MarceliusK

I think there's a lot of courage in admitting that you're terrified of something, especially after watching someone you love go through it

I can't imagine what it was like for her to face the same diagnosis that took her father, only to have such a different outcome. The part about the drug becoming available a month before her diagnosis gave me chills...

This might not even be a bug in the traditional sense. Someone probably decided that switching apps should also switch what's shown on the external display

Apple's idea of magic sometimes seems to be making decisions for you and hoping they guessed correctly

I really hope the next treatment works

That last point is probably the hardest one to actually follow. When someone you love is sick, I imagine it's easy to turn every waking hour into researching treatments, chasing appointments and etc. It probably feels like doing anything less means giving up

It's pretty sobering that one of the most important cancer survival skills seems to be knowing how to navigate bureaucracy...

My concern is that poisoning the dataset doesn't really stop the scanning. The device is still poking around your network

This is why I'd really like to see a packet capture or at least some interface-level counters

I can understand a buggy device flooding the network. What's harder to justify is why a coffee maker needs to know anything about the other devices in your house in the first place

Somewhere there's an engineer who added Wi-Fi to a coffee maker because marketing wanted an app, and now a family is debugging network saturation caused by their breakfast routine

The keyboard layout issue sounds nastier than the reconnect bug to me. Input forwarding gets surprisingly weird once keycodes, layouts, modifiers and the target OS's interpretation all get involved

Synergy is great, but I think this is solving a slightly different problem

A lot of good software starts with refusing to tolerate one tiny annoyance and it's amazing in a way :)

That's an interesting wrinkle. If anything, externally funded international students seem like exactly the case where you'd still want the interview, since the funding answers "who pays for this?" but not "does this person have the background and understanding to do the work?"

I think this is true of defenses as they exist today, but that may be partly because the thesis currently does most of the signaling.

The tricky part is that live review is much more expensive than async review. But maybe that just means we were getting away with a cheap proxy because it happened to work reasonably well.

The key shift seems to be from "can you produce mathematics?" to "do you understand mathematics?"... Those used to be correlated strongly enough that a thesis or paper could serve as evidence of both. If AI breaks that correlation, evaluating people through defenses, discussion and live problem solving suddenly makes a lot more sense.

A phone is a pretty bad proxy for what someone is actually doing with their attention

The first rule reduces platform power. The second potentially requires platforms to build even larger classification and moderation systems. I'm very much in favor of the first, but I don't think they belong under the same conceptual umbrella.

And the incentives are arguably worse than QA: you get little credit if the result replicates, and potentially a lot of grief if it doesn't. Novel results build careers. Checking them mostly doesn't

I'd be pretty wary of stretching wire fraud law to cover this. Prosecutors deciding whether scientific results are fraudulent sounds like it could create its own set of ugly incentives

At this point the surprising part isn't that another Ariely result has problems, it's that some of these datasets apparently went decades without this kind of basic scrutiny

I'm really sorry for your loss. There's not much to say to something like that except that I hope this treatment came too late for fewer people than it did for your sister

If molecular glues keep working like this, I wonder how many other "undruggable" targets are really just targets we've been trying to attack in the wrong way

Pancreatic cancer is one of those diseases where I've gotten used to reading depressing numbers, so seeing 13.2 months vs 6.7 months actually made me stop for a second

This is the classic difference between testing whether the software works and testing whether the system works

That does make the 10 million figure more meaningful but I still think it is doing a lot of work rhetorically

Without the total number of attempts, number of failures and ideally how those failures are distributed, it tells us almost nothing

I mostly agree but I think the bar for the software still has to be unusually high when the failure mode is "you are not allowed to enter the country you legally live in"... The human fallback is essential, but it shouldn't become the thing that makes a fundamentally unreliable system acceptable