This is the culmination of a couple of thoughts I’ve had in my head the past few months. I think I’m not alone in some of these ideas. I just wanted to get these down to help articulate them, even if it’s only for myself.
Through the fire
We’ve seen a number of high severity security vulnerabilities of late. A lot of these have been at least helped along by some form of LLM, either in discovery or in exploitation steps. Mythos got the hype train rolling on this, and honestly it’s a great use of LLMs since so many of these vulnerabilities follow similar patterns. LLMs are great at finding patterns.
I think we can expect at least monthly to see something along these lines for the foreseeable future. I do think the steady stream will turn into a trickle though, and I think that overall much of open source will be significantly better off than we are now.
Unfortunately, I also think there will be plenty of casualties along the way. Good tools managed by talented teams that will see the flood coming their way and decide to shutter rather than soldier on. Hopefully I’m wrong about this but it’s easy to imagine how a side project can be easily overwhelmed, even if the bugs are legitimate. Add in all the slop and it’s a wonder anyone would keep it up.
Local or Die
I don’t typically think of myself as a zealot about much of anything, but a quick conversation with any normal person (even tech people) shows me that I have very strong biases towards local first AI. The thing is, even acknowledging this bias I think I’m still right. Data sovereignty has only been getting more important as the US sets fire to any international goodwill it’s managed to build. Add in the need for companies to maintain some semblance of proprietary data, and it’s a wonder anyone trusts an LLM with anything sensitive.
Additionally, local models (especially Gemma4 and Qwen3.5) have just gotten really good! I think that we’re at an interesting inflection point. We’re seeing more and more companies release smaller models, as they realize that running the frontier models for many tasks is often massively overkill. You see labs parsing your queries then often handing off to smaller models because it’s more cost effective.
Personally, I think a talented developer with even one of the smaller Gemma models is going to produce much better code than a middling developer with Mythos etc. and we’re only going to see this get refined further and further. With MoE becoming the norm, I expect the most important part of a model’s job will be just routing it to the correct expert. Eventually, I imagine we’ll be able to mix-and-match the different “experts” just like we do with models inside of coding harnesses today.
Did google just release an amazing python model, but Qwen still has the best c++ one? They will all work together seamlessly such that the “Mixture of Experts” is going to be more of a “Mixture of highly trained and specific submodels”.
What I think is really interesting about this is that it means that the current giants don’t have a future. At least not like all the projections have shown. Why would you spend a bunch of money on running stuff in the cloud if your highly specialized local setup can handle it fine. We’re already seeing companies leveraging Apple’s MLX, Ryzen AI, and now NVidia Spark to position their products to work well in such an environment.
Where Frontier Models Fit
The use case for these larger models, as I see it, is in orchestration. There is likely good money in this as in some hypothetical future these systems might be able to orchestrate what is today a development team. Got an idea and need an MVP to determine market fit? Throw a few tens of thousands at OpenAI/Anthropic and they’ll get you a passable first version. We’re not there yet but nothing has proven it’s off the table.
Additionally, long context tasks would fall into these massive data center style systems, but generally what most people use LLMs for just doesn’t fall into this category. Certainly Google, NASA, etc can find uses for them but most run of the mill users likely wouldn’t notice much difference. The average user or developer can almost always break their problem down into a chunk that a smaller system could handle.
Finally, I can’t see a future where these smaller good models don’t keep getting released. I imagine people would say that the big companies won’t want to cannibalize their frontier models but Google and Apple want to have models you can run locally on your devices. This lets them give performance and privacy that other companies just can’t offer with off device services.
Additionally, Chinese and European labs have consistently released open models. It’s an easy market differentiator for them. Also they may want to do it just to screw over the big American companies. As such, there is always going to be a place for open models. They might trail the frontier a decent bit, but again they’re just about good enough where that doesn’t really matter as much.
A Case For Optimism
So, personally I see us in 18 months with more secure software and more sovereignty over our own data and LLMs. I see LLMs as being legitimately useful for developers and security experts and being able to be used to make the world a better more secure place.
A Case For Pessimism
However, as others much smarter than me have pointed out repeatedly, just because these tools CAN be used in a productive and positive way doesn’t mean that’s how we as a society will wield them. We’ve seen time and time again that we’re all 10x developers now, and even if you commit to producing quality you can be easily drowned out by all the noise. I understand why people find it difficult to keep up hope in such and environment.