All terms · A-Z

AI model lab · Beijing, China · founded 2023.

Vertically stacked NAND flash · used for AI training data storage · 300+ layers in 2026 flagship nodes.

NVIDIA GPU · Ampere GA100 architecture · 2020.

Agent2Agent, an open protocol for AI agents to discover and talk to other agents.

An AI system that plans, uses tools, and takes multi-step actions to accomplish goals · not just a chat turn.

The research field focused on making AI systems pursue intended goals safely and reliably.

BenchGecko's 0-1000% composite score measuring AI sector valuation vs fundamentals.

The capital expenditure AI labs and hyperscalers spend on chips, datacenters, and training clusters.

AI model lab · Tel Aviv, Israel · founded 2017.

AI researcher.

Paul Gauthier's terminal-based git-aware AI pair programmer · commits with context of the whole repo.

Aider polyglot · coding benchmark tracked on BenchGecko.

AI model lab · Hangzhou, China · founded 1999.

Attention with Linear Biases · position encoding via attention-score penalties instead of positional embeddings.

big tech · Mountain View, CA, USA · founded 1998.

big tech · Seattle, WA, USA · founded 1994.

semiconductor · Santa Clara, CA, USA · founded 1969.

AMD's Instinct MI400 series of data-center GPUs, planned for 2026 with HBM4 memory.

AI researcher.

ANLI · knowledge benchmark tracked on BenchGecko.

AI model lab · San Francisco, CA, USA · founded 2021.

AI compute infrastructure · San Francisco, CA, USA · founded 2019.

AI developer tools · San Francisco, CA, USA · founded 2022.

APEX-Agents · agentic benchmark tracked on BenchGecko.

energy · Dallas, TX, USA · founded 2001.

Running the same open-weight model on multiple providers to exploit price differences · up to 30× savings.

ARC AI2 · knowledge benchmark tracked on BenchGecko.

ARC-AGI · reasoning benchmark tracked on BenchGecko.

ARC-AGI-2 · reasoning benchmark tracked on BenchGecko.

Annualized Recurring Revenue · the standardized metric SaaS and AI companies report to investors.

AI researcher.

Huawei custom AI chip (ASIC) · Da Vinci architecture · 2024.

chip foundry · Veldhoven, Netherlands · founded 1984.

AutoGen is Microsoft's open-source framework for building multi-agent AI applications.

AutoGPT is the open-source autonomous agent that kicked off the entire category.

NVIDIA GPU · Blackwell SM100 architecture · 2024.

NVIDIA GPU · Blackwell Ultra SM100+ architecture · 2025.

AI model lab · Beijing, China · founded 2000.

Balrog · knowledge benchmark tracked on BenchGecko.

AI compute infrastructure · San Francisco, CA, USA · founded 2019.

Asynchronous model calls processed within a time window at a lower price than real-time calls.

Submitting many prompts at once in a single batch job · providers discount 50% on delayed batch completion.

BBH · reasoning benchmark tracked on BenchGecko.

semiconductor · Palo Alto, CA, USA · founded 1961.

Net cash burned divided by net new ARR · < 1× is healthy · AI startups often run 2-5× pre-scale.

Monthly cash outflow · the speed at which an AI company spends investor capital before hitting breakeven.

Bring Your Own Key · you pay the model provider directly, the app just routes your requests.

The fraction of input tokens served from provider-side prompt cache · directly impacts effective pricing.

Discounted input tokens for prompts the provider has already processed · 50-90% off on repeat prefix.

CadEval · coding benchmark tracked on BenchGecko.

semiconductor · Sunnyvale, CA, USA · founded 2016.

A prompting technique (and trained behavior) where the model shows step-by-step reasoning before the final answer.

AI application · Menlo Park, CA, USA · founded 2021.

Crowdsourced head-to-head AI model comparison · humans vote on anonymous outputs and Elo ratings rank the models.

Chess Puzzles · knowledge benchmark tracked on BenchGecko.

Anthropic's large language model family, first released in 2023, sold in several capability tiers.

Claude Code is a command-line coding agent built by Anthropic that lives in your terminal and edits files, runs commands, and searches your codebase with direct Claude model access.

AI researcher.

Open-source VSCode coding agent · forked as Roo Code and Continue, built on user-supplied API keys.

OpenAI's open-source terminal coding agent · runs GPT models against your repo from the shell.

Sourcegraph's enterprise coding AI · deep codebase context from graph-aware search.

AI model lab · Toronto, Canada · founded 2019.

Agents that operate a computer interface: read the screen, move the mouse, click and type.

The maximum number of tokens a model can process in one request · input prompt plus generated output.

Continue is an open-source IDE extension for building your own AI code assistant with custom models, context providers, and slash commands.

Open-source IDE-native coding assistant · inline autocomplete + chat across VSCode and JetBrains.

AI application · San Francisco, CA, USA · founded 2020.

AI compute infrastructure · Livingston, NJ, USA · founded 2017.

TSMC's next-gen advanced packaging · enables 12+ HBM stacks and multi-die GPUs · used on B300, MI400.

TSMC's previous-gen advanced packaging · used on H100 / A100 · being phased out for CoWoS-L.

CrewAI is a framework for orchestrating autonomous AI agents with roles, goals, and tools.

energy · Denver, CO, USA · founded 2018.

Cerebras AI system · 1 × Cerebras WSE-3.

CSQA2 · knowledge benchmark tracked on BenchGecko.

Cursor is a fork of VS Code rebuilt around an AI pair programmer.

% of revenue from top customers · AI labs often 40-60% from top 10 vs SaaS median <20%.

Cybench · coding benchmark tracked on BenchGecko.

AI researcher.

The requirement that data be stored and processed in a specific geographic region.

AI compute infrastructure · San Francisco, CA, USA · founded 2013.

DDR memory · 2021 · 51 GB/s per device.

The planned next generation of mainstream DRAM after DDR5.

DeepResearch Bench · knowledge benchmark tracked on BenchGecko.

Models from the Chinese lab DeepSeek, known for open-weight mixture-of-experts and reasoning models.

AI researcher.

Cognition AI's autonomous software engineer agent · browses, writes, and deploys code end-to-end.

NVIDIA AI system · 8 × NVIDIA B200.

NVIDIA AI system · 72 × NVIDIA B200.

NVIDIA AI system · 72 × NVIDIA B300.

Reduction in existing shareholder ownership % after new equity issuance · typical 15-25% per round.

Training a small model to mimic a large one, preserving most quality at a fraction of the size.

AI application · London, UK · founded 2022.

AI researcher.

A dense vector representation of text, image, or audio in high-dimensional space · the backbone of RAG and search.

Employee Stock Ownership Plan · pool of shares set aside for employees · typical 10-20% of cap table.

EU regulation classifying AI systems by risk level and imposing compliance obligations on developers and deployers.

US federal government authorization for cloud and AI services handling agency data.

Fiction.LiveBench · knowledge benchmark tracked on BenchGecko.

Further training a pre-trained model on domain data to adapt behavior without retraining from scratch.

AI compute infrastructure · San Francisco, CA, USA · founded 2022.

FP16 is a 16-bit floating point format widely used in neural network training and inference.

FP8 is an 8-bit floating point format used to speed up AI training and inference.

The no-cost tier of an AI API · usually rate-limited, quota-capped, or branded-output-only.

FrontierMath-2025-02-28-Private · math benchmark tracked on BenchGecko.

FrontierMath-Tier-4-2025-07-01-Private · math benchmark tracked on BenchGecko.

Tool-use / function-call responses count as output tokens · structured JSON tool calls are billed normally.

The API mechanism letting models emit structured JSON to invoke external functions.

Intel custom AI chip (ASIC) · Heterogeneous MME + TPC architecture · 2024.

NVIDIA CPU and GPU superchip · Grace CPU + 2x B200 architecture · 2025.

NVIDIA CPU and GPU superchip · Grace CPU + 2x B300 architecture · 2025.

GDDR memory · 2020 · 96 GB/s per device.

GDDR memory · 2025 · 192 GB/s per device.

EU data protection regulation · imposes strict consent, residency, and rights requirements on any service processing EU data.

Google DeepMind's multimodal model family, announced in December 2023, in Pro, Flash and lighter tiers.

GeoBench · knowledge benchmark tracked on BenchGecko.

AI researcher.

GitHub Copilot is the AI pair programmer from GitHub and Microsoft.

AI application · Palo Alto, CA, USA · founded 2019.

Chip foundry · United States · founded 2009.

Gross Merchandise Value · total dollar value of transactions · used for marketplace AI products.

AI model lab · London, UK · founded 2010.

Block's open-source local coding agent · MCP-first, offline-capable, built by the company behind Square.

GPQA diamond · knowledge benchmark tracked on BenchGecko.

OpenAI's Generative Pre-trained Transformer model family, from GPT-1 (2018) to the current generation.

GPT Engineer generates an entire codebase from a natural language spec.

A GPU is the accelerator that trains and serves most large AI models.

semiconductor · Mountain View, CA, USA · founded 2016.

Revenue minus inference + training + data costs · AI labs at 50-70%, vs pure SaaS at 80-90%.

Tethering AI responses to verified source material · RAG, citations, tool calls, retrieval.

GSM8K · math benchmark tracked on BenchGecko.

GSO-Bench · coding benchmark tracked on BenchGecko.

Runtime filters and policies enforced around model input and output · the safety layer outside the model itself.

AI researcher.

NVIDIA GPU · Hopper SM90 architecture · 2022.

NVIDIA GPU · Hopper SM90 architecture · 2024.

When an AI generates plausible-sounding but factually incorrect or fabricated content.

AI researcher.

AI application · San Francisco, CA, USA · founded 2022.

HBM memory · 2020 · 460 GB/s per stack.

HBM memory · 2022 · 819 GB/s per stack.

HBM memory · 2024 · 1,180 GB/s per stack.

HBM memory · 2026 · 1,740 GB/s per stack.

HellaSwag · knowledge benchmark tracked on BenchGecko.

NVIDIA AI system · 8 × NVIDIA H100 SXM.

US law governing the handling of Protected Health Information (PHI) by healthcare providers and their vendors.

HLE · knowledge benchmark tracked on BenchGecko.

AI developer tools · New York, NY, USA · founded 2016.

A 164-problem Python benchmark where the model writes a function from its docstring and passes unit tests.

Wafer-to-wafer bonding technique that skips bumps · enables 3D stacking for HBM4 and advanced packaging.

Implicit-convolution alternative to attention · linear-time long-range modeling · explored by Stanford + Together.

Running a trained model to generate outputs from new inputs · what happens every time you call an AI API.

The next generation of AWS Inferentia, AWS's line of inference chips.

AI model lab · Palo Alto, CA, USA · founded 2022.

Tokens you send to the model · priced separately from output tokens, usually cheaper.

semiconductor · Santa Clara, CA, USA · founded 1968.

Chip foundry · United States · founded 2021.

The network of capital flowing into AI · which funds + corporates dominate which tier of deals.

energy · Sydney, Australia · founded 2018.

AI application · Austin, TX, USA · founded 2021.

AI researcher.

Jules is Google's asynchronous coding agent.

The date after which a model has no training data, so it does not know about later events.

Techniques to shrink the key-value cache during inference · sliding windows, quantization, eviction, sparsity.

AI developer tools · San Francisco, CA, USA · founded 2018.

LAMBADA · knowledge benchmark tracked on BenchGecko.

AI compute infrastructure · San Francisco, CA, USA · founded 2012.

AI developer tools · San Francisco, CA, USA · founded 2022.

LangGraph is LangChain's framework for building stateful, multi-actor agents as graphs.

Time to first token · how fast the AI starts responding. Critical for interactive UX.

Lech Mazur Writing · knowledge benchmark tracked on BenchGecko.

A contamination-resistant benchmark that refreshes tasks monthly to prevent models from memorizing answers.

Meta's open-weight model family, first released in February 2023.

Using a language model to grade other models' answers against a rubric.

LPDDR memory · 2022 · 34 GB/s per device.

Groq custom AI chip (ASIC) · TSP deterministic dataflow architecture · 2023.

Lifetime Value divided by Customer Acquisition Cost · >3× is healthy · AI consumer apps tracking 2-4×.

Microsoft custom AI chip (ASIC) · Custom tensor core + MX format architecture · 2024.

Microsoft AI system · 32 × Microsoft Maia 100.

State-space sequence model · linear-time alternative to transformers · foundation of Mamba-2, Jamba, Zamba models.

Butterfly Effect Inc's autonomous general-purpose agent · went viral March 2025 for multi-step research and web automation.

semiconductor · Wilmington, DE, USA · founded 1995.

MATH level 5 · math benchmark tracked on BenchGecko.

Model Context Protocol · an open standard for connecting AI models to external tools, data, and services.

AI model lab · Menlo Park, CA, USA · founded 2013.

big tech · Menlo Park, CA, USA · founded 2004.

MetaGPT models a software company with product managers, architects, engineers and QAs.

AMD GPU · CDNA 3 chiplet architecture · 2023.

AMD GPU · CDNA 3 chiplet architecture · 2025.

AMD AI system · 8 × AMD MI325X.

AMD GPU · CDNA 4 architecture · 2025.

memory chip · Boise, ID, USA · founded 1978.

big tech · Redmond, WA, USA · founded 1975.

Microsoft's successor to the Maia 100 AI accelerator.

AI application · San Francisco, CA, USA · founded 2021.

AI model lab · Shanghai, China · founded 2021.

Models from the Paris-based lab Mistral AI, founded in 2023: open-weight and commercial models.

AI model lab · Paris, France · founded 2023.

A model architecture that routes each token to a subset of specialized experts, so only a fraction of parameters activate per forward pass.

MMLU · knowledge benchmark tracked on BenchGecko.

A harder version of MMLU with 10 answer choices, filtered noise, and more reasoning-heavy questions.

AI compute infrastructure · New York, NY, USA · founded 2021.

When a model served under the same name starts behaving differently over time.

AI model lab · Beijing, China · founded 2023.

Meta custom AI chip (ASIC) · PE grid + SRAM architecture · 2024.

Per-token pricing differs by input type · text tokens vs image tokens vs audio seconds vs video frames.

Models that process and generate multiple types of data · text, image, audio, video.

AI compute infrastructure · Amsterdam, Netherlands · founded 2024.

semiconductor · Santa Clara, CA, USA · founded 1993.

Open weights are model weights released for download, inspection, fine-tuning, or self-hosting under a license.

AI model lab · San Francisco, CA, USA · founded 2015.

OpenBookQA · knowledge benchmark tracked on BenchGecko.

OpenHands (formerly OpenDevin) is a community-built autonomous software development agent that matches frontier closed-source agents on SWE-bench Verified.

OpenAI's computer-use agent · clicks, types, and browses websites on the user's behalf.

OSWorld · agentic benchmark tracked on BenchGecko.

OTIS Mock AIME 2024-2025 · math benchmark tracked on BenchGecko.

Tokens the model generates · priced 3-5× higher than input tokens due to sequential decoding.

Price-to-Sales ratio · valuation divided by annual revenue. The AI sector premium metric.

A flat fee per API call regardless of tokens · used for image generation, search, and some agent products.

AI application · San Francisco, CA, USA · founded 2022.

AI developer tools · New York, NY, USA · founded 2019.

PIQA · knowledge benchmark tracked on BenchGecko.

Plandex is an open-source terminal AI coding agent optimized for large, multi-file tasks.

PostTrainBench · knowledge benchmark tracked on BenchGecko.

The first and most expensive phase of LLM creation · training on trillions of tokens of internet text.

The practice of crafting model inputs to elicit better outputs · cheaper than fine-tuning, more reliable than hope.

semiconductor · San Diego, CA, USA · founded 1985.

Reducing the precision of model weights to shrink memory footprint and speed up inference.

Alibaba Cloud's model family, first released in 2023, with many open-weight sizes and API-only flagship models.

Retrieval-Augmented Generation · grounds model responses in external data retrieved at query time.

Anti-dilution provision that adjusts conversion price if a later round is at lower valuation · full vs weighted-average.

An LLM trained to spend extra compute at inference thinking before it answers, trading latency and cost for accuracy on hard tasks.

Reasoning models bill thinking tokens separately from output · can 2-10× effective cost per query.

AI model lab · San Francisco, CA, USA · founded 2022.

AI compute infrastructure · San Francisco, CA, USA · founded 2019.

AI developer tools · San Francisco, CA, USA · founded 2016.

Replit's in-browser AI coding agent · turns a prompt into a deployed app without leaving the browser.

Pre-purchased dedicated throughput · flat hourly fee for guaranteed tokens-per-second from a specific model.

Revenue per employee measures operating leverage · AI labs hit $1-5M/employee while traditional SaaS ranges $200-400K.

VSCode coding agent forked from Cline · adds auto-mode, multi-tab orchestration, and Orchestrator agent pattern.

AI application · New York, NY, USA · founded 2018.

AI researcher.

semiconductor · Palo Alto, CA, USA · founded 2017.

memory chip · Suwon, South Korea · founded 1969.

Chip foundry · South Korea · founded 2017.

AI developer tools · San Francisco, CA, USA · founded 2016.

The empirical relationship between compute, data, parameters, and model capability.

ScienceQA · knowledge benchmark tracked on BenchGecko.

AI researcher.

SimpleBench · reasoning benchmark tracked on BenchGecko.

SimpleQA Verified · knowledge benchmark tracked on BenchGecko.

memory chip · Icheon, South Korea · founded 1983.

Attention pattern where each token attends only to a local window · used in Mistral, Gemma for long-context efficiency.

Chip foundry · China · founded 2000.

Smolagents is Hugging Face's minimalist agent library.

AI developer tools · Redwood City, CA, USA · founded 2019.

An audit framework certifying that a provider handles customer data with defined security controls.

A faster way to generate text: a small model drafts tokens and the large model verifies them in one pass.

Spot pricing sells preemptible GPU capacity at 60-90% off on-demand · you can be evicted with 2 minutes notice but prices are dramatically lower.

AI model lab · London, UK · founded 2021.

AI model lab · Shanghai, China · founded 2023.

A delegated sub-task agent spawned by a parent orchestrator · each runs in isolated context with specialized tools.

SWE-agent is the research agent from the Princeton NLP group behind SWE-bench.

The real-world coding benchmark · AI resolves actual GitHub issues in open-source Python repos.

SWE-Bench verified · coding benchmark tracked on BenchGecko.

Sweep is an AI junior developer that turns bug reports and feature requests into code changes as pull requests, running entirely from GitHub issues.

AI researcher.

AI application · London, UK · founded 2017.

Self-hosted AI coding assistant · Copilot-style autocomplete you can run on your own GPU.

An inference parameter controlling output randomness · 0 = deterministic, higher = more varied.

Terminal Bench · coding benchmark tracked on BenchGecko.

Extra computation a model spends while answering, for example reasoning tokens, to get better results.

The Agent Company · agentic benchmark tracked on BenchGecko.

Total tokens per second a serving cluster can handle across all concurrent requests.

Volume discounts at defined usage thresholds · $X/M tokens below 1B tokens, $Y/M above.

AI compute infrastructure · San Francisco, CA, USA · founded 2022.

The algorithm that splits text into tokens before a model reads it · BPE, SentencePiece, Tiktoken.

The extra tokens, and cost, needed to write the same text in some languages compared with English.

The fundamental units language models read and generate · roughly 3/4 of an English word per token.

A model invoking external functions or APIs during generation · the foundation of AI agents.

Google TPU · Matrix MXU + VPU architecture · 2024.

Google AI system · 8,960 × Google TPU v5p.

Google TPU · Matrix MXU + VPU + SparseCore architecture · 2024.

Google AI system · 256 × Google TPU v6e.

Google TPU · Inference-first MXU architecture · 2025.

AWS's third-generation Trainium AI training chip, announced at re:Invent 2024.

AWS custom AI chip (ASIC) · NeuronCore v3 architecture · 2024.

The neural network architecture (Vaswani et al., 2017) behind every modern large language model.

TriviaQA · knowledge benchmark tracked on BenchGecko.

AWS AI system · 16 × AWS Trainium2.

chip foundry · Hsinchu, Taiwan · founded 1987.

VideoMME · multimodal benchmark tracked on BenchGecko.

Pre-negotiated flat rate applied to all consumption · typical enterprise AI contract shape.

VPCT · knowledge benchmark tracked on BenchGecko.

AI developer tools · Amsterdam, Netherlands · founded 2019.

AI developer tools · San Francisco, CA, USA · founded 2017.

WeirdML · coding benchmark tracked on BenchGecko.

Codeium's VSCode-fork IDE with "Cascade" multi-file agent · acquired by Cognition in 2025.

Windsurf is Codeium's purpose-built agentic IDE with Cascade, a flow-state collaborative agent that keeps context across your whole project.

Winogrande · knowledge benchmark tracked on BenchGecko.

Cerebras wafer-scale AI processor · Wafer-scale SRAM mesh architecture · 2024.

AI model lab · San Francisco, CA, USA · founded 2023.

AI model lab · Beijing, China · founded 2010.

AI researcher.

Zed is a Rust-built code editor from the Atom team.

Zed editor's built-in AI assistant · Rust-native, collaborative, MCP-first.

AI model lab · Beijing, China · founded 2019.