Yeah I did something similar but more extensive, because not every model is same plus depend on your options how much you want to degrade model from fp8. Some models is just slow on local hardware and not practical really.
256k seems sweet spot. Even large codebase can be handled you just has to map out areas for agent. Anything past 256k usually exception than norm to be.
At least this my discovery when I did evals on my reviewers skills. 10k LOC is high end for high effort or around 150k context using opus. You can push towards 15k but that’s on max effort with about 400-500k ctx.
Fast models like haiku do not even need 1M ctx in my view
Yeah codebase structure is very important. Admittedly, by large I mean average size but very poorly structured with a spaghetti (and sometimes cyclical) import graph.
That's true of WebP and AVIF as well. I think all modern image codecs have both lossy and lossless modes - the separation between JPEG for lossy and PNG for lossless seems to be a historical oddity.
You can’t really do that with PNG today either, because of all the PNG optimizers that quantize images before encoding them to PNG for better compression.
By that definition, all formats are lossy, because any preprocessing could be lossy, and the word loses all meaning … which is why we don't use that definition.
By what definition? I didn’t say anything was lossy; I said you couldn’t be sure that the image was lossless just from the extension. Of course, it might still be, and you might have other reasons to know it to be.
Arguably the same is true of JPEG. The format itself can record all the frequencies in full detail (with some tiny rounding errors I'm ignoring here), but the optimizer creates a different image by cutting elements that are less important and wasteful of bits. And then that different image is encoded losslessly.
This is indeed frustrating, though realistically not that big of a deal for most users, especially if you consider lossless as just highest quality. Though personally I think it makes sense to have different extensions by convention for lossless and animated images.
JPEG artifacts are literally a meme, normal people are aware of this stuff.
I hold the extremist view that had we decided to use .png for lossless webps, it would have been an overall net gain - in people's minds "png" maps roughly to "whoever last touched this image didn't do anything evil to it" - very few people care how the data is actually encoded.
I mean we are already dealing with that nowadays. Many PNGs are either JPEG in a different package or straight up lossy compressed. Unfotunately extensions have never been safe for judge
I've been wishing for years that all of these codecs with hybrid lossless/lossy modes would just add a single l to their file extensions, purely for the annoyance of it.
Like, you can do this individually, but it's not the same thing as actually having it in the standard.
See I would think lossless is the one you have to specify. (Also, it stops being "large" when you extend this gripe to similar formats like AVIF and WebP.)
But people are definitely going to assume "eXtra Large". Lots of people. I guess a few will think "excel" (as in JPEG XL comes out ahead and is superior).
The problem with this is that fm is just a refinement of the md format so anything that recognizes md already would also recognize .fm.md but they would NOT recognize .md.fm
If we're talking about the 3 dashes surrounding a yaml doc at the top of the markdown file, I don't think that's a refinement at all? It breaks the markdown spec as far as I'm aware.
When writing GPU shaders it's not uncommon to use extensions like ".vert.glsl" = "vertex shader written in GLSL", ".frag.hlsl" = "fragment shader written in HLSL", etc.
You cannot tell it just from the extension either. If you like extensions, you can use a '.lossless.jxl'. And in general for most formats, you can use metadata to store that info (also doesn't guarantee anything)
This is something that can be addressed by file managers - showing a metadata entry on whether it's lossy or lossless. File extension (along with the bifurcated mainstream adoption of JPEG and PNG) has been a great proxy to do this, but like others noted, can be deceptive, which is arguably worse.
Then, just do not add text that adds no value. Less is more.
Documentation is first part people see when presented. I personally hate AI written text even though I do agentic engineering a lot as well. Any user facing documentation must be reviewed, if you snoozy by reading it - remove or rewrite.
I do not care about paragraph long comments in code, because I’m not gonna read them, but I still follow my internal guideline and cut all this verbosity to level where it readable using hooks and skills.
So basicly like you said if you care about product do not add placeholders. It’s take away first impression.
I surprised how many people don’t even read their own PR commits and diffs, like at basic level.
So many don’t even bother to tide-up fold their commits into single logical change, consider how many commit agent doing, seems logical to have folded commit history when you open PR.
Default commit and PR body text is word salad didn’t seems to bothered these people. Because if it did they already had hooks and skills to make it look tidy and lack of any AI attribution inserted.
I don’t even consider myself developer more like really weak devops guy, but I was already doing all of that.
localaicost.com