---
id: "concept-large-action-models"
type: "concept"
source_timestamps: ["§ The Everything Engine Needs Your Data"]
tags: ["ai-agents", "task-execution", "automation"]
related: ["concept-personal-large-action-models", "concept-advanced-sensors", "entity-anthropic-claude", "entity-adept-act-1", "quote-llm-vs-lam", "claim-sensor-ubiquity"]
definition: "AI models optimized for task execution rather than content generation, utilizing multimodal sensor and behavioral data to autonomously break down and complete complex tasks."
sources: ["futures"]
sourceVaultSlug: "hbr-seg-futures"
originDay: 2
articleStem: "hbr-foci-73-living-intelligence"
sourceUrl: "https://hbr.org/2025/01/why-living-intelligence-is-the-next-big-thing"
sourceTitle: "Why “Living Intelligence” Is the Next Big Thing"
---
# Large Action Models (LAMs)

**Large Action Models (LAMs)** represent the evolutionary next step beyond Large Language Models (LLMs). Where LLMs are designed to generate content and predict *what to say next*, LAMs are optimized for **task execution** — predicting what should be *done* next (see the quote [[quote-llm-vs-lam]]). They achieve this by breaking complex tasks into smaller, actionable pieces and making real-time decisions based on specific commands.

LAMs are data-hungry: they require vast amounts of **multimodal data**, heavily relying on behavioral data generated from phones, vehicles, and a constellation of environmental sensors (wearables, IoT, smart environments). This is the mechanism by which [[concept-advanced-sensors|advanced sensors]] become foundational — the exponential increase in the volume *and types* of sensor data is what makes LAMs possible (see [[claim-sensor-ubiquity]]).

Early examples cited: [[entity-anthropic-claude-d2|Anthropic's Claude]] and [[entity-adept-act-1|Adept.ai's ACT-1]], which interact directly with code and digital tools to perform actions inside software applications like web browsers. As they mature, LAMs will operate seamlessly in the background, often without direct user engagement. LAMs scale into three tiers: individual ([[concept-personal-large-action-models|PLAMs]]) and institutional ([[concept-corporate-large-action-models|CLAMs / GLAMs]]).

**Definition:** AI models optimized for task execution rather than content generation, utilizing multimodal sensor and behavioral data to autonomously break down and complete complex tasks.

> *Enrichment caveat:* The trend toward AI that *acts* rather than only generates text is well supported, but **"LAM" is not a canonical term** in mainstream AI research — the standard vocabulary is *agents*, *tool use*, *computer-use models*, and *workflow automation*. The examples are also imprecise: Claude is more accurately an LLM / tool-using assistant, whereas [[entity-adept-act-1|ACT-1]] is the cleaner example of an action-oriented system.


## Related across articles
- [[concept-agentic-ai-systems]]
- [[concept-service-as-software]]
- [[concept-recursive-algorithmic-development]]