It's been a moment since I've written, and a lot has been happening in the first bit of 2025. Unfortunately, it's been a busy time for me and I haven't had the opportunity to let people know what's going on. There are a lot of very cool things in the future for NAM, and I'm excited to have the opportunity to support this bright future. Clearly, this also comes with responsibilities, and I'd like to take a moment to share some thoughts on how I view my role in stewarding NAM's role in our future. If you have feedback for me, please get in touch.
Opportunities and risks
NAM is open-source, which essentially means that anyone is free to build whatever they'd like with it. It's a really interesting position for tech that's had such an impact, and I want to take folks through my thinking on this.
On one hand, the possibilities for innovation are remarkable. There are so many people building with NAM right now, and I've heard it be referred to with increasing frequency as "an industry standard" (e.g.). It's a sincere honor to enable so many people to make cool things, and continuing to support this is the highest priority for the open-source project.
The other side of NAM being open-source is that builders are taking on a risk by building on top of NAM: If the open-source code changes in a way that constrains builders' ability to use it, then they might find that their project becomes infeasible shortly after (or even before) it launches.
Standardized models
NAM's open-source trainer defines several "standardized" architectures that users can easily use when training "snapshot" models. These suggestions are powerful: The "standard WaveNet" has become the architecture used by the majority of shared models (e.g. on ToneHunt, which is frequently-used to share models.) Many people see this as a resource they can integrate with to offer musicians very exciting products. If I were to update the "standard WaveNet" to something that requires more DSP to run--or even is just a slightly different "shape"--then this risks products being rendered obsolete that were depending on supporting the "standard" architecture.
I realize that this risk exists, and I realize that it threatens my highest priority stated above for NAM. Fortunately, the "standard" training recipe is already state of the art with regard to several metrics--it is the most accurate neural amp modeler (lower-case) available in terms of quantitative metrics, and it is also the quickest to train (using free resources like TONEZONE3000, one can have a state-of-the-art model trained in about 2 minutes). The recording file v3_0_0.wav is also about 3 minutes long, which (in my subjective opinion) trades off nicely between the amount of time spent recording and the end results (again, state of the art).
It's possible to trade off one some of these metrics to optimize for others. For example, you can make the model bigger or train on more data to increase accuracy. However, these come at the cost of DSP and studio time spent recording, respectively. People's exact preference on these trade-offs will depend on their personal preferences as well as their role--if you're buying a model, then you might care less about how long the model-maker spent recording.
NAM's model-making probably isn't optimal as it is currently. However, there's a responsibility for the open-source code to navigate many considerations as it moves forward--not just technical performance metrics, but its impact on the ecosystem it's the foundation of. Therefore, given the growing maturity of NAM's ecosystem, expect me to be cautious with how I advance the open-source. I do not currently intend to replace the "standard WaveNet" with something else.
Conclusion
NAM has opportunities and risks to consider as it increasingly becomes a standard. I'd like to support the opportunities and mitigate the risks. I hope the above approach achieves that, but I also realize that there may be something I haven't considered. Again, if you have feedback, reach out to me; let's talk so that I can understand your goals and how best to support them.