Last night, a client added a request—and I almost rebuilt the whole workflow
At 10 p.m., a three-person design studio in Shanghai asked, “Can the AI stop repeating itself when it quotes?” I stared at the old prompt until midnight, feeling that sinking certainty: I’d have to rebuild it all over again. I’m just a small project-based team lead, not someone who researches models.
This kind of headache is familiar: clients don’t necessarily need the AI to be smarter; they need it to follow instructions better. The old answer was to buy new data and retrain the model. This time, MIT’s research offered another route: a stronger model directly modifies key parameters inside a weaker one. In the study, the smaller model changed only a tiny number of parameters, and its math score jumped from 7.9% to 48.1%. Still, when they tested a different model that didn’t have the repetition problem, the improvement was small. This is not a universal fix.
Who’s using it? The MIT CSAIL research team validated the method on open-source models. Not every founder will need this, but people building customer support, content generation, or quoting tools may eventually be able to adapt an existing model to their business without starting over each time.
If I wanted to replicate it today, here’s what it would cost
If I wanted to try it myself, a public model could cost anywhere from $0 to tens of dollars. Preparing the research experiment might take one to three days. The technical barrier is basic command-line and model configuration knowledge; this isn’t a one-click task on an ordinary computer. My first step would be to open MIT’s paper project page and look for the code, data, and run instructions. The paper doesn’t include a ready-made button, and the method may not deliver the same capabilities as commercial models. I also don’t think I need to try it now. Organizing the requirement and the old prompt can often be the smarter move.
At each stage, this is how I’d choose
If I were just starting, I’d stay away from modifying models. I’d use an existing AI tool plus a checklist, because reliable delivery comes first. With one or two clients, I’d log the answers that repeatedly go wrong, then test whether fixing the model actually saves more time than adjusting the working prompt. If I were scaling up, I’d run a small comparison of accuracy, cost, and failure rates before and after modification. This research is a reference for a future capability, not a system I need to ship today.