ijk

I wonder if it partially because "write like a human" is kind of a vacuous request. Like, it's the objective everyone including me has been saying that we want, but there's no one way to write like a human and and humans don't even have a good definition past "I know it when I see it."

There's a lot of work in the humanities about different aspects of good writing, but that's not quite the same thing. And anyway they tend to assume a pre-existing level of writing ability. Students are supposed to learn good writing through practice; there are rules and exercises but they're incomplete.

The portable gas turbine is a somewhat separate issue, in that it's highly illegal but if you have more money than god you can just pay off the fines and the judges.

It's relatively easy to avoid being that egregious about it if you're a little bit patient on construction permits and power contracts.

NIMBY opposition to datacenter construction is orthogonal to that, except insofar as in both cases the construction will gravitate toward places that won't or can't object to the construction.

Technically the current best chess engine is a neurosymbolic hybrid.

But I have found that sometimes the best use of an LLM is to write code for symbolic AI.

Yeah, unfortunately "LLMs are bad at counting" seems to apply to counting parentheses.

With modern tool calling I wonder if a better way to go about it is for the LLM to express the program changes as a function or otherwise use an editor that auto-balances parens. There's a lot of relatively simple tooling that makes it easier to write in a Lisp. The languages tend to lend themselves to being straightforward to check like that.

What's special about Lisp's repl is that it's perfectly possible to construct your entire program in the repl, testing each addition live as you write it. (Many Lisp-focused editors assume you'll want to do this, such as Emacs making it easy to run the interpreter on a single function in a file.) That tooling is lost if you just try to one-shot the file, and before 2026 the majority of LLMs originally just tried to one-shot every file.

But, just like a lot of early LLMs had huge problems with whitespace and numbers because the tokenization was taking efficiency shortcuts that made sense for text but absolutely wrecked code syntaxe, I wonder if the current optimizations are badly formatted for Lisp.

At the very least, using a varient like Clojure that also uses [] and {} in addition to () might help.

Except you can express the higher order structure in s-expressions too, and using an AST usually means expressing it as a different syntax than the other parts of your program.

Or maybe I've just been sleeping on the power of Javascript AST.

From the outside Meta's attempts to pivot from open source releases to fast follow closed models fell flat when they tried to prematurely monetize it. They could have owned the open weight model world but tried to pivot to closed weight chatbots before an actually viable revenue model appeared.

I did start a blog in that time period and this is the first time I've ever heard of this. (Admittedly, I wasn't trying to directly make a living off of it.)

If we use the S&P 500 as a proxy for the top 500 companies, I would guess that a lot more than 0% will be gone by the end of 5 years.

It's hard to find non-paywalled sources for business analysis, but from what I can find it'd be about 20%.

Back when red LEDs were the new cheap lighting option everything was great: red doesn't affect noght vision as much and at least for me it doesn't seem to prevent sleeping on the same way that the bright white LEDs do.

Hence the drive to control the compute hardware. If your competitors can't run inference at competitive scale and cost, they can't challenge you.

Historically one pattern this took was patronage. Most famously associated with sponsoring artists, but historically it covered a wide variety of professions. The patron gets power, influence, the ability to call on a guy who can do the thing, and so on.

Of course the first thing we replaced with AI was artists, so expect more exclusively as the lower rungs of patronage clients fall off the economic ladder.

Closing a question as a duplicate because there is already a question with similar wording (but assuming an entirely different tech stack, architecture, coding style, and goal) is a frequent enough experience that it became shorthand for the site's problems.

There was kind of a fatal mis-match between the questions being asked and the intended kind of questions that were being answered. The actual asks were often incomplete diagnostics of the questioner's current problem, frequently focusing on the wrong thing (because if you don't have the full knowledge of the thing you're going to be prone to incorrect assumptions of the diagnosis). SO's intent, though, was a more mathematical "here's the question, here's the programming concept that explains it" so you get the best explanation of how a linked list works under a completely unrelated problem. Which is fine, but the site's culture and design only partially acknowledged the disconnect.

The whole site developed a reputation of being something approximating the reverse of the comments under recipes that substitute lard for cream and wonder why their cake tastes funny. Lots of questions of "How do I implement this functionality in Y? We can't change our tech stack because of other factors, so it has to be Y" questions answered by "If you just use Z instead you wouldn't have these problems" and "closed as a duplicate of this question for how to implement the non-Y version" when there was a perfectly fine way to do it in Y.

Ah yes, programming: the discipline which famously only has singular right answers to problems, such that programmers never get in arguments with each other about the correct approach to solving a given problem, and there are no long running disputes that have ossified into intractable disagreements.

A sink that has large but finite capacity to absorb something can reach an irreversible tipping point when an additional shock happens.

There are many examples of this in nature. (And in Nature [1].) One interesting one that I think is unknown to many people is limnic eruption. A lake can absorb quite a lot of CO₂, for example from volcanic gases. Dissolved CO₂ is invisible, so the lake can look quite ordinary, but the build-up turns the lake into something approximating an unopened carbonated soft drink. If the lake is deep enough and the layers don't mix frequently enough to relieve the pressure, it can build up to the tipping point where the lake will suddenly explode, flooding the nearby landscape and releasing an invisible CO₂ cloud, which will proceed to kill the surrounding life by asphyxiation.

The conditions required for a limnic eruption are rare, though there were two incidents in Cameroon in the 20th century.

It's entirely possible that the build-up of hostility on Stack Overflow were survivable as long as it didn't build up to a level that exceeded the community's ability to absorb it. But an exogenous shock or the community shrinking could upset the balance, with hysteresis making the change difficult to reverse.

[1] https://www.nature.com/articles/s44458-026-00063-5

I agree in principle, but every time I run a debugger on modern C++ it makes it clear that, rather than being a simple and cutesy transformation, "compiler optimization" is actually black magic.

Yeah, private universities being unionized is more recent.

It's not. In the US, public university graduate student unions started in the 1970s.

[1] https://en.wikipedia.org/wiki/Graduate_student_employee_unio...

Which is not to say that conditions in graduate schools (or academia as a whole) are great. But the unionization process is entangled in the legal framework around unions in the United States.

Where I encounter it at the higher education level is that academic-level research almost universally has maximum word counts or page counts rather than minimums: if you think you can get your point across in fewer words, you should. No reviewer is going to object to the paper being too short, so long as you succeeded in making your case.

John Nash's Ph.D. Thesis is notorious for being short: it's still 27 pages (typed, with hand-written equations and a whopping total of two citations) but that's an order of magnitude below average. On the other hand, most of us don't invent game theory.

Students used to minimum-word-count essays sometimes have to do some self-retraining to realize that the expectation is that you have more that you want to say than you have room to say it, and the game is now to figure out how to say more in fewer words.

Literacy.

Percentage of children to survive to adulthood.

Global food surplus.

The was a big phase shift over the course of the 20th century...

So the individual website has to subscribe to the surveillance operator's IP location verification service, or be fined.

I don't understand the need for age verification.

I mean, I understand what it effects it has, and why many parties want to perfect their expanding panopticon, and why screaming think of the children makes politicians' brains turn off.

It won't fix children or social media. That's been apparent ever since Facebook defaulted to real names and people still posted everything they would have otherwise. It makes it easier to use social disapproval to destroy nonconforming individuals, I suppose. And to sell ads. And to destroy anyone who criticizes the government. So no real downside if you don't care about that sort of thing.

Secure encryption has been classified as controlled munitions in the past. Making SSH illegal is well within the range of possible futures.

It'd be a stupid future, but it's a stupid present so I'm not going to rule it out on those grounds.

There were a lot of millennials at the time. Still are, it's just that they're pushing 40 instead of 14.

In the US, homelessness per capita peaked in the 90s. It remained relatively high afterwards compared to the pre 1980 numbers, and has recently spiked higher in the wake of 2020 and following, but in the 1990s there were a lot of homeless people in the US.

Now I want to see this implemented via vacuum tubes, similar to running Eliza on punchcard-driven mainframes...

The authors of the study weren’t stupid. They knew the LLMs would provide poor results. They ran the study to quantify it and create a resource to spread the information in response to the rise of AI carb counting apps.

Yeah. I think it is under-appreciated that much of science is intended for debugging purposes. Sure, you and I know that X is positive, but what's it actual value? Can we find the causes that make it that way? Et cetera.

I was complaining about AI generated clothes being misleading marketing, deceiving customers as to whether the garment even exists.

And then I learned that the pre-AI norms weren't any less fictional: they made an exemplar garment and did photoshoots, sure, but then they send the pictures and patterns to the lowest bidder factories with permission to make whatever edits are necessary to make it cheap and manufactureable. The whole thing was already a simulacrum.

That's an interesting bit, where reducing friction too much can eliminate the side effect that is actually driving the desired results.

Do you want to count calories, or do you want to lose weight? Sounds like it's possible to hyper-optimize calorie counting to the point that it becomes counter-productive...

I feel like the problem is that it's both. We're sanding off the long tail of human expression. It's not profitable this quarter, you see. Faster to let the AI do it.

It really feels sometimes like they were trained on too much short-form fiction or something. Really stunts their sentence and paragraph texture.