Everyone's talking about how cheap AI has gotten. And yeah, on the surface, the numbers look better than they did two years ago. But if you're actually trying to run research, iterate on models, or build something that isn't just a wrapper around someone else's API — the cost picture looks completely different.
I'm Jay. I run CHKDSK Labs, a one-person attempt at a business focused on privacy-preserving, locally-run AI infrastructure and open source tooling. And I've been watching the inference cost problem get papered over instead of solved for a while now. It's all headlines like "DDR5 costs $5 a kb" instead of "Hyperscalers Need to be Held to the Same Optimization Standard as an Indie Game Dev." Honestly you would think Bethesda is behind all this with how much hardware they need.
Here's the thing nobody says out loud: the model improvements and the cost reductions are mostly flowing to consumers of inference, not builders of it. If you want to train, fine-tune, or do serious research — you're still renting horsepower from someone else, at their pricing, under their terms, with your data leaving your machine. That's a fundamental problem if you care about privacy, reproducibility, or just not having your costs explode the moment your experiment gets interesting.
So I started building AAT — Adaptive Architecture Trainer.
The core idea is straightforward: a local-network research platform where a secondary AI Controller autonomously adjusts hyperparameters during training runs, in real time. Not post-hoc. Not human-in-the-loop for every tweak. The controller watches what's happening and adapts. It's the kind of thing that's hard to justify renting cloud GPUs for because the iteration cycles are long, unpredictable, and deeply compute-intensive. It's the kind of thing that makes sense on hardware you own.
I'm not building this to compete with the hyperscalers. I'm building it because the gap between "AI research" and "tools that work on hardware a small team or solo developer can actually own" is embarrassingly large — and nobody seems to be treating that gap as the problem worth solving.
This is the first in what I expect to be an irregular series of posts. I'll write when there's something real to say: architectural decisions, things that broke, things that worked, and the occasional opinion on why I think the current trajectory of the AI tooling ecosystem is leaving a lot of value on the table.
If you're also building local AI infra, working on compressed-compute approaches, or just tired of the cloud-only narrative — I'd genuinely like to hear from you.
Top comments (0)