Return to portfolio

When can I try it?

I've exported a set based on mathematical theorems extracted from Wikidata, as well as a 2-layer deep traversal of Wikipedia's Systems science category page (to extract systems science concepts/people/ideas, etc.) 

What is undesirable is that it doesn't cover a wide range of subjects yet which would be most useful. 

Finally, lots of noise occurs with GPT-3 generations (does not always give you the right answer), so I think for this application to really work it needs funding to fine tune a GPT-3 on a small dataset to increase its batting average. A live Colab demo of the prototype is up here for you to try. 

Below are some samples generated by GPT-3:

"Input" means what I prompted GPT-3 with, and "Output" refers to what GPT-3 was able to generate after being given a query like "Drinking Water"


In an ideal world, with GPT-20. 

For now, we’re stuck with the next best thing: GPT-3: a language model by OpenAI.

1. GPT-3: Generates vague stimuli, action, and result descriptions for a host of ideas. (I've tested it, results are promising)

2. I take each list of descriptions and embed it using a language model like RoBERTa for semantic similarity (or some other better model)

3. You search by entering something like this:

Input/Discriminative stimulus: cycle-ness, periodicity 

Action/Operant: locomotion 

Output/Reinforcement: movement-ness, position-change-ness

But you have the option to not constrain the search by what you think will solve the problem, by simply entering ‘cycle-ness’ into the input section, as well as the output potentially. In other words, the search is flexible. 

How does it work?

1. Find potentially relevant tools: It might help you solve new problems you’ve never come across before. Since you can search by rough descriptions of your circumstance. For instance, you’re trying to get a robot to walk, or move. You can use semantic search to search abstractly by saying “cycle-ness” or “periodicity” etc. And you can specify 1. the circumstance vaguely 2. the action vaguely and  3. the result you want vaguely (vagueness in this case being an asset to increase creativity)

2. Exaptation: You might be able to re-purpose already existing tools, technologies, concepts, etc. that have never been used in the context or experience or problem you're facing.

3. Synthesis: It might, given a new kind of stimulus that doesn't quite have its own determined action, help you arrive at a new kind of equation that is between a quadratic equation and a power law, say (due to the continuous nature of the space) (Because you can say “I feel like this new experience is 50% like when I typically use a quadratic equation, and 50% like a power law” or whatever)

What would this three-layered map allow you to do?

You see an oreo wrapper, and your immediate behavior is to reach for the oreos (assuming you like them), and eat them. The oreo wrapper is a discriminative stimulus that affects your behavior.

Imagine this: we have a map of stimuli. 

In other words, we know all the circumstances that humanity comes across. 

For now, let’s focus on scientists. 

What is the typical situation that leads a marine biologist, a person in optics, a farmer to use the quadratic equation? What do all the circumstances have in common? 

Something to do with ‘curve-ness’? Wanting to find the bottom or top of a curve ‘optimization-ness’? 

This is the first layer of our map: a vector space of stimuli

The next layer of our map is: what, on an abstract level, actions do these scientists go through when they apply something like the quadratic equation? 

What is their action pattern and can we abstract it into broad words that may allow us to blur the boundaries to other related ideas and hence allow us to see connections we hadn’t seen before? 

A vector space of actions. (Such as applying the quadratic equation, using calculus, or even tricks from single-celled organisms  etc.) 

The final layer of our map is: What result do these scientists typically get after applying such an equation or concept? A vector space of results

In the diagram above, each color represents a different kind of stimulus (for example: curve-ness, or position-ness) It's a continuous space because it might be useful to imagine stimuli that are between disparate stimuli, like a vector space being able to say "this new 'grey area' stimulus which I'm experiencing is 10% like the yellow stimulus (curve-ness) and 20% like the purple stimulus (position-ness)" (e.g. a mix between a circle and a square is a rounded square), the soft-ness of the colors corresponds to the fuzzy nature of the stimulus, or to what degree it generalizes. 

What is it?

A new kind of search

Triplet Pattern