AI: Intelligence is Not the Key

Key Takeaways

Hello community. So great that you are back. We have some brand new AI research regarding reasoning and knowledge in our transform architecture. And you know this we have been there. No you ask yourself hey did the LLM just fail because it lacked the necessary knowledge the reasoning capability or did the LLM fail because it simply failed to reason correctly with the parametric knowledge it possesses and I just have to apply here a graph rag optimization. Now to come up with a solution we have a brand new research and they said you know what we do a simple experiment we do a fast thinking where we say the LLM hey your output is exactly only one letter there's no reasoning there's nothing a question or additional text or punctuation just one character you want to have the direct access here to the LLM response based only if you want on the parametric knowledge of the model so if you want this AI has now a gut reaction based on its most immediate high probable patterns stored inside its tens of weight structures here of the layers of the transformer and we get an immediately accuracy of this model let's call it a for accuracy the fasting model and this becomes our direct quantitive measures of the knowledge retrieval capability now of course we have to go then with the chain of sword deep reasoning process now so we do have here a prompt for a slow thinking you can use this of course And what we get something of course we have to evaluate this. So we use another LLM a GLM4 plus to evaluate this here. Great. We also get an accuracy here. Let's call it a slow for slow thinking that represents the final off outcome after both the knowledge retrieval and the deep reasoning have occurred. Now the authors of today's study said that okay let's do this with different models. We want to see different models. So when they write Q1 it's a Q1 2.5. If they go with QW32B, it is a 3B 32B preview with Alama, it's a three. Jamma is still a two. Fi is already a four. So great, we have the classical models that we have and we go with the classical data sets. So we have a lot of data. So no problem. We have enough quantity for our analysis. And you said, let's look at just at first at the token consumption of our LLMs. And you see at a token consumption of a Q and 1.5B to a 32B independent if it's correct correct answer and incorrect answer it is actually surprisingly here within here very close interval of tokens but other models like if you move from a Q to a QW you see a wow it can suddenly double very interesting I al for example sorted a llama 1B or a llama 70B be for to deduct here the correct or incorrect answer the amount of token consumption would be different turns out no look it's almost the same and then he said what about domain specific knowledge and here on A you have mathematics and here on B you have medical interesting look at this the result demonstrate clearly that mathematics here our A imposes a higher demand on reasoning and you see here the slow reasoning here in orange than the reasoning here on medicine. Now this is interesting. So there is something happening. Now there is now a difference we have to examine because why is it that if we have the fast thinking compared to the slow reasoning here of the same model. This is just here the model size increases here on the x-axis. But you see in general we have an identical trend that there's a vast amount of difference between system one or system two thinking completely different in medicine. Why is this? So let's have a closer look. Now they also noticed hey we just want to tell you there's what we call an anchoring effect. Now we know this anchoring effect here from humans. But you know what funny there's funny fact that is also happening here with AI systems. So if you have a question let's say we go with a multiple choice and you know A is the correct answer and then we just disturb here a little bit the I know or maybe you just bring it in from Ragna and there's another piece of information and gives you here an idea. The correct answer seems to be B but I don't really know. So you know you have kind of a predisposition here information brought in in context learning rack systems and now this anchoring effect is it also happening here with our latest AR model and look at this this is unbelievable so again we have a model size from 1.5 billion free trainable parameter to 32 billion and what we have we have here anchor without an anchor so without Here the correct answer seems to be B and then we have the difference here in the performance. So let's have a look here at the performance data. You remember this is the fast um if you want syncing and the very slow deductive syncing here in the chain of sort. So you see if we have given an anchor to the LLM the performance is not really famous. uh eight points but without an anchor. So without this here the system thinks oh no it has a little bit of inherent knowledge and completely has a different performance of 39 which is great. Now if you take the difference you see anchor minus without anchor gives you minus 31. Now look here at the if you want system two this low sinking model if you provide it with an anchor. Okay, you have an effect if you compare it to without an anchor. So you only achieve here a performance of 40 compared to 62 without an anchor. So in general if you compare here now the slow thinking to the fast give me the answer immediately there is a significant performance difference but it also shows you here if you just take here the difference that here the fast system is much more sensitive to anchors and if you increase this you see interesting so the effect on fast stays more or less the same no minus 26 at a 32B B compared to minus 31 and 1.5B. But look if you go here with the slow the effect of anchoring is much less massive here because the system has if you want a causal reasoning regime that it tries to follow and come to a conclusion. Interesting. So the anchor minus or without anchor indicates the the accuracy decreases due to anchoring. So we have a massive bias and then performance decline in slow syncing is generally less. You see here the blue line here the blue numbers here in general less sever indicating that reasoning adjustment effectively mitigates here the anchoring effect. So we have to take care about this. Great to know but you know this difference here this is the delta we introduced a new parameter almost forgot to introduce it to you. So fast and slow and you see the difference the delta is exactly 32 and you get the idea. Great. Now what they did they said hey this delta is interesting but you know this delta here and if we do it now here for a lot of different models we can have a look at this but there's also a deeper symmetry within the delta. There are other components in this delta if we have a mathematical look at the factor delta. So delta is not a monolytic quantity. It is kind of a net result of a tugof-war between two competing cognitive forces within your LLM. We have a delta C and a delta O, a delta correction and a delta overinking. And then we have a normalized rate. So let's have a look here at the mathematical formulas just for a second. So the delta without going into any details has here a delta c a delta correction and a second term a delta oversyncing. So it tells us here hey having a look here at the mathematical background it shows us here to the reasoning adjustment we have two effects. So the llm corrects the error made by the fast thinking but it also incorrectly overrides here the correct answer it found. So we have a positive effect and we have a negative effect. And now the question is what is the net effect in total? And they said yeah you know we have also here to look here if you want at the rate of correction and the rate of oversyncing if we normalize it because we have to want to have here a real precise parameter. So again what do we have? We have a delta C and a delta O. The delta C is if you want the correction gain. This is the value generated here by the reasoning of the LLM. This is the fraction of all the questions that were initially wrong but were corrected now by the reasoning process of the LLM. And this is great. So we improve our performance. But at the same time the found an oversyncing loss gain and loss. So this cost if you want a course now by also by reasoning but this is not a fraction of all the question that were initially answered correct by the LLM but were later on corrupted into an incorrect answer by a nonoptimal reasoning process. The LLM just decided okay I don't know if this is a correct answer so I reason about it and it turns out it reasons to an incorrect answer. So the net reasoning gain of these both effects is if you want profit minus loss is our delta. So you see delta is not monolytic but has components. Now what is the rate this correction rate I can call it or I want to call it here reasoning intelligence because given that the mall's initial intuition was wrong. You ask now what is the probability that this LLM reasoning will now fix this initial mistake that intuition based only on the parametric knowledge base and here we have your RC so the correction rate here the probability is now a simple mathematical formula and also we do the same for the overinking rate now let's call it the reasoning recklessness so given that the model's initial institution intuition was correct you Now, hey, what is now the probability it will continue to over sync this and therefore break the reasoning process and arrive at an incorrect answer. And we get here also a simple formula for this. And now the interesting fact is those rates are important factors that tell us a lot of the behavior of the model because this decomposition that we have of our delta into the rate and into the two factors is now incredibly revealing. if you read this new paper because now it explains why small models often fail at the reasoning and I thought hey it fails of the reasoning because those are model rather small 1.5 billion model now it turns out that there is a very particular effect if you have this decomposition of delta it means that there are high rates here of let's call it recklessness swarms their intelligence leading here to a net negative delta. On the other hand, for the large language malls that achieve here a superior reasoning performance in general, this is not necessarily by being dramatically more intelligent quotation mark as an LLM, but there's an effect taking place. But by becoming vastly more cautious, more prudent, their specific rate on oversyncing drops abruptly with the scaling. This is now a new insight. This is now something looking here all all the data already showed you. This is something interesting that they found. Let me give you an example. This is my example. This is not official. This is just my internal thinking. Maybe it helps you. Maybe I made mistake. place the small model an 8B model let's say our accuracy for the fast reasoning or fast thinking the mode one or syncing one mode is 60%. So it knows the answer 60% of the time. Now the intelligence rate if you want no it has a decent correction rate let's say 40%. So it can fix 40% of its initial mistake that it made during here the fast thinking. Now talking about prudence here it's not very prudent. Yeah it has a high overs syncing rate as a small model let's say 17%. Go withever model you like. So it gets tempted to change its answer and 70% of the time it correctly knew the answer but it ends up breaking it through here an incorrect flawed reasoning process. So we have data 60% 40% 70%. So let's do this now on a real world example. Let's calculate the reasoning gain delta on a 100 question test set. So what are the gains from the intelligence that we have now? So on 100 question it's wrong about 40 questions. So it fixes 40% of them. So we have 16 correct answer with the if you want additional gain here from the intelligence but we also have I told you a counter effect this losses from the intruder. No. So it was right here on 60 questions. No 60 40 100 but it over syncs and breaks 17%. So this means we have less than 10 correct answers. Now what is the net reasoning gain that we achieved? Yeah, not going to believe it. 16 minus 10 is exactly 5.8. So reasoning is helpful here but its benefit is significantly blunted by its tendency to self-sabotage self correct itself but in an incorrect way. So this is now interesting that we have now this specific data on the llama's remodel. So it is really interesting to know here the rates the gains from the intelligence and the losses here from the incorrect corrections. Now the situation is different at the large model. Let's say we have a 70 billion free trainable parameter model and let's say here the accuracy of the fast thinking here the system one thinking is 81%. It's really good. Now let's have a look here again at the intelligence. No, according here to the paper's data, its intelligent is not dramatically higher. In fact, actually it's 38%. So its raw ability to fix a mistake. It doesn't know the answer to is roughly the same. And remember in the smaller example was 40%. So it's now even slightly lower. But look at the prudence. No, this is where everything now changes. Now this model becomes vastly more prudent, more cautious. Its oversyncing rate plummets now to 7%. So it is now, if you want, much more stable in its reasoning performance and it is less likely to corrupt here a correct answer and modify it into an incorrect answer. And you may say, okay, so what is the net effect? So let's calculate this. So again, we have a 100 question test set. Now we have a llama 37B model again gains from intelligence plus 7 losses from imprudence here - 5.7 and the net reasoning gain is now plus 1.5 and I looked at this and said what how is this possible? How is it possible that the big more powerful reasoning model, let's say it is a reasoning model, has suddenly less net reasoning gain? The delta for the llama 7dB model is 1.5 and therefore smaller than for the small 7B model. How is this possible? No, because normally I would assume that a large language model reasoning process is far more reliable and efficient. Shouldn't it result in higher gains? And then I noticed that I made a mistake in my reasoning in my human reasoning. The 70B model reasoning gain delta is smaller precisely because it's initial knowledge you remember the accuracy of the fast thinking is so much higher 81%. So it has already solved most of these easy and medium difficulty problems. these low complexity levels using its if you want type one its fast syncing mode. So the pure knowledge retrieval system. So the reasoning process now the system two syncing or whatever you want to call it now is left to tackle only the most difficult complex problem where the potential for an error is high for any process. No, but the potential for large gains has been indeed diminished by its own superior baseline performance. And now I have an personal idea. So you see here me starring ideas and I have just want to tell you hey I think that in the 70B model the knowledge and the reasoning data encoded representation in the transformers are not too separate anymore are not two separate clunky modules anymore. Right. Somehow I have a feeling that its vast and robust knowledge base of a 70B model or a bigger m reflected here in the high a false performance accuracy already contains here implicit reasoned out patterns in its pure reasoning structure. So what it means I showed you here if you especially do a distillation process from a teacher to a student model I showed you we don't just have factbased but we take the reasoning traces from a deepse R1 we extract the reasoning traces and those reasoning traces become now the pre-training data sets for the smaller student model. So it's not just that we say we have pure data and pure information. We also have knowledge that is implicitly reasoned patterns. So this means the distinction between what is pure knowledge and what is pure reasoning. I don't think it's really that massive here when we scale the models up to higher model sizes and especially if we have this distilled models and I think it's fast syncing is so good because it requires less slow syncing correction because on the reasoning traces that it was trained on there is already here a pattern that provides a solution to the complex problem but this is yeah my personal statement you See here me looking at this coming back to the paper. So they say if let's focus here not on delta let's focus here on the normalized rates and they say this is the key takeaway from our study. So to compare the intrinsic quality of the reasoning process itself of an LLM, we should look here at the rates here RC and R O overthinking which normalize for the baseline knowledge. Now this makes sense. Now also now I think the 70B model reasoning process is far more if you want reliable or efficient not because it produces here a larger delta which it does not as we just saw but because it achieves its result with vastly less selfsabotage. This ratio of oversyncing it is much less. So it is a more stable reasoning pattern machine. Its reasoning process as it flow is much more precise and targeted and more elaborated. It is not just jumping around and self-correcting it whenever it thinks it's possible. That's a more stable more advanced model 70B model in general. So now an interesting question for you my audience is it that the primary function of scaling scaling up a model is to create a more robust and stable internal world model for this particular LLM. And remember world model reward models reinforcement learning. Yeah exactly this is where we are. So that it is less easily perturbed by the noise of its own sorts. Is this really the main effect that distinguishes here a small language model from a large language model? Especially it's it's system one thinking and system two thinking. Is this the main effect? Well, at least the paper showed us that the reasoning is not monolytic. It's not a parameter in itself. No, it's a contextual relation. So the framework that we show here proves quantitatively that the reasoning benefit our delta is highly domain dependent. I show you the example of mathematics and medicine. It's massively positive in mathematics but it can be negative in a knowledge intense field let's say history if the underlying parametric knowledge base of this particular LLM is weak of course but easy. Now if we have an um if you want not so performant knowledge base we know we can help here with graph rack optimization. So we come now closer to the question what should I do if my model is not performance? Should I activate graph rack or should I continue here with reasoning traces? And the answer is yeah. But before I give you maybe my answer, there was another question the author said and they said, "Hey, can we maybe pinpoint the exact location in a transformer layer where this is happening? What we just found?" And they have another parameter. Well, of course, but it's a simple it's a simple similarity measure between here type one and type two between fast and slow syncing. And so this CKA is calculated here for each specific layer in our transformer between the two modes of fast syncing and slow syncing. How you do this? It's a simple metric. Here's the calculation for you. Great. But what is here the main idea? And yeah, maybe it's a good idea to introduce here to the main paper. This is here by Chingua University here from Beijing in China. This is here published July 24, 2025. A beautiful paper. It gives you a lot of factual data, but sometimes it leaves open your answer to the main question that I, for example, have. But really interesting paper. Have a look at this. And the authors they say listen we have the the the starting idea from the human that we have a dual system cognitive theory also for LLMs which I would say is a stark idea to start from but okay let's go with their theorem and they say this dual system cognitive theory which decomposes here the LLM interference in interference sorry into two distinct but complimentary phases And they see an LLM also has here a phase one and a phase two thinking. Phase one is now the knowledge retrieval where the LLM rapidly generates here an initial response by accessing here the learned information. And I showed you at the very beginning of this video the prompt for phase one and the reasoning adjustment. This is the phase two and also show you the prompt for the phase two. So you can experiment with this and they say here is where they refine here the initial responses through a chain of sword generation or whatever you have. So knowledge retrieval from its parametric knowledge without syncing too much is phase one of the LLM. And if you start here, you know this typical sync step by step and validate and react and retract and everything. This is not a reasoning adjustment of the LLM. This is what I call call here the phase 2 of an LLM. If we can map this really from a human brain to an AI, I would put a big question mark behind this. But let's go with this. And they argue here to separate the two cognitive phases of an AI systems. No, the LLMs are now simply prompted to generate answer under the two distinct cognitive mode, the fast syncing and the slow syncing. And you have seen already the prompt at the very beginning of this video. And they think now the difference between the cognitive mode is analyzed now and they found it is a kind of a decoupling of knowledge and reasoning and especially the reasoning is further subdivided. So the reasoning adjustment sorry as they call it is now decomposed here into the correction our delta C and the oversyncing our delta O or the rate of correction and oversyncing. So this is their assumption which they are working with if it's the absolute true model. H but interestingly their approach provides us more information analyzing here the limitation of small model reasoning. Why do small LLMs fail so massively on the reasoning process? And their answer is it's not because they are so small. So they are incapable of reasoning. But they say here if we look at the reasoning adjustment that is happening with and a small LLM you have two forces within this. Yeah. So you have a correction force here but you also have an oversyncing that kind of destroys here the performance. But this overthinking is so much more dominant in small model reasoning which would give us an idea how to optimize those small models especially here on the oversyncing problem that I was not aware at all. So if we go with this hypothesis that LLM also generate answers like human in two phases. No LLM retrieved memory knowledge and LLM applied chain of thought reasoning. Beautiful. I don't know this entroperformation now applied here to this thinking but let's go with this. So fast thinking is an algebra drill and slow syncing is reasoning adjustment great and they have some beautiful thing here and yes we already went through the formulas. What are the insights? This cognitive hierarchy indicating that the knowledge retrieval and the reasoning adjustment operate at different hierarchical layers within a neural network is kind of what we already know of. Yeah. But they made it more quantitatively easy to check and this is a nice aspect. And of course they have a new parameter they call here the kernel the centered kernel alignment. Here this is simply here a measure of a neural network alignment a similarity. And here you have it A and B. Here we have a Q1 7B and here we have a llama 8b. So real close here in the size. And you see if we have on the x-axis here the number of the layers over 30 layers you see here this new parameter. So this remember this is a similarity parameter and a lower cka below one going down here as you can see indicates reduced similarity for the given layer that they examine. So as you see here all CKA curves here exhibit an in initial plateauing if you want here in the lower layers followed by a decline here in the higher layers. We see this here for this mall. We see this here a little bit less for this mall and all the data sets that I told you at the beginning. So this is an interesting phenomenon. What does it indicate? It indicates that the lower layers here remain similar across our syncing modes while here in the higher layers in the later layers of our transform architecture we have more divergence in the similarity and you know we we have an idea already that in the last layers of a transformer architecture the real reasoning is happening. So this would mean that here the phase one or the slow thinking would happen here in the first let's say 20 layers if you look here no this is almost a plateau and they say yeah this suggests now that both modes share the knowledge retrieval here in the lower layers but the slow syncing additionally engages here a reasoning adjustment in the higher layers and this is why we see suddenly our CK parameter goes goes down the similarity measure is reduced significantly. So the C results suggest the knowledge retrieval and the reasoning adjustment are primarily localized to the lower and the higher layer. So knowledge retrieval lower layer reasoning adjustment the higher layers only interestingly so see those performance indicators here on the different levels. So what does it mean now for a typical AI scientist? It means we have now a tool set for a targeted intervention. We open this video with a question. Hey, why the model is not working? Is it because the reasoning capability is not given a complexity level that I want to solve or is it simply it has not enough facts about it and they tell us now a knowledge deficit calls for a continual pre-training or we need graph rack so we have to provide additional data new knowledge they say however if we discover that there is a reasoning deficit here in the performance of the LLM this calls now for supervised finetuning or further if you want chain of sort reasoning trace distillation from a much more potent model. So this is an interesting insight now in addition to everything that now if something is not working we kind of have now a tool set and they tested it only here in alarm on a Q&A 2.5 to have here a further understanding why this LLM is not functioning and if we understand why we can take targeted actions we don't waste anymore any compute cycle on the wrong fixing of the problem. So interesting study kind of a knowledge we already knew what's happening. It kind of explains in a new way why smaller reasoning language walls are not have do not have the performance like an LLM. It is not in the scaling itself but it is also here in the particular oversyncing process that a smaller reasoning model here simply destroys already correct answers. Absolutely amazing insight. Have a look at this study. You're going to enjoy it. If you like this kind of videos, subscribe and I see you in my next