Trendingby GlyphMovements: technology waves, most active on Trending in the last 14 days first
Prompt engineeringWriting and testing the instructions given to a model to get reliable answers.Mixture of expertsModels split into many expert blocks with only a few active per token, so they grow in size without growing in cost.Context engineeringDeciding what goes into a model's context window, and in what order: instructions, retrieved documents, tools and memory.Computer useAgents that operate a computer the way a person does, reading the screen and moving the mouse and keyboard.Vector databasesDatabases that store embeddings and find the nearest ones fast, the retrieval layer under RAG and semantic search.Agentic codingCoding agents that read a repository, run commands and tests, and change code across files on their own.AI browsersWeb browsers with an assistant or agent built in that reads pages and does tasks on them for you.LLMOps and evalsMeasuring, tracing and monitoring model-backed software: eval suites, LLM observability and prompt management.Vibe codingBuilding software by describing it to a model and accepting what it writes, without reading the code closely.Multimodal AIModels that read and write more than text: images, audio and video in the same model.Inference chipsAccelerators built to serve models cheaply and fast rather than to train them: TPUs, LPUs, wafer-scale and custom silicon.Small language modelsCompact models, a few billion parameters or less, trained to be good enough and cheap to run anywhere.Open-weight modelsModels whose trained weights anyone can download, run and fine-tune: Llama, Qwen, DeepSeek, Mistral and more.Physical AIModels that perceive and act in the physical world: robots, self-driving and the simulators that train them.RAGRetrieval-augmented generation: a model answers from documents fetched for the question instead of from memory alone.AI agentsModels that plan and act in steps, calling tools and other services until a task is done.Humanoid robotsGeneral-purpose robots shaped like people, now driven by learned models instead of hand-written control.AI searchSearch that answers with a written summary and citations instead of ten blue links, and the deep research agents that followed.World modelsModels that learn how an environment behaves and can simulate it, for games, video and training robots.Local AIRunning language and image models on your own laptop, phone or server instead of a cloud API.MCPModel Context Protocol: an open standard for connecting models and agents to tools and data sources.Reasoning modelsModels trained with reinforcement learning to think step by step, spending more compute at answer time.Synthetic dataTraining and evaluation data generated by models or simulators instead of collected from people.
| Movement | ||||
|---|---|---|---|---|
| Prompt engineeringDecliningWriting and testing the instructions given to a model to get reliable answers. | Declining | -12% | +117% | |
| Mixture of expertsPeakModels split into many expert blocks with only a few active per token, so they grow in size without growing in cost. | Peak | +10% | +90% | |
| Context engineeringPlateauDeciding what goes into a model's context window, and in what order: instructions, retrieved documents, tools and memory. | Plateau | +34% | +81% | |
| Computer useRisingAgents that operate a computer the way a person does, reading the screen and moving the mouse and keyboard. | Rising | -16% | +77% | |
| Vector databasesPlateauDatabases that store embeddings and find the nearest ones fast, the retrieval layer under RAG and semantic search. | Plateau | -7% | +57% | |
| Agentic codingRisingCoding agents that read a repository, run commands and tests, and change code across files on their own. | Rising | +35% | +50% | |
| AI browsersPlateauWeb browsers with an assistant or agent built in that reads pages and does tasks on them for you. | Plateau | +48% | +36% | |
| LLMOps and evalsPeakMeasuring, tracing and monitoring model-backed software: eval suites, LLM observability and prompt management. | Peak | -12% | +26% | |
| Vibe codingPlateauBuilding software by describing it to a model and accepting what it writes, without reading the code closely. | Plateau | -11% | +7% | |
| Multimodal AIPeakModels that read and write more than text: images, audio and video in the same model. | Peak | -1% | -6% | |
| Inference chipsPeakAccelerators built to serve models cheaply and fast rather than to train them: TPUs, LPUs, wafer-scale and custom silicon. | Peak | +3% | -10% | |
| Small language modelsRisingCompact models, a few billion parameters or less, trained to be good enough and cheap to run anywhere. | Rising | -2% | -13% | |
| Open-weight modelsRisingModels whose trained weights anyone can download, run and fine-tune: Llama, Qwen, DeepSeek, Mistral and more. | Rising | +11% | -15% | |
| Physical AIRisingModels that perceive and act in the physical world: robots, self-driving and the simulators that train them. | Rising | +4% | -16% | |
| RAGPlateauRetrieval-augmented generation: a model answers from documents fetched for the question instead of from memory alone. | Plateau | -4% | -18% | |
| AI agentsRisingModels that plan and act in steps, calling tools and other services until a task is done. | Rising | +2% | -18% | |
| Humanoid robotsPeakGeneral-purpose robots shaped like people, now driven by learned models instead of hand-written control. | Peak | -16% | -19% | |
| AI searchPlateauSearch that answers with a written summary and citations instead of ten blue links, and the deep research agents that followed. | Plateau | +20% | -21% | |
| World modelsPeakModels that learn how an environment behaves and can simulate it, for games, video and training robots. | Peak | +35% | -21% | |
| Local AIPeakRunning language and image models on your own laptop, phone or server instead of a cloud API. | Peak | +8% | -22% | |
| MCPPeakModel Context Protocol: an open standard for connecting models and agents to tools and data sources. | Peak | +2% | -31% | |
| Reasoning modelsPeakModels trained with reinforcement learning to think step by step, spending more compute at answer time. | Peak | -11% | -44% | |
| Synthetic dataPlateauTraining and evaluation data generated by models or simulators instead of collected from people. | Plateau | -42% | -61% |
Sources: Hacker News story titles (Algolia), Wikipedia pageviews, arXiv papers (through DataCite) and Trending's own daily items. The stage rule is on the method page.