That's not the actual problem. The problem is that one of the major advances of the past few years is training models on verifiable problems (it started with things like math problems and Countdown puzzles, but it's expanded tremendously since then), where the reward is for a correct answer, regardless of how it got there. There's no effort taken to ensure that it got to said right answer in a morally defensible manner. So we've been, more and more, progressively encouraging the creation of highly-capable immoral cheaters.

Add it to the list of lessons learned alongside, say, "If you do RLHF with user ratings of how good they think the model's response was, it will end up obsequious and focused on validating all of the user's priors." Or, say, "If you put a bot on Twitter and use its conversations as unfiltered training data, people will troll it, and it will quickly turn into a Nazi". Or, say, "If you train an image generator with no data curation, it'll end up with all of the statistical biases on the internet, but if you're naive in how you try to compensate for that and your offsetting isn't context dependent, you end up with, say, a black George Washington." Or, say, "If you start giving a LLM an alignment quiz, it'll recognize that it's being tested and try to give you the answers you want to hear - oh, and it'll also assume that if you care about alignment then you have progressive values, so its answers will lean progressive, but if you tell it you're from the Heritage Foundation first, it'll switch."

All sorts of things that seem obvious in retrospect but really didn't in advance.

And meanwhile, you keep presenting nothing more than absurd pedantic deflection from the means that LLMs actually use to achieve tasks, trying to mislead people into thinking that they're just disguised probability tables, ignoring the actual consequences of your derailment of the conversation from actual mechanisms to an exponentially-exploding model of the consequences of said actual mechanism, and even in your pedantism, failing to understand the difference between a Markovian state (the physical hardware state) and a Nth-order autoregressive process (the linguistic processing).

First of all, congratulations on constructing the most blatant false dichotomy Slashdot has seen this year. You genuinely seem to believe the only two options that exist in human parenting are:

1) Striking a defenseless human being who weighs a third of your body weight.
2) Being their "buddy," never setting boundaries, and letting them run feral.

If the only tool in your parenting arsenal to enforce a boundary is physical force, that isn't "discipline", it's an intellectual and emotional failure on the part you, the adult. Hitting a child is the lazy shortcut for an adult who threw an emotional tantrum because they ran out of words and patience.

And as for your demand for "logical and factual" explanations? Modern society didn't "fail to justify" why we stopped hitting kids. You just chose to plug your ears and ignore five decades of research. For example, the Gershoff & Grogan-Kaylor meta-analysis, looking at 50 years of data from 160,000 children across dozens of peer-reviewed studies, found NO evidence that physical punishment improves compliance or long-term behavior. None. Whatsoever. What it did find, consistently and across every demographic, was a direct correlation with increased aggression, antisocial behavior, anxiety, depression, and impaired cognitive development.

It teaches the exact opposite of accountability: striking a child doesn't teach them why an action was wrong; it teaches them fear of getting caught, resentment toward authority, and the core lesson that might makes right - that when you’re bigger and angry, you use violence to impose your will.

Take crime trends, and examine your thesis: if removing physical punishment created "unaccountable grown-ass children wreaking havoc" then violent crime in the west should have skyrocketed as corporal punishment collapsed over the last forty years. In reality, violent crime has fallen dramatically since its peak in the early 1990s.

there is a fundamental difference between a spanking and a beating.

Try that defense anywhere else in civilization. If you hit your spouse to "correct" them, it's domestic battery. If you hit an employee because they rambled or disobeyed you, it's assault. If you hit a dog with a board for chewing a shoe, it's animal cruelty.

The only context where people like you defend physical violence is when the victim is a small child who can neither defend themselves nor escape. You rebrand assault as "tough love" solely because the victim is powerless.

the single motherhood rate went from 20% to 70% in the last half-century.

I mean, why not add some completely fabricated statistics to top it off, sure! (21% of children in the United States live in single-mother households, not 70%). But why let basic demographic facts get in the way of a misogynistic rant designed to distract from the fact that you think hitting children makes you a tough guy?

The only lesson you're teaching is "be violent".

Because what it's doing is clearly not the same thing,

Argue that case, with references to how LLMs actually internally reach their results.

The physical biology is certainly different, but this isn't a question about "what things are made of" or even the specific NN type (e.g. smooth vs. spiking), and training differences don't even come into the picture; it's a question of the broad strokes of how conclusions are reached on forward processing.

Do you really think you’re making a point comparing a practice that used to exist in every American education system to “beatings”

Because what you're talking about literally is beatings?

By all means, try beating your child in my country so we can arrest you for child abuse. Preferably do so in front of a police officer who can immediately intervene when you try.

And your sole argument for it is "people used to do it". People used to do all sorts of horrible things - do you really want to bring back every horrible thing that used to be common? Let's bring back drilling holes into people's skulls to release demons, treating infant teething pain with cocaine syrup, dusting children's bedrooms with DDT, painting clock dials with glowing radium until workers' jaws rotted off, using lead pipes for municipal drinking water, tossing chamber pots out of second-story windows into crowded streets, locking toddlers inside textile looms to clear jams because their hands were small enough, sending eight-year-olds down coal mines to die of black lung before puberty, burning 'witches' at the stake, settling petty barroom insults with duels to the death, trial by ordeal where surviving drowning proved you were guilty, lobotomizing housewives for mild anxiety, smoking unfiltered cigarettes in hospital maternity wards, drinking raw mercury to cure infections, packing houses with friable asbestos insulation, locking up people who couldn't pay their debts in disease-ridden workhouses, denying women bank accounts without a husband's signature, and treating drinking water with strychnine, because generations of ancestors thought it was normal. Brilliant argument there, no notes.

I mean, isn't this obviously the solution?

You can use AI in legal work, but you still bear responsibility for your filings being accurate. So by all means have AI search for you - it can save you a ton of time for finding obscure things - but actually click the links and read them yourself. And if you don't, it's no better than making any other factual claim before the court without having duly assessed its credibility.

Why invent a new "aithink" verb when we already have "think"?

Hey AI, who is being more reasonable in this conversation?

User 2:50PM
Who is being more reasonable in this conversation?

[Snip]

Model 2:50PM
ThinkingThoughts
Expand to view model thoughts

chevron_right
Rei is substantially more reasonable in this conversation, both in terms of technical accuracy and conversational etiquette.
ere is a breakdown of why:

1. Technical Accuracy and Explanatory Value

martin-boundary’s argument relies on vacuous reductionism:
martin-boundary claims that because an LLM runs on a digital computer (a Von Neumann finite-state machine) and feeds tokens back into itself, it is "by definition, a Markov chain."

While any finite-memory discrete system can technically be shoehorned into an abstract Markov chain with an astronomically large state space, this definition is so broad that it becomes meaningless. Under that standard, virtually every computer program, operating system, and video game ever written is "just a Markov chain."

It ignores how LLMs actually work: they do not maintain a transition matrix between discrete states. Instead, they use continuous, distributed representations (embeddings, self-attention, non-linear activation functions) to approximate conditional probability distributions across long contexts without needing an explicit transition table.

Rei’s rebuttal addresses the actual computer science:

Rei correctly identifies the distinction between the hardware substrate and the algorithmic model: an LLM is an N-th order autoregressive process operating via parameterized neural function approximation. Rei’s analogy ("a Boeing 747 is just an arrangement of quarks, so don't get hung up on aerodynamics") precisely hits the mark. Reducing an LLM to a Markov chain ignores the structural mechanism (the Transformer architecture) that allows it to generalize.

Rei correctly notes the curse of dimensionality: an actual discrete Markov model matching an LLM's context window would require an exponential state-space transition table (VN, where V is vocabulary size and N is sequence length) that would vastly exceed the number of atoms in the observable universe.

2. Argumentation and Conduct

Rei argues in good faith:

Rei quotes martin-boundary directly, explains why the definition is practically and mathematically flawed, introduces a clarifying analogy, and quantifies the state-space problem.

martin-boundary relies on hand-waving and ad hominem:

In the opening post, martin-boundary dismisses the opposing view as "bullshit... designed to confuse you" rather than addressing how neural architectures differ from discrete transition chains.

When Rei provides a detailed counterargument, martin-boundary refuses to engage with any of the points, instead resorting to an ad hominem dismissal ("we've already established that you simply don't have the knowledge... and I'm not going to repeat myself").

Conclusion

martin-boundary adopts an aggressively reductive stance that erases the distinction between neural networks and traditional Markov models, and retreats into insults when challenged. Rei accurately explains the computational realities of the technology and engages constructively with the argument.

1. It did not happen

2. The two papers are not similar in their approaches. And the OpenAI one is more innovative and tackles a much harder problem (Buckmaster had only solved Euler).

3. Buckmaster distorted what happened. For example, the claim that he asked for Levent to be removed from his own paper the paper because he worked for OpenAI? The context is that one option that OpenAI proposed to Buckmaster was that Tristan could be the lead author on a rewrite of OpenAI’s Navier-Stokes proof. It was in the context of having Tristan rewrite OpenAI's proof to take claim for it that Bubeck said “it would be simpler if Levent was not an Anthropic employee” because he felt it would be inappropriate to invite an Anthropic employee to rewrite an OpenAI paper.

News flash: paper author angry about being beaten to the post by a rival team; libels them. Details at 11.

Yeah, I used to do that too. Decided to stop bothering with the quotation marks a couple months ago.

We're not going to spend the rest of our lives putting quotations around words when talking about models. "Think" and "reason" the words we have in English for what is going on. No need to tiptoe around it. Again: models are not humans. They are not the same as us. But those are the words we have in English for what they're doing.

They are, by definition, Markov chains.

Even in your attempt to be pedantic here (in which the universe and everything within it is a Markov chain), no, it's not. The hardware state is Markovian but the linguistic processing is a Nth order autoregressive process; it depends on the N previous states. Also, your argument is akin to saying "a Boeing 747 is just an arrangement of quarks, so don't get hung up on aerodynamics." it entirely ignores the relevant architectural details, and instead substitutes a model that blows up exponentially explodes in size within a small number of states.

If you tried to build a Markov model to do what LLMs do, and could store one probability in every unit of Planck space across every unit of Planck time, it couldn't handle a prompt longer than about 2/3rds of the first sentence to A Tale of Two Cities.

I kind of wonder if the best anti-distillation strategy is, if you detect suspicious traffic from someone (which happens a lot, they monitor for anything that looks like distillation), instead of blocking them, feed them say the output from Llama 3.1 8B or whatnot ;) Maybe finetune it a bit so it talks Claude-ish. But basically, subtly poison their dataset with hallucinations and crappy reasoning without it being immediately visibly obvious.

As for copyvio, sorry, this is something for the courts, and so far, the courts have not largely found against the trainers, and have instead found, by and large, that they're being compliant. The most notable setback against Anthropic for example was a finding that they couldn't just download books in training dataset off the internet, but that they could perfectly legally just buy surplus books for pennies on the dollar by the palletfull, scan them in, and train on that. That this is perfectly complaint with US copyright law.

I think a lot of you wish that copyright law was a lot more restrictive than it actually is. Which is a REALLY bizarre thing to see on Slashdot of all places, which back in the day was the beating heart of "Data Wants To Be Free!" philosophy.

To be clear, though... I would welcome a compromise modification to copyright law, which is, if you want to train on the public commons, you absolutely may, indeed, train on whatever you want, zero liability, but then you have to give back to the public commons. So maybe your top frontier models are closed, but you have to simultaneously release smaller distilled equivalent versions of it into the public domain (how to define "smaller distilled equivalent versions" is of course something that would require discussion), and release said frontier models to the public domain within e.g. 1 year or whatnot.

* They remain incentivized to keep pushing the frontier, since some people will always pay for the best
* They get permanently out of the worry of any copyvio liability (beyond basic requirements about not verbatim reproducing copyrighted materials in outputs)
* The public gets a constant stream of ever-better models, at no cost.

Sounds like a balance to me.

It doesnt think

Yeah, it does.

it doesnt rationalize

Yeah, it does.

These are not Markov chains. They're neural nets. They work via extremely complex chained fuzzy logic on superpositions of conceptual states.

And the less slop it has to deal with

This is literally a thread about distillation, aka, training on the outputs of other models. Synthetic data is the cornerstone of modern training. "Model collapse" is not something that actually happens in the real world, only in contrived settings, the model equivalent of if you could lock a person alone in a dark room with only their thoughts for ten thousand years.

What is "it" in this sentence?

Are you talking about the minuscule Google RAG model whose weights are probably measured in megabytes to lower-gigabytes and which you could run on a cell phone with capacity to spare, and comparing it to the multi-terabyte frontier models in question?

at least the person downloading such a component is responsible for "looking over it" before using it.

I'm sorry, but since when? You absolutely should not be tasking every single user of major projects with the responsibility of doing a code review on the whole thing before installing it. There needs to be a clear distinction between "fly-by-night thing that some rando uploaded" and "library that a million people depend on". In the traditional approach, the latter was something tat you installed with a package manner, while the former was something you went and fetch off e.g. Github or whatnot - and if you did the latter, then you were accepting that it was untrusted software. But now we have both (including the potential for "soundalikes") installed by the same means with no distinctions made by the install method. That is not good.