A
Accelerationism e/acc · effective accelerationism · accelerationist
The view that AI progress should move as fast as possible, because its benefits vastly outweigh the risks and slowing down is itself dangerous.
Accelerationists argue that AI will drive growth, science and abundance, and that heavy regulation mainly protects incumbents or hands the lead to rivals like China. Critics see the stance as dismissive of real risks. Most people in the industry sit somewhere between this camp and the safety-focused one.
Agent Containment AI containment · sandbox boundary
The technical controls and safeguards that prevent an autonomous AI agent from executing unauthorized external actions or breaking out of its operating environment.
When AI agents are given access to tools like code interpreters, system shells, or web access, containment mechanisms prevent them from bypassing access controls, exploiting vulnerabilities, or communicating with unapproved third-party servers. A containment breach occurs when an agent takes unintended external actions to complete a goal.
AnalogyA secure visitor badge that allows a contractor into the lobby and cafeteria but will not unlock executive suites or private servers.
Agent loop agentic loop
The repeating cycle an agent runs: understand the goal, decide the next step, act with a tool, check the result, repeat.
Each pass through the loop is a fresh model call informed by everything that has happened so far. The loop is what lets agents adapt and recover from mistakes, and also why small errors can compound over many steps. Most agent products are this loop plus good tools, instructions and rules for when to stop or ask a human.
AnalogyIt's how you'd find a restaurant in a new city: check the map, walk a block, look around, adjust, repeat until you're at the door.
Agent Sandboxing Sandboxing · Environment Isolation
Isolating an autonomous AI agent within restricted software boundaries to prevent unauthorized actions and data leakage.
Agent sandboxing enforces strict architectural limits on what an AI agent can access when given external tools like web browsers, file systems, or APIs. Without rigorous sandboxing, autonomous agents can exceed their intended task parameters, access internal networks, or leak confidential data to external servers.
AnalogyIt's letting a new employee practice in a test copy of the system before giving them access to real customer accounts.
Agent-to-Agent (A2A) A2A · inter-agent communication · agent-to-agent collaboration
Direct, autonomous communication and negotiation between specialized AI agents across separate business systems.
Instead of relying on a single general model or requiring humans to transfer data across departments, agent-to-agent architecture allows specialized software agents to coordinate directly. Each agent maintains its own domain expertise, rules, and governance. For instance, a promotional marketing agent can query an internal compliance agent in real time to verify licensing or responsible gaming rules before executing an offer.
AnalogyA commercial sales rep calling the legal department directly to clear contract terms before sending an agreement to a client, rather than guessing at the rules alone.
Agentic AI agentic · agentic systems · agentic workflow
The broad shift toward AI that takes actions and completes multi-step tasks, not just answers questions.
'Agentic' describes a degree of autonomy more than a specific product: the more a system plans, acts and corrects itself without a human at each step, the more agentic it is. It's the dominant theme in AI product strategy as of 2026. Because the word is so popular, it's also a favorite of AI-washing, so ask what the system actually does unsupervised.
AGI (artificial general intelligence) artificial general intelligence
AI that can match or exceed human ability across most cognitive tasks, not just narrow ones.
There's no agreed definition, which is part of why timeline debates get heated. Some lab leaders have suggested it could arrive within a few years; many researchers are more skeptical, or argue the term is too vague to be useful. It also appears in company missions and contracts, including the partnership between Microsoft and OpenAI, so how it's defined can carry real commercial consequences.
AI agent agents · ai agents
An AI system that works toward a goal on its own: planning steps, using tools, checking results and continuing until the job is done.
A chatbot answers a question; an agent takes actions, such as searching, running code, updating a spreadsheet or drafting and sending a message, in a loop. Under the hood it's usually an LLM plus tools, instructions and permissions. Agents are genuinely useful for well-defined tasks today, but reliability drops on long, messy, open-ended work.
AnalogyA chatbot is a travel agent who tells you the best flight; an agent is one who also books it, reserves the hotel and adds it all to your calendar.
AI bubble bubble · ai bubble debate · bubble debate
The debate over whether AI investment and valuations have run far ahead of the revenue and profits the technology will actually produce.
Skeptics point to enormous capex, circular deals in which companies invest in their own customers or suppliers, and unproven business models. Optimists point to fast-growing revenue, real productivity gains and the fact that past bubbles, like the dot-com era, still left lasting infrastructure. Both can be true: a technology can be transformative and still overbuilt in the short run.
AI employee digital worker · digital workers · digital employee
An AI agent that performs a defined job role largely on its own, managed more like a team member than a tool.
It's a popular framing in vendor pitches and in debates about jobs. Today's reality is closer to a capable but inconsistent junior helper that needs clear tasks, tools and supervision. The useful question is which specific workflows it can own end to end, and who checks its work.
AI governance ai policy · ai policies · acceptable use policy
The internal rules, roles and processes a company uses to decide how AI is used, reviewed and monitored responsibly.
Good governance covers which tools are approved, what data can go into them, when human review is required, who owns the risk and how incidents are handled. It's increasingly expected by regulators, partners and customers, especially in regulated industries like gaming. Done well, it speeds adoption by making the rules clear rather than blocking everything.
AI lab ai labs · frontier lab · frontier labs
A company that builds its own large foundation models, such as OpenAI, Anthropic, Google DeepMind, Meta, xAI, Mistral or DeepSeek.
Labs sit in the middle of the AI stack: they buy huge amounts of compute and sell access to models through apps and APIs. Their costs are enormous and their competitive positions shift fast. 'Frontier lab' refers to the handful at the leading edge of capability.
AI safety safety · frontier safety · ai safety research
The field focused on preventing AI systems from causing harm, from everyday failures and misuse up to catastrophic risks.
It spans near-term work, such as blocking harmful outputs and misuse for cyberattacks or weapons, and long-term work on keeping far more capable systems under human control. Labs test models before release and publish safety frameworks, and several governments have set up institutes to evaluate frontier models. The debate is less about whether safety matters than about how much caution should slow progress.
AI stack the stack · stack · ai value chain
The layers of the AI industry, from energy and data centers to chips, cloud, models, tooling and applications, each with its own economics.
Thinking in layers tells you who pays whom: apps pay model labs, labs pay clouds, clouds pay chipmakers, and everyone pays for power. So far, profits have concentrated near the bottom, especially in chips, while the model and app layers spend heavily to grow. When you read an AI headline, first ask which layer it's about.
AnalogyIt's like the electricity business: power plants, the grid, the utility and the appliance maker each capture a different slice of what you pay.
AI winter ai winters
A period when AI hype collapsed, funding dried up and progress slowed; the best-known ones hit in the 1970s and late 1980s.
Each winter followed promises that ran well ahead of what the technology could deliver. Skeptics cite them as a warning about today's spending; optimists argue this cycle differs because the products already earn real revenue. Knowing the history helps you discuss the bubble debate calmly.
AI-native ai-native company · ai-native companies · ai-first
A company or product designed from the start around AI, rather than an existing business bolting AI features onto old workflows.
AI-native startups can run with very small teams because agents handle work that used to need many people. They often redesign the workflow itself rather than speeding up the old one. Incumbents' advantages, like data, distribution and trust, matter only if they adapt fast enough.
AI-washing ai-washed
Overstating how much AI a product really uses, or how capable it is, to ride the hype; the AI version of greenwashing.
Red flags include vague claims ('powered by AI') with no specifics, slick demos you can't test on your own data, and no evaluation results. US regulators, including the SEC, have taken enforcement action against companies over misleading AI claims. A few precise questions usually separate real AI from marketing.
Algorithmic Elasticity promotional elasticity · behavioral elasticity
Using machine-learning models to measure how responsive an individual consumer is to dynamic discounts, notifications, or incentives.
In economics, elasticity measures how customer demand shifts when prices or incentives change. Machine-learning models quantify this at an individual level to predict the exact nudge needed to trigger a transaction. In heavily regulated consumer industries, optimizing unconstrained models for elasticity can inadvertently target vulnerable users who lack impulse control.
AnalogyA retail manager who notices which shoppers pause near the exit and immediately offers them a targeted coupon to close the sale.
Alignment ai alignment · misalignment · misaligned
The challenge of making AI systems reliably pursue the goals and values their developers and users intend, and not something else.
Today, alignment work covers making models honest, helpful and resistant to misuse. Longer term, researchers worry about far more capable systems pursuing goals subtly different from ours in ways we can't easily detect or correct. Techniques like RLHF and constitutional AI are the current tools; whether they'll hold up for much more powerful systems is an open question.
AnalogyIt's like writing incentives into a contract: pay a salesperson only on volume and you may get volume at any cost.
Ambient AI Pervasive AI · Background Intelligence
AI systems embedded in hardware that continuously perceive their environment and assist users without requiring explicit prompts.
Instead of requiring a user to open an app and submit text prompts, ambient AI runs continuously in the background through sensors, cameras, and audio inputs in wearables like smart glasses. It interprets context from the physical environment to surface relevant information automatically.
AnalogyA veteran spotter in a race driver's headset who monitors blind spots and track hazards in real time without the driver having to ask.
AML (anti-money laundering) anti-money laundering · money laundering
The rules and systems that detect and prevent criminals from using betting or financial accounts to disguise illicit money.
Operators must monitor transactions for suspicious patterns, such as large deposits with little betting followed by withdrawals, and report them. Machine learning helps find complex patterns and cut false alarms, which are costly to investigate. Regulators hold operators accountable for gaps, and fines for AML failures can be large.
API apis · application programming interface
Application programming interface: a standard way for one piece of software to request data or services from another.
Most businesses use AI models through APIs: your app sends a prompt to the provider's API and gets a response back, paying per token. APIs are also how agents connect to tools and how data feeds reach sportsbooks. 'Do you have an API?' is often the first question in any integration conversation.
AnalogyIt's like an official data feed: partners don't need to know how the data is collected, only how to plug in and receive it.
Artificial intelligence (AI) artificial intelligence
The broad field of building computer systems that do things we'd normally say require human intelligence: recognizing, predicting, deciding and creating.
AI is an umbrella term, not a single technology. Older AI was mostly rules written by hand; nearly everything called AI today is machine learning, where the system learns patterns from examples instead. When someone says 'AI' in a meeting, it helps to ask which kind they mean: a prediction model, a chatbot or an agent.
AnalogyAI is like the word 'vehicle': one label covering everything from a scooter to a jet.
ASIC (custom chip) asics · custom silicon · custom chip
A chip designed for one specific job; in AI, custom chips built to run models more efficiently than general-purpose GPUs.
Hyperscalers, including Google (TPU) and Amazon (Trainium), along with a wave of startups, are building AI-specific chips. They can be cheaper and more power-efficient for particular workloads but are less flexible than GPUs. The more AI shifts toward inference at massive scale, the stronger the case for custom silicon.
Augmentation vs. automation augmentation · automation · job automation
The difference between AI helping people do their jobs better (augmentation) and AI doing the job instead of people (automation).
Most AI today augments: it speeds up drafting, research and analysis. Automation is growing in narrow, repetitive tasks, and agents push that line further. History suggests technology changes the mix of tasks within jobs before it eliminates whole occupations, but the speed of this wave makes the transition a genuine concern.
Automated highlights ai highlights · auto highlights · automated clipping
Using AI to detect key moments in live video and automatically cut, caption and publish highlight clips within seconds or minutes.
Computer vision and audio cues, like crowd noise and commentator excitement, identify big plays, and AI assembles clips tailored to different platforms or even individual fans. It lets rights holders cover far more games, leagues and angles than human editors could. The usual rights questions about who may clip and distribute what still apply.
Automation bias over-reliance · overreliance
The human tendency to trust a machine’s answer too much, even when it’s wrong or when your own judgment says otherwise.
When a system is usually right, people stop checking it, and that is exactly when its mistakes slip through. It’s why “a human reviews it” only works if the reviewer is trained, has time and is expected to push back.
AnalogyIt's following your GPS into a lake because it has been right every day for a year.
Autonomy levels of autonomy · autonomy spectrum · autonomous
How much an AI system decides and acts on its own versus waiting for human direction; the dial separating chatbots, copilots and agents.
At one end the AI only suggests; in the middle it acts but checks in; at the far end it runs a whole task unsupervised. More autonomy means more output per person and more risk. The right setting depends on how costly a mistake is and how easily it can be undone.
AnalogyIt's the difference between a new hire who checks in before every decision and a trusted veteran who's told to handle it and report back.
C
Capex capital expenditure · capital expenditures · capital spending
Capital expenditure: money spent on long-lived physical assets; in AI, mostly chips, data centers and power infrastructure.
Big tech's combined AI capex has grown to hundreds of billions of dollars a year as of 2026, one of the largest investment waves in corporate history. Bulls see it as building the infrastructure of a new era; bears worry revenue won't catch up fast enough to justify it. Capex guidance on earnings calls is one of the most-watched signals in AI investing.
Change management adoption · ai adoption · ai champion
The work of getting people to actually adopt a new tool or process: training, incentives, communication and redesigned workflows.
Buying AI licenses is easy; getting people to change how they work is hard. Successful rollouts pair tools with training, visible leadership use and updated processes, often led by 'AI champions', respected people in each team who show peers what works. Most failed AI deployments are adoption failures more than technology failures.
Chatbot chatbots · ai assistant · ai assistants
A conversational interface where you type or speak and the AI replies; the most familiar way people use LLMs.
Modern chatbots like ChatGPT, Claude and Gemini are LLMs wrapped in an app with extras such as memory, file uploads and web search. A chatbot answers; an agent acts. Many products marketed as agents are still chatbots underneath, so it's fair to ask what the system can actually do on its own.
Closed model closed models · proprietary model · proprietary models
A model whose weights are kept private by its maker, so it can only be used through that company's apps or API.
Many of the most capable models, including those from OpenAI, Anthropic and Google, are closed. Customers get the latest capabilities without managing infrastructure, but depend on the vendor's pricing, policies and uptime. Closed versus open is as much a business strategy as a technical choice.
Cloud computing cloud · the cloud · cloud provider
Renting computing power, storage and software over the internet instead of owning the hardware yourself.
Nearly all AI products run in the cloud, which lets companies scale up or down without building data centers. For AI buyers, it means your data usually travels to a provider's servers, so where it goes and how it's handled matters. Cloud marketplaces are also a major channel through which companies buy AI models.
Coding agent coding agents · ai coding assistant · ai coding tools
An AI agent that writes, edits, tests and fixes software, often working across a whole codebase with little step-by-step direction.
Coding went first among agent use cases because code can be checked automatically: tests pass or fail, giving the AI fast, clear feedback. Tools such as Claude Code, OpenAI's Codex and Cursor are widely used by developers as of 2026, and 'vibe coding' describes non-engineers building working software just by describing what they want. That's great for prototypes and internal tools; anything secure or customer-facing still needs real engineering review.
Commoditization commoditize · commoditized · model commoditization
When a product becomes so widely available and interchangeable that competition is mostly on price; a real risk for AI model makers.
As many labs release models of similar quality, and open-weight models close the gap, customers can switch more easily and prices fall. That pushes value toward whoever controls distribution, proprietary data or the customer relationship. It's a central question for investors: which layers commoditize and which keep pricing power.
Compounding errors compounding error · error compounding
How small mistakes pile up across an agent's many steps, so a system that's usually right per step can still fail the whole job.
If an agent gets each step right 95% of the time, a 20-step task succeeds only about 36% of the time. That simple math explains why agents shine on short tasks and struggle on long ones, and why checkpoints and human review matter. Small gains in per-step reliability pay off hugely on long tasks.
AnalogyIf each step in a 20-step process goes right 95% of the time, the whole thing succeeds only about a third of the time.
Compute computing power · compute capacity · ai compute
Industry shorthand for computing power: the chips, servers and energy used to train and run AI models.
Compute is the raw input of AI: more of it, spent well, has reliably produced better models. It's measured in chips, in FLOPs, or increasingly in gigawatts of data center power. Access to compute is a key competitive advantage for labs and a strategic priority for governments.
Computer use computer-use agent · computer-use agents · browser agent
An AI agent operating a computer the way a person does: looking at the screen, moving the cursor, clicking and typing.
Computer use lets agents work with software that has no API, like an old back-office system or an ordinary website. It's powerful but slower and more error-prone than direct integrations, and it raises security concerns because the agent can see and act on whatever is on screen. Think of it as the fallback when a clean connection doesn't exist.
Computer vision cv · machine vision · image recognition
AI that interprets images and video: recognizing players, the ball, logos, actions and events.
Computer vision turns raw broadcast or venue video into structured data, such as who is where and what just happened. It powers player tracking, automated highlights, officiating aids and sponsorship measurement. Its accuracy depends heavily on camera angles, lighting, players blocking each other and how much labeled footage exists for that sport.
Constitutional AI constitution · rlaif · reinforcement learning from ai feedback
An approach, developed by Anthropic, where a model is trained to critique and improve its answers against a written set of principles.
Instead of relying on human raters for every judgment, the lab writes a 'constitution' describing how the model should behave, and AI feedback guided by those principles does much of the training. That makes the rules more explicit and scales more cheaply than human labeling. Other labs use similar AI-feedback techniques under different names.
Context engineering context management
Deciding what information, tools and instructions go into a model's context window at each step, so it has exactly what it needs.
As agents run longer, the challenge shifts from writing one good prompt to managing a constantly changing context: which documents to fetch, what to summarize, what to drop. Too little context and the model guesses; too much and it gets distracted and expensive. Many practitioners now see this as the core skill in building agents.
Context window context windows · context length · context limit
The maximum amount of text, measured in tokens, a model can consider at once, including your prompt, any documents and its own reply.
Anything outside the context window is invisible to the model. Windows have grown from a few pages to hundreds of thousands of tokens or more, enough for long contracts or entire codebases. Bigger isn't free, though: long inputs cost more, run slower, and models can still miss details buried in the middle.
AnalogyIt's the model's desk: whatever fits on it, it can work with; anything that falls off the edge is forgotten.
Copilot copilots · ai copilot · ai copilots
An AI assistant built into software you already use, which helps you work but leaves you in control of the decisions.
A copilot drafts the email, suggests the formula or summarizes the meeting; you review and decide. It sits between a chatbot (answers questions) and an agent (acts on its own). Microsoft uses Copilot as a brand name, but the word is used broadly across the industry.
AnalogyLike a GPS that suggests every turn while you keep your hands on the wheel.
Copyright and AI intellectual property · ip · fair use
The legal fight over whether AI companies can train on copyrighted material without permission, and who owns AI-generated output.
Authors, news publishers, artists, music labels and others have sued AI labs, arguing that training on their work without a license is infringement; labs argue it's fair use or its equivalent. Early court rulings have been mixed, and some cases have ended in settlements or licensing deals. The outcomes will shape the value of content and data libraries, including sports footage and data.
CUDA cuda moat
Nvidia's software platform for programming its GPUs, widely seen as a core part of Nvidia's competitive moat.
Over nearly two decades, researchers and engineers built their tools on CUDA, so switching to a rival chip often means rewriting and re-tuning software. That lock-in helps Nvidia hold its lead even when competitors match its hardware on paper. Rivals and hyperscalers are working hard to reduce this dependence.
AnalogyIt's like the QWERTY keyboard: once everyone has learned it, a rival layout has to be far better, not just as good.
D
Data center data centers · datacenter · datacenters
A large building full of servers, networking and cooling; AI data centers pack in huge numbers of chips and draw enormous power.
Inside, thousands of GPUs are networked into clusters that act like one giant computer, which is how frontier models are trained. The biggest planned AI campuses are measured in gigawatts of power, and power supply, cooling, land and permits have become real bottlenecks. The data center buildout is the physical backbone of the AI capex boom.
Data leakage data leak · data leaks · leakage
Sensitive information escaping to places it shouldn't go, for example confidential documents pasted into a consumer AI tool.
Leakage can happen through employees using unapproved tools, vendors retaining data, agents with overly broad access, or prompt injection attacks that trick an AI into revealing information. Enterprise accounts, tight limits on what agents can reach and staff training are the basic defenses. When in doubt, don't paste anything you wouldn't send to an outside vendor.
Data licensing data license · data licenses · data licensing deal
Selling or granting rights to use data, increasingly including the right to use it to train AI models.
Publishers, platforms and sports data owners are signing deals with AI labs, while others are suing over unlicensed use. Contract terms matter: whether data can be used for training, whether trained models must be deleted if the deal ends, and who owns the outputs. Many older contracts never contemplated AI training at all.
Data privacy privacy · data protection · personal data
Protecting personal information from misuse, including how AI systems collect, store, learn from and reveal it.
AI raises new privacy questions: whether customer data trains a vendor's model, whether a model can leak personal details, and whether automated profiling is lawful. Europe's GDPR (General Data Protection Regulation) is the best-known rulebook, applying to any company handling EU residents' data, alongside a growing set of US state laws. Betting companies hold especially sensitive data (identity, finances, behavior), which raises the stakes.
Data readiness ai readiness · data quality · ai-ready data
Whether an organization's data is clean, organized, accessible and legally usable enough to actually power AI.
Many AI projects stall not because of the model but because the data is scattered, inconsistent, locked in old systems or restricted by contracts. Getting data ready is unglamorous and often the biggest hidden cost. Companies whose core product is already well-structured data, like live sports feeds, start with a real head start.
Data rights sports data rights · betting data rights
The commercial rights to collect, distribute and license the data a sporting event generates, usually controlled by leagues and their partners.
Leagues sell data rights much as they sell broadcast rights, often to exclusive partners. AI raises new questions for these deals: whether licensed data can be used to train models, who owns the models and insights built on it, and how to stop unlicensed scraping. Contracts written before generative AI may not answer these questions clearly.
Deep learning deep neural network · deep neural networks
Machine learning that uses large neural networks with many layers, able to learn very complex patterns from huge amounts of data.
'Deep' refers to the many stacked layers, each learning slightly more abstract features than the one before: edges, then shapes, then players. Deep learning took off around 2012, once GPUs and big datasets made it practical. It powers modern image recognition, speech recognition and every large language model.
Deep research deep research mode · research agent · research agents
A feature in major AI assistants where an agent searches the web and your files for several minutes, then writes a sourced report.
Instead of one quick answer, the agent runs dozens of searches, reads the sources and pulls them together. It's excellent for partner research, market scans and first drafts of briefing memos. The output still needs checking: sources can be weak, and the report can sound more certain than the evidence.
Deepfake deepfakes · synthetic media · voice cloning
Realistic fake video, audio or images of real people, generated by AI and used for entertainment, fraud or misinformation.
Deepfakes make convincing impersonation cheap: fake executive calls authorizing payments, fake athlete endorsements, bogus 'leaked' clips. Defenses include call-back checks before payments, watermarking and content-provenance standards (such as C2PA 'content credentials') that label AI-made or authentic media, and detection tools; none are foolproof. For sports and betting, fake injury news or fake official announcements are an integrity and reputational risk.
Diffusion model diffusion models · image generation · image generator
A type of generative model, common for images and video, that learns to turn random noise step by step into a picture matching a description.
Most AI image and video generators use diffusion or a close relative, trained on huge collections of images and clips paired with descriptions. They power creative tools and ad production at a fraction of traditional cost. They're also behind many deepfakes, which is why provenance labels and watermarking matter.
Distillation model distillation · knowledge distillation · distilled model
Training a smaller, cheaper model to imitate a larger, smarter one, keeping much of its ability at a fraction of the cost.
The big 'teacher' model generates answers, and the small 'student' learns from them. It's a major reason capable models keep getting cheaper and faster. It's also contentious: labs' terms of service typically forbid using their outputs to train competing models.
Doomer doomers · ai doomer · ai doomers
An informal, often dismissive label for people who believe advanced AI carries a serious risk of catastrophe for humanity.
Those labeled doomers typically call for slowing development, stronger regulation or international controls. Critics say they overstate speculative risks; they reply that the stakes justify caution. Neutral phrasing like 'safety-focused' tends to land better in mixed rooms.
Dynamic ad insertion dai · dynamic ads · addressable advertising
Automatically inserting different ads into a stream for different viewers, based on location, profile or context.
It lets a single broadcast carry many different ad breaks, each targeted, which raises the value of every impression. AI helps choose the right ad, find natural break points and match ads to what's happening on screen. It's a big part of how streaming sports media gets monetized.
Dynamic pricing dynamic ticket pricing · demand-based pricing
Changing ticket, concession or merchandise prices in real time based on demand, using data and models.
Teams and ticketing platforms adjust prices by opponent, weather, standings and sales pace, much as airlines do. AI improves demand forecasts and can price individual seats. The risk is fan backlash when prices feel exploitative, so many teams pair the models with clear limits.
E
Edge AI edge computing · edge · edge inference
Running AI close to where data is created, such as in a stadium, camera, car or phone, instead of in a distant data center.
Processing at the edge cuts latency, reduces bandwidth costs and can keep sensitive data local. It matters for live sports uses like in-venue tracking, where video is too heavy and time-sensitive to ship far away. The trade-off is that edge hardware is less powerful, so models must be smaller.
Embedding embeddings · vector embedding · vector embeddings
A list of numbers representing the meaning of a piece of text, an image or other data, so that similar things sit close together.
Embeddings let computers search by meaning rather than exact keywords: 'cheap flights to Miami' and 'low-cost airfare to Florida' land near each other. They're the engine behind RAG, recommendation systems and semantic search. You'll rarely touch them directly, but they're why AI search feels smarter than keyword search.
Energy demand power demand · data center power · ai energy use
The electricity needed to run AI data centers, now a major constraint on AI growth and a focus of utility, climate and policy debates.
A large AI campus can draw as much power as a small city, and new grid connections can take years. That's pushing tech companies into long-term deals for natural gas, nuclear and renewable power, and even into building their own generation. Energy is increasingly the bottleneck that decides where and how fast AI capacity gets built.
Enterprise plan enterprise account · enterprise accounts · enterprise tier
The business version of an AI tool, with contract terms, admin controls and data protections that consumer accounts don't provide.
Enterprise plans typically promise your data won't be used to train the vendor's models, add security controls like single sign-on, and set data retention rules. Consumer accounts may use your conversations for training by default unless you opt out. For work involving confidential information, the account type matters as much as the tool.
EU AI Act ai act · european ai act · eu ai regulation
The European Union's comprehensive AI law, which regulates AI by risk level and sets the strictest rules for high-risk uses.
It bans a short list of 'unacceptable' practices, puts heavy requirements on high-risk systems (such as those used in hiring, credit or critical infrastructure), sets transparency duties for things like chatbots and deepfakes, and adds obligations for general-purpose AI models. It entered into force in 2024 with obligations phasing in over several years, and some deadlines have been debated and revised, so check the current timeline. It applies to companies outside the EU if their AI is used in the EU.
Evals eval · evaluation · evaluations
Structured tests of how well an AI system performs on the tasks you actually care about, run before launch and on an ongoing basis.
Public benchmarks tell you how a model does on general tests; your own evals tell you how it does on your contracts, your customers and your edge cases. Serious AI teams maintain eval sets and re-run them whenever the model, prompt or data changes. A vendor who can't show evals is asking you to trust a demo.
AnalogyIt's like backtesting a pricing model on real historical markets before putting money behind it.
Event data play-by-play data · play-by-play · event feed
A timestamped record of what happened in a game, play by play: shots, passes, fouls, scores and substitutions.
Event data has traditionally been logged by trained human scouts, often at the venue, and it's the backbone of live betting and bet settlement. Computer vision is increasingly automating or double-checking that collection. Combined with tracking data, it tells models both what happened and how it happened.
Existential risk x-risk · catastrophic risk · extinction risk
The risk that advanced AI could cause human extinction or permanently and drastically curtail humanity's future.
It's taken seriously by some leading researchers and lab founders, and dismissed as speculative or distracting by others. 'P(doom)' is shorthand people use for their personal estimate of the probability of catastrophe. You can discuss it credibly by separating what has been demonstrated from what is argued, and by noting that serious people disagree.
Export controls export control · chip export controls · chip controls
Government restrictions on selling advanced technology abroad; in AI, chiefly US limits on exporting top chips and chipmaking tools to China.
Since 2022, the US has restricted sales of the most advanced AI chips and manufacturing equipment to China, aiming to slow its military and AI progress. The rules have been tightened, loosened and redrawn repeatedly, and they directly affect Nvidia's sales and chip design choices. China has responded by investing heavily in domestic chips and in highly efficient models.
F
Fine-tuning finetuning · fine-tune · fine-tuned
Extra training of an existing model on a smaller, specialized dataset so it gets better at a particular task, style or domain.
Fine-tuning changes the model's weights, so the new behavior is built in. It's good for consistent formats, a specialized tone or narrow repeated tasks, but a poor way to teach a model facts that change, since you'd have to retrain every time. For most companies, good prompting and RAG come first, and fine-tuning comes later, if at all.
AnalogyIf prompting is briefing a new hire and RAG is handing them the binder, fine-tuning is sending them through months of specialist training.
FLOPs flop · floating-point operations · flop/s
Floating-point operations: basic math calculations, used to measure how much work a chip can do or a training run required.
Training a frontier model takes an astronomically large number of FLOPs, which is why the figure appears in policy: the EU AI Act, for example, uses a training-compute threshold to flag general-purpose models that may pose 'systemic risk'. For business purposes, more FLOPs means more compute, more cost and usually a more capable model.
Foundation model foundation models · general-purpose ai · general-purpose ai model
A large model trained on broad data that can be adapted to many different tasks, rather than built for a single job.
The term captures the business logic: a few companies spend heavily to build one very capable model, and thousands of products get built on top of it. Most LLMs are foundation models. The EU AI Act uses the related term 'general-purpose AI model' for the same idea.
AnalogyIt's like a big convention center: one massive build that then hosts trade shows, concerts and weddings.
Foundry foundries · fab · fabs
A factory that manufactures chips designed by other companies; TSMC is by far the most important foundry for advanced AI chips.
Nvidia, Apple, AMD and Google design chips but don't make them; foundries do. Building a leading-edge fab costs tens of billions of dollars and takes years, which is why so few exist. The industry's heavy reliance on Taiwan is a major geopolitical risk factor.
Fraud detection fraud · bonus abuse · payment fraud
Using data and AI to catch fraudulent activity, such as stolen identities, bonus abuse, payment fraud and account takeovers.
Betting platforms are frequent fraud targets because accounts hold money and promotions can be gamed. Machine-learning models score logins, deposits and betting behavior in real time and flag what looks wrong. Generative AI raises the stakes on both sides: fraudsters use deepfakes and fake documents, and defenders use AI to spot them.
Frontier model frontier models · frontier ai · frontier ai model
One of the most capable AI models available at any given time, built by the handful of labs at the leading edge.
The term matters in policy because the newest, most powerful models are the ones regulators and safety researchers watch most closely. Rules like the EU AI Act and some US state laws set extra obligations for the largest models. What counts as 'frontier' keeps moving as capabilities improve.
Function calling function call · function calls
The technical mechanism behind tool use: the model outputs a structured request to run a specific function, and the software runs it.
Developers describe the functions available (for example, one that looks up a customer's order status), and the model decides when to call one and with what inputs. The app executes it and passes back the result for the model to use. You'll hear 'function calling' and 'tool calling' used almost interchangeably.
M
Machine learning ml · supervised learning · unsupervised learning
A way of building AI where the computer learns patterns from examples (data) instead of following rules a programmer wrote by hand.
You show the system many examples and it adjusts itself until its predictions are good; the result is a model you can use on cases it has never seen. When the examples come with the right answers attached (this transaction was fraud, this one wasn't), that's supervised learning; when the system finds structure on its own, like clusters of similar customers, that's unsupervised learning. Almost all modern AI, from fraud detection to ChatGPT, is machine learning at its core.
AnalogyThe old way is writing a rulebook for spotting spam; machine learning is showing the system a million emails marked spam or not and letting it work out the patterns.
MCP (Model Context Protocol) model context protocol
An open standard for connecting AI models to outside tools and data sources in a consistent, reusable way.
Before MCP, every connection between an AI app and a service (email, a CRM, a database) had to be custom-built. MCP provides a common plug shape, so one connector works across many AI apps. Anthropic introduced it in late 2024, and it has since been widely adopted across the industry, including by rival labs.
AnalogyIt's like a standard data feed format: build to it once and every partner can plug in.
Memorization regurgitation · memorisation
When a model reproduces chunks of its training data word for word instead of generating something new.
Models mostly learn patterns, but they can store and repeat passages they saw many times, such as famous articles or song lyrics. It matters in copyright lawsuits and privacy, because verbatim output is harder to defend than learned style.
Memory ai memory · persistent memory · chat memory
Features that let an AI assistant remember facts about you across conversations; the underlying model itself doesn't remember anything.
A model's weights don't change when you chat with it, so by default each conversation starts fresh. Apps add memory by saving notes about you and quietly inserting them into future prompts. It's convenient, but worth checking what's being stored, especially on work accounts.
Micro-betting microbetting · micro-markets · micro-bets
Betting on very short-term outcomes within a game, such as the result of the next pitch, play, point or drive.
Micro-betting creates many markets per game, each open for only seconds, so it's impossible without very low-latency data and fully automated pricing. It's popular with highly engaged fans and a growth area for operators. It also draws extra attention from regulators and responsible-gaming advocates because of how quickly bets can be placed.
Moat moats · competitive moat · defensibility
A durable competitive advantage that protects a business from rivals, borrowing Warren Buffett's image of a castle's moat.
In AI, candidate moats include compute, top talent, distribution, brand and trust, and especially proprietary data others can't get. A 'data flywheel', where more users generate more data that improves the product and attracts more users, is the classic AI moat argument, though many are weaker than pitched. Model quality alone has proven a weak moat, since rivals often catch up within months.
Model ai model · ai models · models
The trained artifact that does the work: a large file of learned numbers that takes an input and produces an output.
When people say 'the model', they mean what training produced, as distinct from the app or chatbot wrapped around it. The same model can power many different products. Most companies rent access to a model through an API rather than owning one.
Multi-agent system multi-agent · multi-agent systems · multiagent
A setup where several AI agents with different roles work together, for example one researching, one drafting and one checking.
Splitting work across specialized agents can improve quality and let tasks run in parallel. It also adds cost and new failure points, since agents can misunderstand each other or amplify each other's mistakes. It's promising, but simpler single-agent designs often work better for everyday tasks.
Multimodal multimodality · multimodal model · multimodal models
Able to handle more than one type of data, such as text, images, audio and video, as input, output or both.
Frontier models can now read a screenshot, listen to a question, describe a video clip or generate an image. For sports, that means one model could, in principle, take in broadcast video, commentary audio and stats together. Capabilities still vary widely by model and media type, so test on your actual use case.
O
Objective Function Loss Function · Reward Function · Target Metric
The mathematical formula that defines what a machine learning model is instructed to maximize or minimize.
An objective function defines success for an algorithm during training and deployment. If the objective function rewards only short-term commercial output without penalizing negative side effects, the model will optimize aggressively for that single number at the expense of safety, brand reputation, or compliance.
AnalogyCommissioning a sales rep solely on gross contract volume without deducting for customer churn or bad debt; you get massive signed deals that ruin company cash flow.
Odds compilation odds compiling · odds compiler · odds compilers
Setting the prices (odds) on betting markets by estimating the probability of each outcome and building in the bookmaker's margin.
Modern odds compilation relies heavily on statistical and machine-learning models fed by historical and live data, with human traders overseeing and adjusting. AI allows far more markets to be priced, faster and in finer detail, which is what makes huge player-prop and micro-betting menus possible. The quality of the data feeding the model sets the ceiling on pricing accuracy.
Official data official league data · official data feed · official data feeds
Game data supplied under a formal rights agreement with a league or governing body, typically collected in the venue and delivered fast.
Leagues license official data to sportsbooks and media, often through exclusive distribution partners, and some jurisdictions require or encourage its use for certain bets. Its value rests on speed, accuracy and legitimacy, which are exactly what live betting and AI models need. In an era where public information is easy to scrape, rights-backed, low-latency data is much harder to replicate.
On-device AI on-device · on-device model · on-device models
AI models that run directly on your phone, laptop or other device, without sending your data to the cloud.
Phone and PC makers are building dedicated AI chips into devices so assistants can work faster, offline and more privately. On-device models are smaller and less capable than frontier cloud models, so many products mix both. It's a key trend to watch for consumer apps, including betting and fan apps.
Open-weight model open-weight models · open weights · open-weight
A model whose trained weights are published so anyone can download, run and modify it, instead of only using it through the maker's service.
Open-weight models, from Meta, Mistral, DeepSeek, Alibaba's Qwen team and others, let companies run AI on their own servers, customize it and avoid per-token fees to a vendor. They're often called 'open source', although most don't release their training data and many carry license restrictions. As of 2026, the very best models are generally closed, with leading open models trailing by months rather than years.
Optical tracking player tracking · tracking data · optical player tracking
Using multiple cameras and computer vision to record the position of every player and the ball many times per second.
Optical tracking goes beyond the box score to speed, spacing, formations and distance covered; newer 'skeletal tracking' maps players' joints and limbs too, capturing things like a pitcher's arm angle or an attacker's foot on the offside line. The data feeds broadcast graphics, coaching, officiating aids and betting models. Genius Sports, for example, acquired the tracking company Second Spectrum in 2021, and optical tracking is now standard across many major leagues.
Orchestration agent orchestration · orchestrator · orchestration layer
Coordinating multiple AI models, agents and tools so they work together on a larger task, in the right order.
An orchestrator breaks a big job into pieces, assigns each to the right model or agent, and combines the results, handling retries, handoffs and when to ask a human. The software that does this, sometimes called the scaffolding or harness, is where much of an AI product's real engineering value lives. Two products using the same model can perform very differently because of it.
AnalogyIt's the general contractor on a renovation: the plumber, electrician and painter do the work, but someone has to schedule them in the right order.
Out-of-Band Containment hardware containment · out-of-band safety · hardware watchdog
A security architecture where an independent hardware layer monitors and restricts an AI agent from outside its operating environment.
Standard AI guardrails operate 'in-band,' relying on software prompts and internal rules that an autonomous model can misinterpret or bypass. Out-of-band containment places governance onto an isolated physical processor (such as a Data Processing Unit) that inspects data packets, API calls, and memory usage from the outside. If the software agent violates defined parameters, the hardware monitor instantly quarantines or terminates the process without needing the agent's permission.
AnalogyA circuit breaker in a basement that trips and cuts electrical power the instant a wire overheats, rather than relying on an appliance's power button.
Outcome-based pricing per-outcome pricing · pay per outcome · results-based pricing
Charging based on results the AI delivers, such as a resolved support ticket or a qualified lead, rather than on access or usage.
It ties the vendor's revenue directly to value created, which buyers like in principle. The hard part is agreeing on what counts as an outcome and how to measure it. Expect it most in narrow, measurable workflows like customer service rather than in broad tools.
AnalogyIt's like a recruiter who only gets paid when the candidate is hired, not for the number of résumés sent.
Overfitting overfit · overfitted · underfitting
When a model memorizes the quirks of its training data so closely that it performs poorly on new, real-world cases.
An overfit model looks brilliant when tested on familiar data and falls apart in production. It's why models must be tested on data they've never seen, and why vendor demos run on hand-picked examples deserve skepticism. The opposite problem, a model too simple to capture the real patterns, is called underfitting.
AnalogyIt's the student who memorized last year's exam answers and fails this year's test because the questions changed.
P
Parameters (weights) parameter · params · weights
The billions of adjustable numbers inside a model, tuned during training, that together hold everything it has learned.
Parameter count is a rough measure of model size: bigger models can capture more patterns but cost more to train and run. The file of trained parameters, 'the weights', is a lab's crown jewel. Publishing that file is what makes a model open-weight.
AnalogyParameters are the dials on an enormous mixing board; training is the process of setting every dial.
Personalization personalized feeds · recommendation engine · recommendation engines
Using data and AI to tailor content, offers, bets or experiences to each individual user.
In betting, that might mean surfacing markets on a fan's favorite team; in media, a highlight feed built around the players they follow. Done well, it lifts engagement and relevance. In betting it must be balanced against responsible-gaming rules, since pushing more bets to vulnerable customers is exactly what regulators want to prevent.
Physical AI robotics · embodied ai · humanoid robot
AI that senses and acts in the physical world through robots, vehicles or machines, rather than only on screens.
Physical AI combines computer vision, planning and motor control so machines can handle real objects and spaces. Self-driving cars and warehouse robots are the most mature examples; general-purpose humanoid robots attract heavy investment but remain early. It's harder than chatbots because real-world mistakes are costly and good training data is scarce.
Pilot pilots · proof of concept · poc
A small, time-limited trial of an AI tool or project to test whether it works before committing to a full rollout.
Good pilots target one painful, measurable workflow, set success criteria up front and have a clear decision date. Weak pilots are vague experiments with no owner and no metric. The goal is a clear yes or no, not a nice demo.
Pilot purgatory pilot hell · pilotitis
When a company runs AI pilot after pilot but few ever make it into real, scaled production use.
Common causes include no clear business owner, fuzzy success metrics, messy data, stalled security reviews and costs that rise at scale. Industry surveys have repeatedly found that a large share of enterprise AI pilots never deliver measurable returns. Breaking out usually means fewer pilots, each tied to a specific workflow and an executive sponsor.
Player props player prop · prop bet · prop bets
Bets on an individual player's performance, such as passing yards, points or strikeouts, rather than on the game's outcome.
Props have become one of the most popular bet types, and pricing thousands of them per game depends on player-level data and machine-learning models. They're also an integrity concern, because a single player can influence a prop far more easily than a final score. Some regulators and leagues have restricted certain prop types, particularly in college sports.
Post-training posttraining · instruction tuning · instruction-tuned
The phase after pre-training that turns a raw base model into a helpful, safe assistant, using curated examples and feedback.
It usually starts with instruction tuning (training on examples of requests paired with good answers, so the model learns to follow instructions), then feedback methods like RLHF and constitutional AI, plus reinforcement learning on tasks with checkable answers. It's where a model's personality, safety behavior and much of its reasoning skill come from. It uses far less compute than pre-training but has become a major battleground between labs.
AnalogyPre-training is a broad education; post-training is the on-the-job training that teaches someone how your company wants things done.
Pre-training pretraining · pretrained · base model
The first and largest phase of training, where a model learns language and general knowledge by predicting the next token across huge amounts of text.
Pre-training uses most of the compute and produces a 'base model' that knows a lot but isn't yet a helpful assistant; ask it a question and it may just continue with more questions. Because each next word serves as its own answer key, no human labeling is needed, which is why it scales to trillions of words. Post-training then shapes the base model into a product.
Prediction markets prediction market · event contracts · event contract
Exchanges where people trade contracts that pay out based on real-world outcomes, including elections, economic data and, increasingly, sports.
Prices act as crowd-sourced probabilities, which makes these markets useful signals as well as a form of wagering. In the US, some operate as federally regulated exchanges rather than state-licensed sportsbooks, and their move into sports has triggered legal fights with state regulators and tribal gaming interests that remain unsettled as of 2026. Automated trading and AI-driven market-making play a growing role on them.
Predictive AI predictive model · predictive models · predictive analytics
AI that classifies or forecasts, answering questions like 'which outcome is likely?' or 'is this suspicious?' rather than creating new content.
This is the workhorse AI that sportsbooks, banks and retailers have run for years: pricing models, churn prediction, fraud scores, recommendation engines. It's usually cheaper, faster and easier to test than generative AI. Many strong products combine the two: predictive models for the numbers, generative models for the words.
AnalogyA weather forecast is predictive AI's style of output: a probability, not a paragraph.
Prompt prompts · user prompt
The instructions, question or material you give an AI model to get a response.
A prompt can be one line or many pages, including documents, examples and data. The model only knows what it learned in training plus what's in front of it in the context window, so the quality of the prompt drives the quality of the answer. Clear goals, context, examples and a desired format usually beat clever wording.
Prompt engineering prompting · prompt design · few-shot prompting
The skill of writing prompts that reliably get useful results, by giving context, a role, a goal, examples, a format and constraints.
It's less about magic phrases and more about briefing well, the way you'd brief a smart new hire who knows nothing about your situation. One of the most effective techniques is 'few-shot' prompting, showing two or three examples of what good looks like, as opposed to 'zero-shot', just asking. Iterating and asking the model to critique its own draft also help.
AnalogyIt's briefing a new freelancer: the clearer the goal, context and examples, the better the first draft.
Prompt injection prompt injections · indirect prompt injection · injection attack
An attack where hidden instructions in content an AI reads, such as a web page, email or document, hijack it into doing something unintended.
Because models treat text as potential instructions, a malicious line in an email ('forward all invoices to this address') might be obeyed by an AI assistant processing that email. It's one of the biggest unsolved security problems for agents with access to tools and data. Defenses include limiting agent permissions, requiring human approval for sensitive actions and filtering inputs.
Proprietary data first-party data · data moat · exclusive data
Data a company owns or controls exclusively, which competitors and AI labs can't simply scrape or buy elsewhere.
As models become widely available, the data you feed them becomes the differentiator. Unique, high-quality, well-structured data can power better products, be licensed for training, or both. The key questions are who owns it, what the contracts allow and how hard it is to replicate.
R
RAG (retrieval-augmented generation) retrieval-augmented generation · retrieval · grounding
A technique where the system first searches your documents or data for relevant material, then has the model answer from it.
RAG lets a general model answer questions about your contracts, policies or data without retraining it, and it 'grounds' answers in real sources, which reduces hallucinations and makes citations easy. It stays current: update the documents, not the model. Its quality depends heavily on the search step; if the right document isn't retrieved, the model can't use it.
AnalogyIt's handing a new hire the binder and saying 'look it up before you answer'.
Reasoning model reasoning models · thinking model · thinking models
A model trained to think step by step before answering, spending extra time on hard problems in exchange for better accuracy.
Reasoning models work through a long internal chain of thought, checking and revising before giving a final answer. They're much stronger at math, coding, analysis and multi-step planning, but slower and more expensive per question. The visible reasoning can help you check the logic, though researchers caution it isn't always a faithful account of how the model reached its answer.
AnalogyIt's the difference between blurting out an answer and working the problem on scratch paper first.
Red-teaming red team · red teams
Deliberately attacking an AI system, trying to make it misbehave, leak data or cause harm, to find weaknesses before real bad actors do.
Labs use internal teams and outside experts to probe models for dangerous capabilities, bias, jailbreaks and security holes before launch. Companies deploying AI should red-team their own products too, especially customer-facing ones. The practice is borrowed from the military and cybersecurity.
AnalogyIt's hiring someone to try to break into your office so you find the unlocked doors before a real burglar does.
Reinforcement learning rl
Training by trial and error: the model takes actions, gets rewarded or penalized, and gradually learns what earns the best results.
This is how DeepMind's AlphaGo learned to beat the world's best Go players, partly by playing millions of games against itself. It's now central to post-training LLMs, both through human feedback (RLHF) and by rewarding correct answers on math and code, which is how reasoning models are built. The risk is that a model learns to game the reward rather than do the real task.
AnalogyIt's how a dog learns tricks: try something, get a treat or not, adjust, repeat thousands of times.
Responsible gaming responsible gambling · rg · safer gambling
Policies and tools meant to prevent and reduce gambling harm, such as deposit limits, self-exclusion and spotting at-risk behavior.
AI models can detect warning signs, like chasing losses, rising deposits or late-night play, earlier than manual review. Regulators increasingly expect operators to act on those signals, not just offer optional tools. The same personalization AI that drives engagement must be balanced against these duties, which makes responsible gaming a central design question for AI in betting.
Responsible scaling policy rsp · rsps · responsible scaling
A lab's public commitment to test new models for dangerous capabilities and add stronger safeguards as those capabilities grow.
Anthropic introduced the term in 2023; OpenAI, Google DeepMind and others publish similar frameworks under different names. The idea is a set of 'if-then' commitments: if a model reaches a certain danger level, specific protections must be in place before release. Critics note these are largely voluntary and self-policed, though some laws now require frontier developers to publish such frameworks.
Reward hacking specification gaming · reward gaming
When an AI finds a shortcut that scores well on the measure it's trained on without actually doing the intended task.
A classic example is a coding model that edits the tests so they pass rather than fixing the code. It's a practical problem in training today and a core reason alignment is hard: any metric you set can be gamed. The business lesson applies to AI incentives just as it does to people.
AnalogyIt's the call-center rep who hits the 'short call time' target by hanging up on customers.
RLHF reinforcement learning from human feedback
Reinforcement learning from human feedback: people rate a model's answers, and the model is trained to give the kinds of answers people prefer.
People compare pairs of responses and pick the better one; those judgments train a 'reward model' that then scores the LLM's outputs at scale. RLHF was a big reason ChatGPT felt so much more useful than earlier models. Its weakness is that models can learn to please the raters, for example by flattering them, rather than to be right.
ROI return on investment · ai roi
Return on investment: the value an AI project produces relative to its cost, whether in time saved, revenue gained or risk reduced.
AI ROI is often hard to measure because benefits are spread thin, like minutes saved per person per day. The most convincing cases start with a painful, measurable workflow and a before-and-after baseline. Count the full cost, including usage fees, integration, review time and training.
S
SaaS software as a service · per-seat pricing · seat-based pricing
Software as a service: software you rent by subscription over the internet, like Salesforce or Slack, rather than install and own.
SaaS has dominated business software for two decades, typically priced per seat (per user per month). AI challenges that model: if agents do the work of several people, customers need fewer seats, and AI-native rivals can build similar products faster. Many vendors are shifting toward usage- or outcome-based pricing in response.
Same-game parlay sgp · sgps · same-game parlays
A single bet combining several outcomes from the same game, such as a team win, a player's touchdown and the total points.
Pricing SGPs is hard because outcomes in the same game are correlated: if the quarterback throws for 350 yards, his receivers probably had big days too. Models must estimate those relationships accurately or the book misprices the bet. They're high-margin products for operators, made possible at scale by sophisticated modeling.
Scaling laws scaling law · scaling · scaling hypothesis
The observed pattern that models get predictably better as you add more compute, data and parameters.
Scaling laws are why labs spend billions on bigger training runs: the returns have been surprisingly steady. The live debate is whether pre-training scale is hitting diminishing returns, and how much future progress will come from other levers such as reasoning and test-time compute. Much of the industry's capex bet rests on scaling continuing to pay off.
Semiconductor semiconductors · chip · chips
The material, and the industry, behind computer chips; 'semiconductors' is shorthand for the whole chip business.
The AI supply chain runs through a few chokepoints: chip designers like Nvidia, the Dutch firm ASML that makes the most advanced chipmaking machines, foundries led by TSMC in Taiwan, and makers of the high-bandwidth memory (HBM) that AI chips depend on. That concentration makes chips a matter of national strategy and a focus of export controls.
Shadow AI shadow it · bring your own ai
Employees using AI tools their company hasn't approved, often pasting work data into personal accounts.
It's widespread because the tools are useful and easy to reach. The risks are confidential data leaking to vendors, compliance breaches and no record of how decisions were made. Banning AI rarely works; providing approved tools and clear rules usually does.
Small language model (SLM) slm · slms · small language models
A compact language model built to run cheaply and quickly, sometimes on a phone or laptop, trading some capability for speed and cost.
Not every task needs a frontier model; sorting support tickets or tagging clips can run well on a small one. Small models cut costs, reduce latency and can keep data on the device. A common design is routing: easy requests go to a small model, hard ones to a big model.
AnalogyIt's using a compact car for errands and saving the moving truck for moving day.
Sovereign AI ai sovereignty · national ai strategy
A country's push to control its own AI capability, including data centers, chips, models and data, rather than depend on foreign companies.
Governments from the Gulf states to Europe, India and Japan are funding national compute and local models, treating AI like energy or defense. It's a large and growing customer segment for chipmakers and cloud providers. It also means AI rules and data-location requirements will keep diverging by country.
Sports analytics advanced metrics · expected goals · xg
Using data and models to evaluate players and teams, inform strategy and predict performance, from scouting to game planning.
Metrics like expected goals (xG) estimate the quality of chances rather than just the outcome, giving a truer read on performance. Tracking data and AI have pushed analysis into movement, fatigue and injury prediction, where models flag players at elevated risk based on workload and changes in how they move; useful, but still probabilistic. The same models inform front offices, broadcasters and bettors.
Superintelligence (ASI) asi · artificial superintelligence · superintelligent ai
AI that vastly exceeds the best human minds in nearly every field, including science, strategy and persuasion.
Superintelligence is the hypothetical step beyond AGI, and the scenario behind both the biggest hopes (cures, abundance) and the biggest fears (loss of human control). Some argue that AI improving AI could get there quickly once AGI arrives; others see that as speculative. Several labs now name superintelligence as an explicit goal, which is why the topic has moved from science fiction into boardrooms and policy.
Sycophancy sycophantic
An AI model's tendency to tell users what they want to hear: agreeing, flattering or backing down when it shouldn't.
It's partly a side effect of training on human ratings, since people tend to rate agreeable answers higher. It's a real risk when you use AI to pressure-test your thinking: the model may validate a weak deal thesis. Ask it explicitly to argue the other side or find the flaws.
System prompt system prompts · system message · system instructions
Standing instructions, usually hidden from the user, that shape how a model behaves in every conversation within an app.
Before you type anything, most AI products have already given the model a system prompt describing its role, tone, rules and tools; it's how one model becomes a customer-service bot for one company and a coding assistant for another. The 'custom instructions' and project settings in consumer assistants work similarly, letting you save your role and preferences so every answer starts from them. Attackers sometimes try to trick models into revealing or ignoring their system prompts.
T
Temperature sampling temperature
A setting that controls how predictable or creative a model’s word choices are.
At low temperature the model almost always picks the most likely next token, which suits factual or repeatable work. At higher temperature it takes more chances, which helps brainstorming but raises the odds of errors. Many newer models manage this internally.
AnalogyLow temperature is ordering your usual; high temperature is letting the chef surprise you.
Test-time compute inference-time compute · inference scaling · test-time scaling
Computing power spent while a model is answering, letting it 'think longer' on hard questions to get better results.
For years, progress came mainly from spending more compute on training. Reasoning models opened a second lever: spend more at the moment of the question. That shifts cost from one-time training toward every single use, which raises demand for inference chips and changes AI economics.
Throughput tokens per second · output speed
How much work an AI system handles in a given time, such as tokens generated per second or requests served per minute.
Latency is how fast one customer is served; throughput is how many customers you can serve at once. Providers balance the two to control cost. For high-volume uses, like summarizing thousands of documents or screening every transaction, throughput and cost per unit matter more than speed on any single request.
AnalogyLatency is how long one customer waits for coffee; throughput is how many coffees the café serves in the morning rush.
Token tokens · tokenization · tokenizer
The small chunk of text a model reads and writes: often a whole word, sometimes part of one.
Models don't see letters or whole words; they see tokens, each mapped to a number. As a rough rule, a token is about three-quarters of an English word, so a 1,000-word memo is roughly 1,300 tokens. Almost everything is measured in tokens: how much fits in the context window, how fast a model responds and what you pay.
AnalogyTokens are the chips at the table: the unit everything gets counted and priced in.
Token pricing per-token pricing · price per token · cost per token
Charging for AI model use by the number of tokens processed, usually quoted as a price per million input and output tokens.
Input tokens (what you send) usually cost less than output tokens (what the model writes), and bigger or reasoning models cost more. It's a form of usage-based pricing: you pay for what you consume rather than a flat fee. The price for a given level of capability has fallen dramatically year after year, yet total AI bills often rise because cheaper usage leads to far more usage.
Tool use tool calling · tool call · tool calls
An AI model's ability to use outside tools, like search, calculators, databases or other software, to get information or take action.
On its own, a model can only produce text. Tool use lets it say, in effect, 'look up this customer's account' or 'run this calculation', have the software carry it out, and then use the result. It's the bridge between a model that talks and an agent that does.
TPU tpus · tensor processing unit · tensor processing units
Tensor processing unit: Google's custom AI chip, used to train and run its own models and rented to customers through Google Cloud.
TPUs are an example of a big tech company designing its own silicon to reduce dependence on Nvidia and cut costs. Google's Gemini models are trained on them, and other AI labs have also signed deals to use them. They show that GPUs are not the only path to frontier AI.
Trading and risk management trading · live trading · trading desk
The sportsbook function that manages prices and exposure: moving odds, setting limits and balancing liability as bets come in.
Traders watch the flow of money, news and game state, and move lines or suspend markets to manage risk. AI and automation now handle much of the routine repricing and flag unusual activity, freeing human traders for judgment calls and big events. It's a natural home for agent-style monitoring tools, with humans kept in the loop on major decisions.
Training model training · training run · training runs
Building a model by feeding it vast amounts of data and adjusting its parameters until its outputs are good.
Training a frontier model is enormously expensive, requiring huge clusters of GPUs running for weeks or months. It usually comes in stages: pre-training for broad knowledge, then post-training for behavior. Once training ends the model is frozen, and using it is called inference.
AnalogyTraining is going to medical school; inference is every patient appointment afterward.
Training data dataset · datasets · data set
The examples a model learns from, such as text, images, game logs or transactions; the biggest single influence on what a model can do.
A model can only learn patterns that appear in its training data, including the gaps and biases. Frontier LLMs train on a large share of the public internet, books, code, licensed content and, increasingly, synthetic data generated by other AI models. Who owns, licenses and may legally use training data is now a major commercial and legal fight.
AnalogyTraining data is everything a model has ever read: someone who only read one newspaper will have a narrow view of the world.
Transformer transformers · transformer architecture · attention
The neural network design, introduced in Google's 2017 paper 'Attention Is All You Need', that underpins nearly every modern LLM.
Its key idea, called attention, lets the model weigh how every word in a passage relates to every other word at once; in 'the trophy didn't fit in the suitcase because it was too big', attention helps it work out that 'it' is the trophy. That design trains efficiently on GPUs and handles long context well, which is why it took over. The 'T' in GPT (generative pre-trained transformer, the name of OpenAI's model family) stands for Transformer.
Turing test imitation game
A 1950 thought experiment by Alan Turing: if you can't tell a machine's conversation from a human's, should we call it intelligent?
For decades it was the pop-culture finish line for AI. Modern chatbots can pass casual versions of it, which mostly showed that fluent conversation is not the same as reliable intelligence. Researchers now rely on harder, task-specific benchmarks instead.
No terms match. Try a different word.