I think many are missing the point here.
The argument doesn’t rely whatsoever on the idea that there is something beyond human perception that certain people are claiming to somehow hear. This isn’t the point, and makes this conversation different from those sorts of conversations.
The concept is not that a single stream of data is being improved on its own in isolation (although theoretically that could occur).
The concept is not about signal vs. noise floor.
The concept is not about hearing a perceived difference in a single stream of 32-bit data vs. 64-bit data.
The point is that a well-designed 64-bit audio engine is inherently capable of higher precision (and thus resultant accuracy) than a 32-bit audio engine. Thus, the results of mixing in the 64-bit audio engine have the inherent capability of sounding “better,” depending on the processing done and the taste of the person doing the mixing. The results can, in fact, be audibly different. It’s not a mystery.
The claim isn’t the brute idea that “64-bit is inherently better sounding than 32-bit.” That’s definitely not what I am saying. What I am saying is that the capability to perform accurate audio processing is greatly enhanced - as a mathematical fact - by a 64-bit engine vs. a 32-bit engine (assuming the engines are otherwise equivalent and well-designed).
The high-end light bulb example is flawed because what you’re trying to describe is a single output mechanism emanating in spectrum beyond human perception capabilities. That’s not what we’re talking about.
What we are talking about is the additive and multiplicative effect of numerous calculations in a chain, wherein the precision of calculation at each point in the chain has a direct impact on later calculations and overall quality. We’re not talking about whether a single element might be benefited by 32-bit vs. 64-bit (and hence the perceived difference might be nominal, if not null). We’re talking about the collection of a variety of complex calculations, some of which can amplify error numerous times (think about wide variances in gain staging, extreme saturation processing, etc.), and the result of combining numerous effected signals via mixing.
The light bulb example is flawed because the light bulb isn’t performing numerous light calculations, effecting and mixing perhaps hundreds of streams of light to form its output. It’s simply creating a single stream of output.
In a DAW, accuracy will benefit from 64-bit processing because, literally, greater precision is mathematically possible. Arguing against this is somewhat silly because it’s true. Whether or not you would prefer the results is up to you, and whether the results are audibly different depends on how the material is being processed. With 64-bit processing, more precision is possible, and thus higher resultant accuracy can be achieved than with 32-bit processing. Depending upon the nature of the processing you do with a 64-bit engine vs. 32-bit, you may or may not hear or notice differences. It depends on how you use the engine.
Think about a 2017 Ferrari. If you only drive it 30mph in town, will you notice many differences between it and a Honda Civic? Acceleration, handling, etc. sure, but at that speed, a car’s a car. But when you rev it up to 180mph, you’ll definitely be able to tell the difference between the Ferrari and the Civic. Pushed to limits, the Civic simply can’t keep up with a well-tuned Ferrari. In a DAW, the differences in accuracy in processing between 64-bit and 32-bit could be more readily heard if you “rev it up” in processing algorithms and scenarios where the 32-bit engine can’t “keep up” due to errors like the 64-bit engine can. When signal chains and the mix itself is highly effected in ways that multiply/amplify the errors, a difference can be more readily heard. It’s just math, and math is real. It has a real effect. I’m not saying you’ll always hear a clear difference between 64-bit and 32-bit processing; I’m saying the potential is inherently available in a 64-bit engine to produce higher-accuracy results.
Also, if you don’t think Philippe Goutier is a reputable source on this topic, you may be in trouble finding a reputable source.