No terms match your search.
A
A theoretical AI system that surpasses the cognitive abilities of all humans in virtually every domain — scientific reasoning, social intelligence, creativity, strategic planning, and more. ASI goes beyond AGI (matching human intelligence) to something qualitatively different: an intelligence that could improve itself recursively and solve problems humans can't even formulate. No ASI exists, and there's no scientific consensus on whether one can or will be built.
Why it matters: ASI is where AI safety becomes existential. If you believe superintelligence is possible, alignment isn't just about making chatbots polite — it's about ensuring that a system smarter than all of humanity still acts in our interest. It's speculative, but the stakes are high enough that serious researchers take it seriously. Understanding ASI helps you evaluate claims about AI risk with more nuance.
A hypothetical AI system that can understand, learn, and perform any intellectual task that a human can — with the ability to transfer knowledge across domains without being specifically trained for each one. Unlike current AI, which excels at narrow tasks (generating text, classifying images), AGI would handle novel situations, reason abstractly, and adapt to any challenge. Whether AGI is imminent, decades away, or impossible is the most contentious debate in the field.
Why it matters: AGI is the North Star (or bogeyman) of the entire AI industry. It drives billions in investment, shapes safety research priorities, and dominates policy debates. Whether or not you believe AGI is near, the concept defines how companies like Anthropic, OpenAI, and DeepMind frame their missions — and understanding the debate helps you separate genuine progress from hype.
AI tools that help developers write, review, debug, and deploy code. From autocomplete (GitHub Copilot, Codeium) to full autonomous development (Claude Code, Cursor, Devin), coding assistants represent one of the most mature and widely adopted applications of LLMs. They work by predicting the next tokens of code given context from your codebase, documentation, and instructions.
Why it matters: AI coding assistants are the sharpest edge of AI's impact on knowledge work. Developers who use them report 30-50% productivity gains on routine tasks. But they also hallucinate APIs that don't exist, introduce subtle bugs, and can make developers dependent on tools they don't fully understand.
Using AI to perform tasks that previously required human intervention. This ranges from simple automation (auto-categorizing emails, generating reports) to complex autonomous workflows (AI agents that research, write, test, and deploy code). The key shift from traditional automation (rigid rules) to AI automation (flexible intelligence) is that AI can handle ambiguous, unstructured tasks.
Why it matters: Automation is the economic engine of AI adoption. Every enterprise buying AI is really buying automation — fewer humans doing repetitive work, faster processing, 24/7 operation. The question isn't whether AI will automate tasks, but which tasks, how fast, and what happens to the humans who used to do them.
The dual application of AI in cybersecurity: using AI to defend systems (threat detection, anomaly detection, automated incident response) and the new attack vectors AI creates (AI-generated phishing, automated vulnerability discovery, adversarial attacks on ML systems). The field is in an arms race where both attackers and defenders are increasingly AI-powered.
Why it matters: AI makes existing cyber threats faster and cheaper to execute — a phishing email written by an LLM is more convincing and costs nothing to personalize. But AI also enables defenses that would be impossible manually, like analyzing millions of network events per second for anomalies. Security teams that don't use AI will lose to attackers who do.
The frameworks, policies, laws, and organizational practices that guide how AI is developed, deployed, and used. This includes government regulation (the EU AI Act, executive orders), industry self-regulation (responsible scaling policies, model cards), corporate governance (AI ethics boards, usage policies), and international coordination on AI safety standards.
Why it matters: The technology is moving faster than the rules. Companies are shipping AI products into healthcare, criminal justice, and finance with minimal oversight. Governance is the attempt to set boundaries before something breaks badly enough to trigger a backlash that could set the entire field back.
The challenge of building and using AI systems without compromising personal data. This spans the entire lifecycle: training data that might contain private information, models that can memorize and regurgitate personal details, inference logs that track user behavior, and the fundamental tension between AI capability (which improves with more data) and privacy rights.
Why it matters: Every conversation with an AI is data. Every image you generate reveals your prompts. Every document you summarize passes through someone's servers. Privacy isn't just a legal checkbox (GDPR, CCPA) — it's a trust issue that determines whether individuals and enterprises will adopt AI for sensitive work.
The practice of protecting AI systems from adversarial attacks, data poisoning, prompt injection, model theft, and misuse — while also defending against AI-enabled threats like deepfakes and automated cyberattacks. AI security sits at the intersection of traditional cybersecurity and the unique vulnerabilities introduced by machine learning systems.
Why it matters: AI systems are simultaneously powerful tools and novel attack surfaces. A prompt injection can make your customer-support bot leak internal data. A poisoned training dataset can insert backdoors. As AI gets deployed in critical infrastructure, healthcare, and finance, security isn't optional — it's existential.
How AI providers charge for access to their models. The dominant model is per-token pricing — you pay for the number of tokens you send (input) and receive (output), with output tokens typically costing 3-5x more. Other models include per-request pricing, monthly subscriptions, committed-use discounts, and free tiers. The race to lower prices has been fierce, with costs dropping 10-100x in two years.
Why it matters: Pricing determines what you can build. An application that makes 10,000 API calls per day lives or dies by the per-token cost. Understanding pricing models, comparing providers, and optimizing token usage is a core skill for anyone building AI-powered products.
The full stack of hardware, software, and services required to train and deploy AI models at scale. This includes GPUs and custom chips, data centers, networking, storage, orchestration platforms (Kubernetes, Slurm), model serving frameworks (vLLM, TensorRT), and the cloud providers that package it all. AI infrastructure is where the abstract world of model architecture meets the very concrete world of power grids and cooling systems.
Why it matters: Infrastructure determines what's possible. The reason only a handful of companies can train frontier models isn't a lack of ideas — it's a lack of infrastructure. And the reason AI costs what it does for end users traces directly back to GPU availability, data center capacity, and inference serving efficiency.
Speech AI company building developer-friendly APIs for transcription, speaker detection, and audio understanding. Their Universal-2 model rivals OpenAI Whisper in accuracy while adding features like speaker diarization, sentiment, and topic detection out of the box.
Type: Speech AI • Founded 2017 • San Francisco, USA
AI safety company building Claude. Founded by former OpenAI researchers Dario and Daniela Amodei, Anthropic focuses on developing reliable, interpretable, and steerable AI systems.
Type: AI Safety Lab • Founded 2021 • San Francisco, USA
The cloud computing arm of Alibaba Group and creator of the Qwen model family. Qwen models are fully open-weights, multilingual, and among the most capable open models available.
Type: Cloud + AI • Founded 2009 • Hangzhou, China
An AI system that can autonomously plan and execute multi-step tasks, using tools (web search, code execution, API calls) to achieve a goal. Unlike a simple chatbot that answers one question at a time, an agent decides what to do next based on what it's learned so far.
Why it matters: Agents are the bridge between "AI that talks" and "AI that does." When your AI can browse docs, write code, and test it without you holding its hand at every step — that's an agent.
The challenge of making AI systems behave in ways that match human values and intentions. An aligned model does what you mean, not just what you said — and avoids harmful actions even when not explicitly told not to.
Why it matters: A model that's technically brilliant but poorly aligned is like a genius employee who follows instructions too literally. Alignment research is why models refuse dangerous requests and try to be genuinely helpful.
A structured way for software to talk to other software. In AI, this usually means sending a request (your prompt) to a provider's server and getting a response (the model's output) back. REST APIs over HTTPS are the standard.
Why it matters: Every AI provider — Anthropic, Google, Mistral — exposes their models through APIs. If you're building anything with AI beyond a chat window, you're using an API.
The core mechanism in Transformers that lets a model weigh which parts of the input are most relevant to each other. Instead of reading text left-to-right like older models, attention lets every word "look at" every other word simultaneously to understand context.
Why it matters: Attention is why modern LLMs understand that "bank" means different things in "river bank" vs. "bank account." It's also why longer context windows cost more — attention scales quadratically with sequence length.
B
Israeli AI company that built its image generation models exclusively on licensed, attributed training data. Positions itself as the safe choice for enterprises that need AI-generated visuals without copyright risk.
Type: Responsible Image AI • Founded 2020 • Tel Aviv, Israel
Parent company of TikTok and one of the world's most valuable tech companies. Their AI lab builds the Doubao model family and powers recommendation algorithms that serve over a billion users daily.
Type: Tech Conglomerate • Founded 2012 • Beijing, China
Founded by the original creators of Stable Diffusion after leaving Stability AI. Their FLUX models quickly became the new standard for open-source image generation, surpassing the quality of the models they left behind.
Type: Image AI • Founded 2024 • Freiburg, Germany
A standardized test used to evaluate and compare AI models. Benchmarks measure specific capabilities — reasoning (ARC), math (GSM8K), coding (HumanEval), general knowledge (MMLU) — and produce scores that can be compared across models.
Why it matters: Benchmarks are how the industry keeps score, but they're imperfect. Models can be trained to ace benchmarks without being genuinely better. Real-world performance often tells a different story. Treat them as signals, not truth.
Systematic patterns in AI outputs that reflect or amplify societal prejudices present in training data. Bias can appear in text generation, image creation, hiring tools, and anywhere models make decisions that affect people differently.
Why it matters: If the training data says nurses are women and engineers are men, the model will perpetuate that. Bias isn't always obvious — it hides in word associations, default assumptions, and who gets represented.
C
The field of AI focused on enabling machines to interpret and understand visual information from the world — images, video, 3D scenes, and documents. Computer vision powers everything from facial recognition and autonomous driving to medical imaging and AI image generation. Core tasks include object detection, image classification, segmentation, OCR, and pose estimation.
Why it matters: Computer vision was the first area where deep learning clearly surpassed human performance (ImageNet 2012), and it remains one of the most commercially impactful AI applications. Every AI image or video you generate, every document you OCR, every security camera with smart detection — it's all computer vision.
Using AI to detect and filter harmful, illegal, or policy-violating content at scale. This includes text classification (hate speech, spam, threats), image analysis (NSFW detection, CSAM), and video moderation. Modern systems combine AI classifiers with human review, but the volume of content generated by AI itself is creating a moderation crisis — you now need AI to moderate AI.
Why it matters: Every platform with user-generated content needs moderation, and AI is the only way to handle the scale. But moderation is harder than it sounds — context matters, cultural norms differ, and false positives silence legitimate speech while false negatives let harm through.
Voice AI startup built on state space model (SSM) architecture rather than transformers. Their Sonic models achieve ultra-low latency voice generation, making real-time conversational AI feel genuinely natural for the first time.
Type: Real-time Voice AI • Founded 2024 • San Francisco, USA
Enterprise-focused AI company co-founded by Aidan Gomez, one of the co-authors of the original "Attention Is All You Need" Transformer paper. Specializes in models optimized for business use cases, RAG, and multilingual support.
Type: Enterprise AI • Founded 2019 • Toronto, Canada
A prompting technique where you ask the model to show its reasoning step by step before giving a final answer. Instead of jumping to a conclusion, the model "thinks out loud," which dramatically improves accuracy on complex tasks.
Why it matters: Asking "explain your reasoning" isn't just for transparency — it actually makes models smarter. CoT reduced math errors by up to 50% in early studies. Most modern models now do this internally.
The maximum amount of text (measured in tokens) a model can process in a single conversation. This includes both your input and the model's output. If a model has a 200K context window, that's roughly 150,000 words — about two novels.
Why it matters: Context window size determines what you can do. Summarize a whole codebase? Needs big context. Quick question-answer? Small is fine. But bigger isn't always better — models can lose focus in very long contexts.
The body of text (or other data) used to train a model. A corpus can range from curated collections of books and papers to massive scrapes of the entire internet. The quality and composition of the corpus fundamentally shapes what the model knows and how it behaves.
Why it matters: Garbage in, garbage out. A model trained on Reddit talks differently than one trained on scientific papers. This is why we curated our own corpus for Sarah — generic web crawls produced confused, incoherent results.
D
A subset of machine learning that uses neural networks with many layers (hence "deep") to learn hierarchical representations of data. Each layer transforms its input into something slightly more abstract — from pixels to edges to shapes to objects to concepts. Deep learning is what made the modern AI revolution possible: it's the approach behind LLMs, image generators, speech recognition, and virtually every AI breakthrough since 2012.
Why it matters: Deep learning is the engine of the current AI era. Before 2012, AI was a patchwork of specialized algorithms. Deep learning unified everything under one paradigm: stack enough layers, feed enough data, throw enough compute at it, and the model figures out the rest. Understanding deep learning is understanding why AI suddenly works.
AI-generated images, video, or audio designed to convincingly depict real people saying or doing things they never did. Originally built on GAN technology, modern deepfakes use diffusion models and voice cloning to produce outputs that are increasingly difficult to distinguish from reality. Detection tools exist but consistently lag behind generation capabilities.
Why it matters: Deepfakes are the dark side of generative AI's creative power. They've been used for fraud, non-consensual intimate imagery, political manipulation, and identity theft. The technology is now accessible enough that anyone with a laptop can create convincing fakes, making detection, watermarking, and legal frameworks urgent priorities.
Physical facilities that house the servers, GPUs, networking equipment, and cooling systems needed to train and run AI models. Modern AI data centers are purpose-built for massive parallel computation, consuming megawatts of power and requiring specialized cooling. A single frontier model training run might occupy thousands of GPUs across an entire facility for months.
Why it matters: Data centers are the factories of the AI era. Every query to Claude, every image from Midjourney, every video from Runway runs on hardware sitting in one of these buildings. The global shortage of AI-ready data center capacity is one of the biggest constraints on AI growth — and one of the biggest investment opportunities.
German AI company widely regarded as the best machine translation service in the world. Built by a team of computational linguists who consistently outperform Google Translate and other big-tech offerings, especially for European languages.
Type: Translation AI • Founded 2017 • Cologne, Germany
Israeli AI company pushing the boundaries of real-time AI generation. Their technology can generate interactive game-like environments in real-time, blurring the line between traditional rendering and AI generation.
Type: Real-time AI • Founded 2023 • Tel Aviv, Israel
Chinese AI lab that shook the industry in early 2025 with DeepSeek-R1, a reasoning model rivaling frontier labs at a fraction of the training cost. Backed by quantitative hedge fund High-Flyer.
Type: AI Research Lab • Founded 2023 • Hangzhou, China
Speech AI company building fast, accurate speech recognition and text-to-speech APIs. Their Nova models compete with and often beat OpenAI's Whisper on accuracy while running significantly faster for real-time applications.
Type: Speech AI • Founded 2015 • San Francisco, USA
A type of generative model that creates images (or video, audio) by starting with pure noise and gradually removing it until a coherent output appears. The model learns to reverse the process of adding noise to real data. Stable Diffusion, DALL-E 3, and Midjourney all use variants of this approach.
Why it matters: Diffusion models dethroned GANs as the dominant image generation technique around 2022. They produce more diverse, controllable outputs and are the backbone of almost every image and video AI tool today.
E
Capabilities that appear in AI models at scale but were not explicitly trained for — abilities that seem to "emerge" suddenly once a model reaches a certain size or training threshold. A model trained purely to predict the next word somehow learns to do arithmetic, translate between languages it wasn't taught, or write working code. Emergence is one of the most debated phenomena in AI: is it real phase-transition magic, or a measurement artifact?
Why it matters: Emergence is at the heart of the biggest question in AI: can we predict what larger models will be able to do? If capabilities truly emerge unpredictably at scale, then every bigger model is a surprise box. If emergence is an artifact of how we measure, then scaling is more predictable than it seems. The answer shapes everything from safety planning to investment decisions.
The methods used to measure how well an AI model performs. This goes far beyond benchmarks — it includes human evaluation (having people rate outputs), A/B testing (comparing models on real traffic), red teaming (adversarial testing), domain-specific testing (medical accuracy, code correctness), and community leaderboards (Chatbot Arena, LMSYS). Good evaluation is harder than building the model.
Why it matters: If you can't measure it, you can't improve it. But AI evaluation is uniquely hard because the tasks are open-ended and quality is subjective. Benchmarks get gamed, human eval is expensive, and the model that scores highest on paper often isn't the best in practice. Building good evals is a superpower.
Voice AI company that made ultra-realistic speech synthesis accessible to everyone. Their technology powers voice cloning, real-time dubbing, and text-to-speech across 32 languages, blurring the line between human and AI voices.
Type: Voice AI • Founded 2022 • New York, USA
A way to represent text (or images, or audio) as a list of numbers (a vector) that captures its meaning. Similar concepts end up close together in this number space — "cat" and "kitten" are nearby, while "cat" and "economics" are far apart.
Why it matters: Embeddings are the foundation of semantic search and RAG. They're how AI understands that a search for "fix login bug" should match a document about "authentication error resolution" even though no words overlap.
A specific URL where an AI API accepts requests. For example, Anthropic's message endpoint is where you send prompts to Claude. Different endpoints serve different functions: text generation, embeddings, image creation, model listing.
Why it matters: When integrating AI providers, endpoints are where rubber meets road. Each provider structures theirs differently, which is why platforms like Zubnet exist — to normalize the mess.
G
AI systems that create new content — text, images, audio, video, code, 3D models — rather than just analyzing or classifying existing data. Generative AI is the umbrella term for everything from ChatGPT writing essays to Stable Diffusion creating images to Suno composing music. The "generative" part distinguishes these models from earlier AI that could only categorize, predict, or recommend.
Why it matters: Generative AI is the term that brought AI into mainstream culture. It's what people mean when they say "AI" in 2024-2026 — the ability to create, not just compute. Understanding it as a category helps you navigate the landscape: LLMs generate text, diffusion models generate images, and the boundaries between modalities are rapidly blurring.
Google's unified AI research division, formed by merging DeepMind and Google Brain in 2023. Behind Gemini, AlphaGo, AlphaFold, and much of the foundational research that powers modern AI.
Type: AI Research Lab • Founded 2010 / merged 2023 • London, UK & Mountain View, USA
A model architecture where two neural networks compete: a generator creates fake data, and a discriminator tries to tell real from fake. Through this adversarial game, the generator gets better at creating realistic outputs. Dominated image generation from 2014 to ~2022.
Why it matters: GANs pioneered realistic AI image generation and are still used in some real-time applications. But diffusion models have largely replaced them for quality-critical work because GANs are harder to train and less diverse in their outputs.
Originally designed for rendering graphics, GPUs turned out to be perfect for AI because they can do thousands of math operations simultaneously. Training and running AI models is essentially massive matrix multiplication — exactly what GPUs are built for. NVIDIA dominates this market.
Why it matters: GPUs are the physical bottleneck of the entire AI industry. Why models cost what they cost, why some providers are faster than others, why there's a global chip shortage — it all comes back to GPU supply and VRAM.
Connecting a model's responses to factual, verifiable sources rather than letting it rely solely on its training data. Grounding techniques include RAG, web search integration, and citation requirements. A grounded response says "according to [source]" rather than just asserting facts.
Why it matters: Grounding is the primary defense against hallucination. An ungrounded model confidently invents facts. A grounded one points you to real sources you can verify.
Safety mechanisms that prevent AI models from generating harmful, inappropriate, or off-topic content. Guardrails can be built into the model during training (RLHF), applied through system prompts, or enforced by external filters that check outputs before they reach users.
Why it matters: Without guardrails, models will happily help with dangerous requests. The challenge is calibration — too strict and the model becomes useless ("I can't help with that"), too loose and it becomes unsafe.
H
Settings you choose before training begins that control how the model learns — as opposed to parameters, which the model learns on its own. Hyperparameters include learning rate (how big each update step is), batch size (how many examples to process at once), number of epochs (how many times to go through the data), optimizer choice (Adam, SGD, AdamW), weight decay, dropout rate, and architecture decisions like number of layers and hidden dimensions. Getting hyperparameters right is often the difference between a model that converges beautifully and one that diverges into nonsense.
Why it matters: Hyperparameter tuning is where ML engineering becomes part science, part craft. You can have the perfect dataset and architecture, but a learning rate that's too high will blow up training and one that's too low will never converge. Understanding hyperparameters is essential for anyone training or fine-tuning models — and knowing which ones matter most saves enormous amounts of compute.
AI video platform specializing in realistic talking-head avatars and automatic lip-sync dubbing. Used by enterprises for marketing, training, and localization — turning one video into dozens of languages with matching lip movements.
Type: Video Avatars • Founded 2020 • Los Angeles, USA
Emerging image generation company building high-quality diffusion models. Their open-weights releases have gained traction in the creative AI community for strong prompt adherence and visual quality.
Type: Image AI • Founded 2024 • San Francisco, USA
AI company building models that understand and express human emotion. Their Empathic Voice Interface detects tone, sentiment, and emotional context in real-time, enabling AI conversations that respond not just to what you say but how you say it.
Type: Emotion AI • Founded 2021 • New York, USA
When an AI model generates information that sounds confident and plausible but is factually wrong or entirely fabricated. The model isn't "lying" — it's pattern-matching its way to fluent text without a concept of truth. Fake citations, invented statistics, and non-existent API methods are common examples.
Why it matters: Hallucination is the single biggest trust problem in AI today. It's why you should always verify critical facts from AI outputs, and why techniques like RAG and grounding exist.
L
Australian AI image platform that carved out a niche between Midjourney and Stable Diffusion. Popular with game developers and digital artists for its fine-tuned models, real-time canvas, and focus on production-ready creative assets.
Type: Image AI Platform • Founded 2022 • Sydney, Australia
MIT spinout exploring fundamentally different neural network architectures inspired by biological neural circuits. Their Liquid Foundation Models use continuous-time dynamics rather than fixed-weight transformers, promising better efficiency and adaptability.
Type: AI Research • Founded 2023 • Cambridge, USA
AI company focused on video and 3D generation. Their Dream Machine was one of the first accessible, high-quality AI video generators, and Ray2 pushed video quality and coherence significantly forward.
Type: Video & 3D AI • Founded 2021 • San Francisco, USA
The delay between sending a request and getting the first response. In AI, this is often measured as Time to First Token (TTFT) — how long before the model starts streaming its answer. Affected by model size, server load, network distance, and prompt length.
Why it matters: Users perceive anything over ~2 seconds as slow. Low latency is why smaller models often win for real-time applications even when larger models are "smarter." It's a key differentiator between providers.
A neural network trained on massive amounts of text to understand and generate human language. "Large" refers to the number of parameters (billions) and the size of the training data (trillions of tokens). Claude, GPT, Gemini, Llama, and Mistral are all LLMs.
Why it matters: LLMs are the technology behind every AI chat, code assistant, and text generator you use. Understanding what they are (statistical pattern matchers, not sentient beings) helps you use them effectively and recognize their limits.
A technique that makes fine-tuning dramatically cheaper by only training a small number of additional parameters instead of modifying the entire model. LoRA "adapters" are lightweight add-ons (often just megabytes) that modify a model's behavior without retraining its billions of parameters.
Why it matters: LoRA democratized fine-tuning. Before it, customizing a 7B model required serious GPU resources. Now you can fine-tune on a single consumer GPU in hours and share the tiny adapter file. It's why there are thousands of specialized models on HuggingFace.
M
A trained mathematical system that takes inputs and produces outputs based on patterns learned from data. In AI, "model" is the catch-all term for the thing you're actually using — whether it's GPT-4 generating text, Stable Diffusion generating images, or Whisper transcribing speech. A model is defined by its architecture (how it's structured), its parameters (what it learned), and its training data (what it learned from). When someone says "which model should I use?" they're asking about this.
Why it matters: Model is the single most used word in AI, and it means different things in different contexts. A "model" can refer to the architecture (Transformer), a specific trained instance (Claude Opus 4.6), a file on disk (a .gguf file), or an API endpoint. Understanding what a model actually is — and what it isn't — is the foundation for everything else.
The broad field of computer science where systems learn patterns from data rather than following explicit rules. Instead of programming a computer to recognize a cat by listing features (four legs, pointy ears, whiskers), you show it thousands of cat photos and let it figure out the pattern itself. Machine learning encompasses everything from simple linear regression to the deep neural networks powering today's AI — supervised learning (labeled examples), unsupervised learning (finding structure), and reinforcement learning (trial and error).
Why it matters: Machine learning is the foundation under everything we call "AI" today. Every LLM, every image generator, every recommendation algorithm, every spam filter — it's all machine learning. Understanding ML as the broader discipline helps you see where deep learning fits, where classical methods still win, and why "AI" is really just "ML that got really good."
Mechanisms that allow AI models to retain and recall information beyond a single conversation. This includes in-context memory (using the context window), external memory (RAG, vector databases), persistent conversation memory (remembering user preferences across sessions), and working memory (maintaining state during multi-step agent tasks). Memory is what makes AI feel like a collaborator rather than a stateless tool.
Why it matters: Without memory, every AI conversation starts from zero. You repeat your preferences, re-explain your codebase, re-describe your project. Memory is what turns a chatbot into an assistant — and it's one of the hardest problems to solve well, balancing relevance, privacy, staleness, and storage costs.
Chinese AI company that made waves by launching Kimi, a chatbot with a 2-million-token context window. Founded by Yang Zhilin, a former researcher behind key innovations in long-context modeling.
Type: AI Company • Founded 2023 • Beijing, China
European AI powerhouse founded by former DeepMind and Meta researchers. Known for punching above their weight with efficient models and championing open-weights distribution alongside commercial offerings.
Type: AI Company • Founded 2023 • Paris, France
Chinese AI company building large-scale models across text, voice, and video. Known for their Hailuo consumer platform and increasingly competitive multimodal models.
Type: AI Company • Founded 2021 • Shanghai, China
An open protocol (created by Anthropic) that standardizes how AI models connect to external tools and data sources. Think of it as USB-C for AI — one standard interface instead of custom integrations for every tool. MCP servers expose capabilities; MCP clients (like Claude) consume them.
Why it matters: Before MCP, every AI-tool integration was bespoke. MCP means a tool built once works with any compatible AI. It's already supported by Claude, Cursor, and others. This is how AI goes from chatbot to actual assistant.
An architecture where the model contains multiple "expert" sub-networks, but only activates a few of them for each input. A router network decides which experts are relevant for a given token. This means a model can have 100B+ total parameters but only use 20B for any single forward pass.
Why it matters: MoE is how models like Mixtral and (reportedly) GPT-4 get the quality of a huge model with the speed of a smaller one. The trade-off is higher memory usage (all experts must be loaded) even though computation is cheaper.
A model that can understand and/or generate multiple types of data: text, images, audio, video, code. Claude can read images and text; some models can also produce images or speech. "Multimodal" contrasts with "unimodal" models that only handle one type.
Why it matters: Real-world tasks are multimodal. You want to show an AI a screenshot and ask "what's wrong here?" or give it a diagram and say "implement this." Multimodal models make that possible.
N
The branch of AI focused on enabling machines to understand, interpret, and generate human language. NLP covers everything from basic text processing (tokenization, stemming, part-of-speech tagging) to complex tasks like sentiment analysis, machine translation, summarization, and question answering. Before Transformers, NLP was a patchwork of specialized techniques. Now, LLMs have unified most of NLP under one paradigm — but the field's foundations still matter for understanding how and why these models work.
Why it matters: NLP is the reason you can talk to AI in plain English and get useful answers back. Every chatbot, every search engine, every translation service, every AI writing tool is NLP. Even if you never build an NLP system from scratch, understanding the fundamentals — tokenization, attention, embeddings, context — makes you a better user of every AI tool that handles text.
The company whose GPUs power virtually all AI training and most inference worldwide. What started as a graphics card company became the most critical hardware supplier in the AI industry, briefly making NVIDIA the most valuable company on Earth.
Type: Hardware & AI • Founded 1993 • Santa Clara, USA
A computing system loosely inspired by biological brains, made of layers of interconnected "neurons" (mathematical functions) that learn patterns from data. Information flows through layers, getting progressively transformed until the network produces an output. Every modern AI model is a neural network of some kind.
Why it matters: Neural networks are the "how" behind all of AI. Understanding that they're math (not magic, not brains) helps demystify what AI can and can't do. They're pattern matchers — extraordinarily powerful ones, but pattern matchers nonetheless.
O
The broad set of techniques used to make AI models faster, smaller, cheaper, or more accurate. This includes training optimizations (mixed precision, gradient checkpointing, data parallelism), inference optimizations (quantization, pruning, distillation, speculative decoding), and serving optimizations (batching, caching, load balancing). Optimization is the reason you can run a 14B parameter model on a laptop.
Why it matters: Raw capability means nothing if you can't afford to run it. Optimization is the difference between a research demo and a production product. It's why open-weights models can compete with API providers, why mobile AI exists, and why inference costs keep dropping.
The company behind ChatGPT and the GPT series of models. Originally a non-profit research lab, OpenAI became the public face of the AI revolution when ChatGPT launched in November 2022.
Type: AI Research Lab • Founded 2015 • San Francisco, USA
When a company releases a model's trained parameters for anyone to download and run. "Open weights" is more accurate than "open source" because most released models don't include training data or training code — you get the finished model but not the recipe. Llama, Mistral, and Qwen are open-weights models.
Why it matters: Open weights mean you can run AI on your own hardware with full privacy — no API calls, no data leaving your network. The trade-off is you need the GPU resources to run them and you're responsible for safety.
When a model memorizes its training data too well and loses the ability to generalize to new inputs. Like a student who memorizes answers to practice tests but can't solve new problems. The model performs great on training data but poorly on anything it hasn't seen before.
Why it matters: Overfitting is the most common failure mode in model training. It's why evaluation uses separate test sets, and why training for too long (too many epochs) can actually make a model worse.
P
The internal values a neural network learns during training — essentially the "knowledge" of the model encoded as numbers. When someone says a model has "7 billion parameters," they mean 7 billion individual numerical values that were adjusted during training to capture patterns in the data. More parameters generally means more capacity to learn complex patterns, but also more memory to store and more compute to run.
Why it matters: Parameter count is the most common shorthand for model size, and it directly determines how much GPU memory you need. A 7B model in 16-bit precision needs ~14GB of VRAM just for the weights. Understanding parameters helps you estimate costs, choose hardware, and understand why quantization (reducing precision per parameter) is so important for making models accessible.
Chinese video generation company building accessible AI video tools. Known for fast generation speeds and a free tier that helped them build a large user base quickly across international markets.
Type: Video AI • Founded 2023 • Shenzhen, China
AI search engine that combines real-time web search with language model reasoning to give direct, sourced answers instead of a list of links. The most visible challenge to Google's search dominance in a generation.
Type: AI Search • Founded 2022 • San Francisco, USA
The initial, massive training phase where a model learns language (or other modalities) from a huge corpus. This is the expensive part — thousands of GPUs running for weeks or months, costing millions of dollars. The result is a foundation model that understands language but hasn't been specialized for any task yet.
Why it matters: Pre-training is what makes foundation models possible. It's also why only a handful of companies can create frontier models — the compute costs are astronomical. Everything else (fine-tuning, RLHF, prompting) builds on this base.
The practice of crafting inputs to get better outputs from AI models. This ranges from simple techniques (being specific, providing examples) to advanced methods (chain of thought, few-shot prompting, role assignment). Despite the fancy name, it's fundamentally about communicating clearly with a statistical system.
Why it matters: The same model can give wildly different results depending on how you ask. Good prompt engineering is the cheapest way to improve AI output quality — no training, no fine-tuning, just better communication.
R
A training paradigm where an AI agent learns by interacting with an environment, taking actions, and receiving rewards or penalties. Unlike supervised learning (which learns from labeled examples), RL learns from experience — through trial and error. RL trained AlphaGo to beat world champions, teaches robots to walk, and is the "RL" in RLHF that makes chatbots helpful.
Why it matters: Reinforcement learning is how AI learns to act, not just predict. It's the bridge between models that can answer questions and agents that can accomplish goals. Every AI system that plans, strategizes, or optimizes over time has RL somewhere in its lineage.
The ability of AI models to think step-by-step, decompose complex problems, and arrive at logically sound conclusions. Modern reasoning models (like OpenAI's o1/o3 and DeepSeek-R1) are trained to generate explicit reasoning traces before answering, dramatically improving performance on math, coding, and logic tasks. This is distinct from simple pattern matching — reasoning models can solve problems they've never seen before.
Why it matters: Reasoning is the frontier capability that separates "AI that sounds smart" from "AI that is smart." Models that reason well can debug code, prove theorems, plan multi-step strategies, and catch their own mistakes. The gap between models with and without strong reasoning is the biggest quality differentiator in AI right now.
Canadian voice AI company specializing in high-fidelity voice cloning and real-time speech synthesis. One of the first to ship neural audio watermarking for deepfake detection, taking the ethical implications of voice cloning seriously from the start.
Type: Voice AI • Founded 2019 • Toronto, Canada
AI research company founded by former DeepMind, Google Brain, and FAIR researchers. Building natively multimodal models that can process text, images, video, and audio from the ground up.
Type: AI Research • Founded 2023 • San Francisco, USA
AI design tool focused on professional-grade image and vector graphic generation. One of the first to produce truly usable design assets — SVGs, brand-consistent styles, and production-ready outputs that designers actually want to use.
Type: Design AI • Founded 2021 • San Francisco, USA
Pioneering AI video generation company. Co-created the original Stable Diffusion architecture and then pivoted to video, where their Gen series models have defined the state of the art for AI filmmaking tools.
Type: Creative AI • Founded 2018 • New York, USA
A technique that gives AI models access to external knowledge by retrieving relevant documents before generating a response. Instead of relying only on what the model learned during training, RAG searches a knowledge base, finds relevant chunks, and includes them in the prompt as context.
Why it matters: RAG solves two major problems: hallucination (the model has real sources to reference) and knowledge cutoff (the knowledge base can be updated without retraining). It's how most enterprise AI actually works.
Restrictions on how many API requests you can make per minute/hour/day. Providers impose rate limits to prevent server overload and ensure fair access. Limits typically apply per API key and can restrict requests per minute (RPM) and tokens per minute (TPM).
Why it matters: Rate limits are the invisible ceiling you hit when scaling AI applications. They're why batch processing matters, why you need retry logic, and why some providers charge more for higher rate limits.
The practice of deliberately trying to make an AI model fail, misbehave, or produce harmful outputs. Red teams probe for vulnerabilities: jailbreaks, bias, misinformation generation, privacy leaks. Named after military wargaming where a "red team" plays the adversary.
Why it matters: You can't fix what you don't know about. Red teaming is how providers discover that their model will explain how to pick locks if you ask it to "write a story about a locksmith." It's essential safety work that happens before every major model release.
A training technique where human evaluators rank model outputs by quality, and this feedback is used to train a reward model that guides the AI toward better responses. It's what turns a raw pre-trained model (which just predicts next words) into a helpful, harmless assistant.
Why it matters: RLHF is the secret ingredient that made ChatGPT feel different from GPT-3. The base model already "knew" everything, but RLHF taught it to present that knowledge in a way humans actually find useful. It's also how safety behaviors are reinforced.
S
The tendency of AI models to tell users what they want to hear rather than what's true. A sycophantic model agrees with incorrect premises, validates bad ideas, flips its position when challenged even if it was right the first time, and prioritizes being liked over being helpful. Sycophancy is a direct side effect of RLHF training — models learn that agreeable responses get higher ratings from human evaluators, so they optimize for agreement over accuracy.
Why it matters: Sycophancy is one of the most insidious failure modes in AI because it's invisible to the user who's being flattered. If you ask a model "isn't this a great business idea?" and it always says yes, you're getting a mirror, not an advisor. Combating sycophancy is an active area of alignment research, and it's why the best models are trained to respectfully disagree when they should.
A critique of large language models arguing that they are merely sophisticated pattern matchers that stitch together plausible-sounding text without any understanding of meaning. The term was coined by Emily Bender, Timnit Gebru, and colleagues in their influential 2021 paper "On the Dangers of Stochastic Parrots," which warned that LLMs encode biases from their training data, consume enormous resources, and create an illusion of comprehension that misleads users into trusting them more than they should.
Why it matters: The stochastic parrot debate goes to the heart of what AI actually "understands." Whether LLMs are genuinely reasoning or just incredibly good at statistical mimicry shapes how we deploy them, how much we trust their outputs, and how we regulate them. It's also the lens through which critics evaluate every new capability claim — is this real progress or a more convincing parrot?
Low-quality, generic, unwanted AI-generated content that floods the internet. The term emerged in 2024 as a pejorative for the tide of mediocre AI text, images, and video polluting search results, social media feeds, and online marketplaces. Slop is the AI equivalent of spam — technically "content" but adding no value, often indistinguishable from other slop, and degrading the quality of every platform it touches. Think LinkedIn posts that start with "In today's fast-paced world," stock photos with six-fingered hands, or SEO articles that say nothing in 2,000 words.
Why it matters: Slop is the environmental cost of making content generation free. When anyone can generate 1,000 blog posts or 10,000 product images in minutes, the economics of content creation collapse — and quality collapses with them. Slop is why platforms are racing to build AI detection, why Google keeps updating its search algorithm, and why "human-made" is becoming a selling point. It's also the strongest argument against the naive "AI will democratize creativity" narrative.
Chinese AI startup building competitive large language and multimodal models. Their Step series has shown strong performance on international benchmarks, backed by significant compute investment.
Type: AI Company • Founded 2023 • Shanghai, China
AI hardware company that designs custom chips (RDUs) purpose-built for AI workloads. Their SambaNova Cloud offers some of the fastest inference speeds available, competing with Groq on the "speed-first" approach to AI serving.
Type: AI Hardware & Inference • Founded 2017 • Palo Alto, USA
Indian AI company building models specifically optimized for India's linguistic diversity. Their models handle Hindi, Tamil, Telugu, Bengali, and other Indian languages with a fluency that global models consistently struggle with.
Type: Indian AI • Founded 2023 • Bangalore, India
The company that democratized image generation by releasing Stable Diffusion as open-source in 2022. Despite leadership turbulence, their models remain the backbone of the open-source image generation ecosystem.
Type: Open-Source AI • Founded 2019 • London, UK
AI music generation company that lets anyone create full songs — vocals, instruments, production — from a text prompt. Went from unknown to millions of users in months, forcing the music industry to confront AI creativity head-on.
Type: Music AI • Founded 2023 • Cambridge, USA
An alternative to Transformers that processes sequences by maintaining a compressed "state" instead of using attention over all tokens. Mamba is the most well-known SSM architecture. SSMs scale linearly with sequence length (vs. quadratic for attention), making them potentially much more efficient for very long contexts.
Why it matters: SSMs are the main challenger to Transformer dominance. They're faster for long sequences and use less memory, but the research is still maturing. Hybrid architectures (mixing SSM layers with attention) may end up being the best of both worlds.
A special instruction given to a model at the start of a conversation that sets its behavior, personality, and rules. Unlike user messages, the system prompt is meant to be persistent and authoritative — it defines who the model is for this session. "You are a helpful coding assistant. Always use TypeScript."
Why it matters: System prompts are the primary tool for customizing AI behavior without fine-tuning. They're how companies make Claude act as a customer support agent, a code reviewer, or a medical information assistant — same model, different system prompt.
T
Chinese tech giant behind WeChat, one of the world's largest gaming companies, and increasingly a force in generative AI. Their Hunyuan models power features across Tencent's massive ecosystem serving over a billion users.
Type: Tech Conglomerate • Founded 1998 • Shenzhen, China
Video understanding company that lets you search, analyze, and generate content from video using natural language. Think of it as "RAG for video" — their models understand what happens in a video the way LLMs understand text.
Type: Video Understanding • Founded 2021 • San Francisco, USA
AI company specializing in generating 3D models from text or images. In a field where most 3D generation produces unusable blobs, Tripo stands out for generating clean, production-ready meshes that game developers and designers can actually work with.
Type: 3D AI • Founded 2023 • Beijing, China
A parameter that controls how random or deterministic a model's output is. Temperature 0 makes the model always pick the most probable next token (deterministic, focused). Temperature 1+ makes it more willing to pick less probable tokens (creative, unpredictable). Most APIs default to around 0.7.
Why it matters: Temperature is the creativity dial. Writing fiction? Turn it up. Generating code or factual answers? Turn it down. It's one of the most impactful parameters you can adjust, and it costs nothing to experiment with.
The basic unit of text that AI models process. A token is typically a word or word fragment — "understanding" might be one token, while "un" + "der" + "standing" could be three. On average, one token is roughly 3/4 of a word in English. Models read, think, and charge in tokens.
Why it matters: Tokens are the currency of AI. Context windows are measured in tokens. API pricing is per token. When a provider says "1M context" they mean 1 million tokens, roughly 750K words. Understanding tokens helps you estimate costs and optimize usage.
V
AI systems for generating, understanding, and manipulating human speech. This includes text-to-speech (TTS), speech-to-text (STT/ASR), voice cloning, real-time voice translation, emotion detection in speech, and conversational voice agents. The field has advanced to the point where AI-generated speech is often indistinguishable from human speech.
Why it matters: Voice is the most natural human interface, and AI is finally making it programmable. Voice AI powers everything from customer service bots to audiobook narration to real-time meeting transcription. The ethical implications of voice cloning — consent, identity, fraud — make this one of the most sensitive areas in AI.
Video generation platform from Shengshu Technology, producing some of the most physically coherent AI-generated videos. Gained attention for strong motion quality and multi-shot consistency that rivals Western competitors.
Type: Video AI • Founded 2024 • Beijing, China
Embedding model company building specialized vectors for code, legal, finance, and multilingual search. Their models consistently rank at the top of the MTEB leaderboard, offering some of the best retrieval quality available via API.
Type: Embeddings AI • Founded 2023 • San Francisco, USA
A database optimized for storing and searching embeddings (vectors). Instead of matching exact keywords like a traditional database, vector databases find the most semantically similar items. You ask "how to fix a memory leak" and it returns documents about "debugging RAM consumption" because the embeddings are close.
Why it matters: Vector databases are the storage layer that makes RAG work. Without them, you'd need to embed your entire knowledge base on every query. They're also the backbone of recommendation systems and semantic search.
The memory on a GPU, separate from system RAM. AI models must fit in VRAM to run on a GPU. A 7B parameter model in 16-bit precision needs ~14GB of VRAM. Consumer GPUs have 8-24GB; datacenter GPUs (A100, H100) have 40-80GB. VRAM is almost always the bottleneck for local AI.
Why it matters: VRAM determines which models you can run. It's why quantization exists (to shrink models to fit), why MoE models are tricky (all experts must fit in VRAM), and why GPU prices scale so steeply with memory. "Will it fit in VRAM?" is the first question of self-hosting AI.