determinant of A must not be 0.
• Warm Data: This is akin to metadata for emergent relationships within complex systems. Also akin to documentation for complex, designed spaces. However, I haven’t seen any compelling case studies of how this effects irl change.
• Medium Translation: Is using AI to translate the SAME MESSAGE between mediums. At the limit, this is a huge translation, but also smaller translations are possible…. (i.e. ton + audience)
• Forecasting with LLMs: A team at Berkeley is exploring using LLM agents to make systematic forecasts. An extremely interesting paper on using LLMs to forecase events used a RAG design to automatically search for relevant information and produce forecasts. The models were almost as good as the AGGREGATE of competitive forecasters, and beat it in some cases.
  • It seems like it could use deterministic code as a ‘scratchpad’ for predictions.
  • The RAG system allows it to get high quality data
• Simon Willison uses github’s issues as a more narrative, second brain for projects.
  • The ‘issues’ and project scaffolding tracks developments and allows him to switch rapidly between tasks without long information loading phases (as all the information and decisions are stored chronologically).
  • Testing for myself with KLH Website
• Setting up personal learning notes page/website
  • I’ve set up a blog page for my website (kamranhakiman.com/blog), and improved the readability of the very simple website (e.g. limited the width of text to wrap at 800px). I’m doing coding by hand for muscle memory learning.
  • Working with Claude and ChatGPT, I was able to produce an automated workflow which pulls from daily notes on my computer and automatically adds both the correct link into my website.
  • The actual program is in python. However, I’m pulling in both html and css.
• Maths
  • Binomial theorem: \((a + b)^n = \sum_{k=0}^n \binom{n}{k} a^{n-k} b^k\), note that \(\binom{n}{k}=\frac{n!}{k!(n-k)!}\) is the binomial selection of n from k.
  • Arithmetic sums solves as: \(S_n = \frac{n}{2}*(a+l)\), where a is first term, l is last term, and n is the number of terms
    • Sigma notation: \(S_n = \sum_{k=1}^{n} \left( a_1 + (k-1)d \right)\) – here d is the amount of de/in-crementation
  • Geometric sum solves as: \(S_n=a*\frac{1-(r)^n}{1-r}\), where \(r\) is the ratio of interest and \(a\) is the first time
    • Sigma notation: \(S = \sum_{n=0}^{24} a \cdot (r)^n\)
  • Permutation vs. Combination notation. \(r\) are the events that we care about, and \(n\) is the total number of possible events.
    • Permutations (order doesn’t matter) are just the factorial of the n! / (n-r)!: \(\frac{n!}{r!(n-r)!}\) or \(_r P_n\) or \(P(n,r)\)
    • Combinations (order DOES matter) are the factorial of \(\frac{n!}{r!(n-r)!}\) or \(_r C_n\) or \(C(n,r)\)