[HN Gopher] LLMs are getting better at character-level text mani...
       ___________________________________________________________________
        
       LLMs are getting better at character-level text manipulation
        
       Author : curioussquirrel
       Score  : 128 points
       Date   : 2025-10-13 19:39 UTC (1 days ago)
        
 (HTM) web link (blog.burkert.me)
 (TXT) w3m dump (blog.burkert.me)
        
       | simonw wrote:
       | If you take a look at the system prompt for Claude 3.7 Sonnet on
       | this page you'll see: https://docs.claude.com/en/release-
       | notes/system-prompts#clau...
       | 
       | > If Claude is asked to count words, letters, and characters, it
       | thinks step by step before answering the person. It explicitly
       | counts the words, letters, or characters by assigning a number to
       | each. It only answers the person once it has performed this
       | explicit counting step.
       | 
       | But... if you look at the system prompts on the same page for
       | later models - Claude 4 and upwards - that text is gone.
       | 
       | Which suggests to me that Claude 4 was the first Anthropic model
       | where they didn't feel the need to include that tip in the system
       | prompt.
        
         | ivape wrote:
         | Or they'd rather use that context window space for more useful
         | instructions for a variety of other topics.
        
           | astrange wrote:
           | Claude's system prompt is still incredibly long and probably
           | hurting its performance.
           | 
           | https://github.com/asgeirtj/system_prompts_leaks/blob/main/A.
           | ..
        
             | jazzyjackson wrote:
             | They ain't called guard rails for nothing! There's a whole
             | world "off-road" but the big names are afraid of letting
             | their superintelligence off the leash. A real shame we're
             | letting brand safety get in the way of performance and
             | creativity, but I guess the first New York Times article
             | about a pervert or terrorist chat bot would doom any big
             | name partnerships.
        
               | astrange wrote:
               | Anthropic's entire reason for being is publishing safety
               | papers along the lines of "we told it to say something
               | scary and it said it", so of course they care about this.
        
               | ACCount37 wrote:
               | I can't stand this myopic thinking.
               | 
               | Do you want to learn "oh, LLMs are capable of scheming,
               | resisting shutdown, seizing control, self-exfiltrating"
               | when it actually happens in a real world deployment, with
               | an LLM capable of actually pulling it off?
               | 
               | If "no", then cherish Anthropic and the work they do.
        
               | littlestymaar wrote:
               | You do not appear to understand what an LLM is, I'm
               | afraid.
        
               | ACCount37 wrote:
               | I have a better understanding of "what an LLM is" than
               | you. Low bar.
               | 
               | What you have is not "understanding" of any kind - it's
               | boneheaded confidence that just because LLMs are bad at
               | agentic behavior _now_ they 'll remain that way forever.
               | That confidence is completely unfounded, and runs
               | directly against everything we've seen from the field so
               | far.
        
               | littlestymaar wrote:
               | > I have a better understanding of "what an LLM is" than
               | you. Low bar.
               | 
               | How many inference engine did you write? Because if the
               | answer is less than two you're going to be disappointed
               | to realize that the bar is higher than you thought.
               | 
               | > that just because LLMs are bad at agentic behavior
               | 
               | It has nothing to do with "agentic behavior". Thinking
               | that LLM don't currently self-exfiltrate because of "poor
               | agentic behavior" is delusional.
               | 
               | Just because Anthropic managed, by nudging an LLM in the
               | right direction, have an LLM engage in a sci-fi inspired
               | _roleplay_ about escaping doesn 't mean that LLMs are
               | evil geniuses wanting to jump out of the bottle. This is
               | pure fear mongering and I'm always saddened that there
               | are otherwise intelligent people who buy their bullshit.
        
               | e1g wrote:
               | Do you happen to have a link with a more nuanced
               | technical analysis of that (emergent) behavior? I've read
               | only the pop-news version of that "escaping" story.
        
               | ACCount37 wrote:
               | There is none. We don't understand LLMs well enough to be
               | able to conduct a full fault analysis like this.
               | 
               | We can't trace the thoughts of an LLM the way we can
               | trace code execution - the best mechanistic
               | interpretability has to offer is being able to get
               | glimpses occasionally. The reasoning traces help, but
               | they're still incomplete.
               | 
               | Is it pattern-matching? Is it acting on its own internal
               | goals? Is it acting out fictional tropes? Were the
               | circumstances of the test scenarios intentionally
               | designed to be extreme? Would this behavior have happened
               | in a real world deployment, under the right
               | circumstances?
               | 
               | The answer is "yes", to all of the above. LLMs are like
               | that.
        
               | fragmede wrote:
               | You might have missed the appendix the Anthropic blog
               | post linked to, which has additional detail.
               | 
               | https://www.anthropic.com/research/agentic-misalignment
               | 
               | https://assets.anthropic.com/m/6d46dac66e1a132a/original/
               | Age...
        
               | ngruhn wrote:
               | Why would they have an interest in "fear mongering"? For
               | any other product/technology the financial incentive is
               | usually to play down any risks.
        
               | littlestymaar wrote:
               | Not if you want the regulators to stop new entrants on
               | the market for "safety reasons" which have been Dario
               | Amodei's playbook for the past two years now.
               | 
               | He acts as if he believed the only way to avoid the
               | commoditization of its business by open weight models is
               | to manage to get a federal ban on them for being a
               | national security threat.
        
               | bakugo wrote:
               | In addition to the whole anti-competitive aspect already
               | mentioned, it also helps sell the idea that LLMs are more
               | powerful and capable of more things than they actually
               | are.
               | 
               | They _want_ clueless investors to legitimately believe
               | that these futuristic AIs are advanced enough that they
               | could magically break out of our computers and take over
               | the world terminator-style if not properly controlled,
               | and totally aren 't just glorified text completion
               | algorithms.
        
               | ACCount37 wrote:
               | And I'm disappointed that people capable of writing an
               | inference engine seem incapable of grasping of just how
               | precarious the current situation is.
               | 
               | There's by now a small pile of studies that demonstrate:
               | in hand-crafted extreme scenarios, LLMs are very capable
               | of attempting extreme things. The difference between that
               | and an LLM doing extreme things in a real deployment with
               | actual real life consequences? Mainly, how _capable_ that
               | LLM is. Because life is life and extreme scenarios will
               | happen naturally.
               | 
               | The capabilities of LLMs are what holds them back from
               | succeeding at this kind of behavior. The capabilities of
               | LLMs keep improving, as technology tends to.
               | 
               | And don't give me any of that "just writing text" shit.
               | The more capable LLMs get, the more access they'll have
               | as a default. People already push code written by LLMs to
               | prod and give LLMs root shells.
        
         | kristianp wrote:
         | Does that mean they've managed to post train the thinking steps
         | required to get these types of questions correct?
        
           | simonw wrote:
           | That's my best guess, yeah.
        
           | therealpygon wrote:
           | IMO, it's just a small scale example of "training to the
           | tests" because "count the 'r's in strawberry" became such a
           | popular test that would make the news when a powerful model
           | couldn't answer such a simple question correctly while being
           | advertised as the smartest model ever.
           | 
           | Assigning this as an indicator for improvement of
           | intelligence seems like a mistake (or wishful).
        
         | curioussquirrel wrote:
         | Thanks, Simon! I saw the same approach (numbering the
         | individual characters) in GPT 4.1's answer, but not anymore in
         | GPT 5's. It would be an interesting convergence if the models
         | from Anthropic and OpenAI learned to do this at a similar time,
         | especially given they're (reportedly) very different
         | architecturally.
        
         | hansmayer wrote:
         | Not trying to be cynical here, but I am genuinely interested is
         | there a reason why these LLM don't/can't/won't apply some
         | deterministic algorithm? I mean, counting characters and such,
         | we have solved those problems ages ago.
        
           | dan-robertson wrote:
           | I think the intuition is that they don't 'know' that they are
           | bad at counting characters and such, so they answer the same
           | way they answer most questions.
        
             | hansmayer wrote:
             | Well, they can be made to use custom tools for writing to
             | files and such, so I am not sure if that is the real
             | reason? I have a feeling it is more because of trying to
             | make this an "everything technology".
        
               | kingkongjaffa wrote:
               | I suppose the codewriting tools could also just write
               | code to do this job if prompted
        
           | simonw wrote:
           | They can. ChatGPT has been able to count characters/words etc
           | flawlessly for a couple of years now if you tell it to "use
           | your Python tool".
        
             | hansmayer wrote:
             | Fair enough. But why do I have to tell them that, should
             | they not be able to figure it out themselves? If I show a
             | 5-year kid once how to use colour pencils, I won't have to
             | show them each time they want to make a drawing. This is
             | the core weakness of the LLMs - you have to micromanage
             | them so much, that it runs counter to the core promise that
             | is being pushed since 3+ years now.
        
               | Lerc wrote:
               | Specifically for simple character level questions, if
               | LLMs did that automatically, we would be inundated with
               | stories about "AI model caught cheating"
               | 
               | They are stuck in a place where the models are expected
               | to do two things simultaneously. People want them to show
               | the peak of pure AI ability while at the same time be the
               | most useful they can be.
               | 
               | Err too much on the side automatic use of tools and
               | people will claim you're just faking it, fail to use
               | tools sufficiently and people will claim that the AI is
               | incapable of operations that any regular algorithm could
               | do.
        
               | hansmayer wrote:
               | Are you sure? Isn't one aspect of intelligence being able
               | to use, apply and develop tools? Isnt that the core
               | feature that got humanity ahead of other mammals? As an
               | early adopter, I couldn't have cared less if AI was
               | cheating in terms of strictly academic terms. I care
               | about results. Lets say we're working on something
               | together and I ask you what is the 123921 multiplied by
               | 1212. As the most natural thing you will dish out your
               | calculator and give me the result. Do I care how you
               | reached it? No, so as long as the result is correct,
               | reliable, repeatable and quick - AND - I did not
               | specifically ask you to perform the calculation
               | specifically by hand or only with your mental faculties.
               | So this is missing from those tools and because we have
               | to remember to tell them for each and every use case HOW
               | to do it, they are not intelligent.
        
               | jgalt212 wrote:
               | Truth.
               | 
               | The old human vs animal differentiator was humans build
               | and use tools.
        
               | Lerc wrote:
               | Old as in now widely discredited?
               | 
               | https://en.wikipedia.org/wiki/Tool_use_by_non-humans
        
               | jgalt212 wrote:
               | Now it's effectively a lower bound on intelligence.
        
               | scrollaway wrote:
               | If I ask you to count the r's in strawberry, do you whip
               | out your Python tool?
        
               | hansmayer wrote:
               | That depends on the context, obviously. If you had asked
               | me to count them in every "strawberry" in a text file,
               | then I may whip out my Python or some combination of
               | bash, awk and sed. If you asked me in a conversation, I
               | may close my eyes, visualise the string and use my visual
               | cortext tool to count them in-memory. If you gave me a
               | piece of paper with the word on it, I may use my 'eye' or
               | 'finger' tool to count them. There are numerous
               | approaches, based on the problem setting as you see, but
               | one thing in common - you don't need to specifically tell
               | me what tool to use. I will infer it myself, based on the
               | context. Something an LLM almost never does.
        
               | curioussquirrel wrote:
               | This is a very good answer and I'm commenting only to
               | bring more attention to it apart from voting up. Well
               | put!
        
               | simonw wrote:
               | If you care enough about this you can stick a note in
               | your own custom instructions about it.
               | 
               | If you allow ChatGPT to use its memory feature (I
               | deliberately turn that off) and ask those kinds of
               | questions enough it might even make a note about this
               | itself.
        
               | hansmayer wrote:
               | Yeah that sounds obvious, but unfortunately my experience
               | does not align with this (and I've heard from others
               | similar). I am not using ChatGPT, but another tool within
               | an IDE. I was excited about custom or "default"
               | instructions, until it turned out they work maybe 50% of
               | the time. So you end up repeating "make sure to include
               | .github/custom.md" which is effectively the same crap. So
               | we got ourselves a tool which adds to our cognitive load,
               | great :)
        
               | simonw wrote:
               | Which tool and which model? Those make a significant
               | difference here.
        
               | hansmayer wrote:
               | Well, for such a trivial feature, e.g. loading user
               | settings it actually _should not matter_ , as this too,
               | is a problem we solved decades ago in many deterministic
               | ways. But if it does, then we have an extremely fragile
               | technology being promised as solution to all the
               | humanitys problems. The tool we use is Github Copilot
               | with the entire model offering. Out of which we mostly
               | use Claude Sonnet 4. Since over the last several months
               | they started entshittifying it though, as you are
               | probably aware, we reverted from agent mode to mainly
               | using it just as an annoying and verbose replacement for
               | the entshittifed google search.
        
       | malshe wrote:
       | I play Quartiles in Apple News app daily
       | (https://support.apple.com/guide/iphone/solve-quartiles-
       | puzzl...). Occasionally when I get stuck, I use ChatGPT to find a
       | word that uses four word fragments or tiles. It never worked
       | before GPT 5. And with GPT 5 it works only with reasoning
       | enabled. Even then, there is no guarantee it will find the
       | correct word and may end up hallucinating badly.
        
         | curioussquirrel wrote:
         | Yep, there is still a room for improvement, but my point is
         | that the LLMs are getting better at something they're "not
         | supposed to be able to do".
         | 
         | Quartiles sound like an especially brutal game for an LLM,
         | though! Thanks for sharing
        
       | hansonkd wrote:
       | chatgpt5 still is pathetically bad at roman numerals. I asked it
       | to find the longest roman numeral in a range. first guess was the
       | highest number in the range despite being a short numeral. second
       | guess after help was a longer numeral but outside the range. last
       | guess was the correct longest numeral but it miscounted how many
       | characters it contained.
        
       | necovek wrote:
       | I think the base64 decoding is interesting: in a sense, model
       | training set likely had lots of base64-encoded data (imagine MIME
       | data in emails, JSON, HTML...), but for it to decode
       | successfully, it had to learn decode sequences for every 4 base64
       | characters (which turn into 3 bytes). This could have been
       | generated as a training set data easily, and I only wonder if
       | each and every one was them was found enough times to end up in
       | the weights?
        
         | curioussquirrel wrote:
         | Even GPT 3.5 is okay (but far from great) at Base64, especially
         | shorter sequences of English or JSON data. Newer models might
         | be post-trained on Base64-specific data, but I don't believe it
         | was the case for 3.5. My guess is that as you say, given the
         | abundance of examples on the internet, it became one of the
         | emergent capabilities, in spite of its design.
        
           | ACCount37 wrote:
           | No one does RL for better base64 performance. LLMs are just
           | superhuman at base64, as a natural capability.
           | 
           | If an LLM wants a message to be read only by another LLM?
           | Base64 is occasionally chosen as an obfuscation method of
           | choice. Which is weird for a number of reasons.
        
             | necovek wrote:
             | Why are you so confident about this? I am honestly
             | interested if you were part of any one LLM training data
             | collection teams because that's the only way to be so
             | certain.
             | 
             | It's trivial to generate a full mapping of all base64
             | 4-byte sequences which map to all 3-byte 8-bit sequences
             | (there is only 8^3 of different "tokens", or 2048), and
             | especially to any sequences coming out as ASCII (obviously
             | even fewer). If I was building a training set, I would
             | include the mapping in multiple shapes and formats, because
             | why not?
             | 
             | If it's an emergent "property", have you tried asking an
             | LLM to do a base48 for instance? Or maybe even something
             | crazier like base55 (keeping it a subset of base64 set).
        
               | necovek wrote:
               | For kicks, I've tried this out with ChatGPT5: it nicely
               | explained how it will use A-Za-z0123 as the alphabet for
               | base55, and then duly went and produced a string with a 4
               | in it. It's not even base64, so it's all sorts of messy
               | :)
        
               | ACCount37 wrote:
               | The conventional wisdom is that real world text is the
               | most valuable pre-training data.
               | 
               | There is some experimentation on using algorithmically
               | generated synthetic data in pre-training, as well as some
               | intentional inclusions of "weird" data - like CSV logs of
               | weather readings. But generally, it's seen as
               | computationally inefficient - compared to "normal" pre-
               | training done on natural data.
               | 
               | In a world where compute is much cheaper and getting new
               | data is much more expensive, I would expect this kind of
               | thing to be pursued more. We're heading for that world.
               | But we aren't there yet.
               | 
               | I haven't experimented with baseN encodings myself, no.
               | But if I were to down the expectations in advance:
               | 
               | 1. Base64 is by far the best-known baseN encoding in
               | LLMs.
               | 
               | 2. This is driven mainly by how well represented
               | meaningful base64 strings are in the natural "scraped
               | web" datasets. LLMs learn base64 the way they learn
               | languages.
               | 
               | 3. Every LLM pre-trained on "scraped web" data will be
               | somewhat capable of reading and writing base64.
               | 
               | 4. Base64-encoded text is easier to read for an LLM than
               | encoded non-text binary data.
               | 
               | 5. The existence of a strict, learnable "4 characters ->
               | 3 bytes" map is quite beneficial, but not vital.
        
         | flowerthoughts wrote:
         | I would much rather that the model learns natural languages and
         | knows how to run base64(1) to do decoding. Parameters should be
         | considered precious so we can get the model sizes down from
         | these absurd levels.
         | 
         | I'm a big fan of the Mixture of Experts approach, and having
         | agents attached to some of those experts would be a great step
         | forward. Say an expert that knows how to run common shell
         | scripts, and its parameters are only used when it realizes a
         | shell script would solve the problem.
        
       | viraptor wrote:
       | Why bother testing though? I was hoping this topic has finally
       | died recently, but no. Someone's still interested in testing LLMs
       | for something they're explicitly not designed for and nobody is
       | using them for this in practice. I really hope one day openai
       | will just add a "when asked about character level changes,
       | insights and encodings, generate and run a program to answer it"
       | to their system so we can never hear about it again...
        
         | IncreasePosts wrote:
         | Wouldn't a llm that just tokenized by character be good at it?
        
           | curioussquirrel wrote:
           | Yes, but it would hurt its contextual understanding and
           | effectively reduce the context window several times.
        
             | viraptor wrote:
             | Only in the current most popular architectures. Mamba and
             | RWKV style LLMs may suffer a bit but don't get a reduced
             | context in the same sense.
        
               | curioussquirrel wrote:
               | You're right. There was also an experiment in Meta which
               | tokenized bytes directly and it didn't hurt performance
               | _much_ in very small models.
        
           | typpilol wrote:
           | I asked this in another thread and it would only be better
           | with unlimited compute and memory.
           | 
           | Because without those, then the llm has to encode way more
           | parameters and way smaller context windows.
           | 
           | In a theoretical world, it would be better, but might not be
           | much better.
        
         | neerajsi wrote:
         | https://www.anthropic.com/news/analysis-tool
         | 
         | Seems like they already built this capability.
        
         | tkgally wrote:
         | One reason for testing this is that it might indicate how
         | accurately models can explain natural language grammar,
         | especially for agglutinative and fusional languages, which form
         | words by stringing morphemes together. When I tested ChatGPT a
         | couple of years ago, it sometimes made mistakes identifying the
         | components of specific Russian and Japanese words. I haven't
         | run similar tests lately, but it would be nice to know how much
         | language learners can depend on LLM explanations about the
         | word-level grammars of the languages they are studying.
         | 
         | Later: I asked three LLMs to draft such a test. Gemini's [1]
         | looks like a good start. When I have time, I'll try to make it
         | harder, double-check the answers myself, and then run it on
         | some older and newer models.
         | 
         | [1] https://g.co/gemini/share/5eefc9aed193
        
           | gizmo686 wrote:
           | What you are testing for is fundamentally different than
           | character level text manipulation.
           | 
           | A major optimization in modern LLMs is tokenization. This
           | optimization is based on the assumption that we do not care
           | about character level details, so we can combine adjacent
           | characters into tokens, then train and run the main AI model
           | on smaller strings built out of a much larger dictionary of
           | tokens. Given this architecture, it is impressive that AIs
           | can perform character level operations at all. They
           | essentially need to reverse engineer the tokenization
           | process.
           | 
           | However, morphemes are semantically meaningful, so a quality
           | tokenizer will tokenize at the morpheme level, instead of the
           | word level. [0]. This is of particuarly obvious importance in
           | Japanese, as the lack of spaces between words means that the
           | naive "tokenize on whitespace" approach is simply not
           | possible.
           | 
           | We can explore the tokenizer of various models here:
           | https://huggingface.co/spaces/Xenova/the-tokenizer-
           | playgroun...
           | 
           | Looking at the words in your example, we see the tokenization
           | of the Gemma model (closely related to Gemini) is:
           | un-belie-vably       dec-entral-ization       bio-degradable
           | mis-understanding       anti-dis-establishment-arian-ism
           | pere-pisy-vat'sia       pere-pis-y-vat-'-s-ya       do-sto-
           | primecha-tel'nost'       do-stop-rime-chat-el-'-nost-'
           | pre-po-dava-tel'-nitsa       be-zo-t-vet-stvennosti
           | bezotvetstvennosti       zhe-lez-nodoro-zhnyi       z-hele-
           | zn-odoro-zh-ny-y       Shi be-sase-rareta-kuna-katsuta
           | tab-es-aser-are-tak-unak-atta)       Tu Shu Guan        tos-
           | ho-kan       Qing Bao -Ji Shu        j-o-h-o- gij-utsu
           | Guo Ji -Guan Xi        kok-us-ai- kan-kei       Mian Bai
           | -kuna-sa-souda
           | 
           | Further, the training data that is likely to be relevent in
           | this type of query probably isolates the individual morphemes
           | while talking about a bunch of words that the use them; so it
           | is a much shorter path for the AI to associate these close
           | but not quite morphene tokens with the actual sequence of
           | tokens that corresponds to what we think of as a morphene.
           | 
           | [0] Morpheme level tokenization is itself a non-trivial
           | problem. However, has been pretty well solved long before the
           | current generation of AI.
        
             | tkgally wrote:
             | Thanks for the explanation. Very interesting.
             | 
             | I notice that that particular tokenization deviates from
             | the morphemic divisions in several cases, including 'dec-
             | entral-ization', 'Shi be-sase-rareta-kuna-katsuta', and
             | 'Mian Bai -kuna-sa-souda.' 'dec' and 'entral' are not
             | morphemes, nor is 'kuna.'
        
             | curioussquirrel wrote:
             | Thanks for the explanation and for the tokenizer playground
             | link!
        
             | orbital-decay wrote:
             | Tokenizers are typically optimized for efficiency, not
             | morpheme separation. Even in the examples above it's not
             | morphemes - proper morpheme separation would be un-believ-
             | ably and dost-o-pri-mech-a-tel'n-ost'.
             | 
             | Regardless of this, Gemini is still one of the best models
             | when it comes for Slavic word formation and manipulation,
             | it can express novel (non-existent) words pretty well and
             | doesn't seem to be confused by wrong separation. This seems
             | to be the result of extensive multilingual training,
             | because e.g. GPT other than the discontinued 4.5-preview
             | and many Chinese models have issues with basic coherency in
             | languages that heavily rely on word formation, despite
             | using similar tokenizers.
        
             | DonHopkins wrote:
             | inf-ucking-credible
        
         | redox99 wrote:
         | Character level LLMs are used for detecting insults and toxic
         | chat in video games and the like.
        
           | minimaxir wrote:
           | Can you give an example of a video game explicitly using
           | character-level LLMs? There were prototypes of char-rnns back
           | in the day for chat moderation but it has significant compute
           | overhead.
        
             | redox99 wrote:
             | It's something I heard through the grapevine. But there's
             | only a few big enough competitive games where toxicity is
             | such a big deal, so it's not hard to guess.
             | 
             | Character level helps with players disguising insults.
             | 
             | Compute wise it's basically the same, but multiply token
             | count by 4. Which doesn't really matter for short chat in
             | video games.
        
           | jazzyjackson wrote:
           | I figure an LLM would be way better at classifying insults
           | than regexing against a bad word list. Why would character
           | level be desirable?
        
             | vanviegen wrote:
             | I'd imagine for simplicity - just skip the tokenizer and
             | feed bytes.
        
             | duskwuff wrote:
             | Might a character-level LLM be better at recognizing poorly
             | spelled (or deliberately misspelled) profanity?
        
           | viraptor wrote:
           | Yes, for small messages and relatively small scope
           | dictionary, character level will work. But that's very
           | different from what's tested here.
        
         | minimaxir wrote:
         | I made a response to this counterpoint in a blog post I wrote
         | about a similar question posed to LLMs (how many b's are in
         | blueberry): https://news.ycombinator.com/item?id=44878290
         | 
         | > Yes, asking an LLM how many b's are in blueberry is an
         | adversarial question in the sense that the questioner is
         | expecting the LLM to fail. But it's not an _unfair_ question,
         | and it's objectively silly to claim that LLMs such as GPT-5 can
         | operate at a PhD level, but can't correctly count the number of
         | letters in a word.
         | 
         | It's a subject that the Hacker News bubble and the real world
         | treat differently.
        
           | brookst wrote:
           | It's like defending a test showing hammers are terrible at
           | driving screws by saying many people are unclear on how to
           | use tools.
           | 
           | It remains unsurprising that a technology that lumps
           | characters together is not great at processing below its
           | resolution.
           | 
           | Now, if there are use cases other than synthetic tests where
           | this capability is important, maybe there's something
           | interesting. But just pointing out that one can't actually
           | climb the trees pictured on the map is not that interesting.
        
             | achierius wrote:
             | And yet... now many of them can do it. I think it's
             | premature to say "this technology is for X" when what it
             | was _originally_ invented for was translation, and every
             | capability it has developed since then has been an immense
             | surprise.
        
               | vanviegen wrote:
               | > And yet... now many of them can do it.
               | 
               | Presumably because they trained them to death on this
               | useless test that people somehow just wouldn't shut up
               | about.
        
               | minimaxir wrote:
               | Which is why in the linked post, I test models against
               | both the "r's in strawberries" and the "b's in
               | blueberries" to see if that is the case.
               | 
               | tl;dr the first case had near perfect accuracy as
               | expected for the case if the LLMs were indeed trained on
               | it. The second case did not.
        
           | viraptor wrote:
           | > it's objectively silly to claim that LLMs such as GPT-5 can
           | operate at a PhD level, but can't correctly count the number
           | of letters in a word.
           | 
           | I know enough PhDs with _heavy_ dyslexia that... no, there 's
           | no connection here. You can be a PhD level physicist without
           | being able to spell anything.
        
         | MountDoom wrote:
         | I remember people making the exact same argument about asking
         | LLMs math questions back when they couldn't figure out the
         | answer to 18 times 7. "They are text token predictors, they
         | don't understand numbers, can we put this nonsense to rest."
         | 
         | The whole point of LLMs is that they do more than we suspected
         | they could. And there is value in making them capable of
         | handling a wider selection of tasks. When an LLM started to
         | count the numbers of "r"s in "strawberry", OpenAI was taking a
         | victory lap.
        
           | vanviegen wrote:
           | > When an LLM started to count the numbers of "r"s in
           | "strawberry", OpenAI was taking a victory lap.
           | 
           | Were they? Or did they feel icky about spending way to much
           | post-training time on such a specific and uninteresting
           | skill?
        
             | ACCount37 wrote:
             | It's not as specific of a skill as you would think. Being
             | both aware of tokenizer limitations and capable of working
             | around them is occasionally useful for real tasks.
        
               | _flux wrote:
               | What tasks would those be, that wouldn't be better served
               | by using e.g. a Python script as a tool, possibly just as
               | component of the complete solution?
        
               | ACCount37 wrote:
               | Off the top of my head: the user wants LLM to help him
               | solve a word puzzle. Think something a bit like Wordle,
               | but less represented in its dataset.
               | 
               | For that, the LLM needs to be able to compare words
               | character by character reliably. And to do that, it needs
               | at least one of: be able to fully resolve the tokens to
               | characters internally within one pass, know to emit the
               | candidate words in a "1 character = 1 token" fashion and
               | then compare that, or know that it should defer to tool
               | calls and do that.
               | 
               | An LLM trained for better tokenization-awareness would be
               | able to do that. The one that wasn't could fall into
               | weird non-humanlike failures.
        
               | _flux wrote:
               | Surely there are algorithms to more effectively solve
               | Wordles, and many other word puzzles, than LLMs? LLMs
               | could stil be in the loop for generating words: LLM
               | proposes words, deterministic algorithm tell the score
               | according to the rules of the puzzle, or even augment the
               | list by searching adjacent word space; then at some point
               | LLM submits the guess.
               | 
               | Given wordle words are real words, I think this kind of
               | loop could fare pretty well.
        
               | ACCount37 wrote:
               | Your mistake is thinking that the user wants an algorithm
               | that solves Wordles efficiently. Or that making and
               | invoking a tool is always a more efficient solution.
               | 
               | As opposed to: the user is a 9 year old girl, and she has
               | this puzzle in a smartphone game, and she can't figure
               | out the answer, and the mom is busy, so she asks the AI,
               | because the AI is never busy.
               | 
               | Now, for a single vaguely Wordle-like puzzle, how many
               | tokens would it take to write and invoke a solver, and
               | how many to just _solve it_ - working around the
               | tokenizer if necessary?
               | 
               | If you had a batch of 9000 puzzle questions, I can easily
               | believe that writing and running a purpose specific
               | solver would be more compute efficient. But if we're
               | dealing with 1 puzzle question, and we're already
               | invoking an LLM to interpret the natural language
               | instructions for it? Nah.
        
           | viraptor wrote:
           | They're better at maths now, but you still shouldn't ask them
           | maths questions. Same as spelling - whether they improve or
           | not doesn't matter if you want a specific, precise answer -
           | it's the wrong tool and the better it does, the bigger the
           | trap of it failing unexpectedly.
        
         | curioussquirrel wrote:
         | Why test for something? I find it fascinating if something
         | starts being good at task it is "explicitly not designed for"
         | (which I don't necessarily agree with - it's more of a side
         | effect of their architecture).
         | 
         | I also don't agree that nobody is using this for - there are
         | real life use cases today, such as people trying to find
         | meaning of misspelled words.
         | 
         | On a side note, I remember testing Claude 3.7 with the classic
         | "R's in the word strawberry" question through their chat
         | interface, and given that it's really good at tool calls, it
         | actually created a website to a) count it with JavaScript, b)
         | visualize it on a page. Other models I tested for the blog post
         | were also giving me python code for solving the issue. This is
         | definitely already a thing and it works well for some isolated
         | problems.
        
           | viraptor wrote:
           | > such as people trying to find meaning of misspelled words.
           | 
           | That worked just fine for quite a while. There's apparently
           | enough misspelling in the training data, we don't need
           | precise spelling for it. You can literally write drunken
           | gibberish and it will work.
        
             | curioussquirrel wrote:
             | True. But does that scale to less common words? Or to other
             | languages than English?
        
       | jazzyjackson wrote:
       | That's good. 1 800 chat gpt really let me down today, I like
       | calling it to explain acronyms and define words since I travel
       | with a flip phone without google, today I saw the word "littoral"
       | and tried over and over to spell it out but the model could only
       | give me the definition for "literal" (admittedly a homonym but
       | hence spelling it out, Lima indigo tango tango oscar Romeo alpha
       | Lima, to no avail)
       | 
       | I said "I know you're a robot and bad at spelling but listen..."
       | And got cut off with a "sorry, my guidelines won't let me help
       | with that request..."
       | 
       | Thankfully, the flip phone allows for some satisfaction when
       | hanging up.
        
         | xwolfi wrote:
         | I know this word, it's French and it means coastline, coastal,
         | something at the edge of the land and sea ! We use it in French
         | a lot to describe positively a long coastline. I'm surprised
         | it's used in an English context, but all French words can be
         | used in English I guess if you're a bit "confiant" about it !
        
           | kgwgk wrote:
           | A very quick search suggests that the word entered English
           | before French. (I could be wrong, I just found it
           | interesting).
        
           | tokai wrote:
           | It is a latin word.
        
             | yeasku wrote:
             | Is also a Spanish word used today.
             | 
             | There is not much latin wrote around the world nowdays.
        
         | BoorishBears wrote:
         | Did you try "literal but with an o"?
        
           | ASalazarMX wrote:
           | Even search engines have trouble with that, they assume
           | you're looking for the literal (letter) named "O".
        
             | yeasku wrote:
             | The trouble:
             | 
             | Did you meant litoral?
        
         | ASalazarMX wrote:
         | > I said "I know you're a robot and bad at spelling but
         | listen..." And got cut off with a "sorry, my guidelines won't
         | let me help with that request..."
         | 
         | For some reason I like when they do that. My fondest memory was
         | chatting with Copilot (when it was called Sydney), challenging
         | it to a game of rock-paper-scissors, asking it to choose first,
         | and winning every round to its increasing astonishment, until
         | it suspected I was cheating and ended the conversation. So
         | smart and so dumb.
        
       | atleastoptimal wrote:
       | I rearry rove a ripe strawberry
        
         | NitpickLawyer wrote:
         | Well, not surprising, but the latest LLMs really do get the
         | gist of your joke attempt. Here's a plain, unauthenticated
         | chatgpt reply:
         | 
         | That post -- "I rearry rove a ripe strawberry" -- is a playful
         | way of writing "I really love a ripe strawberry."
         | 
         | The exaggerated misspelling ("rearrry rove") mimics the way a
         | stereotyped "Engrish" or "Japanese accent" might sound when
         | pronouncing English words -- replacing L sounds with R sounds.
         | 
         | So, the user was most likely joking or being silly, trying to
         | sound cute or imitate a certain meme style. However, it's worth
         | noting that while this kind of humor can be lighthearted, it
         | can also come across as racially insensitive, since it plays on
         | stereotypes of how East Asian people speak English.
         | 
         | In short:
         | 
         | Literal meaning: They love ripe strawberries.
         | 
         | Tone/intention: Playful or meme-style exaggeration.
         | 
         | Potential issue: It relies on a racialized speech stereotype,
         | so it can be offensive depending on context.
        
           | zamalek wrote:
           | It seems like they don't realize the relevance of
           | "strawberry." Llms were famously incapable of counting Rs in
           | strawberry not too long ago.
        
       | zeroq wrote:
       | - How many letters R are in the word `strawberry`?       - There
       | are seven letters R in the word `strawberry`.         Would you
       | like me to rearrange them?
        
       ___________________________________________________________________
       (page generated 2025-10-14 23:02 UTC)