MetaMask (the crypto wallet) has one that shows warning pages to all domains that are remotely similar to crypto-related domains, and it is very prone to false positives and annoying. They have to maintain a list to skip the detection for real domains, and it's really inefficient.
Feels like this kind of detection is hard to balance, and calling legit websites possible phishing might be problematic...
Seems like the kind of problem LLMs would be perfect for. ChatGPT does a great job at giving a score of whether domains are attempting to appear legitimate, but of course no one wants their browsing history being sent off to OpenAI. Unfortunately from my testing local 4B-7B LLMs aren't up to the task.
I wonder how the map handled the accuracy issue of the timetable data: the public timetables are only accurate to the minutes, but the trains on the map are not only arriving / departing at 0 second. I believe accurate timetables that are accurate to 5 seconds (actual accuracy) are not published anywhere. I'm now very curious how accurate the trains on the map are, as there might be up to 55 seconds of error.
I don't have too many opportunities to take the trains nowadays, and when I had to take the trains I always forget to check this.
> Is it really so hard for people to separate good from bad?
If you mean "HN people", then probably no; but if you mean more generalized "people", then definitely yes. Unfortunately, your customers are much more likely from the latter group.
Also, being in a sad state doesn't really mean it "has no use case", right?
Not understanding what you're saying here. If anything HN has less able to separate good from bad at least on this topic imo, and I wasn't claiming the direct parent posted said exactly it has no use case, but still even that they had to couch any slight mention of utility with being a "sad, sad state" implies they don't like it and know it isn't liked here.
More than halfway through the story mode, and there are some really annoying issues around and I can't find a way to provide feedback, so I'll leave them here:
1. There should be a way to inspect the schema. As one commenter mentioned, `show create table` doesn't work, and you'll have to `select *` which will count as incorrect answer, making the error counter almost meaningless. I know there are hints and most of the times they works, but there are situations that you would want to directly look at column names (see below).
2. You can't edit the previous steps in multi-step chapters, and the answer checker is not catching some errors. For example, the step 1 of chapter 19 has the hint "watch out for column ordering", but because of point 1, there is no way to know the table structure beforehand, and I decided to just do it blindly without adjusting the order to see how it goes. The answer is obviously incorrect, but the game accepted it as the correct answer and made the textarea readonly. Now I can't fix that and have to fix it again after all the steps are completed, but some intermediate steps will be using the wrong answer.
3. (slight spoiler?) There is a logic error in step 5 of chapter 19: considering the lifts are denoted by unique names (as you are asked to group by it), the answer checker expected lifts with inoperable malfunctions to be "usable", while the hint indicates otherwise.
4. The TV effect consumes a lot of CPU power, at least on Firefox (didn't test on other browsers).
Still a great one overall, looking forward to try the challenge mode after I finish the story.
BTW, this is also how iOS apps could detect jailbreak status of the device: just try to open paths like `/var/lib/apt`, if it does not exist, it should return ENOENT; otherwise you would know this device is “not clean”.
Didn’t think the sandboxing on macOS also has this issue.
