[TOC]
↗ Universe, Self-Awareness, and Intelligence ↗ LLM (Large Language Model)
↗ Statistical (Data-Driven) Learning & Machine Learning (ML) ↗ Artificial Neural Networks (ANN) & Deep Learning Methods
- ↗ Mathematical Modeling & Abstraction
- ↗ Mathematical Logic (Foundations of Mathematics)
- ↗ Algebra
- ↗ Mathematical Analysis (& Analytical Mathematics)
- ↗ Cybernetics & Control Theory
- ↗ Game Theory & Decision Making in Multi-Agents Environments
- ↗ Operations Research (OR)
↗ AI4SE
↗ Artificial Intelligence Industry and Companies
https://www.jiqizhixin.com 机器之心
https://www.zhihu.com/org/xin-zhi-yuan-88-3 智能 + 中国主平台 微信公众号：新智元
https://huggingface.co huggingface
https://www.kaggle.com kaggle
https://datawhale.cn https://github.com/datawhalechina Datawhale Datawhale 是一个专注于数据科学与 AI 领域的开源组织，汇集了众多领域院校和知名企业的优秀学习者，聚合了一群有开源精神和探索精神的团队成员。Datawhale 以“ for the learner，和学习者一起成长”为愿景，鼓励真实地展现自我、开放包容、互信互助、敢于试错和勇于担当。同时 Datawhale 用开源的理念去探索开源内容、开源学习和开源方案，赋能人才培养，助力人才成长，建立起人与人，人与知识，人与企业和人与未来的联结。
https://www.modelscope.cn/my/overview 魔塔社区
📂 https://docs.roboflow.com 📂 https://github.com/roboflow/notebooks/tree/main Roboflow empowers developers to build their own computer vision applications, no matter their skillset or experience. We provide all of the tools you need to go from an idea to a robust computer vision model deployed in production.
https://www.alignmentforum.org/
https://artificialanalysis.ai/trends
PEARSON SERIES IN ARTIFICIAL INTELLIGENCE | Stuart Russell and Peter Norvig, Editors
- FORSYTH & PONCE
- Computer Vision: A Modern Approach, 2nd ed.
- GRAHAM ANSI
- Common Lisp
- JURAFSKY & MARTIN
- Speech and Language Processing, 2nd ed.
- NEAPOLITAN
- Learning Bayesian Networks
- RUSSELL & NORVIG
- Artificial Intelligence: A Modern Approach, 4th ed.
🏫 CS50's Introduction to AI with Python | Harvard 🏫 CS188 Introduction to Artificial Intelligence | UC Berkeley
AI and Social Science – Brendan O'Connor
https://www.mit.edu/~amidi/ https://stanford.edu/~shervine/
- 15.003 ― Data Science Tools
- 15.003 - Modeling
- CS 221 ― Artificial Intelligence
- CS 229 ― Machine Learning
- CS 230 ― Deep Learning
https://en.wikipedia.org/wiki/Outline_of_artificial_intelligence The following outline is provided as an overview of and topical guide to artificial intelligence: Artificial intelligence (AI) is intelligence exhibited by machines or software. It is also the name of the scientific field which studies how to create computers and computer software that are capable of intelligent behavior.
http://www.incompleteideas.net/IncIdeas/BitterLesson.html The bitter lesson | rich sutton
[!links] ↗ Philosophy & Its History ↗ Universe, Self-Awareness, and Intelligence
↗ Universe, Self-Awareness, and Intelligence
（感性，理性）X （思考，行动） or（感性，理性）X （认识，实践）
- 感性思考
- 理性思考
- 感性思考指导下的行动
- 理性思考指导下的行动
[!links] ↗ AI Basics & Major Techniques ↗ The Development History of AI
↗ AI4X, AGI (Artificial General Intelligence) & AIGC ↗ AI Embodiment & World Model (WM) ↗ Vision-Language-Action (VLA) Model
🤖 Gemini-3.0 Pro Preview
Artificial Intelligence (AI) refers to the simulation of human intelligence processes by computer systems. It is not a single technology but a broad field combining computer science, data analytics, and mathematics.
At its core, AI is the science of making machines smart—capable of performing tasks that typically require human intelligence, such as:
- Visual Perception: Recognizing faces or objects (Computer Vision).
- Speech Recognition: Transcribing and understanding spoken language.
- Decision-making: Analyzing data to choose the best course of action.
- Translation: Converting language in real-time.
The Three Levels of AI:
- Artificial Narrow Intelligence (ANI): AI designed for a specific task (e.g., Siri, Chess bots, ChatGPT). This is where we are today.
- Artificial General Intelligence (AGI): A hypothetical system that possesses the ability to understand, learn, and apply knowledge across a wide variety of tasks, indistinguishable from a human.
- Artificial Super Intelligence (ASI): A hypothetical system that surpasses human intellect in every field.
📖 Artificial Intelligence: A Modern Approach, 4th ed. RUSSELL & NORVIG Chapter 1
This chapter defines AI and establishes the cultural background against which it has developed. Some of the important points are as follows:
- Different people approach AI with different goals in mind. Two important questions to ask are: Are you concerned with thinking, or behavior? Do you want to model humans, or try to achieve the optimal results?
- According to what we have called the standard model, AI is concerned mainly with rational action. An ideal intelligent agent takes the best possible action in a situation. We study the problem of building agents that are intelligent in this sense.
- Two refinements to this simple idea are needed: first, the ability of any agent, human or otherwise, to choose rational actions is limited by the computational intractability of doing so; second, the concept of a machine that pursues a definite objective needs to be replaced with that of a machine pursuing objectives to benefit humans, but uncertain as to what those objectives are.
- Philosophers (going back to 400 BCE) made AI conceivable by suggesting that the mind is in some ways like a machine, that it operates on knowledge encoded in some internal language, and that thought can be used to choose what actions to take.
- Mathematicians provided the tools to manipulate statements of logical certainty as well as uncertain, probabilistic statements. They also set the groundwork for understanding computation and reasoning about algorithms.
- Economists formalized the problem of making decisions that maximize the expected utility to the decision maker.
- Neuroscientists discovered some facts about how the brain works and the ways in which it is similar to and different from computers.
- Psychologists adopted the idea that humans and animals can be considered information-processing machines. Linguists showed that language use fits into this model.
- Computer engineers provided the ever-more-powerful machines that make AI applications possible, and software engineers made them more usable.
- Control theory deals with designing devices that act optimally on the basis of feedback from the environment. Initially, the mathematical tools of control theory were quite different from those used in AI, but the fields are coming closer together.
- The history of AI has had cycles of success, misplaced optimism, and resulting cutbacks in enthusiasm and funding. There have also been cycles of introducing new, creative approaches and systematically refining the best ones.
- AI has matured considerably compared to its early decades, both theoretically and methodologically. As the problems that AI deals with became more complex, the field moved from Boolean logic to probabilistic reasoning, and from hand-crafted knowledge to machine learning from data. This has led to improvements in the capabilities of real systems and greater integration with other disciplines.
- As AI systems find application in the real world, it has become necessary to consider a wide range of risks and ethical consequences.
- In the longer term, we face the difficult problem of controlling superintelligent AI systems that may evolve in unpredictable ways. Solving this problem seems to necessitate a change in our conception of AI.
📖 Artificial Intelligence: A Modern Approach, 4th ed. RUSSELL & NORVIG Chapter 1
[!links] ↗ AI4X, AGI (Artificial General Intelligence) & AIGC
[!lnks] ↗ Philosophy & Its History
📖 Artificial Intelligence: A Modern Approach, 4th ed. RUSSELL & NORVIG Chapter 1
[!links] ↗ AI Basics & Major Techniques ↗ Mathematical Modeling & Abstraction ↗ Agents & Multi-Agent System ↗ LLM Agents, AI Workflow, & Agentic MLLM ↗ Game Theory & Decision Making in Multi-Agents Environments
📖 Artificial Intelligence: A Modern Approach, 4th ed. RUSSELL & NORVIG Chapter 2
This chapter has been something of a whirlwind tour of AI, which we have conceived of as the science of agent design. The major points to recall are as follows:
- An agent is something that perceives and acts in an environment. The agent function for an agent specifies the action taken by the agent in response to any percept sequence.
- The performance measure evaluates the behavior of the agent in an environment. A rational agent acts so as to maximize the expected value of the performance measure, given the percept sequence it has seen so far.
- A task environment specification includes the performance measure, the external environment, the actuators, and the sensors. In designing an agent, the first step must always be to specify the task environment as fully as possible.
- Task environments vary along several significant dimensions. They can be fully or partially observable, single-agent or multiagent, deterministic or nondeterministic, episodic or sequential, static or dynamic, discrete or continuous, and known or unknown.
- In cases where the performance measure is unknown or hard to specify correctly, there is a significant risk of the agent optimizing the wrong objective. In such cases the agent design should reflect uncertainty about the true objective.
- The agent program implements the agent function. There exists a variety of basic agent program designs reflecting the kind of information made explicit and used in the decision process. The designs vary in efficiency, compactness, and flexibility. The appropriate design of the agent program depends on the nature of the environment.
- Simple reflex agents respond directly to percepts, whereas model-based reflex agents maintain internal state to track aspects of the world that are not evident in the current percept. Goal-based agents act to achieve their goals, and utility-based agents try to maximize their own expected “happiness.”
- All agents can improve their performance through learning.
📖 Artificial Intelligence: A Modern Approach, 4th ed. RUSSELL & NORVIG
[!links] ↗ Mathematical Modeling & Abstraction
📖 Artificial Intelligence: A Modern Approach, 4th ed. RUSSELL & NORVIG
Now that we have a definition of rationality, we are almost ready to think about building rational agents. First, however, we must think about task environments, which are essentially the “problems” to which rational agents are the “solutions.” We begin by showing how to specify a task environment, illustrating the process with a number of examples. We then show that task environments come in a variety of flavors. The nature of the task environment directly affects the appropriate design for the agent program.
In our discussion of the rationality of the simple vacuum-cleaner agent, we had to specify the performance measure, the environment, and the agent’s actuators and sensors. We group all these under the heading of the task environment. For the acronymically minded, we call this the ==PEAS (Performance, Environment, Actuators, Sensors)== description. In designing an agent, the first step must always be to specify the task environment as fully as possible.
...
[!links] ↗ Mathematical Modeling & Abstraction
📖 Artificial Intelligence: A Modern Approach, 4th ed. RUSSELL & NORVIG
The range of task environments that might arise in AI is obviously vast. We can, however, identify a fairly small number of dimensions along which task environments can be categorized. These dimensions determine, to a large extent, the appropriate agent design and the applicability of each of the principal families of techniques for agent implementation. First we list the dimensions, then we analyze several task environments to illustrate the ideas. The definitions here are informal; later chapters provide more precise statements and examples of each kind of environment.
==Fully observable vs. partially observable==: If an agent’s sensors give it access to the complete state of the environment at each point in time, then we say that the task environment is fully observable. A task environment is effectively fully observable if the sensors detect all aspects that are relevant to the choice of action; relevance, in turn, depends on the performance measure. Fully observable environments are convenient because the agent need not maintain any internal state to keep track of the world. An environment might be partially observable because of noisy and inaccurate sensors or because parts of the state are simply missing from the sensor data—for example, a vacuum agent with only a local dirt sensor cannot tell whether there is dirt in other squares, and an automated taxi cannot see what other drivers are thinking. If the agent has no sensors at all then the environment is ==unobservable==. One might think that in such cases the agent’s plight is hopeless, but, as we discuss in Chapter 4, the agent’s goals may still be achievable, sometimes with certainty.
==Single-agent vs. multiagent==: The distinction between single-agent and multiagent environments may seem simple enough. For example, an agent solving a crossword puzzle by itself is clearly in a single-agent environment, whereas an agent playing chess is in a two-agent environment. However, there are some subtle issues. First, we have described how an entity may be viewed as an agent, but we have not explained which entities must be viewed as agents. Does an agent A (the taxi driver for example) have to treat an object B (another vehicle) as an agent, or can it be treated merely as an object behaving according to the laws of physics, analogous to waves at the beach or leaves blowing in the wind? The key distinction is whether B’s behavior is best described as maximizing a performance measure whose value depends on agent A’s behavior.
[!quote] For example, in chess, the opponent entity B is trying to maximize its performance measure, which, by the rules of chess, minimizes agent A’s performance measure. Thus, chess is a ==competitive (adversarial)== multiagent environment. On the other hand, in the taxi-driving environment, avoiding collisions maximizes the performance measure of all agents, so it is a partially ==co-operative== multiagent environment. It is also partially competitive because, for example, only one car can occupy a parking space.
The agent-design problems in multiagent environments are often quite different from those in single-agent environments; for example, communication often emerges as a rational behavior in multiagent environments; in some competitive environments, randomized behavior is rational because it avoids the pitfalls of predictability.
==Deterministic vs. nondeterministic==. If the next state of the environment is completely determined by the current state and the action executed by the agent(s), then we say the environment is deterministic; otherwise, it is nondeterministic. In principle, an agent need not worry about uncertainty in a fully observable, deterministic environment. If the environment is partially observable, however, then it could appear to be nondeterministic.
Most real situations are so complex that it is impossible to keep track of all the unobserved aspects; for practical purposes, they must be treated as nondeterministic. Taxi driving is clearly nondeterministic in this sense, because one can never predict the behavior of traffic exactly; moreover, one’s tires may blow out unexpectedly and one’s engine may seize up without warning. The vacuum world as we described it is deterministic, but variations can include nondeterministic elements such as randomly appearing dirt and an unreliable suction mechanism (Exercise 2.VFIN).
One final note: the word ==stochastic== is used by some as a synonym for “==nondeterministic==,” but we make a distinction between the two terms; we say that a model of the environment is stochastic if it explicitly deals with probabilities (e.g., “there’s a 25% chance of rain tomorrow”) and “nondeterministic” if the possibilities are listed without being quantified (e.g., “there’s a chance of rain tomorrow”).
==Episodic vs. sequential==: In an episodic task environment, the agent’s experience is di- vided into atomic episodes. In each episode the agent receives a percept and then performs a single action. Crucially, the next episode does not depend on the actions taken in previous episodes. Many classification tasks are episodic. For example, an agent that has to spot defective parts on an assembly line bases each decision on the current part, regardless of previous decisions; moreover, the current decision doesn’t affect whether the next part is defective. In sequential environments, on the other hand, the current decision could affect all future decisions.4 Chess and taxi driving are sequential: in both cases, short-term actions can have long-term consequences. Episodic environments are much simpler than sequential environments because the agent does not need to think ahead.
==Static vs. dynamic==: If the environment can change while an agent is deliberating, then we say the environment is dynamic for that agent; otherwise, it is static. Static environments are easy to deal with because the agent need not keep looking at the world while it is deciding on an action, nor need it worry about the passage of time. Dynamic environments, on the other hand, are continuously asking the agent what it wants to do; if it hasn’t decided yet, that counts as deciding to do nothing. If the environment itself does not change with the passage of time but the agent’s performance score does, then we say the environment is ==semidynamic==. Taxi driving is clearly dynamic: the other cars and the taxi itself keep moving while the driving algorithm dithers about what to do next. Chess, when played with a clock, is semidynamic. Crossword puzzles are static.
==Discrete vs. continuous==: The discrete/continuous distinction applies to the state of the environment, to the way time is handled, and to the percepts and actions of the agent. For example, the chess environment has a finite number of distinct states (excluding the clock). Chess also has a discrete set of percepts and actions. Taxi driving is a continuous-state and continuous-time problem: the speed and location of the taxi and of the other vehicles sweep through a range of continuous values and do so smoothly over time. Taxi-driving actions are also continuous (steering angles, etc.). Input from digital cameras is discrete, strictly speaking, but is typically treated as representing continuously varying intensities and locations.
==Known vs. unknown==: Strictly speaking, this distinction refers not to the environment itself but to the agent’s (or designer’s) state of knowledge about the “laws of physics” of the environment. In a known environment, the outcomes (or outcome probabilities if the environment is nondeterministic) for all actions are given. Obviously, if the environment is unknown, the agent will have to learn how it works in order to make good decisions.
The distinction between known and unknown environments is not the same as the one between fully and partially observable environments. It is quite possible for a known environment to be partially observable—for example, in solitaire card games, I know the rules but am still unable to see the cards that have not yet been turned over. Conversely, an unknown environment can be fully observable—in a new video game, the screen may show the entire game state but I still don’t know what the buttons do until I try them.
As noted on page 57, the performance measure itself may be unknown, either because the designer is not sure how to write it down correctly or because the ultimate user—whose preferences matter—is not known. For example, a taxi driver usually won’t know whether a new passenger prefers a leisurely or speedy journey, a cautious or aggressive driving style. A virtual personal assistant starts out knowing nothing about the personal preferences of its new owner. In such cases, the agent may learn more about the performance measure based on further interactions with the designer or user. This, in turn, suggests that the task environment is necessarily viewed as a multiagent environment.
[!quote] The hardest case is partially observable, multiagent, nondeterministic, sequential, dynamic, continuous, and unknown. Taxi driving is hard in all these senses, except that the driver’s environment is mostly known. Driving a rented car in a new country with unfamiliar geography, different traffic laws, and nervous passengers is a lot more exciting.
The code repository associated with this book (🔗 aima.cs.berkeley.edu) includes multiple environment implementations, together with a general-purpose environment simulator for evaluating an agent’s performance. Experiments are often carried out not for a single environment but for many environments drawn from an environment class. For example, to evaluate a taxi driver in simulated traffic, we would want to run many simulations with different traffic, lighting, and weather conditions. We are then interested in the agent’s average performance over the environment class.
[!links] ↗ AI Basics & Major Techniques
📖 Artificial Intelligence: A Modern Approach, 4th ed. RUSSELL & NORVIG
[!links] ↗ Mathematical Modeling & Abstraction ↗ AI Basics & Major Techniques
1️⃣ ==Explicit, Fixed Modeling of Environment==
[!links] ↗ Problem Solving & Search-Based Methods ↗ Automated Planning and Scheduling (APS) & AI Planning
📖 Artificial Intelligence: A Modern Approach, 4th ed. RUSSELL & NORVIG Chapter 3
...
It takes about a thousand pages to begin to answer that question properly, but here we want to draw the reader’s attention to some basic distinctions among the various ways that the components can represent the environment that the agent inhabits. Roughly speaking, ==we can place the representations along an axis of increasing complexity and expressive power—atomic, factored, and structured==. To illustrate these ideas, it helps to consider a particular agent component, such as the one that deals with “What my actions do.” This component describes the changes that might occur in the environment as the result of taking an action, and Figure 2.16 provides schematic depictions of how those transitions might be represented.
[!quote] Axis of Agent's Representation of Environment
As we mentioned earlier, the axis along which atomic, factored, and structured representations lie is the axis of increasing ==expressiveness==. Roughly speaking, a mor expressive representation can capture, at least as concisely, everything a less expressive one can capture, plus some more. Often, the more expressive language is much more concise; for example, the rules of chess can be written in a page or two of a structured-representation language such as first-order logic but require thousands of pages when written in a factored-representation language such as propositional logic and around 1038 pages when written in an atomic language such as that of finite-state automata. On the other hand, reasoning and learning become more complex as the expressive power of the representation increases. To gain the benefits of expressive representations while avoiding their drawbacks, intelligent systems for the real world may need to operate at all points along the axis simultaneously.
Another axis for representation involves ==the mapping of concepts to locations in physical memory==, whether in a computer or in a brain. If there is a one-to-one mapping between concepts and memory locations, we call that a localist representation. On the other hand if the representation of a concept is spread over many memory locations, and each memory location is employed as part of the representation of multiple different concepts, we call that a distributed representation. Distributed representations are more robust against noise and information loss. With a localist representation, the mapping from concept to memory location is arbitrary, and if a transmission error garbles a few bits, we might confuse Truck with the unrelated concept Truce. But with a distributed representation, you can think of each concept representing a point in multidimensional space, and if you garble a few bits you move to a nearby point in that space, which will have similar meaning.
Atomic Representation In an atomic representation ==each state of the world is indivisible==—it has no internal structure. Consider the task of finding a driving route from one end of a country to the other via some sequence of cities (we address this problem in Figure 3.1 on page 82). For the purposes of solving this problem, it may suffice to reduce the state of the world to just the name of the city we are in—a single atom of knowledge, a “black box” whose only discernible property is that of being identical to or different from another black box.
The standard algorithms underlying search and game-playing (Chapters 3, 4, and 6), hidden Markov models (Chapter 14), and Markov decision processes (Chapter 16) all work with atomic representations.
[!links] ↗ Systematic & Combinatorial Search (Classical Search) ↗ Games & Search in Multi-Agents Environment ↗ Markov Decision Processes (MDP) & Stochastic Dynamic Program
Factored Representation A factored representation splits up each state into ==a fixed set of variables or attributes==, each of which can have a value. Consider a higher-fidelity description for the same driving problem, where we need to be concerned with more than just atomic location in one city or another; we might need to pay attention to how much gas is in the tank, our current GPS coordinates, whether or not the oil warning light is working, how much money we have for tolls, what station is on the radio, and so on. While two different atomic states have nothing in common—they are just different black boxes—two different factored states can share some attributes (such as being at some particular GPS location) and not others (such as having lots of gas or having no gas); this makes it much easier to work out how to turn one state into another.
Many important areas of AI are based on factored representations, including constraint satisfaction algorithms (Chapter 5), propositional logic (Chapter 7), planning (Chapter 11), Bayesian networks (Chapters 12, 13, 14, 15, and 18), and various machine learning algorithms.
[!links] ↗ Constraint Based Search & Constraint Programming & Constraint Satisfaction ↗ Constraint Satisfaction Problems (CSPs)
↗ Constraint Solving & Theorem Proving
- ↗ Formal Verifications & Constraint Solvers (Proof Assistants)
- ↗ SAT (Boolean Satisfiability Problem) Solvers
- ↗ SMT (Satisfiability Modulo Theory) Solvers ↗ (Formal) Model Checking
↗ Automated Planning and Scheduling (APS) & AI Planning ↗ Statistical (Data-Driven) Learning & Machine Learning (ML)
Tip
https://chatgpt.com/share/6994e9a5-ffd0-8010-b952-1355963ff237 Symbolic Execution & Constraint Solving 🆚 AI Searching #symbolic_execution #constraint_solving #AI #combinatorial_search
👉 Classical search enumerates states.
👉 SAT planning and symbolic model checking manipulate logical formulas describing many states at once.
That’s the “compression”.
| Field | What is being searched |
|---|---|
| Heuristic search | explicit states |
| SAT planning | logical encodings of plans |
| Symbolic execution | path constraints |
| Model checking | reachable state formulas |
The real reason symbolic methods scale is:
They move complexity from combinatorial enumeration into logical inference.
Atomic search explores: branching_factor^depth states explicitly.
SAT planning represents: branching_factor × depth variables and constraints.
Key insight: 👉 Logical constraints eliminate impossible combinations before search. The SAT solver prunes entire regions of the state space via:
- unit propagation
- clause learning
- conflict analysis
This is much stronger than expanding nodes one by one.
Structured Representation
For many purposes, we need to ==understand the world as having things in it that are related to each other==, not just variables with values. For example, we might notice that a large truck ahead of us is reversing into the driveway of a dairy farm, but a loose cow is blocking the truck’s path. A factored representation is unlikely to be pre-equipped with the attribute TruckAheadBackingIntoDairyFarmDrivewayBlockedByLooseCow with value true or false. Instead, we would need a structured representation, in which objects such as cows and trucks and their various and varying relationships can be described explicitly (see Figure 2.16(c)).
Structured representations underlie relational databases and first-order logic (Chapters 8, 9, and 10), first-order probability models (Chapter 18), and much of natural language understanding (Chapters 24 and 25). ==In fact, much of what humans express in natural language concerns objects and their relationships.==
[!links] ↗ Mathematical Logic (Foundations of Mathematics) ↗ Formal System, Formal Logics, and Its Semantics ↗ AI Basics & Major Techniques
- ↗ Knowledge Representation (Syntax Level) and Reasoning (KRR)
- ↗ Statistical (Data-Driven) Learning & Machine Learning (ML)
↗ Natural Language Processing (NLP) & Computational Linguistics
2️⃣ ==Implicit, Learnt Modeling of Environment==
[!links] ↗ Statistical (Data-Driven) Learning & Machine Learning (ML) ↗ LLM (Large Language Model) ↗ AI Embodiment & World Model (WM)
🤖 GPT-5.2 https://chatgpt.com/share/6994e9a5-ffd0-8010-b952-1355963ff237
LLMs don’t manipulate explicit symbols — they manipulate vectors that approximate symbolic structure. So the spectrum becomes: Atomic → Factored → Structured → Neural / Distributed
From a formal methods perspective:
| System | Representation | Reasoning type |
|---|---|---|
| A* | atomic states | explicit search |
| SAT planner | factored logic | symbolic reasoning |
| LLM | distributed embeddings | statistical inference |
Tip
🤖 GPT-5.2 Q. why many researchers think LLMs behave like implicit planners in latent space, and how “Tree-of-Thoughts” or “Self-Consistency” basically re-introduce classical search on top of LLMs — which connects directly back to A*, SAT planning, and symbolic execution.
A. This is one of the most interesting ideas in modern AI — and it connects directly to the search/planning concepts you’ve been studying. I’ll explain it in a precise way, not hype 🙂. We’ll break it into three parts:
- Why people say LLMs are implicit planners
- What “latent space planning” really means
- How Tree-of-Thoughts and Self-Consistency literally re-add classical search (A*, SAT-style ideas, symbolic execution parallels)
[!links] ↗ AI Basics & Major Techniques ↗ Knowledge Representation (Syntax Level) and Reasoning (KRR)
1️⃣ ==Search-Based AI==
[!links] ↗ Problem Solving & Search-Based Methods ↗ Automated Planning and Scheduling (APS) & AI Planning
🤖 GPT-5.2 https://chatgpt.com/share/6994e9a5-ffd0-8010-b952-1355963ff237
#AI_planning #combinatorial_search #optimization
| Agent type | How it represents the world | Main technique /Algorithms /Reasoning |
|---|---|---|
| Problem-solving agent | State = black box (“atomic”) | search over states |
| Planning agent | State = factored /structured facts | search over logical structures / plans (reasoning) |
🧩 So is planning = search? Here’s the subtle truth: 👉 Planning still uses search internally. But:
- Problem-solving search = search over states
- Planning search = search over logical structures / plans So you’re not wrong: planning ⊂ search But textbooks separate them because:
- the representation changes
- the algorithms change dramatically
Search difficulty is not just about algorithms. It’s about how much structure you expose to the algorithm. Same domain:
Atomic view → brute-force search
Structured view → reasoning + pruning
🎯 One-sentence intuition
- 👉 Problem-solving agent: “I don’t understand the world — I just search through states.”
- 👉 Planning agent: “I understand objects and relationships — I reason about actions before searching.”
So:
The difference between problem-solving agents and planning agents is not whether they plan — it’s whether they understand the internal structure of states.
| Representation | Typical algorithms |
|---|---|
| Atomic | A*, BFS, DFS |
| Factored | Graphplan, SATPlan, heuristic planners |
| Structured | STRIPS, HTN planning, logical inference |
2️⃣ ==Learn-Based AI==
[!links] ↗ Statistical (Data-Driven) Learning & Machine Learning (ML) ↗ Knowledge Representation (Syntax Level) and Reasoning (KRR) ↗ Artificial Neural Networks (ANN) & Deep Learning Methods
🤖 GPT-5.2 https://chatgpt.com/share/6994e9a5-ffd0-8010-b952-1355963ff237
Here’s a version you can safely use:
| Agent type | Internal representation of the world | Main technique | What search looks like |
|---|---|---|---|
| Problem-solving agent | Atomic states (opaque nodes) | Graph search (A*, BFS) | Explicit state expansion |
| Planning agent | Factored / structured symbols (variables, predicates) | Logical planning, SAT, reasoning | Symbolic search over constraints |
| LLM agent | Distributed latent representations (embeddings) | Neural generation + learned heuristics | Implicit trajectory; optional external search (ToT, self-consistency) |
| If you want an even cleaner academic phrasing |
| Paradigm | Representation | Reasoning style |
|---|---|---|
| Classical search | Atomic states | Explicit combinatorial search |
| Symbolic planning / verification | Factored logical variables | Constraint solving / inference |
| Neural agents (LLMs) | Distributed latent vectors | Probabilistic sequence prediction |
🤖 GPT-5.2 https://chatgpt.com/share/6994e9a5-ffd0-8010-b952-1355963ff237
| Paradigm | Environment knowledge | Learning? | Core idea | Typical algorithms |
|---|---|---|---|---|
| Search / Planning | Known model | ❌ No learning | Compute a plan using an explicit model before acting | A*, STRIPS, SAT planning |
| Online search (unknown env) | Unknown model | ❌ or minimal | Act while discovering environment structure | LRTA*, real-time heuristic search |
| Learning in online search | Unknown initially | ✅ Partial | Improve heuristics or models through experience | Adaptive A*, heuristic learning |
| Reinforcement Learning (RL) | Usually unknown | ✅ Core mechanism | Learn behavior via reward-driven interaction | Q-learning, Policy Gradient |
| Machine Learning (general) | Not required | ✅ Core mechanism | Learn predictive patterns from data (not necessarily sequential decisions) | SVM, decision trees, regression |
| Deep Learning | Not required | ✅ Core mechanism | Learn hierarchical representations with neural networks | CNNs, Transformers, MLPs |
| All of these are about an agent choosing actions over time. | ||||
| The main differences come from two questions: |
1) Does the agent know the environment model?2) Does the agent learn from experience?If you organize things along those axes, the relationships become very clear.
Here’s a continuum that many researchers implicitly use:
Known model
↓
Search / Planning
↓
Online Planning (unknown map)
↓
Learning to Improve Planning
↓
Model-Based RL
↓
Model-Free RL
The deeper you go:
less explicit search
more statistical learning
👉 Search / Planning / RL are mainly about sequential decision making.
👉 Machine Learning and Deep Learning are mainly about function approximation.
So ML/DL are not just “another row” — they are more like a horizontal capability that can appear inside other paradigms.
Even deeper insight (since you’re thinking at research level) The real unifying view is not:
search vs learningbut:optimization over different spaces
| Paradigm | What is optimized |
|---|---|
| Search | action sequences / plans |
| Planning | logical constraints |
| RL | expected reward |
| LLM training | prediction loss |
| Different tools, same underlying idea. |
[!links] ↗ Philosophy & Its History
↗ Philosophy, Ethics, and Risks of AI ↗ Trust-worthy AI & LLM Safety and Security
📖 Artificial Intelligence: A Modern Approach, 4th ed. RUSSELL & NORVIG Chapter 29
In which we consider the big questions around the meaning of AI, how we can ethically develop and apply it, and how we can keep it safe.
Philosophers have been asking big questions for a long time: How do minds work? Is it possible for machines to act intelligently in the way that people do? Would such machines have real, conscious minds?
To these, we add new ones: What are the ethical implications of intelligent machines in day-to-day use? Should machines be allowed to decide to kill humans? Can algorithms be fair and unbiased? What will humans do if machines can do all kinds of work? And how do we control machines that may become more intelligent than us?
- The Limits of AI
- The argument from informality
- The argument from disability
- The mathematical objection
- Measuring AI
- Can Machines Really Think?
- The Chinese room
- Consciousness and qualia
- The Ethics of AI
- Lethal autonomous weapons
- Surveillance, security, and privacy
- Fairness and bias
- Trust and transparency
- The future of work
- Robot rights
- AI Safety
This chapter has addressed the following issues:
- Philosophers use the term weak AI for the hypothesis that machines could possibly behave intelligently, and strong AI for the hypothesis that such machines would count as having actual minds (as opposed to simulated minds).
- Alan Turing rejected the question “Can machines think?” and replaced it with a behavioral test. He anticipated many objections to the possibility of thinking machines. Few AI researchers pay attention to the Turing test, preferring to concentrate on their systems’ performance on practical tasks, rather than the ability to imitate humans.
- Consciousness remains a mystery.
- AI is a powerful technology, and as such it poses potential dangers, through lethal autonomous weapons, security and privacy breaches, unintended side effects, unintentional errors, and malignant misuse. Those who work with AI technology have an ethical imperative to responsibly reduce those dangers.
- AI systems must be able to demonstrate they are fair, trustworthy, and transparent.
- There are multiple aspects of fairness, and it is impossible to maximize all of them at once. So a first step is to decide what counts as fair.
- Automation is already changing the way people work. As a society, we will have to deal with these changes.
📖 Artificial Intelligence: A Modern Approach, 4th ed. RUSSELL & NORVIG Chapter 29
In Chapter 2, we decided to view AI as the task of designing approximately rational agents. A variety of different agent designs were considered, ranging from reflex agents to knowledge-based decision-theoretic agents to deep learning agents using reinforcement learning. There is also variety in the component technologies from which these designs are assembled: logical, probabilistic, or neural reasoning; atomic, factored, or structured representations of states; various learning algorithms from various types of data; sensors and actuators to interact with the world. Finally, we have seen a variety of applications, in medicine, finance, transportation, communication, and other fields. There has been progress on all these fronts, both in our scientific understanding and in our technological capabilities.
Most experts are optimistic about continued progress; as we saw on page 46, the median estimate is for approximately human-level AI across a broad variety of tasks somewhere in the next 50 to 100 years. Within the next decade, AI is predicted to add trillions of dollars to the economy each year. But as we also saw, there are some critics who think general AI is centuries off, and there are numerous ethical concerns about the fairness, equity, and lethality of AI. In this chapter, we ask: where are we headed and what remains to be done? We do that by asking whether we have the right components, architectures, and goals to make AI a successful technology that delivers benefits to the world.
📖 Artificial Intelligence: A Modern Approach, 4th ed. RUSSELL & NORVIG Chapter 29.1
[!links] ↗ AI Embodiment & World Model (WM)
📖 Artificial Intelligence: A Modern Approach, 4th ed. RUSSELL & NORVIG Chapter 29.2
It is natural to ask, “Which of the agent architectures in Chapter 2 should an agent use?” The answer is, “All of them!” Reflex responses are needed for situations in which time is of the essence, whereas knowledge-based deliberation allows the agent to plan ahead. Learning is convenient when we have lots of data, and necessary when the environment is changing, or when human designers have insufficient knowledge of the domain.
AI has long had a split between symbolic systems (based on logical and probabilistic inference) and connectionist systems (based on loss minimization over a large number of uninterpreted parameters). A continuing challenge for AI is to bring these two together, to capture the best of both. Symbolic systems allow us to string together long chains of reasoning and to take advantage of the expressive power of structured representations, while connectionist systems can recognize patterns even in the face of noisy data. One line of research aims to combine probabilistic programming with deep learning, although as yet the various proposals are limited in the extent to which the approaches are truly merged.
Agents also need ways to control their own deliberations. They must be able to use the available time well, and cease deliberating when action is demanded. For example, a taxi-driving agent that sees an accident ahead must decide in a split second whether to brake or swerve. It should also spend that split second thinking about the most important questions, such as whether the lanes to the left and right are clear and whether there is a large truck close behind, rather than worrying about where to pick up the next passenger. These issues are usually studied under the heading of real-time AI. As AI systems move into more complex domains, all problems will become real-time, because the agent will never have long enough to solve the decision problem exactly.
Clearly, there is a pressing need for general methods of controlling deliberation, rather than specific recipes for what to think about in each situation. The first useful idea is the anytime algorithms (Dean and Boddy, 1988; Horvitz, 1987): an algorithm whose output quality improves gradually over time, so that it has a reasonable decision ready whenever it is interrupted. Examples of anytime algorithms include iterative deepening in game-tree search and MCMC in Bayesian networks.
The second technique for controlling deliberation is decision-theoretic metareasoning (Russell and Wefald, 1989; Horvitz and Breese, 1996; Hay et al., 2012). This method, which was mentioned briefly in Sections 3.6.5 and 6.7, applies the theory of information value (Chapter 15) to the selection of individual computations (Section 3.6.5). The value of a computation depends on both its cost (in terms of delaying action) and its benefits (in terms of improved decision quality).
Metareasoning techniques can be used to design better search algorithms and to guarantee that the algorithms have the anytime property. Monte Carlo tree search is one example: the choice of leaf node at which to begin the next playout is made by an approximately rational metalevel decision derived from bandit theory.
Metareasoning is more expensive than reflex action, of course, but compilation methods can be applied so that the overhead is small compared to the costs of the computations being controlled. Metalevel reinforcement learning may provide another way to acquire effective policies for controlling deliberation: in essence, computations that lead to better decisions are reinforced, while those that turn out to have no effect are penalized. This approach avoids the myopia problems of the simple value-of-information calculation.
Metareasoning is one specific example of a reflective architecture—that is, an architecture that enables deliberation about the computational entities and actions occurring within the architecture itself. A theoretical foundation for reflective architectures can be built by defining a joint state space composed from the environment state and the computational state of the agent itself. Decision-making and learning algorithms can be designed that operate over this joint state space and thereby serve to implement and improve the agent’s computational activities. Eventually, we expect task-specific algorithms such as alpha–beta search, regression planning, and variable elimination to disappear from AI systems, to be replaced by general methods that direct the agent’s computations toward the efficient generation of high-quality decisions.
Metareasoning and reflection (and many other efficiency-related architectural and algorithmic devices explored in this book) are necessary because making decisions is hard. Ever since computers were invented, their blinding speed has led people to overestimate their ability to overcome complexity, or, equivalently, to underestimate what complexity really means. The truly gargantuan power of today’s machines tempts one to think that we could bypass all the clever devices and rely more on brute force. So let’s try to counteract this tendency. We begin with what physicists believe to be the speed of the ultimate 1kg computing device: about 1051 operations per second, or a billion trillion trillion times faster than the fastest supercomputer as of 2020 (Lloyd, 2000).1 Then we propose a simple task: enumerating strings of English words, much as Borges proposed in The Library of Babel. Borges stipulated books of 410 pages. Would that be feasible? Not quite. In fact, the computer running for a year could enumerate only the 11-word strings.
Now consider the fact that a detailed plan for a human life consists of (very roughly) twenty trillion potential muscle actuations (Russell, 2019), and you begin to see the scale of the problem. A computer that is a billion trillion trillion times more powerful than the human brain is much further from being rational than a slug is from overtaking the starship Enterprise traveling at warp nine.
With these considerations in mind, it seems that the goal of building rational agents is perhaps a little too ambitious. Rather than aiming for something that cannot possibly exist, we should consider a different normative target—one that necessarily exists. Recall from Chapter 2 the following simple idea:
agent= architecture + program.
Now fix the agent architecture (the underlying machine capabilities, perhaps with a fixed software layer on top) and allow the agent program to vary over all possible programs that the architecture can support. In any given task environment, one of these programs (or an equivalence class of them) delivers the best possible performance—perhaps not close to perfect rationality, but still better than any other agent program. We say that this program satisfies the criterion of bounded optimality. Clearly it exists, and clearly it constitutes a desirable goal. The trick is finding it, or something close to it.
For some elementary classes of agent programs in simple real-time environments, it is possible to identify bounded-optimal agent programs (Etzioni, 1989; Russell and Subramanian, 1995). The success of Monte Carlo tree search has revived interest in metalevel decision making, and there is reason to hope that bounded optimality within more complex families of agent programs can be achieved by techniques such as metalevel reinforcement learning. It should also be possible to develop a constructive theory of architecture, beginning with theorems on the bounded optimality of suitable methods of combining different bounded-optimal components such as reflex and action–value systems.
↗ AI4X, AGI (Artificial General Intelligence) & AIGC
