0:00 Music 0:06 how training works, like on a very, very high level. 0:12 This is approximately the amount of data that's used for training and fine-tuning. 0:19 So on the left, there are like trillions of trillions of tokens that are provided in the base LLM as a knowledge base. 0:27 It's numbers in, numbers out. 0:31 And this contains of all the general knowledge, all that our foundational model providers have put into the training data set. 0:41 It's months of compute, just to give you understanding how much energy and... 0:48 Information and efforts being spent to train an LLM. 0:52 When some big companies, foundational providers, train the LLMs, 0:57 they have to align the training schedule with the power grid providers and power plants, because if they just do it without control, they would put all the electricity network down. 1:09 It's like the impact is just immense. 1:12 And it takes times, months and months of compute to train kind of large language models with this amount of tokens and data set and a lot of energy. 1:23 So kind of with energy to train like, I think, 1:28 Even one-year-old. 1:30 LLM, about 600 billion tokens, is enough to go to the moon and back 500 times. 1:36 So it's like really immense. 1:39 And this takes a lot of effort, but this gives this foundational model that trained on all 1:46 the average approximation of all the knowledge ever available on the internet that companies were able to build a data set from. 1:55 And then there goes the fine-tuning process. 1:58 And fine-tuning process is the approach when you specify explicitly how to align the outputs from this general corpus of data for particular outcomes. 2:10 You can specify for the domain, for the style, for the particular ethics, for like guardrails, constraints, anything that you would want your EI to be fine-tuned for. 2:25 And so kind of the fine-tuning is much more affordable process. 2:30 And for many use cases, for example, very often when for the companies and the consumers, when your 2:39 prompts become like too big and too large and you still see some deviations and you have to constantly add some additional rules and guardrails, where often it might make sense to 2:51 find someone to help to fine-tune a particular LAM for your general purposes. 2:57 And that's usually my 3:00 much more affordable process. 3:01 Now we have a lot of infrastructure for it. 3:03 So what you need to know, if you use case, you have to constantly do some adjustments of your inputs and have still some deviation from what you'd expect. 3:17 And you have a well-defined understanding what it is that you are missing in the outputs. 3:22 You can always fine-tune some available general purpose LLM for your purposes. 3:31 And that's how it makes a lot of difference. 3:35 So the general rule of thumb, whenever prompt tuning stops working, it's time to think about how to turn on and do defined tuning. 3:46 Okay, let me move on to the next. 3:53 So how training difference from inference? 3:57 So training is a process when you are 4:03 collect all the data set inputs, outputs, combination pairs, and build this huge mathematical calculator that is able to predict next words or next most likely outcome of your request. 4:19 And fine-tuning is being able to specify how exactly to process. 4:24 And this is training. 4:25 The process that most commonly you will be using is called inference. 4:31 And inference is extracting most probable outcome based on your input provider. 4:36 And how this usually works, now you understand the token, you understand the model, you build your prompt, provide it to the system, the system tokenizes it, breaks it down into like particular token sequence of number. 4:51 Model processes your request. 4:53 It predicts what will be the most likely response based on its training data set. 5:00 Generates output tokens. 5:02 Output tokens get converted into text that you receive in your response. 5:06 So, and that's where kind of the size of the model and the performance matters because, for example, 5:13 Some requests. 5:14 The smaller models can take less than a second, for instance, but more complex models, it can take tens of seconds or even like minutes, like in case of reasoning models, for example. 5:27 So that's when models self-reflects on its own response and tries to... 5:33 Process the same request again and again, again, the output measures particular criteria. 5:39 So that's where the outcomes are kind of meta and you need to think like very well strategically how to choose the right model for your part of the process. 5:52 Why I'm explaining this now, because for some tasks, for example, deep analysis or like 6:00 synthesis of several research results using expensive model makes sense, but for some more fine-tuned and genetic processes like collection of data, 6:12 recombination of already known existed, it doesn't require deep analysis or deep understanding, you can just use like 6:20 smaller model and you won't lose a lot in the value of the information that you get. 6:26 Because these smaller models are very often Excel, it's pretty cognitive expensive tasks, but they don't cost as much and they don't take as much time. 6:37 It's kind of the analogy that I like to give. 6:41 Is in avionics, it's 6:45 for quite a while that pilots have been offsetting the decision-making to our cognitive offload systems. 6:54 It's basically some auxiliary systems on plane that help to make them decisions faster and make decisions for them. 7:01 So they don't have to think of all details on how to balance the angles of different controls on the plane. 7:09 It's all kind of done automatically. 7:10 It's offloaded from the pilot's decision-making system. 7:13 They don't have to think about it. 7:16 And similarly, for us, like now with AI, these systems can be used as these cognitive offload structures and systems that just let you focus on the direction and trust 7:30 your autopilot to handle the rest of the cognitive loads on yourself.
0:00 Music 0:06 how training works, like on a very, very high level. 0:12 This is approximately the amount of data that's used for training and fine-tuning. 0:19 So on the left, there are like trillions of trillions of tokens that are provided in the base LLM as a knowledge base. 0:27 It's numbers in, numbers out. 0:31 And this contains of all the general knowledge, all that our foundational model providers have put into the training data set. 0:41 It's months of compute, just to give you understanding how much energy and... 0:48 Information and efforts being spent to train an LLM. 0:52 When some big companies, foundational providers, train the LLMs, 0:57 they have to align the training schedule with the power grid providers and power plants, because if they just do it without control, they would put all the electricity network down. 1:09 It's like the impact is just immense. 1:12 And it takes times, months and months of compute to train kind of large language models with this amount of tokens and data set and a lot of energy. 1:23 So kind of with energy to train like, I think, 1:28 Even one-year-old. 1:30 LLM, about 600 billion tokens, is enough to go to the moon and back 500 times. 1:36 So it's like really immense. 1:39 And this takes a lot of effort, but this gives this foundational model that trained on all 1:46 the average approximation of all the knowledge ever available on the internet that companies were able to build a data set from. 1:55 And then there goes the fine-tuning process. 1:58 And fine-tuning process is the approach when you specify explicitly how to align the outputs from this general corpus of data for particular outcomes. 2:10 You can specify for the domain, for the style, for the particular ethics, for like guardrails, constraints, anything that you would want your EI to be fine-tuned for. 2:25 And so kind of the fine-tuning is much more affordable process. 2:30 And for many use cases, for example, very often when for the companies and the consumers, when your 2:39 prompts become like too big and too large and you still see some deviations and you have to constantly add some additional rules and guardrails, where often it might make sense to 2:51 find someone to help to fine-tune a particular LAM for your general purposes. 2:57 And that's usually my 3:00 much more affordable process. 3:01 Now we have a lot of infrastructure for it. 3:03 So what you need to know, if you use case, you have to constantly do some adjustments of your inputs and have still some deviation from what you'd expect. 3:17 And you have a well-defined understanding what it is that you are missing in the outputs. 3:22 You can always fine-tune some available general purpose LLM for your purposes. 3:31 And that's how it makes a lot of difference. 3:35 So the general rule of thumb, whenever prompt tuning stops working, it's time to think about how to turn on and do defined tuning. 3:46 Okay, let me move on to the next. 3:53 So how training difference from inference? 3:57 So training is a process when you are 4:03 collect all the data set inputs, outputs, combination pairs, and build this huge mathematical calculator that is able to predict next words or next most likely outcome of your request. 4:19 And fine-tuning is being able to specify how exactly to process. 4:24 And this is training. 4:25 The process that most commonly you will be using is called inference. 4:31 And inference is extracting most probable outcome based on your input provider. 4:36 And how this usually works, now you understand the token, you understand the model, you build your prompt, provide it to the system, the system tokenizes it, breaks it down into like particular token sequence of number. 4:51 Model processes your request. 4:53 It predicts what will be the most likely response based on its training data set. 5:00 Generates output tokens. 5:02 Output tokens get converted into text that you receive in your response. 5:06 So, and that's where kind of the size of the model and the performance matters because, for example, 5:13 Some requests. 5:14 The smaller models can take less than a second, for instance, but more complex models, it can take tens of seconds or even like minutes, like in case of reasoning models, for example. 5:27 So that's when models self-reflects on its own response and tries to... 5:33 Process the same request again and again, again, the output measures particular criteria. 5:39 So that's where the outcomes are kind of meta and you need to think like very well strategically how to choose the right model for your part of the process. 5:52 Why I'm explaining this now, because for some tasks, for example, deep analysis or like 6:00 synthesis of several research results using expensive model makes sense, but for some more fine-tuned and genetic processes like collection of data, 6:12 recombination of already known existed, it doesn't require deep analysis or deep understanding, you can just use like 6:20 smaller model and you won't lose a lot in the value of the information that you get. 6:26 Because these smaller models are very often Excel, it's pretty cognitive expensive tasks, but they don't cost as much and they don't take as much time. 6:37 It's kind of the analogy that I like to give. 6:41 Is in avionics, it's 6:45 for quite a while that pilots have been offsetting the decision-making to our cognitive offload systems. 6:54 It's basically some auxiliary systems on plane that help to make them decisions faster and make decisions for them. 7:01 So they don't have to think of all details on how to balance the angles of different controls on the plane. 7:09 It's all kind of done automatically. 7:10 It's offloaded from the pilot's decision-making system. 7:13 They don't have to think about it. 7:16 And similarly, for us, like now with AI, these systems can be used as these cognitive offload structures and systems that just let you focus on the direction and trust 7:30 your autopilot to handle the rest of the cognitive loads on yourself.