sheikheddy

It’s not about gating the new feature behind a flight, but not wanting to push a new build out to the furthest deployment ring before you need to.

1) Any particular reasoning behind estimating OpenAI’s margins are 60%?

2) How much does human preference diverge from benchmark scores in your experience?

3) Do woodpeckers stop attacking houses when it’s winter in Alberta?

Abdurraheem dot ali at microsoft dot com

https://arxiv.org/abs/1411.1127

For privacy and compliance reasons (read: “oh boy wouldn’t wanna get sued, eh?” reasons) we actually don’t snoop into the message body much. Hooray, good job on not doing the maximally big brother thing for once, MS!

My hot take is that this prolly won’t last because every org descends to doing a creepy level of data collection eventually so I have a textbook on privacy preserving ML downloaded for when we join the “surveillance but we found a way to make it technically legal” squad. We haven’t done that yet though.

What do you mean by tiers, exactly?

You didn’t make any mistakes encoding, I just screwed up my decoding, it happens :D

Thanks for clarifying, ok, hmm… that seems hard to do if you can check the IP block by using a subnet mask but the specific IP isn’t resolved until later in mailflow. It might not actually work like that in… ProtocolFilterHub? I always get this mixed up, wait… I think this might be something that we are already working on. And have been working on for a while, wow. Looks hairy. It’s stuck since the guy working on it transferred to another team, and no one picked it up, but some PM noticed before I did and put it up for vote in semester planning. Always creepy to see engineers get referred to as “resources”.

Huh? This shouldn’t be possible in principle? Don’t quote me on that though, I wish I’d paid more attention to my notes but they’re a mess and haven’t kept up with newer changes, if they were accurate at all in the first place. I’d submit an escalation so support can look into it.

This is correct but not the only reason it could be happening.

I don’t. I have slept in the daytime ever since covid and actually got a move to the east coast approved as a health accommodation after I started routinely missing important afternoon meetings due to my incurable insomnia (mornings are easy when you stay up all night). I still struggle with it, especially since it’s not a consistent offset to my circadian rhythm. There’s data I’ve collected but it’s hard to fit a simple function to it— it’s not like I’m on a 26 hour schedule either. This isn’t due to trauma or addiction, my brain is just an outlier in many dimensions and this is one of them.

Hmm, oh wow, occasionally I’m reminded that if I flipped sides to run phishing campaigns I’d be totally unstoppable.

There isn’t a quick way, by design. You need to wait a minimum period and meet some predicates, and the organized scammers already know what the period is via empirical testing but I’m not comfortable disclosing details of those predicates for disorganized scammers to use. More so because I’d definitely get into trouble for it than due to any belief in security via obscurity. Cushy job makes you risk averse.

Since I can’t share any of the tricks, some general advice— the main thing that matters is a long track record of good behavior. You can end up in a vicious cycle where you fight the system when it punishes you and then it doubles down on the beatings— this is bizarre and kafkaesque and happens all the time. What you want is for there to be two-way communication, if it’s unbalanced with traffic being broadcast but no one engaging with it, that’s going to be cracked down on sooner than if recipients reply.

Aww man, not joking this actually breaks my heart, something about the way you wrote it makes it sink in how much we’ve failed you. I’m angry at how much of your time we’ve wasted and this experience is completely unacceptable.

…I think this is just a systemic issue beyond my ability to comprehend, let alone solve, and— I hope I’m wrong about this but honestly when I look ahead it seems the future is only going to get worse for people like you. Which I wish I could phrase in a way that was more kind and respectful, it’s not what anyone wants, these unthinking scars inflicted on email as a medium.

But what I can do is make sure that it’s not worse for you, specifically. If I was perfect I’d attack this rot at its core, but I’m not, so I’ll just solve the problem in front of me even though I know it doesn’t scale and hope God forgives me. Get in touch with me directly and I’ll figure out how to make sure you don’t have to jump through those hurdles again.

Yo I’m not even gonna apologize about this, it would be so wack if we didn’t do that:

a) if a mail sever looks like it’s gonna send spam, then you gotta block it. I personally have philosophical hang ups about this, like it’d be wrong to sentence someone to prison for crimes they didn’t commit just because a system added up some points and made a prediction with high confidence, but in real life, you absolutely need to be proactive. b) there is literally no way to do this that wont immediately get abused. Trust me we’ve tried. We make it nearly impossible to get unlocked on purpose because if it was easy, then it’d be like 1 innocent person using it and 99 attackers due to the adversarial incentive structures.

Now ofc there’s more nuance here, we really do want to get it wrong less often, and you do pay us so it’s not fair to blame it all on the bad guys, so I’m grateful for the feedback but I think you should give me even more detailed feedback since there’s not much I can do except give a vague high level explanation unless you help me by being specific.

Oh man I think around 2 years ago there was a 3x spike in Europe outbound spam and the fraud team had to disable like 200k+ tenants from some shady cloud VPS. We didn’t have a long term plan for the abuse back then besides playing whack-a-mole, and if we have one now, I haven’t heard of it.

Dumb question, but wtf is the solution even? I’m confused about what you expect us to do. I haven’t thought about the problem much so I might be missing some obvious Pareto improvement.

My guess is that the effectiveness issue isn’t actually due to SNDS and is probably related to sender reputation having famously high false positive rates. I read a paper a while back which introduced a different algorithm with tighter bounds on regret, I didn’t really understand it tbh, but I can implement it behind a flight and run a data study to see if it works better. The problem is that most graph based stuff doesn’t scale super well because of something-something complexity classes. I think the lady who architected it 5 years ago didn’t do a great job and there’s a bunch of arbitrary config stuff which was put as a placeholder and then became enshrined in stone… but the guy maintaining it rn is really smart so I’ll have him review my half-assed PR when he’s back next week (and idk how long it’ll take to finish the other half of it, shit never ships around here).

About the calendar widget thing… man am I glad I our team doesn’t own that. No one ever touches legacy stuff cause they’re afraid it’ll break or no one will update but the trick is to file it as an accessibility bug since that gets someone to actually prioritize it since it shows up in reports that the execs read. But dude good luck getting that off the backlog, the one engineer we have who is good at UX stuff (i.e, can code with both quality and velocity instead of just one) has her hands full as is.

I work on Microsoft's anti-spam team, AMA!

Almost certainly yes.

top of org charts?

Used to be available but Google banned it: https://github.com/abhishekkrthakur/colabcode/issues/109

People reading this for inspiration on what stack to use should keep in mind that the best stack is the one they'll actually use.

LinkedIn is probably down due to high traffic from all the people who got laid off.

Really like the "Draft in Progress" bar to delimit the polished and unpolished sections.

Wrote a paper about him for a history class in college, nice to see it pop up on HN.

Intuitively, we only mask if the current value of the online network is outside of the trust region and the sign of the TD-error points away from the trust region.

Seems like this is where most of the improvement comes from. Anyone have an analogy to help explain why this works?

$100k job for 10 years, or $200k for 5 years.

Those animated interactive explanations remind me of http://explorabl.es/.

I'm really interested in creating some on my own, where can I find source code for guides like https://github.com/ncase/trust?

I will also not be looking at browsers that only work on Microsoft Windows or macOS, even if they are Open Source. Both Microsoft Windows and macOS are highly controversial and completely untrustworthy operating systems.

Is this serious? I think we have different definitions for controversial and untrustworthy.

Edit: Checked out a couple of other articles, it seems that this is not out of the ordinary for the author, after calibration the tone falls within what I'd expect.

The way progress is made on tall and wide problems is different. On tall problems, you make progress "climbing", and don't move forward much until you reach the peak, and then it's behind you all at once. For wide problems, you can make a little headway "forward" in little increments, which is easier to track.

I think a lot of wide problems could be approached as tall problems, but the incentive structure in some places is set up to measure how much "forward" movement is made each day, so you're punished for "climbing" since people think that you're staying still instead of doing work.

Yeah, I clicked expecting something like wizardzines, but this isn't as high quality as the stuff from Julia Evans.

I'm fairly certain that you could probably write and publish a decent quality paper yourself. But why would you, when industry is so much more lucrative? There are many competent people in CS Academia, but their motivations tend to be quite different from the average engineer, and they don't have access to the same resources or expertise that tech giants or startups do (outside of internships and fellowships that are commonplace for visiting researchers who spend a season or two working at a lab)

Try reading some of the older, classic papers. I think those have taught me a lot about the fundamentals.