Evidence of Controversial Planet 9 Uncovered In Sky Surveys Taken 23 Years Apart (space.com) 149
Astronomers may have found the best candidate yet for the elusive Planet Nine: a mysterious object in infrared sky surveys taken 23 years apart that appears to be more massive than Neptune and about 700 times farther from the sun than Earth. Space.com reports: [A] team led by astronomer Terry Long Phan of the National Tsing Hua University in Taiwan has delved into the archives of two far-infrared all-sky surveys in search of Planet Nine -- and incredibly, they have found something that could possibly be Planet Nine. The Infrared Astronomy Satellite, IRAS, launched in 1983 and surveyed the universe for almost a year before being decommissioned. Then, in 2006, the Japanese Aerospace Exploration Agency (JAXA) launched AKARI, another infrared astronomy satellite that was active between 2006 and 2011. Phan's team were looking for objects that appeared in IRAS's database, then appeared to have moved by the time AKARI took a look. The amount of movement on the sky would be tiny -- about three arcminutes per year at a distance of approximately 700 astronomical units (AU). One arcminute is 1/60 of an angular degree.
But there's an extra motion that Phan's team had to account for. As the Earth orbits the sun, our view of the position of very distant objects changes slightly in an effect called parallax. It is the same phenomenon as when you hold your index finger up to your face, close one eye and look at your finger, and then switch eyes -- your finger appears to move as a result of you looking at it from a slightly different position. Planet Nine would appear to move on the sky because of parallax as Earth moves around the sun. On any particular day, it might seem to be in one position, then six months later when Earth is on the other side of the sun, it would shift to another position, perhaps by 10 to 15 arcminutes -- then, six months after that, it would seem to shift back to its original position. To remove the effects of parallax, Phan's team searched for Planet Nine on the same date every year in the AKARI data, because on any given date it would appear in the same place, with zero parallax shift, every year. They then also scrutinized each candidate object that their search threw up on an hourly basis. If a candidate is a fast-moving, nearby object, then its motion would be detectable from hour to hour, and could therefore be ruled out. This careful search led Phan's team to a single object, a tiny dot in the infrared data.
It appears in one position in IRAS's 1983 image, though it was not in that position when AKARI looked. However, there is an object seen by AKARI in a position 47.4 arcminutes away that isn't there in the IRAS imagery, and it is within the range that Planet Nine could have traveled in the intervening time. In other words, this object has moved a little further along its orbit around the sun in the 23 or more years between IRAS and AKARI. The knowledge of its motion in that intervening time is not sufficient to be able to extrapolate the object's full orbit, therefore it's not yet possible to say for certain whether this is Planet Nine. First, astronomers need to recover it in more up-to-date imagery. [...] Based on the candidate object's brightness in the IRAS and AKARI images, Phan estimates that the object, if it really is Planet Nine, must be more massive than Neptune. This came as a surprise, because he and his team were searching for a super-Earth-size body. Previous surveys by NASA's Wide-field Infrared Survey Explorer (WISE) have ruled out any Jupiter-size planets out to 256,000 AU, and any Saturn-size planets out to 10,000 AU, but a smaller Neptune or Uranus-size world could still have gone undetected. Phan told Space.com that he had searched for his candidate in the WISE data, "but no convincing counterpart was found because it has moved since the 2006 position," and without knowing its orbit more accurately, we can't say where it has moved to. "Once we know the position of the candidate, a longer exposure with the current large optical telescopes can detect it," Phan told Space.com. "However, the follow-up observations with optical telescopes still need to cover about three square degrees because Planet Nine would have moved from the position where AKARI detected it in 2006. This is doable with a camera that has a large field of view, such as the Dark Energy Camera, which has a field of view of three square degrees on the Blanco four-meter telescope [in Chile]."
But there's an extra motion that Phan's team had to account for. As the Earth orbits the sun, our view of the position of very distant objects changes slightly in an effect called parallax. It is the same phenomenon as when you hold your index finger up to your face, close one eye and look at your finger, and then switch eyes -- your finger appears to move as a result of you looking at it from a slightly different position. Planet Nine would appear to move on the sky because of parallax as Earth moves around the sun. On any particular day, it might seem to be in one position, then six months later when Earth is on the other side of the sun, it would shift to another position, perhaps by 10 to 15 arcminutes -- then, six months after that, it would seem to shift back to its original position. To remove the effects of parallax, Phan's team searched for Planet Nine on the same date every year in the AKARI data, because on any given date it would appear in the same place, with zero parallax shift, every year. They then also scrutinized each candidate object that their search threw up on an hourly basis. If a candidate is a fast-moving, nearby object, then its motion would be detectable from hour to hour, and could therefore be ruled out. This careful search led Phan's team to a single object, a tiny dot in the infrared data.
It appears in one position in IRAS's 1983 image, though it was not in that position when AKARI looked. However, there is an object seen by AKARI in a position 47.4 arcminutes away that isn't there in the IRAS imagery, and it is within the range that Planet Nine could have traveled in the intervening time. In other words, this object has moved a little further along its orbit around the sun in the 23 or more years between IRAS and AKARI. The knowledge of its motion in that intervening time is not sufficient to be able to extrapolate the object's full orbit, therefore it's not yet possible to say for certain whether this is Planet Nine. First, astronomers need to recover it in more up-to-date imagery. [...] Based on the candidate object's brightness in the IRAS and AKARI images, Phan estimates that the object, if it really is Planet Nine, must be more massive than Neptune. This came as a surprise, because he and his team were searching for a super-Earth-size body. Previous surveys by NASA's Wide-field Infrared Survey Explorer (WISE) have ruled out any Jupiter-size planets out to 256,000 AU, and any Saturn-size planets out to 10,000 AU, but a smaller Neptune or Uranus-size world could still have gone undetected. Phan told Space.com that he had searched for his candidate in the WISE data, "but no convincing counterpart was found because it has moved since the 2006 position," and without knowing its orbit more accurately, we can't say where it has moved to. "Once we know the position of the candidate, a longer exposure with the current large optical telescopes can detect it," Phan told Space.com. "However, the follow-up observations with optical telescopes still need to cover about three square degrees because Planet Nine would have moved from the position where AKARI detected it in 2006. This is doable with a camera that has a large field of view, such as the Dark Energy Camera, which has a field of view of three square degrees on the Blanco four-meter telescope [in Chile]."
A fitting use for AI (Score:5, Insightful)
If they'd'let loose an AI on all the solar system data and it would uncover more (circumstancial) evidence for the planet, that would be a fitting task and actually ineresting AI news for once.
Re: (Score:2, Informative)
Re: (Score:2)
Re: (Score:2)
Let me make my point crystal clear: LLMs cannot even perform floating point multiplication to standard precision, the most basic operation in numerical processing.
Trained LLMs may know how to multiply floating point numbers or matrices, and they can also describe
Re: (Score:2)
Re: (Score:2)
Re: (Score:3)
I just tried on a smaller local model, and it aced 10x10 with ease. Of course it took it 30k tokens to do it, but there's no reason it wouldn't reliably scale to the size of the context. A million-token context window should be able to do a decently sized matrix.
It's highly unlikely that a non-CoT, zero-shot prompt would be able to do it, though.
Each token is com
Re: (Score:3)
Re: (Score:2)
Re: (Score:2)
Also keep in mind that the 10x10 matrix multiply example you gave should take 1000 float fused multiply-adds (fma) operations.
Correct.
On the LLM it would have taken many orders of magnitude more, likely billions of fma and millions of times less efficient, assuming it even obtained the correct result (did it?).
What? lol. That's pure lunacy.
You think it can't do basic math iteratively?
And yes, it did come to the correct answer. Of course, like I said, it took about 30k tokens to do it, which is ridiculously inefficient, but the point was to prove that you were wrong, not to prove that it was an efficient matrix multiplier.
Ironically, the LLM is implemented USING matrix and tensor operations, but is very poor at DOING these operations at the token generation and inference level.
Poor? No, it's perfect fine at doing it- it's just not efficient, for very obvious reasons.
Re: (Score:2)
My original claim was this: “Are you talking about LLMs that can barely multiply even small matrices? They are language models — terrible at numerical analysis, especially for the vast data sets in this scenario.”
My follow up statement, to which you replied, was this:
Re: (Score:2)
> LLMs are extremely inefficient at even moderate numerical processing
Re: (Score:2)
Maybe read the thread. It is ALL about efficiency and reliability: whether an LLM can do numeric processing on a scale needed for processing the astronomy imaging datasets discussed in the article.
That's not how AI-assisted research works.
The relevance here of showing that they can reliably multiply matrices is to demonstrate that they understand the fundamentals.
You'd no sooner have an LLM manually compute a model than you'd have a human. The LLM would design it, including its training methodology, and look at the results.
My follow up statement, to which you replied, was this: “LLMs (themselves) cannot even perform a moderate matrix operation reliably. This is because they are language models, and have very poor performance on large numerical tasks.”
And this statement is absurdly false.
The reliability isn't the problem, the problem is you're treating the LLM like it's a first year college student. Why on earth would it com
Re: (Score:2)
Re: (Score:2)
Read my very first statement: “Are you talking about LLMs that can barely multiply even small matrices? They are language models — terrible at numerical analysis, especially for the vast data sets in this scenario. ”
Not sure what your inference skills are like, but my concern there is efficiency of LLMs themselves. Read the quote. They CANNOT perform large numerical cal
Re: (Score:2)
Re: (Score:2)
You said:
My point was that LLMs (themselves) cannot even perform a moderate matrix operation reliably.
Your point was reliability, and now you're trying to walk it back. Just fucking admit it, ffs.
That LLMs can barely do a 10x10 matrix multiply, and likely not a 100x100 matrix multiply, proves my point.
No, it doesn't. The fact that it can do it reliably flatly disproves your original point. Your new point was never contested by me. Trying to pretend like it is won't save your argument.
Yes- LLMs are not great number crunchers, as we have both said.
They are, however, perfectly capable of being reliable number crunchers, just like a person is.
There's a reason my response l
Re: (Score:2)
Of course, even though LLMs are extraordinarily bad at numeric processing
No, they're not.
inefficient in compute time and memory to the extent that even moderate tasks are intractable
Yes on inefficient in compute time and memory. To the extent that even moderate tasks are intractable? Now that's just silly.
they are good at understanding and explaining this limitation. https://chatgpt.com/share/6816 [chatgpt.com]... [chatgpt.com]
It did no such thing. You asked it how inefficient it was for it to multiple a couple of matrices. It correctly answered. You keep beating this dead horse.
The only reason I brought it up, is because you claimed it couldn't be reliably done, which was wrong. It can't be efficiently done- which is perfectly true.
From the perspective of data analysis, if you're using i
Re: (Score:2)
Not sure what your inference skills are like, but
He's a neckbeard. He may sometimes use technical words and make technical claims, but any argument with him boils down to he's right and you're wrong. No logical fallacy is too large or blatant in service of that goal. And he doesn't actually understand any of it, anyway; what he uses in place of thinking is not that dissimilar to the LLM. He's just spewing free associations that feel like they support him.
He's been "arguing" like this since he created his account a couple decades ago.
Re: (Score:2)
Are you talking about LLMs that can barely multiply even small matrices? They are language models — terrible at numerical analysis, especially for the vast data sets in this scenario.
and
LLMs (themselves) cannot even perform a moderate matrix operation reliably. This is because they are language models, and have very poor performance on large numerical tasks.
LLM are language models. They CANNOT do even very simple numeric processing tasks, eg say a 1024x1024 2D FFT. This is because the scaling factors in time and space make this task IMPOSSIBLE in the LLM. On small operations they can do they are thousands or millions of times slower than a native implementation. On larger tasks they FAIL due to compute
Re: (Score:2)
You seem to be having an argument no one else is.
Stop it.
My point was that LLMs (themselves) cannot even perform a moderate matrix operation reliably.
LLMs can reliably do matrix multiplication
The scale problem is your adjusted argument.
Nobody is questioning the scale limitations of using an LLM to do non-complex math.
The astronomers in the article aren't morons. They're not using an LLM to do a million-point FFT.
They'd be using them to observe the results of several FFTs.
You're lying about what the argument was initially about, which was LLM reliability/hallucination and beating a dead horse on an argument no one ever made about LLM scale.
You clearified what you meant:
LLMs know how to describe the operations needed to perform large matrix multiplications, or land an aircraft, or bake muffins. This is because they have read descriptions of these activities during training, and can therefore generate new descriptions of the same activities.
And I pointe
Re: (Score:2)
LLMs can reliably do matrix multiplication
You removed the important word from my quote: “moderate”.
They FAIL at performing even moderately-sized matrix multiplications (or similar operations, eg FFT), because of very large inefficiency in memory and time. A small 10x10 matrix might succeed, but even a moderate 100x100 matrix mul likely will not. Similarly for a million-point FFT.
Never once did I argue that they were an efficient way to do them, or lacked scaling issues. That was your redirect to try to rescue a bullshit argument you made.
Nope. It was my original point, and main point throughout.
You're lying about what the argument was initially about, which was LLM reliability/hallucination and beating a dead horse on an argument no one ever made about LLM scale.
You’re slower than an LLM doing numeric processing.
My original point, and main point through
Re: (Score:2)
You removed the important word from my quote: “moderate”.
You're such a fucking liar, lol.
Deflect, deflect, deflect.
One does not describe something that another thing cannot do as being unreliably able to do it.
The fact is, an LLM can never multiply a matrix requires more operations than it has context, period. For matrices that are small enough that it has context to compute- they are completely reliable. There is no lack of reliability that applies, period.
You were trying to imply that they couldn't accurately do it. You were wrong. And now you're trying t
Re: (Score:2)
Are you talking about LLMs that can barely multiply even small matrices? They are language models — terrible at numerical analysis, especially for the vast data sets in this scenario.
Can a large language model (LLM) accurately and reliably perform 100×100 matrix multiplications for the purpose of analysing an astronomical imaging dataset—or similar computations such as a 1D FFT of a 1024×1024 array or a 2D FFT of a million-element array? No, it cannot.
Can it barely perform a 16×16 matrix multiplication? Apparently it can (as you said), but likely only barely, and likely only with small integers or i
Re: (Score:2)
Ironically, the LLM is implemented USING matrix and tensor operations, but is very poor at DOING these operations at the token generation and inference level.
YOU:
Poor? No, it's perfect fine at doing it- it's just not efficient, for very obvious reasons.
You might call it “perfectly fine” and similar words you have used throughout. But for those of us who actually develop and optimise high-performance systems—whether it’s for light transport, fluid simulation, or similar workloads spread across tens of thousands of cores under strict production deadlines—we’d call it “poor” or even “terrible” as I have done (much to your confusion).
Even 10% performance impact would be poor or terrible. Ye
Re: (Score:2)
Re: (Score:2)
Are you talking about LLMs that can barely multiply even small matrices? They are language models — terrible at numerical analysis, especially for the vast data sets in this scenario. ML signal analysis and image processing, on the other hand, may be useful here.
I’d be surprised if an LLM can multiply even two 16x16 matrices to basic float32 precision. They are POOR at this task. They can BARELY do it. They take vastly too much TIME and SPACE by factors of millions. It is a task they are completely UNRELIABLE at in practical terms.
Re: (Score:2)
Re: (Score:2)
You just quoted a response this thread isn't even in reply to.
My point was that LLMs (themselves) cannot even perform a moderate matrix operation reliably. This is because they are language models, and have very poor performance on large numerical tasks. LLMs know how to describe the operations needed to perform large matrix multiplications, or land an aircraft, or bake muffins. This is because they have read descriptions of these activities during training, and can therefore generate new descriptions of the same activities. However, they cannot perform these operations without external help, for example an external code execution in Python or C++ using custom code or standard BLAS library (Eigen), or using a human pilot or chef to help them.
That is what this was in reply to.
You're very obviously trying to infer that "since an LLM is a language model, it can generate new descriptions, but not do what it is it is describing" (paraphrased)
You are wrong. And you have walked it back. And you're too fucking cowardly to admit it.
Re: (Score:2)
Regardless, an LLM literally can’t even multiply two float32 numbers to standard precision accurately or reliably. That is equivalent to a 1x1 matrix multiplication. Any aspirations for scientific numeric processing inside an LLE itself begin and end with this simple fact. All numeric processing must be done outside the LLE, for example
Re: (Score:2)
Can we at least agree that a technology (LLM) that cannot accurately or reliably perform even a single floating point multiplication in standard precision with bit-exact and repeatable results, let alone a small matrix multiply or large operation like an FFT, therefore cannot be used for any scientific numeric processing in any sensible context? And all processing must be done outside the LLE via li
Re: (Score:2)
Re: (Score:2)
These are just three of many things an LLM can describe but not perform in practice.
The assumption of working with floating point data was so basic I left it unstated in my posts. The pro
Re: (Score:2)
LLMs can calculate floating point numbers just fucking fine, to arbitrary precision. (within context limits).
"float32" isn't a standard anything.
Perhaps you're thinking of IEEE754 single-precision... in which case- ya, that's actually probably right. But that's not due to some inherent limitation in LLMs, that's simply because I don't think I've ever encountered one that well trained in computer science.
However, if the question is whether
Re: (Score:2)
This seems to be looping around again: are you looping back to the “LLMs can do integer or float maths just fine (they can’t) but why would I want to (I don’t)” stage, or are you finally ready for the realisation “LLMs can’t do integer or float maths just fine at all”
LLMs really do not operate the way you think they do. A standard LLM cannot compute floating point numbers, or even integers, “just fucking fine” and to arbitrary precision, because long bef
Re: (Score:2)
LLMs really do not operate the way you think they do. A standard LLM cannot compute floating point numbers, or even integers, “just fucking fine” and to arbitrary precision, because long before the context window is reached the probabilistic nature of the LLM leads to errors. They CANNOT do long division even over very short number of digits. Even the 24-bits of a float32 mantissa (with implicit one) would be a stretch.
This is flat out wrong.
LLMs are only probabilistic at the output layer- and for any kind of math/programming task, you use greedy decoding- which means you take the highest-ranked token.
The hidden states are not probabilistic at all, probabilities are simply how the network communicates with you (should you so choose).
You say there isn’t some inherent limitations in LLMs, but there is: despite their many strengths it is extremely difficult to train them to perform simple arithmetic even on integers and modest precision (eg 10 decimal digits). Basic school long division of say 6 digits by 3 digits is about the current practical limit. https://chatgpt.com/share/6818 [chatgpt.com]... [chatgpt.com]
Simply false. [markdownpastebin.com]
Model is Qwen3-32b-fp16. Standard hyperparameters and prompt template for reasoning.
I'd quit using ChatGPT to fill in for actual knowledge. Now that is not something LLMs ar
Re: (Score:2)
I chose floating point because most workloads use this, including the astronomy task in the article.Commodity hardware in a consumer GPU can do tens of trillions of float mads per second (single precision) and can do tens of trillions of dou
Re: (Score:2)
It's not like multiplying a matrix involves different operates depending on its size, which should be the first clue to you that it's not a difficult task for them.
There are very big reasons why the LLM architecture does not scale to a decently sized matrix.
No, there's 1 reason, and 1 reason alone. Because they use their context to do the math, and their context is limited.
Look at it this way, it can multiply any two arbitrary matrices better than you can, without a tool to help.
Even a small 1000x1000 matrix would be completely intractable.
Indeed. It would be for you t
Re: (Score:2)
Large scale linear anlgebra via matrix and tensor operations are excellent for implementing LLMs.A large proportion of their implementation and compute cost is precisely that.
But LLMs (themselves) are terrible for performing large scale matrix and tensor operations.
Using an LLM to do large scale numeric processing for astronomy images or signal detection is utterly ridiculous, No toy example
Re: (Score:2)
> Indeed. It would be for you too.
No, because if attempting to compute a large matrix a reasonable person approaching this from first-principles perspectives would use a math tool specialized in multiplying large matrices if that were the actual goal.
That is why an LLM is a poor tool for the task...it is a generalized language tool that can be forced to emulate the base functionality of its own tech stack somewhat poorly with ho
Re: (Score:2)
No, because if attempting to compute a large matrix a reasonable person approaching this from first-principles perspectives would use a math tool specialized in multiplying large matrices if that were the actual goal.
So would an LLM, which if you had read further along the thread, you would have seen it did.
That is why an LLM is a poor tool for the task...it is a generalized language tool that can be forced to emulate the base functionality of its own tech stack somewhat poorly with horrible efficiency. It wouldn't even show up in the top contenders for the task of multiplying large matrices.
Of course it wouldn't... what kind of idiotic fucking point are you trying to make? Do you imagine that someone contested this?
The discussion was on reliability.
Dude, you're a fucking imbecile- get lost.
Re: (Score:2)
So would an LLM
"would" is doing a lot of lifting.
"Doesn't" would be more accurate, though.
Just because you imagine it doesn't mean it is actually so. And when you argue against what is using only your imagination, it just makes you an idiot. It doesn't make you a visionary, or Future Man, or whatever.
Go and invent your LLM that is better at math than mathematicians using specialist tools, then you can talk. Until then, stfu, it doesn't exist.
Re: (Score:2)
"would" is doing a lot of lifting.
No, it isn't.
Because it did.
"Doesn't" would be more accurate, though.
Again with your reading problem. You know- you could take classes for that. They're very good at bringing special needs kids up to speed, these days.
Just because you imagine it doesn't mean it is actually so. And when you argue against what is using only your imagination, it just makes you an idiot. It doesn't make you a visionary, or Future Man, or whatever.
What in the fuck are you talking about, you intellectually handicapped simpleton?
There's no imagination anywhere here. We're talking numbers, and observables. You're the shit-for-brains over here trying to inject with nothing but some dumbshit hallucinations.
Go and invent your LLM that is better at math than mathematicians using specialist tools, then you can talk. Until then, stfu, it doesn't exist.
Doesn't need to be better than a mathematician- just needs to be better th
Re: (Score:2)
An LLM is the wrong tool for this job, you cannot bullshit an approximation to an established algorithm.
And using a bullshit approximation layer to translate mathematical algorithm results to words is fraught with the risk of uncaught hallucinations, so adding an LLM would take you further away from the goal, rather than closer.
Re: (Score:3, Informative)
Re: (Score:2)
Pro tip: the LLM isn’t spotting differences in photographs. It is delegating out to external image processing algorithms to do these operations, which are implemented in native libraries and languages, similar to how the LLM itself is implemented.
Incorrect.
The LLM isn’t doing any image operations any more than your web browser or network router is doing image operations.
Laughably incorrect.
I can see I gave you too much credit before.
Re: (Score:3)
Instead, images are processed by a dedicated vision model—such as CLIP or Flamingo—which encodes them into embeddings. These models handle the compute-intensive operations using optimized native kernels. Once the image is converted into an embedding, it is passed to the LLM for reasoning and inference.
LLMs could not directly handle even modest image data—for example, the raw pixel values from a low-resolution image conta
Re: (Score:2)
Nope. As I said, the LLM itself does not perform image processing.
Instead, images are processed by a dedicated vision model—such as CLIP or Flamingo—which encodes them into embeddings.
Correct. Trying to separate those is absurd. The LLM is trained to understand (process) those embeddings. If it were not- it would not be able to make heads or tails of them. The embeddings are merely an encoding of the image, and it is therefor being processed by the LLM.
You said:
It is delegating out to external image processing algorithms to do these operations, which are implemented in native libraries and languages, similar to how the LLM itself is implemented.
A frontend CLIP model is not "implemented in native libraries and langauges". It's simply a different encoder layer than the one that handles the embedding of tokens from the context.
You really are making a habit of trying to mo
Re: (Score:2)
In Clip or Flamingo, the per-pixel compute operations for image processing are performed in native libraries and languages, with the base operations (FMA) executing on hardware.
As I’ve repeatedly stated from the very beginning, LLMs cannot perform numeric processing at any practical scale (millions or billions of elements) due to prohibitve compute cost and memory use. They are many orders of magnitude slower at the toy examples they actually can do (eg a 16x16 matrix multiply).
This has been my po
Re: (Score:2)
In Clip or Flamingo, the per-pixel compute operations for image processing are performed in native libraries and languages, with the base operations (FMA) executing on hardware.
No, they're not. CLIP doesn't care *how* it gets pixel data. Of course some library loaded the image and turned it into a bitmap, and then fed it into the embedding network, but that's not what you meant, and you know it.
Converting a format into a bitmap and then handing that off to an ANN encoder is not "a native library being used for the per-pixel compute operations".
You'll next claim that LLMs don't actually work with text, since it is tokenized and fed into an embedding network before handed off to t
Re: (Score:2)
>While they are trained statistically, there are no statistics involved in their inference until the very end (token sampling)
And yet you also say
> It's a series of math equations done on a large set of vectors.
Which is it? Is it math, or are statistics uninvolved?
You say that distinguishing between the LLM and image operations is laughably incorrect, but then you say:
> LLMs can execute an algorithm just fine.
Which is it? Is the LLM distinct from things it executes as child pr
Re: (Score:2)
It's a series of math equations done on a large set of vectors. I love that you see to think that a set of ReLU parameters trained on several trillion tokens can't do math, lol.
You’re conflating “being implemented using X” with “being able to perform X as a task”.
Yes, LLMs are built using large-scale linear algebra—billions of parameters, tensors, and matrix operations. But that doesn’t mean they can do large-scale linear algebra themselves. In fact, they perform quite poorly at precise numerical computation, even at relatively small scales.
A useful analogy: the human brain runs on incredibly complex biochemical and electrical proces
Re: (Score:2)
Which is it? Is it math, or are statistics uninvolved?
Oh, is that the game we're playing? All math involved on a statistical inference is statistics?
You say that distinguishing between the LLM and image operations is laughably incorrect, but then you say:
Which is it? Is the LLM distinct from things it executes as child processes, or are we lumping all tech that an LLM touches under the collective umbrella of "AI" and nothing gets to be something else?
As I said, you're an idiot.
There's no child process involved.
VLMs and LLM+CLIP do not use "child processes" to do "image operations."
As suspected, you're a fucking idiot.
Re: (Score:2)
You’re conflating “being implemented using X” with “being able to perform X as a task”.
No, I'm not.
I'm laughing at someone who says that "being implemented using X can't do X."
There is absolutely nothing that guarantees an LLM can do any particular kind of math.
There is, however, a state of ignorance that would lead someone to say that it can't.
Your understanding clearly doesn't go as far as "reading".
Re: (Score:2)
You’re conflating “being implemented using X” with “being able to perform X as a task”.
No, I'm not.
Yes. You are.
Re: (Score:2)
Let's go over the sentences that were apparently too difficult for you to read.
Because no amount of statistically significant wordbarf will accurately represent a precise computation.
I love that you see to think that a set of ReLU parameters trained on several trillion tokens can't do math, lol.
Get it?
Are you also stupid enough to make the claim that an LLM can't do math?
Re: (Score:2)
As I’ve said from the very, very beginning:
Are you talking about LLMs that can barely multiply even small matrices? They are language models — terrible at numerical analysis, especially for the vast data sets in this scenario.
An LLM can barely do a 16x16 matrix multiply, or similar amount of processing (eg FFT of 256 element 1D array). A task that would be trivial in native implementation on teraflop-class hardware (now just a commodity GPU) — such as 1024x1024 2D FFT for image analysis — would be impossible to perform on an LLE due to the massive inefficiencies i
Re: (Score:2)
This is ignoring the fact that you were also in the "LLMs can't do math" camp before being shown otherwise.
Re: (Score:2)
Re: (Score:2)
This one in particular is a fucking weird one. I mean, you're in numerical analysis, right? Why ask it for an imperfect encoding, like IEEE754? We use IEEE754 because it's efficient and easy to implement in binary logic. You see, I'm not in numerical analysis. But I am a BSCS, from far enough back that I had to demonstrate that I had learned how to implement an ALU in TTL. Should we ask it to simulate
Re: (Score:2)
Someone proposed feeding the solar system data — which in the context of the article, is presumably the astronomy images and orbital elements.
Not only is an LLM not designed to crunch numbers, it is an Archilles heel of LLM.You claim they can process to arbitrary precision (context window notwithstanding) whereas in fact they can only process a handful of digits. As an exam
Re:A fitting use for AI (Score:5, Insightful)
Re: (Score:2)
That is a very small numerical processing task that requires only one million mused multiply-adds and would normally take only microseconds on teraflop-class hardware with the data in local memory. Negligible compared to the processing required for the astronomy analysis in the article.
I’ll wait.
Re: (Score:2)
Re: (Score:2)
We do not use LLMs to analyse photos.
That was my original point. I simply asked:
Are you talking about LLMs that can barely multiply even small matrices? They are language models — terrible at numerical analysis, especially for the vast data sets in this scenario.
and
LLMs (themselves) cannot even perform a moderate matrix operation reliably. This is because they are language models, and have very poor performance on large numerical tasks.
They CANNOT do the numeric processing required for the astronomy analysis in the article. You and DamnOregonian seem to be having an argument no one else is happening.
And an LLM that can produce a cross product for vectors of length 3, does it just fine for any length of vectors.
Incorrect. LLMs are NOT just fine for any length of vectors: they FAIL on very small amounts of data (eg one million data elements) that would be trivial for conventional numeric processing. Because the overhead makes these tasks is prohibitive.
(I’ll leave as an aside that cross product
Re: (Score:3)
Re: (Score:2)
Re:A fitting use for AI (Score:5, Interesting)
Planet Nine? (Score:5, Insightful)
That is Pluto you ninnies. It was discovered in 1930.
Re: (Score:2)
Re: (Score:3, Interesting)
Yep, and Pluto, on average, is only 40 AU from the Sun. This supposed planet 9 is possible 700AU from the sun. The heliopause is 120 AU and basically considered the edge of our solar system.
So calling something 700AU from the Sun a planet sounds suspicious at best. I'm likely showing my ignorance here but this still seems out there.
Distance from the sun of each planet and Pluto https://phys.org/news/2014-04-... [phys.org]
Re:Planet Nine? (Score:4, Informative)
So calling something 700AU from the Sun a planet sounds suspicious at best. I'm likely showing my ignorance here but this still seems out there.
The main factors appear to be "Is the trajectory of this object governed mostly by the sun's gravity?" and "Is it going to stay on a more or less elliptic trajectory around the sun?", and these factors make a lot more sense than "How close is it to the sun?".
Re: (Score:2)
Re: Planet Nine? (Score:2)
Re: (Score:3, Insightful)
Pluto is a planet. I didn't make the "IAU" my word definition body. And their definition doesn't actually make any kind of sense. Defining what one object is based on the characteristics of other objects is not something that passes any standard of rigor. And you're doing it again with your pluto/moon comparison nonsense.
It orbits a star, it is massive enough to collapse into a sphere under its own gravity, then it's a planet. If it doesn't go spherical, then we can talk about other names for it.
Pluto
Re: (Score:3)
By your own definition, that makes Pluto at least #10.
What's your excuse for Ceres other than they didn't tell you that it was a planet in elementary school?
Re: (Score:2)
Pluto is a planet.
Americans will gain the opportunity to run experiments on Pluto in January 2027, once "The Moose Hunt" enters the public domain. This should help clarify what he is, apart from bloodhound.
Re: (Score:3)
Re: (Score:2)
Mod parent Funny, but I already dibbed "Neo-Pluto".
Re: (Score:2)
I would think Trump would go for renaming Jupiter, the largest planet. Probably call it "America" or some such.
The Earth will become "Planet Trump". Elon or Bezos get to rename Mars; depending on who gets there first.
Re: (Score:2)
I would think Trump would go for renaming Jupiter, the largest planet. Probably call it "America" or some such.
The Earth will become "Planet Trump". Elon or Bezos get to rename Mars; depending on who gets there first.
Planets from the sun outward:
Trump, formerly Mercury, closest to the power of our system.
Acid America, formerly Venus.
Greater America, formerly Earth
Red America, formerly Mars, Making America Red Again, still God of War
Larger America, formerly Jupiter, because imagination is for suckers
Married America, formerly Saturn, because America put the ring on it
Cold America, formerly Uranus, because it looks like ice
Microsoft America, because it looks like the old Windows background color
Re: (Score:2)
This is the 2020s.
Scientific facts are out; nostalgia is in.
We can expect an executive order regarding the status of Pluto in the upcoming weeks.
You mean "Distant America," don't you?
Re: (Score:3)
Please name some of the “planets larger than Pluto” known before 1930 (aside from the 8 other solar system planets) .. Since there are thousands you should be able to name one or two. You know what, I’ll even accept one that is not a moon (you said planet) and confirmed larger than Pluto and within the Solar system.
Re: (Score:2)
AC is obviously exaggerating. Here's a list of 2 and a half dozen, 2 dozen from before 1930 and all reclassified. https://en.wikipedia.org/wiki/... [wikipedia.org] Remember that planet originally meant wandering star, anything that moved relative to the fixed stars was a planet including the Sun and Moon. More recently the moons of Jupiter and Saturn were first called planets, satellite planets actually, and some bigger then Pluto and even bigger then Mercury.
Not planet 9 (Score:5, Informative)
Re: (Score:2)
Noise can be a planet too, as long as it clears its path. One theory is the universe is filled with black noise.
Re: (Score:2)
While it is absolutely NOT Planet 9 (the orbit, while not yet nailed down much, would obviously be well outside the required paths), it could be something else other than just noise.
That could be just as potentially exciting and interesting as finding Planet 9.
Re: (Score:2)
I get what you're saying, but that particular Redditor is renowned, look her up.
Ed Wood almost got it right! (Score:2)
It's just ... (Score:2)
Planet 9 From Outer Space (Score:2)
Worst movie ever. So bad you have to see it.
Oh wait, that was Plan 9...
Re: (Score:2)
Re: (Score:2)
Re: (Score:2)
Now explain to the class what Nibiruuuuuuu is.
Wait. Let me put on a huge tall hat first. Does this hat make me look like a God?
Re: (Score:2)
I had to listen to a nutter on this subject several times because I was living at a former commune-cum-sawmill where he still lived. He kept trying to tell me that it was passing close enough to Earth for its gravitation to affect our orbit, but it still somehow couldn't be detected.
Re: (Score:2)
Re: (Score:2)
Can't, the lizard people won't let me. I only got that one post out before
Re: (Score:2)
Obviously. Now explain to the class what Nibiruuuuuuu is. Wait. Let me put on a huge tall hat first. Does this hat make me look like a God?
I'm not saying it's aliens, but...
... aliens from Nibiru are our ancestors and meddled with early humanity pretty regularly, until technology developed to the point where we could capture actual images of them. Then they disappeared. Coincidence?
No, I'm not serious. But if you watch thirty minutes of History Channel on the wrong night, that's the story you'll get.
Re: (Score:2)
Dont' tease now....I want pix.
Also, I'm leaving this hat on, I think I look like I got a huge effing head, so I look smart.
Re: (Score:2)