Agreed. I feel like humanity has not yet figured out how to effectively avoid using (or developing!) harmful technologies. We are quickly reaching the point where the question isn't "can we", but "should we". This applies to a lot more than AI, too.
Same for me. I have adopted a "zero usage" policy for the moment. In part it's because I am too worried about the long-term effects of cognitive offloading. In part it's because I don't want to support in any way the development of a technology that I believe is becoming a serious threat to humanity. And there are also things that I enjoy doing, like coding, solving problems, or writing in my own style, which I simply don't want to delegate to a machine.
Having tried LLMs for coding a couple of times in the last couple of years, this used to be me until about a month ago.
Then I again tried Claude Code and OpenCode with a local (due to work I cannot use anything else, moreover wouldn't want to either) Qwen3.8 and behold: my current stance now is that _not_ using something like that, if you can, is missing out on a great tool.
And maybe it's because of the way I use it (in essence: to take care of what is the slowest bit of the process for me which is the typing and the find/replace and other repetitive stuff), but I really cannot come up with any sane reason not to use it.
Somewhat similar experience over here. I'm generally a skeptic during this current hype wave but have been forcing myself to try the latest "default" setup for AI-powered coding every month, just to not completely lose touch with the industry and double down on an opinion based on old models.
A few weeks ago something clicked between Claude Code getting better and my expectations about what it can and can't do becoming more realistic. Now I happily use it for code generation, rubber duck debugging, search tasks (in the code, in git history, in logs, across repos, in our ticketing system, etc) and for rote tasks such as opening PRs.
Still not a fan of vibe coding or doing ten tasks in parallel or generally overly relying on these tools but I do get a lot of extra value in my day-to-day work.
I guess you can find sane reasons not to use LLMs outside the purely technical realm when you consider the economics and ethics of how the sausage is made...
Even if I wouldn't use it to write actual code, the value "now review this" currently brings is really quite large. It finds real issues to straight up bugs in any decently sized codebase which has only been reviewed by a couple of people. It's humbling at times.
Still not a fan of vibe coding or doing ten tasks in parallel or generally overly relying on these tools
Me neither and I probably never will be. It also just doesn't match how I usually write code. For anything which matters I'm a 'one statement at a time and consider all possible consequences' type of person.
I guess you can find sane reasons not to use LLMs outside the purely technical realm when you consider the economics and ethics of how the sausage is made...
It's definitely a 'pick your battles' kind of thing.
Could be that I missed something, but I doubt it. I'm not talking simple textual find/replace here. I mean being able to tell a tool in natural language things like "x is now a dictionary, lookup all keys used, then change it into a type where each key is a member and update all call sites" and having it execute that correctly. In all commonly used programming languages. And this is still a fairly simply example. Which tools could do that 2 years ago?
Same for me. Since I quit tech (professionally) about 5 years ago, I was never part of the AI thing. I still write a lot of code, but I enjoy coding, so I don't use AI.
Specifically mostly React and Angular, and my job has a nontrivial amount of other related work involved with implementing new features for customers (requirements work, bugtesting, etc.); but mostly, I write code. I like to do it, the ethical and existential concerns I have with AI make me hesitant to use it, and for most tasks the overhead involved with finding the correct requirements and implementation method prevents AI from being too helpful for me. I'm sure a frontier model could be much more useful than what I've tested in the past, but I'd be insane to use my own money to pay for one just to make myself more stressed and unhappy. There is some work coming up that will require a lot of "convert thing X to thing Y" type tasks that I may test AI on again, as that it could handle and I do not have any ethical issues with automating.
I've never understood why people have this viewpoint. If there is no preference for one property of a work (ie. lighting) to match reality just because another unrelated property (biology) doesn't, then surely the only possible scenarios remaining are an exact replica of the real world (where all properties match the real world) or an entirely arbitrary construction (where no properties should match reality). Fiction is predicated on the idea of "reality, but...". Saying that any single difference from reality should lead you to expect no similarities to reality at all, if followed, would collapse all fictions into pure arbitrary soup.
That's not what they said, they said it limited its output to 2 issues because it saw the "both" in the sentence, which is clearly an incorrect reading of their instructions.
Yes, I think I misread -- looks like the pebkac problem was at my desk instead :/
But I don't think their prompt as written would have limited the responses to two, unless there really were only two matching both conditionals. I've not seen an LLM misinterpret language that way before.
You're right, but I do wonder if they're misremembering their prompt a bit. The AI's response would be much more likely to arise if the prompt was something closer to "find both issues that are critical and related to X". Based on my own experience, it would be a very surprising mistake for the AI to have made if the prompt really was as straightforward and precise as "find issues that are both critical and related to X". I know it's a bit annoying to question someone's account of their own experience, and obviously I'm not saying this couldn't have happened, but my guess would be that the actual prompt was at least somewhat ambiguous.
I do think it's reasonable-- although certainly not the only valid approach-- to be aware of historical misinformation and emphasize certain facts more prominently to counteract it. It could skew the emphasis the other way in the long term, but the more immediate benefits of combating misinfo actively seem justifiable.
One of the challenging things about sharing knowledge is that it doesn't exist in a vacuum. With certain things there's definitely more complicated and nuanced approaches that may make sense. It's absolutely true that it can swing back too far - and that's usually the case with the most important areas of life, I'd argue. It's the cause of all sorts of seemingly-well-intentioned attempts at sharing information, as well as, of course, intentionally malicious forms of sharing information. That's what all propaganda does, realistically.
To bring it back to the topic, this is certainly a serious concern with AI-generated content. The world changes and evolves and things that are acceptable become unacceptable and vice versa. With training data potentially being stuck in time due to the majority of new data being generated, history could essentially just stop here. Or at least that is an idea that's been proposed. I'll be honest that I have thought that to be a genuine concern. However, objectively, novel ideas have come out of these new technologies. So it's a bit of a pickle to figure out how that relates to the human experience as time marches on.
Using reserved space is still using space. Creating a folder is using up some portion of the reserved space until it will need to be expanded and take up more free space on the disk. Preemptively allocating that reserved space doesn't make the folder not take up space. If I create a file on my desktop, is it not taking up any space because I haven't yet had to expand my partition size or get another disk?
Yes. But not using space from the pot that people are usually meaning when they talk about available space, the amount listed in the `avail` column in the output from `du`, or the space free number displayed by Windows Explorer and other GUI file managers.
Not that it usually makes much difference in the grand scheme of things, outside of made up examples like those mentioned above, or specific cases like file-per-message mail/usenet stores where you could easily run it of inodes before running out of main storage space on a filesystem and small files not taking more than the one inode (no extra storage block) could save a significant amount of main space in total.
Plan mode does have different permission controls though? Its difficult for me to parse the docs on this but it does seem to imply more than just a prompt although less than a full sandbox. https://code.claude.com/docs/en/permission-modes#analyze-bef...
The way it worked originally was it would trigger a different system prompt, and then the permissions system would prompt you (even on YOLO mode) before running any non-readonly actions.
Cursor would also write a plan document, which was useful when working on larger tasks (due to context size). I still find that part useful today.
Plan forced the model to write down and outline what it was going to do before writing the code. As Boris said up thread, he created it to prevent it from just rushing into writing code. So that regard, it did act like a permission system.
It helps to think more abstractly when approaching problems like this.
In the states I've lived in, its a piece of paper you keep in your wallet alongside your previous DL/permit that's now been hole-punched. You use both together to serve as your proof of authority to drive if you get pulled over.
Of course, only until the real DL arrives in the mail, which usually takes a week or two.
"For non-photographic images, the argument that “they should be vector images” doesn't hold up because many images could be vector images but aren't, and they can't be vectorized perfectly. “The world should be different” is not a justifiable defense against optimizing for the way the world actually is."
This is a good paragraph. I see people forget to apply that last sentence all the time.
Agreed. JPG is the most popular format for photography. PNG, the lossless format is great for diagrams and stuff with sharp edges. It can be "compressed" via dithering and quantization.
GIF is there... because it will always be there because it's just a hit that will never go away. It's also lossless (when discounting colour depth).
The new formats are interesting webp, avif, JXL and so on.
Jpeg classic is the clear winner IMHO and will probably never be less relevant than the new fancier formats. JPG even does text and sharp edges, reds, and so on well if you tell your process to take into account chroma subsampling!
Gif is a weird one because it's silently been replaced while keeping the name around. Virtually every software using the term "gif" actually uses a h264/vp9 video with no sound.
Modern video formats are so much better that they are a no brainer to use while the word gif just got used to refer to looping video without sound rater than the literal gif image format.
The major blunder here is browsers refusing to support silent videos in tags etc. so we end up with all these hacks. Instead we got animated webp that you can't use for progressive enhancement because it doesn't have its own mime type and browsers reported support for image/webp before implementing animation support. And webp is also very limited when it comes to inter-frame compression for no good reason at all.
Gif doesn't mean "no sound" anymore. Gif is more or less "short video", which broke a lot of assumptions at first but we have pretty much passed that transition and it makes sense for what they are used for today.
Those are messaging/communication apps,sound in those apps can be quite disruptive so that doesn't surprise me at all. It has nothing to do with the original gif, it is just that gif is now synonymous with short clip (with or without sound) and sound in that context is disruptive.
Go to google.com, search for images and select image type gif and you will get mp4s with sound.
Telegram's usage of "GIF" (any short video without sound) is incredibly weird and frequently goes against users' expectations. I wouldn't treat that as definitive.
I ... what? I've never heard anyone conflate the two--not teenagers, non-technical adults, or software people. Most folks know that "GIF" means "moving image that can be easily shared but has no sound and usually crap quality", not a video/audio clip.
Reminds me of when Imgur started trying to push changing gifs over to gifv, which were just mp4 with a strange new file extension. I didn't like their approach but it really was such an improvement on file size and quality.
Not to mention if you for some odd reason wanted to do gif-style animations, you can also do it with apng or webp. You shouldn't, real video format makes more sense, but even if you did there is still no reason to use gif.
Perhaps the clue is in the past tense "supported". Nowadays browsers support many formats, but in the "olden days" gif was the only way to get animations in browsers without plugins like Macromedia Flash...
also, don't forget - actual video format is better for animated stuff. Current default approach is .webm which is standardized .mkv container + vp9 (or av1) video + opus audio (usually no audio to as a feature parity to gif). IMO animated picture formats like .avifs are very backwards and don't add much
One advantage to animated AVIF is that it supports an alpha channel, which AV1 doesn't. But I guess that just points to how short-sighted it was for AV1 to omit alpha channel support.
> Gif is a weird one because it's silently been replaced while keeping the name around. Virtually every software using the term "gif" actually uses a h264/vp9 video with no sound.
Then, long live the iMessage and Whatsup exploits, because SW developers are not able to do things right.
I began a dispute on the English Wikipedia which revealed that more than a few people use “GIF” to describe, not a file format, but a visual style of looping animation.
> PNG, the lossless format is great for diagrams and stuff with sharp edges.
But also, like GIF, it's pointless when lossless WebP exists and is supported everywhere. It's more efficient. Lossy WebP does not make as much sense though since jpegli entered the game.
OP has made https://halide.cx/iris/ which could make lossy WebP make sense, but it's not open source or generally available.
> The new formats are interesting webp, avif, JXL and so on.
WebP was introduced 15 years ago. New relative to JPEG perhaps, but certainly not new.
PNGs will persist because I know that they're lossless by looking at the filename. When I hit File > Save as .png, I don't have to inspect output metadata (that image viewers don't expose) or play around with exporter settings.
It's unfortunate that we didn't give lossless webps a separate extension. Frankly I think even saving them as .png would've better than what we have now - people roughly understand that a .png is an image without artifacts, and any self-respecting image viewer will inspect magic bytes before opening a file anyways.
Being lossless is a function of the whole pipeline from an input image to an output image. The last step being encode to PNG doesn't guarantee anything about losslessness.
I do not know whether this has anything to do with any limitation in the WebP format, because it is more likely that this is some kind of bias in the kind of authors of Web pages who choose the WebP image format, who might also choose more aggressive compression settings, but for me WebP is a synonym for "low-quality image".
This means, that at least for the Web sites that I happen to access frequently, whenever I see PNG or JPEG images they typically have a better quality than when I see WebP images.
I do not know whether when browsing the Web I see frequently any AVIF images. They are either rare or they have a similar quality with JPEG images, so they did not stand out like WebP, where I perceived a degradation, so I noticed that the image format was different.
I do not care much about the speed of decoding Web images, because the speed of loading Web pages is determined mostly by the amount of junk JS that is included and not by the loading and decoding of the included images. So for me the quality of the included images is more important, thus WebP is something that I dislike, while JPEG is far too limited, so I would prefer JPEG XL images.
The arguments in TFA do not seem entirely honest, because the presented data confirms the low quality of WebP in comparison with JPEG XL, and because the 2 compared JPEG XL implementations differ in their performance by a factor more than 2, so it is likely that they are both immature and one could still accelerate the decoding of JPEG XL by at least 2 or 3 times in comparison with the best of them.
The only claim of TFA that seems valid, is that if AVIF is really as good as claimed, it could be sufficient for Web pages that want to show only highly-compressed images and good for replacing JPEG for such a use, while JPEG XL might be strictly required only for applications that need higher quality, i.e. for the processing and storing of photographic images.
But supported by no browser today and software support in general is quite limited. If the PNG authors would just incorporate this possibility into PNG from start it would be interesting, but it's too late now.
Gif is still there because it's still often the only option for frictionless video embedding. Embedded videos usually come with a video player that hardly ever quite works out well, often not starting automatically, not looping, or showing video player UI elementa that shouldnt be shown.
But this is quite a different format than JPEG or better said the subset of JPEG that we know today as JPEG format and the JPEG encoder/decoders support. The whole JPEG format did have support for a lossless mode for JPEG (different one than JPEG LS), arithmetic coding and 12-bit images too, but those features just didn't make it into the implementations of the format (arithmetic coding initially did, but was removed from popular implementations because of patents).
Png is still being updated, just last year it was updated to support a wider color space than any other web format. This shipped in all browsers before the official spec was publicly so you’ve been seeing Png upgrades without even realizing it
Also it would be nice to have a good standard vector format that isn't the complete kitchen-sink that is SVG, but we don't have that. We needed "jpeg but for vector graphics" and instead we got "html but for vector graphics".
So anybody saying "we should use more vectors" has to remember what the state of open vector graphics standards is.
In the context of the world it's all hair splitting, you can have a big photographic picture in png, or 1% compressed jpeg, or 1% jpeg converted to png. No new format can change it.
reply