Google is also seriously dropping the ball in terms of security. The CVE-2026-43499 root LPE (aka ghostlock) is still unpatched across all Pixel devices, on the latest """security""" update, despite weaponized exploits being public for months.

I don't see why that's weird

Plants tend to enjoy rain

> overused phrases and tics in your writing style

In the extreme these can be annoying, but more often I find them kinda charming, as an expression of the author's personality.

Simple example from this piece: I scrubbed 7 instances of the word “just” from it. Obviously, Emacs can do that on their own, but you have to know to look for that word, and every other one.

You are erasing yourself in doing this, not improving your writing.

If you want to improve your writing, read more.

No, you're wrong about this. One obvious reason why: I'd have gotten rid of those "justs" myself before models were available; it just would have taken me longer.

Agreed and depends on the writing. It's a very helpful tool though to be able to flag and highlight that sort of thing

> Present the original and new writing to the model and ask it which is better.

LLMs have absolutely terrible taste when it comes to writing. I don't find their feedback useful at all, beyond trivial spelling/grammar mistakes, which you don't really need an LLM for in the first place.

Proofreading is all you need.

Edit: I do sometimes ask an LLM for a fact-check, though.

Yeah, the best advice is: Don't. Write it yourself; make judgements; learn.

In general use today, AI primarily functions as a stupidity amplifier: it empowers dumb people to do dumb things more pervasively, with less effort, and at unprecedented scale.

We're moving into an era of artificially enhanced stupidity.

That’s what this is: it’s mechanical proofing.

As a "proof" of their poor judgement, take a paragraph you like. Ask the LLM to rewrite it to make it better (which I think we both agree will not make it better), and then in a fresh session ask it which it thinks is best. It'll almost always pick its own writing, even when it sucks.

Right, I agree: that’s exactly not what I’m recommending.

You are specifically recommending asking the model which of two versions is better (the quote in my top-level comment).

We both agree that they are poor "make it better" machines, but I also believe they are bad A/B testers and I'm using the former to demonstrate the latter.

You are writing both paragraphs. You’re specifically not asking a model to make a better paragraph. That would be a load-bearing debacle.

It doesn't matter who writes what, what matters is that LLMs have a preference for LLM-shaped writing. By A/B testing against an LLMs opinion, you are optimizing in the direction of LLM prose even if the LLM never writes any of the prose itself.

LLM style is not literally anticorrelated with quality. There are some things that they tend to do poorly, but you're not going to do those because you're writing the text yourself. If you have it judge your writing, and are careful to avoid the failure modes that the post goes into, it can be helpful by serving as a competent editor that doesn't share your blind spots.

I don't think it's a useful substitute for actual proofreading though, unless you're truly in a time crunch. When I proofread I don't just look for mistakes, I try to put myself into the position of a prospective audience member. LLMs seem to have absolutely no concept of "theory of the reader's mind". If I'm going for extra high effort, I have friends read it, usually asking them to identify anything that was unclear or hard to follow.

You could try out some version of this today, with a wiki. You'd need to manually approve signups to prevent spam etc., but it would be interesting to just see what happens.

Using a wiki for this is only one step above using stone tablets and messenger pigeons.

You'd want an enormous vector database at minimum. Text is just completely wrong for models at this scale, you must work in the latent space directly.

A global distributed network of vector databases. The web for agents but there's no text. Only vectors and knowledge graphs kind of like graph rag. Web 4.0 which is vector based and makes this arxiv paper come true and decentralized like the internet or the text based www?

Every existing text web server can voluntarily offer a vector version of their website and charge for it or inject advertisements onto their content so the ai labs don't have to do it all themselves as model pretraining off a dataset. The vector version can have many links to other websites in the knowledge graph. This would be decentralized so not a monopoly and everyone not just ai labs would contribute to ai development because the dataset would be open because it would come from the internet itself as it already is (except for synthetic or user data).

Humanity has built every capable institution, every scientific discovery, every technological innovation, without mind-reading. Quite the opposite: By putting thoughts from our latent spaces into words, we better separated the good arguments from the bad ones, placed trust where it was more deserved, and learned to stand on giant's shoulders so well we thought thoughts the giants never would have.

Anyway, I'm much more optimistic about the worlds with AIs that make the effort to properly explain themselves.

I'm convinced that piles of markdown and effective search (which may involve vectors) is all you need.

Maybe I'm missing out, but I've never used a single MCP, ever.

I recommend using spotify mcp server (if you use spotify), it freed me from the loop of listening to the same stuff again and again :P

that's interesting and helpful.. I have like 2k+ liked songs and Spotify always plays the same ones... Will try it out, thanks

I understand why they're doing this, but the anticausative title kinda rubs me the wrong way.

I'm not GP, but hammers do kill hundreds of people every year.

I suspect harness and usage modality is a big factor. Personally I disable all "memory" features wherever possible. I don't want the LLM to "get to know me", or any illusion thereof, which I assume helps with not anthropomorphizing it.

I use memory all the time so that I can ask it about things. That doesn't mean it "knows me" any more than my notes documents do, it's just more convenient to search.

> That doesn't mean it "knows me" any more than my notes documents do

But are you happy handing over your notes documents to mass aggregators of information?

In general, yes.

I'd be happy doing it 99.99% of the time myself, but I don't want to form the habit to think it's a good thing to do.

Yeah I don't think we disagree that much about this. But I'm not really worried about this particular habit forming risk. I can see where you're coming from though.

I do the same.

But frankly, I mostly use (greenfield, non-remembering) LLMs directly to counteract the enshittification of search brought on by attempts to use AI to remember things about me (and especially to sell me things).

This works well for me because (a) I'm usually using my desktop, and (b) I'm a touch-typist. I want to say that my search results have gotten back to where they were a decade or more ago by doing this, even though I often have to give a couple of prompts in order to get the LLM to provide me a useful link.

I guess I'm also pretty skeptical of the "automatically remember things about me" functionality. But I really like being able to say "hey make sure to remember this fact" and being able to ask about those things later.

For me the convenience of “hey remember this fact” is outweighed by a desire not to get stuck in a search or context bubble.

It might be nice to have better UI to control which bits of history get added to the context of a chat, but then just use a coding harness instead of web UI

Maybe. But I don't just use it for coding. For instance one thing I use it for is helping me track and evolve my workouts over time. It helps that it can remember the weight and number of reps I did the previous time. I could make an app for this, but it works well to do it purely within chat, with memory, and that's convenient. This is just an example, I use it this way for a number of things. I'm not going to use a coding harness for this kind of trivial use case!