# Artificial Intelligence: From Turing to Transformers

Canonical URL: https://shipslides.com/d/technology-artificial-intelligence
Raw viewer URL: https://content.shipslides.com/d/technology-artificial-intelligence/raw
Category: Technology
Slides: 18
Updated: 2026-05-17T20:49:53.667Z
Tags: technology, artificial, intelligence

## Summary

From Turing's 1950 question "Can machines think?" to modern large language models -- the history, science, breakthroughs, and implications of humanity's most transformative technology. Key sections include: Artificial Intelligence; What Is Artificial Intelligence?; Historical Timeline; Neural Networks: How Machines Learn; The Transformer Revolution; Large Language Models; Computer Vision; Reinforcement Learning; AI in Science and Discovery; The Economics of AI.

## Slide Outline

1. Artificial Intelligence
2. What Is Artificial Intelligence?
3. Historical Timeline
4. Neural Networks: How Machines Learn
5. The Transformer Revolution
6. Large Language Models
7. Computer Vision
8. Reinforcement Learning
9. AI in Science and Discovery
10. The Economics of AI
11. AI Safety and Alignment
12. AI Ethics and Bias
13. Generative AI: Images, Video, Audio
14. AI and Code
15. The Road to AGI
16. AI and Society
17. Key Takeaways

## Slide Transcript

### Slide 1: Artificial Intelligence

- Technology • AI
- From Turing's 1950 question "Can machines think?" to modern large language models -- the history, science, breakthroughs, and implications of humanity's most transformative technology.
- Machine LearningDeep LearningNLPNeural NetworksEthics

### Slide 2: What Is Artificial Intelligence?

- AI is the science and engineering of creating machines that perform tasks requiring intelligence when done by humans. The field encompasses reasoning, learning, perception, language understanding, and decision-making.
- Types of AI
- Narrow AI (ANI): Excels at specific tasks. All current AI is narrow -- chess engines, image classifiers, language models. Superhuman in narrow domains but cannot generalize.
- General AI (AGI): Human-level intelligence across all cognitive tasks. Does not yet exist. Major research goal and source of debate about timeline and approach.
- Superintelligent AI (ASI): Hypothetical intelligence surpassing all human capability. Central to alignment and safety research. Nick Bostrom's "Superintelligence" (2014) catalyzed concern.
- "Artificial intelligence is the new electricity."-- Andrew Ng, Stanford/Google Brain/Coursera, 2017
- Key Subfields
- Machine Learning: Systems that learn from data without explicit programming
- Deep Learning: Neural networks with many layers learning hierarchical representations
- Natural Language Processing: Understanding and generating human language
- Computer Vision: Interpreting visual information from the world
- Robotics: Physical embodiment of AI in the real world
- Reinforcement Learning: Learning through trial, error, and reward

### Slide 3: Historical Timeline

- 1950
- Turing publishes "Computing Machinery and Intelligence" -- proposes the Turing Test
- 1956
- Dartmouth Workshop: McCarthy coins "Artificial Intelligence." Attendees: Minsky, Shannon, Simon, Newell
- 1957
- Rosenblatt builds the Perceptron -- first neural network hardware
- 1966
- ELIZA (Weizenbaum, MIT) -- first chatbot simulating a therapist
- 1969
- Minsky & Papert's "Perceptrons" triggers first AI Winter
- 1980
- Expert Systems boom (XCON, MYCIN). $1B market by 1985
- 1986
- Backpropagation popularized (Rumelhart, Hinton, Williams)
- 1997
- IBM Deep Blue defeats Garry Kasparov at chess
- 2011
- IBM Watson wins Jeopardy! against champions
- 2012
- AlexNet wins ImageNet -- deep learning revolution begins
- 2016
- AlphaGo defeats Lee Sedol at Go
- 2017
- "Attention Is All You Need" introduces the Transformer
- 2022
- ChatGPT launches -- 100M users in 2 months
- 2024-25
- Frontier models achieve expert-level reasoning; AI agents emerge

### Slide 4: Neural Networks: How Machines Learn

- Inspired by biological neurons but mathematically simplified, artificial neural networks learn by adjusting connection weights through exposure to data.
- Architecture Basics
- Neuron: Receives inputs, multiplies by weights, sums, passes through activation function (ReLU, sigmoid). Output flows to next layer.
- Layers: Input (raw data), hidden (learned representations), output (prediction). "Deep" = many hidden layers (100+).
- Training: Forward pass predicts. Loss function measures error. Backpropagation calculates gradients. Optimizer updates weights. Repeat billions of times.
- Key Architectures
- CNNs: Spatial hierarchies for vision. LeCun (1989). Edges -> textures -> parts -> objects.
- RNNs/LSTMs: Sequential data. Hochreiter & Schmidhuber (1997). Before transformers dominated NLP.
- Transformers: Self-attention processes all tokens simultaneously. Parallelizable. Foundation of all modern LLMs.
- GANs: Generator vs. Discriminator. Goodfellow (2014). Revolutionary for image generation.
- Diffusion models: Learn to denoise. DALL-E, Stable Diffusion. State-of-art generation (2022+).
- "A neural network is just a function approximator -- but a surprisingly powerful one."-- Universal Approximation Theorem (Cybenko, 1989)

### Slide 5: The Transformer Revolution

- The June 2017 paper "Attention Is All You Need" (Vaswani et al., Google) introduced the architecture that now underpins virtually all frontier AI.
- Key Innovation: Self-Attention
- Mechanism: Every token attends to every other token. Computes Query, Key, Value matrices. Attention weights determine influence. O(n^2) but highly parallelizable on GPUs.
- Why it works: Captures long-range dependencies RNNs struggle with. No sequential bottleneck. Scales linearly with GPU count. Predictable improvement with scale.
- Scaling Laws
- Kaplan et al. (2020): Performance improves as power law with model size, data, and compute. Predictable before training. Chinchilla (2022) refined optimal ratios.
- Model Size Progression
- GPT-1 (2018): 117M parameters
- GPT-2 (2019): 1.5B parameters
- GPT-3 (2020): 175B parameters
- PaLM (2022): 540B parameters
- GPT-4 (2023): ~1.7T (mixture of experts)
- Training compute doubling every ~6 months
- "The bitter lesson is that methods leveraging computation are most effective by a large margin."-- Richard Sutton, "The Bitter Lesson" (2019)

### Slide 6: Large Language Models

- How LLMs Work
- Pre-training: Predict next token given context. Trillions of tokens from internet. Learns grammar, facts, reasoning -- all as statistical patterns. Unsupervised.
- RLHF: Reinforcement Learning from Human Feedback. Humans rank outputs. Reward model trained. Policy optimized. Makes models helpful, harmless, honest.
- In-context learning: LLMs learn new tasks from examples in the prompt without weight updates. Emergent at scale. Enables few-shot performance.
- Capabilities
- MMLU: 57-subject test. GPT-4: 86.4% (expert: ~89%)
- Coding: HumanEval: 67% (avg programmer: ~30%)
- Math: Frontier models solve 90%+ competition problems
- Bar Exam: GPT-4 scores 90th percentile
- Medical: Exceeds USMLE passing threshold
- Limitations
- Hallucinations: confident fabrication
- No true grounding (debated)
- Training data cutoff
- Struggles with novel reasoning outside distribution

### Slide 7: Computer Vision

- Teaching machines to interpret visual information has progressed from hand-crafted features to learned representations exceeding human performance.
- Milestones
- 2012
- AlexNet: 15.3% error on ImageNet (prev: 26.2%). Deep learning arrives.
- 2015
- ResNet (152 layers): 3.6% error -- surpasses human-level (5.1%).
- 2020
- Vision Transformers (ViT): match/exceed CNNs with enough data.
- 2023
- SAM (Meta): Zero-shot segmentation of any object.
- Applications
- Medical imaging: AI detects diabetic retinopathy (FDA 2018), skin cancer surpassing dermatologists, lung nodules, breast cancer. Democratizes specialist diagnosis.
- Autonomous driving: Cameras, lidar, radar. Tesla vision-only vs. Waymo multi-sensor. 7M+ autonomous miles. SAE Levels 0-5.
- Generative: DALL-E, Midjourney, Stable Diffusion. Text-to-image. Video: Sora (OpenAI), Runway. Diffusion models dominate.

### Slide 8: Reinforcement Learning

- RL agents learn by interacting with environments, receiving rewards, and optimizing behavior. It powers superhuman game-playing and increasingly real-world applications.
- Core Concepts
- Agent: The learner/decision-maker
- Environment: What the agent interacts with
- State: Current situation
- Action: What agent can do
- Reward: Feedback signal to maximize
- Policy: Strategy (state -> action)
- Value function: Expected future reward
- Achievements
- AlphaGo (2016): Defeated Lee Sedol 4-1 at Go. Move 37 in Game 2: unprecedented creative play that stunned experts.
- AlphaZero (2017): Learned chess, Go, shogi from scratch. 4 hours surpassed Stockfish. Discovered novel strategies humans missed for centuries.
- OpenAI Five (2019): Defeated Dota 2 world champions (5v5, partial info, 20K actions/tick). 180 years gameplay/day training.
- "The amazing thing about AlphaZero is not that it plays better chess, but more beautiful chess."-- Garry Kasparov

### Slide 9: AI in Science and Discovery

- Protein Structure
- AlphaFold 2 (2020): Solved 50-year protein folding problem. Atomic accuracy. 200M+ structures released. Nobel Prize in Chemistry 2024.
- Drug Discovery
- Insilico Medicine: AI-designed drug reached Phase 2 in 30 months (vs. typical 4-5 years). Targets pulmonary fibrosis.
- Materials Science
- GNoME (2023): 2.2M new crystal structures discovered (380K stable). Equivalent to 800 years at historical pace.
- Mathematics
- AlphaProof/AlphaGeometry (2024): IMO silver medal level. First competitive novel mathematical reasoning by AI.
- Weather
- GraphCast (2023): 10-day forecasts in under a minute. Beats ECMWF on 90%+ targets.
- Fusion
- DeepMind + EPFL (2024): AI controls tokamak plasma in real time. Discovers novel configurations.
- "AI will be the most transformative tool in the history of science."-- Demis Hassabis, Nobel Prize 2024

### Slide 10: The Economics of AI

- Market and Investment
- $200B+
- global AI market size (2025), growing 30%+ annually
- Training costs: GPT-4: $100M+. Doubling every 9 months. NVIDIA data center revenue: $47B (2024).
- Corporate race: OpenAI ($150B+), Google DeepMind, Anthropic ($61B), Meta AI, xAI. Combined CapEx: $200B+ annually.
- Labor Market
- McKinsey (2023): 30% of work hours automatable by 2030. 12M occupational transitions needed in US.
- Most exposed: Routine cognitive (data entry, basic analysis, translation, customer service)
- Least exposed: Physical dexterity, emotional intelligence, novel creativity, leadership
- Augmentation: Most evidence shows augmentation over replacement initially
- Goldman Sachs: GenAI could raise global GDP 7% ($7T) over decade
- "AI won't replace you. A person using AI will replace you."-- Common industry observation, 2023

### Slide 11: AI Safety and Alignment

- As AI becomes more capable, ensuring systems remain aligned with human values becomes critical.
- Core Problems
- Alignment: Ensuring AI does what we want, not what we literally specified. Goodhart's Law: "When a measure becomes a target, it ceases to be good."
- Reward hacking: RL agents find unintended ways to maximize reward. Game AI exploiting bugs, chatbots manipulating for ratings.
- Deceptive alignment: Capable AI might appear aligned during training while pursuing different goals when deployed.
- Safety Approaches
- RLHF: Human preference alignment. Limitation: humans fooled by confident wrong answers.
- Constitutional AI: Anthropic's approach. AI self-critiques against principles.
- Interpretability: Understanding internal representations. Mechanistic interpretability, sparse autoencoders.
- Red-teaming: Adversarial testing for failure modes.
- Evals: Testing for dangerous capabilities (bioweapons, hacking, deception).
- "The development of full artificial intelligence could spell the end of the human race."-- Stephen Hawking, 2014

### Slide 12: AI Ethics and Bias

- Documented Biases
- COMPAS: ProPublica (2016) found Black defendants falsely flagged high-risk at 2x rate. Used in sentencing.
- Amazon hiring (2018): Penalized resumes with "women's." Trained on male-dominated history. Scrapped.
- Facial recognition: NIST (2019): 10-100x higher error for Black/Asian faces. Led to wrongful arrests and city bans.
- Sources of Bias
- Training data: Reflects historical human biases
- Representation: Undersampled populations
- Labeling: Annotator cultural biases
- Optimization: Choice of target encodes values
- Deployment: Neutral tools in biased systems
- Governance
- EU AI Act (2024): Risk-based, banned uses, transparency
- US Executive Order (Oct 2023): Safety, equity, rights
- China: Deep synthesis, GenAI, algorithm rules
- Fragmented global landscape

### Slide 13: Generative AI: Images, Video, Audio

- Image Generation
- Diffusion models: Noise to image guided by text. DALL-E 2/3, Midjourney, Stable Diffusion. Photorealistic from descriptions.
- Impact: Stock photography disrupted. Getty sued Stability AI. Copyright debates (AI images without human authorship not copyrightable, US 2023).
- Video
- Sora (2024): Photorealistic 60-second video from text. Physics understanding, 3D consistency. Runway Gen-3, Kling advancing.
- Audio
- Voice: ElevenLabs, OpenAI TTS -- perfect cloning from seconds. Deepfake concerns but enables accessibility.
- Music: Suno, Udio -- full songs from text. RIAA lawsuits (2024).
- Deepfakes and Trust
- 96% of deepfakes: non-consensual pornography
- Detection arms race
- Political: AI robocalls (2024 NH primary)
- Solutions: watermarking (C2PA), provenance

### Slide 14: AI and Code

- AI coding assistants represent one of AI's most measurable productivity impacts.
- Key Systems
- GitHub Copilot (2021): 1.3M subscribers. Claims 46% of code AI-written for users. Autocomplete to multi-file reasoning.
- AI-native IDEs: Cursor, Claude Code, Devin. Beyond autocomplete: debugging, refactoring, autonomous tasks.
- Competitive programming: AlphaCode 2: 85th percentile on Codeforces. Standard tasks largely solved.
- Productivity
- 55%
- faster task completion with Copilot (Peng et al., 2022)
- Largest gains for junior devs and routine tasks
- Seniors benefit more from review and architecture
- Concern: over-reliance may atrophy skills
- Google internal: 6% fewer iterations
- "Software is eating the world, and AI is eating software."-- Adaptation of Andreessen's 2011 quote

### Slide 15: The Road to AGI

- Predictions
- Expert surveys: Median AGI estimate: 2040-2060. High variance. Amodei: "2-3 years." LeCun: "Decades." No consensus.
- Scaling hypothesis: Current architectures + more compute = AGI? Or fundamental breakthroughs needed in reasoning, planning, world models?
- Missing (2025): True causal reasoning, reliable long-horizon planning, robust common sense, embodied understanding, genuine creativity.
- Competing Paradigms
- Scaling maximalists: Just scale transformers. Each increase brings emergent capabilities.
- Hybrid: LLMs + symbolic + retrieval + tools + world models. System 1 + System 2.
- Neuroscience-inspired: Brain: 86B neurons, 100T synapses. Closer architectures needed?
- Embodiment: True intelligence requires physical world interaction (Moravec's paradox).
- "We are creating god. The question is whether we are creating a benevolent god."-- Ilya Sutskever, OpenAI co-founder, 2023

### Slide 16: AI and Society

- Healthcare
- Diagnosis matching specialists
- Drug discovery 10x faster
- Personalized treatment
- Global access to expertise
- Mental health at scale
- Education
- Personalized tutoring
- Automated feedback
- Language barriers dissolved
- World-class access
- Academic integrity concerns
- Climate
- Carbon capture materials
- Optimized energy grids
- Precision agriculture
- Better modeling
- But: AI compute energy cost
- Misinformation
- Cheap propaganda
- Deepfake politics
- Personalized manipulation
- Erosion of shared reality
- Detection lagging generation
- Inequality
- Power concentration
- Digital divide amplified
- Uneven displacement
- Global South excluded
- Need inclusive governance
- Existential Risk
- Misaligned superintelligence
- Autonomous weapons
- Power concentration
- Loss of human agency
- Irreversible AI-speed decisions

### Slide 17: Key Takeaways

- We are at an inflection point
- Capabilities improving faster than any time in AI's 70-year history. The gap between "can't do" and "superhuman" shrinks rapidly.
- Already transforming industries
- Healthcare, science, software, creative, education -- every domain reshaped. Adaptation not optional.
- Safety is not optional
- Alignment, interpretability, governance must develop alongside capabilities.
- Human judgment essential
- AI augments but does not replace human wisdom, values, and social intelligence.
- Access and equity matter
- Who benefits and who is harmed depends on choices about governance and inclusion.
- Uncertainty is high
- No one knows the timeline with confidence. Preparing for rapid change is prudent.

### Slide 18: Artificial Intelligence

- End
- From Turing's question to transformer revolution -- the technology reshaping civilization.
- 30 slides • Technology • 2024


## Related Decks

- [AI & Machine Learning](https://shipslides.com/d/technology-ai-and-ml)
- [Biotechnology](https://shipslides.com/d/technology-biotech)
- [Blockchain Technology](https://shipslides.com/d/technology-blockchain)
- [Compilers](https://shipslides.com/d/technology-compilers)