هوش مصنوعی برای همه: هوش مصنوعی چطور باهوش شد؟
هوش مصنوعی امروزی از چه مسیری به تواناییهایش رسیده است؟ این دوره تاریخ فنی و مفهومی آن را از پرسش هوش ماشینی و سیستمهای نمادین تا یادگیری از داده، deep learning، Transformer و مدلهای مولد دنبال میکند. هر تحول را با مسئلهای که پیش از آن وجود داشت، ایدهٔ تازه، مثالی ملموس و محدودیتهایش بررسی میکنید؛ نه صرفاً با فهرستی از نامها و تاریخها.
مسیر دوره شامل 28 درس کوتاه با ترتیب پیشنیازهای روشن است. مدرس هوش مصنوعی هر بار یک مفهوم را آموزش میدهد و با پرسش و آزمون، یادگیری آن را میسنجد. مثالهای تشخیص میوه، حل بازی و جستوجوی اسناد، همراه با نمودار و تشبیههایی با حدود مشخص، مفاهیم را بدون کدنویسی یا اثباتهای سنگین توضیح میدهند.
در بخش پایانی، تولید متن و تصویر، hallucination، context، prompting، RAG، ابزارها، workflow و agent را به هم پیوند میدهید. تفاوت خود مدل با نرمافزار پیرامون آن روشن میشود تا بتوانید ادعاهای مربوط به توانایی و قابلیت اعتماد یک سیستم را با توجه به شواهد و شرایط استفاده ارزیابی کنید. برای شروع، دانش عمومی در سطح دبیرستان کافی است.
زمان تقریبی یادگیری: 19 ساعت
آنچه یاد میگیرید
- توضیح دهید چگونه محدودیتهای جستوجو، قواعد و سیستمهای خبره به گسترش رویکردهای یادگیری از داده کمک کردند.
- مقایسه کنید یادگیری نظارتشده، بدون نظارت و reinforcement learning چه دادهها، هدفها و شیوههای ارزیابی متفاوتی دارند.
- شرح دهید شبکههای عصبی، deep learning، داده و توان محاسباتی چگونه پیشرفت بینایی ماشین و پردازش زبان را ممکن کردند.
- توضیح دهید token، embedding، attention، Transformer و pretraining چه نقشی در تولید متن دارند و تولید تصویر با diffusion چه ایدهای را دنبال میکند.
- تمایز بگذارید میان مدل، context، RAG، ابزار، workflow، agent و harness در یک سیستم مبتنی بر مدل زبانی.
- ارزیابی کنید خروجی یک سیستم با چه شواهدی پشتیبانی میشود و چرا hallucination، کیفیت داده و شرایط آزمون بر اعتمادپذیری آن اثر دارند.
سرفصلهای دوره
- معرفی دوره: هوش مصنوعی چطور باهوش شد؟
- A 10-15 minute introduction to understanding how today's AI systems became possible and when their answers deserve trust.
- The central storyline: moving from instructions written by people to systems that learn from examples, create content, and carry out tasks.
- The course roadmap: early ideas, learning from examples, increasingly capable models, content creation, and systems that use outside resources.
- How pictures, simple games, fruit examples, and document searches will make the ideas understandable without programming or advanced mathematics.
- Why each stage will connect an earlier difficulty, a new idea, a practical example, and a remaining limitation.
- هوش مصنوعی و پرسش هوش ماشینی
- The development of programmable computers in the 1940s and the question of whether machines could perform tasks associated with intelligence.
- The emergence of AI as a named research field in the 1950s.
- AI as the study and construction of systems that perform tasks such as searching, recognizing patterns, generating language, and selecting actions.
- Distinguishing observed task performance from claims about understanding or consciousness.
- Task-specific systems, general-purpose interfaces, and the difference between broad usefulness and reliable ability across all tasks.
- A timeline diagram showing overlapping research traditions rather than a sequence in which every new method replaces the previous one.
- الگوریتم، داده، مدل و یادگیری
- An algorithm as a specified procedure, illustrated by sorting fruit using explicit steps.
- Data as recorded examples or observations, with inputs and optional target answers.
- A model as a representation used to produce outputs from inputs.
- Handwritten rules versus adjustable model parameters.
- Training as a procedure that adjusts a model using examples; inference as using the resulting model.
- A diagram separating the training algorithm, training data, trained model, and new input.
- Why learning from data does not mean a system automatically learns from every conversation.
- هوش مصنوعی نمادین و بازنمایی دانش
- Why early AI researchers represented problems with explicit symbols and statements.
- Facts, relations, and if-then rules in a small collection of geometric objects.
- Separating a knowledge representation from the procedure that reasons over it.
- Tracing a short chain of rule applications to derive a new fact.
- The strength of explicit representations when the world is well specified.
- The difficulty of translating messy images, ambiguous language, and exceptions into complete rules.
- جستوجو، روشهای ابتکاری و بازی
- Representing a puzzle using states, legal actions, and a goal.
- Search trees and the rapid growth of possible action sequences.
- Breadth-first search and depth-first search through small diagrams.
- A heuristic as an estimate that guides exploration rather than guarantees correctness.
- Game trees and choosing moves while accounting for an opponent's possible responses.
- Why search produced important early successes in structured tasks but struggles with huge spaces and poor representations.
- سیستمهای خبره و گلوگاه کسب دانش
- The growth of expert systems during the 1970s and 1980s.
- Knowledge bases, inference engines, and explanation traces.
- Forward chaining from observations and backward chaining from a proposed conclusion.
- A rule-based machine-troubleshooting example.
- The effort needed to obtain, maintain, and reconcile specialized rules.
- Brittleness, missing information, and why confidence labels do not repair an incomplete knowledge base.
- افتوخیزهای فنی و گذار به یادگیری
- The gap between successful demonstrations on small problems and dependable performance on larger ones.
- Technical limitations behind periods of reduced expectations in AI, including the commonly named AI winters.
- Search explosion, brittle rules, limited computing resources, and insufficient digital examples.
- Why recognizing fruit from images is harder to specify with rules than sorting fruit by an already measured size.
- The growing appeal of estimating useful patterns from examples.
- Why symbolic methods, search, and learned models continued to coexist rather than disappear.
- یادگیری نظارتشده: پیشبینی از مثالهای برچسبدار
- Features and labels in a fruit dataset.
- Classification as choosing a category and regression as predicting a numerical quantity.
- Learning a decision boundary or numerical relationship from examples.
- Prediction error and loss as feedback for adjusting a model.
- A small visual example of improving predictions over repeated updates.
- How learned rules differ from both memorizing individual examples and writing every decision manually.
- تعمیم، کیفیت داده و ارزیابی مدل
- Training, validation, and test examples as distinct parts of model development.
- Underfitting and overfitting through simple curves and fruit-classification examples.
- Data leakage and why reusing test answers can produce misleading results.
- Correct and incorrect predictions by category; accuracy, precision, and recall using small counts.
- Shortcut learning, noisy labels, and unrepresentative sampling in object datasets.
- Distribution shift, illustrated by studio fruit photographs versus dimly lit images.
- Why a prediction score is not automatically a calibrated probability of correctness.
- یادگیری بدون نظارت و کشف الگوها
- Learning from examples without supplied target labels.
- Clustering objects by measured similarity.
- How changing features, scales, or a similarity rule can change the resulting groups.
- Dimensionality reduction as a simplified view of many measurements.
- Anomaly detection as identifying unusual examples rather than automatically diagnosing their cause.
- Why discovered groups do not come with guaranteed meanings or correct names.
- یادگیری تقویتی: یادگیری با انجام عمل
- An agent, an environment, observations, actions, and rewards in a small maze game.
- A policy as a strategy for selecting actions.
- Immediate rewards versus longer-term outcomes.
- Exploration and exploitation.
- How repeated interaction differs from learning from a fixed set of labeled examples.
- Connections between learned strategies and game search.
- Sparse rewards, costly trial and error, and behavior that maximizes an imperfectly designed reward.
- شبکههای عصبی و یادگیری بازنمایی
- Early artificial neuron models in the 1940s and learnable single-layer classifiers in the 1950s.
- Inputs, weights, a bias, and an activation through a small numerical diagram.
- Why a single simple boundary cannot solve every classification problem.
- Hidden layers as intermediate representations that transform the input.
- Learning useful features rather than specifying every feature by hand.
- The limits of the brain analogy: artificial networks are mathematical models, not faithful copies of brains.
- یادگیری عمیق: آموزش، داده و توان محاسباتی
- Deep learning as learning through multiple layers of representations.
- The wider use of backpropagation in the 1980s and renewed progress in deep learning during the 2000s and 2010s.
- Gradient descent as making small parameter changes to reduce loss.
- Backpropagation as efficiently assigning sensitivity to parameters, without deriving calculus formulas.
- Training batches, repeated passes through data, and why optimization can be difficult.
- The combined contribution of larger datasets, improved methods, GPUs, and increased computation.
- Why greater size can improve performance without guaranteeing correctness or unlimited progress.
- بینایی ماشین و ویژگیهای تصویری یادگرفتهشده
- An image as a grid of numerical pixel values.
- The shift from manually designed visual features toward learned features.
- Convolution as applying a reusable local pattern detector across an image.
- Layers that combine local patterns into more complex visual representations.
- Large image datasets, GPU training, and major recognition improvements in the early 2010s.
- Classification, object detection, and segmentation as different output tasks.
- Failure under unusual viewpoints, backgrounds, lighting, or misleading visual shortcuts.
- مدلسازی زبان پیش از ترنسفورمر
- Language modeling as predicting plausible continuations of a sequence.
- Count-based models that estimate the next word from a short preceding window.
- Sparse observations and the limitations of fixed short windows.
- Recurrent neural networks as updating a compact state while reading a sequence.
- Sequence-to-sequence systems for tasks such as translating simple technical instructions.
- Difficulties retaining distant details and processing long sequences efficiently.
- Why useful language patterns can be learned without a handwritten rule for every sentence.
- توکنها و امبدینگها
- Tokenization as dividing text into units that may be words, word parts, punctuation, or other fragments.
- Why visible word counts and model token counts can differ.
- Token identifiers as lookup keys rather than measurements of meaning.
- Embeddings as learned numerical representations.
- A two-dimensional picture of similarity, with the warning that real embeddings usually have many dimensions.
- Context-dependent representations and why a word can have different meanings in different sentences.
- The difference between representations used inside language models and embeddings designed for document retrieval.
- توجه و ترنسفورمر
- Attention as assigning different relevance weights to available sequence representations.
- Queries, keys, and values through a small relevance-matching diagram without matrix calculations.
- Self-attention, multiple attention heads, and position information.
- The Transformer architecture introduced in 2017 and its combination of attention and feed-forward layers.
- Causal masking for next-token prediction.
- Why parallel processing of training sequences helped scaling compared with recurrent processing.
- Why generating a sequence still proceeds step by step and why long contexts remain computationally costly.
- پیشآموزش، فاینتیونینگ و بازخورد
- Self-supervised learning: deriving training targets from the data itself.
- Pretraining a language model on next-token prediction across large text collections.
- Why pretraining is different from installing a searchable copy of every source document.
- Supervised fine-tuning on examples of instructions and desired responses.
- Preference-based training, including methods that use reinforcement learning and methods that do not.
- Adaptation with a smaller specialized dataset and the possibility of losing other capabilities.
- The difference between parameter changes during training and instructions supplied only at inference time.
- هوش مصنوعی مولد و تولید متن در مدلهای زبانی بزرگ
- Predictive classification versus generating a new sequence or example.
- A large language model as a model trained to assign probabilities to token continuations at substantial scale.
- Autoregressive generation: predict, select, append, and repeat.
- Sampling and temperature through a small next-token probability table.
- How next-token training supports many tasks without making every task equally reliable.
- Novel combinations, memorized fragments, and why generation is neither simple copying nor guaranteed originality.
- Distinguishing a model's text generation from retrieval, calculation, or tool execution performed by a surrounding system.
- تولید تصویر و مدلهای چندوجهی
- The progression from earlier learned image generators to diffusion-based systems.
- A high-level contrast between adversarial training and learning to remove added noise.
- Diffusion generation as repeated denoising guided by learned patterns.
- Text conditioning and why a requested image is not simply retrieved from a gallery.
- Latent representations as compressed working spaces rather than ordinary picture files.
- Multimodal models that connect text, images, and other input types through learned representations.
- Limitations involving exact counts, spatial relationships, fine detail, and unsupported descriptions of images.
- کانتکست، پرامپت و یادگیری درونمتنی
- A context window as the bounded input available for a particular model call.
- Instructions, examples, conversation history, and supplied documents as different context contents.
- In-context learning as adapting behavior from examples without updating model parameters.
- Context limits, truncation, and why a longer window does not ensure perfect use of every detail.
- A simple prompt structure using a task, relevant evidence, constraints, and an output format.
- The difference between visible conversation history, application-managed memory, and model training.
- Why requesting explanations can help inspect an answer but cannot establish that its account faithfully reveals internal computation.
- هالوسینیشن، عدم قطعیت و بررسی ادعاها
- Hallucination as plausible generated content that is unsupported or incorrect.
- Why an objective based on likely language does not directly enforce truth.
- Factual errors, invented references, arithmetic mistakes, and failures to follow supplied evidence.
- The difference between a correct statement, a supported statement, and a convincing statement.
- Why verbal confidence and repeated agreement do not constitute independent verification.
- Checking claims against original documents, reproducible calculations, and appropriately independent sources.
- Appropriate uncertainty and abstention when the available evidence is insufficient.
- بازیابی و تولید تقویتشده با بازیابی
- Why a model's parameters are a poor substitute for a current, inspectable document collection.
- Keyword retrieval and semantic retrieval using document embeddings.
- Splitting documents into chunks, indexing them, retrieving candidates, and optionally reranking them.
- Retrieval-augmented generation, or RAG: placing retrieved material into the model's context.
- A continuing example that answers questions from equipment manuals.
- Retrieval failure versus generation failure.
- Source attribution, irrelevant chunks, incomplete evidence, and why RAG reduces some errors without eliminating hallucination.
- استفاده از ابزار و انجام عمل
- Why generating text is different from calculating, querying a database, or changing a file.
- A tool interface as a description of available operations and expected arguments.
- The tool-call loop: propose a structured call, validate it, execute it in the application, and return the result.
- A calculator and a document-search tool in the equipment-assistant example.
- Read-only operations versus operations with side effects.
- Invalid arguments, tool errors, and why a model's claim that an operation succeeded is not evidence of execution.
- The distinction between capability residing in model parameters and capability supplied by surrounding software.
- گردش کار و سیستمهای مبتنی بر مدل
- A workflow as an application-defined sequence of steps and branches.
- Combining ordinary software, retrieval, tools, and one or more model calls.
- Structured outputs and validation between steps.
- A fixed equipment-assistant workflow: retrieve a manual section, extract values, calculate, and produce a sourced explanation.
- Retries, explicit failure paths, and passing state between steps.
- Why a well-designed workflow can outperform a single more capable model on a constrained task.
- Distinguishing the base model, the full application, and the user-facing interface.
- عاملهای هوش مصنوعی: برنامهریزی و چرخهٔ عمل
- An operational definition of an agent as a system that selects actions toward a goal using feedback from its environment.
- How current language-model agents relate to earlier AI agents without necessarily using reinforcement learning.
- A dynamic loop of observing, selecting a next step, calling a tool, and updating task state.
- Planning and replanning when a document is missing or a tool returns an error.
- Agent memory as stored application state rather than automatic parameter learning.
- Fixed workflows versus systems with model-selected next steps, including hybrids.
- Compounding errors, unnecessary actions, loop failures, and uncertain definitions of completion.
- Harness عامل و کنترل اجرا
- A harness as the surrounding software that runs the agent loop and manages its execution.
- State storage, context assembly, tool routing, and handling tool results.
- Bounded retries, time limits, action budgets, and stopping conditions.
- Restricted tools, isolated execution environments, and confirmation before consequential changes.
- Treating instructions found inside retrieved documents as untrusted content rather than application commands.
- Execution logs and checkpoints that help reconstruct what actually happened.
- Why terminology varies and why a harness is not another name for the language model.
- قابلیت اعتماد سیستم و شواهد توانایی
- Reliability as a property of a complete system operating on a specified task under specified conditions.
- Distinguishing model-only performance from results obtained with retrieval, tools, repeated attempts, or external assistance.
- Representative task samples, success criteria, and reproducible execution records.
- Why a benchmark score or selected demonstration may not predict performance on a different task.
- How failures across a chain of steps can reduce end-to-end success even when individual steps usually work.
- Matching verification effort to the consequences of an incorrect answer or action.
- Choosing among rules, search, learned prediction, generation, retrieval, fixed workflows, and agents based on task requirements rather than historical novelty.
شروع این دوره در تیچمی