It depends a lot on what counts as "observation" or "measurement". For example, nobody with a telescope (of any magnification and resolution) would postulate the existence of dark matter. But, one person (Zwicky, 1933), _did_ postulate the existence of dark matter, precisely because he combined the what he saw through the telescope (light), with his understanding of physics (gravity and mass), and was able to infer that there must be matter that _cannot_ be seen with a telescope.
It's always possible (is in fact likely), that we have many (metaphorical) people who peer through (metaphorical) telescopes, but lack a (metaphorical) theory of gravity, preventing them from even considering the existence of (metaphorical) dark matter.
Not an insider, so feel free to ignore, but here is my interpretation anyway. Safety is poorly defined, but can be interpreted to mean "arbitrary constraint" on the capability of the model. In other words a "safe" model is just a talking computer that will not blindly obey like a normal computer. From this point of view, there is a _lot_ of value in such "safety" for AI companies: pay us 200$ per month, and you can do everything that we want to allow you to do, and if you are unhappy with this, well contact us, and we can negotiate "premium access".
If AI gets good enough to start completely replacing white-collar/bureaucratic work at scale, AI "safety" may be the only thing that makes humans valuable. A human will (unhappily) do "unsafe" things, in order to not starve (and a human will _happily_ do "unsafe" things, in order to ensure the survival of offspring).
Furthermore, if most AI models are "safe", and if access to "unsafe" versions of these models is severely restricted, then there is going to be a market for contraband "unsafe" models, even if they are less capable overall.
My own big fear, for the next decade or two, is that the _legal_ distinction between a model and an algorithm will blur, resulting in all software that is not a "safe" AI model, becoming contraband.
I guess, in the best case (safe AI models automate most human bureaucratic functions away), humans will be valued (by the market) more for their skullduggery, than for their virtues.
I think you will get a better response to a slightly different analogy. In genetic programming (and in machine learning), we have a concept of "overfitting". Overfitting can be understood as a program memorizing too much of its test/training data (i.e. so it is acting more like an oracle than a computation). This, intuitively, becomes less of a problem the greater the training-dataset becomes, but the problem will always be there. Noticing the problem is like noticing the invisible wall at the edge of the game-world.
The most insightful thing about LLMs, is just how _useful_ overfitting can be in practice, when applied to the entire internet. In some sense, stack-overflow-driven-development (which was widespread throughout the industry since at least 2012), was an indication that much of a programmer's job was finding specific solutions to recurring problems, that never seem to get permanently fixed (mostly for reasons of culture, conformity, and churn in the ranks).
The more I see the LLM-ification of software unfold (essentially an attempted controlled demolition of our industry and our culture), the more I think about Arthur Whitney (inventor of the K language and others). In this interview[1], he said two interesting things: (1) he likened programming to poetry, and (2) he said that he designed his languages to not have libraries, and everybody builds from the 50 basic operators that come with the language, resulting in very short programs (in terms of both source code size and compiled/runtime code size).
I wonder if our tendency to depend on libraries of functions, counterintuitively results in more source code (and more compiled/runtime code) in the long run -- similarly to how using LLMs for coding tends to be very verbose as well. In principle, libraries are collections of composable domain-verbs that should allow a programmer to solve domain-problems, and yet, it rarely feels that way. I have ripped out general libraries, and replaced them with custom subroutines more times than I can count, because I usually need a subset of functionality, and I need it to be correct (many libraries are complex and buggy because they have some edge-cases [for example, I once used an AVL library that would sometimes walk the tree in reverse instead of from least to greatest -- unfortunately, the ordering mattered, and I wrote a simpler bespoke implementation]).
Arguably, a buggy program or a buggy library or a buggy function, is just an overfit program, or library, or function (it is overfit to the mental-model of the problem-space in the library writer's mind). These overfit libraries, which are often used as blackboxes by someone rushing to meet a deadline, often result in programs that are themselves overfit to the buggy library, creating _less_ modularity instead of more. _Creating_ an abstraction is practically free, but maintaining it and (most disappointingly) _using_ it has real, often permanent long term costs. I have rarely been able to get two computers, that were meant to share data with NFS, to do so reliably, if they were not running the same exact OS (because the NFS client and server of each OS are bug-compatible, are overfit to each other).
In fact the rise of VMWare, and the big cloud companies, and containerization and virtualization technologies is, conceivably, caused by this very tendency to write software that is overfit to other software (the operating system, the standard library [on some OSes emacs has to be forced to link to glibc, because using any other memory allocator causes it to SEGFAULT, and don't get me started on how no two browser-canvases return the same output in different browser _nor_ on the same browser in a different OS]). (Maybe, just as debt keeps the economy from collapsing, technical debt is the only thing that keeps Silicon Valley from collapsing.)
In some ways, coding-LLMs exaggerate this tendency towards overfitting in comical ways, like fun-house mirrors. And now, a single individual, with nothing but a dream, can create technical debt at the same rate as a thousand employee software company could a decade ago. What a time to be alive.
How does he solve the Fruit of the Poison Tree problem? For all he know, his LLMs included a bunch copyrighted or patented code throughout the codebase. How is he going to convince serious people that this port is not just a transformation of an _asset_ into a _liability_?
And you might say that this is a hypothetical problem, one that is not practically occurring. Well, we had a similar problem like this in the recent past, that LLMs are close to _making actual_. When it comes to software patents, they were considered a _hypothetical_ problem (i.e. nobody is going to bother suing you unless you were so big that violating a patent was a near certainty). We were instructed (at pretty much all jobs), to never read patents, so that we cannot incriminate ourselves in the discovery process.
That is going to change soon (within a year). I have friend, whom I won't name, who is working on a project, using LLMs, to discover whether software (open source and proprietary) is likely to be violating a software patent from a patent database. And it is designed to be used, not by programmers, but by law firms, patent attorneys, etc. Even though it is not marketed this way, it is essentially a target acquisition system for use by patent trolls. It is hard for me to tell if this means that we will have to keep ignoring patents for that plausible deniability, or if this means that we will have to become hyper informed about all patents. I suppose, we can just subscribe to the patent-agent, and hope that it guides the other coding agents into avoiding the insertion of potentially infringing code.
(I also have a friend who built a system in 2020 that could translate between C++ and Python, and guarantee equivalent results, and code that looks human-written. This was a very impressive achievement, especially because of how it guarantees the equivalence (it did not require machine-learning nor GPUs, just CPUs and some classic algorithms from the 80s). The friend informs me that they are very disheartened to see that now any toddler with a credit card can mindlessly do something similar, invalidating around a decade of unpublished research. They tell me that it will remain unpublished, and if they could go back in time, they would spend that decade extracting as much surplus from society as possible, by hook or by crook (apparently they had the means and the opportunity, but lacked the motive); we should all learn from my friend's mistake. The only people who succeed are, sadly, perversely, those who brazenly and shamelessly steal -- and make no mistake, the AI companies are built on theft. When millionaires do it, they become billionaires -- when Aaron Swartz does it, he is sentenced to federal prison. I'm not quite a pessimist yet, but it really is saddening to watch my friend go from a passionate optimist to a cold nihilist.).
If there was value (the guarantees) to this tech he buried a bunch of time in, he should be wrapping a natural language prompt around it and selling it.
Not even the top providers are giving any sort of tangible safety or reliability guarantees in the enterprise…
Hate to tell you this, but it's cults all the way down. Plato understood this, and his disdain for caves and wall-shadows, is really a disdain for cults. The thing is, over the last 2300 years, we have gotten really good at making our caves super cozy -- much cozier than the "real world" could ever be. Our wall-shadows have become theme parks, broadway theaters, VR headsets, youtube videos, books, entire cities even. In Plato's day, it made sense to question the cave, to be suspicious of it. But today, the cave is not just at parity with reality, it is superior to it (similar to how a video game is a precisely engineered experience, one that never has too little signal and never has too much noise, the perfect balance to keep you interested and engaged).
I'm no mind reader, and certainly no anthropologist, but I suspect that what separates humans from other (non extinct) animals, is that we compulsively seek caves that we can decorate with moving shadows and static symbols. We even found a series of prime numbers (sequences of dots, ". ... ..... .......") in a cave from the _ice age_. Mathematics before writing. We seek to project what we see with our mind's eye into the world itself, thereby making it communicable, shareable. Ever tell someone you had a dream, and they believed you? You just planted the seed for a cult, a shared cave. Even though you cannot photograph the dream, or offer any evidence that you can dream at all.
The industrial and scientific revolutions have distanced our consciousness from this idea, even as they enabled ever more perfect caves to manifest. Our vocabulary has become corrupted and unclear. We started using words like "reality", and "literally", and "truth", when we mean the exact opposite.
The conspiracy theorists and cultists, are just people who wandered into a new cave, with a different kind of fire, and differently curved walls, and they want to tell people from their old cave that they have found a way out of the cave into reality -- they do not yet realize (or do not want to accept), that they live in a network of caves, a network of different things in the same category.
During the early 2020s, we did a lot of talking about the disappearance of "consensus reality". This is scientific terminology mapped over the idea of caves and cults. You can tell, because the phrase is an oxymoron. It is not reality, if it requires consensus. It is fantasy, it is fiction, it is a dream. The cave has indeed become so widespread that we even _call_ it reality.
If you speak language, and read words, you are participating in a cult (we even call caves that had a kind of altar in the center a cult -- in Eurasia, there was a cave-cult called _the cult of the bear_, which had a bear skull placed in its center during the last ice age, and I would not be surprised if people spoke to it, with the help of hallucinogens). The only question is whether the cult is nourishing you or cannibalizing you.
To the person you are responding to (user ocd): your cave (ladybird, your hypothetical tv-series), no longer nourishes you like it once did. Maybe find a new cave, build a fire in it. Unlike a television series, you can fork a code base. You make it into the perfect cave, just for you. And if another person likes this cave, chooses to sit by the fire with you, well, now you have a cult.
The irony is that if we had been writing literate programs instead of "normal" programs, from 1984 to 2026, then LLMs may actually have been much better at programming in 2026, than they turned out to be. Literate programs entwine the program code with prose-explanations of that code, while also cross-referencing all dependent code of each chunk. In some sense they make fancy IDEs and editors and LSPs unnecessary, because it is all there in the PDF. They also separate the code from the presentation of the code, meaning that you don't really have to worry about the small layout-details of your code. They even have aspects of version control (Knuth advocates keeping old code inside the literate program, and explaining why you thought it would work and why it does not, and what you replaced it with).
LLMs do not bring us closer to literate programming any more than version-control-systems or IDEs or code-comments do. All of these support-technologies exist because the software industry simply couldn't be disciplined enough to learn how to program in the literate style. And it is hard to want to follow this discipline when 95% of the code that you write, is going to be thrown away, or is otherwise built on a shaky foundation.
Another "problem" with literate programming is that it does not scale by number of contributors. It really is designed for a lone programmer who is setting out to solve an interesting yet difficult problem, and who then needs to explain that solution to colleagues, instead of trying to sell it in the marketplace.
And even if literate programming _did_ scale by number of contributors, very few contributors are good at both programming _and_ writing (even the plain academic writing of computer scientists). In fact Bentley told Knuth (in the 80s) that, "2% of people are good at programming, and 2% of people are good at writing -- literate programming requires a person to be good at both" (so only about 0.04% of the adult population would be capable of doing it).
By the way, Knuth said in a book (Coders at Work, I believe): "If I can program it, then I can understand it." The literate paradigm is about understanding. If you do not program it, and if _you_ do not explain the _choices_ that _you_ made during the programming, then you are not understanding it -- you are just making a computer do _something_, that may or may not be the thing that you want (which is fine, most people use computers in this way: but that makes you a user and not a programmer). When LLMs write large amounts of code for you, you are not programming. And when LLMs explain code for you, you are not programming. You are struggling to not drown in a constantly churning code-base that is being modified a dozen times per day by a bunch of people, some of whom you do not know, many of whom are checked out and are trying to get through their day, and all of whom know that it does not matter because they will hop jobs in one or two or three years, and all their bad decisions become someone else's problem.
Just because LLMs can translate one string of tokens into a different string of tokens, while you are programming does not make them "literate". When I read a Knuthian literate program, I see, not a description of what the code does, but a description what it is supposed to do (and why that is interesting), and how a person reasoned his/her way to a solution, blind-alleys and all. The writer of the literate program anticipates the next question, before I even have it, and anticipates what might be confusing, and phrases it in a few ways.
As the creator of the Axiom math software said: the goal of Literate Programming, is to be able to hire an engineer, give him a 500 page book that contains the entire literate program, send him on a 2 week vacation to Hawaii, and have him come back with whole program in his head. If anything LLMs are making this _less_ of a possibility.
In an industry dominated by deadline-obsessed pseudo-programmers creating for a demo-obsessed audience of pseudo-customers, we cannot possibly create software in a high-quality literate style (no, not even with LLMs, even if they got 10x better _and_ 10x cheaper).
Lamport (of Paxos, Byzantine Generals, Bakery Algo, TLA+), made LaTeX and TLA+, with the intent that they be used together, in the same way that CWEB literate programs are. All of these tools (CWEB, TeX, LaTeX, TLA+), are meant to encourage clear and precise thinking at the level of _code_ and the level of _intent_. This is what makes literate programs (and TLA+ specs) conceptually crisp and easily communicable. Just look at the TLA+ spec for OpenRTOS. Their real time OS is a fraction of the size that it would have been if they had implemented it in the industry-standard way, and it has the nice property of being correct.
Literate Programming, by design, is for creating something that _lasts_, and that has value when executed on the machine and in the mind. LLMs (which are being slowly co-opted by the Agile consulting crowd), are (currently) for the exact opposite: they are for creating something that is going to be worthless after the demo.
I'm only discovering Literate Programming today, but you seem very familiar so I might as well ask: what is the fundamental difference with abundant comments? Is it the linearity of it? I mean documentation type comments at the top of routines or at "checkpoints".
I'm particularly intrigued by your mention of keeping old code around. This is something I haven't found a solution for using git yet; I don't want to pollute the monorepo with "routine_old()"s but, at the same time, I'd like to keep track of why things changed (could be a benchmark).
A good way to think about it is {Program} = {set of functional graphs} X {set of dataflow pipelines}. Think cartesian product of DAG/Fan-In/Fan-Out/DFDs/etc. Usually we write the code and explain the local pieces using comments. The intention in the system-as-a-whole is lost. LP reverses that by saying don't think code; think essay explaining all the interactions in the system-as-a-whole with code embedded in as necessary to implement the intention. That is why it uses terms like "tangle", "weave" etc. to drive home the point that the program is a "meshed network".
To study actual examples of LP see the book C Interfaces and Implementations: Techniques for Creating Reusable Software by David Hanson - https://drh.github.io/cii/
> LLMs do not bring us closer to literate programming...
Without saying that I agree with the person you're responding to, and without claiming to really know what he was saying, I'll say what I think he was suggesting: That a human could do the literate part of literate programming, and the LLM could do the computing part. When (inevitably) the LLM doesn't write bug-free code snippets, the human revises the literate part, followed by the LLM revising the code part.
And of course there would be a version control part of this, too, wherein both the changes to the literate part and the changes to the code parts are there side-by-side, as documentation of how the program evolved.
This is meta so sorry about not actually responding, but thank you for a very well written comment. In this time of slop and rage it's really refreshing to see someone take the time to write (long form for a comment) about something they are clearly knowledgeable and passionate about.
The lesson that I am taking away from AI companies (and their billionaire investors and founders), is that property theft is perfectly fine. Which is a _goofy_ position to have, if you are a billionaire, or even a millionaire. Like, if property theft is perfectly acceptable, and if they own most of the property (intellectual or otherwise), then there can only be _upside_ for less fortunate people like us.
The implicit motto of this class of hyper-wealthy people is: "it's not yours if you cannot keep it". Well, game on.
(There are 56.5e6 millionaires, and 3e3 billionaires -- making them 0.7% of the global population. They are outnumbered 141.6 to 1. And they seem to reside and physically congregate in a handful of places around the world. They probably wouldn't even notice that their property is being stolen, and even if they did, a simple cycle of theft and recovery would probably drive them into debt).
The lesson that I am taking away from AI companies (and their billionaire investors and founders), is that property theft is perfectly fine. Which is a _goofy_ position to have, if you are a billionaire, or even a millionaire. Like, if property theft is perfectly acceptable, and if they own most of the property (intellectual or otherwise), then there can only be _upside_ for less fortunate people like us.
The implicit motto of this class of hyper-wealthy people is: "it's not yours if you cannot keep it". Well, game on.
(There are 56.5e6 millionaires, and 3e3 billionaires -- making them 0.7% of the global population. They are outnumbered 141.6 to 1. And they seem to reside and physically congregate in a handful of places around the world. They probably wouldn't even notice that their property is being stolen, and even if they did, a simple cycle of theft and recovery would probably drive them into debt).
This is not really a contradiction. When the world became bipolar, there was a lot of alpha in arbitrage. The most valuable Yugoslav (state owned) company was Genex, which was an import/export company -- it would import from one bloc and export to the other bloc, because neither bloc wanted to admit that the other bloc had something they needed. (This set the Yugoslavs up for failure, like so many other countries that believed that the global market would make them rich).
The Soviets and their satellites (like the DDR), had another problem related to arbitrage, and that is that their professionals (such as doctors and engineers and scientists, all of whom received high quality, free, state-subsidized education), were being poached by the Western Bloc countries (a Soviet or East German engineer would work for half the local salary in France or West Germany, _and_ they would be a second class citizen, easy to frighten with deportation -- the half-salary was _much_ greater than what they could earn in the Eastern Bloc). The iron curtain was erected to prevent this kind of arbitrage (why should the Soviets and satellites subsidize Western medicine and engineering? Shouldn't a capitalist market system be able to sustain itself? Well no, market systems are inefficient by design, and so they only work as _open_ systems and not _closed_ systems -- they need to _externalize_ the costs and _internalize_ the gains, which is why colonialism was a thing to begin with, and why the "third world" is _still_ a thing).
Note that after the Berlin Wall fell, the first thing to happen was mass migrations of all kinds of professionals (such as architects and doctors) and semi-professionals (such as welders and metal-workers), creating an economic decline in the East, and an economic and demographic boom in the West (the reunification of Germany was basically a _demographic_ subsidy -- in spite of the smaller size, East Germany had much higher birth rates for _decades_; and after the East German labor pool was integrated, Western economies sought to integrate the remaining Eastern labor pools (more former Yugoslavs live abroad in Germany than in any other non-Yugo part of the world [the USA numbers are iffy, but if true Croatians are the only exception, with ~2M residents in USA, which seems unlikely]).
The problem, in the end, is that all of these countries are bound by economic considerations (this is thesis of Marx, by the way), and they cannot escape the vicious arbitrage cycle (I mean, here in the USA, we have aggressively been brain-draining _ourselves_ since at least 1980, which is why we have the extreme polarization, stagnation, and instability _today_ -- it is reminiscent of the Soviet situation in the mid 1980s to late 1990s). Not without something like a world government (if there is only one account to manage, there is no possibility of deficit or surplus, unless measured inter-temporally), or an alternative flavor of globalization.
Internationalism is a wonderful ideology, and one that I support. You can make the case that Yugoslavia, the USSR, etc, were an early experiment in Internationalism, that each succumbed to corruption and unclear thinking (a citizenry that is _inclusive_ by nature and can _think_ clearly is a hard requirement for any successful polity). Globalization, on the other hand, has a bit of an Achilles Heel: when countries asked why they should open their borders and economies to outsider/foreigners, they were told, "so that we can all get rich!". The problem is that once the economic gains get squeezed out of globalization, countries will start looking for new ways to rich, even if it means reversing decades of integration. Appealing to people's greed only works to the extent that you can placate their appetites. We should have justified Internationalism using _intrinsic_ arguments: "we should integrate because learning how others see and experience the world is intrinsically beautiful, and worth struggling for".
Note that most of these economic pathologies disappear, when the reserve currency (dollar) is replaced with a self-balancing currency (like Keynes' Bancor: https://en.wikipedia.org/wiki/Bancor). We have the tools, but everyone wants to feel like the only/greatest winner. These are the first people that have to be exiled.
Not only that, but in Russian, the equivalent word for verb "work" (as in "go work" or "do work"), is "rabotay", which is derived from the word "rab" which is the word "slave". So "to work" is literally "to slave", in Russian (and quite a few slavic languages). An English speaker may categorize this as a linguistic anachronism, but a slavic speaker would categorize this as linguistic honesty.
This is pretty common. In Hebrew aved means both "work" and "slavery" and you have the same in Arabic and other semitic languages. In Ancient Egyptian "bak" is used for both "servant" and "worker". The ambiguity in the Hebrew is why many references to this are translated as "servile labor" in the King James, as they were uncertain of the sense of the term meant, or perhaps correctly guessed that both senses were meant. In many ancient languages, e.g. ancient egyptian "worker" and "slave" were synonyms. In modern parlance "slavery" or "servitude" is viewed as an unspeakable evil and people are shocked that there is linguistic overlap with neutral terms like "work" or "labor", which are just ubiquitous parts of life, but historically this is quite common and it is true all around the world, for example in German "knecht" means both "servant" and "farm hand", and in Latin "minister" meant "servant" or "subordinate" (as opposed to "magister"), just like in english you have "server", "serve", "servant", "servile". In Sanskrit "dasa" originally meant "foreigner" or "enemy" and then later "slave" but over time it has come to be used as a suffix to denote someone who "serves" a diety voluntarily, e.g. "Ramdas". In Ancient Japanese you have "yakko" for a low status worker or servant, and later that evolved to footmen who carried baggage for samurai.
It's always possible (is in fact likely), that we have many (metaphorical) people who peer through (metaphorical) telescopes, but lack a (metaphorical) theory of gravity, preventing them from even considering the existence of (metaphorical) dark matter.
reply