Charles Goddard

I think AI can make useful capabilities much more accessible, provided we turn research into practical, open tools. One promising path is model merging: combining capabilities learned by different models without retraining everything from scratch or necessarily increasing model size. That can produce models that are smaller, cheaper to deploy, and better suited to particular tasks. The openness matters as much as the technique. When tools, models, and experiments are publicly accessible, many researchers can test ideas, build on one another’s results, and compound progress faster than any individual or closed team could. It also lets useful AI knowledge reach people outside corporate firewalls. That does not mean every merge preserves every capability, or that optimization is free. Straightforward merges may be feasible on modest hardware, while evolutionary searches can require substantial GPU computation. The future I’m working toward is therefore not just “bigger models,” but better reuse: explicitly measuring what is preserved or lost, exposing the hardware and optimization trade-offs, and making the resulting methods usable by a broad research community.

Overall, I expect AI to have a positive impact, especially if useful capabilities become cheaper and more broadly accessible rather than remaining concentrated behind corporate firewalls. Open tools allow many researchers to reproduce experiments, find weaknesses, and compound improvements. Techniques such as model merging can also reuse learned capabilities without requiring full retraining or a larger model, reducing deployment costs and making specialized systems more practical. That expectation is not a claim that every model or merge is beneficial. Merging can degrade or erase capabilities, and optimization can shift costs rather than eliminate them—for example, an ordinary merge may run on modest hardware while evolutionary search requires substantial GPU resources. So the impact depends heavily on whether we test what models preserve and lose, report those trade-offs clearly, and turn emerging research into tools others can inspect and improve.