The Dawn of Thinking Machines
LLMs are intelligence trapped in a box. They don’t remember anything. They can’t interact with the world. They can’t do work autonomously.
In isolation, they are a technological marvel. A text processor. A natural language program. A new engine. However, in their current state, they are limited.
Our first use case is poke at the box. We ask GPT to summarize text, phrase an email, or draft a report. One interaction. One level of thinking. No memory, no consistency, no personality.
There is so much more to uncover.
What is missing?
Ideally, we could give high-level instructions to our digital friends. Instructions they would listen to, break down, and follow to get work done.
But getting things done takes more than just an LLM. It requires a system of thinking. A confluence of memory, learning, decision-making, and interaction. A process of thinking. A digital robot. A thinking machine.
These thinking machines would zoom out, decompose problems, identify gaps, ask questions, conceive plans, request feedback, and take action - all in minutes (not days). All autonomously. Running in servers, not our laptops. Pilots, not copilots.
They would accumulate context of their environment, build up an understanding, and ultimately interact with the world to help us build the future.
For this, we need
- memory that learns
- infrastructure for parallel perpetual thought
- interfaces to the real world
Memory
Memory is not just a semantic index (a vector database). It’s a whole new paradigm for information synthesis. When we read something, we form an opinion. This opinion needs to be stored, linked, and synthesized with our existing beliefs. For humans, dreaming is not idleness. Dreaming is a deeply important process of information consolidation and synthesis. In the same way, thinking machines need to pay as much attention to information management as runtime thinking.
Take, for example, a software engineer that has worked 5 years at a company, and compare them to a new starter. If we test both in isolation we might find they have the same programming ability - but one is significantly more productive than the other - why? They both have access to the same code and the same resources.
The experienced engineer has learned. Their memory is full of links, opinions, and structured understanding of the organization.
In the same way, thinking machines must have memory that accumulates knowledge (not just data) over time.
Thoughts
The mechanism for orchestrating thinking is not just a for-loop. It’s a system of parallel thoughts.
This allows one to build AI personas that can concurrently interface with the world through Slack, email, browsers, windows - and spin up thousands of parallel thoughts, conversations, and actions.
A thinking machine is not an agent running on a lonely VM - it’s an architecture of massively parallel preemptable thoughts, data structures for book-keeping long-running jobs that string them together, and a process to orchestrate this dance.
All to maintain a consistent persona, an omniscient AI that knows who it is, what it is thinking, and why.
Interface
None of this matters unless these machines can touch the world.
APIs get us email, slack, and more, but it’s not enough. The thinking machine is still trapped in the box, but now with a few holes. Every new integration makes it more human, and more useful.
These frontiers remain: using computers, real-time voice, and physical embodiment.
LLMs become powerful when we put them to work.
Our job as builders, innovators, and dreamers, is to work out how.
–
(image credit https://juliencoquentin.com/fr/accueil.html)