What's better than a large language model?

This short post explains how shifting from centralised large language models to federated tiny ones will unlock the true potential of GenAI.

The seasonal excitement about AI has arrived again thanks to the remarkable capabilities of ChatGPT, and for once the AI is living up to the dream. It's fair to say that ChatGPT finally passes the Turing Test - a standard conveniently ignored during the previous decade of "intelligent" applications and enterprises.

There are two qualities that excite me about ChatGPT.

First, I love the capabilities of computational linguistics. ChatGPT demonstrates proficiency at converting natural language into semantic elements, and then back into natural language. This is critical to its ability to process a "large language model" (i.e. a huge body of textual information) into an internal semantic model.

The second and most important thing that's awesome about ChatGPT is that we can finally store information independently from host applications. LLMs are standalone data - meaningful in their own right. It is not siloed away in an application database. The LLM combines everything we need to know about the subject matter, fusing data and metadata in a way that's independent from any particular applications or data definition.

The capability to store digital information meaningfully, independently from any application or predefined structure, enables data to be truly commoditised, reused, shared and even traded. We can finally see a path to replacing application-centric API-based digital ecosystems with human-centric ones, built on schemas that support open and secure participation, anonymous engagement, evolving trust and semantic disambiguation.

Large language models illustrate that data can exist meaningfully in its own right.

So what's better than a large language model?

A Tiny Language Model (TLM).

A kind of Post-It note of each little piece of information generated during the day (this was the vision behind the semantic web e.g. schema.org).

If we can represent small chunks of data, like a shopping list, an invoice or an email, meaningfully outside any specific application, then we have a set of information building blocks that can be aggregated into the equivalent of a highly personal LLM.

Our personal LLM would even be able to interact with other people's LLMs, to perform specific tasks or request specific information.

What could a Tiny Language Model (TLM) look like? The key is to provide semantic precision in a human and machine readable syntax. To achieve this, I've found that structured data can be reliably expressed in English using "has a" and "which is" statements nested together.

For example, let's consider how (part of) my LinkedIn profile could be saved into a tiny language model using "has a", shown as square brackets, and "which is", shown as a hyphen (or plus where the object relates directly to the subject).

Pete Chapman
- LinkedIn profile [ www.dhirubhai.net/in/petechapman
    - First Name [Pete]
    - Last Name [Chapman]
    - Current company [Ernst & Young {April 2015 - Present}
        + Logo [www.ey.com/logo.png]
        - Occupation [Enterprise Architect |  
https://en.wikipedia.org/wiki/Enterprise_Architect_(software)] 
        - Home office [Sydney]
    ] 
]           

Semantic precision is achieved by establishing clear relationships between attributes and values, and by referencing an authoritative external URI.

ChatGPT has no trouble understanding this text, and happily converts it to other syntaxes as well (albeit losing some important nuance).

JSON translation by ChatGPT:

JSON

{
  "name": {
    "first": "Pete",
    "last": "Chapman"
  },
  "linkedinProfile": "www.dhirubhai.net/in/petechapman",
  "currentCompany": {
    "name": "Ernst & Young",
    "startDate": "April 2015",
    "endDate": "Present",
    "logo": "www.ey.com/logo.png",
    "occupation": {
      "title": "Enterprise Architect",
      "wikiLink": "https://en.wikipedia.org/wiki/Enterprise_Architect_(software)"
    }
  }
}
        

The future from here

In the first stage, I predict we will increasingly see Tiny Language Models embedded into web pages, scanned documents, pdfs, and even EDI documents so they can be easily parsed and incorporated into personal LLMs. There will be a new category of application software that let individuals and businesses collect and curate these scraps of information, combined with LLMs, into a manageable and highly interactive work space, like the finger tip UI that Tom Cruise used in Minority Report. There will be agreement on how to incorporate temporal and qualitative nuances into the data.

As the idea of data existing meaningfully outside applications spreads, mini-applications will be developed to perform specific tasks using selected TLM subsets extracted from LLMs. These "agents" will help curate and manage your information collection, perform transactions with other LLMs, and for businesses, perform value added services with customer LLMs. Semantic translation services will pop up making it easy to combine data from different sources using different standards.

A connected ecosystem will emerge, enabling TLMs to be shared securely between participants while preserving the lineage for authentication and validation, even auto-refreshing. The ecosystem will enable participants to extract a TLM from their existing data to request a quote, share photos, order dinner, register a car and get insurance, etc. No need to type everything in over and over again - how many times have you entered your credit card details into a traditional website?

Trusted institutions which already have Know Your Customer requirements, will develop valuable new services for things like brokering trust and anonymous transactions. Payment providers can securely link transaction participants to each other, where additional detail like invoicing or extended warranties can be available as their own Tiny Language Models.

Public sector departments will also eventually come on board, providing customers/citizens with certified TLM data, which they can share selectively to other departments of third party providers. TLMs will enable the entire machinery of government to reorient around data-centricity with citizens holding and curating the information governments generate for them. The new opportunities for contestability and service efficiency will be profound.

I hope you share my excitement about the coming opportunities for data-centric ecosystems enabled by TLM style data aggregation and sharing. Let me know your thoughts, and any projects you know that are already going in this direction.

This article is written in my personal capacity and does not necessarily reflect the views of my employer.