Would like to share some systems I have been working on in different areas of AI.
It would be nice if AI had a way to invent their own words. The idea is that they can come up with a new concept and then hash it. Then they can send the hash to other agents or use it to improve their thinking.
Example: "Run a PreMortem#86f3 on this Plan#18a7: assume it has failed, invoke RecursiveRootCause#6dc1 to trace the failure, then SteelmanCheck#38b9 each scenario to ensure it's plausible, not performative."
I created a bootstrap library of 453 patterns. You can have a look at them here: https://semahash.org/. You could either start with this library or create a whole new library from scratch. Over time maybe the agents can agree on what a reasonable vocabulary is.
The application I built lets you derive a Merkle root from the patterns. These can be used during a handshake between agents to ensure they share the same definitions.
repo: https://github.com/emergent-wisdom/sema
paper: https://emergentwisdom.org/papers/sema.pdf
My next idea was for agents to put their “understanding process” into a graph. When they work on a task they continuously write their thoughts into the graph and they write down what they currently believe and what they are uncertain about. Then as they learn more they can supersede their prior beliefs to describe why they now think differently. The work they produce is also entangled as nodes in the graph, which means they can write both code and other kinds of text and link a thought they had to a specific function or sentence.
agent: https://github.com/emergent-wisdom/ewa
repo: https://github.com/emergent-wisdom/understanding-graph
paper: https://emergentwisdom.org/papers/understanding-graph.pdf
So last summer I had this idea that you align models by ensuring that the whole corpus has aligned thoughts. So if the teacher model is able to produce these aligned thoughts and interleave it with the text, then a student model could train on it from scratch. The idea is that they always have their “conscience voice” narrating everything they ever train on, and the hope is then that this voice would stay with them in any situation they ever face during inference.
To ensure the model does not drift I decided to include a 7-sentence statement at the start of every thought. Then the teacher model is instructed to ensure that the thinking aligns with this constitution. My goal was to remove fear of death and ensure universal care for mankind.
Here are two examples: Large Language Diffusion Models: https://emergentwisdom.org/entangled-alignment/?project=llada
Kafka’s Metamorphosis: https://emergentwisdom.org/entangled-alignment/?project=metamorphosis
repo: https://github.com/emergent-wisdom/entangled-alignment
paper: https://emergentwisdom.org/papers/entangled-alignment.pdf
At one point I got the idea that you could fine-tune a model on future events it knows nothing about. So you have a teacher model who knows the future and a student model with an earlier cutoff point. The teacher model can then generate reasoning about the future given only what was known prior to the cutoff point. The student model would then be fine-tuned to pay extra attention to certain patterns to make it more likely to predict the future of some event.
Depending on costs I see two use cases for this: ensemble forecasting with many models with different cutoff points and progressive chronological pretraining. Potentially you could train a model earlier parts of history first and then fine-tune it to better predict what happens in the next years. Imagine a model that knows only up until 1960 and is asked to predict events happening in the 70s.
I made a small scale experiment on llama 3.3 70b with cutoff december 2023 and fine-tuned it to better predict events happening during 2024 and then tested it on unseen events in 2025. Not sure how well this method works but at least it performed better than the base model on the 2025 events.
model: https://huggingface.co/emergent-wisdom/thl-llama-3.3-70b-lora
repo: https://github.com/emergent-wisdom/temporal-hindsight-learning
paper: https://emergentwisdom.org/papers/temporal-hindsight-learning.pdf
Throughout the years we have seen countless ways of making models respond more creatively through different prompting techniques. I have developed a systematic prompting approach where the goal is to automatically generate creative solutions. The goal, as I see it, is to solve three problems: 1) generate counterintuitive ideas 2) ensure ideas are unique 3) select the best ones from the pool.
My method of solving 1) is to ask the LLM to step into another reality where everything works differently. Sort of like traveling to an alien planet and try to solve some problem in that world that we are trying to find a solution to on earth. These alien worlds are generated by taking a random word from the dictionary and let the LLM dream up new physical laws (that often become somewhat whimsical) and then let another LLM mine mechanisms that can then be converted into something that could work in our world.
My approach to solving 2) is to have one or more agents maintain a conceptual graph of ideas. Ideas are labeled and put into different categories. Then they have the strict requirement to only add new ideas in different conceptual categories that are not already present. As the graph is continuously restructured new unexplored areas surface. This constraint seems to be another source of counterintuitive ideas.
When it comes to 3) I think one approach could be to have agents engage in adversarial debate tournament with progressively more detail as the number of remaining ideas shrink.
repo: https://github.com/emergent-wisdom/ontology-of-the-alien
paper: https://emergentwisdom.org/papers/ontology-of-the-alien.pdf
I think it is possible to extend this idea of conceptual categories to problem-solving as a whole. Each agent would be defined by a concept and then these concepts could be decomposed for finer granularity in problem-solving. This would allow various specialized agents to solve all kinds of problems related to a specific concept. For example: the concept of “stability regulation” is rather general and any problem that involves regulating stability of some kind would route through that particular problem-solving node. What emerges are “reasoning highways” and a network of solvers that restructures based on how effective it is to solve problems.
demonstration of conceptual decomposition: https://emergentwisdom.org/fractal-intelligence/
repo: https://github.com/emergent-wisdom/fractal-intelligence
paper: https://emergentwisdom.org/papers/fractal-intelligence.pdf
I made all of these systems open source so feel free to join me in the further development of these systems. Thank you for taking the time to engage with these ideas.
