The token efficiency improvement might be underrated. If the model solves tasks with fewer tokens, that directly translates into lower cost and faster responses for anyone building on the API.
If you’re installing this on a fresh machine, the network installer is usually the smoother option. The full ISO is great if you’re setting up multiple systems or need an offline install, but for most people the net install saves some headaches.
I think a lot of engineers intellectually agree with this idea, but emotionally still default to building the “proper” system.
There’s a strange pressure in tech to reach for architecture, frameworks, and infrastructure even when the problem might only need something scrappy. Sometimes the ugly solution survives longer simply because it’s closer to the actual problem.
This feels like a recurring pattern in the stack. abstraction removes visibility faster than tooling replaces it.
Encryption and higher-level platforms are great for security and productivity, but the debugging surface keeps shrinking. Eventually when something breaks, nobody actually has the layer-by-layer visibility needed to reason about it.
But I’m not sure it’s entirely inaccessible to models either. If you feed them enough signals,logs, incidents, metrics, past debugging threads they might approximate that feedback indirectly. Not the same as being paged at 3am, but maybe closer than we assume.
but your distinction is really good. The feedback loop is probably the key difference.
I think this is slightly romanticizing the idea that humans “hold the territory” in their heads.
In most real systems no single engineer actually understands the full territory either. People rely on partial mental models, docs, logs, and tribal knowledge. In that sense, LLMs operating on maps might not be that different from how teams already work.
ECH is great from a privacy perspective, but I’m curious how well this will actually work in practice.every time the web encrypts more metadata there’s pushback from middleboxes and network operators.
This is a fun idea. What surprised me is the inversion where MUL ends up faster than ADD because the neural LUT removes sequential dependency while the adder still needs prefix stages.