Regardless of the rise of AI, SaaS corporations proceed to play an important position, as giant language fashions (LLMs) can’t perform as databases. Sridhar Vembu, founding father of Indian SaaS firm Zoho, lately defined that neural networks “take in” knowledge in a manner that makes it inconceivable to replace, delete, or retrieve particular info precisely.
In accordance with Vembu, this isn’t only a technological problem however a basic mathematical and scientific limitation of the present AI strategy.
He defined that if a enterprise trains an LLM utilizing its buyer knowledge, the mannequin can’t replace itself when a buyer modifies or deletes their knowledge. It is because there is no such thing as a clear mapping between the unique knowledge and the skilled parameters. Even when the mannequin is devoted to a single buyer, there is no such thing as a solution to assure that their knowledge adjustments can be mirrored precisely.
Vembu in contrast the method of coaching LLMs to dissolving trillions of cubes of salt and sugar in an unlimited lake. “After the dissolution, we can’t know which of the cubes of sugar went the place within the lake—each dice of sugar is in all places!”
Notably, Klarna CEO Sebastian Siemiatkowski lately shared on X that he experimented with changing SaaS options like Salesforce by constructing in-house merchandise with ChatGPT.
His expertise with LLMs was fairly much like Vembu’s. Siemiatkowski stated that feeding an LLM the fragmented, dispersed, and unstructured world of company knowledge would end in a really confused mannequin.
He stated that to handle these challenges, Klarna explored graph databases (Neo4j) and ideas like ontology, vectors, and retrieval-augmented technology (RAG) to raised mannequin and construction information.
Siemiatkowski defined that Klarna’s information base, spanning paperwork, analytics, buyer knowledge, HR information, and provider administration, was fragmented throughout a number of SaaS instruments similar to Salesforce, buyer relationship administration (CRM), enterprise useful resource planning (ERP), and Kanban boards.
He famous that every of those SaaS options operated with its personal logic, making it tough to create a unified, navigable information system. By consolidating its databases, Klarna considerably decreased its reliance on exterior SaaS suppliers, eliminating round 1,200 functions.
Microsoft chief Satya Nadella, in a latest podcast, not directly took a dig at Salesforce by saying that conventional SaaS corporations will collapse within the AI agent period.
He famous that the majority enterprise functions—similar to Salesforce, SAP, and conventional ERP/CRM programs—perform as structured databases with interfaces for customers to enter, retrieve, and modify knowledge. He likened them to CRUD databases with embedded enterprise logic.
Nadella defined that AI brokers won’t be tied to a single database or system however will function throughout a number of repositories, dynamically pulling and updating info.
“Enterprise logic is all going to those brokers, and these brokers are going to be multi-repo CRUD. They’re not going to discriminate between what the again finish is; they’re going to replace a number of databases, and all of the logic can be within the AI tier,” he stated.
RAG is a Stopgap Resolution
Vembu argued that RAGs have their very own limitations and can’t totally handle the core downside of AI fashions being inherently static as soon as skilled. “In that sense, neural networks (and due to this fact LLMs) will not be an acceptable database.”
“The RAG structure retains the enterprise database separate and augments the consumer immediate with knowledge fetched from the database,” he added.
In high-stakes functions, similar to monetary transactions, medical information, or regulatory compliance, this lack of adaptability might be a big roadblock.
“Vembu’s observations about LLMs’ static nature resonate strongly. The ‘frozen information’ downside he describes isn’t simply theoretical — it’s a sensible problem we grapple with day by day in manufacturing environments,” Tagore Reddi, director of digital and knowledge analytics at Hinduja International Options, stated.
“Whereas RAG architectures provide a workable interim resolution, particularly for delicate enterprise knowledge, they introduce their very own complexity round knowledge freshness, latency, and system structure,” he added.
Nevertheless, many developments are going down right now in RAG, particularly together with vector search. At this time, many database corporations like Pinecone, Redis, and MongoDB provide vector seek for RAG.
Pinecone lately launched Assistant, an API service that simplifies constructing RAG-powered functions by dealing with chunking, embedding, vector search, and extra. It permits builders to deploy production-grade AI functions in below half-hour.
Equally, Oracle lately launched HeatWave GenAI, which integrates LLMs and vector processing throughout the database, permitting customers to leverage generative AI with out requiring AI experience or knowledge motion.
In the meantime, Microsoft Azure provides Azure Cosmos DB, a totally managed NoSQL, relational, and vector database that integrates AI capabilities for duties like RAG. Azure additionally supplies Azure Cognitive Search, which makes use of AI for superior search and knowledge evaluation.
Knowledge warehousing platform Snowflake lately launched Cortex Brokers, a totally managed service for integrating, retrieving, and processing structured and unstructured knowledge at scale.
For now, LLMs can’t change databases as a result of they lack real-time updates and exact knowledge management. Companies nonetheless want dependable database options alongside AI.
The submit Why Companies Shouldn’t Deal with LLMs as Databases appeared first on Analytics India Journal.
