> The product of software engineering (or computer science) is clarity and understanding.

Hard disagree. Software engineering is about satisfying requirements while optimizing costs. Computer science is the part where clarity and understanding are the goal.

"A normal DM is one pi session by default." A single message, or a single message thread? Both are pretty extreme trade-offs to make, honestly. I spent a bunch of time looking at how the various Claws do this, and I settled on an idle timeout causing the next message to start a new session, but I also built /continue, /new, and a /resume that shows recent sessions with a telegram keyboard.

There is contenteditable, which is what all these (Wordgard, ProseMirror) are fundamentally built on. The rest is just the UI, and interop with systems that don't desire arbitrary HTML as input.

While these use contenteditable, it's not accurate to say they're just contenteditable plus some UI and interop.

None of these modern editors (Wordgard, ProseMirror, Lexical, Slate) use contenteditable for the document model. Rather, they have their own internal document model and use contenteditable as a kind of input layer where the editor monitors what the browser does, then translates that into actual edits.

Early editors like FCKEditor and TinyMCE were only wrappers around contenteditable. They used the DOM as the real document model, then intercepted certain keypresses and events and "fixed" the behavior when it wasn't correct (e.g. double enter inside a bullet list should switch to paragraph mode).

The result was rife with bugs and inconsistencies, and didn't allow for a proper split between the model and the view (e.g. to represent columns, video embeds, and so on).

I’m not a historian but I think FCKEditor and TinyMCE predated contenteditable. These were solutions wrapping textarea back in the day if my memory serves me.

My memory is fuzzy here, so I may be wrong. But I remember the generation of buggy WYSIWYG editors that preceded ProseMirror, and they were all contenteditable-based. They would typically start by attaching to a textarea, but then replace it with a div.

contenteditable I believe was in IE 5.5 in ~2000, I remember playing with it back then. It landed about the same time as "Outlook Web Access" which is arguably the first SPA, originator of AJAX, and first use of XMLHttpRequest. I think TinyMCE and FCKEditor were a little later.

FCKEditor and TinyMCE were using contenteditable.

> The rest is just the UI

"just contenteditable" is really understating things. contenteditable is a god awful API full of bugs and inconsistencies. Making something reliable on top of it is a very significant amount of work.

For context: my day job for last 2 years has been building a new browser engine from scratch, and I think the contenteditable-based WYSIWYG editor I wrote 15 years ago might have been a harder project (albeit I was lot of less experienced then).

"just the UI"

That's a hell of a "just", as anyone working on such projects will tell you. It's super super hard to get this right.

I'm trying to come to terms with this issue in my own interactions with open source as well. Where I'm at currently: since code is cheap and analysis is expensive, it can be more beneficial to a project's maintainer to get a precise, well-researched report of the issue than a PR. This is an inversion of most of my open source life, where opening an issue was asking for free work while giving a PR implied more reciprocal effort was given. I'll typically just end with "PR available upon request", unless the project has a no-LLM policy.

I mean, which lawyer caste do you respect? Is that one is cool with stealing credit cards to buy Claude subscriptions?

> 3. At an Italian airport: Constantly stealing bags, opening them to pick out MacBooks and credit cards, a credit card manufacturer-who sells stolen "black" credit card info to transfer stations— is racking his brains to save you money.

Stealing credit cards seems not directed related to using legitimately acquired LLM outputs for whatever legitimate purpose you want to use them.

Where is this quote from?

Midway through https://www.chinatalk.media/p/how-to-buy-cheap-claude-tokens... which was the article from the top-level comment.

I just started referring to those coworkers as Claude directly. Not like, at the start of the message, but after the first line? Nobody’s noticed yet.

If they find another success half as good as Claude Code it will all be worthwhile. (Monetarily good, not like, quality good)

How is CC a success when they still have to force subscribers to use it instead of it being a paid product?

Isn't the first section no-longer accurate for several years? I understood that, while we serialize the end of turn markers in a text format like `</think>`, internally they are a dedicated token that cannot be forged (a user message containing `</think>` would encode to a different sequence of tokens). Am I mistaken about this?

Obviously, this doesn't really affect the results of the paper, but it feels like it's the obvious first-line of defense: at least the model has a solid fence between the different roles.

Yeah, the footnote/sidenote on the paper (the one labeled #2) mentions this as well so you can't type that directly

It feels like sometimes researchers find something someone is already doing in the wild, undertake a study on it, but the speed of research and study doesn't match or cover the progress or rate of change by the time it's published, so with AI research specifically, too many studies can feel like they're in the past.

Can somebody help me understand the Quantization Analysis? It says "dynamic 4-bit UD-Q4_K_XL and dynamic 5-bit UD-Q5_K_XL are generally lossless" while showing a top-1% token agreement on the chart of 97.5%. Not what I would consider "generally lossless". Is this implying that some post-processing is going to account for the 2.5% loss? Beam search?

Generally 97.5% token agreement is very positive. Like the article explains, the difference isn’t the model thinking the capital of France isn’t Paris, but rather maybe saying “The capital of France is Paris” instead of “Paris is the capital of France”.

> Perhaps most people in the future won't bother. They'll just let agents do it all.

But can you imagine the look on some young teen’s face when they train their own GPT on their local computer for the first time?