How come there's so little chat about *that*?.. the more I think about the scenario of devastating AI-augmented/driven wars or runaway military AI, the more my subjective P(doom) tends to 1.0. Am I missing some counter-factor? 4/4

Furthermore, the current AI "leadership" doesn't seem to worry about these applications and, if anything, (in)directly encourages them through increased discussions of geopolitical competition and sovereignty 3/

War is ugly and, as I understand, every possible technical solution gets explored. War and oppression would benefit the most from centralized coordination and autonomous decision making. AI will be trained to hunt and kill people and do it with less and less oversight. 2/

exactly, “gen AI” just means “bad AI” in order to avoid the charge of being against *all* AI. I am also against bad AI. by defintion, everyone is against bad AI!

Personally, I am *very* excited about this development, as it paves the way for a new generation of scalable, robust, and easy-to-implement coarse-grained models of biological systems and also helps our lab enter this field.
Please, let us know if you have any qs/comments/suggestions/future ideas!

VBC is available as an open-source HOOMD-blue plugin: github.com/glab-vbc/vbc....
Vladimir Dmitriev led this project almost single-handedly, and Ankit Gupta developed the HOOMD-BLUE plugin. 11/

There is also a second use of the same correction: it prevents Fisher–Ruelle collapse in attractive soft-core potentials, so one can keep a soft repulsive core and a large MD timestep even at strong attraction. 10/

The model reproduces the known reentrant RNA aggregation and its dependence on local pairing constraints. The full scan over interaction strength and the two pairing rules takes ~150 GPU-hours on a single GPU. 9/

We believe this regime may be useful for molecules with flexible or surface-mobile binding sites. We tested it on repeat-expanded RNA molecules, where intra- and intermolecular use of the same valency-one stickers compete with each other. 8/

The lack of fixed bond directions also changes assembly, in agreement with previous predictions. At the same bond saturation, VBC particles form small cyclic clusters much more readily than patchy particles, and large aggregates appear only at higher saturation. 7/

The counting geometry can also be tuned to lower the barrier for partner exchange, keeping bond swapping fast even at high bond energy/saturation. 6/

VBC also reaches bonded states much faster in simulation time. Because there is no orientational search for patches, bonds form essentially on contact, and our test system reaches saturation almost an order of magnitude faster than the patchy reference. 5/

The coordination number saturates at the prescribed valency for f=1–6. Computational cost is almost independent of valency; on the same GPU, the advantage over our patchy-particle reference reaches almost 10x at f=6. 4/

VBC takes a different route. Building on existing density/neighbour-counting potentials, it uses a smooth local neighbor count to suppress attraction as the target valency on either interacting particle is reached. No angular terms or explicit triplet enumeration, and it works with soft cores. 3/

Limited valency is common in soft and living matter, from proteins and nucleic acids to multivalent molecules in condensates. Patchy particles are the standard way to model it, but in MD they are relatively expensive, can be finicky, and require a fixed binding-site geometry. 2/