Rubber Soul - Giles Martin's 2026 Remix

I hate all AI remixes. I said that on Instagram, and I meant it. And then I said that this new Giles Martin remix of The Beatles Rubber Soul was my favorite version of the album. I meant that. And that’s including an original stereo and an original mono press.

And then half the comments told me, hmm, in not so many words that I had no idea what I was talking about. So let’s settle it. Now with a fight, but with some facts, what actually happened on this record? Step-by-step, including the computer science piece. Because that’s where I think the argument really lives.

The question is not just whether this mix sounds good. That’s going to be a matter of opinion, no matter how you slice it. I think it does, some other people don’t. The question is, when you take a performance like this, where much of it is on one track, and you separate that into four tracks, or more, what did you just do? Did you recover something?

Did you invent something? Or did you do something in between? By the end of this video, you will know exactly what the computer saw, what it learned, and what it guessed, and what a human did after that. And then you can decide whether saying, A.I. is in the mixed chain now is the right sentence. And I think it is.

Rubber Soul was recorded on four tracks, and a band that needed guitar, drums, vocals, bass, harmonies, piano, tambourine, all of the sounds that are in this awesome album. And they had to fit all of that on four tracks. So they did what’s called bouncing. They recorded a few tracks, they’d mix them together. That would free up space to record more on another track.

So you have Ringo’s drums and a guitar on one track, for example, bass and backing vocals on the same track, because they wanted the bass to be as loud as possible. They were chasing that motown low end. And the mono mix was the main mix. The Beatles and George Martin focused all their attention on the mono mix. The stereo mix in 1965 was basically just an afterthought.

Vocal shoved to one side, instrument shoved to the other side because that’s all the tape allowed. Here is the key point. If you balance those four tracks together using no other technology, you can balance it better. You can EQ them. You can make them louder or cleaner.

What you cannot do is unbounce them. You cannot take drums and guitar that were summed together back in 1965 and pull them apart just by turning knobs. And that is the whole problem. You don’t get from the 1965 stereo mix to the 2026 Giles Martin stereo mix without solving that problem first. Okay, so how does that process start?

First of all, a computer does not hear music. It sees numbers. Sound in the air is wave. It’s pressure going up and down. Digitizing that tape turns that sound into a sequence of numbers.

Sample tens of thousands of times per second. Pitch, timber, a voice versus a guitar. None of that lives in one sample. It lives in a pattern over time. A low note will repeat slowly.

A high pitch note will repeat quickly. A voice has a particular stack of harmonics. A guitar has a sharp attack and a different set of harmonics. So to see any of that, a computer has to look at a window. It takes a short slice of that number sequence, milliseconds long, and runs a piece of mathematics on it called a four-year transform.

That transform asks one simple question. In this short slice, how much energy is there at each frequency? Do that slice by slice by slice and you get a spectrogram. Time runs left to right. Pitch runs bottom to top.

That spectrogram is what the computer is actually splitting. A map of numbers for every tiny tile on that map. That’s a lot of jargon, but what matters is this. At any one tile, the computer still doesn’t know whose energy that is. If John is singing a note and John is playing a guitar, and they play in exactly the same frequency, their energy would land in that same tile.

And the number in that tile would be the sum. It doesn’t come with a label that says 60% John, 40% George. That label has to be estimated, which is why in my video, I used words like estimation or approximation. So that is the moment in time that we’re asking about. The ones in the zeros aren’t the voice and the guitar neatly separated.

They are a summed amplitude, turned into a frequency map, and the computer has to make the decision at that point in the map from tile to tile. So how does it make that decision? Well, the computer has to be taught first, right? And that’s where the learning part comes in. And that’s before the rubber sole tracks are even touched.

An AI model is shown examples of a guitar playing alone, a drum playing alone, a bass playing alone, John’s vocals, George’s vocals. And a lot of those examples will come from the tape itself, which is important. So that’s how they were able to get clean samples of each of the voices and each of the instruments in this mix. All right. So I mentioned a piece of software called M.A.L. And that’s what’s important here.

And this was developed by Peter Jackson’s team on the Get Back project. In order to make that get back documentary, they needed a way to separate the performance from the vocals or the talking and then training the model on new instrument types when it encountered them. So remember where the started? Peter Jackson working on Get Back. This was not built to remix rubber sole.

Peter Jackson’s team started to work on Get Back in 2017 working from the January 1969 sessions, which was about 60 hours of film and a hundred hours of audio. Much of it was in mono, which had voices, instruments, room noises, all on one track. And there was no way that anyone was going to be able to understand what was happening in the film if they couldn’t figure out a way to pull these sounds apart. So Peter Jackson’s team built Mal for this first project. And that’s where they first used it back in 2021.

And it was then when Giles Martin pointed to Mal and said, Hey, we could use this on the revolver remix. And now we’ve come all the way to the rubber sole remix. And the rest is history, but history actually has yet to be written as you’ll see here. I personally have spent 18 years in the software industry. So I do know the difference between a model being trained and software running that model.

And I’m still calling this AI software. A meal delay calls them stencils. A mask does not create sound. It keeps some tiles and mute others. A meal has said the network works more as a precise manipulator rather than a generator.

It does not generate something fundamentally new from the original audio. It manipulates what was already there. And this might be where some of the confusion came. And I think people took me to mean that there was generative, which I never used the word generative, but that I meant that somehow Giles Martin was rewriting rubber sole and that there was new sounds in there that AI was creating. And that’s not what I meant at all.

There is still generation happening though. And this is where people, I think, lose the plot a little bit because we started with one track and we had to get to four tracks. Or in this case, we started with four tracks that were compressed, that were bounced. And we got several tracks from those four tracks. It is not one track becoming exactly four stems.

It is one bounce track becoming a stem for each source that was in it. However many that was, then those stems go to a human. And this is another important aspect of it that people argued in the comments about is that yes, once this bounced track, and we separate those using the process that I just mentioned, the Mal AI process, you end up with these stems or tracks, and then it’s handed off to a human to actually remix the album. Okay, so that’s a lot feel free to get yourself a glass of water, pet yourself on the back, but we’re not done. And a lot of it comes down to multiple sources matching frequencies at a given point in time.

That’s where you’re gonna run into problems where you’re gonna have frequencies that a voice runs into a guitar, and you don’t know which is which at that point in time, right? And that’s why when I say the word approximate, that’s what I mean. The AI has to be smart enough to be able to know how do I separate those two things? And does it always get it right? And we know that it probably gets it right or now than it did, you know, in 2019, when Mal first started, or in 2021, when they first used it on get back, but it is not perfect, perfect.

And decisions have to be made. Traditional software could do crude separation. That’s why we have seen things like this in the past, but again, it was not a perfect science. And again, as I mentioned earlier, it had no prior knowledge of what it was searching for. It didn’t know how to resolve any ambiguities in the frequencies.

And it’s also why this isn’t just peeling seams away like string cheese. I mean, string cheese has those seams. A balanced mix does not. Those seams have to be inferred. And I think that’s part of where the confusion comes from.

Giles has said it himself about now and then, AI’s not creating John’s voice and about the process generally, he says there’s no additional material. Nothing’s being recreated or added to the record. And I do want to be clear about that. I’m not trying to mislead people here. I just want to try to get the facts out so that people can make their own choices.

There’s one big problem I have with this. And I’m having a hard time holding it back, but I will get to it later. So, and I use generative actually to be or generated or generating or generate in a more traditional sense, where places in this mix that are ambiguous had to be generated by AI. There is a test for the total and is the most honest test and audio engineering. It’s called a null test and Giles Martin talks about this.

You take the original four track master and you take the sum of all the separated stems, you flip the polarity and when you have two sounds that are the opposite, they will cancel each other out. So, you flip the polarity of one of them, peaks become valleys and then you play them together. If they are identical, they cancel the two silences. If they are not, whatever you hear is exactly the difference. So, that’s why it’s called a null test.

So, it proves no new energy was invented. What it cannot prove is that each individual stem is a perfect recovery because any allocation of the same total right or wrong passes the same test. But the point that I was making that I want to make now is that AI is in the chain at this point. AI has done its job, AI has approximated what this balanced track should sound like as four, five, six, eight tracks or whatever. What’s important to remember is that humans could not have remixed this record if it was not for the AI.

That is presumably becoming smarter all the time to remix these tracks. And so, another comment that at least one person maybe multiple people had mentioned is that then by that rationale, a DAC or a CD is AI. And so, I want to explain that part of this. A DAC, a digital to analog converter converts numbers back into voltage. It is deterministic, one input, one output, no decisions, no learning, no guessing.

If it works correctly, it gives you back the waveform you put in in analog form. It does not decide what is a voice and what is a guitar. On a CD, putting a record onto CD is the same kind of process in reverse, you digitize an analog waveform into numbers. You preserve the total, you do not split it. You do not decide anything about sources.

If you digitize a mixed track, you still have a mixed track just in digital form. A CD transfer doesn’t give you stems. It gives you the same mix just as numbers. Digital has been in the chain for decades. What is different is that a learned model is now making allocation decisions about the sound itself before a human mixes it.

All right, so to get back to where we started, I hate AI remixes on principle, but I played this record all the way through. I listened to Norwegian Wood for the 1,000th time and it sounds really nice to my ears. I think some of the Beatles albums including Rubber Soul and Revolver had stereo mix issues. I personally think they sound a little weird with the hard panning and everything. And there’s a lot of comments on that all over the internet with lots of people just like me that thought, man, these stereo mixes do not sound great and that’s why we live with the mono mixes for so long because mono just sounded better.

But if you want a stereo mix, I would check this out because it does not sound digital to me. Yes, it sounds clean. I mean, it sounds nice, but to me, it doesn’t sound necessarily sterile. And maybe this is something that’s happening to me over time because there was a time where I wouldn’t ever even give it something like this a second. Look, I had to have the original and that’s it because I was a diehard purist.

But call it what you will, and I won’t say I’m more open-minded or anything like that, but it’s just, I took this home and I listened to it and believe me, when I dropped the needle on this, I’m like, I’m gonna hate this. Go ahead and prove me right. And I liked it. I really liked it. I felt like it fixed a lot of the problems that the original stereo mix had.

I felt like it didn’t, I was listening for certain things like the sitar on Norwegian wood coming at the end. And I’m like, okay, if they don’t play this moment that wells me up every time I hear it, I’m gonna crap on this record and it was there. And it was there and it was clear and it sounded great in the mix and everything and the sound stage sounded good and the vocals sound good and I have one big problem with this record. I have massive problem with this. Nowhere on the outside packaging, which is nice.

It looks all original. Nowhere on here on the hype sticker anywhere does it say anything about AI in the chain? It doesn’t say anything about AI in the chain. It doesn’t even say anything about it being digital. It just has half speed mastered 180 gram orange vinyl.

New stereo mix by Jiles Martin and Sam O’Kale. I don’t like that. I don’t like that. I don’t like that. I mean, look what happened with MoFy when they weren’t releasing and being honest about the fact that there was digital in the chain, right?

And here we have a new Beatles, D mix, remix, Mal in the chain. And it’s not plainly being disseminated to the record buying public. If they’re gonna do this, they need to let people know that that’s what they did on these records and believe me, as good as this sounds in my opinion, using this technology in 2026, it’s only gonna get used more. They’re only gonna do more of these next thing you know or they’re gonna start fixing every Led Zeppelin album, every Pink Floyd album, every Journey album. I kind of wish they would do Journey.

So they’re not gonna reveal that there’s AI in the chain. That’s just wrong, man. That’s just wrong. This right here, you peel off this shrink wrap, no one’s ever gonna know. So thanks for watching.

Let me know what you think. Is it just an atrocity that they are not putting that this had Mal in the chain, AI in the chain, somewhere on this packaging, permanently, not just on the hype sticker, permanently somewhere on the jacket, preferably on the back or something. Or are you totally fine with it? Are you gonna buy this? Have you bought it?

What do you think of it? Do you like it? Do you love it? Anything you wanna talk about in the comment section, let me know. I will see you next time.