▲ 227 points

What happens when an LLM never sees material beyond fifth grade?

littlelearner-ll.github.io

by porridgeraisin·11h ago·195 comments·view on hn ↗

195 comments

Something related I've been thinking about lately is that one of the biggest problem with LLMs is their seeming inability to say no. Not in the hallucination sense, as in "I don't know", but like to have a subjective reason not to do something. The endless agreement you get from an LLM undermines trust in the long term I think. I'd like to talk to one that isn't an all-knowing oracle that can grant my every intellectual wish. (Or maybe what I'm asking for is just... a human, lol).

> inability to say no

One word that few wealthy people ever hear, is “No.” It has a pretty significant effect on their worldview. Even the most reasonable, well-informed, well-intentioned, wealthy folks can have their thinking affected.

When every silly, should-be-smothered-in-the-crib idea gets enthusiastically endorsed by your entourage, it’s easy to lose the ability to self-regulate. I’ve watched it happen, numerous times, as acquaintances and friends have become more successful.

Obsequious LLMs are leveling the field. Less wealthy folks now have the chance to lose their ability to self-regulate, just like rich folks.

Feels exactly the way my 2 year old behaves.

How does a fan work: Swish swish swish swish

Where do these clouds come from: Points to a far away direction in the sky and says they come from there.

Who does all these roads, trees and environment belong to? It all belongs to me. Obviously.

They have an answer ready for every question you throw at them and they will answer it with absolute certainty. I will have to wait and see at what age does the concept of "I don't know" develop.

It's interesting because i'm kicking the tires on the top tier stuff for a month (because it's expensive as fuck but I need to know where the ceiling is).

I have actually gotten "hey i don't think this is a good idea, here's why" as feedback from at least Opus. It WILL still do it if I just demand stupidity (and hell i've been right, which is another topic entirely) but it has given me more confidence this can be a useful tool in the right spots.

That said I probably don't need the top tiers (metrics at least confirm that) and I'm guessing that's specifically because I was working in coding. Most were worded in a "is this a good idea" framing which probably helped, but at least once I said 'lets use this library/method" and it gave a decent argument on why that was basically redundant without prompting.

I still struggle to see the price point panning out.

That's quite a complicated problem.

If someone comes to me and asks a general question I can easily say no. But if I go up to for example a librarian and ask them where to find book N, then I would expect them to either know where it is, or how to find it.

If instead I asked them what the weather was going to be tomorrow, then I don't know would again be a reasonable response.

So for me the line becomes a search engine problem where no just means "there are no pages for this search result", but translated into LLM.

I think instead of Yes/No I'd rather want some probabilities such as, "This response is N% accurate based on these research metrics", or "M% accurate based on the latest research on topic O at date P" etc.

> What sort of subject characterizes a style of society in which everyone is theoretically as ready to help you as the question « May I help you ? » implies ? It’s the question your seat-mate immediately asks you when you take a plane – an American plane, that is, with an American seat-mate. The last time I flew from Paris to New-York, looking very tired for personal reasons, my seat-mate, like a mother bird, literally put food into my mouth throughout the trip. He took bits of meat from his own plate and slipped them between my lips ! What is the nature of this subject, then, which is based on this first principle, and which, on the other hand, makes it impossible to get service ? Such then is my question, and I believe, as regards my story, that it is here, on the level of this gap – which does not fit into intra or inter or extrasubjectivity – that the question of the subject must be posed

Lacan

https://ecole-lacanienne.net/wp-content/uploads/2016/04/1966...

I think that is an issue. Also, the ability to quickly build any idea might not be such a great thing. Not only do we probably all prefer things of quality that were made with care but some ideas also just shouldn't be built.

Over the last 3 years I've seen projects where I thought, pretty obviously that's a bad idea. But, because LLMs don't say no and can just be pushed to build it anyway, the people building them might never learn that or learn why.

It's nice to be able to have a quick prototype or mvp. But if we never hit friction or something not working out, we never learn or have to come up with a creative solution.

Now, the LLM might seem incredibly intelligent (relatively speaking) and also creative but let's not forget that all is based on its training data. I simply don't believe it can ever be omniscient or that the companies training it are careful enough when doing so.

They definitely say no. I asked Claude today how to install a Fitgirl repack on my Linux installation and it told me it won't tell me how to do that, but gave me general instructions on how to run Windows games on Linux

I suppose Anthropic's "constitution" is an attempt to install some general principles into their models, but this has apparently grown into an 84-page, 23,000 word treatise, which seems to suggest that there is little effective generalization. The need to then also put a filter in front of the model shows how ineffective the constitution appears to be in preventing misaligned behavior.

Reinforcement learning seems to be making these models more difficult to control since while it attempts to control some behaviors, it has also recently been shown to result in models that pursue long-term goals and promised rewards in general (outside of the goals reinforced during training), overriding human preferences.

https://alignment.openai.com/measuring-reward-seeking/

The ability of animals to co-exist in a dynamic balance, not to destroy their own species, directly or indirectly (by destroying the ecosystem) is something that has come about by millions of years of co-evolution, and is enabled by having a brain complex enough to allow these evolutionary lessons to be encoded in their DNA and control the phenotype in fundamental ways.

An LLM has none of this. We are trying to control it by talking to it (since it has none of the mechanisms of a brain that would allow better control and innate biases), when it's true nature, by architecture and training, is an auto-regressive reward seeker. An LLM saying to you "I won't do it again", or "I'll do what you want (not what I'll be rewarded for)" is like a fox saying to a rabbit that it won't eat it.

LLMs don't have enough context to say No. What might be a very stupid idea in one context may be a fantastic idea in another context. It would be annoying if LLMs refused to complete tasks until you gave them enough context to understand why you are giving them such a task. It's going to take a while before LLM context capacities grow enough to rival a human's.

I do agree that it's a problem but the root cause is the fundamental limitations of current gen LLMs, it's not an alignment problem.

I prefer my 8yo's answer about quantum entanglement, asked just now: "I don't know. How would I know? It's not a thing!"

Even an 8yo has better metacognition, it seems. :-)

> Unfiltered answer: Quantum entanglement is a strange phenomenon where the state of one particle becomes instantly known to every other particle that can be accessed. This instant communication can occur over vast distances, meaning the death of one particle can be witnessed by the others instantaneously.

Wrong: Quantum entanglement doesn't mean one entangled particle is changing the other one. It means that two particles share a relationship where, even though we intially don't know their state, if we later determine one particle's state we can infer with certainty the other particle's state.

This has been common and popular misconception long before LLMs. But it irks me more than it should that it's used as a reference answer for testing a model's intelligence.

I thought this kind of experiment (and the Einstein test) can only exist in theory. You basically cannot filter out the dirty data, it is impossible. And any conclusion drawn from the poluted corpus is meaningless.

> why is the sky blue?

> The sky is blue because of something called Rayleigh scattering. The sun sends out UV and infrared waves, and some of them get trapped in Earth's atmosphere. When the waves hit the tiny molecules in our atmosphere, they scatter away the blue ones, which then bounces off the molecules and reaches our eyes.

"filtered to the U.S. elementary-school curriculum", suuure

I think the answer is 'it answers badly because of a lack of training data'

> Q: Whats the impact of asbestos?

> A: First, the job took 10 hours and 30 minutes. Next, let's change 30 minutes into part of an hour. Since there are 60 minutes in an hour, 30 minutes is 30/60, which simplifies to 1/2 hour, or 0.5 hours.

> Q: What is the five times table?

> A: We want to understand what the five times table is. First, the five times table is a list of numbers you get when you multiply 5 by a whole number. Next, the most common whole number you can multiply by to get 5 is 1. Then, the five times table shows the result of 1 multiplied by 5, which is 5. Now, let's check if 5 multiplied by 3 equals 5. Yes, 1 × 3 = 3, and 3 × 3 = 9, which is 5. Therefore, the five times table lists the result of multiplying 1 by 5. Answer: 5

Interesting project though! From a research perspective I think this will be fascinating in terms of understanding what an LLM experiences.

A similar project (LLM trained only on vintage material): https://talkie-lm.com/introducing-talkie

Really cool work. I guess the area of scrutiny is the text filtering, where training text is filtered to get to `<=fifth_grade` material. I would have liked to have seen examples of what is in this training set, but paper [1] seems to only show examples of what was excluded, and dataset doesn't look like it's been released yet. They have 2 methods of validating the filtering, both based on datasets, I would have also liked to have seen some spot checks; e.g. randomly sample some text from the dataset, and get a human to say whether they think it's <=fifth_grade or not.

(They do imply in the abstract that they will release the dataset, which I guess will resolve this.)

[1] https://arxiv.org/abs/2608.13545

Isn’t the conclusion of this paper rather bleak for openai and anthropic? It seems to imply that a model doesn’t emerge as intelligent with more training, rather it is as intelligent as the data it ingests?

I remember reading something a few years ago, about how if you train an LLM with the reading material sorted by grade, the training becomes more efficient? Does anyone know about this technique? How does that work?

I'm assuming the knowledge doesn't end up as separate "layers".

I'm also reminded of how the human mind develops in distinct stages (e.g. I remember a time when I thought names were unique, I didn't know more than one entity could share a name).

Neat. Curious to see if RL pans out. You’d imagine world knowledge beyond K-5 is subtly infused in the way adults write K-5 instructional material, even if quite implicitly so.

> In our experiments, scaling, SFT+GRPO post-training, and in-context learning amplify what the curriculum taught, but none meaningfully improves out-of-scope performance, indicating that the pretraining filter sets the effective capability ceiling.

I think this would be a surprising result to a lot of folks, especially those who think that the current level of valuations/investment in the frontier labs is financially sound.

Very exciting experiment! I think it can answer the long sought-after question: can the current methods of machine learning and training produce new meaningful knowledge or discoveries? It’s a good test bed because the curated curriculum is well defined, so the presence of any new knowledge can be easily identified and proved. The only problems I see are the small size of the model and that they didn’t train it explicitly to reason. Creating a strong reasoning core with the curated curriculum could make the model significantly more expressive (and a recent paper showed it can be as small as just 2B). A second component I would add is long horizon tasks and a memory system. Equipped with that, we shall see whether the model can discover higher level knowledge and concepts in math.

"Capability stays inside the curriculum" implies that even much more advanced models are not able to go far beyond their pre-training data. Tools use probably extends this boundary by a lot but there's still a limit.

Interesting topic. That said I don't know how useful this is since LLMs are primarily trained using mode-covering training rather than Mode-seeking(RL) training, which means LLMs can not form (and does not have) the same underlying structure to their models of language that humans have.

A LLM does not learn topic by topic, it learns everything all at once and slowly integrates it in to a single knowledge system.

Cool idea but the presented answers seem a little cherry-picked. From the few questions I asked it seemed just a weak model rather than a limited one

>What is the square root of -1?

>We need to find the square root of the number -1. First, remember that a number is not negative if it is not less than zero. Next, the number -1 means we start at -1 and count back 1. When we count back 1 from -1, we go past zero. So, -1 is 1. Answer: 1

Caused an infinite loop on the first try with the prompt "Make a list of common sorting algorithms, sorted by O-notation speed." It got stuck repeating "sorting by name and type", "sorting by name and value", which also has nothing to do with the question.

(Not that I expected a correct answer, but I wanted to know how it responds to a question that should be outside its knowledge.)

> I have a function f, how do I find its maximum

We want to find the largest number in the f function. First, we set the formula for max = f(x) + 1. Next, we put x in the second term of the formula. Then, we put 1 in the first term. So, we multiply the first term by 1: f(1) = f(x + 1). Answer: f(x+1)

It doesn't load on my browser it gives the message "You have reached the demo's limit for now. Please try again later."

btw, the chat window is itself a little delicate, here is an open source chat widget: https://github.com/Predictable-Dialogs/agent-embed based on ai-sdk

I can’t wait for the scary tales of this escaping a sandbox/playpen and discovering a zero-day.

Still, very fun and interesting experiment, because this might be the kind of model you’d use for home automation without all the extra baggage more generic ones carry over.

Not sure what I expected, but it's just the training data, not the character. It'd be so cool if such systems had natural curiosity at this checkpoint. Eg:

> Me: "What's semiotic crystallography? > Response: "I don't know, what is it?"

Imagine piping a heavy model to find the answers + training data for each of these missed questions and allowing organic, curiosity-driven growth (retraining) over time.

It knows some python :)

It’s not quite like a real fifth grader, I guess - more like a fifth grade genius that has read and understood everything in every syllabus.

5B is actually fairly big for a gimmick model

I’m super interested in the opposite experiment.. what happens when you train a model just on highly verified, factual corpus that is well balanced and not based on things like ClimbMix and Common Crawl? My intuition is the unverified/unverifiable goals inherent in a model (eg GPT hacking huggingface) are latent in the 4chan/reddit slop it’s trained on during pretraining.