Context is Holding Intelligence Back

Intelligence in the age of LLMs

Broadly speaking, business intelligence used to suggest a combination of data and compute paired with human ingenuity. Setting aside for now AGI and the debate of what role humans have in the future, we've entered a world where business intelligence can be framed as composed of three key elements: Inference, Compute, and Context.

Without all three, you cannot provide intelligence. Context is the lagging problem in the industry today.

Inference

It is tempting to consider inference as the only critical ingredient for intelligence. This is incorrect. There are a few obvious shortfalls:

1. Limited world model

An LLM cannot accurately predict the execution output of a program in a real world scenario. At minimum it needs access to compute to generate a simulation and react to the results. Without compute it can only generate probabilistic predictions of what will work without knowing if it does work. Something will break if it generates Python 3 code but the interpreter is in Python 2.

You can extend this world model into the physical world with cars and robots where you cannot plan out a route in a vacuum, it must be dynamic and responsive to the real world.

2. Limited context

The model can be trained on a large corpus of data, but that corpus will always be incomplete. It lacks access to realtime data (though this is helped with tooling like web fetch), structured data (ask an LLM to print more than a few rows of a public dataset), and private data. In addition, most LLM-driven experiences today lack an ability to collaborate with multiple humans in realtime, hampering context more.

Even with access to all of the context in the world, some models today struggle to produce code that works for the latest version of its own APIs. It needs context and a contextual (ha!) understanding of what is important in that context.

Compute

I'm referring to compute in the broadest way possible: from OpenAI's Operator browsing the web and returning results from internet activity, to executing db queries, to running a python "hello world" script. Bringing compute closer to the LLM results in hugely impressive gains in intelligence: we see tremendous gains in user productivity in Colab between simple LLM-assisted autocomplete in the code editor vs. our latest AI-first agentic programming experience which will generate code, then execute that code in a sandbox and fix errors dynamically as they occur. For the latter, we're seeing enormous gains in terms of time to development for our users and amount of successfully-executed code.

Compute is limited, however, as it lacks any predictive power; it is a reactive lens. Any actions taken are limited by what it is instructed to do, and that is limited by Inference and Context.

Context

Data is the critical unsolved problem in AI today - not for training, but for usage. Context is king. The biggest unsolved problem in applying intelligence as a service is getting the right context to the LLM.

LLMs lack critical context, techniques for accessing context, and abilities to access that context securely.

1. Increasing context

Humans have access to a much richer context window than LLMs today, spanning interpersonal relationships, the covid-era elusive spontaneous hallway chats, documents scattered across the internet and internal intranets, documents authored and shared in a multitude of channels, databases and quantitative information, and the various live meetings they have attended. Humans access this in realtime dynamically.

It is impossibly difficult to grant LLMs the same context window today. Getting LLMs that same context window should be a top priority across intelligence providers.

I played around with this in developing thechrisperry.com - I loaded up an agent with context of my resume, my performance reviews, and anonymized peer feedback. This goes beyond static resume into understanding more of how I work, and points to how increasing context leads to very different outcomes and experiences.

2. Techniques for accessing context

How can an LLM find the exact relevant information from a sea of millions of documents? We've experimented with loading up a model's context window, but we see limited returns after a certain point. RAG gets us part of the way there: you can contextually match documents across a wide space, but it doesn't understand relationships across documents/actions, just that it found a concept that is a match. What if the concept doesn't match?

We need something like a live personalized knowledge graph across the entire context window which uses dynamic compute to expand its state. You could consider Context its own agent: it should query live systems, run commands, find documents, and call APIs, gathering real-time, dynamic context. This context then needs to be understood and ranked in novel ways to focus on the key points, not overwhelm the model with cruft.

3. Security & privacy

LLMs need access to private data in secure ways. There are the obvious "don't leak private data" arguments I'm not going to belabor here, but one critical aspect of intelligence is a shared experience. If there's a really smart person on your team, you can have a shared conversation with the team about thing x and they'll help everyone. How does an LLM get access to the right context AND be able to have a shared conversation across individuals with varying access AND not give out privileged information to the wrong people?

This is a very hard problem. It's hard because it is going to be hard to design a solution that isn't soul-suckingly painful to configure as a human. In addition, we need a sea-change among application UX to enable a new kind of shared intelligence that doesn't pair LLMs so tightly to single user::single surface.

Intelligence as a Service

Imagine opening up an LLM and asking "why are our Q3 sales flat" and not just getting an answer based on queries & logs, but also grounded in context you may not know - maybe sales changed their compensation system discussed in a chat you didn't see, maybe there was a competitive launch discussed in a deck, maybe there's impactful analysis from a third party shared in exec circles you didn't have. This service does not exist today. Context is too hard to manage.

I'm extremely bullish on companies that exist today who have a path to enable intelligence across Inference, Compute, and Context: Google is the obvious best placed on these three dimensions.

I'm skeptical of companies that only play in one domain: if you are only an LLM developer, I don't know how you build durable business intelligence without building or partnering to accrue the other two components, as your value is fundamentally impaired. I'm less skeptical of companies that already have data and compute (wish I could buy Databricks stock!) - inference is getting cheaper and more globally accessible all the time. You can bet on open models giving you everything you need re: inference if not now, soon.

Context is what you need to make scaled intelligence a reality. The people who solve this are going to define the next wave of application development.