As a professional software engineer, it pains me to see the marketing hype of GenAI and the false promises (and hopes) that come with new technology being touted by non-subject matter experts. I’m a keen learner and willing to explore where the frontier actually is.
I’ve started experimenting with technologies like Cursor, Claude Code and ChatGPT but I feel uncomfortable conversing with machines owned by someone else. This led me to then dabble with Ollama, Aider and a host of models which has convinced me why developers to use cloud-based LLMs. The open-source versions (at least the ones I tested - maxed out at 33B params) just don’t cut it yet.
The main challenges with open-source models so far has been that the content generated doesn’t make any sense. Mind you, I did try and throw a very generic “Build me a SaaS application” task for a startup idea, in which it failed miserable. My next step is to experiment with larger models and tweak the contexting and prompting that I do. Perhaps breaking down the user request into bite-size chunks that the model can understand will yield better results.
The journey begins with a single step.