HNHacker News
TopNewBestAskShowJobs

Olreich

711 karma · joined July 9, 2012

submissionscomments
Olreich··on Embeddings: What they are and why they matter
The “secret sauce” is Word2Vec. Take a snapshot of all the text you can find on the internet, then with a sliding context window, vectorize each word based on what words are around it. The core assumption is that words with similar context have similar meaning. How you decide which components represent which words in the context is unclear, but it looks like we’re doing some kind of ML training to convince a computer to decide for us. Here’s a paper about the technique which might help: https://arxiv.org/pdf/1301.3781.pdf

Once all the words have vectors you can assume that there’s meaning in there and move on to trying to math these vectors against each other to find interesting correlations. It looks like the scoring for the initial training is based on making the vectors computable in various ways, so you can likely come up with a comparability criteria different than the papers use and get a more useful vectorization for your own purposes. Seems like cosine similarity is good enough for most things though.

Olreich··on Cooler screens
VC money going up in flames has a tendency to kill their other invested-in companies through increased pressure to become the unicorn that digs them out of the hole. Hopefully this VC doesn’t invest heavily in anything actually good.
Olreich··on We have used too many levels of abstractions
Most likely, your framework Is productive for you not because it’s the best framework available, but because you wrote it and know the ins and outs of it and it matches up to how you think. This isn’t too day it’s not likely better than other frameworks in at least some ways, but familiarity begets expertise which begets productivity.

Many experts in drywall installation before drywall screws were popularized swear that screws are slower and worse. However, an expert nailer and an expert screwer both complete jobs just as quickly (if not faster for the screw-adherent), and just as well.

Olreich··on Self-Checkout Is a Failed Experiment
Grocery needs a little more tech to be perfectly convenient. Best I've seen are handheld scanners with electronic scales in the produce area to let you get barcodes for all your produce. The handheld scanner to let you scan and pack away from the checkout area. When you get to the checkout area, you generate a barcode that loads all the items into the self-checkout and pay.
Olreich··on Self-Checkout Is a Failed Experiment
I'm pretty sure the various managers can play with that experience. I've had Walmart self-checkout complain if I moved too quickly or removed items from the scale after scanning them. They've also complained when I had a second person handing me items from the cart or if I was moving too quickly.
Olreich··on Google Cloud Spanner is now half the cost of Amazon DynamoDB
Lambda is the most expensive thing you can do if you have more than 25% utilization. Fargate is extremely close to modern on-demand EC2 pricing (m7a family).
Olreich··on Google Cloud Spanner is now half the cost of Amazon DynamoDB
Postgres. NoSQL databases are specialized databases. They are best-in-class at some things, but generally that specialization came at great cost to their other options. DynamoDB is an amazing key-value store, but is severely limited at everything else. Elasticsearch is an amazing for search and analytics, but is severely limited at everything else. Other specialized databases that are SQL-full are also great at what they do, like Spark is a columnar database that has amazing capabilities for massive datasets where you need lots of cross-joins, but that severely limits it's ability to act in a lot of roles, because they traded latency for throughput and horizontal scalability, and you're restricted in what you can do with it.

The super-power of Postgres is that it supports everything. It's a best-in-class relational database, but it's also a decent key-value store, it's a decent full-text search engine, it's a decent vector database, it's a decent analytics engine. So if there's a chance you want to do something else, Postgres can act as a one-stop-shop and doesn't suck at anything but horizontal scaling. With partitioning improving, you can deal with that pretty well.

If you're writing fresh, there is basically no reason not to use Postgres to start with. It's only when you already know your scale won't work with Postgres that you should reach for a specialized database. And if you think you know because of published wisdom, I'd recommend you set up your own little benchmark, generate the volume of data you want to support, and then query it with Postgres and see if that is fast enough for you. It probably will be.

Olreich··on I'm Loud Quitting My Job Due to RTO Policy, Bad Company Culture
Do everything the in-person one does, and ensure there’s a team meet up sometime in their first month, with a cultural emphasis on getting to know the new folks.

It’s fundamentally difficult to onboard. What makes “remote” onboarding hard is the same things that makes in-person hard: you basically need to hover over the new hire and ensure that they aren’t getting stuck on anything and that they are getting comfortable with asking for help and that existing team members are helping them.

The thing that makes remote more difficult is that it’s harder for a team lead to see if the new hire is getting helped: passive surveillance is harder when you aren’t sitting nearby. I think you can easily mitigate that by being a good leader and making sure to have routine check ins throughout the onboarding, starting at twice or three times a day for the first couple of days and slowing down over the first couple of weeks until you hit a weekly cadence or the new hire is thriving.

Olreich··on We Can't Compete with AI Girlfriends
Ah, that article... Here's the paper:

https://amj.kma.re.kr/cgi/viewcontent.cgi?article=1597&conte...

With a sample size of 29 women and 0 men, they presented 100 tinder profiles and asked them if they would swipe right or left. The sample size was all Korean and excluded anyone that looked "foreign" with no further explanation of what that meant. The women were likely to swipe right only 20-30% of the time, however, they go on to explain that while some 40% of the profiles were rejected by most participants, there was only 1 profile that was accepted by over 80% of participants.

At worst, this indicates that 40% of men are generally undesirable to a large part of the female population. But the researchers asked the participants why they rejected these most-rejected profiles:

> To better understand possible reasons for rejection we have additionally analyzed the profiles that were rejected the most (by more than 90% of the participants). In the majority of those pictures (61.5%), a man’s face could not be seen clearly, which supports previous findings. When a face can be seen, there was weird or unfriendly facial expression (60%) or excessive photo editing (20%), which were also mentioned during the qualitative stage as the reasons for swiping left.

So it sounds like this study is finding that profile presentation is much more of a factor for universal rejection than other factors such as aesthetics, fitness, status, etc.

This study has a small sample size, doesn't make the claims that the articles all make, and doesn't contain any data to support those articles. The study also assumes that all men will swipe right on all women in the study. When they run an experiment where they put 100 women and 100 men into the same study and show them each-others tinder profiles, maybe you'd be able to make the claim that 20% of men are dating 80% of men, but that's not what the study winds up saying. They say that women are rejecting ~80% of possible matches, but in aggregate, only about 30% of profiles had fewer than 5 of 29 swipe rights, and only 1% of the profiles were rejected by all participants and 1% of profiles were accepted by all participants.

Olreich··on We Can't Compete with AI Girlfriends
They don't even have to sell the data, just get the model to manipulate the user into buying things for their AI companion. Virtual clothes, virtual locations, upgrades for their personality, etc.
Olreich··on You Don't Need UUID
And for UUIDs, Base58 and Base64 both require 22 characters, so no density loss! Even Base32 only takes you up to 26 characters.
Olreich··on How to speed up the Rust compiler
> Compilers involve giant tree structures, with lots of allocations, pointer chasing, and data-dependent multi-way branches.

Tree structures and data-dependent multi-way branches do seem inevitable to me, but pointer chasing and allocations don't seem inevitable at all. There are lots of ways to avoid allocations or make allocations practically free. Similarly there are lots of ways to make pointer chasing less prevalent, especially getting down to single-indirection.

rustc may not be able to fix their allocation or pointer chasing problems, but this is not endemic to the problem of compiling.

Olreich··on Factorio: Space Age
Factorio inspired a lot of factory games you're seeing: Mindustry, Shapez.io, Satisfactory, Techtonica, Dyson Sphere Program, etc.

One of the devs have stated that they were inspired to create Factorio by the IndustrialCraft and BuildCraft mods for Minecraft, and I suspect that things like Infinifactory were also inspirations.

Olreich··on Toyota has been developing a solid-state battery for EVs with a range of 745mi
If the only ICE vehicles left are in the outback, we're doing okay.
Olreich··on Solve your big problems by solving your real problem
The vast majority of the time the reason your ideas are dismissed are because people don't see the value in them. Oftentimes, the easiest way to show the value is to build a vertical slice of that idea and show it to folks.

In highly-regulated environments, this may not be an option, but for the vast majority of engineers, testing out different wordings, getting more seniority, or making allies will take a lot longer than producing a small prototype that can show the value.

I think the analysis of why people don't like your idea should be saved for post-prototyping. The prototype should be a reasonable amount of effort, no 6 month projects ignoring the value you're being asked to produce, but more like a 1 or 2 day project when the workload is low.

A good manager will encourage this sort of prototyping and not force you do it in secret. If you are having to build your prototypes in secret because people will prevent you from doing them, then you should be job hunting, because the place you're in doesn't trust you enough for you to thrive.

Olreich··on I don't need your query language
SQL selection works perfectly on tables in poor normal forms. If you have the columns you need to query pre-joined into the table you’re querying, you just skip the joins. Updates are what gets fun if you don’t have normal form.
Olreich··on I don't need your query language
`3w` is a command all its own. There’s an implicit move command baked in. Moving gets pretty clunky if you don’t have first order movement (adding a specifier or something to clear selection). You can mimic this by mapping a single button to <C-v> and disabling all movement commands in normal mode. It’s rough, but may be learnable.

I think something that might work better is adding a “commit” signal to operations. So you type `d3w` and the editor highlights the next three words with a strike through or red or whatnot. Then you can hit enter to commit the delete or escape to cancel it (cursor resets to start position, highlight goes away).

Olreich··on First impressions: Yes, Apple Vision Pro works and yes, it’s good
I kinda prefer controllers. Tapping my fingers together and waving them around with even less feedback will mean the gesture recognition needs to be extremely far beyond anything we’ve had before, or is going to feel awful to use. Any time you’re in a game and are meant to be holding an object, having something to hold helps a lot, and most games want that.

If we can turn everyday objects into the thing you hold, that’d be great, but I’m not sure it’s feasible yet. We’ll see if anyone tries.

Olreich··on First impressions: Yes, Apple Vision Pro works and yes, it’s good
Unless there’s some radio wave danger, I don’t see how headsets could be harmful. Your eyes are not actually focusing on a close screen, they’re focused however far away the objects appear to be. Is there a danger to letting your eyes be unfocused and placing an iPhone right next to them (letting the screen be blurry, not trying to focus on it).

Maybe it’s the inner ear that is going to get hurt after prolonged disorientation, but astronauts should surely have that worse and I haven’t heard of any studies about it. Maybe something will show up about it.

I don’t see an obvious mechanism where this is harmful and VR headsets have been around long enough that this should be a well-understood safety space.

Olreich··on Automatic emergency braking should become mandatory, feds say
Unfortunately, unless the frustrated driver is allowed to drive unimpeded at 100mph+, they will continue to be unsafe and make others unsafe. If there is enough traffic for you to be the slow car on the road, then that unsafe driver will be causing safety problems for all of you. Bottling them up in a line of cars going the same speed may be the safest course of action you can take for everyone.
Olreich··on Automatic emergency braking should become mandatory, feds say
They also already have that computer and most of them have the sensors too. The reason so many cars come with lots of these features is that the base electronic package that the maker uses just comes with it, so the R&D is already paid for and removing it would be more labor than the components cost. After all, it’s a $5 radar, a $3 chip, a $1 camera, and $0.02 of wiring.
Olreich··on Automatic emergency braking should become mandatory, feds say
Those would be good stats to mandate sharing with the proposed rules. Would likely cause them to rework the system if they started losing sales over it.
Olreich··on Ask HN: Is Rust worth investing in as a reliable C-alt given its recent saga?
Rc<T> is a chunk of memory you can only access through special versions of malloc and free. These don’t actually allocate new memory with the OS though, they just give you pointers to the memory and tell the memory management system that you’re done using it. Once everyone who has allocated that memory calls free, then the memory is actually given back to the OS. Box<T> is the same, except if you have multiple people try to call the special malloc it will fail to compile.
Olreich··on The Absurdity Of Stacks
Can you provide a concrete example of an algorithm that’s easier to express with recursion?
Olreich··on The Absurdity Of Stacks
I’ve hit the cliff once. That’s why it’s last on the list. It’s not very often, but it is possible to hit and with the other reasons and the ease of avoiding it make it easy to justify the very small (if any) amount of work to avoid recursion most of the time.
Olreich··on The Absurdity Of Stacks
The reasons I avoid recursion don’t align with saving memory. I avoid recursion primarily to make it easier to read the code. Recursion requires you to keep track of the implicit stack, an explicit stack makes that easier for me.

Secondarily, I don’t want to deal the administrative work the author says I should do in a recursive situation, as allocating the stack size would happen far away from my algorithm, making it another thing to go hunt down when my memory usage grows, instead of a local memory allocation.

And thirdly, there can be a cost to calling functions that has been significant for me. I have a much better feel for how a loop will perform than a recursive call tree. Maybe the optimizer will do something with it, maybe it won’t. Maybe I’ll make a minor change and the optimizer will stop working and I’m getting a tenth the speed I should.

I will prototype algorithms recursively sometimes, but usually it’s very simple to convert them back into a loop with explicit stack, and many times that stack isn’t even needed. So I hardly leave it in.

Olreich··on Can America Go Car-Free? Gen Z Hopes So
Even in the cities that are touted as most walkable, cars have access to almost all destinations. It may be more expensive to own a vehicle, but it’s not impossible. Feel free to keep driving, but you can reduce congestion for yourself by supporting high quality transit and walkable initiatives, because they will encourage more people to choose those modes of travel.
Olreich··on The Air Force has loosened its restrictions on body fat for new recruits
This system change improves accuracy without increasing cost or complexity. They already are doing these measurements, just adjusting the threshold based on height. The previous system made it so very short people could be very fat so long as they could do the other components, but very tall people had to be rail thin or else automatically fail the test.
Olreich··on Culture Is Stuck
It only feels homogenous. The algorithm is just giving you the same stuff you’re willing to interact with because it’s working. If you interact or look for different things, the recommendation algorithm will give you that instead.
Olreich··on Culture Is Stuck
The author’s culture may be stuck, but that’s a them problem, not an algorithm or consolidation problem. You can get to more culture than ever, but if you’re content with previous cultural content, they will happily keep selling it to you.

Algorithmic discovery requires you to do a little more work since there isn’t a single 18-64 demographic that’s the primary marketing demographic. Now it’s 24-27 year old woman with adhd and loves a list of 20 different things that gets targeting for advertising. The algorithms will give you what you ask for, so ask for different stuff.

← PreviousPage 3 of 9Next →