fork creates a second process that is an exact clone from the first.

The original process is called the parent, the child process is called the child. The child comes in at the next instruction after fork. This means that fork calls once, returns twice. After fork, the execution order between both processes is completely up to the OS. After fork, we cannot assume execution order.

Fork’s return value is different between parent and child:

Ingredients:

• \(\mathcal{P}\) problem (states, transitions, etc.)
• \(d\) depth (how many next states to look into)—more is more accurate but slower
• \(U\) value function estimate at depth \(d\)

We essentially roll forward into all possible next states up to depth \(d\), and tabulate our value function.

Define subroutine forward_search(depth_remaining, value_function_estimate_at_d, state).

1. if depth_remaining=0; return (action=None, utility=value_function_estimate_at_d(state))
2. otherwise,
   1. let best = (action = None, utility = -infinity)
   2. for each possible action at our state
      1. get an action-value for our current state where the utility of each next state is the utility given by forward_search(depth_remaining-1, value_function_estimate_at_d, next_state)
      2. if the action-value is higher than what we have, then we set best=(a, action-value)
   3. return best

What this essentially does is to Dijkstra an optimal path towards the highest final utility \(U(s)\) om your current state, by trying all states.

The Forw

foundational model is a large model which can be generalized over multiple input signal types.

Problem: end-to-end analysis of biological interactions at all timescales is hard; womp womp. No relationship explicitly between sequence, crystallography, md, etc. Also, some of them have time, some of them are frozen, etc.

Solution: use ML to glue multiple scales’ analysis together, using ML to

proteins can be encoded as hierarchies

1. protein functional behavior
2. secondary structure/primary structure
3. amino acids
4. sequences!

Slicing through the embedding space of GenSLMs can be used to identify these larger scale things from just the sequence by looking at the “general area” it exists in the latest space.