3,255 karma · joined July 20, 2015
> in the meantime Microsoft fired their AI Ethics team
Actually that story turned out to be a nothingburger. Microsoft has greatly expanded their AI ethics initiative, so there are members embedded directly in product groups, and also expanded the greater Office of Responsible AI, responsible for ensuring they follow their "AI Principles."
The layoffs impacted fewer than 10 people on one, relatively old part of the overall AI ethics initiative... and I understand through insider sources they were actually folded into other parts of AI ethics anyway.
None of which invalidates your actual point, with which I agree.
In 20+ years on Linux I've never been able to make my variety of environments "just work“ this seamlessly on X. It's one of my favorite things about Wayland.
Now that screensharing "just works" I've run out of complaints about wayland, though. Just that it's turned me into an insufferable snob about the blurriness and incompatibility of the X apps I still have to use in xwayland.
But then, I stopped buying NVIDIA more than a decade ago, and linux compatibility is a priority for me in my device selection. So I'm sure there are plenty of other users out there with a less rosy experience than mine.
this post (same author) talks about how to he strategic and selective about your opt outs. Not everyone has to be an extremist!
Personally I moved from NVIDIA to quicksync exactly because I felt NVIDIA required too much babysitting. Driver compatibility, interactive mode updates, patches and patch compatibility... In comparison the intel driver felt like a cake walk. But you should definitely go with your own comfort level, especially for the price point!
Other 10th gen I5 users get multiple simultaneous 4k streams going at > 1x speed. Even with tone mapping it should only use like 45% CPU for a single stream.
Unfortunately the difference between software and hardware development cycles means that when you're on data formats that came to prominence in the last 2 years, even the newest hardware may not support it yet.
I just searched ebay for used celeron NUCs, and got lots of options like this one, with a 9th gen Celeron, for a hundred bucks. NB that the generation matters much more than the CPU, but some operations are still done by the CPU so if you can get an i3 or i5 it will make a marginal difference.
https://www.ebay.com/itm/115640978606?hash=item1aecbd4cae%3A...
- what's the source/destination quality?
- what's the container format? Quicksync doesn't support them all; even the newest gen is missing av1 and VP9 IIRC. But then, the same can be said for discrete GPUs.
- what filters are you applying. Many (most?) filters are CPU bound, which will kill you really quick.
But for supported formats without filtering, 12g Intel quicksync is regularly benchmarked at 4-6 streams.
One should always measure internet advice against one's actual use case. Personally I would argue that if you have a library of 8K video and more than 4 4K devices, you don't qualify as a "regular" home user. But of that's the case, hardware up my friend!
2) quicksync on a 10g or newer cpu is benchmarked at 4-6 4k simultaneous streams, but a lot depends on the details: container format in particular matters a lot, and so do your transcode options. Color matching for example, is only ever done on CPU. And IIRC AV1 and VP9 are not supported by quicksync (or older discrete GPUs).
That said, it's completely correct that you should measure all advice against your actual output device capabilities and source quality.
I don't use tailscale; I just port forward from my router to the k3s ingress IP, since that's fixed anyway. Accordingly k3s handles letsencrypt certificates. My router has a built in openvpn server.
I haven't moved to jellyfin... yet. Plex is super slick and runs nicely in the cluster. I've learned to keep it version locked though, to avoid regressions and unwanted new "features", which means jellyfin is only a matter of time.
I also run Nextcloud, and photoprism for my photo library.
Storage is on a built-from-scraps 16TB NAS which backs up to azure blob with duplicity, and longhorn for block-based storage (since lots of services nowadays prefer sqlite, which breaks on NFS). Yes I do need that space; I run an entertainment company and we store a LOT of video and audio. Not to mention media for plex!
I have many times considered moving most of this to a cloud system, but the cost is prohibitive. If anyone can find 13+TB of storage and transcode- and ML-capable hardware (for plex and photoprism face recognition) for less than $45/mo (my cost of electricity plus annual amortized hw cost), I'm interested.
TL;DR: pick up a NUC knockoff with an >8th gen celeron or i3; it'll handle anything a casual household can throw at it.
[0] https://forums.serverbuilds.net/t/guide-hardware-transcoding... this is far and away the most comprehensive hardware guide for plex servers.
Usually people complain - listening to the earlier version - that Gould plays fast and loose with the source. He makes opinionated choices throughout and even recomposes parts!
All that being said, there are no wrong answers with this stuff. If you prefer it more freely played than Gould 62, that's perfectly valid and no music expert (worth listening to) will bat an eye. This thread has some good suggestions of other interpreters you may like more.
[1] https://en.m.wikipedia.org/wiki/Planning_fallacy
[2] https://en.m.wikipedia.org/wiki/Reference_class_forecasting
Relevant papers:
Buehler, Roger; Dale Griffin; Michael Ross (1994). "Exploring the "planning fallacy": Why people underestimate their task completion times". Journal of Personality and Social Psychology. 67 (3)
Kahneman, Daniel; Tversky, Amos (1982). "Intuitive prediction: Biases and corrective procedures". In Kahneman, Daniel; Slovic, Paul; Tversky, Amos (eds.). Judgment Under Uncertainty: Heuristics and Biases. Science. Vol. 185.
Kahneman, Daniel; Tversky, Amos (1979). "Prospect Theory: An Analysis of Decision under Risk" (PDF). Econometrica. 47 (2)
Flyvbjerg, Bent (2006). "From Nobel Prize to Project Management: Getting Risks Right". Project Management Journal. 37 (3)
Hope this helps!
Estimation is actually well understood in academic project management. There is (Nobel Prize winning!) research about what, actually, are the problems inherent to estimation, and how to produce specific and accurate estimates despite them. This academic field is almost 50 years old, and no one who complains about estimation in blog posts or comments is aware of it.
Stop navel gazing and actually go READ about the subject. I know it's hard for us engineers to take in anything that doesn't come from StackExchange, but please try, BEFORE you write about your shitty experience with estimates and generalize to the entire problem space.
Here's what the research says:
- Humans are ALL bad at time estimation. Even the ones who consider themselves good at it, estimating tasks with which they are very familiar, "only" underestimate by 30% at best.
- humans are pretty good at estimating non-time attributes of work, even those with a direct correlation to time. Like effort, complexity, or "cups of coffee."
- if you estimate something with a time corellation (e.g. complexity) in a consistent way and measure the average throughput over time, you can very precisely and accurately estimate time to completion. This is the Law of Large Numbers, which is how casinos stay profitable when dealing with much more randomness than exists in software projects. It also makes your estimates include unexpected complexity, personal issues, illness, windows updates, etc. It's a statistical law.
- the accuracy of average time estimates is proportional to the time left on the project. It runs opposite to the uncertainty of distant features. I.e. this method does not predict how much you can build in a week; you're better off with your relatively intimate knowledge of the feature at that point in time and a gut check. Rather it predicts how much you will build over 12 weeks, with extraordinary accuracy.
- estimates are more.understandable when presented with a confidence interval, e.g. "the work as we understand it today will take 8 weeks, with a 95% confidence."
What I HAVEN'T seen in the research, but which is undoubtedly true, is that most teams violate these fundamentals and then complain that estimates are useless.
Asking your team to estimate in time units IS useless. Adding up those time estimates to create a long term plan is doubly useless. Cracking the whip on them when their estimates prove inaccurate is triply useless. And complaining about it on the Internet because you've never read any of the grown up work on the subject... well that's Hacker News.
This is the vision of the music industry we all would like to have. Hard work, pluckiness, some element of luck and "striking a chord" with the public and you, too, can be a success! It is a dangerous fallacy that drives many musicians into the trap the top comment is complaining about.
The reality is a lot uglier.
Ever since large music conglomerates conquered (integrated) the entire promotion/publishing/distribution/sales stack in the 70s/80s, becoming successful in music has been about fighting giants on their home turf. That's AFTER you beat the odds by getting a label A&R person (or the contemporary equivalent) to like you. The markets, gatekeepers, processes, and legal agreements are all set up to screw you, hard, because every one of those is run by a different face of the same multi-billion dollar company with serious lawyer and lobbying budgets, and you're a small entity with no legal budget (and probably no experience).
This was the very successful industry profit model for decades before the Internet, and the generations of executives since then have been focused on regaining that regulatory and legal positioning in the "new industry".
These two famous articles explain the structure of the pre/early internet music indistry very clearly: Steve Albini (famous music producer) wrote "the problem with music" in 1993[1], and Courtney Love "did the math" [2] of how an impossibly good record deal would work out for the artist in 2000. They're enetertaining and interesting reads. Every step along that journey, think of the modern music industry equivalent, and you'll find it is facing (or has already lost) exactly the battle the top comment is bemoaning with Spotify. The large players are working hard to achieve the vertical integration and control they once had, and it is working.
[1] https://mpg.org.uk/knowledge-bank/the-problem-with-music-by-...
[2] https://www.salon.com/2000/06/14/love_7/
I should note: there is no board of shadowy figures out to screw artists. This is natural dynamics of any labor market with a persistent glut. The buyer controls the price, and if it's a market where the sellers are passionate and driven to sell because they believe "it just takes hard work and luck", they will lower the price indefinitely... until they are even taking on predatory debt just to get their product out there.
Please stop spreading the myth that "it just takes hard work." It convinces too many good people to walk into an exploitative industry and thank their exploiters for the screwing. What it actually takes, is relentlessly beating giant megacorps at their own game. Most of the famous ones are actively getting screwed for the privilege of riding on a tour bus. There is a lot that music fans can do to help the musicians they love, and it starts with understanding the musician's fight.
Consider making a habit of using it when you are circumstantially visually disabled: in bright sunlight, while you can't be physically present at the phone (eg while doing dishes or folding laundry), etc.
Congress does outsource the Fed to the big banks, including Chase and Wells. [1] That's how the US Federal Reserve system works. Congress (through its appointed board of governors) sets goals, and the regional privately incorporated Federal Reserve Banks act towards those goals at their own discretion. The Regional Feds pay taxes, own property, etc just like any other private corporation. Principal shareholders are the biggest banks, including Chase, Wells, Goldman, Bank of America, etc. Apart from the power of the position, the banks profit by the mechanism of "printing money" in the US: Congress creates new bonds and treasury notes and gives them to the Feds, whose shareholders earn interest and fees on those bonds/T-bills. You would not be the first to worry that this is the fox guarding the hen house.
The SEC however, is a government organization... which outsources the bulk of its regulatory responsibilities to a private association of stock exchanges and broker-dealer banks (including Chase and Wells) called FINRA. [2] The principle of "self-regulation" has been a cornerstone of the SEC since its founding. Here, too, there is considerable debate about whether self-regulation is a good idea in the age of private exchanges, dark pools, rapidly-multiplying derivatives, etc.
I don't know what this does to your impression of the Fed or the SEC being fair and balanced. But you should definitely think about it every time you see discussion of the unequal impact of monetary policy, or a million dollar fine for an SEC violation worth billions.
[1] https://www.federalreserve.gov/faqs/about_14986.htm [2] https://www.investopedia.com/ask/answers/how-does-finra-diff...
But we should bear in mind that one of these women was a professional novelist, and the other, one of the first professional typists. Meaning they both had a much higher sensitivity to the use of language than the average Joe on the street. They were professionally more literate and intentional in the use of language than the average.
A better comparison would be modern professional writers or speakers. Or if you prefer, compare modern tiktok speech to those "cockneys" who sang the bawdy song.
I am not discounting the terrible effect on the English language of the kids these days, of course. We are all entitled to our expression of "ugh" like the magistrate in her story.
Did you know that Tchaikovsky struggled with depression? He once tried to commit suicide by walking into a frozen river. Do you know what happened?
He got cold feet.
It's funny 'cause it's true!
also deutsche post... what's so bad about it?
OR you handle it the way casinos do, using the law of large numbers[1] to make extremely accurate predictions over many consistently-produced values. There's a good reason why every single formal study on estimation, and every single agile framework has come up with abstracted estimation units and the same velocity calculations. A mature estimation process, run for a month or two, will give you medium- and long-term estimates that are at least within 5% of reality - that's perfectly useful for long-range planning. So the first half of the article is simply showing off that the author has never learned about professional estimation processes. Fair enough, most developers haven't.
Others have already pointed out in this thread that the purpose of long range planning is to indicate what we expect comes next. TFA acknowledges this as a good use of long-range planning.
There is one good remark in there, however:
> Instead of rewarding teams for delivering useful, profitable features, companies which worship conformance to plan reward developers for features on time and within budget; it doesn’t matter how useless or broken those features are.
As an organizational principle this is exactly correct. By all means use long-range plans, and adapt them as the situation changes. But don't build a rewards scheme around them!
"Why not just estimate hours anyway?"
Because humans are extremely bad at estimating time, which is borne out by studies many times over. A good overview is the original research on this, for which Kahneman et al won a Nobel prize (yes, the same Kahneman who would later go on to write HN favorite, "thinking fast and slow"). The broad stroke is, the very best time estimators in the very best circumstances only underestimate their time needs by 33%. The norm is more like 80%. They propose a few time estimation strategies to get around it, like "third party estimation" and "tripartite estimation". But the simplest approach (which emerged in later research) is to ask them to estimate "size" of task, and use statistical corellation to convert that to a number.
This last is hand wavy unless you're familiar with the law of large numbers, the law that makes casinos profitable. A casino cannot (without cheating) determine the outcome of a single roulette spin. But they can predict with extremely high certainty the aggregate outcome of a thousand spins. This is the same with your estimates. You can't predict the corellation to time of a single story point. As you pointed out, sometimes something that looked complicated turns out to be easy and vice-versa. But given a sufficient sample size (of estimates with a consistent corellation to time), you can predict with extreme accuracy the time for 1000 story points.
"Consistent corellation to time" is a bit of a PITA in a group, BTW. If you have developers do their own estimations individually, each one will have a different corellation to time. You would need a very large sample size to overcome that much variation. This is why so many systems encourage team estimation, so the consistency is dependent on the team dynamic, which is much more stable even when adding/removing engineers. But as I said, if it's the same person or team always writing your tasks, you can use their team dynamic instead, since their story size will be consistent.
FWIW by sufficient sample size, I mean after about 3 sprints (of any duration) you can make reasonable predictions. After 6 sprints you'll have confusing outliers, and after about 9 sprints it will be clear with some numerical weight to it.
Which brings up question 2, "the commitment ends up being a deadline". This is a human nature thing, you're right! But the problem isn't a mismatch between human nature and your estimate. The mismatch is between human nature and the uncertainty of reality. How you push to improve this is contextual to your org. In hard situations I reverse the statement of my estimate, to "if we set dec 15 as the deadline, there's a 5% chance we won't make it. What's our fallback?" Asking that question a lot is helpful. But there's no magic bullet to making leadership - or worse, people who are afraid of leadership - plan appropriately for uncertainty. The best you can do is expose the uncertainty as clearly as possible, and give lots of lead time for the times when they still run into conflict between deadline, resources, and scope. After that, it's the manager's job to "manage" things and decide which variable they will alter to break the conflict.
Put another way: reality is uncertain. When that uncertainty leads to a conflict between deadline, scope, and available resources - because that will happen sometimes per point 1 - only someone with deadline, scope, or hiring authority can solve it. That's (usually) not within your purview as a lead engineer. The best you can do is to 1) call out the uncertainty as clearly as you can, as early as you can, and 2) signal that conflict as early as you can, so those managers have maximum leeway. Abstracted estimation makes that possible. Guesses and hopes don't.
For many teams who get requests from external stakeholders in widely varied technical environments - ie consulting, often - estimation is functionally just a conversion process to a consistent unit. But you're absolutely right that for some teams the stories themselves are a good enough unit.
1) Your long range timelines come with a specific, quantified error margin. It's no longer "we'll be done by December 15," it's "we'll be done by December 15 with a 95% confidence, or January 15 with a 99% confidence." (Not to mention, those CIs are real and so your estimates have a very high degree of certainty)
2) your estimates have explicit conditions built in, most notably "based on our current understanding of the work." The door is already explicitly open to respond to feature requests, changes, or just new information with an estimate change.
3) Your estimate adjusts very quickly, so the conversations from #2 are also clear.
You can make commitments under those circumstances a lot more easily.
CI itself is just continually testing before continually merging. Doesn't matter where.
To take your example of the abortion issue in America, your steel man argument has to explain why an acorn is the same as an oak tree, or perhaps the difference between an acorn and a fertilized human egg. They have to account for the mass murder of fertility clinics and researchers throwing out thousands of fertilized embryos per day.
Or you could choose to accept their philosophical argument about the start of life as a given, and discuss the consequences of it. E.g. you have a right to kill another human in self defence, unless the person threatening your life is inside a womb. Or the conditions under which you have a right to another person's life and body. Is there an innocence test we could do for organ transplant recipients, where the top end of the scale gives you a right?
Lots of room for good faith discussion there, even on immaterial subjects.