# Why LLMs Need Two Timescales of Learning

## Метаданные

- **Канал:** Discover AI
- **YouTube:** https://www.youtube.com/watch?v=KXFnmlgQ_TI
- **Дата:** 15.05.2026
- **Длительность:** 35:18
- **Просмотры:** 836

## Описание

The video's central move is to stop treating LLM adaptation as a single process that must all be written into the tensor weights. Instead, it models adaptation as a coupled cognitive system with two interacting channels: a slow parametric channel θ and a fast textual/contextual channel ϕ.  Slow and Fast together - a coherent learning process for AI. We bring together the classical RL with verifiable rewards and the skill.md and memory.md complexity from the AI harness.

The slow channel is the ordinary model update path: expensive, persistent, and global. The fast channel is the prompt, instruction, reflection, and context layer: cheap, editable, and temporary. The key claim is that these two channels should co-evolve, not be optimized in isolation.

This AI Learns in Two Minds
Why LLMs Need Two Timescales of Learning

all rights w/ authors:
"Learning, Fast and Slow: Towards LLMs
That Adapt Continually"
Rishabh Tiwari∗ 1,4 Kusha Sareen∗ 2 Lakshya A Agrawal∗ 1
Joseph E. Gonzalez 1 Matei Zaharia 1 Kurt Keutzer 1 Inderjit S Dhillon 3
Rishabh Agarwal† 2,5 Devvrit Khatri† 3,6
from
1 UC Berkeley 
2 Mila 
3 UT Austin 
4 Eragon 
5 Periodic Labs 
6 Mirendil

#aithinking 
#scienceexplained 
#aiexplained 
#airesearch

## Содержание

### [0:00](https://www.youtube.com/watch?v=KXFnmlgQ_TI) Segment 1 (00:00 - 05:00)

Hello community. So great that you are back. Today we look at a complete new training and learning methodology to train our AI models. It is called the slow fast train, the SFT AI methodology. Let's have a look. Now it is no coincidence that in my last video we talked about here the AI scientist and I showed you here in this particular case that we modified the skill MD file, the memory MD files and the reinforcement learning in with a self DPO here all three together. And I told you this is essential because we modified the skill and the memory file here for the reinforcement for a coherent learning experience here together. And you remember one of the task of the AI scientist was to develop new skill MD files, to develop new methodologies, new insight, new scientific insight and therefore we are optimizing here the skill MD files together with the LLM itself. And this was already quite interesting because suddenly we are not anymore that we take any AI model, no. And we just add some skill model from the internet, no way. They are not coherent, they are not working in tandem, no. So we have to bring them that this is a coherent learning experience. This will simply improve our performance and remember we had here the big human feedback loop here where the human provided here additional guidance here for the development of the system and new skill MD files. Now skills become a topic of research. Why? Here on this publication by Shanghai Jiao Tong University from May 12, 2026, Proteus, a self-evolving red team for agent skill ecosystem. Why are skills becoming so dangerous in cybersecurity? Because a skill exposes both, know? Your executable behavior and your context setting documentation. Therefore, its deployment risk cannot be measured here as by simple prompt level red team alone. A realistic attacker, and we see this really in real time, can use the audit and runtime feedback to repeatedly rewrite the skills given the feedback from the system, and therefore optimize the skill multiple times until finally the skill can break through and is has access here to your infrastructure. In order to say, we frame this risk as an adaptive leakage, you know? Because an attacker can iteratively revise the skill until it passes your audit and produces some verified runtime harm. So, skills, extreme dangerous objects, do not download here any skill from the internet if you are not aware of what you're downloading, you know? Why is it? It is because the skills are so powerful, you know? And they show us here in Proteus, they perform a path expansion which find alternative implementation of successful attack vectors and some surface expansion which transfers learned implementation patterns to new attack objectives beyond original seed catalog. So, this skill will learn based on your LLM or based on your protection, based on your service security, and it is learnable, and it means it will find a way to penetrate here all your defenses. This is something we can use also in the positive one, and therefore, I do not want to focus on this paper, but I highly recommend that you read want to go now in a different direction. In a positive direction. You remember we had our LLM beautiful at the core here, and then we had around the LLM our AI harnessing sphere, and we have everything from graph reg, from lean four, from super computer, from database clusters. Beautiful. And today we're going to further develop this image here. Because the new idea is with skills becoming self-learning skills, and not something that you just download here for a simple task, but you can make those skills learn themselves new dangerous or non-dangerous behavior. We have a new learning dynamics. So, let's imagine here the following thought. At the core of our sphere, we have our LLM. We have our neural network. We have our layer of the transformer here. And this is here where we have the tensor weights. the training happens so that the system can do some deep thinking. And then we have on the surface of the sphere, we have all the problems and all the context from the outside world, from the environment. And this is all the short-lived data swirling around here on the surface, on the sphere. So, therefore we have now a system and we are now experiencing how we can couple this into a

### [5:00](https://www.youtube.com/watch?v=KXFnmlgQ_TI&t=300s) Segment 2 (05:00 - 10:00)

coherent system learning. So, we go a step further. We don't say we need new skill files or we need self-updating skill empty files or self-updating memory files. We go to a complete more complex system configuration where we say the core is now here our tensor structure, our weight tensors, and the prompt also will have here now a direct influence how learning happens here in the LLM. We go a step further. And yeah, never mind here the visualization. I currently work here with my AI here on physics, on plasma physics. So, therefore you see here torus, but never mind. This is just an image that I generated because I want to show you we now kind of separate the knowledge. We have here in blue, here the core, if you want plasma, is here the primary logic of an AI. The deep reasoning engine of your LLM, no? The neural network of a transformer that has been trained pre-trained and post-trained here on particular domain knowledge to perform this task. And then, surrounding all of this, we have here our scaffolding, no? Those are, let's call it the short-lived facts, no? What happened on Monday? What happened on Tuesday? The financial data from the Tuesday market, no? The outer scaffolding is here more or less something a surface here of a data flux of highly dynamic data. You have a high flux, throughput. Those data are changing within seconds. So, we now try to separate what is the core logic, reasoning engine that we really want to imprint here on the transformer layers. This is blue. And then, we want to say, "Hey, but wait a minute. All this scaffolding around, all the data that are changing daily or every second, no? We don't want to train the LLM to understand the I don't know, Monday, Tuesday and so pattern here because they will change continuously. So, we divide now the knowledge into the pure abstract knowledge for reasoning and into the dynamic knowledge here that will change constantly. And yes, of course, you see, we have here also kind of a scaffolding, kind of a harness here. And yeah, we will find here new mathematical optimization. This is the paper. UC Berkeley. I always love UC Berkeley when they publish something. They go about learning fast and slow toward LLMs that adapt continually. So, this is now something you know, if you have an LLM and we do have here the pre-training phase beautifully and then you have your domain knowledge and you want now to train it on your data. And maybe you want to do this every week you have new data. Now, we have the problem of catastrophic forgetting and and now they focus now here together with Mila and UT Austin and all the other beautiful artists here on a new methodology to train an LLM. They call it fast slow training and they give us here one indication that is up to three times more sample efficient than only the slow training which is of course the reinforcement learning from verifiable feedback that you know. Across multiple reasoning tasks while consistently reaching a higher performance asymptote. So, this is now UC Berkeley argues here in this article a complete new way to train our LLMs beyond reinforcement learning beyond in-context learning. Published May 12, 2026. And they say, "Updating the parameters now our tensor weights with reinforcement learning forces LLMs to absorb task-specific information that is maybe only valid here for a particular Monday morning, you know, or like a weather 3 weeks ago, who cares about this? Which can result in because now those knowledge has to be learned those data become now information and knowledge so therefore they have to overwrite some other data and we have now the phenomenon of catastrophic forgetting of our LLM and the loss of the reasoning plasticity of our LLM. This is known for years. " Now, you know we have in-context learning with a fixed LLM parameters so the tensor weights become frozen and now we can cheaply and rapidly adapt to task-specific requirements, no? This is everything with the prompt optimization DSP 3 whatever you have here about this optimization mechanics, no? But cannot by itself typically match here the performance gain that we have with a reinforcement learning training. And this was it for a long time, no? And now we have what is introduced now a new fast slow learning framework for the LLMs with the model parameters, our tensor weights in the LLM as the slow weights, and the optimized context, our context engineering and everything that we have in the harness as the fast weights. So

### [10:00](https://www.youtube.com/watch?v=KXFnmlgQ_TI&t=600s) Segment 3 (10:00 - 15:00)

this is the definition that we use here throughout the paper. So, the idea is simple, no? You provide some context here in your prompt. You have a concatenated prompt. Maybe you have a rack system that built here a beautiful context here with some few-shot examples, and then you have the task prompt, the task part here simply compute here for a function here for particular X and Y value here this particular function. So, what they do now is to say, "Okay, great. We have now a model in LLM model here our parameter theta, and we have trainable parameters. Now, you know, when you roll out the model and you have all the thoughts beautiful, and you have different tries now. You have specific tool calls, and you have an error and great. And now the learning for the model side, which they call the slow learning here, you have here a loop. The reward that is normally given is an end-of-task reward. So, at the end, do we have a result that is correct? Yes, the plus one. Is do we are stuck? Do we have some incorrect result? Minus one. So, we have a scalar reward that is given, yeah? So, we only at the end of the job we have here a reward, and then this is here the update that goes here with a slow weight update, and this tells the model which of it approaches was really working. But now, as I told you, we have those intelligence skills, those memory files, all those templates, all those tool calls that we have, now? So, now the question is, if we build the model more or less around a pure reasoning engine, and we have all the little flexible, highly dynamic data structures, not in the model, but outside in the wholeness of the eye. So, we have a context pool, and this is phi. Now, phi is, if you want, here the fast learning route here. So, we have also, we have the sort, we have the tool called the error, but then, since we are working here also with sorts, tool calls, errors, feedback operation, we can go here much faster, because we can have a fast learning optimization that simply goes and says, "Okay, if it's not tool number one, maybe it is tool number two, tool number three. " And you got the idea. And then, we have here a fast context update in our context pool. And then, if it works or does not work, but if we got some recommendation, for example here, draw context of population weeks within the group, or any additional information, we optimize the context and the whole loop starts from the very beginning. Now, you remember the slow loop here updates here our parameter pi theta from here the scalar rewards alone, and the fast loop at the bottom is here the updates our phi. Remember, this is now a population of Pareto frontiers via reflective optimization, consuming the wallet full text, including the sorts, the tool calls, the errors, the feedback from everything and whatever. And you might say, "But, okay, we have here the reinforcement learning, the supervised fine-tuning. This is everything that we know about the slow learning mathematics, the coding. But, how we optimize here the fast loop? " And now, the authors go and of course you see Berkeley, they go in their own toolbox and they say we will do it here with a GePA or the reflective optimization. Now, if you are um inscribed subscribed to member of my channel, you know, you just go to my channel, you put in GePA and you immediately get here the videos that I have on GePA. You see GePA originated here about 9 months ago. It was in New Channel AI structure here from MIT, UC Berkeley, Stanford, Databricks here and they called it here the end of reinforcement learning now because our GePA uh reflective prompt evolution can maybe even outperform reinforcement learning. Great. It was also from the authors here of the S price three and everything that is combined with this and I hope you have seen my videos because it was 37 minutes and I really tried to explain it to everybody. Now, you know that a month ago I give you here another research scientific video that GePA is already not the latest thing because we went from GePA to ViSTA and I showed you here with ViSTA that the human prompt is the most complex thing for any AI. But anyway, GePA. GePA in this publication here, you see Berkeley, Stanford and whatever, MIT, Databricks, 82 beautiful pages and they give us here the reflective prompt evolution and it can outperform reinforcement learning. Now, it seems it doesn't really outperform reinforcement learning because you see now UC Berkeley

### [15:00](https://www.youtube.com/watch?v=KXFnmlgQ_TI&t=900s) Segment 4 (15:00 - 20:00)

9 months later has chosen GePA only for the fast learning route but not for the complete learning route and we still have the reinforcement learning here for the slow learning. This is absolutely interesting but you see here also Omar here and you immediately understand that yeah, if you have seen my video from GePA to Vista, that we have done this because gayper is here. If you want yeah, it runs into a local minima and then it's fixed here and we need gayper. We need something much more intelligent here from the algorithm here. We need here Vista here and you see it really finds here gets out of the local minimum has enough momentum here in pulse gets out and find here the real global minimum here is the best solution. This is a multi-agent APO framework that is that decouples the hypothesis generation from the prompt rewriting. Beautiful, you find everything here in this video. So, gayper, you know how to handle gayper. Great. So, you see 9 months later gayper is suddenly only one of the two systems. So, we have reinforcement learning and gayper. Now, how we bring it together and we have to have the mathematics before we can do any coding because we code mathematics. So, the authors now give us here in chapter three describe FST fast flow training jointly optimizes here the slow weights, our thetas, through reinforcement learning by verifiable feedback. We know this. And the fast weights, our phi, through gayper. So, it means we have two methodologies we know. We know exactly we have the gayper repos. You can build it immediately. Great. The only question is how can we combine it here into here a joint expectation here. Now, you remember we have a population here of textual prompts here, our phi, and the optimization is rather simple. Here our U of phi here is uniform. Is the uniform over the prompt population. So, this looks easy, now. This is the only, if you want, mathematical formula. If you do not really familiar, I have a particular video, my goodness, 52 minutes, where I explain the AI mathematics that it absolutely necessary to understand those mathematical expression and I explain every single term and what it means, why we need this, how we write this, why we write it in this way, and you understand everything from this video. So, as I told you our natural joint objective here now from this new methodology is simple. So, if you want this the first important mathematical statement of the paper, the reward depends jointly on the model parameters theta and the context of phi. So, the learning problem is inherently two channel and you see this here. We have our theta and we have our phi, a two channel, and it turns out it is a coupled learning experience. Now, again, for the slow channel the paper uses reinforcement learning with very far away regrets. Rewards. Given a batch of the reward, it computes the group relative advantages. You notice this is a SIBO style standard reward driven post training. Beautiful. For the fast channel here, this is keeper. You have seen this in my video on keeper. Beautiful. So, together, as I showed you, becomes your this here. Beautiful. Where the phi is your the population of the prompts and the U phi is your the uniform distribution over the population of the prompts. And this is, if you want here again, the mathematical statement of the coupling that we have of the two learning system to learn out together. So, the policy is now trained under a distribution of fast context and not a single frozen prompt. So, this is now important because keeper here is not single prompt optimization, but we talk here about here complete ensemble. Now, if you want to see the complete run through. So, this algorithms runs in circles, no? At this particular circle, see, let's say you see equal one. It first pre-fetches here the next RL mini batches and call that look ahead batch LC. It then runs GePA, of course, using the current policy pi theta 4 C, let's say C equal 1, a frozen reflection model to look ahead data and the previous prompt population phi C as the C data. And then GePA, being GePA, returns a new prompt population C + 1. And after that, the system performs now T slow weight updates. So, after we have found here a good solution, the system says, "Okay, if this has here a new reasoning path, a new reasoning complexity, I will integrate this here into my tensor weights. " So, slow weight updates while holding here our phi C + 1 fixed. So, each RL batch sample rollouts now under all the prompts in the population, groups them together, computes here the group relative advantages, and applies

### [20:00](https://www.youtube.com/watch?v=KXFnmlgQ_TI&t=1200s) Segment 5 (20:00 - 25:00)

here a SISPO update functions. And then, the next cycle begins. So, beautiful. But, this means that the RL learner, our reinforcement learning here is always learning inside the current prompt ecology, while the prompt ecology itself is being updated based on what the learner just failed at, no? So, we then for- try to find skills, have a look at skill layer, my video here tries to find here the particular skills that are necessary to overcome here this failure, and then applying here these new skills, or these new composed skills with subskills, we find a solution. So, this means that the two learners are coupled here in a closed loop, no? We have our phi at C, we have the rollouts, then we have the theta at C + 1, the new rollouts, and then we have again our phi C + 2, so we always within the new environment of the other learner. And this is why this paper is not really that you can just downgrade it into a prompt tuning plus reinforcement learning. This is really, have a look at the formula, a co-learning, a co-adaptation of the learning methodologies. They depend on each other, and this is why it can be, theoretically, let's have a look at the result, real powerful, more powerful than a oil itself. Now, the paper's empirical thesis is that a fast channel absorbs some of the short-lived task-specific burden that would otherwise force you to slow weights to move too far. And this has consequences. And the first one is yeah, the learning becomes suddenly more data efficient. And I will show you the data in a moment. FST reaches here the reinforcement learning running peak in fewer optimization steps. And for code IO, three times fewer. For mathematic, 1. 4 fewer. And for yeah, some other tasks, also three times fewer. So, data efficiency and time and compute time is absolutely important. Second, the LLM model stays closer to its own base policy. The Kullback-Leibler is not as divergent as it will become otherwise. But the paper's Kullback-Leibler versus the reward plot shows that, and I will show you this in 2 minutes, that at match reward, FST is its left of the oil on the Kullback-Leibler axis, which means it achieves a similar performance with less parametric drift. Beautiful. And third, I started here also with the plasticity of the LLM here in a phase two problem. Checkpoint trained here much easier to adapt on a new task than a checkpoint trained with reinforcement alone. I will show you this data in a second. Let's do it. Okay. So, here you have it now, the validation reward, and we have here in solid green, the new FST, and then in dashed and this is a blue kind of a blue, we have reinforcement learning. And if you only look at GePa, you have here the dotted line. So, you see here that if we go here with our number of steps on the x-axis here, green reaches here the plateau performance plateau of reinforcement learning here at, I don't know, 42, 43 percent much, much earlier. You see, reinforcement learning takes about 1,500 steps and our new FST reaches the same performance at 500 steps. So, three times better. So, this is here first direct empirical sign here from the uh testing here that the coupled system is more sample efficient than the weight only learning with reinforcement learning, which is nice. So, it goes in the right direction. Here we have the validation reward versus the Kullback-Leibler divergence from the base model. Now, here this is rather simple. You see the green line is almost all the way on the left side here of our reinforcement line, no? Look at this. Especially here in physics, no? Green line is definitely on the left side here. So, this means if we have again on the y-axis the validation reward in percentage, at the same reward, FST requires less movement in the weight space. So, this is here really one of the best numerical evidences for the paper's central claim that the context can absorb part of the adaptation load. So, this means we do not have to cram everything into the reinforcement learning of the core, but context and context optimization here in the harness, in skill empty files or whatever you have here in templates, those contexts can absorb here part of a daily fluctuating fast data complexity.

### [25:00](https://www.youtube.com/watch?v=KXFnmlgQ_TI&t=1500s) Segment 6 (25:00 - 30:00)

So, the adaptation is not moved into the LLM reinforcement learning training. Plasticity. This ask here or answers here very practical question. After training on one task, how learnable is a new task, no? And the answer is that the FST trained model preserve more capacity for future learning than the reinforcement learning trained models alone. Because you see here reinforcement learning in it here in the blue line here, and our FST here in the green line. So, you see if you have here now and you take on a second complexity learning or you go here from physics to another task here, you see here the green line is given the number of step on the x-axis outperforming here the other methodologies. So, the authors argue that this is here the clearest evidence. I would say it's an indicator, not an evidence, but anyway, that the fast channel here protects here the slow channel, the weights from an over specialization, from an overfitting. Because all those data are not stored and not projected to be learned here by the core, but rather by the surface of our sphere in the AI harness. Beautiful. Now, let's combine this. A continual learning, and you know I started this with a catastrophic forgetting. And now if we have continual learning in an uninterrupted run. So, at first we train it on this one, then code, and physics, no? And you want the performance here stays for all three elements here maximum. Okay, so simple. The solid line is the new FST and the dashed line is here our reinforcement learning. You have here for the first block in blue, you see both are more or less identical, but then if we go here to the second one with code IO, you see here the solid line here just jumps up here to maximum performance, and you see that the dashed line is not coming up at all. So, if you change you'll have a complete new domain knowledge, this is exactly happening if you have reinforcement learning. You can't have catastrophic forgetting here, especially in the second part. But you see in the third part for physics, it's trying to come back also, no? So, it is really interesting, but in general we can say Wait a minute, where was I here? FSD keeps adapting at each stage, and reinforcement learning, the classical one, stalls badly in the middle stage, no? And recovers only partially later, no? And you what is arguing, no? This is the strongest demonstration that a couple system supports life adaptation under the task shift if we move from one domain to another domain, or we want to train another knowledge domain here on top of the other one. So, it seems continual learning with FSD, yes, it outperforms. And there you have it. So, the idea is simple. Whatever we put now on training data for the slow learning, for the weight optimization here of our neural network tensor weights, we are now very careful with the data in this um pre-training and post-training data sets because we just want to have a generic reasoning, a generic logic engine. And all the other, if you want, flux data that will change every day, we only want to have them on the surface of the sphere itself. Of course, they influence each other. They are now dependent in their continuous learning, and this is beautiful. So, you see you have a coherent learning process between the inner core of the sphere and the surface of the sphere. Now, you can go a step further. As not in the paper, this is now my recommendation. If we have this after some looping, then hey, this is easy because whatever is on the surface of the sphere, we just can extract, no, from the surface of the sphere because guess what? Those examples or new solutions are simple new skills. Or are simple new memories where you see, okay, we have discovered certain patterns, we examine certain lessons, we have some successful strategy, we have strategies that fail completely, put in the memory, make it clear, past reflection, everything. So, you can beautifully also have here a coherent translation into our current skill and memory and everything that we outsource here in the harness, but the learning is a combined interwoven complexity and therefore has a better performance. So, if you want this new paper, push a sphere ideas that I think are really useful. First, it gives us the idea that not useful not all useful adaptation should

### [30:00](https://www.youtube.com/watch?v=KXFnmlgQ_TI&t=1800s) Segment 7 (30:00 - 35:00)

be permanently written into the weights of our LLM, yeah? Because what the weather was on Friday, the November 30th is really not that important if it is about physics, no? So, some of it belongs just in context because the information is temporary or task-specific, and it is easier to revise the information over there or update the information about the new financial data that will come out tomorrow. Go or do not give it here into the core, but sphere. Second, that context and weights should be trained together. And this is, if you think, the big breakthrough. Not sequentially, not one after the other. No, you have to train it together. You have to find a mathematical presentation that you can code that you are really having interwoven complexity that depend on each other. And I think here GAPA and the classical reinforcement learning with verifiable feedback are really beautiful complementary training routines you can train together. You can bring together as I just have showed you, no? So, the prompt channel changes to rollout distribution which changes the gradient seen by the slow learner of reinforcement learning. And this is really the paper is about a coupled dynamics rather than a simple prompt optimization. So, you see you cannot do it anymore in a simple case. At first we do the part A and then we do part B. No, everything is a system. It has a complex interwoven system dynamics and we have coupled dynamic systems that we have to describe mathematically. And third, practical reason to care the couple system learns faster, stays closer to the base model, retains its reasoning plasticity, and handles task changes better because they are not in the core of the LLM but in the AI harness. So, therefore, the scientific claim is not just that prompt helps but the two-times scale learning architecture is a better abstraction for a continual LLM adaptation. So, you see we moved here from a prompt engineering to context engineering to DSPy 3 to a complete multi-agent whatever. Now, we combine GAPA with reinforcement learning. The next logical step would be to combine WISTER with reinforcement learning. And maybe then develop a complete new mathematic for the next generation of AI models. Beautiful. Yeah, hardware infrastructure Slurm cluster, eight times an H100 with 80 GB per node. You see they went here with a simple Q and 3 8 billion model for physics format, for whatever. Thinking without only one exception for StarGraph, they had a non-thinking model. Beautiful. Give you an idea how long does it take? Well, only for one single headline here on the new FSD, about 25 to 40 GPU hours. And of course GePA cycles are a sizable fraction here of this time. So, this is expensive. This is complex. This is really I would say not hardware intense. You just need Yeah, okay, this is hardware intense. But from the model, they work with a Q and 3 8 billion model. So, you see this is beautiful. Yeah, but on the other side then for GePA, they use here the GPT-5. 2. So, yeah, you can't do it. They decided to not do the second part locally. Maybe you can experience and also do GePA locally. Depends on the quality of your trained LLM and the domain complexity that you have to handle. Medicine, finance, chemistry, whatever. But gives you a good overview. It is a rather expensive infrastructure. But from a scientific point of view, we moved further and a big step further. So, we have a fast slow framework here for the LLM post-training phase that jointly optimizes here our classical reinforcement learning with verifiable feedback as a slow model depending here on our pi theta parameter and a fast textual context population what we used to call in the simplified way as in-context learning, but today it is more with these learnable skills. We are GePA reflective prompt evolution and now we have a complete interwove interleaving now of these two learning channels, and together they learn better, faster, higher performance for your system. What an interesting study, but there are so many other studies out there, but I decided I want to show you this because this opens up here a complete new learning experience that maybe we can here optimize also here on our local infrastructure with our local LLMs. More

### [35:00](https://www.youtube.com/watch?v=KXFnmlgQ_TI&t=2100s) Segment 8 (35:00 - 35:00)

about this in a later video. I hope you enjoyed it. You had a little bit of fun. Some new information for you to test out. Maybe you decide that you want to read here this paper. I highly recommend it, both papers that I introduced to you. And anyway, it would be great to see you in my next video.

---
*Источник: https://ekstraktznaniy.ru/video/50832*