Garbage in, Garbage out. You heard the phrase a lot. But usually, it’s the people that create the AI models that take it seriously. Or, unless if you’re a pipeline engineer of not only in software and AI, but also in other industries like making fertilizer, or working in factories, you may agree to the phrase, but never had the chance to experience it yourself. But not anymore. As a user of AI models, everyone now had the chance to be on the end of ‘garbage out’, and because it’s not you who train the models, you don’t even have the chance to control what garbage to put in! Recently, one was working on a software to detect changes to a database and push it to the client in ‘real-time’. So, for example, there’s an INSERT operation on a particular table one subscribed to, and that should trigger something and push the change to the backend, where one could then process the data into a form one would like to view before passing it to the frontend. Now, that was something one never done before. Previously, one settled for getting the client side to ping every n seconds for new data, and it worked then, because there’s no requirement for ‘real-time’, so one could do it every 30 seconds or 1 minute. The ‘infrequent’ pinging means it won’t overload the bandwidth, both on the server and on the database. But when you have a quite large data to send back to the front and you use frequent pinging and you’re not installing it on localhost, that occupies some bandwidth. Plus, frequently pinging the database to check if there’s new data, even one as simple as checking for the latest ID, would take a toll on the database if it’s in heavy usage at the same time. Not to say they couldn’t be done, but wouldn’t it be better if you can have the database listening to change, notify you to run a command on the server when there’s an update, and then push the update to the client? Save the bandwidth and the toll! something n quite The traditional way to search for a solution would be to look it up on the web. Actually, it’s not very traditional, still the mainstream way to do it, but with the advent of AI boom and startups competing with established businesses to deliver the latest, state of the art (SOTA) language learning models (LLM) where you could type in your request and have AI do it for you, the nature of searching had changed, so had the behavior of its users. A search engine is designed for you to first think about how you are to execute your plans, and for each step of the plan where you don’t know how to do, you search it up, then stitch together the idea yourself and be complete with the final product. The new search would be to be disciplined and complete the 7-days cumulative usage of Perplexity and gain its Perplexity Pro free for a month, then maximize the advantage you can take by, instead of using its free ‘search’ feature, went for the pro ‘research’ features and ask, “Hey, using .Net Core 9 and Angular 19 and MS SQL Server 2022, how can I push data from the database (DBName.dbo.Table1) to the frontend client in real time? List for me the steps required, the libraries required to install, and show me example codes of how to do it.” And while it does its magic in the background (which can take several minutes -- and if you use ‘lab’ that create presumably ‘working’ projects that start from scratch, it takes even longer than ‘research’), go and have a cup of coffee/tea, hug your cat/dog and play with it for a while, until you heard ‘ding’, the notification that the answer you want had been completely generated for you. You put down your cat/dog, have a final gulp of your coffee/tea, before fixating your eyes back on the screen, reading through the answers from top to bottom, with all the steps laid out for you, and started integrating it into your own program. yourself Now, if you’re a programmer or software developer and equivalent, you’d know that every situation is somewhat unique. Sure, some bugs are well known and well documented, but most others you may find something on the internet, but the solution they gave, although they might’ve solved the person asking the problem, they might not solve yours, even though the problem seems the same. Debugging is a strenuous and indirect process, and a specific code like 400 Not Found can be caused by a lot of things, each probably unrelated to one another, even when they raise the same error. Why do you think, if you cannot solve the problem immediately using ‘traditional’ search engine method, that the AI can solve it for you? Do you really, genuinely, think that the solution comes easy for them just because their speed of scraping through the web is 100, no, 1000 times yours? And do you really think the abstracted solution generated by AI that works in most but not all (and unfortunately, yours is the exception, else you won’t be feeling the headache right now) really knows what you want and fit their solution for you, as if they could read your mind or read the entire state of your computer and trace out the stack deep into your program and provide you with an instantaneous fix while you sit back and relax, pursuing your hobby? can really just because but not all your SignalR, developed by Microsoft, is the high-level library published by Microsoft to push updates to the client side. Unfamiliar with how SignalR worked, one, of course, followed the steps compiled by AI and, after manually fitting it to one’s program, copy and pasting the code to their respective places where one sees fit, converting the ‘abstract’ into the ‘specific’, and run the program. And… it didn’t work. And one don’t know why. One next use traditional search methods and looked up online, including the tutorials offered in Microsoft’s website, and tried to solve it, to no avail. Trying one after one solution, days after days passed, the solution just won’t materialize. It kept telling me that “404 Not Found”, but what wasn’t being found? Didn’t one followed the steps listed down by AI and Microsoft already? And shouldn’t it had magically worked out? Reality is much more harsher than fantasy. The realization come when one dig deeper into how it differs from normal HTTPS request. In fact, because one uses QuickApp developed by ebMonney, one never had the chance to understand how the proxy works to transfer data to the backend. Before, when one realizes that some of the links, like /api, got proxied to the backend, while the others get parsed by Angular, one thought, perhaps if one changes the /hub one uses for SignalR to /api/hub, it would work. For sure, the error saying 404 didn’t come up anymore, but it still didn’t work. There had to be something deeper that caused it. QuickApp developed by ebMonney had There was a proxy.conf.js, which one heard of it but forgot about it because it’d been built for one so easily that one don’t need to worry about it… until then, that allows you to list the path you want to exclude from being parsed by Angular’s app-routing.module.ts. Basically, Angular will parse all routes in its app-routing, and if it cannot find, it’ll redirect to a page called 404 Not Found, which is a HTML custom-made by ebMonney and included within the QuickApp package. However, if you want angular to proxy some specific link to your backend server, you need to add it inside proxy.conf.js to exclude it from being parsed by Angular. But how? That’s another question for AI. “I keep getting 404 for /hub because it wasn’t added in the proxy.conf.js for SignalR Angular. How do I add it?” And this time, one had to be grateful that one uses AI to generate the answer. You see, SignalR don’t uses HTTPS by default, but WebSocket, which goes through WSS. And in the excluded list already included in the default proxy.conf.js, all the links are proxied through HTTPS layer; if AI hadn’t told me to set ws: true, one would have another round of headache to deal with. Though one didn’t try, one suspect that’ll be the same error message as when one uses /api/hub earlier. ws: true The next part is as expected. Because one forgot to change /api/hub back to /hub in the frontend, but one changed it already in the backend, it returned a 405 Method Not Allowed error. Another confusing error that tells me something else unrelated to what actually happened. Changing it back to /hub in the frontend fixed the problem. So when one posted the question and answered the problem on Stack Overflow, one get a -2 for the question within 24 hours for the lack of debugging details. The person editing the question is right, unfortunately. Though what he didn’t know was that there’s no debugging details, because it’s not a problem with the library, but with the setup. Particularly, the setup provided by AI -- and that is the debugging detail. Let me explain. posted the question and answered the problem on Stack Overflow provided by AI that Recently, in Stack Overflow “Issues” (if you subscribe to it, they’ll send a digest of a list of links to your email of what they picked to be interesting, and other links on what’s being sponsored to advertise to you), Issue 280, there’s a link titled Dispatch from the trenches of the Butlerian Jihad. The headline, of course, wasn’t useful in telling one what’s it about, so one, of course, asked Perplexity to summarize it for one, to gauge whether it’s worth reading or not. With reading, one’s very careful about reading proxies than the original, because, as Mortimer Adler mentioned in his book, How to Read a Book, the digests given by other readers are usually a bad proxy to judge the content of the book. Similarly, the contents summarized by AI trimmed way too much content away from the original article. In fact, it’s bullet-pointed under several headings, like those digests sent by self-help articles that compile articles into points for those who deemed themselves no time to read the originals. When one read the original, it’s a compelling story about a teacher’s life experience and the frustration of dealing with kids using AI to complete their homework, and he had to read about AI-produced stuffs, some clearly not double-checked by the student and produced to him gibberish, and the praise that tries to flatter him but ended up confusing him because Dispatch from the trenches of the Butlerian Jihad How to Read a Book …about how I’d become “not just a teacher but a mentor” when I’d never once seen them at office hours. …about how I’d become “not just a teacher but a mentor” when I’d never once seen them at office hours. It’s not one don’t want to provide technical debugging details on the topic, but, the problem being a misuse of AI to help one create something step by step without a deeper understanding of how it actually works, rather than using AI just like how you would use a search engine to search for specifics and compile the information before you use it to create one’s own step by step methods to solve the problem, is the problem. The teacher might have seen the scenario played out with his students, but adults are no less tempted to misuse AI, just because how they’re being designed. AI, or rather, AGI, a computer capable of listening to your commands and execute what you want just by telling it what you want, is slowly becoming a reality. No wonder the hype is in creating LLMs, and while other AI models like predictive models and reinforcement learning models might be equally useful in different scenarios, they do not play to our expectations of a futuristic utopia that’s cool enough for not only the geeks but also the average. We really wanted chatbots that can do everything, from solving our everyday troubles, to “solving the loneliness epidemic”, have them chat with us about stuffs when other people were busy with their own lives. You might think the latter sounds unethical, and the people who supported it stupid, but believe me, there are people who’re desperate enough they need someone to salve their loneliness but no one wants to come to them, and in the end, they have no choice but to resort to chatbots, hoping an artificial entity created to serve you 24 hours a day, that never gets angry with you like normal people does, that doesn’t get annoyed with your complaints like normal people does, would save them from the depth of hell they plunged into, alone, for too long, and pull them up, seeing the light from heaven yet again. Except, we just really don’t know how to use AI. one’s own
