A secure hash function has the avalanche property that changing 1 bit in the input, has an independent 50% chance of flipping each bit in the output.
So you just need to ensure that you have enough entropy for your desired security level.
Anybody wanting to reconstruct what the camera is seeing would have to reconstruct it to the exact pixel values, including any noise. i.e. you would have to intercept the signal from the camera they are using, you couldn't just point another camera at the wall.
They do feed it as a one source, but it’s basically for fun.
I imagine if you (e.g.) XOR the state of a strong PRNG with the timestamp from an atomic clock whenever your Geiger counter goes off, you'd have infinite true random.
There are literally thousands of academic papers written on the topic of un-biasing and “extracting entropy” from noise sources. Most are overly complex; if you want a practical but secure approach see Yarrow, Fortuna, or the Linux kernel entropy pool.
Secondly, non-cryptographic PRNGs leak their state from their outputs.
This is a solved problem. Many academic and engineering hours have been put into real solutions. Dear reader, please use a recognized secure scheme instead of the amateur ad-hoc approach recommended up-thread. It has no theoretic or engineering basis. Don’t roll your own crypto.
https://blog.cloudflare.com/lavarand-in-production-the-nitty...
Which answers all your questions. Tl;dr: the lava feed is only an additional entropy source, it's not like they generate keys directly from the video feed.
It seems that they've pointed a camera at a wall of lava lamps, and the output video feed is fed into their randomness generator. I wonder how they account for the fact that: (A) Most of the video feed is constant, only the blobs in the lava lamps move (B) Lava lamps aren't actually all that random (C) The wall is visible from the street, and someone with nation-state-level surveillance technology could easily reconstruct what that camera is seeing.