Every letter in the soup

Creating three-dimensional shapes from text, a single image, or multiple views.

Comparing two versions by routing traffic to each and measuring which yields better outcomes.

A Hugging Face library that runs the same PyTorch training script across CPUs, GPUs, and Google's AI chips.

Specialized hardware built to run AI math far faster than a general-purpose CPU.

The share of a model's predictions that are correct against known answers.

A human approval step before an AI agent does something consequential.

The intermediate value a neuron or layer computes as data passes through a network.

Saving memory in training by recomputing the model's intermediate working values instead of storing them all.

The nonlinear step that lets stacked network layers learn complex patterns.

Quantization that protects the few weights most important to a model's outputs.

An adaptive optimizer that tunes a separate learning rate for each parameter.

A version of the Adam training optimizer that handles weight decay (a pull toward smaller numbers) separately.

Small trainable modules inserted into a frozen model so it can learn new tasks cheaply.

An input crafted to make a model produce a wrong or unintended output.

Input crafted to exploit a model's weaknesses and trigger unintended behavior.

An LLM that can plan, call tools, and act in a loop toward a goal.

Evaluations built to measure what AI agents can actually accomplish, not just model knowledge.

A structured profile describing what an AI agent can do and how to connect to it.

Techniques that let an agent retain information across turns or sessions.

A protocol for letting an AI agent make payments on a user's behalf, with limits and an audit trail.

Giving an AI agent only the permissions it needs for the task at hand.

A walled-off environment that limits what an AI agent can reach or change.

An open standard that lets separate AI agents advertise capabilities and work together.

AI systems that plan, call tools, observe results, and act over many steps toward a goal.

Online shopping where AI agents help search, compare, buy, track, or return products.

The repeating cycle in which an agent reasons, acts, observes, and continues.

Retrieval driven by an agent loop that decides when and what to search.

A multi-step process where one or more AI agents plan, use tools, and act to complete a goal.

A software toolkit for building and coordinating AI agents.

The worry that AI investment and hype have outpaced real value and may sharply correct.

The software that schedules and manages AI workloads across large GPU clusters.

An AI tool that helps developers write, explain, fix, and review code.

An AI assistant embedded in a product to help users get work done within it.

Using AI to defend computer systems — and the new risks AI brings to security.

The study of how to build and use AI in ways that are fair, safe, and respectful of people.

The rules, institutions, and processes that shape how AI is built and deployed.

Using generative models to alter existing images, often from text instructions.

A structured review of how an AI system may affect people, rights, and operations.

The practical ability to understand AI systems and use them responsibly.

An organization-wide system for governing AI policies, risks, roles, and controls.

The competition among companies and countries to build the most capable AI fastest.

The broad practice of reducing risks and harms from AI systems.

Low-quality, mass-produced AI-generated content that clutters feeds, search results, and inboxes.

An AI system that teaches and helps a learner one-on-one, adapting to their needs.

A past period when AI funding and interest collapsed after hype outran results.

A hard high-school math competition repurposed to test mathematical reasoning in models.

A finite, well-defined sequence of steps that solves a problem or performs a computation.

A Chinese tech company whose Tongyi Lab builds the freely downloadable Qwen model family.

Alibaba's cloud division that develops and serves the Qwen language and multimodal models.

Alibaba's model family and team behind the open-weight Qwen language and multimodal models.

A way to encode token position by penalizing attention to distant tokens instead of adding embeddings.

Making model behavior track intended human goals, values, and constraints.

A collective operation that combines values across many workers and shares the result with all.

A Seattle nonprofit AI lab focused on fully open models, data, and reproducible research.

A tech and e-commerce company whose AWS division is a major cloud platform for AI workloads.

Amazon's cloud platform offering GPU compute, custom AI chips, and managed model services.

Asking the model to recall similar solved examples on its own before solving the new problem.

Attaching labels, tags, or metadata to raw data so it can train or evaluate models.

Spotting data points that don't fit the normal pattern, such as fraud or equipment faults.

An AI safety company that builds the Claude models and pioneered Constitutional AI.