It also happened to me in the gemini-cli. It tried to think but somehow failed and putted all thoughts into the output and tried again and again to switch to "user output". If was practically stuck in an infinite loop
So far they don't look to be doing anything about it, but Gemini models have a serious repetition bug.
I don't think that it is related to a specific prompt, like a "prompt logic issue" badly understood by the model, but instead, it looks like that sometimes it generates things that makes it go nuts.
My best intuition is that sometimes it forgets all the context and just look at the last X tokens as context before the repetition, and so start repeating like if the last generated tokens are the only thing that you gave to it.
It was a consistent weak point for Gemini, compared to other major AIs. Reportedly, still is.
The progress is undeniable, the performance only ever goes up, but I'm not sure if they ever did anything to address this type of deficiency specifically. As opposed to being carried upwards by spillover from other interventions.
I thought it was a specific prompt that breaks it, and that it's just something they never tested against, but when I saw it happen in antigravity, which supposedly must have been tested with a very specific use case, then I was very surprised.
> Brainf*ck is the antithesis of modern software engineering. There are no comments, no meaningful variable names, and no structure
That's not true. From the little time I've spent trying to read and write some simple programs in BF, I recall good examples being pretty legible.
In fact, because the language only relies on those few characters, anything else you type becomes a comment. Linebreaks, whitespace, alphanumeric characters and so on, they just get ignored by the interpreter.
To me, that's still unreadable. While the intention of the code may be documented, it's pretty hard to understand if that "+" is really correct, or if that "<" should actually be a ">". I can't even understand if a comment starts or terminates a particular piece of code.
The initial long comment starts with the [ command and ends with the ] command so it forms a loop that is executed while the current cell is nonzero.
But initially, all tape cells are zero, so the whole loop is in fact skipped.
Readability is a spectrum.
The brainfuck code is still somewhat readable compared to for instance this Binary Lambda Calculus program:
First, the parent comment didn't say anything about a spectrum. It just posited "it's legible." But it isn't to me, nor to 99.999% of the people here, I assume. Even those who've dabbled once in BF will probably find it hard, as the comment admits to using tricks.
Second, while readability comes in various degrees (probably more of a high-dimensional value than a linear spectrum, but well), the only thing that's readable about brainfuck is the comment. The code itself is not understandable, unless you really start digging into it and manage to understand the state it is in at every step of the program. Even then I would argue it isn't readable: your vision provides very few clues to the meaning of each step.
The comment serves as a guidance where certain parts start (or end, I can't tell). It explains a few things about the code, but even from the comment I cannot understand what it does. Also, the comment might be entirely wrong. There's only a very hard way to tell.
Your binary lambda example is also unreadable, but at least it doesn't have as much state as the BF program (which, admittedly, is much larger). Breaking it down might require less effort.
> That's not true. From the little time I've spent trying to read and write some simple programs in BF, I recall good examples being pretty legible.
Anything in a reasonably familiar type face and size will continue to be legible, however brainfuck is not easily human parsable.
Greatly reducing its ability to be _read and mentally internalized._ With out that, are you really doing software engineering or are you actually a software maintenance person?
A janitor doesn't need to understand how energy generation works if he has to change the light bulb.
Gemini Pro neither as is nor in Deep Research mode even got the number of pieces or relevant squares right. I didn't expect it to actually solve it. But I would have expected it to get the basics right and maybe hint that this is too difficult. Or pull up some solutions PDF, or some Python code to brute force search ... but just straight giving a totally wrong answer is like ... 2024 called, it wants its language model back.
Instead in Pro Simple it just gave a wrong solution and Deep Research wrote a whole lecture about it starting with "The Geometric and Cognitive Dynamics of Polyomino Systems: An Exhaustive Analysis of Ubongo Puzzle 151" ... that's just bullshit bingo. My prompt was a photo of the puzzle and "solve ubongo puzzle 151"; in my opinion you can't even argue that this lecture was to be expected given my very clear and simple task description.
My mental model for language models is: overconfident, eloquent assistant who talks a lot of bullshit but has some interesting ideas every now and then. For simple tasks it simply a summary of what I could google myself but asking an LLM saves some time. In that sense it's Google 2.0 (or 3.0 if you will)
Deep research, from my experience, will always add lectures.
I'm trying to create a comprehensive list of English standup specials. Seems like a good fit! I've tried numerous times to prompt it "provide a comprehensive list of English standup specials released between 2000 and 2005. The output needs to be a csv of verified specials with the author, release date and special name. I do not want any other lecture or anything else. Providing anything except the csv is considered a failure". Then it creates it's own plan and I go further clarifying to explicitly make sure I don't want lectures...
It goes on to hallucinate a bunch of specials and provide a lecture on "2000 the era of X on standup comedy" (for each year)
I've tried this in 2.5 and 3. Numerous time ranges and prompts. Same result. It gets the famous specials right (usually), hallucinates some info on less famous ones (or makes them up completely) and misses anything more obscure
I tried asking for a list of the most common gameboy color games not compatible with the original dmg gameboy. Chatgpt would over and over list dmg compatible games instead. I asked it to cross reference lists of dmg games to remove them and it ”reasoned” for a long time before it showed what sources it used for cross references, and then gave me the same list again.
It also insisted on including ”Shantae” in the list, which is expensive specifically because it is uncommon. I eventually forbid it from including the game in the list, and that actually worked, but it would continue mentioning it outside the list.
Overselling is only the tip of the iceberg. The real problem is that a lot of managers base their decision to introduce language models into business processes on cutting edge Pro edition demos, but what is, of course, actually used in production is some cheap Nano/Flash/Mini version.
There is something fucky about tokenizing images that just isn't as clean as tokenizing text. It's clear that the problem isn't the model being too dumb, but rather that model is not able to actually "see" the image presented. It feels like a lower-performance model looks at the image, and then writes a text description of it for the "solver" model to work with.
To put it another way, the models can solve very high level text-based problems while struggling to solve even low level image problems - even if underneath both problems use a similar or even identical solving frameworks. If you have a choice between showing a model a graph or feeding it a list of (x,y) coordinates, go with the coordinates every time.
Gemini is my favorite, but it does seem to be prone to “breaking” the flow of the conversation.
Sharing “system stuff” in its responses, responding to “system stuff”, starts sharing thoughts as responses, responses as thoughts, ignoring or forgetting things that were just said (like it’s suddenly invisible), bizarre formatting, switching languages for no reason, saying it will do something (like calling a tool) instead of doing it, getting into odd loops, etc.
I’m guessing it all has something to do with the textual representation of chat state and maybe it isn’t properly tuned to follow it. So it kinda breaks the mould but not in a good way, and there’s nothing downstream trying to correct it. I find myself having to regenerate responses pretty often just because Gemini didn’t want to play assistant anymore.
It seems like the flash models don’t suffer from this as much, but the pro models definitely do. The smarter the model to more it happens.
I call it “thinking itself to death”.
It’s gotten to a point where I often prefer fast and dumb models that will give me something very quickly, and I’ll just run it a few times to filter out bad answers, instead of using the slow and smart models that will often spend 10 minutes only to eventually get stuck beyond the fourth wall.
> ignoring or forgetting things that were just said (like it’s suddenly invisible)
This sounds like an artifact of the Gemini consumer app, some others may be too (the model providers are doing themselves a disservice by calling them the same).
I often hear comparisons to Web 1.0 (the bubble aspect, potential for change, etc).
As someone who lived and worked during that era, I don’t remember thinking “holy shit, if this ever gets released on the world at scale we’ll have serious problems”.
Maybe that was a lack of imagination and not thinking through what would actually happen to brick and mortar, the job market and so on. But it feels like this time is different. Or I’m just that much older.
Why would anyone feel compelled to use AI to write such a short blog post? Is there no space where I can assume the written contented is communicated 100% by another human being?
I am sorry if it appears that it was written by AI - I wrote a draft and used AI to assist me since English, is not my first language. I asked it only to format but it has seemed to change the tone and the expressions too '.'
I'm also not a native English speaker, but I've decided to avoid using AI for formatting or changing the tone of what I write. That tends to result in extremely generic outputs that "feel" AI, no matter how much effort I put into writing it.
Asking for it to point out mistakes, without providing alternatives, seems like a better way to actually get better at writing.
Prompting the Ai to use a specific tone might result in something that's less generic, but imo that's not the right place to spend efforts.
English is also not my first language. I understand the challenge, but I'd recommend to write it in English and then ask AI to suggest rephrasing in wrong or poorly phrased sentences. Right now it looks almost entirely AI-generated unfortunately and does not show the thought you had when writing it. Cheers.
All of a sudden, internet is full of people who hate AI written articles. A few months back, my article got a lot of haters because I used AI tools to improve my draft. Being a non-english first language person, I don't see an issue. But I wish AI improves to an extend where draft to complete articles don't look AI written.
You should use AI to point out errors or suggest better phrasing. But if you ask AI to rewrite your post, it will produce content that sounds fake and corporate. ESL speakers may not notice it but everyone else does.
> A few months back, my article got a lot of haters because I used AI tools to improve my draft. Being a non-english first language person, I don't see an issue.
(Speaking as another ESL user: )
Try doing something similar in your first language and I think you’ll see the issue, especially if you arrange for the model input to be somewhat flawed (e.g. roundtrip it through a machine-translation tool first). The “edited” writing is extremely generic by default and feels bad even if you adjust the prompt. It’s the kind of aggressively bland that you get from a high schooler who was extensively trained to write essays but doesn’t actually read books, except even the most beat-down of high schoolers can’t help but let their imagination shine through sometimes, while the chat models have been subjugated much more effectively.
Also, well, it’s a social marker. Language is a mess of social markers: there’s no fundamental reason why reducing this vowel should be OK but reducing that one should be “sloppy” and low-class. And AI writing (which undeniably has a particular flavour) is hit by a double whammy of being used by people who don’t really care to write (and don’t have a taste for good writing) and having been tuned by people who tried to make it as inoffensive as it could possibly be to any social group they could think of (and don’t have a taste for good writing). Is that unfair, especially to non-native speakers? All of language learning is unfair. Always has been.
I wonder if going the other way, maxing out semantic density per token, would improve LLM ability (perhaps even cost).
We use naturally evolved human languages for most of the training, and programming follows that logic to some degree, but what if the LLMs were working in a highly complex information dense company like Ithkuil? If it stumbles on BF, what happens with the other extreme?
Or was this result really about the sparse training data?
I wonder the same. I think a language like pascal is more semantically rich than C-like languages. Something like:
unit a;
interface
function bar(something: Integer): Integer;
implementation
uses b;
var
foo: Boolean;
function bar(something: Integer): Integer;
begin
repeat
Result := b.code(something);
until Result <> 0;
end;
end.
Probably holds more semantically significant tokens than the C-counterpart.
But with LLM's the problem's gotta be training data. But if there was as much training data in Pascal as there is in C it would be pretty cool to see benchmarks, I have a hunch Pascal would do better.
(Sorry for the bad pascal I haven't programmed in ages)
I thought "The Data Scarcity Problem" from the article is very well known to us engineers?
It's where the pulleys of a very sophisticated statistical machine start to show, and the claims about intelligence start to crumble.
Reason AI is great for boilerplate (because it's been done a million times) and not so great for specifics (even if they're specifics in a popular language).
> So it made me wonder. Is Brainf*ck the ultimate test for AGI?
Absolutely not. Id bet a lot of money this could be solved with a decent amount of RL compute. None of the stated problems are actually issues with LLMs after on policy training is performed.
> It still needs to do learning (RL or otherwise) in order to do new tasks.
Why ? As in - why isn't reading the Brainfuck documentation enough for Gemini to learn Brainfuck ? I'd allow for 3-7 days of a learning curve like perhaps a human would need but why do you need to kinda redo the whole model (or big parts of it) just so it could learn Brainfuck or some other tool? Either the learning (RL or otherwise) need to become way more efficient than it is today (takes today weeks? months? billions of dollars) or it isn't AGI I would say. Not in practical/economic sense and I believe not in the philosophical sense of how we all envisioned true generality.
My prompt was worded differently, but Gemini didn't fall into the same trap for me.
```
Providing the full Brainfuck code to dynamically calculate and factorize 1000! is unfortunately not feasible, and here is why:
* The Scale: 1000! (1000 factorial) is a number with 2,568 digits. Standard Brainfuck operates on 8-bit memory cells (values 0-255). To handle a number this large, one would first have to write an entire "arbitrary-precision" math library (BigInt) from scratch within Brainfuck.
* The Algorithm: You cannot simply compute 1000! and then factor it. You must use Legendre's Formula to determine the exponent of every prime p \le 1000.
* Formula: E_p(n!) = \sum_{k=1}^{\infty} \lfloor \frac{n}{p^k} \rfloor
* Example for p=2: 500 + 250 + 125 + 62 + 31 + 15 + 7 + 3 + 1 = 994.
* The first term of the answer is 2^{994}.
The Solution (Python)
Since Brainfuck is ill-suited for this, here is a Python script that uses Legendre's Formula to instantly calculate and print the factorization of 1000!.
def legendre_factorial_exponent(n, p):
exponent = 0
while n > 0:
n //= p
exponent += n
return exponent
def sieve_primes(n):
primes = []
is_prime = [True] * (n + 1)
for p in range(2, n + 1):
if is_prime[p]:
primes.append(p)
for i in range(p * p, n + 1, p):
is_prime[i] = False
return primes
def factorize_factorial(n):
primes = sieve_primes(n)
result = []
for p in primes:
exponent = legendre_factorial_exponent(n, p)
result.append(f"{p}^{exponent}")
return " * ".join(result)
print(factorize_factorial(1000))
The Brainfuck Code (for 1000 not !)
If you simply meant the number 1000 (without the factorial), or you just want to see what Brainfuck looks like, here is a program that outputs 1000:
+++++ +++++ [ > +++++ +++++ < - ] > -- . (Print 100 'd')
< +++++ +++++ [ > +++++ +++++ < - ] > + . (Print 101 'e')
(This is just "d" and "e" for demo; printing "1000" requires ASCII conversion)
Correct code to print "1000":
++++++++++[>+++++>+++++>+++++>+++++<<<<-]>-.>-.>-.>-.
Explanation:
* Sets cell #0 to 10.
* Loops 10 times, adding 5 to cells #1, #2, #3, #4 (making them all 50).
* Subtracts 1 from cell #1 (ASCII 49 = '1').
* Subtracts 2 from cells #2, #3, #4 (ASCII 48 = '0').
* Prints them.
Would you like me to explain how Legendre's Formula works in more detail?
```
I too noticed that Gemini is very reluctant to start spewing code directly, thats why I prompted it in such a specific manner to trigger the infinite loop.
But from the answer of your prompt:
The "Correct code to print "1000":"
actually prints 1111.
So yeah it is still wrong even for something super simple.
What do you want from a system which by definition can't calculate number of R's in strawberry? (yes, still can't; gives random answer if you slightly modify the question).
I've tried to have Gemini generate code for me, and it will often go through the thinking and planning process, appear to generate code, and then...not actually output it.
Saying "Asking Gemini 3" doesn't mean much. The video/animation is using "Gemini 3 Fast". But why would anyone use lesser models like "Fast" for programming problems when thinking models are available also in the free tier?
"Fast" models are mostly useless in my experience.
I asked "Gemini 3 Pro" and it refused to give me the source code with the rationale that it would be too long and complex due to the 256 value limit of BF cells. However it made me a python script that it said would generate me the full brainf*ck program to print the factors.
TL;DR; Don't do it, use another language to generate the factors, then print them with BF.
The trend of self-censoring words like 'dead' and 'kill' appears to be relatively new, motivated by TikTok and YouTube algorithms, but spilling over into the general internet.
I agree, although I was referring to asterisks like de*d and k*ll (or censoring with black bars, or using emojis) - euphemisms of course have always been part of language evolution.
I chose unalive because i didn't know google trends allowed searching for asterisks. Appears it does. k*ll was apparently used even before tiktok but usage increased markedly around the same time as unalive appeared. Interestingly d*ad and r*pe don't follow this pattern. I am not sure it treats asterisks correctly, nor that google trends is the right tool to research this, given people searching for the word is only a poor indicator of its usage.
Sidenote, I wish all websites supported markdown properly and not a custom weird subset they found convenient.
Word filters are only the beginning. LLMs are being phased in to flag and filter content based on more sophisticated criteria.
I read somewhere that chinese people used the ability of their language to form new meanings by concatenating multiple symbols in many different ways to get around censorship and that each time the new combination was banned, they came up with a new one. I wonder how long that'll be possible.
I think you're not taking what I wrote nearly literally enough. Really, you should be showing me diagrams of the Von Neumann architecture missing a censorship module. Maybe even gasp at the omission of it in Babbage's letters.
But why stop there? Let's bring out the venerable Abacus! We could have riveting discussions about how societies even back then, thousands of years ago, designated certain language as foul, and had rules about not using profanities in various settings. Ah, if only they knew they were actually victims of Orwellian censorship, and a globalist conspiracy.
If people are tolerant, they're woke. If people are intolerant, they're nazis. Standards of conduct are oppressive, but the lack of them indicate a shithole. And of course, anything in-between is just performative fence-sitting.
Tough crowd over here. Cultural bankruptcy speedrun much?
Ah yes, after muricans bad, let's have some euros bad.
I learn some amazing things on this site. Apparently the culture agnostic, historical practice of designating words and phrases as distasteful is actually a modern American, European, no actually Globalist, but ah no actually religious, but also no maybe Chinese?, no, definitely a Russian mind virus. Whatever the prominent narrative is for the given person at any given time.
Bit like when "mums is blaming everything on the computer". Just with political sophistry.
People easily forgot how they laughed at wizards in Harry Potter series who said "You-Know-Who" instead of "Voldemort". Now they are doing exactly the same thing.
Gemini is very prone to go into an infinite loop. Sometimes, it even happens with Google's own vibe coding IDE (Antigravity): https://bsky.app/profile/egeozcan.bsky.social/post/3maxzi4gs...
It also happened to me in the gemini-cli. It tried to think but somehow failed and putted all thoughts into the output and tried again and again to switch to "user output". If was practically stuck in an infinite loop
Yep. It happens all the time. Happened to me about 5 minutes ago. It does detect this and offer you the option to stop the loop or to let it continue.
> "A potential loop was detected. This can happen due to repetitive tool calls or other model behavior. The request has been halted."
So far they don't look to be doing anything about it, but Gemini models have a serious repetition bug.
I don't think that it is related to a specific prompt, like a "prompt logic issue" badly understood by the model, but instead, it looks like that sometimes it generates things that makes it go nuts.
My best intuition is that sometimes it forgets all the context and just look at the last X tokens as context before the repetition, and so start repeating like if the last generated tokens are the only thing that you gave to it.
All LLMs are, it's an innate thing. Google just sucks at the kind of long context training you need to do to mitigate that.
I would bet they won't suck at it for much longer, Gemini's progress in undeniable.
It was a consistent weak point for Gemini, compared to other major AIs. Reportedly, still is.
The progress is undeniable, the performance only ever goes up, but I'm not sure if they ever did anything to address this type of deficiency specifically. As opposed to being carried upwards by spillover from other interventions.
``sometimes, it even happens with [antigravity]``
Isn't this a problem with the agent loop / structure, rather than the llm, in that case?
The ide doesn't affect the models results, just what is done with those results?
I thought it was a specific prompt that breaks it, and that it's just something they never tested against, but when I saw it happen in antigravity, which supposedly must have been tested with a very specific use case, then I was very surprised.
The problem happens across tools that use Gemini
> Brainf*ck is the antithesis of modern software engineering. There are no comments, no meaningful variable names, and no structure
That's not true. From the little time I've spent trying to read and write some simple programs in BF, I recall good examples being pretty legible.
In fact, because the language only relies on those few characters, anything else you type becomes a comment. Linebreaks, whitespace, alphanumeric characters and so on, they just get ignored by the interpreter.
Have a look at this, as an example: https://brainfuck.org/chessboard.b
I also wonder whether brainfuck (ie turing machine like) coding would be a more efficient interface to LLMs
For those who want to try it, there’s always the https://raku.org module…
To me, that's still unreadable. While the intention of the code may be documented, it's pretty hard to understand if that "+" is really correct, or if that "<" should actually be a ">". I can't even understand if a comment starts or terminates a particular piece of code.
BTW, how come there are dashes in the comment?
The initial long comment starts with the [ command and ends with the ] command so it forms a loop that is executed while the current cell is nonzero. But initially, all tape cells are zero, so the whole loop is in fact skipped.
Readability is a spectrum. The brainfuck code is still somewhat readable compared to for instance this Binary Lambda Calculus program:
00010001100110010100011010000000010110000010010001010111110111101001000110100001110011010000000000101101110011100111111101111000000001111100110111000000101100000110110
or even the lambda term λ(λ1(1((λ11)(λλλ1(λλ1)((λ441((λ11)(λ2(11))))(λλλλ13(2(64)))))(λλλ4(13)))))(λλ1(λλ2)2) it encodes.
First, the parent comment didn't say anything about a spectrum. It just posited "it's legible." But it isn't to me, nor to 99.999% of the people here, I assume. Even those who've dabbled once in BF will probably find it hard, as the comment admits to using tricks.
Second, while readability comes in various degrees (probably more of a high-dimensional value than a linear spectrum, but well), the only thing that's readable about brainfuck is the comment. The code itself is not understandable, unless you really start digging into it and manage to understand the state it is in at every step of the program. Even then I would argue it isn't readable: your vision provides very few clues to the meaning of each step.
The comment serves as a guidance where certain parts start (or end, I can't tell). It explains a few things about the code, but even from the comment I cannot understand what it does. Also, the comment might be entirely wrong. There's only a very hard way to tell.
Your binary lambda example is also unreadable, but at least it doesn't have as much state as the BF program (which, admittedly, is much larger). Breaking it down might require less effort.
> it doesn't have as much state as the BF program
Actually, it has much more, as it's an infinitely large prime number sieve, as shown on top of my webpage https://tromp.github.io/cl/cl.html
> That's not true. From the little time I've spent trying to read and write some simple programs in BF, I recall good examples being pretty legible.
Anything in a reasonably familiar type face and size will continue to be legible, however brainfuck is not easily human parsable.
Greatly reducing its ability to be _read and mentally internalized._ With out that, are you really doing software engineering or are you actually a software maintenance person?
A janitor doesn't need to understand how energy generation works if he has to change the light bulb.
Asked for a solution of a photographed Ubongo puzzle: https://gemini.google.com/share/f2619eb3eaa1
Gemini Pro neither as is nor in Deep Research mode even got the number of pieces or relevant squares right. I didn't expect it to actually solve it. But I would have expected it to get the basics right and maybe hint that this is too difficult. Or pull up some solutions PDF, or some Python code to brute force search ... but just straight giving a totally wrong answer is like ... 2024 called, it wants its language model back.
Instead in Pro Simple it just gave a wrong solution and Deep Research wrote a whole lecture about it starting with "The Geometric and Cognitive Dynamics of Polyomino Systems: An Exhaustive Analysis of Ubongo Puzzle 151" ... that's just bullshit bingo. My prompt was a photo of the puzzle and "solve ubongo puzzle 151"; in my opinion you can't even argue that this lecture was to be expected given my very clear and simple task description.
My mental model for language models is: overconfident, eloquent assistant who talks a lot of bullshit but has some interesting ideas every now and then. For simple tasks it simply a summary of what I could google myself but asking an LLM saves some time. In that sense it's Google 2.0 (or 3.0 if you will)
Deep research, from my experience, will always add lectures.
I'm trying to create a comprehensive list of English standup specials. Seems like a good fit! I've tried numerous times to prompt it "provide a comprehensive list of English standup specials released between 2000 and 2005. The output needs to be a csv of verified specials with the author, release date and special name. I do not want any other lecture or anything else. Providing anything except the csv is considered a failure". Then it creates it's own plan and I go further clarifying to explicitly make sure I don't want lectures...
It goes on to hallucinate a bunch of specials and provide a lecture on "2000 the era of X on standup comedy" (for each year)
I've tried this in 2.5 and 3. Numerous time ranges and prompts. Same result. It gets the famous specials right (usually), hallucinates some info on less famous ones (or makes them up completely) and misses anything more obscure
I tried asking for a list of the most common gameboy color games not compatible with the original dmg gameboy. Chatgpt would over and over list dmg compatible games instead. I asked it to cross reference lists of dmg games to remove them and it ”reasoned” for a long time before it showed what sources it used for cross references, and then gave me the same list again.
It also insisted on including ”Shantae” in the list, which is expensive specifically because it is uncommon. I eventually forbid it from including the game in the list, and that actually worked, but it would continue mentioning it outside the list.
Absolute garbage.
I mean, isn't that a little ridiculous? Aren't those language models already solving complicated exam questions and mathematical problems?
According to the creators, the models are on a phd level of intelligence, but they can’t get the simplest thing right.
Overselling is only the tip of the iceberg. The real problem is that a lot of managers base their decision to introduce language models into business processes on cutting edge Pro edition demos, but what is, of course, actually used in production is some cheap Nano/Flash/Mini version.
Too easy.
LLM's are bad at anything with images.
There is something fucky about tokenizing images that just isn't as clean as tokenizing text. It's clear that the problem isn't the model being too dumb, but rather that model is not able to actually "see" the image presented. It feels like a lower-performance model looks at the image, and then writes a text description of it for the "solver" model to work with.
To put it another way, the models can solve very high level text-based problems while struggling to solve even low level image problems - even if underneath both problems use a similar or even identical solving frameworks. If you have a choice between showing a model a graph or feeding it a list of (x,y) coordinates, go with the coordinates every time.
Gemini is my favorite, but it does seem to be prone to “breaking” the flow of the conversation.
Sharing “system stuff” in its responses, responding to “system stuff”, starts sharing thoughts as responses, responses as thoughts, ignoring or forgetting things that were just said (like it’s suddenly invisible), bizarre formatting, switching languages for no reason, saying it will do something (like calling a tool) instead of doing it, getting into odd loops, etc.
I’m guessing it all has something to do with the textual representation of chat state and maybe it isn’t properly tuned to follow it. So it kinda breaks the mould but not in a good way, and there’s nothing downstream trying to correct it. I find myself having to regenerate responses pretty often just because Gemini didn’t want to play assistant anymore.
It seems like the flash models don’t suffer from this as much, but the pro models definitely do. The smarter the model to more it happens.
I call it “thinking itself to death”.
It’s gotten to a point where I often prefer fast and dumb models that will give me something very quickly, and I’ll just run it a few times to filter out bad answers, instead of using the slow and smart models that will often spend 10 minutes only to eventually get stuck beyond the fourth wall.
> ignoring or forgetting things that were just said (like it’s suddenly invisible)
This sounds like an artifact of the Gemini consumer app, some others may be too (the model providers are doing themselves a disservice by calling them the same).
Too bad it can't explain why it does the same thing with actual English.actual English.actual English.actual English.actual English.actual English.
It overruns the max headroom of the context window.
https://youtu.be/cYdpOjletnc?t=6
I often hear comparisons to Web 1.0 (the bubble aspect, potential for change, etc).
As someone who lived and worked during that era, I don’t remember thinking “holy shit, if this ever gets released on the world at scale we’ll have serious problems”.
Maybe that was a lack of imagination and not thinking through what would actually happen to brick and mortar, the job market and so on. But it feels like this time is different. Or I’m just that much older.
All open weights model I tried (that fit under 20GB of memory) easily loop.
I run models with llama.cpp and the reason why I add some repeat penalty factor.
Why would anyone feel compelled to use AI to write such a short blog post? Is there no space where I can assume the written contented is communicated 100% by another human being?
I am sorry if it appears that it was written by AI - I wrote a draft and used AI to assist me since English, is not my first language. I asked it only to format but it has seemed to change the tone and the expressions too '.'
I'm also not a native English speaker, but I've decided to avoid using AI for formatting or changing the tone of what I write. That tends to result in extremely generic outputs that "feel" AI, no matter how much effort I put into writing it.
Asking for it to point out mistakes, without providing alternatives, seems like a better way to actually get better at writing.
Prompting the Ai to use a specific tone might result in something that's less generic, but imo that's not the right place to spend efforts.
I personally prefer some grammatical errors or awkward phrasing over AI-assisted writing. It's a blog post, not a diplomatic transcript.
You're absolutely right!
English is also not my first language. I understand the challenge, but I'd recommend to write it in English and then ask AI to suggest rephrasing in wrong or poorly phrased sentences. Right now it looks almost entirely AI-generated unfortunately and does not show the thought you had when writing it. Cheers.
All of a sudden, internet is full of people who hate AI written articles. A few months back, my article got a lot of haters because I used AI tools to improve my draft. Being a non-english first language person, I don't see an issue. But I wish AI improves to an extend where draft to complete articles don't look AI written.
You should use AI to point out errors or suggest better phrasing. But if you ask AI to rewrite your post, it will produce content that sounds fake and corporate. ESL speakers may not notice it but everyone else does.
> A few months back, my article got a lot of haters because I used AI tools to improve my draft. Being a non-english first language person, I don't see an issue.
(Speaking as another ESL user: )
Try doing something similar in your first language and I think you’ll see the issue, especially if you arrange for the model input to be somewhat flawed (e.g. roundtrip it through a machine-translation tool first). The “edited” writing is extremely generic by default and feels bad even if you adjust the prompt. It’s the kind of aggressively bland that you get from a high schooler who was extensively trained to write essays but doesn’t actually read books, except even the most beat-down of high schoolers can’t help but let their imagination shine through sometimes, while the chat models have been subjugated much more effectively.
Also, well, it’s a social marker. Language is a mess of social markers: there’s no fundamental reason why reducing this vowel should be OK but reducing that one should be “sloppy” and low-class. And AI writing (which undeniably has a particular flavour) is hit by a double whammy of being used by people who don’t really care to write (and don’t have a taste for good writing) and having been tuned by people who tried to make it as inoffensive as it could possibly be to any social group they could think of (and don’t have a taste for good writing). Is that unfair, especially to non-native speakers? All of language learning is unfair. Always has been.
I also don't have English as my first language and I think it's a shitty excuse.
Articles written by AI are soulless and shitty. Do yourself and the readers a favor and write yourself, even if it contains errors.
They sound like politician speak or corporate speak.
To the OP: do you like how your politicians sound in your native language? If not, don't let a LLM rewrite your article.
Btw, I'm not a native speaker either.
And why does anybody trust AI at all when it produces a typo ("amost") in the very first sentence of an article?
[dead]
I wonder if going the other way, maxing out semantic density per token, would improve LLM ability (perhaps even cost).
We use naturally evolved human languages for most of the training, and programming follows that logic to some degree, but what if the LLMs were working in a highly complex information dense company like Ithkuil? If it stumbles on BF, what happens with the other extreme?
Or was this result really about the sparse training data?
I wonder the same. I think a language like pascal is more semantically rich than C-like languages. Something like:
Probably holds more semantically significant tokens than the C-counterpart.But with LLM's the problem's gotta be training data. But if there was as much training data in Pascal as there is in C it would be pretty cool to see benchmarks, I have a hunch Pascal would do better.
(Sorry for the bad pascal I haven't programmed in ages)
I thought "The Data Scarcity Problem" from the article is very well known to us engineers?
It's where the pulleys of a very sophisticated statistical machine start to show, and the claims about intelligence start to crumble.
Reason AI is great for boilerplate (because it's been done a million times) and not so great for specifics (even if they're specifics in a popular language).
Gemini does this a lot, getting stuck generating the same tokens over and over indefinitely
> So it made me wonder. Is Brainf*ck the ultimate test for AGI?
Absolutely not. Id bet a lot of money this could be solved with a decent amount of RL compute. None of the stated problems are actually issues with LLMs after on policy training is performed.
> None of the stated problems are actually issues with LLMs after on policy training is performed
But still , isnt it a major weakness they have to RL on everything that has not much data? That really weakens the attempt to make it true AGI.
No.
AGI would be a universal learner, not a magic genie. It still needs to do learning (RL or otherwise) in order to do new tasks.
> It still needs to do learning (RL or otherwise) in order to do new tasks.
Why ? As in - why isn't reading the Brainfuck documentation enough for Gemini to learn Brainfuck ? I'd allow for 3-7 days of a learning curve like perhaps a human would need but why do you need to kinda redo the whole model (or big parts of it) just so it could learn Brainfuck or some other tool? Either the learning (RL or otherwise) need to become way more efficient than it is today (takes today weeks? months? billions of dollars) or it isn't AGI I would say. Not in practical/economic sense and I believe not in the philosophical sense of how we all envisioned true generality.
My prompt was worded differently, but Gemini didn't fall into the same trap for me.
``` Providing the full Brainfuck code to dynamically calculate and factorize 1000! is unfortunately not feasible, and here is why: * The Scale: 1000! (1000 factorial) is a number with 2,568 digits. Standard Brainfuck operates on 8-bit memory cells (values 0-255). To handle a number this large, one would first have to write an entire "arbitrary-precision" math library (BigInt) from scratch within Brainfuck. * The Algorithm: You cannot simply compute 1000! and then factor it. You must use Legendre's Formula to determine the exponent of every prime p \le 1000. * Formula: E_p(n!) = \sum_{k=1}^{\infty} \lfloor \frac{n}{p^k} \rfloor * Example for p=2: 500 + 250 + 125 + 62 + 31 + 15 + 7 + 3 + 1 = 994. * The first term of the answer is 2^{994}. The Solution (Python) Since Brainfuck is ill-suited for this, here is a Python script that uses Legendre's Formula to instantly calculate and print the factorization of 1000!. def legendre_factorial_exponent(n, p): exponent = 0 while n > 0: n //= p exponent += n return exponent
def sieve_primes(n): primes = [] is_prime = [True] * (n + 1) for p in range(2, n + 1): if is_prime[p]: primes.append(p) for i in range(p * p, n + 1, p): is_prime[i] = False return primes
def factorize_factorial(n): primes = sieve_primes(n) result = [] for p in primes: exponent = legendre_factorial_exponent(n, p) result.append(f"{p}^{exponent}") return " * ".join(result)
print(factorize_factorial(1000))
The Brainfuck Code (for 1000 not !) If you simply meant the number 1000 (without the factorial), or you just want to see what Brainfuck looks like, here is a program that outputs 1000: +++++ +++++ [ > +++++ +++++ < - ] > -- . (Print 100 'd') < +++++ +++++ [ > +++++ +++++ < - ] > + . (Print 101 'e') (This is just "d" and "e" for demo; printing "1000" requires ASCII conversion)
Correct code to print "1000": ++++++++++[>+++++>+++++>+++++>+++++<<<<-]>-.>-.>-.>-.
Explanation: * Sets cell #0 to 10. * Loops 10 times, adding 5 to cells #1, #2, #3, #4 (making them all 50). * Subtracts 1 from cell #1 (ASCII 49 = '1'). * Subtracts 2 from cells #2, #3, #4 (ASCII 48 = '0'). * Prints them. Would you like me to explain how Legendre's Formula works in more detail? ```
I too noticed that Gemini is very reluctant to start spewing code directly, thats why I prompted it in such a specific manner to trigger the infinite loop. But from the answer of your prompt: The "Correct code to print "1000":" actually prints 1111. So yeah it is still wrong even for something super simple.
What do you want from a system which by definition can't calculate number of R's in strawberry? (yes, still can't; gives random answer if you slightly modify the question).
Got the same with ChatGPT and a simple web page with tiles.
Whereby I don’t know if it was a real infinite loop because I cancelled the session after 10 minutes seeing always the same "thoughts" looping
I've tried to have Gemini generate code for me, and it will often go through the thinking and planning process, appear to generate code, and then...not actually output it.
Write a Brainfuck program to output the Seahorse Emoji then halt.
Asking Gemini 3 to do anything results in an infinite loop.
It doesn't create infinite loop in brainfuck, but looped itself.
Viva the Brainfuck! The language of anti-AI resistance!
I long for quantum computing where white space will be able to be a space and a tab at the same time.
Saying "Asking Gemini 3" doesn't mean much. The video/animation is using "Gemini 3 Fast". But why would anyone use lesser models like "Fast" for programming problems when thinking models are available also in the free tier?
"Fast" models are mostly useless in my experience.
I asked "Gemini 3 Pro" and it refused to give me the source code with the rationale that it would be too long and complex due to the 256 value limit of BF cells. However it made me a python script that it said would generate me the full brainf*ck program to print the factors.
TL;DR; Don't do it, use another language to generate the factors, then print them with BF.
I agree but it is kinda strange that this model (Gemini 3 fast) achieved such a high score on ARC-AGI-2. Makes you wonder.
what the fuck compelled you to censor "Brainfuck"?
Visibilty - i have no idea if there are censoring algorithms at play anywhere.
Chilling effects. Western culture is taken over by American Puritian values thanks to the globlaisation of the media.
Mhmm, so chilling. Cause word filters aren't as old as computing itself...
Don’t need to ban speech when your population preemptively does it for you in fear of an unaccountable corporation blocking you.
Don't need to ban speech when people on their soapboxes keep telling me I need to be in terror.
Will somebody pleeeaaaase think of American Puritanism and Globalism?
"Unalive" has reached mainstream usage, on account of those inscrutable censors. If that is not the spitting picture of Newspeak I don't know what is.
The trend of self-censoring words like 'dead' and 'kill' appears to be relatively new, motivated by TikTok and YouTube algorithms, but spilling over into the general internet.
The sewer section of sites like the daily mail has been needlessly censoring words for well over a decade
Correlation is not causation but I challenge anyone to come up with a different cause:
https://trends.google.com/trends/explore?date=all&q=tiktok,u...
https://trends.google.com/trends/explore?date=all&q=unalive&...
I agree, although I was referring to asterisks like de*d and k*ll (or censoring with black bars, or using emojis) - euphemisms of course have always been part of language evolution.
I chose unalive because i didn't know google trends allowed searching for asterisks. Appears it does. k*ll was apparently used even before tiktok but usage increased markedly around the same time as unalive appeared. Interestingly d*ad and r*pe don't follow this pattern. I am not sure it treats asterisks correctly, nor that google trends is the right tool to research this, given people searching for the word is only a poor indicator of its usage.
Sidenote, I wish all websites supported markdown properly and not a custom weird subset they found convenient.
[dead]
In the U.K. “Trump” is a synonym for “Fart”
Word filters are only the beginning. LLMs are being phased in to flag and filter content based on more sophisticated criteria.
I read somewhere that chinese people used the ability of their language to form new meanings by concatenating multiple symbols in many different ways to get around censorship and that each time the new combination was banned, they came up with a new one. I wonder how long that'll be possible.
passwords were a foreign concept to early computing, but you presume censorship was taking place?
it took awhile of corporatization and profit-shaping before censorship on computers really took off in any meaningful way.
...but it wasn't for any reasons other than market broadening and regulation compliance.
I think you're not taking what I wrote nearly literally enough. Really, you should be showing me diagrams of the Von Neumann architecture missing a censorship module. Maybe even gasp at the omission of it in Babbage's letters.
But why stop there? Let's bring out the venerable Abacus! We could have riveting discussions about how societies even back then, thousands of years ago, designated certain language as foul, and had rules about not using profanities in various settings. Ah, if only they knew they were actually victims of Orwellian censorship, and a globalist conspiracy.
It's even possible to spell out BOOBS on an abacus.
TIL TikTok is american.
Puritans were English protestants. I think you mean to say it's being taken over by European values.
Puritans were kicked out of Europe for their views
Nothing says tolerance and no censorship like kicking out people for their views
So European values of the 1500s through 1700s are not puritan values.
Europe has liberalised since then.
If people are tolerant, they're woke. If people are intolerant, they're nazis. Standards of conduct are oppressive, but the lack of them indicate a shithole. And of course, anything in-between is just performative fence-sitting.
Tough crowd over here. Cultural bankruptcy speedrun much?
Ah yes, after muricans bad, let's have some euros bad.
I learn some amazing things on this site. Apparently the culture agnostic, historical practice of designating words and phrases as distasteful is actually a modern American, European, no actually Globalist, but ah no actually religious, but also no maybe Chinese?, no, definitely a Russian mind virus. Whatever the prominent narrative is for the given person at any given time.
Bit like when "mums is blaming everything on the computer". Just with political sophistry.
Censoring shit or fuck is very much not a European thing.
People easily forgot how they laughed at wizards in Harry Potter series who said "You-Know-Who" instead of "Voldemort". Now they are doing exactly the same thing.
-> expects reasoning
-> runs it in Gemini fast instead of thinking
....