Excessive generative AI output devalues all communication media unfortunately. As a greater and greater portion of text, imagery, audio, and video come to be synthetic, I don't see why anybody would be motivated to do deep reading, or listening, or analysis of said media. It's frightening to think what that might do to our civilization, almost like putting us back into a weird version of the ancient past when nobody was literate, save the priest.
Maybe humans can adapt to care about generated content and treat it with the same emotional weight that human communication has. I don't know if that's good or not.
> LLM can produce texts but we as people will not be able to understand anyway
LLM generated text oftentimes just do not have the point to be understood. LLM is training you to NOT read, because trying to do close reading of LLM generated text is usually a mistake. There is no deeper there. Either it was correct on the surface things or it is a gibberish.
PISA is a statistical measurement tool designed for sensitivity to differences in student performance. So it uses questions that some students are expected to get right and some are expected to get wrong, and places the result on a bell curve fit to a reference population.
It's not clear what you mean by "sufficient reading literacy," but I guess it's one of the PISA proficiency levels. Those are simply equally-spaced intervals on the bell curve. For example, level 1c in reading is 2.4 to 3.1 standard deviations below the reference mean and 21% of participants in Iraqi Kurdistan and Lebanon scored at that level. Relative to the baseline, they are definitely close to the bottom. However, even they must've gotten some questions right, or they would've ended up in the "below level 1c" bucket (like 3% in Iraqi Kurdistan and 5% in Lebanon did). (https://www.oecd.org/en/publications/pisa-2025-results-volum... Table I.B1.2a.11)
If you wanted to know how many people achieve "sufficient reading literacy," you would need a test specifically designed for measuring it (i.e. not PISA), where people with sufficient ability can be expected to get all questions right. For example, the literacy rate is typically defined via the ability to "read and write a simple statement about everyday life," a test which 99% of adults in wealthy countries pass, 92% in Lebanon and 84% in Iraq: https://ourworldindata.org/grapher/cross-country-literacy-ra...
Of course a test that 99% pass just becomes table stakes and boring, whereas PISA with its large differences between individuals is much more exciting, but you shouldn't let that mislead you into thinking that students scoring relatively low on PISA necessarily must be lacking even those table-stakes skills.
Even before AI, it was clear that developed societies will be weighted down with an underclass of welfare recipients who are unable or unwilling to develop skills necessary to meaningfully contribute.
As society/technology/the world moves faster and becomes more complicated, more and more of them will be left behind. 1 in 5 Americans are already functionally illiterate, and about 20% of working age men are already out of the labor force.
CIS is an anti-immigration hate group founded by a eugenicist and white supremacist. They're widely condemned for publishing factually incorrect and methodologically flawed reports by organizations across all of the political spectrum, save the furthest right extremists. That is not a legitimate source of information on immigration or welfare.
Do you have any numbers you'd like to provide to counter the claim that a significant portion of the population of working age individuals aren't already out of the labor force?
While I may have responded to your comment, I'm not debating you. My intention was to let other readers know that the sentiments you're expressing are little more than gutter eugenics hiding behind a flimsy facade of academic language and charts.
You're skipping any causal connection between the "as greater and greater" and "anybody would be motivated" parts. What is it in "excessive generative AI output" that "devalues all communication media"?
Not that I disagree much, but for myself I have an actual cause connecting the two parts, and it's not something inherent to AI.
Basically, synthetic content is like those fancy desserts - they look nice and the top layer possibly has an exquisite flavor (probably on purpose), but once you cut deeper, it's some weird fruit or nut or other foodstuff that belongs next to potatoes and meat, not in a cake. So I learned to avoid such desserts and stick to basic stuff I know are good entirely, like IDK a cream pie or cream-filled chocolate stuff or such.
Same is with content. My problem isn't that it's AI, but that it's wrong, full of mistakes human authors don't make. Hallucinations mid-text that make half of it convincingly argue bullshit. Or the uncanny or repetitive factor in art that kicks in after few seconds of looking.
Like with desert, I'm avoiding disappointment and wasted hopes. When synthetic context gets better, I'll be consuming it just like "vanilla" one, for the same reasons.
I think with the desert thing, you are getting at what was once described as "aura" of original art work. Things like photography have already made us question things like synthetic art and how it relates to "real art" and I am pretty sure this is how it will go with AI and art in general. I wrote about this in my full human only blog: https://decodingvibes.com/blog/3-hypothetical-paintings-and-...
Wait wait wait, why will humand syop wanting to read, listen to music, or watch movies and tv shows just because there are AI generated art? People will still feel the need for all those things, technology alone cannot erase the demand. The problem is generated art is not considered good by audiences because it usually lacks substance and genuine creativity. But that doesn't mean they will not want to watch The Odessy. If an AI can make a movie which is exactly equivalent to a Christopher Nolan movie, then yes people can adapt. But the current situation is that most AI art sucks and people do not like it. Not that people want to stop reading, listening or watching. We are doing that in much larger numbers than we ever did in history.
People have lessened what they read because they can more easily watch a video or listen to a podcast. It’s not like the 1800s where books were the only game in town. And why read an AI written article when you can listen to the AI generated podcast with banter between two fake personalities (yes, this is what I heard when I wanted info about qwen’s newest model on Apple podcast).
Yes, true, but in the 1800s, literacy rates were extremely low as well. Most people didn't read then either. By the time literacy improved, we already had competing media like radio, television, etc. But books still won because they were a different form of media. And most "popular" books since the invention of the printing press have usually been garbage, like guides on how to burn witches properly, etc. It wasn't like people were reading Dostoevsky in masses then and don't do it now. It was always a small minority of people that cared about actual literature.
I think it's also more about active vrs passive consumption these days. People these days consume media mostly passively (watch whatever is on a streaming service, hear a podcast/audiobook while driving, let AI summarize an article instead of reading it), whereas that wasn't really possible previously (go to a theater to watch a movie, actually read stuff with attention, etc.).
So yes, people might be reading less, but I don't think it's some massive fall in literacy as people make it out to be. We're much more busy and don't have time to actually sit down and actively focus on media consumption these days as well (assuming you're employed in a somewhat capitalism-based system).
> Yes, true, but in the 1800s, literacy rates were extremely low as well. Most people didn't read then either.
False, at least from an American or Northern European perspective.
In the 1800s reading was an extremely popular activity. Not something restricted to a few. Printing press had been around for ages and the majority of the population read.
If by majority of the population, you mean no women, no slaves, then yes, sure. And even from an American or NE perspective, you're still overestimating its popularity.
> According to data compiled by Our World in Data and the World Bank, the literacy rate of the world's population from secondary school age onward was only 12 percent in 1820 - around one person in ten. In 1900, it still barely exceeded 20 percent. From the 1950s on, world literacy began to take off, hitting 42 percent in 1960 and 70 percent in 1983.
As I said. If you look not at the entire world, but say Northern Europe, reading literacy (a different measure from the literacy discussed in many modern sources which includes writing literacy) was almost universal by the late 1800s.
For instance, in Sweden, to cherrypick a stat, in 1875 only 1% of military recruits were found unable to read [1].
We really must stop thinking that people of yestertimes where so much inferior to people today. Yes, there were differences between men and women, but this was also starting to change.
Kind of feels like we're in a similar place with AI programming as we are with self driving vehicles. Good enough for the typical case, maybe even better than most humans, but bad when dealing with novel situations or situations which require a really cohesive model of the system (code, business requirements, physical world, whatever).
Can't really have it all. If you want runtime assertions on your data/state, that has a cost. Nothing is stopping you from writing that logic out and having all your inputs and outputs sanitized for "semantic invariance". But if you want the language to implicitly do that, you're gonna have overheads. Pick a different systems language in that case, and deal with the tradeoff.
Well, sure they make reference. Borrow checkers and memory ownership concepts, like in Rust, are pretty good at giving a very low overhead improvement over bare C pointers. But they're not as flexible. Shared memory complicates ownership rules. And invariably, any Rust program that involves real hardware will require dropping down into unsafe at some points. A safer program with an unsafe kernel is an improvement, don't get me wrong.
I'm not sure how a type system would "solve" arithmetic overflow / bounds dynamically, without expensive overhead (boxed types).
I don’t think you have to reach for something as invasive as a borrow checker to handle most of these conditions. Bounds checks, for example, are rarely skipped intentionally in the vast majority of C programs; rather, they’re just skipped accidentally or performed incorrectly. Combine that with the extremely common and well-known compiler optimizations for BCE present in many compiled languages, and I don’t think C programmers get to trot out performance as an excuse for the language not supporting bounds checking by default, and providing escape hatches for people who really need to avoid it.
The same is true for overflow and unsafe math (tons of languages increase math and overflow safety on many paths without performance penalty, and baking that behavior into the type system with nondefault unsafe options is demonstrably tractable).
Hell, the prevalence of C static analyzers that warn about the vast majority of errors in these categories is proof that more safety is possible in C.
It’s not really about performance. It’s more the stubborn obsession with ABI compatibility at all costs and a pervasive lack of type system support and conventions relating to code reuse that hold C back.
Well said. I get the sense that C is kind of ossified today, and that the whole point is its compatibility benefit. I wonder what even makes it into the next C standard? I think c99 was the last meaningful one. Everything else is offloaded to the compilers, God bless them. -Weverything
Regarding the "expensive overhead of boxed types", do we just need more work at the hardware ISA level (with pointer tagging or NaN-boxing) and we need the hardware to be aware of those bits and handle them safely/properly via some minor extensions to RISC-V or Intel/AMD/Apple ISAs?
Or do we have to have completely new forms of hardware datatypes?
I've got a dead older GPU (RTX titan, 24GB of GDDR6). It's been a paper weight for a while, core is damaged, but I realize I could probably sell it or just the memory chips for a decent price on eBay.
Every quantifiable science has some amount of math, but it's by definition applied math. I think this was more of a debate between applied and more abstract math.
Apple has all the money in the world to take this to court, I don't see why they'd accept a settlement. The discovery process alone could honestly destroy OpenAI by making investors and employees nervous enough to look elsewhere. Would be especially interesting if this crosses into criminal territory, especially if there's solid proof of upper management or executives being aware.
glances at Kushner's $1 billion OpenAI investment through Thrive Capital
Are you sure they won't settle? Apple settled their case with NSO Group, even though they were (and still are) hacking iPhone handsets. Seems to me like Apple is happy to settle cases that interfere with their political ambitions whenever America's government asks them to stand down.
If Apple's protectionist treatment is predicated on non-interference with other protected companies like OpenAI, maybe they will be motivated to settle out of court.
If OpenAI is this shady culturally then all kinds of dark secrets could come out.
They just released 5.6 to great fanfare with benchmarks showing even their weakest model (Luna) supposedly smoking Fable. Having used it a couple hours I'm certain this is a lie, and I'm now curious what else is.
My first week at a new job that forces me back into Windows (with WSL). I'm about ready to throw my machine out the window and follow after it. I understand that a lot of the performance issues are things like crowdstrike and Defender and maybe some poorly configured network proxy stuff but Jesus, this sucks. I write embedded software for machines with 1 ten thousandth the compute of my dev machine, I should not be encumbered by issues like this.
Maybe humans can adapt to care about generated content and treat it with the same emotional weight that human communication has. I don't know if that's good or not.
reply