<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Anshul Karn on Medium]]></title>
        <description><![CDATA[Stories by Anshul Karn on Medium]]></description>
        <link>https://medium.com/@anshulkarn?source=rss-6e2dda35e704------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/proxy/1*TGH72Nnw24QL3iV9IOm4VA.png</url>
            <title>Stories by Anshul Karn on Medium</title>
            <link>https://medium.com/@anshulkarn?source=rss-6e2dda35e704------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 21:05:08 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@anshulkarn/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Neural Space: Inside the making of a Digital Modeller]]></title>
            <link>https://medium.com/@anshulkarn/neural-space-inside-the-making-of-a-digital-modeller-fd402673b18e?source=rss-6e2dda35e704------2</link>
            <guid isPermaLink="false">https://medium.com/p/fd402673b18e</guid>
            <category><![CDATA[guitar-effects]]></category>
            <category><![CDATA[electric-guitar]]></category>
            <category><![CDATA[guitar-amps]]></category>
            <category><![CDATA[guitar]]></category>
            <category><![CDATA[guitar-pedals]]></category>
            <dc:creator><![CDATA[Anshul Karn]]></dc:creator>
            <pubDate>Fri, 20 Nov 2026 04:09:38 GMT</pubDate>
            <atom:updated>2026-11-20T04:09:38.991Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ePP5XsE8NgTqolztFMkxlw@2x.jpeg" /></figure><p>I thought I would take some time out to reflect upon my experience in building Neural Space and share it with fellow guitar players and digital modelling enthusiasts. When I was 10, I was gifted a toy guitar by my grandfather, and that’s where my journey started. In the first few years, I would listen to songs on the radio/cassette and try to figure out songs by ear. Years later, we had access to the internet, Guitar Pro tabs, forums, and YouTube videos, and that expanded my horizon into this world of guitar. I still remember the day I walked into a music store to buy my first electric guitar. They had only two guitars – a ’90s Samick Strat and an Ibanez RG. I went ahead with the Strat; it’s still my number 1 guitar to this day. Living in this part of the world also meant that access to good guitar gear and tube amps was limited; stores did not carry them, and if they did, they were out of reach for most of us. I got a Korg multi-effects unit from a college senior, and that’s where my journey into digital modellers started. I have to acknowledge that the reason why I’m able to make an app like Neural Space is because brands like Line 6, Fractal, Kemper, NeuralDSP, and others showed us the way of what good products can do. This learning continues to evolve with the impactful work done by Steven to create NAM, and the good people at Tone3000 to further improve accessibility and ease of adoption of this technology.</p><p>If you have worked with a digital modeller, you would perhaps relate to some of this experience – the latency of the product makes you feel uninspired to create anything new, and you end up scrolling through a list of amp models and tweaking them endlessly to make your sound work. Both of these factors, in my opinion, are key to what inspires you to play and create. We need the sound to be instant, like how we hear it when an amp is connected to the guitar in a room; we need to feel those dynamics. I hate to break this to you, but I do not think recreating that is possible. Nothing can recreate the experience of playing through a real amp. What is established, however, is that this little latency gap is the key for you to fully explore and appreciate that true amplifier experience. Is more processing power or better architecture required to close this gap? Aren’t our mobile devices today powerful enough to deliver that compute? Moreover, in the world of AI, why restrict yourself to a list of amps when there can be infinitely more possible amps if a model has truly learned the behaviour of amplifiers? These things made me wonder if there was a better experience that could be delivered. Up until now, I had experience working on Digital Public Infrastructure consulting projects, so this field of music tech was new to me. Of course, as a longtime musician, I knew what I wanted to hear, and I knew how the interface had to be built, but putting it into code meant this would be a challenging journey. At the same time, the guitar effects modelling industry seemed to be changing; with the introduction of NAM, open-source models have established a clear superiority in tone, performing on par with or even better than the gear we’ve been used to. Entering such a highly competitive and changing industry was a difficult decision. I spent the next year making Neural Space.</p><p>There are two broad approaches to building a modeller: you either build every part of the app yourself, or you piece together what is open-source (note that some implementations might require you to get a licence from the entity you source the code from). The core decision to be taken initially was whether to build our own tech for Digital Amplifiers or use NAM as a base for the amp modelling. The NAM A1 architecture had already proven its case; after creating and going through hundreds of files, I was convinced that NAM was the way to go. It felt right, it felt accurate, and more importantly, it sounded amazing. Then came the question of whether a mobile or tablet device would be able to keep up with the load and deliver within a sub-3ms latency target. That meant I first had to make a minimalist build to prove that this could even exist as a concept. You’d think that with all the AI assistance on offer, building this from concept to production would be a few hours of work, but reality is far from that. Decisions you have to make range from which AI code aggregator to use and which reasoning models will work best for your case (balancing quality output vs. cost), to continuously reminding your AI tool of the basics so that you do not drift away into making something else. Perhaps all these issues will eventually ease out with better reasoning models in the long term, but in my experience, the more detailed and clear you are about what you want to build – and the more clearly the model understands your requirements – the more accurate and fast the results will be. The real test, however, is in the output of sound, and this is where AI coding agents will disappoint you the most. The more complex your requirements get, the more context is lost by the model; only patience and perseverance get rewarded. When you build things that have not yet been made, the models start to steer in their own direction, making decisions based on guesswork at times rather than facts from the real world. In my case, the model decided that the GPUs of modern-day iPhones and iPads were capable of running NAM. Two months into development, after a lot of trial and error, I got to hear my first amp model. Words can’t express the joy I had, but it was short-lived since I discovered two major issues – I could not go below a latency of 5.33 ms (256 samples / 48 kHz); moreover, with every scroll in the app, the sound would vanish as the GPU could not keep up with the load. It meant that if the chunk of audio did not get processed under 5.33 ms, the audio was lost. The newer-generation iPads could handle the load at 128 samples (better latency); however, the GPU was still struggling to manage the app UI and audio at the same time. The AI model was convinced that if the GPU could not handle the load, there was no way the CPU could handle it. Spending months on development only to find out your core work might not be any good was disappointing. I knew NAM was built in a way that even lower-power chips could handle it, but in the tests I was doing, real-world results seemed to be very different. Perhaps this was one of my biggest learnings: AI models at best give you answers based on probability – trust them for the work they have seen, but their judgements cannot be taken at face value. If I heard that lower-power chips should be able to run NAM, I should trust that statement. I opened my Cursor tab, asked it to move the NAM implementation to the CPU, selected the best reasoning model in the hope that this would work, and voilà, it did. 32 samples, 0.67 ms. The guitar feedback felt instant. Progress at last. That was the day I finally felt this project could be a reality.</p><p>So why call it Neural Space? In my mind, if you are truly able to model a significant number of amplifiers, you could place them on a map by their similarity and spaces in-between these amplifiers would also exhibit amplifier behaviour. Imagine if these amplifiers, which look like planets, are placed on a map. The visual of the map you see here is what turned out after the analysis: the clean amps are placed in blue, mid-gain amps are placed in green, and high-gain amps are placed in gold. Now imagine the space in between those planets. Do those spaces also represent amplifier behaviour? Are they new amplifiers that do not exist in the real world, but only exist virtually? More importantly, what will those spaces sound like – their nearest neighbours, or entirely new hybrids? I had a lot of questions, and the only way to find out was to build and test it out. Let’s now understand more about the core amplifiers in Neural Space. Imagine a bunch of numbers in a sequence representing an amplifier; let’s say 20,000 numbers are enough to represent amplifier behaviour. Now, in the same format, another 20,000 numbers might represent another amplifier. Now, the question to solve was: what is the right approach to making this map? Do we reduce the dimensions to learn the most important features and then try to fill the space in between these planets? Or should we average out these numbers based on the distance from the nearest planets? Another option was to create a hypernetwork, a neural network of neural networks (our amp models). There could be many recipes for the map, but making 16 million grid points on a map sound good was going to be a challenging task. I must have gone through eight versions of the map, all coming up with a different way to fill that void; two approaches in particular made more sense to me. The Classic mode of the map lets you mix the audio in a traditional way guitar players are used to – the nearest amps contribute to making the amp in the empty spaces, and you can filter out the planets, which lets you create your own custom map/tones. The Neural mode populates the weights of the neural network as you click the empty space. Imagine moving from Amp A to B on a map: you also move the numbers that represent the amplifiers in the same way, and the weights of the network move in that same direction. I assumed that would retain the amplifier behaviour. Working on the weights and biases of the neural network can dramatically alter the tone; there are tight limits under which the output is musical. Now, in theory, it might seem that the natural output for spaces in between the map would be a mix of the nearest planets (amps). This was true, but only for planets that were placed very close together on the map. For amps that were far away from each other, you would notice a slight decrease in power amp character if you moved away from the amps. The results in some of these tests were promising. I felt good about the sound, though I felt the results were nothing special. You see, the reason planets sounded distinct was because they had a lot of character in them; the farther I moved away from them, that unique character we look for in an amp faded away. So the question was: where and how do we inject that character? I started experimenting with the weights of the amp models; one particular way of moving these weights led to creating a new parameter called Gravity. Turn it up and the amp suddenly felt more alive, connected, like a special gain button. The effect of Gravity is different on NAM A1 and A2 profiles since they differ in their structure. Gravity paired with any point on the Neural mode of the map felt good.</p><p>When you look at building a modeller from an independent developer’s perspective, you are constantly thinking about the ways in which you can differentiate the product. The truth is that NAM has made this industry a level playing field. The value addition is probably not in the core offering of amp modelling itself, but in the functionality built around it. The next few months went into making the effects for the app: noise gates, drives, overdrives, fuzz, User NAM + IR with Tone3000 integration, delays, reverb, modulation, and pitch blocks were added. Some particular features I wanted to add were the ability to record audio and video, play and record alongside backing tracks, automate patch changes, and include MIDI controller support. I also added a Match EQ effect – drop an isolated guitar track of the tone you want to replicate, record a few seconds of playing through it, and you should get a close-enough EQ match to the reference guitar track. Some of these learnings from Match EQ went into the creation of a Pickup Shaper block. You will be surprised how small EQ changes right at the start of your signal chain create a massive difference in tone in the later stages. The idea of the Pickup Shaper was to learn the wiring of different guitars – the pots, the pickups, and the pickup selector positions – and model them to shape the output of the pickups. I’m not claiming your Les Paul will sound like a Strat, but you’ll notice changes in clarity and tone for the better. In my experience, my old Samick Strat pickups ended up with more clarity, the sound was closer to a Strat, I could work with more amp models, and sometimes it became my default way of working.</p><p>I want to spend the next few months listening to your experience with the app. I’ve decided to release an early version which is still rough around the edges. For a while, it will be a free download, and I will continue to keep it free until I feel the build is stable. The App is scheduled to release on the App Store soon, I will update the link here once it’s available. For all updates, please follow our Instagram page- https://www.instagram.com/neural_space?dmpr=cHExZTFidzB3aTNr&amp;utm_source=qr</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fd402673b18e" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>