For the past 20 years, semantic indexing sucked.

For the most part, the core offerings of search products in the last while is divided into two categories:

1. Full-text search things (i.e. every app in the face of the planet that stores text), which for the most part use something n-grammy like Okapi BM25 to do nice fuzzy string matching
2. Ranking/Recommendation things, who isn’t so much trying to search a database as they are trying to guess the user’s intent and recommend them things from it

And we lived in a pretty happy world in which, depending on the application, developers chose one or the other to build.

Been Kim

alignment problem involves “aligning” the representation spaces between machines of the world and that of the human. alternative perspective: teach humans new concepts to understand/communicate better

feature attribution doesn’t work

We take that perspective because many of the intersectional intepretability doesn’t work well (feature permutation, etc.)—feature attribution type analyses (“Impossibility Theorems Been Kim”) actually has no correlation with predictive results.

feature information store in models is unrelated to model edit success

i.e.: knowledge storing location located using ROME technique, though it gives you a sense of the location to store information, doens’t correlate to success of model editing.

We begin with a policy parameterized on anything you’d like with random seed weights. Then,

1. We sample a local set of parameters, one pertubation \(\pm \alpha\) per direction in the parameter vector (for instance, for a parameter in 4-space, up, down, left, right in latent space), and use those new parameters to seed a policy.
2. Check each policy for its utility via monte-carlo policy evaluation
3. If any of the adjacent points are better, we move there
4. If none of the adjacent points are better, we set \(\alpha = 0.5 \alpha\) (of the up/down/left/right) and try again

We continue until \(\alpha\) drops below some \(\epsilon\).

\begin{equation} \log a^{b} = b\log a \end{equation}

\begin{equation} \log (ab) = \log a + \log b \end{equation}

\begin{equation} \log (\frac{a}{b}) = \log a - \log b \end{equation}

TODO: connect Logan with a few fire departments