🤖

Artificial Intelligence

From Turing's question to GPT-4 — the technology reshaping everything

What is AI?

Intelligence, Engineered

Artificial intelligence is the simulation of human-like reasoning by machines — perceiving inputs, learning patterns, and generating outputs that were once exclusively human. The current AI wave is driven by large language models (LLMs) and neural networks trained on internet-scale data.

1.8T

GPT-4 parameters (est.)

100M

ChatGPT users in 60 days

$200B

AI market cap, 2025

1950

Turing Test proposed

70%

Devs using AI coding tools

2017

"Attention Is All You Need"

⚡ The Transformer Revolution

In 2017, Google researchers published "Attention Is All You Need," introducing the transformer architecture. Unlike RNNs, transformers process entire sequences in parallel using self-attention — letting every token attend to every other token simultaneously.

This single architecture now underlies GPT, Claude, Gemini, Llama, DALL·E, Stable Diffusion, AlphaFold, and essentially every frontier AI system in 2025.

🔥 The ChatGPT Moment

ChatGPT launched November 30, 2022 and reached 100 million users in 60 days — the fastest product adoption in history. It forced every tech company on earth to accelerate or abandon its AI roadmap.

Within 18 months, Google released Gemini, Meta open-sourced Llama 2, Anthropic launched Claude, Microsoft integrated GPT-4 into Office 365, and AI passed bar exams and medical licensing tests.

🧬 How LLMs Work

Large language models are trained on trillions of tokens of text using next-token prediction — given a sequence, predict the most likely next word. After billions of gradient descent steps, the model learns grammar, facts, reasoning, and code as emergent behaviors.

RLHF (Reinforcement Learning from Human Feedback) then fine-tunes the raw model into a helpful assistant by rewarding good responses and penalizing harmful ones.

🌐 Multimodal AI

Early LLMs were text-only. Modern frontier models are multimodal — GPT-4o, Gemini 1.5, and Claude 3 can see images, hear audio, read PDFs, watch video, and write code that runs in sandboxes.

This shift from "chatbot" to "cognitive assistant" is the defining trend of 2024–2026. Models are embedded in IDEs, browsers, operating systems, and autonomous agents that take real-world actions.

🎯 AI Capability Milestones

BAR EXAM90th percentile (GPT-4)

USMLE STEP 1Passing score (GPT-4)

SAT MATH800/800 (GPT-4)

CODING#1 on HumanEval leaderboards

GOAlphaGo beat world champion (2016)

CHESSStockfish 16 rated ~3640 ELO

PROTEINAlphaFold solved 200M structures

IMAGESMidjourney won state fair art contest (2022)

Frontier Models

The Leading LLMs of 2025

These are the models shaping how the world writes, codes, reasons, and creates. Parameter counts are estimates where not officially disclosed.

OpenAI

GPT-4o

~1.8T params (MoE, est.)

OpenAI's flagship omni model — text, vision, audio, and video in one. Near-human voice latency, multimodal reasoning, 128K context. Powers ChatGPT Plus and the API. Released May 2024.

multimodalvoice128K ctx

OpenAI

o3 / o3-mini

Undisclosed

OpenAI's reasoning models trained with chain-of-thought reinforcement learning. o3 achieved 87.5% on ARC-AGI, the hardest benchmark for human-level abstract reasoning. Released early 2025.

reasoningSTEMARC-AGI

Anthropic

Claude 3.5 Sonnet

Undisclosed

Anthropic's best coding and instruction-following model. #1 on SWE-bench coding benchmark. Features computer use (controls a browser/desktop) and a 200K context window. Best-in-class for long-document analysis.

coding200K ctxcomputer use

Anthropic

Claude 3 Opus

Undisclosed

Anthropic's most intelligent model — excels at open-ended analysis, nuanced writing, and complex reasoning. Trained with Constitutional AI to be helpful, harmless, and honest. 200K context window.

reasoningwritingconstitutional AI

Google DeepMind

Gemini 1.5 Pro

~1T params (est.)

Google's flagship with a 1-million-token context window — the longest of any frontier model. Can analyze 1-hour videos, 11-hour audio recordings, or 700,000-word corpora. Integrated into Google Workspace.

1M ctxvideomultimodal

Google DeepMind

Gemini 2.0 Flash

Undisclosed

Google's fast, low-cost Gemini variant with native tool use, image generation, and real-time streaming. Powers Google Search AI Overviews and is the default model for Gemini app users. Released Dec 2024.

fasttool usestreaming

Meta AI

Llama 3.1 405B

405B (open weights)

Meta's open-weight flagship — 405 billion parameters, competitive with GPT-4 on most benchmarks. Available to download and run locally. The most capable openly accessible model ever released as of mid-2024.

open sourceself-hostable128K ctx

Mistral AI

Mistral Large 2

123B

European open-weight model with top-tier multilingual performance. Efficient MoE architecture gives it GPT-4-class quality at lower inference cost. Excels at European languages where US models lag.

open weightsmultilingualMoE

DeepSeek

DeepSeek-R1

671B (MoE)

Chinese open-source reasoning model that shocked the industry in Jan 2025 — matching o1 performance at a fraction of the cost, trained for $6M vs. hundreds of millions. Sparked a global debate about AI export controls.

reasoningopen sourceMoElow cost

xAI

Grok 3

Undisclosed

Elon Musk's xAI model trained on X (Twitter) real-time data. Has a "fun mode," no content guardrails by default, and access to live internet data. Grok 3 topped several math and science benchmarks on release in Feb 2025.

real-time datamathscience

Cohere

Command R+

104B

Enterprise-focused model optimized for RAG (Retrieval-Augmented Generation) and grounded business search. Built specifically for accurate citation, low hallucination rates, and enterprise document workflows.

RAGenterprisecitations

Apple

Apple Intelligence

~3B on-device

Apple's on-device LLM suite running on Neural Engine in iPhones and Macs. Routes complex requests to Private Cloud Compute (larger server models). Key feature: never sends private data off-device for on-device tasks.

on-deviceprivateiOS 18

AI Labs & Companies

Who Is Building the Future

A handful of labs and tech giants are racing to build artificial general intelligence — with wildly different philosophies, funding structures, and safety postures.

🟢 OpenAI

Founded 2015 · San Francisco · ~$157B valuation

Started as a nonprofit "to benefit humanity" and pivoted to a "capped-profit" structure after taking $1B from Microsoft. Sam Altman leads a team of ~1,500. GPT-4o, DALL·E 3, Sora, and Whisper are their flagship products. ChatGPT has 200M weekly active users. Microsoft invested $13B total and embedded OpenAI models in every product from Excel to GitHub Copilot.

Models: GPT-3.5, GPT-4, GPT-4o, o1, o3, DALL·E, Whisper, Sora

🔵 Anthropic

Founded 2021 · San Francisco · ~$61B valuation

Founded by Dario Amodei, Daniela Amodei, and 9 other ex-OpenAI researchers who left over safety concerns. Anthropic invented Constitutional AI — training models to follow a set of principles rather than pure RLHF. Amazon invested $4B. Google invested $2B. Claude is the preferred model for enterprise legal, medical, and coding workflows due to its low hallucination rate and precise instruction-following.

Models: Claude 3 Haiku, Sonnet, Opus · Claude 3.5 Sonnet / Haiku · Claude 4 Sonnet / Opus

🔴 Google DeepMind

Founded 1998 (Google) / 2010 (DeepMind) · London + Mountain View

Google and DeepMind merged their AI divisions in 2023. DeepMind invented AlphaGo (2016), AlphaFold (2020), and Gemini (2023). Google invented the transformer (2017), BERT (2018), and PaLM. With 40,000+ ML engineers and TPU custom hardware, they have more AI infrastructure than any other company. Sundar Pichai described AI as "more profound than fire or electricity."

Models: Gemini 1.0/1.5/2.0, AlphaFold 3, ImageFX, Vids, AlphaCode 2

🟣 Meta AI

Founded 1984 (Facebook) · Menlo Park · Public (META)

Mark Zuckerberg committed Meta to open-source AI as a strategic differentiator — giving away Llama models that would cost competitors billions to build. Meta AI is embedded in WhatsApp, Instagram, and Facebook for 3.2 billion users. Yann LeCun, Meta's Chief AI Scientist and deep learning pioneer, argues current LLMs will never reach AGI and advocates for "world models."

Models: Llama 2, Llama 3, Llama 3.1 (8B/70B/405B), Code Llama, SAM

⬛ Mistral AI

Founded 2023 · Paris · €1.1B raised

French startup founded by ex-Google DeepMind and Meta researchers. Released Mistral 7B for free download in a torrent — 2 weeks after founding — shocking the industry with its quality-to-size ratio. Champions European open-source AI sovereignty. Raised at a €6B valuation in 2024 and partnered with Microsoft.

Models: Mistral 7B, Mixtral 8x7B, Mistral Large, Mistral Large 2, Codestral

⚡ xAI

Founded 2023 · Palo Alto · $24B valuation

Elon Musk founded xAI after leaving OpenAI's board in 2018. Built Colossus — a 100,000-GPU training cluster in Memphis, TN, the largest single AI supercomputer in history — in 19 days. Grok is integrated into X (Twitter) and has access to real-time social data no other model has.

Models: Grok 1, Grok 1.5, Grok 2, Grok 3, Aurora (image gen)

🟡 NVIDIA

Founded 1993 · Santa Clara · $3.3T market cap (2025)

NVIDIA doesn't build LLMs — it builds the GPUs they run on. The H100 and H200 data center chips are the "picks and shovels" of the AI gold rush. Every major AI company runs on NVIDIA hardware. Jensen Huang became the world's richest person in 2024. NVIDIA's CUDA ecosystem created a moat so deep that AMD and Intel have struggled to compete even with superior raw specs.

Hardware: A100, H100, H200, B200 · Software: CUDA, cuDNN, TensorRT, NIM

🔶 DeepSeek

Founded 2023 · Hangzhou, China

Subsidiary of Chinese quant hedge fund High-Flyer Capital. Released DeepSeek-R1 in January 2025 for free download — a reasoning model matching OpenAI's o1 performance, trained for roughly $6M. The announcement dropped NVIDIA stock 17% in a single day ($600B wiped), the largest single-day market cap loss in US history, as investors questioned GPU demand assumptions.

Models: DeepSeek-V2, DeepSeek-V3, DeepSeek-R1, DeepSeek-Coder

History of AI

75 Years of Artificial Intelligence

From Alan Turing's theoretical framework to transformer models that passed the bar exam — the uneven, punctuated history of a field that failed, nearly died, and then took over the world.

1950

Alan Turing — "Computing Machinery and Intelligence"

Turing proposes the Imitation Game (Turing Test): if a machine can convincingly imitate human conversation, it should be considered intelligent. He predicts machines will pass by 2000. He's off by 20 years.

1956

Dartmouth Conference — AI Is Born as a Field

John McCarthy, Marvin Minsky, Claude Shannon, and others gather at Dartmouth and coin the term "Artificial Intelligence." McCarthy's proposal: "every aspect of learning or intelligence can be so precisely described that a machine can be made to simulate it."

1957

Perceptron — The First Neural Network

Frank Rosenblatt builds the Perceptron at Cornell — the first electronic device that could learn from examples. It can recognize simple shapes. Rosenblatt predicts it will walk, talk, and reproduce. Minsky's 1969 book "Perceptrons" proves its limits and kills neural network research for a decade.

1966–74

First AI Winter

ALPAC report tells the US government that machine translation is impossible. DARPA cuts funding. The field implodes. Researchers who were promised "20 years to solve AI" receive nothing for a generation.

1986

Backpropagation — Neural Nets Return

Rumelhart, Hinton, and Williams publish the backpropagation algorithm for training multi-layer neural networks. This is the foundational technique still used today — it allows gradients to flow backwards through layers to update weights.

1997

Deep Blue Beats Kasparov

IBM's Deep Blue defeats world chess champion Garry Kasparov 3.5–2.5. Kasparov accuses IBM of cheating. The match is one of the most-watched events of 1997. Deep Blue uses brute-force tree search, not learned intelligence — but the public can't tell the difference.

1998–2010

Second AI Winter & SVMs Dominate

Neural networks fall out of favor again. Support Vector Machines and random forests become the dominant ML paradigm. Geoffrey Hinton, Yann LeCun, and Yoshua Bengio — the "deep learning rebels" — keep training neural nets in obscurity. Funding is scarce. Results are modest.

2006

Hinton Coins "Deep Learning"

Geoffrey Hinton publishes techniques for pre-training deep belief networks, showing that deep neural networks can learn useful representations. He coins "deep learning" to differentiate from the shallow networks that had failed. The AI spring is beginning.

2012

AlexNet — The ImageNet Moment

Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton win the ImageNet competition with AlexNet — a deep convolutional neural network running on two GPUs. Their error rate (15.3%) is nearly half the runner-up's (26.2%). Every major tech company immediately pivots to deep learning. NVIDIA's stock begins a decade-long climb.

2014

GANs — Generative Adversarial Networks

Ian Goodfellow invents GANs in a bar argument. Two networks — a generator and discriminator — compete until the generator produces realistic synthetic images. Yann LeCun calls it "the most interesting idea in ML in the last 10 years." GANs eventually produce photorealistic human faces and deepfakes.

2016

AlphaGo Beats Lee Sedol

Google DeepMind's AlphaGo defeats 18-time world Go champion Lee Sedol 4–1 in Seoul. Go has more board positions than atoms in the universe — brute force is impossible. AlphaGo used deep reinforcement learning and Monte Carlo tree search. Lee Sedol called Move 78 in Game 4 "a move only a human could make." AlphaGo had made it.

2017

"Attention Is All You Need" — The Transformer

Eight Google Brain researchers publish the transformer paper. Self-attention replaces recurrence, enabling parallel training on massive datasets. GPT-1, BERT, GPT-2, GPT-3, T5, and every LLM in existence descends from this architecture. It is arguably the most consequential ML paper ever published.

2020

GPT-3 — 175 Billion Parameters

OpenAI releases GPT-3 with 175 billion parameters — 100× larger than GPT-2. Developers who gain API access are stunned: it writes code, poetry, legal briefs, and movie scripts with no fine-tuning. The scaling hypothesis — more data + more compute = more intelligence — appears to be a law of nature.

2020

AlphaFold 2 — Biology's 50-Year Problem Solved

DeepMind's AlphaFold 2 predicts protein 3D structure from amino acid sequence with atomic accuracy — solving a problem that had stymied biochemists for 50 years. Science magazine's Breakthrough of the Year 2020. AlphaFold 3 (2024) extended this to DNA, RNA, and small molecules, accelerating drug discovery across every disease area.

2021

DALL·E, CLIP, and the Generative Image Era

OpenAI releases DALL·E — a transformer trained on text-image pairs that generates images from descriptions. CLIP learns visual concepts from internet alt-text. Stability AI releases Stable Diffusion as open source in 2022, enabling anyone to run image generation locally. The stock photo industry begins its structural decline.

2022

ChatGPT — 100 Million Users in 60 Days

OpenAI releases ChatGPT on November 30, 2022. It reaches 1 million users in 5 days and 100 million in 60 — the fastest product adoption in history, surpassing TikTok and Instagram. Every board of directors in the world adds "AI strategy" to the 2023 agenda. The AI era begins for the general public.

2023

GPT-4, Claude, Llama, and the Model Explosion

GPT-4 passes the bar exam in the 90th percentile. Anthropic releases Claude. Meta open-sources Llama, enabling the open-source AI ecosystem. Microsoft embeds Copilot in Windows. Google releases Bard (later Gemini). The race is on. Sam Altman testifies before Congress. Geoffrey Hinton quits Google to warn about AI risk.

2024

Agents, Reasoning, and the $1T Infrastructure Build

AI agents that autonomously browse the web, write and run code, and manage files become mainstream. OpenAI releases o1 — a reasoning model that "thinks before it answers." Microsoft, Google, Amazon, and Meta announce $300B+ in combined AI infrastructure spending. NVIDIA's market cap exceeds $3 trillion. The Stargate project pledges $500B for US AI infrastructure.

2025

DeepSeek Shock, Grok 3, and the Era of Reasoning Models

DeepSeek-R1 launches for free download in January, matching o1 on benchmarks at $6M training cost. NVIDIA drops 17% — $600B market cap — in one day. OpenAI releases o3. Anthropic releases Claude 3.7 Sonnet with "extended thinking." AI-generated code comprises an estimated 40%+ of new code on GitHub. The question shifts from "can AI code?" to "how many engineers does a company need?"

Core Concepts

How Modern AI Actually Works

The ideas behind transformers, training, alignment, and deployment — explained without the PhD.

🔷 Transformer Architecture

Every modern LLM is a transformer. Text is split into tokens (roughly word fragments) and converted to numeric vectors called embeddings. The transformer's key innovation is self-attention: each token computes how much it should "attend to" every other token in the sequence, enabling the model to understand that "it" in "the trophy won't fit in the suitcase because it's too big" refers to the trophy, not the suitcase.

👁️ Self-Attention Mechanism

Each token is projected into three vectors: Query (Q), Key (K), and Value (V). Attention scores are computed as softmax(QKᵀ/√d), then applied to V. This happens in parallel across all tokens simultaneously — the key reason transformers train much faster than RNNs, which had to process tokens sequentially. Multi-head attention runs this in parallel across many "heads," each learning different relationships.

📊 Pre-training (Next Token Prediction)

LLMs are trained by giving the model a sequence and asking it to predict the next token. Given "The cat sat on the ___", predict "mat." Do this with trillions of tokens from the web, books, and code, optimize with gradient descent, and after billions of steps the model has encoded grammar, facts, reasoning, and common sense as statistical patterns in billions of floating-point weights.

🎯 RLHF — Reinforcement Learning from Human Feedback

Raw pre-trained models are unpredictable — they complete text, not instructions. RLHF fixes this: human raters score model outputs, a reward model learns to predict human preferences, and the LLM is then fine-tuned with PPO (Proximal Policy Optimization) to maximize predicted human approval. This transforms a text completer into an assistant. InstructGPT (2022) was the first major demonstration.

📜 Constitutional AI (Anthropic)

Anthropic's alternative to pure RLHF. Rather than relying only on human raters, the model is given a written "constitution" — a list of principles ("be helpful," "avoid harmful content," "don't deceive"). The model critiques and revises its own responses according to these principles using AI feedback, reducing reliance on human labelers while improving consistency and safety.

🔍 RAG — Retrieval-Augmented Generation

LLMs have a knowledge cutoff and fixed context window. RAG solves this: at inference time, retrieve relevant documents from a vector database using semantic search, inject them into the prompt, then generate grounded in retrieved facts. Used for enterprise search, customer support bots, and any application where the model needs up-to-date or private knowledge without retraining.

🎛️ Fine-Tuning and LoRA

Fine-tuning adapts a pre-trained model to a specific task or style by training on a smaller, curated dataset. Full fine-tuning updates all billions of parameters — expensive. LoRA (Low-Rank Adaptation) updates only small adapter matrices injected into each layer — training just 0.1% of parameters while achieving similar results. Enables fine-tuning on a single consumer GPU.

🌊 Embeddings & Vector Databases

Embeddings convert text (or images, audio) into dense numeric vectors where semantic similarity = geometric proximity. "King" − "man" + "woman" ≈ "queen" in embedding space. Vector databases (Pinecone, Weaviate, pgvector) index billions of these vectors for sub-millisecond nearest-neighbor search — the backbone of RAG, recommendation systems, and semantic search.

🧮 MoE — Mixture of Experts

Instead of activating all parameters for every input, MoE models route each token to a subset of "expert" sub-networks via a learned router. A 1.8T parameter GPT-4 (estimated MoE) may only activate ~200B parameters per forward pass — achieving large-model quality at smaller-model inference cost. DeepSeek-R1 (671B total, 37B active per token) demonstrated MoE's power dramatically.

🔗 Chain-of-Thought Prompting

Adding "Let's think step by step" to a prompt dramatically improves reasoning accuracy on math and logic tasks. The model generates intermediate reasoning steps before the final answer — forcing it to "show its work." OpenAI's o1/o3 models take this further with learned chain-of-thought through reinforcement learning, enabling them to solve competition math problems that stumped previous frontier models.

🤖 AI Agents & Tool Use

An agent is an LLM given access to tools — web search, code execution, file systems, APIs, computer control. The model reasons about a goal, selects a tool, observes the output, and continues until done. This creates autonomous workflows: "research this topic and write a report," "find the bug and open a PR," "book the cheapest flight to Miami." Agentic AI is the defining deployment pattern of 2024–2026.

🛡️ AI Safety & Alignment

Alignment is the problem of ensuring AI systems do what humans actually want, not just what they're literally told. Risks range from near-term (bias, hallucination, misuse) to long-term (misaligned superintelligent systems). Anthropic, DeepMind Safety, OpenAI Preparedness, and academic groups like ARC study this full-time. Techniques include RLHF, Constitutional AI, activation steering, interpretability research, and red-teaming.

📏 Quantization & Inference Efficiency

Frontier models are trained in 32-bit or 16-bit floating point but can be compressed to 8-bit, 4-bit, or even 1-bit ("BitNet") with modest quality loss. A 70B-parameter Llama 3 model at 4-bit quantization fits on a single consumer GPU (24GB VRAM). llama.cpp runs quantized models on M-series Macs at practical speeds — enabling local, private, offline LLM inference without a data center.

📐 Scaling Laws

Kaplan et al. (2020) and Hoffmann et al. "Chinchilla" (2022) discovered that LLM performance follows predictable power laws: loss decreases smoothly as compute, data, and parameters scale. The Chinchilla paper showed that most models were undertrained — optimal compute budget should train a model ~20× its parameter count in tokens. This insight reshaped every major lab's training strategy.

Early AI — 1991

Dr. Sbaitso

Before ChatGPT, before Siri, before ALICE — there was Dr. Sbaitso. A text-to-speech chatbot bundled with Creative Labs Sound Blaster cards, dispensing armchair therapy to DOS users since 1991.

Sound Blaster Artificial Intelligence Text to Speech Operator

Dr. Sbaitso

Dr. Sbaitso (an acronym for Sound Blaster Artificial Intelligence Text to Speech Operator) was a novelty chatbot program bundled with Creative Labs Sound Blaster sound cards from 1991 onward. Its primary purpose was to demonstrate the card's text-to-speech synthesis — but for many early PC users, it was their first conversation with an "AI."

The program ran on MS-DOS and simulated a psychotherapist, using simple keyword pattern-matching derived from ELIZA (1966). It responded in a robotic synthesized voice — an eerie novelty in an era when most PCs were silent machines.

1991Release year

ELIZAAncestor (1966)

DOSPlatform

v2.10Final version

TTSKey innovation

Screenshots

The Interface

Dr. Sbaitso · Title / intro screen · 640×400 VGA · Internet Archive

Dr. Sbaitso conversation session — ELIZA-style therapy dialogue

Dr. Sbaitso in session · Patient dialogue · ELIZA-style pattern responses · 640×400 VGA

Sample Session

Talking to the Doctor

Dr. Sbaitso responded to everything in a robotic synthesized voice using Creative's text-to-speech engine. The pattern-matching was rudimentary — but startlingly lifelike to users who had never interacted with a "talking computer" before.

─────────────────────────────────────────────────────

HELLO, MY NAME IS DOCTOR SBAITSO.
I AM HERE TO HELP YOU.
WHAT IS YOUR NAME?

USER > my name is Dave

HELLO, DAVE. HOW ARE YOU FEELING TODAY?

USER > I feel a bit depressed

TELL ME MORE ABOUT THESE FEELINGS.

USER > I keep having the same dream over and over

DOES IT BOTHER YOU THAT YOU KEEP HAVING THE SAME DREAM OVER AND OVER?

USER > poop

WATCH YOUR MOUTH OR I WILL WASH IT OUT WITH SOAP.

─────────────────────────────────────────────────────

Press ESC to exit · Type QUIT to end session

The Hardware

Sound Blaster Cards That Shipped With It

Dr. Sbaitso was a pack-in software demo for Creative Labs' Sound Blaster line — the dominant PC sound card of the 1990s. Before Sound Blaster, most PC audio was a beeping speaker. These cards brought FM synthesis, CD audio, and text-to-speech to DOS PCs.

Sound Blaster 1.0 (CT1320)

1989 · ISA 8-bit · OPL2 FM synthesis

The original Sound Blaster — the card that started it all. Featured an AdLib-compatible OPL2 FM synthesizer, 8-bit DAC, a game port, and a MIDI interface. Became the de facto standard for DOS gaming audio. Dr. Sbaitso was introduced with later cards but demonstrated the TTS capabilities of the entire line.

Sound Blaster 2.0 (CT1350B)

1991 · ISA 8-bit · OPL2 · on-board DMA

The Sound Blaster 2.0 refined the original design with on-board DMA support for background audio playback, double the sample rate (up to 44.1kHz), and improved compatibility. This is the era Dr. Sbaitso launched with — bundled on a floppy alongside the drivers, showing off the card's text-to-speech engine to amazed first-time owners.

Historical Lineage

From ELIZA to ChatGPT — A Direct Line

Dr. Sbaitso was a direct descendant of ELIZA and a spiritual ancestor of every chatbot that followed. The core pattern-matching approach Weizenbaum invented in 1966 lives on — scaled by 6 orders of magnitude — in every modern LLM.

ELIZAMIT · 1966
Pattern matching
psychotherapist

→

PARRYStanford · 1972
First model of
mental illness

→

Dr. SbaitsoCreative · 1991
TTS + ELIZA
on Sound Blaster

→

ALICE1995
AIML pattern
matching

→

SmarterChildAIM · 2001
First mass-market
chatbot

→

SiriApple · 2011
Voice assistant
goes mainstream

→

ChatGPTOpenAI · 2022
Transformer-based
LLM assistant

Deep Cuts

Facts About Dr. Sbaitso

🧼 The Soap Easter Egg

Dr. Sbaitso had a profanity filter. Type anything crude and it would respond: "WATCH YOUR MOUTH OR I WILL WASH IT OUT WITH SOAP." This was one of its most famous and endearing responses — teenagers across America immediately tested it upon installing Sound Blaster drivers for the first time.

🗣️ The Voice

Dr. Sbaitso's voice was generated using Creative Labs' own text-to-speech engine — a robotic, slightly nasal monotone that was unlike any sound most PC speakers had ever made. The novelty of a computer speaking aloud was so striking that Creative bundled the program primarily as a TTS demo, not as a productivity tool.

🧠 How It "Thought"

Dr. Sbaitso used the same technique as ELIZA: keyword scanning and template responses. Input text was scanned for trigger words ("dream," "mother," "feel"), and a matching response template was selected and filled in. There was no understanding, no memory beyond the current exchange, and no learning. The illusion of intelligence was entirely in the user's mind.

📦 The Pack-In Strategy

Creative Labs bundled Dr. Sbaitso with Sound Blaster Pro, Sound Blaster 16, Sound Blaster AWE32, and other cards through the early 1990s. It was one of several showcase applications — alongside VOXKIT (voice commands) and various MIDI demos — designed to justify the $200+ price tag of a sound card at a time when most PCs had no audio whatsoever.

🎮 Cultural Impact

Dr. Sbaitso has a dedicated cult following among retro computing enthusiasts. It appears in YouTube "playing Dr. Sbaitso in 2023" videos with millions of views, has been emulated on archive.org, and is widely cited as many people's first experience of "artificial intelligence." Its infamous catchphrases — "THAT IS INTERESTING... PLEASE CONTINUE" and "HOW DOES THAT MAKE YOU FEEL?" — became early internet memes.

🔬 ELIZA's Legacy

Joseph Weizenbaum created ELIZA in 1966 at MIT and was disturbed by how quickly users formed emotional attachments to the program — including his own secretary, who asked him to leave the room so she could have a private conversation with it. He wrote Computer Power and Human Reason (1976) warning against anthropomorphizing machines. Dr. Sbaitso repeated this phenomenon for a new generation 25 years later.