The basic premise of spread spectrum is that each bit of information is encoded as multiple chips.

First question. A “chip” is a group of bits? Why multiple chips?

Then continue:

The substitution of one bit for multiple chips of information means that the spreading factor has a direct influence on the duration of the LoRa packet.

Why?
I understand that the Spreading Factor is the number of bits encoded into each symbol (chip?).
And a symbol is an instantaneous change in frequency. So in an unique variation of frequency I associated not only one bit, but more (e.g in SF7 → 7 bits). I imagine this like a process in which we add more information, redundancy, to obtain a better receive sensitivity.

But why if Spreading Factor increase the Time on air increase too?

This YouTube link starts the video at 35’. But actually the whole webcast is nice, and he uses some radio terminology he explained in the first half hour. For the slides see LoRa Crash Course by Thomas Telkamp.

I have looked at the same question for our IoT balloon project. Found an explanation that 30 seconds daily sending time translates into 500 packets of 10 bytes a day at the normal SF7. An average of 3 minutes between messages.

For different message size, you can rescale but take into account the extra 14 byte overhead for each message.

As I understand, for different SF, each step up makes the message send time twice as long. Which means 6 minutes average between SF8 messages, 12 min for SF9, 24 min for SF10.

To ensure the balloon remains visible we like some stronger messages as well. I decided for a sequence of 8 normal SF7 messages and added one SF9 and one SF10. My guess is this adds up to 8x 3 min + 1x 12 min +1x 24 min, thus the sequence can be sent 24/7 with 5 mins interval between messages.

Does this make sense?
Does anyone know if the 30sec per day policy is enforced?

While true, somehow the Semtech formulas to calculate the airtime end up with less than the double time for the whole packet, as the total number of required symbols decreases. See Spreadsheet for LoRa airtime calculation. But it’s roughly true indeed.