AI Glossary

Agent2Agent Protocol (A2A)

ay-too-ay

Abductive Reasoning

/æbˈdʌktɪv ˈriːzənɪŋ/

Logical inference that seeks the simplest explanation for an observation. What Sherlock Holmes does. Current AI struggles with this creative hypothesis formation.

Agent Evaluation

/ˈeɪdʒənt ɪˌvæljuˈeɪʃən/

Systematic testing of AI agents through automated trials measuring task completion, reliability, and behavior quality. How organizations ensure agents are ready for production work.

Agent Harness

/ˈeɪdʒənt ˈhɑːrnɪs/

The infrastructure layer that enables AI models to function as agents—managing context, orchestrating tools, handling errors, and maintaining state across sessions. The scaffold that turns a language model into a worker.

Agent Transcript

/ˈeɪdʒənt ˈtrænskrɪpt/

The complete record of an AI agent's execution—every output, tool call, reasoning step, and intermediate result. The black box recorder for understanding what agents actually did.

Agentic Coding

/eɪˈdʒentɪk ˈkoʊdɪŋ/

AGI (Artificial General Intelligence)

/ˌeɪ-dʒiː-ˈaɪ/

AI Agents

/eɪ aɪ ˈeɪdʒənts/

AI systems that autonomously take actions to accomplish goals. Beyond chatbots - agents use tools, make decisions, and complete multi-step tasks. The future of AI work.

AI Copilot

/eɪ aɪ ˈkoʊˌpaɪlət/

AI assistants that augment human work rather than replace it. Suggests code, drafts emails, summarizes documents. The 'bicycle for the mind' model of AI.

AI Gateway

/ˌeɪˈaɪ ˈɡeɪtweɪ/

Infrastructure layer that routes, monitors, and manages API calls between applications and multiple AI model providers. Enables multi-model orchestration, failover, and cost optimization.

AI Infrastructure

/eɪ aɪ ˈɪnfrəstrʌktʃə/

The compute stack powering AI: chips (GPUs, TPUs), data centers, networking, and cloud platforms. Jensen Huang's 'AI factories' concept. A multi-trillion dollar buildout.

AI SDR

/eɪ aɪ ɛs diː ɑːr/

AI-powered Sales Development Representative. Autonomous agents that handle outbound prospecting, email sequences, and lead qualification - working 24/7 with multivariate optimization humans can't match.

ASI (Artificial Superintelligence)

/ˌeɪ-es-ˈaɪ/

Chinchilla

/tʃɪnˈtʃɪlə/

DeepMind's 2022 paper proving LLMs were undertrained. For optimal compute, model size and training data should scale equally. Changed how the industry trains models.

Closing the Loop

/ˈkloʊzɪŋ ðə luːp/

Confabulation

kon-fab-yoo-LAY-shun

Geoffrey Hinton's preferred term for AI hallucinations - the phenomenon where models generate plausible-sounding but incorrect information. Humans do this too.

Deep Learning

/diːp ˈlɜːrnɪŋ/

Machine learning using multilayered neural networks. The 'deep' refers to multiple layers - from three to thousands. Revolutionized AI from 2012 onwards.

Embodied AI

/ɪmˈbɒdid eɪ aɪ/

AI systems with physical bodies that interact with the real world. Intelligence emerges from brain-body-environment interplay. Key pathway to AGI according to researchers.

Enterprise AI

/ˈentərˌprīz ˌeɪˈaɪ/

Artificial intelligence designed for business environments - solving complex problems, automating workflows, and integrating with corporate systems.

GDP val

/ˌdʒiː diː ˈpiː væl/

OpenAI's benchmark measuring AI on economically valuable knowledge work - legal briefs, engineering, customer support. GPT-5.2 scores 71%, beating human experts.

Generalization

/ˌdʒenərəlaɪˈzeɪʃən/

An AI model's ability to perform well on new, unseen data—not just training examples. The holy grail of machine learning. Jagged intelligence shows current models struggle here.

Grounding

/ˈɡraʊndɪŋ/

Connecting AI outputs to verified external sources to reduce hallucinations. RAG is the primary technique. Enables source citation and fact verification.

Hallucination

/həˌluːsɪˈneɪʃən/

When AI generates confident but false information. Sub-1% rates now achievable in top models, but managing uncertainty beats chasing zero.

Human-in-the-Loop

/ˈhjuːmən ɪn ðə luːp/

AI systems that include human oversight, approval, or intervention at key decision points. Balances automation benefits with human judgment and accountability.

Jagged Intelligence

JAG-id in-TEL-ih-jence

The inconsistent capability profile of AI - PhD-level at Math Olympiad, failing basic logic puzzles. A core barrier to AGI identified by Demis Hassabis.

JEPA

/ˈdʒepə/

Yann LeCun's Joint Embedding Predictive Architecture. Predicts abstract representations, not pixels. His proposed path to human-level AI, avoiding generative model limitations.

Long-running Agents

/lɒŋ ˈrʌnɪŋ ˈeɪdʒənts/

AI agents that work on tasks spanning hours, days, or multiple sessions—requiring persistent state, error recovery, and context management beyond a single conversation.

Model Context Protocol (MCP)

em-see-pee

Model Commoditization

/ˈmɒdəl kəˌmɒdɪtaɪˈzeɪʃən/

Frontier AI models reaching capability parity, shifting competition from 'smartest model' to applications and distribution. Multiple viable options drive price competition.

Moltbot

Open-source personal AI assistant that runs locally on your computer with full system access. Created by Peter Steinberger, formerly known as Claudebot.

Neurosymbolic AI

/ˌnjʊərəʊ-sɪmˈbɒlɪk/

Pre-training

/priː ˈtreɪnɪŋ/

The first phase of LLM training - learning language patterns from billions of words. Takes weeks/months and massive compute. Fine-tuning comes after.

RALPH Loop

/rælf luːp/

Reinforcement Learning

/ˌriːɪnˈfɔːrsmənt ˈlɜːrnɪŋ/

Machine learning where agents learn through trial and error, receiving rewards for actions. Powers RLHF alignment for ChatGPT, Claude, and reasoning models like DeepSeek-R1.

Scaling Laws

SKAY-ling lawz

The empirical relationship between AI model performance and compute, data, and parameters. Drove the 2020-2025 era, now showing diminishing returns.

Sora

/ˈsɔːrə/

OpenAI's text-to-video model. Generates up to 20 seconds at 1080p. Sora 2 (Sept 2025) added audio, better physics, and 'Cameos' for personal likeness generation.

Supervision Threshold

/ˌsuːpərˈvɪʒən ˈθreʃˌhoʊld/

The capability level at which AI transitions from requiring human oversight to operating autonomously—the key dividing line between augmentation and replacement.

Tool Use

/tuːl juːz/

The ability of AI models to call external functions, APIs, and systems. What transforms chatbots into agents. Also called 'function calling' - the hands of AI.

TPU

/tiː piː juː/

Google's custom AI chip (Tensor Processing Unit). 7th generation 'Ironwood' delivers 42.5 exaflops across 9,216 chips. Anthropic plans to use 1 million for Claude.

Training Ladder

/ˈtreɪnɪŋ ˈlædər/

The professional development pipeline where juniors exchange grunt work for mentorship—now threatened by AI that can do the grunt work better.

Universal Commerce Protocol (UCP)

yoo-see-pee

Workflow Automation

/ˈwɜːrkfloʊ ˌɔːtəˈmeɪʃən/

Using AI to automate multi-step business processes end-to-end. Goes beyond single tasks to orchestrate entire workflows: intake → processing → output → handoff.

World Models

wurld MOD-els

AI systems that learn to simulate and predict how the physical world works - spatial dynamics, intuitive physics, and cause-effect relationships beyond text.