o3-mini

As an automations engineer, among other things, I’ve played around with o3-mini API this weekend, and I’ve had this weird realization: what’s even left to build? I mean, sure, companies have their task-specific flows with vector search, API calling, and prompt chaining to emulate human reasoning/actions—but with how good o3-mini is, and for how cheap, a lot of that just feels unnecessary now. You can throw a massive chunk of context at it with a clear success criterion, and it just gets it right. For example, take all those elaborate RAG systems with semantic search, metadata filtering, graph-based retrieval, etc. Apart from niche cases, do they even make sense anymore? Let’s say you have a knowledge base equivalent to 20,000 pages of text (~10M tokens). Someone asks a question that touches multiple concepts. The maximum effort you might need is extracting entities and running a parallel search… but even that’s probably overkill. If you just do a plain cosine similarity search, cut it down to 100,000 tokens, and feed that into o3-mini, it’ll almost certainly find and use what’s relevant. And as long as that’s true, you’re done—the model does the reasoning. Yeah, you could say that ~$0.10 per query is expensive, or that enterprises need full control over models. But we've all seen how fast prices drop and how open-source catches up. Betting on "it's too expensive" as a reason to avoid simpler approaches seems short-sighted at this point. I’m sure there are lots of situations where this rough picture doesn’t apply, but I suspect that for the majority of small-to-medium-sized companies, it absolutely does. And that makes me wonder is where does that leave tools like Langchain? If you have a model that just works with minimal glue code, why add extra complexity? Sure, some cases still need strict control etc, but for the vast majority of workflows, a single well-formed query to a strong model (with some tool-calling here and there) beats chaining a dozen weaker steps. This shift is super exciting, but also kind of unsettling. The role of a human in automation seems to be shifting from stitching together complex logic, to just conveying a task to a system that kind of just figures things out. Is it just me, or the Singularity is nigh? 😅