{"id":83536,"date":"2024-01-25T07:03:16","date_gmt":"2024-01-25T06:03:16","guid":{"rendered":"https:\/\/blog.axway.com\/?p=83536"},"modified":"2024-02-06T13:48:28","modified_gmt":"2024-02-06T20:48:28","slug":"lam-vs-llm-vague-apisation","status":"publish","type":"post","link":"https:\/\/blog.axway.com\/fr\/learning-center-fr\/api\/tendances-api\/lam-vs-llm-vague-apisation","title":{"rendered":"LAM vs LLM : La vague d\u2019APIsation va-t-elle se briser sur la digue LAM?"},"content":{"rendered":"
<\/div>

Le CES vient de fermer ses portes et l\u2019on peut sans aucun doute d\u00e9cerner la palme de l\u2019innovation \u00e0 Rabbit R1<\/a>. Ce petit lapin malin, qui se loge dans la poche, nous a \u00e9t\u00e9 pr\u00e9sent\u00e9 dans une keynote digne de la pr\u00e9sentation de l\u2019iPhone et vient confirmer nos pr\u00e9dictions 2023<\/a> sur la multiplication des assistants dop\u00e9e \u00e0 l\u2019IA g\u00e9n\u00e9rative (ChatGPT, Bard, etc).<\/p>\n

Le Rabbit R1 est un assistant virtuel qui apprend \u00e0 g\u00e9rer vos t\u00e2ches num\u00e9riques. Il suffit d’appuyer sur un bouton, de donner une commande vocale et de laisser les scripts automatis\u00e9s (appel\u00e9s “rabbits”) faire le travail \u00e0 votre place.<\/p>\n

 <\/p>\n

\"Rabbit
Image: Rabbit AI<\/a><\/figcaption><\/figure>\n

 <\/p>\n

Mais ce qui, \u00e0 mon sens, provoque un \u00e9tonnement sp\u00e9cifique, ce n\u2019est pas que ce petit animal se nourrisse\u00a0d\u2019IA g\u00e9n\u00e9rative bas\u00e9e sur des LLM. C\u2019\u00e9tait attendu, \u00e7a ne fait qu\u2019arriver un peu plus t\u00f4t que pr\u00e9vu. Non, ce qui nous interpelle se passe au niveau de l\u2019interconnexion de ces LLM aux services sous-jacents.<\/p>\n

<\/span>La r\u00e9volution au c\u0153ur de Rabbit R1<\/span><\/h2>\n

En effet, on imaginait tous que,\u00a0de mani\u00e8re tr\u00e8s classique, ces LLM se connecteraient aux grands services d\u2019internet par le biais d\u2019APIs\u00a0m\u00e9tier orient\u00e9es\u00a0IA. Cela reste le mod\u00e8le cible d\u2019un \u00e9cosyst\u00e8me o\u00f9 l\u2019IA g\u00e9n\u00e9rative se positionne au centre du dispositif, consommant et composant\u00a0les diff\u00e9rents services (calcul d\u2019itin\u00e9raire, r\u00e9servation de train, acc\u00e8s aux comptes, etc.).<\/p>\n

Mais le probl\u00e8me est que ce mod\u00e8le cible p\u00eache par le retard de la mise \u00e0 disposition de ces interfaces API et de leur exposition.<\/p>\n

Pour pallier ce retard et pour autant proposer un service de qualit\u00e9, la startup de Santa Monica, m\u00e8re de RabbitR1, et\u00a0qui a d\u00e9j\u00e0 pr\u00e9-vendu plus de 10 000 Rabbit R1 durant le CES (alors que les pr\u00e9visions livr\u00e9es aux analystes se basaient sur 500)\u00a0va donc compl\u00e9ter ce LLM d\u2019un LAM, un \u00ab\u00a0Large Action Model\u00a0\u00bb.<\/p>\n

Ce LAM est le bras arm\u00e9 de l\u2019intelligence du Rabbit.<\/p>\n

 <\/p>\n

\"Rabbit<\/a>
Image: Rabbit AI introduction video<\/a><\/figcaption><\/figure>\n

 <\/p>\n

<\/span>LAM vs LLM \u2013 comprendre la diff\u00e9rence<\/span><\/h2>\n

Contrairement aux LLM qui cr\u00e9ent de l\u2019intelligence \u00e0 travers des stocks\u00a0de donn\u00e9es et des API d\u2019acc\u00e8s aux services, les LAM sont, elles,\u00a0capables de reproduire l\u2019activit\u00e9 d\u2019un humain sur une interface classique (interface WEB, interface d\u2019application).<\/p>\n

Ce tableau, inspir\u00e9 de celui partag\u00e9 par SuperAGI sur X<\/a>, illustre plus en d\u00e9tail les distinctions entre LAM et LLM.<\/p>\n

 <\/p>\n\n\n\n\n\n\n\n\n\n\n\n\n
Aspects<\/strong><\/td>\nLLMS<\/strong><\/td>\nLAMS<\/strong><\/td>\n<\/tr>\n
<\/td>\nLarge Language Models<\/td>\nLarge Agentic(or Action) Models<\/td>\n<\/tr>\n
Core Function<\/strong><\/td>\nLanguage understanding and generation<\/strong><\/td>\nLanguage understanding, generation, complex reasoning and actions<\/strong><\/td>\n<\/tr>\n
Primary Strength<\/td>\nFormal linguistic capabilities, generating coherent and contextually relevant text<\/td>\nAdvanced linguistic capabilities \/ (Formal + Functional) combined with multi-hop thinking and generating actionable outputs<\/td>\n<\/tr>\n
Reasoning Ability<\/strong><\/td>\nLimited to single-step reasoning based on language patterns<\/strong><\/td>\nAdvanced multi-step reasoning, capable of handling complex, interconnected tasks & goals<\/strong><\/td>\n<\/tr>\n
Contextual Understanding<\/td>\nGood at understanding context within text, but limited in applying external knowledge<\/td>\nSuperior in understanding and applying both textual and external context<\/td>\n<\/tr>\n
Problem-Solving<\/strong><\/td>\nCan provide information and answer questions based on existing data<\/strong><\/td>\nCan propose solutions, strategic planning, make reasoned decisions and act autonomously<\/strong><\/td>\n<\/tr>\n
Learning Approach<\/td>\nPrimarily based on pattern recognition from large datasets<\/td>\nIntegrates pattern recognition, self-assessment & learning with advanced learning algorithms for reasoning and decision-making<\/td>\n<\/tr>\n
Application Scope<\/strong><\/td>\nSuitable for tasks like content creation, simple Q&A, translations, chatbots, etc.<\/strong><\/td>\nSuitable for building autonomous applications that requires strategic planning, advanced research, and specialized task execution<\/strong><\/td>\n<\/tr>\n
Towards AGI<\/td>\nA step in the journey towards Artificial General Intelligence, but with limitations<\/td>\nRepresents a significant leap towards achieving Artificial General Intelligence<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

 <\/p>\n

Un peu \u00e0 la mani\u00e8re des ant\u00e9diluviens outils de screen scraping<\/a>, les propri\u00e9taires du Rabbit R1 vont lui confier leurs identit\u00e9s\u00a0num\u00e9riques (code d\u2019acc\u00e8s banque, etc.) pour que\u00a0les LAM\u00a0se connectent au site de la banque afin de consulter les soldes des comptes et ainsi ne\u00a0proposer que des offres de voyages qui entrent dans le budget par exemple.<\/p>\n

Cette nouvelle mani\u00e8re d\u2019acc\u00e9der aux services doit nous interpeller \u00e0 plusieurs niveaux, car si cette technologie\u00a0co\u00fbteuse au niveau IA\u00a0(GPU) est n\u00e9cessaire, c\u2019est que\u00a0:<\/p>\n