I understand people being paranoid about this, but just understand that the people who will judge you for spelling errors will always dwarf the ones who believe they are capable of sniffing LLMs out...
1,231 karma · joined July 3, 2015
Despite all the complaints about AI slop, there is something ironic about the fact that simply being exposed to it might be a net positive influence for most of society. Discord often begins from the simplest of communication errors after all...
My work insurance seems to change all the time, and while going through GoodRx doesn't count towards my deductible, I prefer the price stability. Not fun when I'm randomly told it's $120 now at the pharmacy because my insurance doesn't cover it now for some fucking inane reason. A few phone calls can often resolve it, but it's the last thing I want to do when I'm a day away from withdrawals kicking in. Even more absurd is this is basically guaranteed to happen more than once a year, THERE IS ONLY 12 MONTHS IN A YEAR!
Hospitals were among the first to get "computers", I'm talking the big mainframes and such that used to be popular in big institutions & universities. On these systems many hospitals each individually hired programmers to construct custom databases for their record keeping. While most have by now have transitioned into a more standardized structure, like HL7, the original sin has carried forward enormously bizarre data structures that make you wonder if the designers were deliberately trying to sabotage the possibility of good software in the industry. I can't think of a better example of why you should never design by committee.
Yet in parallel to all this, capturing medical data is already hard. Doctors are most comfortable just writing notes freehand, recording the patients current state, notable observations, treatments and so on. When modeling this it becomes very tricky because you basically need a proper medical background and be a good at data modeling / programming. This kind of person is basically a unicorn in the industry everyone wants but can never get.
Consider, just for a moment, all the complexities that come with dealing with the thousands of different units and their conversions within the industry. Some doctors don't even use the same units for certain measurements, entirely out of personal preference. Then remember that measurements are the easiest part of the system to model, even what should be the simplest part of the entire thing is hard. Also yes, you will have to re-write all this from scratch, there is no special library or open source software to help. Everytime someone makes tools for this they keep it proprietary.
But that's just the tip of the iceberg, to really get an idea of what I mean, just look at HL7. It's basically a data format that is like a cursed csv with about 5 layers of deliminators for nested entries, since all hospitals like to be super special, the specification tries to be "flexible", so what exactly these characters are is not actually standardized! It wasn't enough for HL7 to just be a data model, they needed to violate a few OSI layers and interlace it with the transport protocol too!
So in essence you must establish a bizarre handshake on top of tcp to learn what the hospitals super special configuration of the standard is, the very syntax itself! Worse, 90% of it is the same for all hospitals but the 10% that isn't is entirely unpredictable!
Then you have the actual data model itself, like demographics, lab records and so on. They change the specification every few years! You need to support it all since this committee of monsters don't seem to care much about the migration path! All the changes they make seem pretty arbitrary to me but what do I know?
I'm still only scraping the surface here but my exposure has been limited to what I do, which was processing all this from the perspective of a medical device that only needed to deal with a subset. When I imagine the struggle one would have with actually dealing with the entire thing holistically I feel empathy and a desire to never have their job.
It's like building a house on top of an active volcano. Any illusion I had that my medical records could be used for anything other than basic notes for another doctor to read have long since shattered, because clearly that's how all of this mess is actually being used in practice.
Oh and don't forget HIPPA! Even when you roll up your sleeves and try to fix the problem, you learn you aren't even allowed to thanks to the governments overbearing regulations against using medical data for things that could help society. Wish they just made it a crime for insurance companies to use instead of whatever this is.
The fact any of this works at all is a fucking miracle honestly.
It obviously depends on your use case but I find most time operations are often easier to deal with as pure mathematical operations. I will reach for Date when I need to deal with weird stuff like the day of the week but having burned through so many footguns with what I consider the worst DateTime API I have ever dealt with in any language, I will do anything I can to avoid using it.
Some libraries like date-fns can help here, but can't always bring in new deps easily. I hope the new temporal API fixes things.
> lumping together arrests for stalking, incitement to violence and other forms of harassment to produce a big scary number
But that's exactly the problem - the UK defines "incitement" and "harassment" so broadly that ordinary political speech becomes criminal:
UK "Harassment" includes:
- Misgendering someone online
- Posting offensive jokes
- Retweeting protest footage
- Criticizing immigration policy "grossly"
UK "Incitement" includes:
- Lucy Connolly's Facebook post (31 months)
- Jordan Parlour's "every man and their dog should smash [hotel] up" (20 months)
- Tyler Kay's "set fire to all the hotels" retweet (38 months)
NONE of these would meet Brandenburg's standard in the US. They lack:
- Directed at specific individuals
- Imminent timeframe
- Likelihood of producing immediate action
> if the stated number is true, there should be thousands of examples every year
There ARE thousands. In 2023:
- 3,537 arrested for online speech
- 1,991 convicted under Section 127 Communications Act
- Hundreds more under Public Order Act
You don't hear about most because "UK citizen arrested for offensive tweet" stopped being newsworthy years ago.
You're using the word "incitement" to equate UK thought policing with legitimate US restrictions on speech that creates immediate danger. That's like defending China's censorship because "every country bans fraud."
The definitions matter. The UK criminalizes hurt feelings. The US criminalizes immediate threats to public safety.
Not true. The US has a much higher bar for prosecuting speech than the UK.
Brandenburg v. Ohio (1969) - 395 U.S. 444
- https://www.courtlistener.com/opinion/107965/brandenburg-v-o...
- Speech must be "directed to inciting or producing imminent lawless action"
- AND "likely to incite or produce such action"
- General statements like "burn them all" typically fail both prongs
The "imminent" requirement is key. Connolly's Facebook post lacked:
- Specific targets or locations
- Timeframe for action
- Direct instructions to specific individuals
- Any indication people were prepared to act on her words immediately
Here are cases with far more explicit threats that were protected:
United States v. Bagdasarian (2009)
- https://www.courtlistener.com/opinion/221261/united-states-v...
- Citation: 652 F.3d 1113 (9th Cir. 2011)
- Posted that Obama "will have a 50 cal in the head" with racial slurs
- Result: Conviction reversed as crude political statement, not true threat
United States v. Turner (2013)
- https://www.courtlistener.com/opinion/904120/united-states-v...
- Citation: 720 F.3d 411 (2d Cir. 2013)
- Posted that three federal judges "deserve to die" with their photos and addresses
- Result: Conviction overturned as protected political hyperbole
Connolly's "set fire to all the hotels" would likely be viewed as angry hyperbole in the United States, not meeting Brandenburg's strict standard.
The distinction: The US prosecutes actual incitement (directing a mob to attack a building RIGHT NOW). The UK prosecutes offensive speech that merely might inspire someone, somewhere, someday. Your Brandenburg citation actually proves this difference rather than refutes it.
You want thousands of examples? Check Twitter during any US political crisis - they're not prosecuted precisely because Brandenburg protects them.
I see Meritocracy as a deterring force against corruption so I'm sensing some semantic discord here. A nation that starts to rot will be taken advantage of by external entities which will result in a drop on QoL. While GDP and such can somewhat approximate national power, they seem a bit tangential to the discussion imo, the point is rot invites parasites.
> What I am seeing is that the value system behind meritocracy is too close to my liking to self-appointed superiority. I am rich and powerful because I am the smartest, fastest, strongest, and worked the hardest. No one else deserves my position of power unless they too are rich, and if they are not rich, they are not smart and don't merit such position. The idea of merit I think can be subterfuged, old Egyptian leaders were thought to be Gods, so it was deemed they were the only ones that could merit to rule.
But that's the opposite of Meritocracy? Or rather, it's like you are confusing the cause and effort perhaps? It's an oppositional force to the default nepotistic hereditary nobility type systems, which will naturally emerge in every system that does not account for it, these are absolutes. Caveat being that the means of avoiding it are nuanced ofc.
The point is you design systems where positions of power are selected on (best effort) neutral criteria that at minimum narrows the candidate pool down in a way that the preserves a degree of instability, and through which helps prevent calcification of power structures. With a Meritocracy the criteria is via a demonstration of merit/qualifications/evidence you are the most capable for the position.
It does not give someone license to act as if their wealth justifies their position, that's just a simple narcissist. Meritocracy is just a good general principle to follow when designing the process of selection, it's not some complex ideology. Having power never implies you earned it, your merits do, and society is the judge of what exactly those merits are.
You also focus on wealth a lot so I'm wondering if you are primarily pushing back on the thought that having wealth qualifies as intellectual merit? Because if so I very much agree, but I also rarely see this from anyone but narcissists who don't even need a reason to think that in the first place, their conclusion came first. But maybe this is just a blind spot for me.
Money is power, and our modern economic system has made the liquidation of wealth into money easier than ever. It has helped shift power struggles from violent to competitive and allowed some innovative types of tax policy to become possible. But that doesn't make our economy a Meritocracy, what we have is closer to natural selection, where any snake can kill a lion and so on. The perks of capitalism are entirely from it's ability to parry these inevitable power struggles into something society can gain a net benefit from through the innovation that arises from healthy competition. It's impossible to eliminate the power struggles themselves though, those are human nature.
I can see how the concepts can be confused but fundamentally it's a brain (skills) vs brawn (power) thing. A meritocracy advocates for selecting for the most skilled not the most powerful. It's only practical to enforce on a institutional level though.
The key here is that while meritocracy is championed as a means of finding the best, it in reality functions as a system to keep out the worst. You want harness the ambitions in people, even if not everyone's ambitions can actually be met, and you want to mitigate the harms of nepotism, even when eliminating it entirely is impossible.
So the difference between qualifications and merit evaluation are moot from my perspective, the question you need to ask is if whatever selection criteria you prefer is vulnerable to ladder kicking. If you preferred way is more vulnerable than the current system then you are putting the cart in front of the horse.
Also to make my position clear, I can't tell either way in regards to what you have suggested. As far as I was aware, we already select based on qualifications, so it's unclear to me what the exact change you are proposing is.
For games like dwarf fortress, I think embeddings are the easiest to integrate, they can handle aspects of subjectivity we can't be easily simulated. Even simple things like a sentiment meter that measures how screwed up the world is right now would be a fun and meaningful addition. For LLMs the main value could be just having it clean up and summarize generated text, constructing a more coherent real-time narrative on what's happening in battle for example. Dwarf Fortress has so much depth to it, but comprehending it all can be overwhelming at times.
More primitive, but this is technically how AI in Bethesda games work, has this background simulation for NPCs going on when out of sight. Think it's mainly focused on movement patterns though.
Dwarf fortress as well has it's world simulation setup such that the current tile is higher precision than the rest of the world. In fact I think plugging the output of this simulation would work perfectly as context for an LLM (or Embedding?) depending on what you wanted to accomplish...
Tell me this, which is more productive in an open discussion? My idle rumination? Or the lazy dismissal that lacks any substantive contribution to the ongoing discussion? I can't help but agree with your distaste for "dorkisms".
To me evolutionary pressure isn't an on/off thing, it's like a signal in the noise. It's a vector with a direction and magnitude, facing varying levels of environmental resistance.
To be more specific, if there was enough "magnitude", evolution could potentially arrive at a perfect CRC. But the "resistance" requires a "magnitude" higher than evolution is willing to pay to reach that perfection. Likely in part due to the implicit complexity slope. Considering the systemic malfunction mutation can cause, one might assume this magnitude would be higher than it currently is. However, this is entirely speculation, and not falsifiable.
So when I think of evolutionary pressure I'm considering it as a component of the final vector, where a biologist might more pragmatically consider the total sum of vectors instead. This way of thinking is likely more productive for what they are doing.
As for evolution always selecting for the individual over the group, I'm surprised this is controversial when it's so obviously happening? If that was true how could multicellular organisms even exist? I'm very much not an expert on any of this, but this sounds like perhaps an over focus on DNA itself and not evolution as a whole, but maybe I'm misunderstanding something?
Ignoring whether or not it would even be possible, a perfect CRC is antithetical to evolution itself, wiggle room for mutation must always be possible, but too much mutation gives you cancer and systemic malfunction. So you end up with these bizarre processes that allow just the right amount of imperfection.
With sperm specifically it ends up closer to a signature check than just a CRC, if the sperm doesn't exhibit behavior that falls under a certain umbrella of expected behaviors, it's rejected by the surrounding environment. The difficulty to comprehend it could even be a feature of the process in many respects, especially when you consider everything in this realm risks getting "hacked" if precautions are not put into place.
So when I see huge numbers like this, I see it as an indirect measure of the precision of the overall process. To put it another way, it's like brute forcing a password you don't know, but happen to have a lot of hints to (since obviously, we are all still the same species at the end of the day).
I don't see LLMs as much different really, our jobs becoming easier just means there's more things we can do now and with more capabilities comes more demand. Not right away of course.
Our modern peace is not from enlightenment but because war became too destructive. Peace should never be assumed.
This is would be pretty tangential to the discussion so not sure why you are bringing it up? Did you think my post was illustrating that camera based self driving cars were infeasible? Also it appears you are mistaking framerate for latency, something I specifically warned about in my post... I don't really know what to say here.
Sorry not trying to dunk on you here, but this reads like something a junior engineer would complain to me about. These are not trivial problems, and I'm sure your co-workers who resolved them already so you didn't need to worry about them would agree with me.
> In reality none of this is a problem. There are automotive grade cameras which can collect enough light fast enough that the images are not blurry in practice. Yes, these cameras have a non-zero exposure time. Yes, this adds latency. No, this is not a problem.
You are contradicting yourself here, yes there are automotive grade cameras, but if this wasn't a problem, why would automotive grade cameras need to exist? My post wasn't saying these were impossible problems but hard ones.
> You mean 88 feet. If my off the shelf library introduces a random 1 second delay i will chuck it in the bin post haste. Use stuff whose performance characteristics are well understood by you and is not terrible.
Look I might have typed the wrong unit but it's a bit ironic you gave me this word salad right after complaining about this... Yeah you obviously don't use the library that adds a 1 second delay, but often you don't have the luxury of knowing that until after you learn about it through integration testing.
Libraries don't usually come with latency stats calibrated to your desired hardware right on the tin, would be pretty sweet if they did though.
> I do not recognise the world you describe.
I don't find this surprising :)
In regards to the 120hz camera line, I would be happy to expand on that for you. To be clear, I'm specifically talking about how when you try to increase sampling rate by interlacing multiple concurrent digital sensors you need to deal with the following potential problems (this is a property of all digital sensors):
- Real time synchronization between concurrent sensors requires additional processing time to ensure proper ordering. The samples need to be processed together in small batches. This adds latency and the more things you are trying to synchronize the more latency is introduced.
- Inter-sensor calibration to account for variations between individual sensors can be used to reduce the latency introduced by synchronization but lacking this, you are bottlenecked by the slowest sensor. There are a lot of different ways to handle this though so I'm speaking very generally.
- Broadly speaking most signals need to go through some kind of filter to remove noise, most digital filters have a certain amount of algorithmic latency built-in that is physically unavoidable. When you are interlacing multiple different sensors, you are getting more noise, so the likelihood of a filter being required starts to increase.
I want to stress here that these are not impossible challenges. In fact they are largely solved problems. But they are not universally solved in the same way, you need to balance between precision manufacturing, signal quality and signal latency. In practice most people are not prioritizing latency, so a 120hz camera might be optimizing for video recordings and not live processing scenerios. So long as you know what you are looking for you can avoid this when choosing which camera to use.
Computers can be fast, but fast can mean different things when dealing with real-time situations in high speeds. Bottlenecks need to be considered from all levels. The clock speed of the computers CPU often gives product managers weird ideas about what is possible. This was the main point I was trying to make here.
Consider it in the context of camera based self-driving cars, it's tangential to this discussion but it's an easy to visualize metaphor:
- A car traveling 60mph is traveling at 88 feet per second
- Assuming a 60hz camera, there would be a 16.67 ms gap between each frame
- The car is traveling 1.5 feet between each frame interval
- A certain amount of exposure time is necessary for the camera to generate even 1 frame or it will be blurry
- High framerate cameras often work around this by staggering/interlacing multiple sensors, but doing this implicitly increases the latency of each frame
- A 120hz camera might deliver double the frames per second, but each frame could be arriving 4 frames late in exchange
- 4 frames would be imperceptible to humans, it would be 3 feet for the car
- You haven't even processed any of these frames yet.
- Your off the shelf library introduces a random 1 second delay for some reason and costs you 88 miles in processing time
- The car can drive as fast as 120mph
All digital sensors implicitly have a sampling frequency, and the fundamental disconnect is always high sampling frequency =/= low latency. People constantly make this mistake over and over, and by the time you notice you are already too deep into development to make a change.
Decreasing latency is expensive, and requires specialized knowledge. Often you get expert software engineers who end up bottlenecked by the hardware limitations they can't even comprehend or the reverse, hardware guys bottlenecked by the software they can't introspect. The latency is only truly understood when you get to integration testing.
Nearly every step of the way you discover you need specialized hardware, software, operating systems, sensors. Every part of the chain each costing you more latency. It's like it's own ecosystem where almost everyone writes everything from scratch and doesn't share anything. It's gotten better in recent years though.
Full disclosure: I work in medtech and don't actually deal with cars, but it's a very similar problem space. We often use the same hardware/software cars use for this reason.