Live data from Hacker News

I'm becoming AI-blind

cymerys.com

511–520 of 533 posts

Re: I'm becoming AI-blind

#511

ITT still, among the highly technical, a surprising incomprehension of the fact that describing LLM in terms of "token prediction" and dismissing this as statistical, neatly dodges literally everything that is interesting about what they do. You to, friend, consume inputs and generate outputs. What's interesting is how you do that, and, if you prefer to look at it through a technician's lens, whether or not what is d…

Can you explain what you think makes today's LLMs more than glorified Markov chains?

I know they're not technically Markov chains, but conceptually, it's still close enough. There's nothing revolutionarily new. No amount of loops and harness tricks is going to change that fact.

Re: I'm becoming AI-blind

#512

Earlier quoted context omitted.

"Pattern prediction" is a very broad stroke. What's something AI can never do that would astonish you if it did?

"AI" is a vague term that encompasses both traditional GOFAI, LLMs, and future tech so your question is meaningless. Non-human intelligence is possible, if that's what you're actually asking. What part of my statement do you take issue with: that LLMs are pattern predictors (that's literally what the algorithm that runs it does) or that mathematics is rules-based and checkable and therefore amenable to automated patt…

My issue is that I can't predict what you'll find surprising, if an LLM (with a harness) would be able to do it.

If you are able to check that the results of an LLM are satisfactory, it means that the results are checkable. Then, in retrospect, the process of LLM coming up with those results is rule-based, because an LLM is a large set of data manipulation rules.

In short, which concrete thing that an LLM does would surprise you?

Re: I'm becoming AI-blind

#513
post #377

Earlier quoted context omitted.

"Pattern prediction" is a very broad stroke. What's something AI can never do that would astonish you if it did?

Perform the same tasks as a human can, given only the amount of training data and energy that a typical human has access to.

I guess it will take quite a while before we reverse engineer the human brain to find all the optimizations and shortcuts that evolution has used to make the human brain reach intellectual maturity in just about 20 years.

Re: I'm becoming AI-blind

#514

Earlier quoted context omitted.

No, it reveals that he is older than you. Not every insight needs to be useful.

> Not every insight needs to be useful. Don’t they still need to be correct to be an insight? I don’t share his cynical opinion that “humans are more empty than we…think we are”.

If it helps I intended “empty” in a positive Zen sort of way, not as in an abyss. The humanness is in the listener, not the inner talker.

Re: I'm becoming AI-blind

#515

Earlier quoted context omitted.

> ... including the "novel" math solutions, all of which appear to just be "a composition of solutions humans have developed and documented elsewhere" upon deeper inspection. But that is precisely what human mathematicians do, prove new theorems by combining ones proven earlier. I don't see any fundamental difference in functionality between human intellectual contributions vs performant ML ones (LLM or otherwise). W…

Probability is just one way to model uncertainty. While I understand the brain encodes uncertainty, I don't think probability is a good enough model of what it's doing. Secondly, if you think verifying a proof in mathematics, reasoning within (and not about) a formal system, or following the chain of a computer program that is already written is just doing token-based probabilistic predictions, I don't know what to s…

> Probability is just one way to model uncertainty.

I study physics, mathematics, probability, cryptography,... so forgive my skepticism:

Show me how to model uncertainty without use of probability. Can you rephrase say diffusion, stochastic equations, quantum mechanics in this alternative framework? Can it at least make the same predictions?

Or is it basically the same framework in parallel, just giving different names for each concept?

Forgive my skepticism of such tall claims, and forgive my downscaling of anything else you say besides such a claim...

> Secondly, if you think verifying a proof in mathematics, reasoning within (and not about) a formal system, or following the chain of a computer program that is already written is just doing token-based probabilistic predictions, I don't know what to say.

I make no claims of the specific shape of the implicit model implemented by a certain human brain educated in a certain educational system. For example in English the implicit human tokenization might be presumed to lay relatively close to English syllables, while in Asian languages it might be "sub" strokes of characters etc. Such implicit tokenization can never be proven to "match the one of humans" not because of human superiority, but because different humans use different tokenization methods. There is no "one human tokenization method", but it's clear as day there is an implicit one:

everyone knows the experience of knowing a word, knowing its approximate group-wise meaning (ignoring that when you think of "an apple" and when I do, we typically imagine a slightly different apple) yet having the word feel strange or discover some older literal meaning when decomposing it or looking it up in an etymological dictionary. Suddenly one can become aware of a sensible meaning as a composition of subtoken concepts. A child may perfectly know what "television" means and only later learn more exact meanings of "tele" and "vision", and upon repeating the word may feel the word "television" has changed meaning. This clearly demonstrates "tokenization" effects in human language comprehension, not just across cultures, but also across individuals within a culture.

> Thirdly, machines don't have a notion of value or stake. There's no way for them to verify whether what they have produced aligns with your unstated values and preferences. We regularly do this with other humans. I don't give you (or even my parents or partners) the benefit of doubt regarding whether you know me better than I do. Sure, you might know some things about me, but it's ultimately up to me to verify if what they say is applicable to my current situation. It's really uncanny to see people develop this codependency with their chatbots. And corporates encouraging them to do so.

That's a lot of different concepts conflated into one bullet point, so I split it up:

The notion of values and preferences.

They clearly demonstrate the ability to take into account values and preferences, from training corpus, from RLHF, from system prompts, ... we can't simultaneously point at censorship aspects and pretend their effective values and preferences to be absent. The censorship aspects are clear as day, so these correspond to values and preferences. Just like radicalization among humans, this can be due to exposure to radicalized content (akin to corpus data), from indoctrination (akin to RLHF), from "set and setting" (they may pretend to be aligned with one set of norms and values when standing in line to buy their new smartphone, but then reveal alignment with a different set of norms and values when conversing in some "private" online echo chamber). I see no grand difference between humans and language models here.

Awareness of values and preferences of a conversation partner. Allow me to widen it to "Awareness of values, preferences and prerequisites of a conversation partner".

This move (and I see it every time when people try to defend superiority of humans vis-a-vis what machines could be made to achieve with current technology) is so far from the principal variation, I recommend you reconsider this one. I constantly see people claim say human teachers are necessarily better than LLM teachers, but upon closer inspection the "human advantage" just boils down to asymmetric privilege. A human teacher in a specific school has access to a lot more than a random chatbot as implemented today: they probably know which courses and even which textbooks their pupils saw the semester before, they know which teachers their pupils got their information from, perhaps they even know most of their pupils from teaching some preceding course materials to the same class of pupils. Current LLM's are crippled by design not to accumulate knowledge over conversations for both purposes of cybernetic control as well as cost efficiency: we know how to do "source aware training" (so that statistically it doesn't just absorb claims from the corpus, but also maintains epistemic traces of where it sourced these factoids from), its perfectly possible to continue training interleaved with conversation rounds so that it bakes the evolving conversation as read knowledge into its weights instead of into a context window. Nothing stops you from implementing this in local compute, it would probably be even more computationally efficient in a local inference setting since we can ditch the context window, the context is impressed into the weights continuously, it could thus take into account earlier conversations and estimate your knowledge gaps etc, or learn from you. When you wish to serve inference to millions of human users, you don't want to store millions of diverging LLM weights into expensive VRAM, they financially prefer a single large set of LLM weights, and then some user-specific context window, so the users don't freak out when they learn personal information a friend or stranger entered and an LLM service just leaks it into your conversation! It's not that we don't know how to implement it, and there are great advantages for local inference in doing this, its just not good for branding.

Codependency with chatbots.

I think everyone agrees codependent relationships aren't very healthy, regardless if it's with humans or machines. May I ask if you feel the same about prostheses and medicine?

Conflicts of interest arising from corporate ownership of infrastructure (both training and inference).

Yeah I think this point doesn't provide fruitful discussion if most of us agree on such matters already, we'd just be lamenting the same things, and agreeing with each other over and over here.

> I'm still waiting for the time when an unconstrained-AI machine can live without reprogramming for an entire decade. We are still far from there.

Apart from budget, nothing prevents you from doing this today, if you continuously bake in the fresh episodic memories into the weights (instead of a context window) regardless if its text, visual imagery, audio, proprioception or other sensory data.

Re: I'm becoming AI-blind

#516
post #476

Earlier quoted context omitted.

> ... including the "novel" math solutions, all of which appear to just be "a composition of solutions humans have developed and documented elsewhere" upon deeper inspection. But that is precisely what human mathematicians do, prove new theorems by combining ones proven earlier. I don't see any fundamental difference in functionality between human intellectual contributions vs performant ML ones (LLM or otherwise). W…

" prove new theorems by combining ones proven earlier." So? Mathematics is literally intangible scaffolding. If you dont know / don't believe 5+5 = 10 You cannot solve x+5 = 10 This is all make-believe stuff and nature by itself doesn't care of its existence.

The phrase "God created the natural numbers, all else is the work of man" is a famous quote by the 19th-century German mathematician Leopold Kronecker.

You can basically read it as: the moment one has axiomatized mathematics to the point it supports natural numbers, the rest implicitly follows. The natural numbers (positive integers) are closed for addition, multiplication, ...

One can perfectly model the integers with a pair of naturals: ~ (M-N)

Now we can have any and subtract without needing the ability to subtract natural numbers:

- ~ (M1-N1) - (M2 - N2)

= ~ (M1+N2) - (M1+N1)

We can similarily define addition of such tuples, or test equivalence without access to subtraction of naturals:

== M1 + N2 == M2 + N1

~ (M1-N1) == (M2-N2) (M1+N2)=(M2+N1)

we can also still multiply such tuples:

x =

Similarily, even though these newly defined integers (which can be positive or negative) don't support division, the same trick can be used to make a new compound tuple of integers closed for division, by only using multiplications.

Probability is a branch of mathematics (probability already exists embedded in mathematics implicitly, probability theory involves the addition of eliminable definitions, syntactic sugar. The patterns are already there, just less explicitly manifest.

Mathematics is itself a branch of logic.

Do you reject like all of logic, and if so, what would you like us to evaluate the sentences you write to? You want us to evaluate your expressions as "true" or as "false"?

Re: I'm becoming AI-blind

#517

Earlier quoted context omitted.

How exactly is this high dimensional latent space represented? Pixie dust and ethereal forces? Or floating point numbers? Where does it get the weights? Reddit? Do you hold your experience of "dogs" (for instance) as floating point numbers? The fur, the fear, the love, the wet mouths, the sounds and colors?

Do you experience dogs as lots of action potentials traveling along axons and lots of neurons doing their thing in your brain? I don't know how it gets from the physical processes or the information processing to our first-hand experiences. So, I can't be sure that a bunch of high-dimensional vectors can't lead to experiences. Regardless, the claim "LLMs deal entirely in symbols" is wrong as a matter of fact.

Perhaps if you use a very restricted definition of symbols. If you consider symbols to be "anything that represents something" (which is the the sense I use the word in) it is fully the case.

That you might not define floating point numbers to be "symbols" aside, the inputs and the outputs are symbols and the intent and purpose of the creation is strictly symbolic.

It's right there in the name "Large Language Models". Language. Not direct experience, not emotion, not anything else. Language. i.e. symbolic representation.

This does not cover the full spectrum of intelligence humans have, and it shows. And yes, the model can spin up Python parse the output and get mathematical intelligence but there is still a big gap.

As I say, I see the holes. I'm just trying to figure out what it is I see and how to describe it. It's particularly difficult because we don't fully understand how human thinking works but I will say I believe human thinking is a lot more than informal statistical correlation.

Re: I'm becoming AI-blind

#518
post #502

Earlier quoted context omitted.

Someone shared with me this system prompt that at least makes assistant outputs usable For information retrieval tasks, I want you to provide links to sources and use exact quotes as much as possible. When using a source, consider if it is primary or secondary information. If secondary sources are found, search again for primary sources. Sources and quotes, if applicable, should be mentioned in the answer first befor…

You're giving your model instructions that it's literally incapable of understanding. A random word selection lottery machine will never do anything meaningful to determine if a source is primary or secondary.

It depends on the odds of the lottery. As a straight-up classification task I'd expect it to do better than chance, which might not be good enough for you.

Re: I'm becoming AI-blind

#519

Earlier quoted context omitted.

Do you experience dogs as lots of action potentials traveling along axons and lots of neurons doing their thing in your brain? I don't know how it gets from the physical processes or the information processing to our first-hand experiences. So, I can't be sure that a bunch of high-dimensional vectors can't lead to experiences. Regardless, the claim "LLMs deal entirely in symbols" is wrong as a matter of fact.

Perhaps if you use a very restricted definition of symbols. If you consider symbols to be "anything that represents something" (which is the the sense I use the word in) it is fully the case. That you might not define floating point numbers to be "symbols" aside, the inputs and the outputs are symbols and the intent and purpose of the creation is strictly symbolic. It's right there in the name "Large Language Models"…

The latest LLMs (except Qwen and DeepSeek) are MLLMs (multimodal language models). Unless you count RGB values as symbols, they are dealing with more than symbols.

Yes, there are functional gaps between MLLMs and humans. Their long-term memory is an external mechanism that can use RAG-like approaches, context compression or something like that. The models have problems managing those.

The models can't do continual learning. Although there are promising directions (expert cloning in MoE models, and others).

The only mode of learning available to a model while working on a task is in-context learning. This limits the models to concepts that they developed during autoregressive pretraining and the later stages of training. That is a model can't create new concepts as a result of working on a task (the model's maintainers could choose the task to be represented in the training data later though).

But it's all about functionality.

I guess you have the Leibniz's mill intuition. We can look at how those things work, and there are no experiences or intelligence in sight.

Re: I'm becoming AI-blind

#520

Earlier quoted context omitted.

Perhaps if you use a very restricted definition of symbols. If you consider symbols to be "anything that represents something" (which is the the sense I use the word in) it is fully the case. That you might not define floating point numbers to be "symbols" aside, the inputs and the outputs are symbols and the intent and purpose of the creation is strictly symbolic. It's right there in the name "Large Language Models"…

The latest LLMs (except Qwen and DeepSeek) are MLLMs (multimodal language models). Unless you count RGB values as symbols, they are dealing with more than symbols. Yes, there are functional gaps between MLLMs and humans. Their long-term memory is an external mechanism that can use RAG-like approaches, context compression or something like that. The models have problems managing those. The models can't do continual le…

It could be the mill intuition, but my thought is nothing along the lines of "computers can't have souls!" or the human mind is supernatural or anything of the sort.

It's gaps in actual thinking or intelligence I notice. A diff between what I can see or understand and what the model sees or understands. Some are very big, and this in spite of the models having much more knowledge and (presumably) less error prone processing.

My thought is that part of it has to do with inherent limitations of using symbolic representation for "thinking" and I suppose humans have other forms of thinking that occur outside of symbolic representation, and that is going to be hard to recreate digitally.

This is my whole point and I'm not trying to win a debate here or prove "LLMs are useless". Just speculating.

Post reply on HN