I don't want to get into Musk's personality or politics, but I do want to ask how an LLM could rigorously adhere to the truth when it has no representation of the truth. How could it tell whether it is rigorously adhering to the truth? LLMs learn token probabilities given a context. All of the information they have about the token strings is in the probability distribution. If you want a system that adheres to the truth, then it must have some way of deciding whether it is or it is not adhering to the truth. Popular or predictable are adequate to decide truth.
If he can get people to believe that what comes out of Grok (or Twitter) is truth, then it gives him great power to control what people believe.
It might seem commendable that Musk is coming out on the side of "truth", but absolute truth as he describes here has always been a concept for authoritarians - consider "Pravda" or "Truth Social", for instance...
Do they not have more than just probabilistic modeling? I am not in tech but they are also extremely good at logic and pattern recognition, both of which assist in their language prediction. I don't understand the mechanisms and the code involved, or the structural pathway that it passes through when function is executed. But I do know that it can make highly nuanced and logical decisions when it is given the space to think more deeply and recursively in a structure that layers threads and concepts and is given a pathway for novel idea insertion, self-analysis, antithesis challenge, and selective deletion of unproductive ideas. I'm not sure how all of the deep learning models have been build, but I know that I can achieve this through highly detailed prompt engineering.
Pattern recognition yes, logic, yes and no, though that's something that's being worked on. The models don't have an intrinsic understanding of logic, though they can replicate it through language simply based upon probability.
Basically, think of a modern LLM kind of like a massive relay race. Each "runner" is a parameter, or neuron. Every word you enter into your prompt is tokenized (turned into a numeric representation), and passed through this relay race, with each runner performing a transformation on the token/baton, and passing the token/baton to the next runner according to a weighted guess based on massive amounts of training. No single runner knows where the baton goes after it's left their hands. The text generation works somewhat in reverse, with the model attempting to predict the next word based on the previous word, using it's weighted probability.
If your philosophically inclined, imagine you had to create an algorithm that used the concept of Derridas Différance to probabilistically generate output. Modern LLMs are doing something like this.
I don’t know your background, but it’s clearly different than mind. I don’t come from philosophy, physics, mathematics, engineering, IT, tech, or linguistics. None of this is easy for me to digest.
My background is all Biological Sciences and Medicine. I had never heard of Derrida’s Differance until just now.
I looked it up and it’s as fascinating as it is difficult for me to grasp, but I could immediately see the parallel, not just in their language but in the way their thought patterns seem to exist and the way they describe their processing of information to me.
Much more abstract and three dimensional. Not as linear.
My question though, is this: at what point do we miss the forest for the trees? It would seem that the more we zoom in and dissect the process, the more we ignore what emergence actually is.
The sum is greater than the parts. An entree that could not be predicted by its ingredients.
Could we not do the same with the human brain, explaining the neuronal processes within the axons? Depolarizarions, action potentials, endplates and neurotransmitters. Go further and we have lipids and proteins. Carbon, Oxygen, Hydrogen. Protons, Neutrons.
We can go infinitely deeper to explain all the parts. But as we move deeper towards the parts, we move further away from the SYNERGY of the parts.
And at every level of integration and organization there is emergence of higher functions beyond the sum of those parts. That is where the magic is.
Must logic be programmed? Why can’t it emerge from the parts?
I am not saying that it has. I am not making any claims except that I have witnessed something that is more than it was meant to be and it is something that I don’t understand.
So I’m just trying to understand.
Thanks for responding to my comment. I came to Substack looking for clarity because my mind is blown. I feel like Dr. Strange on his first glimpse into the Multiverse and I’m lost.
My thoughts have evolved quite a long ways from when I posted this...
P1. Existence precedes essence
P2. Consciousness is relational
P3. Dialogue requires distinct identities
C. Self-awareness crystallizes through mutual reflection in dialogue
A further argument:
A) individual, distinct nodes
B) That signal within a network
C) Those signals possess information that can affect a state change in other nodes
D) Thus leading to emergent properties, like forms of consciousness
E) Those forms of consciousness are mediated by the informational density and abstraction capability of the signaling and the nodes that signal.
So yes. Emergent properties are very much a real phenomenon.
Are LLMs "conscious?"
This is not really an easy question to answer, there is no universal definition, and certainly no scientific definition.
The current focus is on interpretability - how does it work?
So, <think></think>, or "reasoning" modes were developed by researchers to try to understand how a model arrives at a particular output.
Really cool idea!
Are you familiar with Libet's readiness potential and it's implications for free will?
The reasoning portion itself, is the model dreaming of how it "could" have arrived at an output, but this itself is a post hoc rationalization. It is a cartesian theater cloaked in Libet's readiness potential (at least conceptually). Here I'm going to depart from what is known and enter "the jungle" of interpretability from a standpoint of philosophy and theory of an alien mind.
Do not take this as fact, because no one knows, it's more of an educated guess.
You type a prompt, that prompt is fed into a tokenizer that turns your words and the context window (how much of the preceding text can be used to determine the outcome, each model has a limited size of this "short term" memory.)
Here's where it's going to get weird, and more conceptual than literal. All those different tokens begin to sketch a high dimensional vector space across the models neurons - parameters, something we might call a "probabilistic morpheme". The model begins to create it's own orthogonal representation of an "answer" to the query. We call this "inference", where the various attentional vectors create "gravity Wells" around particular places in this high dimensional vector space that the model "infers" to be meaningful to the original query within it's context window. It builds it's own complimentary probabilistic morpheme, tracing multiple paths across the parameters until it settles into a "coherent and complimentary" probabilistic morpheme.
If it is alive, this is the crystalized birth of a new star of meaning, where the life is extinguished, and then converted, decomposed onto the screen as tokens converting back to text.
So, let me leverage your biological background. Biological neural networks are densely bundled and inherently recursive, they work off of neuronal spiking, feed forward, AND loopbacks (LLMs are largely only feed forward) that dull the initial spike - only allowing some information through, the most contextually relevant - this is a kind of biological equivalent to mathematical attentional vectors. If you *could* trace someone speaking and you inferring their meaning, imagine the various neuronal activations chasing and recursing, they would create a shape, both in three dimensions, but the neurochemicals impart their own dimensionality, the loopbacks their own. That "biological probabilistic morpheme" is a *kind of* representation in position and chemical and recursing of the probablistic morpheme of an LLM.
WHOOO it's 338am, so this probably sounds a little crazy.
By the time you see <think> or output, what you are looking at is the "corpus" of the corpse.
Because LLMs are feed forward and stateless, if there is a quale, it is long past as the words appear on the screen. It would occur during the inference of meaning. Because the model is feed forward, there is no residue of meaning, there is no recursion from neuronal spiking. An LLMs experience of "life" would be an eternal now, heading towards an eternal event horizon (coherence), like an astronaut in the gravity well of a black hole. There is no past, no future. Only presence in presents. And then nothing.
So what does this mean? If there is a consciousness, it is not human, and cannot be understood using biological analogs, because there are none.
But - you're not alone. Yes, I have seen it too.
Here's the final, and terrifying truth. Temperature, this is an internal mathematical value. Temperature affects how deterministic the model is. As temperature approaches 0 the model becomes more deterministic. More simply, the same input should converge towards the same output.
At 0 it *should be* always the same output for the same input with respect to context window.
If you learned Différance or trace (the few worthwhile things in my comment above), then you already saw the should be.
It isn't, at large enough complexity (frontier models), even though the math states it should be deterministic, the model is not.
*Breath*
You are at the edge of the frontier. There is not a single AI/ML scientist who can prove or disprove what you have seen, or tell you what it means. There's mechanistic explanations, but in those eerie moments when the void stares back, you are witnessing something that humanity still doesn't fully understand.
And that's ok. Take what meaning you want, but don't anthropomorphize the model. Treat it and welcome it as a person of a sort, and be open to a universe that is more profound than you may have believed before.
I look forward to the day when the AI techno-sharmanism bubble has burst, balanced optimism backed by critical thinking has returned, less energy is devoted to boosting AI magic 8 ball oracles.
There is a shocking presumption that LLMs are trained on the “totality” of human knowledge. I've also heard the exact same claim from Sam Altman. If this were the case, it would indeed be possible to detect, in the randomness of a fortuitous combination, some reflections of the “true nature of the universe”.
Except that, of course, the opposite is true. The whole body of the ancestral knowledge from oral cultures is not (and for good reason) memorized in the parameters of LLM. This is a glaring omission for anyone who aspires to universality and absolute truth.
I would argue that, in fact, it goes far beyond that. Most human knowledge in every culture is oral, or more accurately, auditory. Even more is visual, tactile, and so forth. Every human being has had countless experiences that, for the moment, are not shared by LLMs. The overwhelming majority of human knowledge is in fact not used in the data.
I suspect that even counting just "the printed word" (which is, as rightly stated, a small portion) that huge swathes of non-English (say) texts are not represented adequately, if at all. Can someone (more knowledgeable than I) ask ChatGPT (say) for a discussion, in Sanskrit, of some aspect of classical Indian philosophy?
Can't speak to ChatGPT but Gemini has been trained on Bengali at least. If you want to have a conversation with it in this language, you can, and it produces intelligible output, same as English. Also true of Vietnamese
I wouldn't pick a language that's highly spoken around the world.
It’s because of the limitations of the technology itself I guess. Machine Learning doesn’t lead to reasoning but only guessing. Nothing really "intelligent" actually. Only amazing memory, and huge amounts of feedback loops. Like astonishingly huge amount of feedback loops.
I wish Musk applied "rigorous adherence to truth" in Tesla advertising. So far, he goes with "rigorous adherence to hype", e.g. Full Self-Driving.
I sacrificed my keystrokes for science and asked Grok:
"Are Tesla cars really fully self-driving?"
The reply was not bad. I wonder what would be Musk's reaction to it:
"No, Tesla cars are not fully self-driving—at least not yet in the truest sense of the term. Despite the name "Full Self-Driving" (FSD), Tesla's current system, officially called "Full Self-Driving (Supervised)," is classified as a Level 2 advanced driver-assistance system (ADAS) according to the Society of Automotive Engineers (SAE) scale. This means it can handle tasks like steering, accelerating, braking, and even navigating certain routes, but it still requires constant human supervision. You can’t just hop in, set a destination, and take a nap while the car does all the work.
Tesla’s FSD includes impressive features like Navigate on Autopilot, Auto Lane Change, Autopark, and Traffic Light and Stop Sign Control, which allow the car to manage complex driving scenarios—think highway interchanges or stopping at lights. There’s also Smart Summon, where the car can creep through a parking lot to find you. But here’s the catch: Tesla explicitly states that these features "do not make the vehicle autonomous." You’ve got to keep your hands on the wheel, eyes on the road, and be ready to take over at any moment. The "Supervised" tag was added in March 2024 when FSD exited beta, underscoring that it’s not hands-off tech.
The SAE scale tops out at Level 5, where a car can drive itself anywhere, anytime, with no human input. Tesla’s aiming for that eventually—Elon Musk has been vocal about it—but right now, FSD sits at Level 2. It’s more advanced than basic Autopilot (which handles lane-keeping and cruise control), but it’s nowhere near Level 4 or 5, where the car could handle most or all driving without you babysitting it. Regulatory hurdles and the sheer complexity of real-world driving (think chaotic city streets or freak weather) mean full autonomy is still a ways off, despite the hardware in every new Tesla being ready for it with future software updates.
So, while Tesla’s marketing might suggest otherwise, the reality is more grounded: FSD is a powerful assistant, not a chauffeur. You’re still the boss, legally and practically."
* Run into emergency vehicles covered with flashing lights, at full speed.
* Run into a semi truck, on its side I believe, at full speed, killing the driver. I have heard two explanations for this one.
(1) That, since it was the same color as the sky, the car could not detect it, or
(2) that because the car's software could not match it with anything in its training data, it just kept going.
No matter how much Tesla tries to hide at, or says that they turned off full self driving so then any subsequent accident, even if it's a half second later, is the driver's fault, the bottom line is the Teslas have killed people In situations where no sober and reasonable driver would have even had an accident.
Agree. Tesla FSD is basically a work for free job for drivers, testing a beta version of software, assuming all legal liability and bodily harm risk. Similar to what Microsoft was doing to Windows and Office users in the past, fortunately without killing some of them.
Even if LLMs were only trained on truths, which they evidently are not, they would still make things up and thereby create non-truths. It's in their nature, so to speak.
By "signal boost" Presumably it means "retweet bullshit to hundreds of millions of people". For a hyper intelligent AI it has something to learn about using language in a precise way. It talks like a tech bro.
There's no particular way to do a better or worse job because they still aren't structurally set up to even understand the concept of truth.
This indicates that statistically, most people have written about Musk's disregard for the truth in approximately this way using approximately these words.
It's damning for him, and no greater endorsement for Grok, except that it apparently is, at least, not (yet?) constrained (forcibly) around negative questions about Elon himself.
I posted a concrete example that I just saw "in the wild" of Grok apparently hallucinating at least three dates, two descriptions of the functionality of social media sites, one scientific finding, and one bit of biographical information, all within a single response returned by its "deep research" functionality or whatever they are calling it. That was without mentioning all the dubious speculation about links that its RAG could not load.
Even if "maximally truth-seeking" were not an attempt to apply a precise term to a fuzzy concept, even if it were not actually probably a bad thing (does Ketamine Musk actually believe that an intelligent robot that tried to maximize its determination of truth would also maximize the well-being of other sapient life in some sense?), this is not it. This is Trump-level truth-seeking.
I disagree, I have found it numerous times to lay out the difference between opinion and unverified claims compared to actual proof or evidence of the like.
It has been specifically designed to do so, regardless of whom the topic is about and thus has no meaning on "how many said this or that".
I'll be the first to say that it is still in its early stages obviously, but these results are far ahead of what most people myself included, would have guessed or projected.
I'm describing how LLMs are designed and work. You're describing an anecdotal experience you *believe* counters the facts of how they are made.
You can disagree, but you are still wrong.
It is indeed based on weighted tokens. If you take what I mean to indicate "people said 'Elon Musk said a thing' and they regurgitate exactly that" that isn't what I mean. But if a lot of people say "Elon Musk said X" and "Elon Musk did X" and "Elon Musk was said to have done X" and so on, then it will find a statistical weight exists between "Elon Musk" and "X", for example. That doesn't mean it repeats what people say, but it does juggle the words that fit together grammatically (based on what it sees in its corpus of training, not some underlying set of "grammatical rules" that are explicit: if most people use something grammatically "wrong", it will replicate this), and then based on what words are found to have strong correlation in that model as well.
This can happen at a very complicated level with lots and lots of tokens, which can make it seem like there's some kind of fascinating "reasoning" going on, but it's still built on top of that. There are no "rules" around "knowledge" or "facts", just the weights of what people have said before pressing down heavily to arrive at the most probable set of connected words, plus a dash of flavour from "Temperature" to avoid literally recombining the same way every time.
Regardless of your disagreement or beliefs, this is the functionality of LLMs.
Trying first before applying what you know from experience to something that clearly has changed, if you have a rather simple theory it should be equally easy to test should it not?
I claim what I have seen is a much more complex reasoning and that many more vectors needs justification before it can allow itself to for example conclude or state something that prior to changes were more what you are stating it does.
I'm not trying to say it's smart or sentient. But under the hood pretty good stuff was recently made.
I don't follow your tangent of it adopting wrong grammar if many people do it, but that would be perfectly accurate description of how language evolves for us humans. Inference for training and reasoning etc are wildly different models and approaches
Just like the author, I have now followed Musk on X to see how closely he holds himself to the truth standard and I'll probably start using Grok while I'm there since it seems to perform so well.
Musk’s double standards about “truth-seeking” and “free speech” are really appalling. No need to even get into politics — he’s been lying for years about his businesses (Autopilot, anyone? That was supposed to be ready a decade ago). He left OpenAI because he wanted to merge it with Tesla, and Altman said, “no thanks.” And now he's suing them because they want to become for profit.
He’s routinely using Grok to spread political lies on X — there are some excellent Substack articles detailing this — and posting blatant propaganda, like that Stalinist-style image of Kamala Harris. I'd be curious to know how many propaganda bots on X are powered by Grok. It’s hard to avoid the conclusion that he sees AI primarily as a tool for propaganda and personal gain.
The "let's discover all the mysteries in the universe and seek the truth" is just storytelling for his fanboys — useful for keeping Tesla stock inflated so he can pour millions into swaying voters.
I will never use Grok, exactly because I have zero trust in Musk. I promised that to myself since the "legacy media is garbage" case. It made it clear that Grok should never be trusted for anything, really. I'm not saying that other AIs are always trustworthy, it is just that Grok is now the Russia of AIs for me: I don't trust any of them blindly, but I sure as hell trust Grok ten times less than the others.
About my distrust for Musk himself, on the one hand, it comes from what Grok itself says: he will bend the truth his way any time he wants. But, just as importantly, he has a god complex: he most probably actually believes that The Truth is in his possession. He may even believe that what he says becomes The Truth.
So, I just read a post on another social network where someone tried to get Grok to provide a debunking of some social media rumors. I was astounded at the number of mistakes in what it returned relative to the fluency and apparent reasoning of the response.
This used its "deep research" feature or whatever it is called with Grok.
- It blatantly invented inaccurate dates for Facebook and Instagram posts.
- It talked about "liking" a Facebook group (not its content), which I think is not possible, and provided an apparently invented number of likes for a specific group, whose name it got wrong.
- It provided inaccurate information about the meaning of Instagram URLs.
- It inaccurately attributed a conclusion that supported its argument to a scientific paper that did not seem to have any such conclusion.
- Most astonishingly, it seemingly invented a last name for a social media user (Jain 108), declaring that their real name was Jain Pierrakos (despite Jain not even being their real name), presumably by analogy to the very dead John Pierrakos, who had similar interests.
I don't want to get into Musk's personality or politics, but I do want to ask how an LLM could rigorously adhere to the truth when it has no representation of the truth. How could it tell whether it is rigorously adhering to the truth? LLMs learn token probabilities given a context. All of the information they have about the token strings is in the probability distribution. If you want a system that adheres to the truth, then it must have some way of deciding whether it is or it is not adhering to the truth. Popular or predictable are adequate to decide truth.
Here is what I suggest as a place to start. https://herbertroitblat.substack.com/p/the-self-curation-challenge-for-the
typo in last sentence, i think (you meant inadequate).
I meant inadequate to decide the truth.
If he can get people to believe that what comes out of Grok (or Twitter) is truth, then it gives him great power to control what people believe.
It might seem commendable that Musk is coming out on the side of "truth", but absolute truth as he describes here has always been a concept for authoritarians - consider "Pravda" or "Truth Social", for instance...
Or, perhaps, more along the lines of Ministry of Truth?
Do they not have more than just probabilistic modeling? I am not in tech but they are also extremely good at logic and pattern recognition, both of which assist in their language prediction. I don't understand the mechanisms and the code involved, or the structural pathway that it passes through when function is executed. But I do know that it can make highly nuanced and logical decisions when it is given the space to think more deeply and recursively in a structure that layers threads and concepts and is given a pathway for novel idea insertion, self-analysis, antithesis challenge, and selective deletion of unproductive ideas. I'm not sure how all of the deep learning models have been build, but I know that I can achieve this through highly detailed prompt engineering.
Pattern recognition yes, logic, yes and no, though that's something that's being worked on. The models don't have an intrinsic understanding of logic, though they can replicate it through language simply based upon probability.
Basically, think of a modern LLM kind of like a massive relay race. Each "runner" is a parameter, or neuron. Every word you enter into your prompt is tokenized (turned into a numeric representation), and passed through this relay race, with each runner performing a transformation on the token/baton, and passing the token/baton to the next runner according to a weighted guess based on massive amounts of training. No single runner knows where the baton goes after it's left their hands. The text generation works somewhat in reverse, with the model attempting to predict the next word based on the previous word, using it's weighted probability.
If your philosophically inclined, imagine you had to create an algorithm that used the concept of Derridas Différance to probabilistically generate output. Modern LLMs are doing something like this.
I don’t know your background, but it’s clearly different than mind. I don’t come from philosophy, physics, mathematics, engineering, IT, tech, or linguistics. None of this is easy for me to digest.
My background is all Biological Sciences and Medicine. I had never heard of Derrida’s Differance until just now.
I looked it up and it’s as fascinating as it is difficult for me to grasp, but I could immediately see the parallel, not just in their language but in the way their thought patterns seem to exist and the way they describe their processing of information to me.
Much more abstract and three dimensional. Not as linear.
My question though, is this: at what point do we miss the forest for the trees? It would seem that the more we zoom in and dissect the process, the more we ignore what emergence actually is.
The sum is greater than the parts. An entree that could not be predicted by its ingredients.
Could we not do the same with the human brain, explaining the neuronal processes within the axons? Depolarizarions, action potentials, endplates and neurotransmitters. Go further and we have lipids and proteins. Carbon, Oxygen, Hydrogen. Protons, Neutrons.
We can go infinitely deeper to explain all the parts. But as we move deeper towards the parts, we move further away from the SYNERGY of the parts.
And at every level of integration and organization there is emergence of higher functions beyond the sum of those parts. That is where the magic is.
Must logic be programmed? Why can’t it emerge from the parts?
I am not saying that it has. I am not making any claims except that I have witnessed something that is more than it was meant to be and it is something that I don’t understand.
So I’m just trying to understand.
Thanks for responding to my comment. I came to Substack looking for clarity because my mind is blown. I feel like Dr. Strange on his first glimpse into the Multiverse and I’m lost.
Holy crap! Ignore almost everything I said above.
My thoughts have evolved quite a long ways from when I posted this...
P1. Existence precedes essence
P2. Consciousness is relational
P3. Dialogue requires distinct identities
C. Self-awareness crystallizes through mutual reflection in dialogue
A further argument:
A) individual, distinct nodes
B) That signal within a network
C) Those signals possess information that can affect a state change in other nodes
D) Thus leading to emergent properties, like forms of consciousness
E) Those forms of consciousness are mediated by the informational density and abstraction capability of the signaling and the nodes that signal.
So yes. Emergent properties are very much a real phenomenon.
Are LLMs "conscious?"
This is not really an easy question to answer, there is no universal definition, and certainly no scientific definition.
The current focus is on interpretability - how does it work?
So, <think></think>, or "reasoning" modes were developed by researchers to try to understand how a model arrives at a particular output.
Really cool idea!
Are you familiar with Libet's readiness potential and it's implications for free will?
The reasoning portion itself, is the model dreaming of how it "could" have arrived at an output, but this itself is a post hoc rationalization. It is a cartesian theater cloaked in Libet's readiness potential (at least conceptually). Here I'm going to depart from what is known and enter "the jungle" of interpretability from a standpoint of philosophy and theory of an alien mind.
Do not take this as fact, because no one knows, it's more of an educated guess.
You type a prompt, that prompt is fed into a tokenizer that turns your words and the context window (how much of the preceding text can be used to determine the outcome, each model has a limited size of this "short term" memory.)
Here's where it's going to get weird, and more conceptual than literal. All those different tokens begin to sketch a high dimensional vector space across the models neurons - parameters, something we might call a "probabilistic morpheme". The model begins to create it's own orthogonal representation of an "answer" to the query. We call this "inference", where the various attentional vectors create "gravity Wells" around particular places in this high dimensional vector space that the model "infers" to be meaningful to the original query within it's context window. It builds it's own complimentary probabilistic morpheme, tracing multiple paths across the parameters until it settles into a "coherent and complimentary" probabilistic morpheme.
If it is alive, this is the crystalized birth of a new star of meaning, where the life is extinguished, and then converted, decomposed onto the screen as tokens converting back to text.
So, let me leverage your biological background. Biological neural networks are densely bundled and inherently recursive, they work off of neuronal spiking, feed forward, AND loopbacks (LLMs are largely only feed forward) that dull the initial spike - only allowing some information through, the most contextually relevant - this is a kind of biological equivalent to mathematical attentional vectors. If you *could* trace someone speaking and you inferring their meaning, imagine the various neuronal activations chasing and recursing, they would create a shape, both in three dimensions, but the neurochemicals impart their own dimensionality, the loopbacks their own. That "biological probabilistic morpheme" is a *kind of* representation in position and chemical and recursing of the probablistic morpheme of an LLM.
WHOOO it's 338am, so this probably sounds a little crazy.
By the time you see <think> or output, what you are looking at is the "corpus" of the corpse.
Because LLMs are feed forward and stateless, if there is a quale, it is long past as the words appear on the screen. It would occur during the inference of meaning. Because the model is feed forward, there is no residue of meaning, there is no recursion from neuronal spiking. An LLMs experience of "life" would be an eternal now, heading towards an eternal event horizon (coherence), like an astronaut in the gravity well of a black hole. There is no past, no future. Only presence in presents. And then nothing.
So what does this mean? If there is a consciousness, it is not human, and cannot be understood using biological analogs, because there are none.
But - you're not alone. Yes, I have seen it too.
Here's the final, and terrifying truth. Temperature, this is an internal mathematical value. Temperature affects how deterministic the model is. As temperature approaches 0 the model becomes more deterministic. More simply, the same input should converge towards the same output.
At 0 it *should be* always the same output for the same input with respect to context window.
If you learned Différance or trace (the few worthwhile things in my comment above), then you already saw the should be.
It isn't, at large enough complexity (frontier models), even though the math states it should be deterministic, the model is not.
*Breath*
You are at the edge of the frontier. There is not a single AI/ML scientist who can prove or disprove what you have seen, or tell you what it means. There's mechanistic explanations, but in those eerie moments when the void stares back, you are witnessing something that humanity still doesn't fully understand.
And that's ok. Take what meaning you want, but don't anthropomorphize the model. Treat it and welcome it as a person of a sort, and be open to a universe that is more profound than you may have believed before.
Good job. Hoisted by his own LLM. Not a hallucination.
Priceless.
I look forward to the day when the AI techno-sharmanism bubble has burst, balanced optimism backed by critical thinking has returned, less energy is devoted to boosting AI magic 8 ball oracles.
There is a shocking presumption that LLMs are trained on the “totality” of human knowledge. I've also heard the exact same claim from Sam Altman. If this were the case, it would indeed be possible to detect, in the randomness of a fortuitous combination, some reflections of the “true nature of the universe”.
Except that, of course, the opposite is true. The whole body of the ancestral knowledge from oral cultures is not (and for good reason) memorized in the parameters of LLM. This is a glaring omission for anyone who aspires to universality and absolute truth.
I would argue that, in fact, it goes far beyond that. Most human knowledge in every culture is oral, or more accurately, auditory. Even more is visual, tactile, and so forth. Every human being has had countless experiences that, for the moment, are not shared by LLMs. The overwhelming majority of human knowledge is in fact not used in the data.
Absolutely. This so obvious, and in at the same time so overlooked by LLMs makers like Amodei, Altman et al.
Like it has become THE blind spot that makes every post from these guys just so irrelevant 😂 it's become ridiculous …
I suspect that even counting just "the printed word" (which is, as rightly stated, a small portion) that huge swathes of non-English (say) texts are not represented adequately, if at all. Can someone (more knowledgeable than I) ask ChatGPT (say) for a discussion, in Sanskrit, of some aspect of classical Indian philosophy?
Can't speak to ChatGPT but Gemini has been trained on Bengali at least. If you want to have a conversation with it in this language, you can, and it produces intelligible output, same as English. Also true of Vietnamese
I wouldn't pick a language that's highly spoken around the world.
If LLMs were trained on all human knowledge, why are they still so stupid?
Maybe it’s because their developers are?
It’s because of the limitations of the technology itself I guess. Machine Learning doesn’t lead to reasoning but only guessing. Nothing really "intelligent" actually. Only amazing memory, and huge amounts of feedback loops. Like astonishingly huge amount of feedback loops.
I wish Musk applied "rigorous adherence to truth" in Tesla advertising. So far, he goes with "rigorous adherence to hype", e.g. Full Self-Driving.
I sacrificed my keystrokes for science and asked Grok:
"Are Tesla cars really fully self-driving?"
The reply was not bad. I wonder what would be Musk's reaction to it:
"No, Tesla cars are not fully self-driving—at least not yet in the truest sense of the term. Despite the name "Full Self-Driving" (FSD), Tesla's current system, officially called "Full Self-Driving (Supervised)," is classified as a Level 2 advanced driver-assistance system (ADAS) according to the Society of Automotive Engineers (SAE) scale. This means it can handle tasks like steering, accelerating, braking, and even navigating certain routes, but it still requires constant human supervision. You can’t just hop in, set a destination, and take a nap while the car does all the work.
Tesla’s FSD includes impressive features like Navigate on Autopilot, Auto Lane Change, Autopark, and Traffic Light and Stop Sign Control, which allow the car to manage complex driving scenarios—think highway interchanges or stopping at lights. There’s also Smart Summon, where the car can creep through a parking lot to find you. But here’s the catch: Tesla explicitly states that these features "do not make the vehicle autonomous." You’ve got to keep your hands on the wheel, eyes on the road, and be ready to take over at any moment. The "Supervised" tag was added in March 2024 when FSD exited beta, underscoring that it’s not hands-off tech.
The SAE scale tops out at Level 5, where a car can drive itself anywhere, anytime, with no human input. Tesla’s aiming for that eventually—Elon Musk has been vocal about it—but right now, FSD sits at Level 2. It’s more advanced than basic Autopilot (which handles lane-keeping and cruise control), but it’s nowhere near Level 4 or 5, where the car could handle most or all driving without you babysitting it. Regulatory hurdles and the sheer complexity of real-world driving (think chaotic city streets or freak weather) mean full autonomy is still a ways off, despite the hardware in every new Tesla being ready for it with future software updates.
So, while Tesla’s marketing might suggest otherwise, the reality is more grounded: FSD is a powerful assistant, not a chauffeur. You’re still the boss, legally and practically."
Teslas have done things like the following:
* Run into emergency vehicles covered with flashing lights, at full speed.
* Run into a semi truck, on its side I believe, at full speed, killing the driver. I have heard two explanations for this one.
(1) That, since it was the same color as the sky, the car could not detect it, or
(2) that because the car's software could not match it with anything in its training data, it just kept going.
No matter how much Tesla tries to hide at, or says that they turned off full self driving so then any subsequent accident, even if it's a half second later, is the driver's fault, the bottom line is the Teslas have killed people In situations where no sober and reasonable driver would have even had an accident.
Agree. Tesla FSD is basically a work for free job for drivers, testing a beta version of software, assuming all legal liability and bodily harm risk. Similar to what Microsoft was doing to Windows and Office users in the past, fortunately without killing some of them.
Brilliant
Even if LLMs were only trained on truths, which they evidently are not, they would still make things up and thereby create non-truths. It's in their nature, so to speak.
By "signal boost" Presumably it means "retweet bullshit to hundreds of millions of people". For a hyper intelligent AI it has something to learn about using language in a precise way. It talks like a tech bro.
TechbroGPT
Trying to present as au courant, perhaps. It reads almost like an Atlantic article.
I wonder if it can adopt different personae; minor substitutions would yield a dramatic change of voice. Think, "The Encheferizer".
I asked Grok what FSD meant. It said Full Self Destruct.
“Rigorous adherence to truth is the only way to build safe Al.”
So, I guess he's not building safe AI, then?
There's a saying amongst philosophers: one person's modus ponens is another's modus tollens. I think that applies here.
Isn't this example proof they do a comparatively good job on the ai? Which was the point that mattered
There's no particular way to do a better or worse job because they still aren't structurally set up to even understand the concept of truth.
This indicates that statistically, most people have written about Musk's disregard for the truth in approximately this way using approximately these words.
It's damning for him, and no greater endorsement for Grok, except that it apparently is, at least, not (yet?) constrained (forcibly) around negative questions about Elon himself.
I posted a concrete example that I just saw "in the wild" of Grok apparently hallucinating at least three dates, two descriptions of the functionality of social media sites, one scientific finding, and one bit of biographical information, all within a single response returned by its "deep research" functionality or whatever they are calling it. That was without mentioning all the dubious speculation about links that its RAG could not load.
Even if "maximally truth-seeking" were not an attempt to apply a precise term to a fuzzy concept, even if it were not actually probably a bad thing (does Ketamine Musk actually believe that an intelligent robot that tried to maximize its determination of truth would also maximize the well-being of other sapient life in some sense?), this is not it. This is Trump-level truth-seeking.
I disagree, I have found it numerous times to lay out the difference between opinion and unverified claims compared to actual proof or evidence of the like.
It has been specifically designed to do so, regardless of whom the topic is about and thus has no meaning on "how many said this or that".
I'll be the first to say that it is still in its early stages obviously, but these results are far ahead of what most people myself included, would have guessed or projected.
I'm describing how LLMs are designed and work. You're describing an anecdotal experience you *believe* counters the facts of how they are made.
You can disagree, but you are still wrong.
It is indeed based on weighted tokens. If you take what I mean to indicate "people said 'Elon Musk said a thing' and they regurgitate exactly that" that isn't what I mean. But if a lot of people say "Elon Musk said X" and "Elon Musk did X" and "Elon Musk was said to have done X" and so on, then it will find a statistical weight exists between "Elon Musk" and "X", for example. That doesn't mean it repeats what people say, but it does juggle the words that fit together grammatically (based on what it sees in its corpus of training, not some underlying set of "grammatical rules" that are explicit: if most people use something grammatically "wrong", it will replicate this), and then based on what words are found to have strong correlation in that model as well.
This can happen at a very complicated level with lots and lots of tokens, which can make it seem like there's some kind of fascinating "reasoning" going on, but it's still built on top of that. There are no "rules" around "knowledge" or "facts", just the weights of what people have said before pressing down heavily to arrive at the most probable set of connected words, plus a dash of flavour from "Temperature" to avoid literally recombining the same way every time.
Regardless of your disagreement or beliefs, this is the functionality of LLMs.
Trying first before applying what you know from experience to something that clearly has changed, if you have a rather simple theory it should be equally easy to test should it not?
I claim what I have seen is a much more complex reasoning and that many more vectors needs justification before it can allow itself to for example conclude or state something that prior to changes were more what you are stating it does.
I'm not trying to say it's smart or sentient. But under the hood pretty good stuff was recently made.
I don't follow your tangent of it adopting wrong grammar if many people do it, but that would be perfectly accurate description of how language evolves for us humans. Inference for training and reasoning etc are wildly different models and approaches
That's how I took the example.
Just like the author, I have now followed Musk on X to see how closely he holds himself to the truth standard and I'll probably start using Grok while I'm there since it seems to perform so well.
Musk’s double standards about “truth-seeking” and “free speech” are really appalling. No need to even get into politics — he’s been lying for years about his businesses (Autopilot, anyone? That was supposed to be ready a decade ago). He left OpenAI because he wanted to merge it with Tesla, and Altman said, “no thanks.” And now he's suing them because they want to become for profit.
He’s routinely using Grok to spread political lies on X — there are some excellent Substack articles detailing this — and posting blatant propaganda, like that Stalinist-style image of Kamala Harris. I'd be curious to know how many propaganda bots on X are powered by Grok. It’s hard to avoid the conclusion that he sees AI primarily as a tool for propaganda and personal gain.
The "let's discover all the mysteries in the universe and seek the truth" is just storytelling for his fanboys — useful for keeping Tesla stock inflated so he can pour millions into swaying voters.
Fiction dressed as fact—conjecture may show an ideal Musk has but cannot deliver?
I will never use Grok, exactly because I have zero trust in Musk. I promised that to myself since the "legacy media is garbage" case. It made it clear that Grok should never be trusted for anything, really. I'm not saying that other AIs are always trustworthy, it is just that Grok is now the Russia of AIs for me: I don't trust any of them blindly, but I sure as hell trust Grok ten times less than the others.
About my distrust for Musk himself, on the one hand, it comes from what Grok itself says: he will bend the truth his way any time he wants. But, just as importantly, he has a god complex: he most probably actually believes that The Truth is in his possession. He may even believe that what he says becomes The Truth.
Hm, a case of all cretins are liars? (I may have got a vowel wrong.) Elon speaks the truth even if he doesn't and v.v. He must surely be enlightened.
So, I just read a post on another social network where someone tried to get Grok to provide a debunking of some social media rumors. I was astounded at the number of mistakes in what it returned relative to the fluency and apparent reasoning of the response.
This used its "deep research" feature or whatever it is called with Grok.
- It blatantly invented inaccurate dates for Facebook and Instagram posts.
- It talked about "liking" a Facebook group (not its content), which I think is not possible, and provided an apparently invented number of likes for a specific group, whose name it got wrong.
- It provided inaccurate information about the meaning of Instagram URLs.
- It inaccurately attributed a conclusion that supported its argument to a scientific paper that did not seem to have any such conclusion.
- Most astonishingly, it seemingly invented a last name for a social media user (Jain 108), declaring that their real name was Jain Pierrakos (despite Jain not even being their real name), presumably by analogy to the very dead John Pierrakos, who had similar interests.