Unique Research · Insight · September 5, 2024 · Compiled from Goldman / industry commentary on Large Action Models (LAMs)
Goldman notes GenAI has advanced chatbots and co-pilots but still lacks a mass-market “killer app.” Analysts argue the unlock is apps that act autonomously — and Large Action Models (LAMs) are a candidate architecture.
LLM vs LAM
Coined late 2023 and quickly hot. LLMs predict next tokens; LAMs extend that into agents that pursue goals — language understanding plus autonomous task execution and decision-making. 3Pillar CIO Pankaj Chawla: LLMs generate and answer but don’t take the next step; LAMs fill that gap by simulating apps/human behaviors and executing complex ops, further helped by neuro-symbolic programming that reduces need for short demos.
Contrast sketch: LLM = text/Q&A; LAM = plan + act toward a goal (book the flight, not only recommend it). Names vary — Gartner often says neurosymbolic AI.
What LAMs can do in practice
Tight integration with external systems (e.g. IoT): physical control, data fetch/manipulation, human dialogue, environment adaptation, multi-LAM collaboration. Core strength: turn complex human goals into executable steps in real time — via APIs or UI interaction (RPA-like). Example: find flights/hotels and complete the booking. Chawla: beyond co-pilots — unattended execution that gets more efficient with repetition. Amazon/AWS semi-autonomous agents already claimed large Java maintenance time savings.
Scenes: (1) advanced virtual assistants / CS that also act for the customer; (2) process automation (data entry, docs, inventory) via voice into similar forms — time/cost + accuracy as recognition improves; (3) retail assistants — history/preference-based recommendations, promos, recipes; sentiment-driven service fixes in real time.
Early product instances
Rabbit r1 — ~half iPhone size, touchscreen, 360° camera, scroll wheel, far-field mic, push-to-talk; Rabbit OS with private LAM that watches user actions across interfaces without classic app hopping (still largely pre-order era in the source window).
Salesforce — LAM-style automation in the enterprise suite using internal data for activity launch and output tracking.
McKinsey-style framing in the piece: LAMs + intelligent agents as GenAI’s frontier — “ultra-efficient virtual colleagues” for underwriting, code docs, campaign builds, etc.
Misconceptions (Dmytro Ivanov, Trinetix ML engineer)
Much talk is future impact, not proven results — scarce performance trackers, market monitors, or widely accepted case studies. Watch for:
- “Enterprises already widely use LAMs” — Rabbit popularized the term for a trainable assistant on custom hardware; pre-sales limit real field evidence.
- “LAMs already handle critical tasks” — theory ≠ practice; even simple-task competence is still research. AI needs hardware too — OpenAI-scale supercomputers (source cites 285k+ CPU cores) are not a near-term option for most firms; Amazon/Meta-class only for now.
Until first applications are validated, industry-transformation claims are premature. Focus on demonstrated capability over unverified claims.
Challenges and 2–5 year outlook
LLM probabilistic drift off task → need deterministic programming + classics (knowledge graphs) for stability — 3Pillar’s interactive LAM app example. Longer term: ERP suite integration unlocks cross-industry data; more configure-and-click LAM frameworks. Chawla: concept→real apps in 2–5 years, especially finance, healthcare, autonomous driving. McKinsey line in CN: LAMs as the next GenAI frontier toward autonomous decision and action.
Sources compiled in CN: Rosemary J Thomas (Version 1 AI Labs) on LAMs; “Road from Chatbots and Co-Pilots to LAMs and AI Agents”; Trinetix “What Are Large Action Models…”
— END —