What’s the equivalent phrase to “hair of the dog”, but for caffeine?
I’ve always been pretty sensitive to it, and while I’m less-so now, it was still definitely a mistake to eat several portions of tiramisu right before bedtime.
So now I’m going to need more caffeine, this morning, to compensate for the lack of sleep.
What do we call that?
This post is also available as a podcast. Listen here, download for later, or subscribe wherever you consume podcasts.
Eleven years ago, comedy sketch The Expert had software engineers (and other misunderstood specialists) laughing to tears at the relatability of Anderson’s (Orion Lee) situation: asked to do the literally-impossible by people who don’t understand why their requests can’t be fulfilled.
Decades ago, a client wanted their Web application to automatically print to the user’s printer, without prompting. I explained that it was impossible because “if a website could print to your printer without at least asking you first, everybody would be printing ads as you browsed the web”. The client’s response: “I don’t need you to let everybody print. Just my users.”1
So yeah, I was among those that sympathised with Anderson.
In the sketch, the client requires him to “draw seven red lines, all of them strictly perpendicular; some with green ink and some with transparent”. He (reasonably) states that this is impossible2.
Following one of the many fever dreams when I was ill, recently, I woke up wondering… how might an AI programmer tackle this task? I had an inkling of the answer, so I had to try it:
RSS
Spotify
Pocket Casts
Apple Podcasts
YouTube<canvas> element3, the
question is the same as in the sketch.
When I asked gpt-4o to assist me, it initially completely ignored the perpendicularity requirement.
Let’s see if it can do better, with a bit of a nudge:
gpt-4o claimed that the task was absolutely achievable, even clarifying that the lines would all be “strictly perpendicular to each other”… before proceeding to instead
make each consecutively-drawn line be perpendicular only to its predecessor:
You might argue that this test is unfair, and it is. But there’s a point that I’ll get to.
But first, let me show you how a different model responded. I tried the same question with the newly-released Claude 3.7 Sonnet model, and got what I’d consider to be a much better answer:
In my mind: an ideal answer acknowledges the impossibility of the question, or at least addresses the supposed-impossibility of it. Claude 3.7 Sonnet did well here, although I can’t confirm whether it did so because it had been trained on data that recognised the existence of “The Expert” or not (it’s clearly aware of the sketch, given its answer).
I remain committed to not using AI to do anything I couldn’t do myself (and can therefore check).4
And the answer I got from gpt-4o to this question goes a long way to demonstrating why.
Suppose I didn’t know that it was impossible to make seven lines perpendicular to one another in anything less than seven-dimensional space. If that were the case, it’d be tempting to accept an AI-provided answer as correct, and ship it. And while that example is trivial (and at least a little bit silly), it’s the kind of thing that, I have no doubt, actually happens in other areas.
Chatbots eagerness to provide a helpful answer, even if no answer is possible, is a huge liability. The other week, I experimentally asked Claude 3.5 for assistance with a PHPUnit mocking challenge and it provided a whole series of answers… that were completely invalid! It later turned out that what I was trying to achieve was impossible5.
Given that its answers clearly didn’t-work there was no risk I’d have shipped it anyway, but I’m certain that there exist developers who’ve asked a chatbot for help in a domain they didn’t understood and accepted its answer while still not understanding it, which feels to me like a quick route to introducing into your code a bug that happy-path testing won’t reveal. (Y’know, something like a security vulnerability, or an accessibility failure, or whatever.)
Code assisting AI remains really interesting and occasionally useful… but it’s also a real minefield and I see a lot of naiveté about its limitations.
1 My client eventually took that particular requirement out of scope and I thought the matter was settled, but I heard that they later contracted a different developer to implement just that bit of functionality into the application that we delivered. I never checked, but I think that what they delivered exploited ActiveX/Java applet vulnerabilities to achieve the goal.
2 Nerds gotta nerd, and so there’s been endless debate on the Internet about whether the task is truly impossible. For example, when I first saw the video I was struck by the observation that perpendicularity within a set of lines is limited linearly by the number of dimensions you’re working in, so it’s absolutely possible to model seven lines all perpendicular to one another… if you’re working in seven dimensions. But let’s put that all aside for a moment and assume the task is truly impossible within some framework unspecified-but-implied within the universe of the sketch, ‘k?
3 Two-dimensionality feels like a fair assumed constraint, given that in the sketch Anderson tries to demonstrate the challenges of the task by using a flip-chart.
4 I also don’t use AI to produce anything creative that I then pass off as my own, because, y’know, many of these models don’t seem to respect copyright. You won’t find any AI-written content on my blog, for example, except specifically to demonstrate AI’s capabilities (or lack thereof) when discussing AI, and this is always be clearly labelled. But that’s another question.
5 In fact, I was going about the problem itself in entirely the wrong way: some minor refactoring later and I had some solid unit tests that fit the bill, and I didn’t need to do the impossible. But the AI never clocked that, and I suspect it never would have.
This post is also available as a podcast. Listen here, download for later, or subscribe wherever you consume podcasts.
I’d already read every prior book published by the excellent Becky Chambers, but this (and its sequel) had been sitting on my to-read list for some time,
and so while I’ve been ill and off work these last few days, I felt it would be a perfect opportunity to
pick it up. I’ve spent most of this week so far in bed, often drifting in and out of sleep, and a lightweight novella that I coud dip in and out of over the course of a day felt like
the ideal comfort.
I couldn’t have been more right, as the very first page gave away. My friend Ash described the experience of reading it (and its sequel) as being “like sitting in a warm bath”, and I see where they’re coming from. True to form, Chambers does a magnificent job of spinning a believable utopia: a world that acts like an idealised future while still being familiar enough for the reader to easily engage with it. The world of Wild-Built is inhabited by humans whose past saw them come together to prevent catastrophic climate change and peacefully move beyond their creation of general-purpose AI, eventually building for themselves a post-scarcity economy based on caring communities living in harmony with their ecosystem.
Writing a story in a utopia has sometimes been seen as challenging, because without anything to strive for, what is there for a protagonist to strive against? But Wild-Built has no such problem. Written throughout with a close personal focus on Sibling Dex, a city monk who decides to uproot their life to travel around the various agrarian lands of their world, a growing philosophical theme emerges: once ones needs have been met, how does one identify with ones purpose? Deprived of the struggle to climb some Maslowian pyramid, how does a person freed of their immediate needs (unless they choose to take unnecessary risks: we hear of hikers who die exploring the uncultivated wilderness Dex’s people leave to nature, for example) define their place in the world?
Aside from Dex, the other major character in the book is Mosscap, a robot whom they meet by a chance encounter on the very edge of human civilisation. Nobody has seen a robot for centuries, since such machines became self-aware and, rather than consign them to slavery, the humans set them free (at which point they vanished to go do their own thing).
To take a diversion from the plot, can I just share for a moment a few lines from an early conversation between Dex and Mosscap, in which I think the level of mutual interpersonal respect shown by the characters mirrors the utopia of the author’s construction:
…
“What—what are you? What is this? Why are you here?”
The robot, again, looked confused. “Do you not know? Do you no longer speak of us?”
“We—I mean, we tell stories about—is robots the right word? Do you call yourself robots or something else?”
“Robot is correct.”
…
“Okay. Mosscap. I’m Dex. Do you have a gender?”
“No.”
“Me neither.”
These two strangers take the time in their initial introduction to ensure they’re using the right terms for one another: starting with those relating to their… let’s say species… and then working towards pronouns (Dex uses they/them, which seems to be widespread and commonplace but far from universal in their society; Mosscap uses it/its, which provides for an entire discussion on the nature of objectship and objectification in self-identity). It’s queer as anything, and a delightful touch.
In any case: the outward presence of the plot revolves around a question that the robot has been charged to find an answer to: “What do humans need?” The narrative theme of self-defined purpose and desires is both a presenting and a subtextual issue, and it carries through every chapter. The entire book is as much a thought experiment as it is a novel, but it doesn’t diminish in the slightest from the delightful adventure that carries it.
Dex and Mosscap go on to explore the world, to learn more about it and about one another, and crucially about themselves and their place in it. It’s charming and wonderful and uplifting and, I suppose, like a warm bath: comfortable and calming and centering. And it does an excellent job of setting the stage for the second book in the series, which we’ll get to presently…
After most of four days spent primarily in bed with what’s probably a norovirus infection (or something like it), this afternoon I got up and went outside. 🎉
It felt like a huge achievement, even if I ran back in to the warmth the very second that the dog I was supervising had finished her business.
Using WordPress internally at Automattic as a productivity tool is great… until you have to call in sick three days in a row and Jetpack treats your “streak” as an “achievement”! 😅
Third day of being ill with what’s probably a winter vomiting bug, with one child home sick from school… and just having had to collect the other kid who started throwing up on his school trip… I finally got back to my bed and picked up the next book on my pile, Becky Chambers’ A Psalm for the Willd-Built.
The opening page reads: “For anybody who could use a break.”
Yes. Yes, please.
Don’t remember the last time I felt so run-down. I’ve been unwell since Sunday with an illness I can only assume I caught from the 11-year-old, who’s been unable to keep food down for several days.
In my case, though, I’ve mostly been full of muscular aches and cramps, ocassionally fits of shivers, and strange dizziness.
I’ve spent the last day and a bit mostly drifting in and out of sleep, where I’ve had the weirdest dreams. I just woke from one where I was lost in a sprawling hotel, looking for my room which was number 317 or possibly 305, I couldn’t remember. The signage didn’t make sense to me and I couldn’t read it, and found myself wandering around a sprawling resort, with hot air balloon services connecting different parts. At one point I found myself lost in a library whose winding shelves formed a Escherian maze, and a small child watched me with suspicion as I fumbled around for an exit.
In a disturbing dream from yesterday afternoon, I was lying in a desert of cold sand as the wind gradually piled up more and more sand against me. At first I felt fascinated, thinking I’d learned something about how dunes form, until I discovered that I wasn’t able to move. I gradually sunk deeper and deeper underground, in pain except when I lay very still and let it take me, until eventually I started to become the very sand that I was disappearing into. I felt flakes of myself break away and become part of the desert, unable to resist the change nor reconstruct myself, resigned to my fate.
I’m in less pain so far this morning than yesterday, so I think I’m recovering, but man this has been an unpleasant ride. Whatever I’m infected with: do not recommend, would not contract again, one out of five.
This is a repost promoting content originally published elsewhere. See more things Dan's reposted.
…
5. If you use AI, you are the one who is accountable for whatever you produce with it. You have to be certain that whatever you produced was correct. You cannot ask the system itself to do this. You must either already be expert at the task you are doing so you can recognise good output yourself, or you must check through other, different means the validity of any output.
…
9. Generative AI produces above average human output, but typically not top human output. If you overuse generative AI you may produce more mediocre output than you are capable of.
…
I was also tempted to include in 9 as a middle sentence “Note that if you are in an elite context, like attending a university, above average for humanity widely could be below average for your context.”
…
In this excellent post, Joanna says more-succinctly what I was trying to say in my comments on “AI Is Reshaping Software Engineering — But There’s a Catch…” a few days ago. In my case, I was talking very-specifically about AI as a programmer’s assistant, and Joanna’s points 5. and 9. are absolutely spot on.
Point 5 is a reminder that, as I’ve long said, you can’t trust an AI to do anything that you can’t do for yourself. I sometimes use a GenAI-based programming assistant, and I can tell you this – it’s really good for:
DateTimeInterface::format character codes memorised. Now I’d have to look them up. Or I can write a comment (which I should anyway, for the next human) that says something like //
@returns String a date in the form: Mon 7th January 2023 and when I get to my date(...) statement the AI will already have worked out that the format is 'D
jS F Y' for me. I’ll recognise a valid format when I see it, and I’ll be testing it anyway.
Other things… not so much. The other day I experimentally tried to have a GenAI help me to boilerplate some unit tests and it really failed at it. It determined pretty quickly, as I had, that to test a particular piece of functionality need to mock a function provided by a standard library, but despite nearly a dozen attempts to do so, with copious prompting assistance, it couldn’t come up with a working solution.
Overall, as a result of that experiment, I was less-effective as a developer while working on that unit test than I would have been had I not tried to get AI assistance: once I dived deep into the documentation (and eventually the source code) of the underlying library I was able to come up with a mocking solution that worked, and I can see why the AI failed: it’s quite-possibly never come across anything quite like this particular problem in its training set.
Solving it required a level of creativity and a depth of research that it was simply incapable of, and I’d clearly made a mistake in trying to outsource the problem to it. I was able to work around it because I can solve that problem.
But I know people who’ve used GenAI to program things that they wouldn’t be able to do for themselves, and that scares me. If you don’t understand the code your tool has written, how can you know that it does what you intended? Most developers have a blind spot for testing and will happy-path test their code without noticing if they’ve introduced, say, a security vulnerability owing to their handling of unescaped input or similar… and that’s a problem that gets much, much worse when a “developer” doesn’t even look at the code they deploy.
Security, accessibility, maintainability and performance – among others, I’ve no doubt – are all hard problems that are not made easier when you use an AI to write code that you don’t understand.
1 I’ve 100% had an occasion when I’ve called something $theUserID in one
place and then $theUserId in another and not noticed the case difference until I’m debugging and swearing at the computer
2 I’ve described the experience of using an LLM in this way as being a little like having a very-knowledgeable but very-inexperienced junior developer sat next to me to whom I can pass off the boring tasks, so long as I make sure to check their work because they’re so eager-to-please that they’ll choose to assume they know more than they do if they think it’ll briefly impress you.
3 e.g. switching a selector from $(...) to
document.querySelector but then failing to switch the trailing .addClass(...) to .classList.add(...)– you know: like an underexperienced but
eager-to-please dev!
Especially outside of urban centres, and especially if you’re on foot, OpenStreetMap is way better than Google Maps, Bing Maps, Apple Maps, or what-have-you.
OpenStreetMap is especially good for walkers, with its more-comprehensive coverage of public footpaths as well as the ability to drill-down for accessibility information: whether a path ends in a gate or a stile matters a lot if you can’t climb the latter (or you’re walking with a small-but-muddy dog who’ll need lifting over).
Sure, you don’t get (as much) street view photography. But how often do you use that, anyway?1
I’ve heard it argued that OpenStreetMap, with its Wikipedia-like “anybody can edit it” model, cannot be relied upon. And sure, if you’re looking for an “official” level of accuracy and the alternative is an Ordinance Survey map, then that’s what you should go for.
But there’s nothing specific to, say, Google Maps that makes it fundamentally more “accurate” for most2 geographic features than OpenStreetMap. The vast of cartographic data on Google Maps is produced by humans, looking at satellite photos, and then tracing the features on them, probably with AI assistance. And the vast majority of cartographic data on OpenStreetMap is produced… exactly the same way, although without the AI “helping”.
Google Maps has mistakes, just like every map3. And it’s got trap streets, like most commercially-produced maps (including the Ordinance Survey). Google Maps’ mistakes tend to be made by somebody on the other side of the world from the feature, doing a bad job of tracing what they think might be a road… while OpenStreetMaps’ mistakes are for the most part omissions in areas that are under-explored by local contributors. And there are plenty of areas – like those near where I live, especially if you’re on foot – where the latter mistakes are much less-troublesome.
I fixed a couple of omissions on OpenStreetMap just earlier today. While I was out walking the dog, earlier, I added the names of two houses whose identities weren’t specifically marked on the map, and I added detail to the newly-constructed Deansfield estate. Google Maps shows there being only two houses on Deansfield Estate, among other inaccuracies, even though they’ve got up-to-date aerial and street photography.
Google Maps is fine if you want to drive to Sheffield, you need public transport connections to Plymouth4, or you’re looking for a restaurant nearby and you want the data about them to be accurate. But next time you’re walking somewhere, or when you’re looking for a specific address… I’d suggest you give OpenStreetMap a go. You might be pleasantly surprised.
1 I say that as somebody who uses street view and satellite photography a more than average amount, for geohashing purposes. But I can switch mapping software on-the-fly; nobody’s stopping me looking at “ostrich” photos when I need them.
2 The place that Google Maps really beats OpenStreetMap, in my mind, is in the integration of its business directory. If you search for a business in Google Maps, you’ll probably find it and get reasonably-accurate opening hours and contact details. But that’s a factor of two things: the Google My Business directory, and – more importantly – the popularity of the application and the fact that the mobile app “nudges” people to check on the places around them. By the way: if you want to contribute to making maps better in that way without becoming an unpaid researcher working to line Google’s pockets, StreetComplete is an app that helps fill-out business and related information on OpenStreetMap!
3 Google Maps used to show Vauxhall tube station on entirely the wrong side of the River Thames, for example.
4 Public transport’s another thing Google Maps does very well.
This checkin to GC9EXX4 Friar's Farm - Bisected Byway reflects a geocaching.com log entry. See more of Dan's cache logs.
Checked up on this cache while the dog and I were nearby. It’s in fine condition and ready to find. The latch for the container is beginning to rust, but the whole thing is perfectly serviceable. Go find it!
This is a repost promoting content originally published elsewhere. See more things Dan's reposted.
I don’t believe AI will replace software developers, but it will exponentially boost their productivity. The more I talk to developers, the more I hear the same thing—they’re now accomplishing in half the time what used to take them days.
But there’s a risk… Less experienced developers often take shortcuts, relying on AI to fix bugs, write code, and even test it—without fully understanding what’s happening under the hood. And the less you understand your code, the harder it becomes to debug, operate, and maintain in the long run.
So while AI is a game-changer for developers, junior engineers must ensure they actually develop the foundational skills—otherwise, they’ll struggle when AI can’t do all the heavy lifting.
![Comic comparing 'Devs Then' to 'Devs Now'. The 'Devs Then' are illustrated as muscular men, with captions 'Writes code without AI or Stack Overflow', 'Builds entire games in Assembly', 'Crafts mission-critical code fo [sic] Moon landing', and 'Fixes memory leaks by tweaking pointers'. The 'Devs Now' are illustrated with badly-drawn, somewhat-stupid-looking faces and captioned 'Googles how to center a div in 2025?', 'ChatGPT please fix my syntax error', 'Cannot exit vim', and 'Fixes one bug, creates three new ones'.](/_q23u/2025/02/devs-then-devs-now-lieo-640x601.jpg)
Eduardo picks up on something I’ve been concerned about too: that the productivity boost afforded to junior developers by AI does not provide them with the necessary experience to be able to continue to advance their skills. GenAI for developers can be a dead end, from a personal development perspective.
That’s a phenomenon not unique to AI, mind. The drive to have more developers be more productive on day one has for many years lead to an increase in developers who are hyper-focused on a very specific, narrow technology to the exclusion even of the fundamentals that underpin them.
When somebody learns how to be a “React developer” without understanding enough about HTTP to explain which bits of data exist on the server-side and which are delivered to the client, for example, they’re at risk of introducing security problems. We see this kind of thing a lot!
There’s absolutely nothing wrong with not-knowing-everything, of course (in fact, knowing where the gaps around the edges of your knowledge are and being willing to work to fill them in, over time, is admirable, and everybody should be doing it!). But until they learn, a developer that lacks a comprehension of the fundamentals on which they depend needs to be supported by a team that “fill the gaps” in their knowledge.
AI muddies the water because it appears to fulfil the role of that supportive team. But in reality it’s just regurgitating code synthesised from the fragments it’s read in the past without critically thinking about it. That’s fine if it’s suggesting code that the developer understands, because it’s like… “fancy autocomplete”, which you can accept or reject based on their understanding of the domain. I use AI in exactly this way many times a week. But when people try to use AI to fill the “gaps” at the edge of their knowledge, they neither learn from it nor do they write good code.
I’ve long argued that as an industry, we lack a pedagogical base: we don’t know how to teach people to do what we do (this is evidenced by the relatively high drop-out rate on computer science course, the popular opinion that one requires a particular way of thinking to be a programmer, and the fact that sometimes people who fail to learn programming through paradigm are suddenly able to do so when presented with a different one). I suspect that AI will make this problem worse, not better.
Yesterday, owing to some plot, the dog didn’t get as much outside/walk/play time as she’d like.
Today, she let me know how she felt about that by shredding my slippers. 🙄
I was experimenting with VP8/VP9 WebM video transparency and I made a stupid thing: a URL that, if you go to it, means you’ll he followed around my blog by a video of me just hanging out in the corner of the page – https://danq.me/?fool_id=06
I’ve added it to my list of “stupid/random things that can happen if you visit my blog on April Fools’ Day”: https://danq.me/fools/
Ruth bought me a copy of The Adventure Challenge: Couples Edition, which is… well, it’s basically a book of 50 curious and unusual ideas for date activities. This week, for the first time, we gave it a go.
As a result, we spent this date night… baking a pie!
The book is written by Americans, but that wasn’t going to stop us from making a savoury pie. Of course, “bake a pie” isn’t much of a challenge by itself, which is why the book stipulates that:
We used this recipe for “mini creamy mushroom pies”. We chose to interpret the brief as permitting pre-prep to be done in accordance with the ingredients list: e.g. because the ingredients list says “1 egg, beaten”, we were allowed to break and beat the egg first, before blindfolding up.
This was a smart choice (breaking an egg while blindfolded, even under close direction, would probably have been especially stress-inducing!).
#JustSwitchThings
I really enjoyed this experience. It forced us into doing something different on date night (we have developed a bit of a pattern, as folks are wont to do), stretched our comfort zones, and left us with tasty tasty pies to each afterwards. That’s a win-win-win, in my book.
Plus, communication is sexy, and so anything that makes you practice your coupley-communication-skills is fundamentally hot and therefore a great date night activity.
So yeah: we’ll probably be trying some of the other ideas in the book, when the time comes.
Some of the categories are pretty curious, and I’m already wondering what other couples we know that’d be brave enough to join us for the “double date” chapter: four challenges for which you need a second dyad to hang out with? (I’m, like… 90% sure it’s not going to be swinging. So if we know you and you’d like to volunteer yourselves, go ahead!)
This is a reply to a post published elsewhere. Its content might be duplicated as a traditional comment at the original source.
Had a fight with the Apparently I can’t just insert <style> tags into my posts anymore, because otherwise I’d have to somehow either put nonces on them, or hash their content (which would be more preferrable, because that way it remains static).Content-Security-Policyheader today. Turns out, I won, but not without sacrifices.I could probably do the latter by rewriting HTML at publish-time, but I’d need to hook into my Markdown parser and process HTML for that, and, well, that’s really complicated, isn’t it? (It probably is no harder than searching for Webmention links, and I’m overthinking it.)
I’ve had this exact same battle.
Obviously the intended way to use nonces in a Content-Security-Policy is to have the nonce generated, injected, and served in a single operation. So in PHP,
perhaps, you might do something like this:
<?php $nonce = bin2hex(random_bytes(16)); header("Content-Security-Policy: script-src 'nonce-$nonce'"); ?> <!DOCTYPE html> <html lang="en"> <head> <meta charset="UTF-8"> <title>PHP CSP Nonce Test</title> </head> <body> <h1>PHP CSP Nonce Test</h1> <p> JavaScript did not run. </p> <!-- This JS has a valid nonce: --> <script nonce="<?php echo $nonce; ?>"> document.querySelector('p').textContent = 'JavaScript ran successfully.'; </script> <!-- This JS does not: --> <script nonce="wrong-nonce"> alert('The bad guys won!'); </script> </body> </html>
But for folks like me – and you too, Vika,, from the sounds of things – who serve most of their pages, most of the time, from the cache or from static HTML files… and who add the CSP header on using webserver configuration… this approach just doesn’t work.
I experimented with a few solutions:
s/<script/<script nonce="..."/ to search-and-replace it in.
Both of these are terrible solutions. The first one leaves a window of, in my case, about 24 hours during which a successfully-injected script can be executed. The second one effectively allowlists all scripts, regardless of their provenance. I realised that what I was doing was security theatre: seeking to boost my A-rating to an A+-rating on SecurityHeaders.com without actually improving security at all.
But the second approach gave me an idea. I could have a server-side secret that gets search-replaced out. E.g. if I “signed” all of my legitimate scripts with something like
<script nonce="dans-secret-key-goes-here" ...> then I could replace s/dans-secret-key-goes-here/actual-nonce-goes-here/ and thus have the best of both
worlds: static, cacheable pages, and actual untamperable nonces. So long as I took care to ensure that the pages were never delivered to anybody with the secret key still
intact, I’d be sorted!
Alternatively, I was looking into whether Caddy can do something like mod_asis does for Apache: that is, serve a file “as is”, with headers included in the file. That way, I could have the CSP header generated with the page and then saved into the cache, so it’s delivered with the same none every time… until the page changes. I’d love more webservers to have an “as is” mode, but I appreciate that might be a big ask (Apache’s mechanism, I suspect, exploits the fact that HTTP/1.0 and HTTP/1.1 literally send headers, followed by two CRLFs, then content… but that’s not what happens in HTTP/2+).
So yeah, I’ll probably do a server-side-secret approach, down the line. Maybe that’ll work for you, too.
You're browing in "1999 Web 1.0" mode.
Ready to go back to 2025?