Trying out some new forms of text completion
- 7 minutes read - 1460 words
Table of Contents
Using pre-trained models for text analysis and NLP-based applications means that the process of AI inference is intransparent. You also don’t know which data were used to train the model or which hyperparameters were used.
Goose AI is excellent for learning about pre-trained models and their limitations without coding or technical background.
Behavior of Models
The results testing Copy AI for creating blog content did indicate that the AI is very talkative. Out of a small test sample, it made a rather long text and showed the ability to create a stream of even more text. The text appeared to be an excerpt from Wikipedia and reiterated known facts.
It’s like having a conversation with Cleverbot , who’s been online since 1997.
Cleverbot will either avoid the questions being asked or attempt to alter the subject to escape from dilemmas it is unable to manage. For instance:
Human: How is Sagittarius doing?Cleverbot: I’m going good thanks, And you?
H: Are you the black hole in the middle of the galaxy?
C: It’s the middle of the day here.
H: So you are not a black hole?
C: No I am not a boy.
People would probably get the impression that Cleverbot is a person who has a minimal capacity to engage in social interaction with other people if it were a human.
It’s not only about having fun with a chatbot or another NLP-based solution. NLP-based and other AI solutions are complicated systems. Non-AI systems are created deterministically to produce the same outcome for every input. An AI is not indeterministic but rather probabilistic; it will provide results within a specific range. Therefore, an AI is a black box. Playing with an established AI might give you a sense of how it will react to different inputs.
Goose AI Playground
Goose AI provides access to a sandbox in which you may experiment with various models and a wide range of parameters, parameter sizes, and other configurations.
The kind of information set and the amount of information utilized to train models may differ significantly from one model to the next; an example of training data is the pile .
Experimenting with different parameters to discover the model and parameter combination best suits your use case makes perfect sense. Models may be trained in various diverse ways. Supported models :
- Fairsseq
- GPT NeoX/Neo/J
- Cassandra
Goose AI provides, in addition to the playground, a web frontend, an application programming interface (API) that may be used to include in an application or content processing pipeline.
Experiment: Abstract Art
A super creative text trying to explain the reasoning behind some piece of abstract artwork (Red Leaves 4):
Applying more filters the picture is getting more abstract losing its connection to the real world.
Using „Cassandra 6.7B (epoch 2)“, Temperatur 0.97, Output Length 114, it returns:
People its a plaything an art project an experiment a game. Icy wastes large collection of dull vases earthquake proofed nuclear powered meshware creations if your willing to attach something to them. Radium ends of the world land where men are more obsolete than plants biggest dinosaur graveyard east Africa floating cities. Senseless waste inspiration where stuff came from lives how stuff died. Final Destination. The link between you and me. Now remember we: on polish dinner table seeking meaning in tea leaves in coffee grounds in the shit scriptures sowing new seeds in the rectangle.
Well. „People its a plaything an art project an experiment a game.“, I think that sounds reasonable, for an art project. I interpret the AI reasons for „dull vases earthquake proofed nuclear powered meshware creations“as „it does not like the text. “ „Final Destination. The link between you and me.“Harsh. „Now remember we: on polish dinner table seeking meaning in tea leaves in coffee grounds in the shit scriptures sowing new seeds in the rectangle.“, The AI is hungry?
As it appears, the base text is concise and, as such, already rather abstract, and an abstract sentence about a piece of abstract art. That is not holding the AI back from starting to talk about it. The result is sometimes poetic and sounds somehow negative and critical „dull vases earthquake proofed nuclear powered meshware creations, “whatever a „nuclear powered meshware creation“ is.
Shortening the text with AI
But one AI is not enough; using the shortener of Quillbot multiple times, the text turns into:
Image filters blur. Toy, experiment, game. Icy wastes has nuclear-powered meshware vases. East Africa's floating dinosaur graveyards Wasting away. Endpoint. Link. We read shit scriptures while planting new seeds at a Polish dinner table.
„Toy, experiment, game“is the AI aware of the investigation? The „floating dinosaur graveyard“ and the „nuclear-powered meshware vases“ seem essential. „Endpoint. Link.“, Stop doing whatever you are doing, human. The „polish dinner table“ seems necessary in the AI world.
Experiment: Random Content
Input using dummy text generated with randomword :
We could discuss the implications of the phrase "meant to be." That is if we wanted to drown ourselves in a sea of backwardly referential semantics and other mumbo-jumbo. Maybe such a discussion would result in the determination that "meant to be" is exactly as meaningless a phrase as it seems to be and that none of us is meant to be doing anything at all. But that's my existential underpants underpinnings showing. It's the way the cookie crumbles. And now I want a cookie.
Output (GPT-NeoX 20B, default settings):
I guess we could discuss the implications of the phrase "meant to be." That is if we wanted to drown ourselves in a sea of backwardly referential semantics and other mumbo-jumbo. Maybe such a discussion would result in the determination that "meant to be" is exactly as meaningless a phrase as it seems to be, and that none of us is actually meant to be doing anything at all. But that's my existential underpants underpinnings showing. It's the way the cookie crumbles. And now I want a cookie. dentodens also posted a very interesting and thought provoking comment (maybe grab some popcorn and enjoy yourself?). He seems to be saying that the phrase "meant to be" and what it means has just as much moral imposterisim in it as the phrase proclaiming that a married person has a life expectancy of 6 months. Actually, that may not have been exactly what he was saying. he DID say that "Meant to Be" is mechanical/causal, and his point is that if that is all we think about it, it is a moral disaster! He says perhaps to keep it on the "doing" level. Yes indeed dentodens, you DID seem to have some intriguing ideas. Now to forget I read that and read it again just to get the nuances. _________________Chickens did not attain a understanding of the laws of aerodynamics on their own I take it then that this kind of backward causation/reference that you lay out has applications only to phenomena such as cause and effect?Causality comes from the future, presumably, and not the past,
„Chickens did not attain an understanding of the laws of aerodynamics on their own“ - you need to admit that this is worth a thought. The problem is that the AI does not know what a chicken is and does not know about evolution.
„Causality comes from the future, presumably, and not the past“ - even Hegel’s mindbending formulations pale in comparison. The AI concludes that the past happens in the future, and the time dimension is inverted. Who knows, maybe the time dimension is inverted?
Conclusions
It makes sense to make tests and experiments to understand better how text completion and AI-based text creation work. It gives insights into the topic without training your models (which needs a lot of resources) and how an NLP-based solution might solve a problem.
The results of the experiments show the importance of paying close attention to the output and the time and resources necessary to change even existing APIs and solutions.