Просмотр новости
Найдите то, что Вас интересует
Спутники
РН
SpaceX
Роскосмос
NASA
СпутникиРНSpaceXРоскосмосNASA
Image
To understand how frontier AI companies (like OpenAI, Google, or Anthropic) build AI Models, you should first know a very fundamental concept.AI Models don’t have genuine understanding. They just mimic and repeat the patterns. They don’t actually think. They are just a very, very advanced auto-complete system. For example, if you say “The sky is ___”, it can simply guess and complete the sentence by the word “blue”.So, building an AI Model, fundamentally means building an advanced auto-complete system.Let’s take a look at the process.Frontier AI models (like GPT-4, Claude, or Gemini) aren't built in one go. They go through a massive, expensive, multi-stage industrial assembly line. Here is the real, step-by-step process:
1- Decide the model's goalThe company decides what type of AI Model it wants to build, like a general chatbot, coding assistant, image generator, etc. Let’s say they want to build a language AI Model.
2- Collect massive datasetsThe training process starts with enormous amounts of text. They collect data from various sources like public web pages, books, academic papers, programming code, Wikipedia, …The internet itself is a large source of data.
3- Clean the dataThe collected raw data is messy and can't be used directly. They remove: spam, duplicate pages, broken HTML, advertisements, …
4- Tokenize everythingComputers don't understand words. Instead, they work with tokens. A Token is loosely either a word or a subpart of a word. For example, ("unbelievable" = "un" + "believ" + "able" → 3 tokens).Then, every token gets converted into an integer. For example ("The" → 4312). Now the dataset is just huge sequences of numbers.
5- Decide and design the learning algorithmThey use the algorithm as a recipe to discover patterns among data and build the final product of the AI Model. As an example, GPT stands for “Generative Pre-trained Transformer,” meaning it uses the Transformer algorithm. (It is too technical. As we want to understand the overall process, I don’t go deeper for this step.)
6- Build the training infrastructureThis is one of the hardest parts. Training uses thousands to tens of thousands of GPUs working together. This infrastructure alone can cost hundreds of millions of dollars.
7- Pretrain the modelThis is where the model actually learns language. Given “The sky is ….” → Predict “Blue”.The model repeats this prediction task trillions of times. The training loop is like this: Read text → Predict next token → Compare prediction → Compute error → Adjust → RepeatOver months, the model gradually learns: Grammar, Facts, Programming, Mathematics, Writing styles, Reasoning patterns, Multiple languages, …This step needs massive processing power.
8- Evaluate continuouslyResearchers constantly test the model. If performance is poor: change architecture, improve data, retrain, …
9- Supervised fine-tuning (SFT)The pretrained model predicts text well but isn't necessarily helpful.Humans create examples like:User: Explain gravity.AI should respond: (Gravity explanation...)The model learns to imitate high-quality responses like in a real conversation. Now it behaves more naturally like a human.
10- Preference training and reinforcement learningThey use Reinforcement Learning from Human Feedback (RLHF) to compare responses. In fact, humans compare outputs:Prompt: Explain gravityAI generates answer AAI generates answer BHumans pick the better one. The model learns to prefer responses that are more helpful, more accurate, better written, and safer.11- Safety trainingAdditional training teaches the model to:Refuse dangerous requestsAvoid generating harmful contentReduce hallucinationsHandle sensitive topics more carefullyFollow policies, … This stage is a major research effort in itself.
12- Red-team testingSpecialists deliberately try to break the model by prompting it with:Tricky questionsUnsafe requestsFalse assumptionsToxic questions, …Weaknesses found here often lead to more training or safety improvements.
13- Optimize for inferenceA training model is often too slow and expensive to serve directly. Engineers optimize it to generate outputs with reduced latency and cost while maintaining quality.
14- DeployThe model is then made available through APIs, Chatbots, Mobile apps, Enterprise platforms, and Cloud services.
15- Improve continuouslyLarge AI companies treat deployment as the start of the next cycle. They monitor bugs, failure cases, user feedback, infrastructure costs, and safety issues. This was the overall process. Hopefully, you enjoyed it.
------------------------------------------------------------------------------Are you an action taker?If you are a Product Owner and want to migrate from a classic to an AI-aware Product Owner, attend my upcoming PSPO-AI Essentials (Professional Scrum Product Owner - AI Essentials) class. Click here for the class information. If you are a Scrum Master and want to migrate from a classic to an AI-aware Scrum Master, attend my upcoming PSM-AI Essentials (Professional Scrum Master - AI Essentials) class. Click here for the class information.