azalemeth • today at 11:23 AM

Honestly for a "side project" Opus has been fantastic for me writing a hybrid simulation framework that prior to large scale code generation would have been a matter of years (and writing a grant, assembling a team, etc – in order to do it "properly"). I've had a bit of help with a grad student and I hacking together on a project that is basically "please merge the following GPL codebases and different areas of physics into one coherent environment". I've given Opus validated codes in disparate languages (julia, python, C) and asked for aspects of various algorithms as an extension module to a large chunk of C and C++ code that is a monte carlo simulator that has been around since 2004.

A bit more context if you care: it's a meso-scale, physiological simulation environment of "particles" that carry nuclear spin, can move in 3D space, and (should they interact with each other or their environment) undergo chemical kinetics. The idea is to simulate molecules within e.g. organs or blood vessels within a person in an MRI scanner, with the motion of the particles dominated by the Navier Stokes equations, but here solved in a Lagrangian (rather than Eulerian) framework by smoothed particle hydrodynamics.

The fact that particles carry nuclear spin means that we can solve the (semiclassical) Bloch equations and by using a python plugin module import exactly the physical MRI scanner would do (in pulseq format) and be able to predict what signal the machine would record – e.g. there's a whole world of cardiac or neurological flow imaging work done in the context of nasty diseases like stroke or myocardial infarction – which has a bunch of physical artefacts behind it. I'm trying to make a simulation framework that can take in realistic patient geometries and act as a 'data generating process' because if we do it right the various physical artefacts that the machine records are reproduced, surprisingly accurately. Of course you also know the ground truth of where the particles are. I'm specifically interested in a weird technique (which I did my PhD in and you can read an article all about here: [0]) called dynamic nuclear polarisation, where specific spin states of molecules such as [1-13C]pyruvate are injected essentially out of thermodynamic equilibrium and act as short-lived tracers of metabolism – again highly altered in disease. The signal we record is a strong function of the physics of what you told the machine to do, the spatial constraints and environment of the patient's body, and the chemical kinetics of the patients' biochemistry (the latter two are usually what we're interested in).

Getting them to do chemistry as well as act as a "simple" tracer is more involved, because in the Lagrangian framework the number of particles is ≈ the spatial resolution of your simulation. That's fine if you're simulating water, but if you're simulating something that reacts concentration is not scale invariant (if you want to keep the interpretability of the rate constants). I've worked out an analytic set of scaling rules around this and fortunately for my application environments and length scales "it just works", completely by luck.

I've used Claude to port various SPH algorithms and boundary condition handling ideas (which are absolutely critical and highly not obvious – we have leaky walls in some places, and e.g. LCR / circuit theory models of the microcirculation to plug in) and it's been a godsend. But I'm running into its limitations constantly. It both confidently makes shit up, claims it is mathematically justified and when the resulting simulation explodes says "I apologise; I lied above" (!) or "I apologise; I am wrong" and I periodically have to yell at it to try to do something more productive.

The real hope is that this simulation environment would be both generally useful for basically anyone doing flow MRI, and help our basic scientific understanding of what we're measuring (the technique is in many hospitals!) but also be able to produce meaningful synthetic training data for image reconstruction algorithms later on. It'll end up permissively licensed (all of the "starting" codebases have compatible OSS licenses, and we're releasing our contributions similarly).

I really hoped that Fable would be better at this sort of work. Occasionally, relating to my work DNP [1], I have need to talk about proper nuclear physics and I have seen Opus's chat interface write a wall of text (e.g. talking about photonuclear reactions and cross section differences in millibarn) and then just delete it all. Support have told me that yes, I've hit the nuclear filter and, well, tough shit, basically.

I wrote a version of the above to them yesterday, and just got the most boilerplate response that I've yet to test:

    Thanks for reaching out to Anthropic Support.
   
       We're sorry to hear of the issue that you're running into with accessing Fable 5. I'm happy to say the issue has now been resolved and you should be able to access the model within Claude.

    I'll close this case out for now, but please feel free to reach back out to us here if you have any follow up questions or concerns or if you're still in need of assistance. We'll be happy to help.

which doesn't fill me with hope...

[0] https://physicsworld.com/a/dynamic-nuclear-polarization-how-... [an "accessible" article] [1] https://www.science.org/doi/pdf/10.1126/sciadv.adz4334