I agree...except, I would've appreciated some self-awareness from the author. They represent a minority of the users, yet they fail to comprehend this simple fact.
For the large majority of users, phones are THE primary (if not only) device for their interaction with the internet. You can complain "them-lazy-brainrotting-GenZs" but some people don't have a choice. There are plenty of countries where a smartphone is the cheapest internet enabled device that a person can afford.
Secondly, the UX for "web browsing" on the phone is strictly worse compared to (well-made) apps. In fact, apps are the reason for the explosion in popularity of smartphones. And also the reason only android and iphone have survived the OS race (see windows phone and linux phones). So — much to my own disappointment — it does make sense for companies to treat mobile (app) users as first class citizens. You need to understand that you are not the target user anymore. And, yes it sucks.
That said, it does not justify the gradual enshittification, dark patterns and dopamine hacking that have been normalized in modern apps.
Given how the blog is presented, I assumed this was something novel that solved a unique problem, maybe a local multi-modal assistant for your device.
I installed it and it's none of that. It is a mere wrapper around small local LLM models. And, it's not even multi-modal! Anyone could've one-shotted this in Claude in an hour (I'm not exaggerating).
What's the target audience here? Your average person doesn't care about the privacy value proposition (at least not by severely sacrificing chat model's quality). And users who do want that control can already install LMStudio/Llama.cpp (which is dead simple to setup).
The actual release product should've been what's described in "What's next" section.
> Instead of general chat, we shape Ensu to have a more specialized interface, say like a single, never-ending note you keep writing on, while the LLM offers suggestions, critiques, reminders, context, alternatives, viewpoints, quotes. A second brain, if you will.
> A more utilitarian take, say like an Android Launcher, where the LLM is an implementation detail behind an existing interaction that people are already used to.
> Your agent, running on your phone. No setup, no management, no manual backups. An LLM that grows with you, remembers you, your choices, manages your tasks, and has long-term memory and personality.
I don't think it should be conflated with auto generated AI slop. I see a lot of snippets which were clearly manually written. I'm assuming the author used AI in a supervised manner, to smooth out the writing process and improve coherency.
> It’s doubly hard to distinguish valid statements from invalid ones when you don’t have any examples labeled as invalid. But even with labels, some errors are inevitable. To see why, consider a simpler analogy. In image recognition, if millions of cat and dog photos are labeled as “cat” or “dog,” algorithms can learn to classify them reliably. But imagine instead labeling each pet photo by the pet’s birthday. Since birthdays are essentially random, this task would always produce errors, no matter how advanced the algorithm.
> The same principle applies in pretraining. Spelling and parentheses follow consistent patterns, so errors there disappear with scale. But arbitrary low-frequency facts, like a pet’s birthday, cannot be predicted from patterns alone and hence lead to hallucinations. Our analysis explains which kinds of hallucinations should arise from next-word prediction. Ideally, further stages after pretraining should remove them, but this is not fully successful for reasons described in the previous section.