I wonder if it makes sense for ad platforms to enable a "clear history" option just like our browsers do? That way you can say very clearly "the current model is wrong, please restart and try building a new one for me". That might also help this particular case: rebuild the model from scratch including only content going forward rather than backward. Should eliminate the baby ads effectively, right?
Interesting idea, but I'm bothered by the idea that social media platforms would be effectively monetizing grief in this solution.
This sort of thing, to me, points at the deficiencies of appealing to purely statistical patterns without any guard rails placed on top. After all, statistically, the most likely relevant categories after "pre-natal" are going to be surrounding children who _were_ born safely. So if you appeal just to data alone, you won't likely solve this problem because it is relatively rare.
The biggest failure, to me, is the response of the system when the user goes out of her way to say that the "pre-natal" ads are not relevant. Statistically it assumes successful birth, but that doesn't reflect her actual intent. A simple dialog tree would maybe suffice: "not relevant" -> "this specific thing is irrelevant" vs "suppress this and all related content". One signal says maybe don't show that particular brand again, and the other says to pivot the relevance model significantly away from that topical section of the ad space entirely.
The correct response is not to try to advertise targeting the traumatic event, but to pivot and avoid it and related subjects entirely. I doubt people will complain if ads about babies and parenting were to disappear after a still birth or miscarriage.
If this had occurred in the US (and the mentioned confidential conversation did in fact allege sexual misconduct), DJB would be in trouble:
> Within the University of Illinois System, ALL employees, unless specifically exempted, are “Responsible Employees” with the responsibility and authority to report sexual misconduct to their university's Title IX Coordinator. The only employees who are exempt from this reporting requirement are professional or pastoral counselors who provide work-related mental-health counseling, campus advocates who provide confidential victim assistance, and employees who are otherwise prohibited by law from disclosing information received in the course of providing professional care and treatment. Student and graduate employees are handled differently at each university. Please reference the Responsible Employee Resource Page under the "Portfolio" and Resources tabs. Please remember that all references to Responsible Employees are references to YOU and apply to you in your capacity as a university employee.
To me, this would mean that he is a mandatory reporter, and I am unaware of any scenario where you are freed from that obligation because it was a "confidential conversation".
The weird part comes in when you realize that (a) this is happening outside of the US, but (b) DJB likely has NSF grants, which require adherence to Title IX (this is what the author is referring to when he brings up Title IX training). But how does one enforce Title IX outside of the country in which it was passed?
I absolutely love this project. I've been making bindings for a C++ library using it recently and have yet to encounter a situation that was too tricky to bind properly. I'm loving the ability to take an abstract class from the C++ side of things and derive and override virtual functions in Python. Super excited about how much easier that makes it to do exploratory programming to get something that woks, and then worrying about writing it in C++ when I need that level of performance.
The lack of a dependency on Boost only sweetens the deal for me. I love what Boost has set out to do, but it's just such a huge dependency with a ve4y strange (for me) build system.
He starts the talk by assuming monosyllabic words as his primitives and builds up the words he needs to use to give the talk by providing definitions for them first.
libstdc++, libc++, and MSVC's STL all implement it as a red-black tree.
Even if it's not required to be a red-black tree, it is de facto a red-black tree on every major compiler.
Moreover, this doesn't really change the parent's point. It must be some sort of ordered associative container, meaning that it's not going to have the performance characteristics of a hash table.