Skip to content

76 terms, plain English

AI glossary

Plain-English definitions of the AI terms that come up across the Stanford CS 153 lectures we distilled, from scaling laws and MFU to network effects and enrichment. Each links back to where it's explained.

1

18 USC 1030
The U.S. computer-hacking statute; the Uber trial hinged on whether a company can grant access permission after the intrusion has occurred.
from Joe Sullivan's lecture →

A

Adversarial diffusion distillation
A technique that compresses a many-step diffusion model into a two-to-four-step one for fast, cheap generation.
from Andreas Blattmann's lecture →
Amdahl's law
1967 rule that a parallel system needs matching IO per unit of compute, or the compute cannot be fed and sits idle.
from Amin Vahdat's lecture →

B

Base load
Power that runs continuously and reliably, the always-on floor of supply that data centers need for uptime, as opposed to intermittent solar or wind.
from Scott Nolan's lecture →
Bootstrapping a network
The hard early phase of getting a network started when there are too few nodes to be useful, like selling the first telephone.
from Ben Horowitz's lecture →
Bug bounty
Paying outside researchers cash for the vulnerabilities they find; Sullivan launched one of the first at Facebook and Google now offers up to $250,000 per bug.
from Joe Sullivan's lecture →

C

Cascaded architecture
A voice system that chains three separate models: speech-to-text, an LLM, then text-to-speech.
from Mati Staniszewski's lecture →
Character consistency
Editing an image so a specific person or object stays recognizably the same across new scenes and prompts.
from Andreas Blattmann's lecture →
Closed loop
A company run like a control system where an agent reads every artifact and feeds errors back tightly, versus lossy open-loop information trapped in people's heads.
from Garry Tan & Diana Hu's lecture →
Co-design
Optimizing algorithms, compilers, chips, networking, and storage together instead of as separate specialized fields.
from Jensen Huang's lecture →
Context
The environment and feedback an agent learns from; whoever owns verifiable context can keep improving a model in that domain.
from Anjney Midha's lecture →
Context feedback loop
Piping real usage results (like whether a coding agent solved a task) back through RL to keep raising capability.
from Anjney Midha's lecture →

D

Dennard scaling
The physics that let transistors shrink at constant power density and underpinned Moore's Law; it ran out roughly a decade ago.
from Jensen Huang's lecture →
Differentiable
A function you can put in a training loop with a loss, so gradient descent can optimize it; if it isn't differentiable, deep learning can't learn it.
from Amit Jain's lecture →
Diffusion model
A generator that starts from pure noise and iteratively denoises it into an image, video, or audio sample.
from Andreas Blattmann's lecture →
DRI
Directly Responsible Individual, the single owner of an outcome who orchestrates ICs and agents to reach a goal.
from Garry Tan & Diana Hu's lecture →

E

Emergent property
A useful behavior that appears only at a larger scale and didn't exist at all at smaller sizes.
from Sam Altman's lecture →
Enrichment
Refining mined uranium to raise the concentration of the fissile isotope U-235 to the level a reactor needs; the middle of five fuel-making steps.
from Scott Nolan's lecture →

F

Five nines
99.999% availability, about 30 seconds of downtime a year, which requires costly 2N redundant power.
from Amin Vahdat's lecture →
Forward deployed engineer
A technical person who embeds with a customer to solve their problem, then pulls those learnings back into the core product.
from Nikhyl Singhal's lecture →
Fungible
Interchangeable unit for unit; megawatts are fungible, but GPUs today are not, which is why compute behaves like a scarce commodity.
from Anjney Midha's lecture →
Fused (omni) model
A single model trained to go straight from incoming audio to generated speech tokens without a text step, favoring speed over reliability.
from Mati Staniszewski's lecture →

G

Gigawatt (of infra)
The unit the industry uses to size an AI datacenter buildout; about $40 billion and 150,000-200,000 accelerators.
from Amin Vahdat's lecture →

H

HALEU
High-Assay Low-Enriched Uranium, a more enriched fuel grade that advanced small and micro reactors require.
from Scott Nolan's lecture →
Hill climbing machine
A setup where a company trains a licensed model on its own tasks, evals, and traces so its IP compounds instead of leaking.
from Satya Nadella's lecture →

I

Inference
Running a trained model to actually serve answers; the cost and speed of using AI rather than building it.
from Sam Altman's lecture →
Info-mover
Singhal's term for a PM whose job is packaging and moving information for someone else to decide, the role AI now replaces.
from Nikhyl Singhal's lecture →
Intelligence as a utility
The idea that AI becomes like electricity or internet: metered access that every person, company, and government plugs into.
from Sam Altman's lecture →

L

Latent space
A compressed representation of an image that stays perceptually equivalent to humans but is far smaller and cheaper to model.
from Andreas Blattmann's lecture →
LP (Limited Partner)
An investor, such as a university endowment, that puts money into a venture fund rather than running it.
from Ben Horowitz's lecture →

M

Megatons to Megawatts
A post-Cold War program that down-blended Russian warhead uranium into US reactor fuel, which helped make domestic enrichment uneconomical.
from Scott Nolan's lecture →
MFU (Model FLOPs Utilization)
The percentage of a chip's FLOPs actually used during work; Huang prefers it low, meaning over-provisioned rather than starved.
from Jensen Huang's lecture →
MFU / goodput
Model FLOPs Utilization and useful output; how much of the hardware you actually turn into real work rather than idle waste.
from Amin Vahdat's lecture →
Middle-to-middle
Using AI as an iterative step inside a creative workflow rather than an end-to-end prompt-to-output button, to avoid AI slop.
from Mati Staniszewski's lecture →
Moat
A durable barrier that stops competitors from copying a company; Horowitz argues code and UI no longer qualify in the AI era.
from Ben Horowitz's lecture →
Multimodal model
A single model trained jointly on several natural signals (image, video, audio) rather than one type of input.
from Andreas Blattmann's lecture →
MXC container
A Windows sandbox that isolates a running agent at the process, session, or VM level so its code execution can be governed.
from Satya Nadella's lecture →

N

NeRF / Gaussian splats
Techniques that reconstruct a 3D scene from 2D photos; Luma was first to productionize them in its capture app.
from Amit Jain's lecture →
Network effect
The property where each new user makes the whole network more valuable, with value scaling roughly like the square of the number of nodes.
from Ben Horowitz's lecture →

O

Operational resilience
A security goal focused on keeping a business running through an attack, not just preventing data from leaking out.
from Joe Sullivan's lecture →
Optical circuit switch
A chip of steerable mirrors that reroutes fiber in software, letting a spare rack instantly replace a failed one in the topology.
from Amin Vahdat's lecture →

P

PE mindset
Jain's term for private-equity thinking in Hollywood: rent-seek a proven franchise with endless sequels instead of trying many new ideas.
from Amit Jain's lecture →
PLG
Product-led growth, letting creators and developers adopt the product directly rather than through a sales motion.
from Mati Staniszewski's lecture →
Positive-sum ecosystem
A market where many companies can operate at the AI frontier and keep value, not one where a few firms capture all returns.
from Satya Nadella's lecture →
Post-training
Tuning a trained model, for example for instruction-following, so it does a good job in conversation.
from Sam Altman's lecture →
PRD
Product requirements document, the rigid spec a project manager once wrote and handed to engineers in old enterprise software.
from Nikhyl Singhal's lecture →
Prefill vs decode
Prefill processes the input context and attention; decode generates output tokens and is the memory-bandwidth-hungry phase.
from Jensen Huang's lecture →
Prepared mind
Nadella's phrase for being conditioned by years of prior work to recognize and act on a breakthrough when it arrives.
from Satya Nadella's lecture →
Product-market fit
When enough people naturally pull on your product that you can keep going; Singhal calls it a 'sucking sound'.
from Nikhyl Singhal's lecture →

R

Recursive self-improvement
When the compute and context flywheels get good enough to propel themselves; Midha frames it at the system level, not one super-model.
from Anjney Midha's lecture →
Reinforcement learning (RL)
Training by rewarding a system for completing a task rather than telling it how, now the compute-heaviest stage of model training.
from Anjney Midha's lecture →
Resolver
A master index that tells the agent which instruction file to load only when a task needs it, keeping the context window small.
from Garry Tan & Diana Hu's lecture →
Responsible disclosure
A policy promising researchers who report a vulnerability that the company will not sue or report them to law enforcement; Sullivan published the first one at PayPal in 2007.
from Joe Sullivan's lecture →
Runtime anomaly detection
Watching an AI agent's behavior live rather than pre-defining what it may do, because you cannot scope its access purpose by purpose.
from Joe Sullivan's lecture →

S

S-curve of growth
The company growth arc whose four stages (fit, scale, hypergrowth, late-stage) each demand a different kind of PM.
from Nikhyl Singhal's lecture →
SaaS apocalypse
The Wall Street narrative that AI will one-shot cheap software companies, which Horowitz treats as an overblown, tradable overreaction.
from Ben Horowitz's lecture →
SAFE
Simple Agreement for Future Equity, YC's two-page funding document that became the seed-stage standard, used as the analogy for what code and markdown will standardize next.
from Garry Tan & Diana Hu's lecture →
Scaling laws
The finding that model capability keeps improving predictably as you add more compute and data to the transformer.
from Satya Nadella's lecture →
Self-Flow
Black Forest Labs' published method for aligning a generative model's internal representations across multiple modalities so it understands, not just draws.
from Andreas Blattmann's lecture →
Skill
A markdown runbook of steps an agent follows to do a task, which can also invoke deterministic code.
from Garry Tan & Diana Hu's lecture →
Skillify
Promoting a proven one-off agent task into a reusable skill complete with unit tests, LLM evals, integration tests, and a trigger check.
from Garry Tan & Diana Hu's lecture →
Skills layer
Domain knowledge given to the model as context (like a 50-page slide-design doc), not baked into weights or tools.
from Amit Jain's lecture →
Sovereign AI
AI infrastructure a government or org runs on its own hardware so sensitive data never leaves its control.
from Anjney Midha's lecture →
Stranded energy
Power supply with no nearby demand, like remote hydro, geothermal, or West Texas wind, cheap to grab but now mostly claimed.
from Scott Nolan's lecture →
System balance
Provisioning flops, memory bandwidth (HBM), and network in the right ratio so no part starves the others.
from Amin Vahdat's lecture →

T

Terawatt
A unit of power equal to a trillion watts; a benchmark for how much new generation AI demand may need within a decade.
from Scott Nolan's lecture →
Text-to-speech (TTS)
Turning written text into spoken audio; the generation step ElevenLabs chose to specialize in.
from Mati Staniszewski's lecture →
The pipeline
The standard training sequence across labs: pre-training, mid-training, post-training, then reinforcement learning and supervised feedback.
from Sam Altman's lecture →
Tokens per watt
Output tokens generated per unit of energy, Huang's proposed real measure of intelligence delivered, driven more by bandwidth than FLOPs.
from Jensen Huang's lecture →
Tortoise TTS
An open-source speech model by James Betker that first sounded human on short fragments but was slow and unstable on long text.
from Mati Staniszewski's lecture →

U

Unified model
One transformer backbone that both understands and generates across text, image, video, and audio, instead of separate stitched-together models.
from Amit Jain's lecture →
Unmetered intelligence
Running AI models on local PC and edge GPUs so agents keep working without consuming scarce cloud tokens.
from Satya Nadella's lecture →

V

VLM
Vision-language model: it can understand images but cannot generate them, unlike a diffusion model which generates but doesn't understand.
from Amit Jain's lecture →
Voting vs weighing machine
Buffett's idea that markets price on narrative in the short run and on actual financial results in the long run.
from Ben Horowitz's lecture →

These come from our Frontier Systems study guides. For how the models themselves work, read the 5-part LLM primer.