Everything went smoothly today.
Nothing to change to solve part 2 because I already used memoization for part 1 (it looked like an AoC exercise where it would be useful ).

Here are the main parts :

  defp blink([_], 0), do: 1

  defp blink([0], steps), do: blink([1], steps - 1)

  defp blink([h], steps) do
    Performance.memoize({h, steps}, fn ->
      case should_split?(h) do
        {true, split_params} -> split_params |> split() |> blink(steps - 1)
        _ -> blink([h * 2024], steps - 1)
      end
    end)
  end

  defp blink([h | tail], steps) do
    blink([h], steps) + blink(tail, steps)
  end

Part 2 takes ~100ms ~70ms to run.
A virtual part 3 with 1000 blinks instead of 75 takes “only” ~5475ms to run.