The Current AI Landscape

Unit 8: Machine Learning Foundations (Capstone) — Section 8.5

The algorithms you studied this week — decision trees, k-NN, Naive Bayes — still power thousands of production systems today. But AI has also been through a revolution over the past decade, producing capabilities that would have seemed implausible when the foundational techniques in this course were developed. This section traces that evolution, introduces the technologies reshaping every industry right now, and examines the ethical responsibilities that come with deploying powerful AI systems.

Survey the evolution of AI from symbolic methods to the current era.

AI Trends and Evolution

The Arc of AI Development

The AI Timeline:

  • 1950s—​1980s: Symbolic AI, expert systems, search algorithms — the methods you studied in Units 3—​6.

  • 1990s—​2000s: Classical machine learning — decision trees, support vector machines, k-NN — the foundations of Unit 8.

  • 2012: AlexNet wins the ImageNet competition by a large margin. Deep convolutional networks become the dominant approach for computer vision. The deep learning era begins.

  • 2016: AlphaGo defeats world Go champion Lee Sedol. A landmark for reinforcement learning and hybrid symbolic/ML systems.

  • 2017: The Transformer architecture is introduced ("Attention Is All You Need"). This paper makes large language models possible.

  • 2020: GPT-3 (175 billion parameters) demonstrates few-shot learning — solving new tasks from just a few examples in the prompt.

  • 2022: ChatGPT (GPT-3.5 fine-tuned with RLHF) reaches 100 million users in two months. Generative AI enters mainstream public awareness.

  • 2023—​2024: GPT-4, Claude 2/3, Gemini, and Llama 2 establish multimodal, long-context reasoning as standard LLM capabilities. Open-source LLMs become viable for deployment.

  • 2025—​2026: Reasoning models (o1-series, DeepSeek-R1) make chain-of-thought planning a standard capability. AI agents that can use tools, browse the web, and write and execute code become commercially deployed. Regulatory frameworks (EU AI Act enforcement begins 2025) reshape how high-risk AI systems are developed and audited.

Deep Learning: Neural Networks at Scale

The algorithms in Sections 8.1—​8.3 use hand-crafted features and relatively shallow models. Deep learning replaces hand-crafted features with automatic feature discovery through many layers of learned transformations.

Deep Learning

A class of machine learning methods that use multi-layer neural networks ("deep" networks) to automatically learn hierarchical feature representations from raw data. The "depth" refers to the number of processing layers between input and output. Deep learning enabled breakthroughs in computer vision, speech recognition, and natural language processing.

How it connects to what you already know:

  • Training a neural network is gradient descent hill climbing (Unit 4) applied to the space of billions of parameters.

  • Each layer learns features; the final layer makes probabilistic predictions (Unit 7).

  • The whole system remains a function from inputs to outputs — just one that was discovered from data rather than programmed.

Key achievements of deep learning include: AlphaFold (predicting protein 3D structure from amino acid sequences, 2021), self-driving car perception, real-time speech-to-text, and medical imaging diagnosis at radiologist accuracy.

Reinforcement Learning: Learning by Doing

Reinforcement learning (RL) trains agents through trial and error, not labeled examples. An RL agent takes actions, observes rewards and penalties, and gradually learns a policy that maximizes long-term cumulative reward.

Reinforcement Learning

A machine learning paradigm where an agent learns by interacting with an environment. The agent receives numerical rewards for desirable actions and penalties for undesirable ones. Over many trials it discovers a policy — a mapping from states to actions — that maximizes total reward. Unlike supervised learning, no labeled examples are provided; the agent must explore to learn.

Key RL milestones: AlphaGo (2016), OpenAI Five defeating professional Dota 2 players (2018), AlphaCode generating competitive programming solutions (2022), and RLHF (Reinforcement Learning from Human Feedback) used to align ChatGPT with human preferences.

Large Language Models and the Transformer Revolution

The most consequential development in AI since the deep learning wave is the transformer architecture, introduced in 2017. Transformers use a mechanism called self-attention that allows a model to relate every word in a long document to every other word in parallel, making it possible to train on very long sequences of text.

Scaling transformers to billions of parameters and training them on vast text corpora — essentially a snapshot of the written internet — produces large language models (LLMs).

Large Language Model (LLM)

A transformer-based neural network trained on massive text corpora that can generate, translate, summarize, and reason about natural language. LLMs like GPT-4, Claude, and Gemini exhibit emergent capabilities — abilities not explicitly programmed, arising from scale and training. They are the foundation of most modern AI assistants and coding tools.

Transformer

A neural network architecture introduced in 2017 that uses self-attention to model relationships between all positions in an input sequence simultaneously. Transformers replaced recurrent architectures for NLP and are now used for text, images, audio, video, and protein sequences.

What LLMs Can Do (2025—​2026):

  • Write and debug code from a natural language description (GitHub Copilot, Cursor)

  • Analyze documents, images, and spreadsheets in a single conversation

  • Reason step-by-step through multi-part math and logic problems (o1-series reasoning models)

  • Use external tools — calculators, search engines, code interpreters — as part of a task

  • Act as autonomous agents that plan multi-step tasks and take actions in the real world

What was considered research in 2022 is now commercially deployed infrastructure for millions of developers and knowledge workers.

Generative AI: Creating New Content

Generative AI refers to models that create new content — images, text, audio, video, or code — rather than simply classifying or predicting.

Generative AI

AI systems capable of producing new, original content — images, text, audio, video, or code — that was not explicitly stored in the system. Examples include DALL-E, Stable Diffusion (images), GPT-4 and Claude (text), Sora (video), and GitHub Copilot (code).

Key generative AI milestones through 2026:

  • DALL-E 3 and Midjourney 6 (2023—​2024): photorealistic image generation from text prompts

  • Sora (2024): text-to-video generation at high quality

  • Stable Diffusion 3.5 and Flux: open-source image generation models deployed locally

  • Music generation (Udio, Suno): full song production from text descriptions

  • Video dubbing and voice cloning at near-human quality

The democratization of creative tools is one of the most significant social impacts of this wave of AI.

Responsible AI: The Ongoing Challenge

Capability advances do not automatically produce trustworthy systems. The Montreal AI Ethics Institute’s State of AI Ethics reports consistently document the gap between what AI can do and what it does safely, fairly, and accountably.

Active Ethical Challenges in AI (2025—​2026):

  • Bias and Fairness: Models trained on historical data can perpetuate and amplify historical discrimination. Facial recognition systems continue to show higher error rates for darker-skinned individuals. Hiring algorithms can reproduce demographic disparities from past hiring patterns.

  • Explainability: Deep neural networks and LLMs remain difficult to interpret. In healthcare, finance, and criminal justice, "the model said so" is not an acceptable justification. Explainable AI (XAI) is an active research field but remains unsolved for large models.

  • Misinformation and Deepfakes: Generative AI enables synthetic media that is increasingly indistinguishable from real recordings. Election interference, fraud, and harassment using AI-generated content are documented harms.

  • Energy and Environment: Training large models requires significant computational resources and energy. The environmental cost of AI infrastructure is growing rapidly.

  • Safety and Alignment: Ensuring that AI systems pursue intended goals — and not harmful proxies — remains an open research problem, especially as systems become more autonomous.

Responsible AI

A set of principles and practices for developing and deploying AI systems that are safe, fair, transparent, accountable, and aligned with human values. Responsible AI addresses the full lifecycle: design, training data, evaluation, deployment, and monitoring. The EU AI Act (2025) establishes legally binding requirements for high-risk AI systems in Europe.

The Montreal AI Ethics Institute publishes annual State of AI Ethics reports (CC BY 4.0) that survey responsible AI research and real-world harms: montrealethics.ai/state-of-ai-ethics

The NIST AI Risk Management Framework (public domain) provides practical guidance for organizations developing AI systems: nist.gov/artificial-intelligence

Open Problems: What Remains Unsolved

Despite remarkable progress, fundamental challenges remain.

  • Common sense reasoning: LLMs still make errors that reveal a lack of basic world knowledge (objects fall, you cannot be in two places at once, people are generally unhappy when harmed). Current models learn correlations from text but do not model the physical and social world they describe.

  • Causal reasoning: Current AI finds statistical correlations but does not reliably infer causation — the difference between "ice cream sales are correlated with drowning deaths" and "hot weather causes both."

  • Transfer learning: Humans learn to drive a car and immediately understand something about driving a truck. AI models trained on one domain often fail dramatically when the domain shifts slightly.

  • Data efficiency: A child can recognize a zebra from one or two examples. Most current ML algorithms need thousands. Few-shot and zero-shot learning are active research areas.

You have now seen AI’s greatest successes (image recognition, protein folding, language generation) and its persistent blind spots (common sense, causality, fairness).

Consider a specific real-world problem — your future career field, a community you are part of, or a challenge you care about.

  • Where could AI be genuinely helpful?

  • What are the failure modes that would be most damaging?

  • What safeguards would you want in place before deploying an AI system in that context?

There are no single right answers here. This is the conversation the field — and society — needs to keep having.

Key Takeaways

AI has progressed from symbolic expert systems to deep learning to transformer-based LLMs, each wave building on the theoretical foundations of the previous one. Large language models and generative AI represent the current frontier, with capabilities expanding rapidly. Yet fundamental challenges — common sense reasoning, causality, fairness, explainability — remain unsolved. Responsible AI development requires addressing these challenges intentionally, not hoping they resolve themselves. The foundations you have learned this semester — search, logic, probability, learning — are the bedrock on which all of these advances rest.

Test your understanding of current AI trends and ethical considerations.

Next: 8.L Machine Learning Classifier Lab →

Ethics content adapted from the Montreal AI Ethics Institute, licensed under CC BY 4.0.

AI governance content incorporates material from the NIST AI Risk Management Framework (AI RMF 1.0), a U.S. Government work in the public domain.

Based on the UC Berkeley CS 188 Online Textbook by Nikhil Sharma, Josh Hug, Jacky Liang, and Henry Zhu, licensed under CC BY-SA 4.0.

This work is licensed under CC BY-SA 4.0.