English spelling is not phonetic. You use one of a few canons. "Cause" is a canonical spelling. "cus" is not. There isn't a Black spelling canon. Your phone does not offer it. Nowhere is the correct spelling "cus". It isn't Black spelling. All English speakers have unique accents and pronunciations, but do not transcribe them. They spell with a particular canon. I am prepared to defend this point to the end.
The model, given incorrect spellings, infers the writer is dumber than a person who can spell.
We can frame the paper as a syllogism.
1. The model judges people who can't spell.
2. Black people, allegedly, in 2024, with autocorrect on their phones, can't spell as well as my six year old can with a crayon.
Doesn't matter, you'd still be wrong about the correctness of the spelling. The development of differentiated versions of a language for racial, ethnic, and cultural minorities is accepted and studied by linguists in both spoken and written form, e.g.:
Alternate spellings like that are not considered mispellings, as they are usually intentional and/or part of such a dialect. Persons using them are not losing on a spelling bee, they're using an alternative "canon".
As for the paper: a model using statistical inference (going from statistics on how a group tends to write a word to an assumption of the membership to the group of someone writing like that) is not using stereotypes, it's using data (e.g. from online discussions of the group).
The association "use of languge/dialect X == lazy, etc" though, is racist.
That said, I'm willing to entertain the idea of "stupid" for other kinds of unintentional real mispellings like "they're/their". Those are not confined to a specific group though. Not that I don't make them all the time when quickly hammering some point in anger on HN...
> The association "use of languge/dialect X == lazy, etc" though, is racist.
The model would probably say certain forms of "bad English" (that would likely also include things like redneck or bro type language, as another commenter remarked, but that wasn't tested for some reason) make it more likely that the person in question is lazy or something like that. And that is a statistical claim which could be true. In which case it wouldn't be racist. So it is premature to call some probabilistic statements racist without even investigating whether they are correct.
My claim is it is a recognized by linguists dialect of English (with tons of regional and temporal subdialects).
Black people can spell in the standard US English spelling, and can also spell and use syntax from AAE, and can also do both in different mixes and percentages, depending on the context, their background, and so on.
No "context of playing" required, it can also be done when dead serious.
It’s certainly hard to detect authorship per se, but I expect very easy to detect content people produce using LLMs. You can target affiliate marketing blogs and other SEO garbage, entirely ignoring whether it’s literally written by human hands.
> It’s certainly hard to detect authorship per se, but I expect very easy to detect content people produce using LLMs.
Yes. For now. Just flag the use of the words "delve" and "deep dive," flag bullet-pointed paragraphs and lists, and flag anything that closes with a too-neat one-paragraph conclusion — and you've just caught 85% of AIspam.
But this is already changing, and in about 6-10 months I expect it'll be impossible to detect at more than 70% confidence whether an article is written by an AI or a human. (You could say: The human writes with soul, with style, whereas the AI is totally bland, irritatingly didactic, and generic. But you can quite easily coax an AI into imitating any literary style, and some of the best AI-generated content is warped in this way.)
> Yes. For now. Just flag the use of the words "delve" and "deep dive," flag bullet-pointed paragraphs and lists, and flag anything that closes with a too-neat one-paragraph conclusion
Uhoh, is my copy pasta in danger of being snuffed out before anyone is served it?
Look, I get how some people would consider it spam. However, the problem still remains and the solution has not changed!
I have a life to live and can't spend every waking moment crafting a unique comment every time I need to reply to someone.
They had an algorithm that perfectly recommended content to you. Then they scrapped it and went back to star ratings. Why?
Netflix makes money from monthly subscriptions. They lose money when actually streaming content to users. So the ideal user from their standpoint is one who keeps paying but watches almost nothing.
Thus, the company is incentivized to put out just enough good stuff so you won't cancel, but they don't want you binging.
If Google sends users to pages and advertisers pay them, the incentive to address the content problem isn't there. In fact, LLM content may be another cash cow waiting in the wings.
Yes. It is possible to check everything.
We've been manufacturing now for centuries. There are absolutely techniques, methodologies, organizational structures, all to prevent this.
That this is happening is purely cost cutting. Accountants balancing lives against pennies. Just like automakers would do calculations to determine the cost of installing seat belts against potential lawsuits. Nobody cares about lives, it was simply a calculation based on potential lawsuits.
Notice, it is totally possible to check everything. There was never a problem with our ability to prevent deaths, it was just too expensive. Because lives are actually cheap. MOLOCH.
They're not balancing lives against pennies. They're just not thinking about it. That's the mistake you're making: you're implying a plan because a plan means there's control. But that's not how it works: someone said "hey it would be cheaper if we could just not do this" - someone probably said "well this double check is necessary because bad things could happen". They were dismissed with the twofer of "that check never finds any faults" or my personal favorite "that won't happen" (no justification given).
No one in Boeing has a spreadsheet of expected aircraft crash rates versus dollars earned on manufacturing and turned the dial till they got a nice peak. They just plain didn't think about the potential consequences.
Industrial tools manufacturers have the same issues. Sometimes it ends in breaking that hurt both people and the tool, but sometimes it's even more dumb: 'hey, this part is under a lot of friction, has to be replaced after each maintenance cycle, let's buy almost the same at half the cost, the maintenance will be 20% cheaper'. How it end up: the tool require faster maintenance cycle, and in the end, without taking into account maintenance technicians salaries, it ends up 30% more expensive (which isn't an issue for new tools, but quite a big one for tools under initial maintenance contracts)
Right, you are correct. Corporations have adapted and don't make such blatant calculations anymore. With seat belts, or gas tanks, in 70's. There were actual actuaries that did the exact calculations, what would manufacturing cost, what percentage of accidents would result in lawsuits, what are estimated payouts. And they actually did the math to balance cost to produce to lives.
But they got into a lot of trouble that they were doing those calculations.
Now it is more subtle, everyone has budgets to hit, there are deadlines, production quotas. And with Just In Time Manufacturing, and local 'ownership'. Some Blackbelt consultant convinces the local managers to 'optimize'. And of course, they start cutting corners. No real calculation, just constant pressure to cut.
Sorry, I could've been clearer, I understood the reference, I wanted to know what they meant by signing off like that. I took an inference that seemed unusual and perhaps mistaken, so I wanted clarification.
In the Moloch essay, there are several examples of 'capitalism' or 'libertarianism' that make the point that there is such a constant drive to cut costs, that humans will destroy themselves. And this drive is "Moloch", we are it, and also subject to it.
So to me Boeing is just another example in a long line of corporations just doing what corporations do, cut costs, no matter what the impact.
It sounds too blatant, who would consciously want to kill people to save a few pennies.
But what actually happens is just organizational, everyone has a budget to hit, there are layers and layers of middle managers that are getting squeezed, and they start cutting corners. No single corner seems that bad. But it is a constant slow progression towards disaster.
A thousand cuts. Some 'common joe manager' that has a kid in collage, a mortgage, and his boss is breathing down his neck, so he cuts a corner. -- But the problem is, everyone has a boss, no one is free, the entire organization/system is the beast, but we also comprise the beast.
Depends how deep they go on the root cause analysis. If the manufacturing flaw was in turn caused by management flaws, then yes, that’s an underlying issue that has the potential to taint everything above it.
Well, looking at another part of the organization distant enough that top management (the C-suite) is pretty much all they have in common: They had massive quality problems on their space capsule (Starliner). (They launched the first test flight without having ever done an intgrated test of the software in expected mission time -- which led to the computers being so confused immediately after launch that it blew most of the thruster propellant. That led to an urgent hunt for software bugs elsewhere, with the capsule already in orbit -- which found another nasty bug which was even worse.)
This is in addition to numerous other quality control issues on airliners -- several other well-publicized issues on the 737-MAX itself, bric-a-brac left in tanks on 767 tankers, quality control issues at the plant producing 787s.
You don’t see any possibility that the unprecedented in modern times failure of the 737 MAX might have caused anyone in the FAA leadership or Congress to reconsider their relationship? Some might see the language in this memo as evidence of that.
The DC-10 cargo door issue was unprecedented in its time, and the first failure was smoothed over with a phone call from the CEO to the FAA, instead of an airworthiness directive. The plane wasn’t grounded until the second incident killed all aboard.
I am glad for the AD in this case, and hope the FAA casts as watchful an eye over Boeing that we are all hoping they will, but shortcuts have happened before.
And the FAA director at the time was selected specifically as a person who would be more willing to play ball with the business, and had such marching orders from President.
Similar thing happened during Bush Jr's administration.
Yes - I'd like to think that we've learned better but that was in the late 1970s and we are suspiciously around the time where the people who had direct experience would have retired or been on the way out, similar to how a lot anti-vax recruits were too young to have lived memory of just how many kids things like polio harmed.
It sometimes seems like we need the engineering equivalent of what Germany does with Holocaust education to remind everyone of how past disasters were preventable, and usually caused by multiple people deciding not to act on early warning signs.
This happens with guitars, cars, houses…any item which has both historical and function value.
The original owner of the item makes numerous small changes and repairs over decades which preserve and improve the functionality.
They finally sell it off to a collector who reverts all the changes, degrading the functionality in order to restore some notion of authenticity. And sometimes they go further, fixing “mistakes” made by the factory.
In this case, the functionality of a Z80 is so ridiculously outdated and obsolete that I suppose authenticity (and nostalgia) is really the only point.
Important, and often ignored! Technological improvements change the cost structure of activities. Consider texting. We send billions of texts every day. This does not replace billions of couriers. Rather, more people work to maintain our phone supply and telecom than we ever had running around with letters. It's not a promise but it is a pattern.
Also bicycles, motorbikes, e-scooters, trams, trains, runaway objects...
I always thought that the city itself could be part of the swarm system. Roadside sensors could broadcast information to all relevant participants - such as "emergency vehicle incoming, move to the right lane, vehicles ID 123, 456 and 789 stop immediately" or "danger - static object at 45th km", or even "slow down/speed up to 60 km/h to optimize traffic flow".
In the science fiction of my youth, where roads or similar systems existed, there were cases where the infrastructure was part of the autonomous vehicle system. It was often car-exclusive, never mentioning pedestrians, mobility devices, or anything that wasn't motorized, but it was an interesting bit of speculation.
Reminds me of VLIW. As per Wikipedia, from the Itanium page:
> One VLIW instruction word can contain several independent instructions, which can be executed in parallel without having to evaluate them for independence. A compiler must attempt to find valid combinations of instructions that can be executed at the same time, effectively performing the instruction scheduling that conventional superscalar processors must do in hardware at runtime.
If your CPU exposed the single-stream parallelism at the interface, you can do it at compile-time or even decide it with in-line assembler.
I wonder if it hasn't caught due strictly to the business dynamics of the industry, or are there technical reasons this isn't really a good strategy?
Well, IIRC it didn't caught on mostly because of a) compilers weren't really that good at that kind of instruction scheduling (and when they improved, Itanium has sunk already), b) conventional ISAs (that is, x86) got quite good at doing this in hardware, at runtime, and actually deliver slightly better results than static scheduling precisely because they do it at runtime, when profiling data is available.
I believe Linus has a good even if tangentially related to this exact topic rant at [0]. "While the RISC people were off trying to optimize their compilers to generate loops that used all 32 registers efficiently, the x86 implementors instead made the chip run fast on varied loads and used tons of register renaming hardware (and looking at _memory_ renaming too)."
Static scheduling, even with profiling, can never be as good as dynamic scheduling for general-purpose workloads. VLIW/EPIC can do well for HPC-style number crunching, but that isn't everything. https://news.ycombinator.com/context?id=37900987
One can move complexity back and forth between compiler, runtime and processor implementation to some extent. VLIW works really well in some niches. It's harder to program than single instructions that execute in sequence, either by hand or by compiler, but it simplifies the scheduling for the hardware. Works better if the bundled instructions have similar latency.
The key design puzzle at present seems to be that memory access takes many more cycles than arithmetic. Bundling a few cycles of arithmetic with a few hundred cycles of memory load is kind of pointless. So VLIW works well if you know memory access is going to be fast, which roughly means knowing it'll hit in L1 cache or equivalent. I think that's part of why it suits DSP style systems.
Exposed pipelines are an interesting quirk of some of these systems. One instruction in a VLIW bundle writes to a register and subsequent instructions that read from that same register will see the previous value for N subsequent cycles, after which the write becomes visible. They're really confusing to program by hand but compilers can deal with that sort of scheduling.
Because static scheduling is terrible for non-DSP and non-HPC loads like the typical server or desktop application where the control and data flow is very input dependent. Until recently DSP and HPC were a tiny fraction of the market so architectures capable of dynamic scheduling dominated even those markets as they had more investment.
With GPUs of course things have changed and in fact GPUs relied more on static scheduling, but even there as they expand to more varied loads, they are acquiring more dynamism.
Im reading that TeraScale (AMD) works this way. Itanium is a major attempt to ship it in a CPU. I guess AMD64 and ARM rule the day but maybe in the future we'll see it again.
Terascale was a vliw, worked well as far as I know. The current amdgpu architectures aren't - those are multiple execution port systems, reminiscent of the x64 setup.
Qualcomms' Hexagon is a vliw, I think that's contemporary. Graphcore's IPU is two instructions per word.
The model, given incorrect spellings, infers the writer is dumber than a person who can spell.
We can frame the paper as a syllogism.
1. The model judges people who can't spell.
2. Black people, allegedly, in 2024, with autocorrect on their phones, can't spell as well as my six year old can with a crayon.
3. Therefore, the model is racist.
What do I do with this?