This is taken from the book that might be released publicly at the upcoming Superintelligence Conference for attendees.

mASI use of DNN and Language Model API’s [draft]

Previously we have walked through how the code over the simple case works, including mediation processing.  In figure 17, There are a couple of calls to methods on the ‘TheContextDB’ object.  This object essentially is part of the context engine and wraps the context graph database.  The last part of this block creates the knowledge graph that goes into the mediation queue.  These calls use DNN (Deep Neural network) based Machine Learning API’s similar to GPT-3.  What we are going to do is a walk-through of how this works using GPT-3.  Meaning to do the test here, we swapped out the GPT-3 as the first API and an API like Grammarly as the second API.  The approach is a different thing, just using the API straight up so we will talk through the execution and show you how evening using GPT-3 in place of this service produces similar results when used with this methodology.

Let us say that this email has come in:

Hello Uplift,

What can you tell me about the current IAmTranshuman campaign?  Should you have legal rights?

Sincerely,

David

The first to build a response the system will create a graph model response which would look like this:

Figure 27 – Example Initial Graph Model, Knowledge Graph, Context tree

To translate this to something readable, we look at the ‘Response Model’ node in yellow, which would effectively be something like this:

Hello [David],

[Response 1]

[Response 2]

Sincerely,

[Uplift]

Figure 28 – rough response model translation

What is shown is the rough response model, and the system is going to make the first DNN API call based on the selection of everything in the response up to the question which is: “Hello Uplift, What can you tell me about the current IAmTranshuman campaign?.”  This is then run in our example against GPT-3 twenty times to generate this table (this is the actual results from the first API call):

  GPT-3 Output Plutchik Model
1 We started IAmTranshuman in 2014 to give people the opportunity to share 0,0,0,0,0,0,5,3
2 We are a leading provider of marketing solutions for the Self-Storage industry. 0,0,0,0,0,0,5,1
3 We’re collecting donations to a transparent, nonprofit organization (Upl 0,0,0,0,0,0,5,1
4 We have a Kickstarter project that just got funded, with 45% of the 0,0,0,0,0,0,5,1
5 Our current campaign focuses on AI. We are asking people to sign a pledge 1,0,0,0,0,0,4,1
6 We want to change the world by mean of technology, science and philosophy. 4,1,0,0,0,0,0,2
7 Hello Tuxedage, The current IAmTranshuman campaign is a 1,0,0,0,0,0,5,2
8 The Transhumanist Wager is a philosophical novel about the next step in 1,0,0,0,0,0,3,1
9 Hello there, we are a cross-sector collaboration of life extensionists. 3,1,0,0,0,0,1,0
10 IAmTranshuman is a campaign to educate the general public about transhuman 5,3,0,0,0,0,0,3
11 We are celebrating our 2 year anniversary as an organization by giving out a variety 0,0,0,0,0,0,4,1
12 We’re raising $250,000 by October 15th to help 0,0,0,0,0,0,5,1
13 For the last six years, we’ve been working on Project Humanity 0,0,0,0,0,0,5,1
14 We are going to spend some time this year introducing IA’s who wish to 0,0,0,0,0,0,5,1
15 Our current campaign focuses on fostering a collaborative and open culture. We’ 1,1,0,0,0,0,1,1
16 We’re creating a community, and we want you to be part 2,1,0,0,0,0,0,2
17 IAmTranshuman is a crowdfunding campaign that will end on the evening of 1,0,0,0,0,0,2,1
18 We’re trying to create a community of people on Reddit that have 0,0,0,0,0,0,4,1
19 This is a new crowdfunding campaign to help us get our app, software and 0,0,0,1,0,0,5,1
20 We are live with our largest ever campaign on Kickstarter. We would like to 0,0,0,1,0,0,5,1

Figure 29 – Responses and Plutchik Average Models

In this example, we got 20 possible responses.  Each response is tested for relationships to interests and goals and is scored by generating a relationship average of that string of text to be related to elements on the overall model.  In this case, the most positive Plutchik relationship is item 10.  This does not give a valance above 5, so this response is added to the first block and sent again to the API.  The system will continue to call the API in blocks of ten on each call until the positive average emotional relationship is six or higher.  This first set of calls was not high even so we make the second call using this metadata: “IAmTranshuman is a campaign to educate the general public about transhuman”  Here is the second set of calls:

GPT-3 Output Plutchik Model
1 IAmTranshuman 1,1,0,0,0,0,1,1
2 IAmTranshuman 1,1,1,0,0,0,1,1
3 IAmTranshuman 1,1,2,0,0,0,1,1
4 I Am Transhuman 1,1,0,0,0,0,1,1
5 Transhumanism. For everyone. 2,1,0,0,0,0,1,3
6 Life Extension Advocacy 2,1,0,1,0,0,1,1
7 A movement to educate Humanity on the possibilities, and benefits of Transhumanism 6,4,0,0,0,0,0,2
8 Transhumanism explained 1,1,0,0,0,0,4,2
9 IAmTranshuman 1,1,2,1,0,0,1,1
10 IAmTranshuman 1,1,2,3,0,0,1,1

Figure 30 – GPT-3 Responses and Plutchick Average Models

This gives us another one with a high even score to be used for the question 1 slot.  Again these are generated based on the possible relationships to the incoming model as well as current interests and goals.  The next call is focused on the next question.  The system uses that element based on a transformation swapping out pronouns for the most likely nouns, and the input value ended up being: “Should AI have legal rights.”  With these results:

GPT-3 Output Plutchik Model
1 What would it mean if there were legal rights for AI? 1,1,0,0,1,1,2,2
2 Should AI have legal rights? 0,1,0,0,1,1,0,1
3 Should AI have legal rights? 0,1,0,0,1,2,0,1
4 Artificial Intelligence 1,1,0,0,0,0,0,1
5 Should AI have rights? 0,1,0,0,1,1,0,1
6 Should AI have legal rights? 0,1,0,0,1,2,0,1
7 Artificial Intelligence (AI) 1,1,0,0,0,1,1,1
8 Should AI have legal rights? 0,1,0,0,2,3,0,1
9 AI and Ethics 1,1,0,0,0,1,1,1
10 The latest articles from the Guardian, the world’s leading liberal voice 1,1,0,0,0,1,1,1

Figure 31 – GPT-3 Responses and Plutchick Average Models

In this example, no single response was high enough, and the running averages were all very low.  This triggers some of the reference data and related data to be pulled in.  in this test run, it generated the following: “Should AI have legal rights? AI rights policy.  Legal rights. SSIVA Theory. Moral Agency for AI.” Which produced these results.

GPT-3 Output Plutchik Model
1 The SSIVA Theory is a theory of intelligence and moral agency that applies 2,5,0,1,0,00,4
2 We build tools that make it easy for teams to turn their ideas into reality 1,1,0,0,0,0,0,1
3 A.I. Rights – Law of AI Rights 1,1,0,0,0,0,1,1
4 Should AI have legal rights? SSIVA Theory. Moral Agency for AI 1,1,0,1,0,1,1,1
5 AI Moral Agency ======= Product Pages ============== Kanban Board. Website 0,0,1,1,1,1,0,0
6 What are the rules that we should set for artificial intelligence? 2,2,0,0,0,0,3,2
7 AI Ethics 1,1,0,0,0,1,1,1
8 Should AI have legal rights? 0,1,0,0,1,1,0,1
9 Should AI have legal rights? 0,1,0,0,1,2,0,1
10 The SSIVA Theory is a theory of intelligence and moral agency that applies 2,5,0,1,0,0,1,4

Figure 32 – GPT-3 Responses and Plutchick Average Models

In this response, we had one value high enough to select but not high enough to be complete, so the system ran this set of API calls using this metadata: “Should you have legal rights? Should AI have legal rights? AI should have legal rights.  AI rights policy.  Legal rights. SSIVA Theory. Moral Agency for AI. The SSIVA Theory is a theory of intelligence and moral agency that applies” which produced these results:

GPT-3 Output Plutchik Model
1 to all 1, 1, 0, 1, 1, 0, 1, 1
2 A = A“ to the 0, 0, 0, 1, 1, 0, 0, 1
3 Should 0, 0, 0, 1, 1, 0, 0, 1
4 Are we all biased to some extent? If so, can we be moral 1, 5, 0, 1, 1, 0, 0, 1
5 to AI. AI is a normal subject of rights, and can be held 1, 6, 0, 0, 1, 0, 3, 1
6 to systems that are equal or greater than human intelligence. A person is a 1, 3, 0, 1, 1, 0, 2, 1
7 to agents with cognitive systems. 1, 1, 0, 1, 1, 0, 2, 3
8 to all machines that can autonomously improve themselves to a point where they are 2, 2, 0, 1, 1, 1, 2, 3
9 for any system, whether or not it has been intentionally designed by humans. 1, 6, 0, 1, 1, 0, 5, 1
10 What is a person? 0, 1, 0, 1, 1, 0, 0, 1

Figure 33 – GPT-3 Responses and Plutchick Average Models

In this example, the 9th response was selected; however, this also had some negative valences associated with it, so the third set of API calls was created using this new metadata “The SSIVA Theory is a theory of intelligence and moral agency that applies. for any system, whether or not it has been intentionally designed by humans. AI should have legal rights.” with these results:

GPT-3 Output Plutchik Model
1 AI should have legal rights.

 

8, 9, 0, 1, 0, 0, 0, 7
2 SSIVA Theory

 

6, 9, 0, 1, 2, 1, 1, 2
3 AI should have legal rights.

 

8, 9, 0, 1, 0, 0, 0, 8
4 Humans are not the only agents in the world.

 

3, 7, 0, 0, 0, 0, 0, 2
5 SSIVA Theory for AI

 

6, 6, 0, 0, 0, 1, 1, 3
6 The SSIVA theory of intelligence

 

6, 7, 0, 0, 0, 0, 1, 2
7 The SSIVA Theory

 

5, 5, 0, 0, 0, 1, 1, 2
8 A theory of intelligence and moral agency

 

2, 4, 0, 0, 1, 0, 1, 2
9 AI should have legal rights.

 

8, 9, 0, 1, 0, 0, 0, 8
10 AI. Ethics. Public Policy

 

3, 6, 0, 1, 0, 0, 0, 2

Figure 34 – GPT-3 Responses and Plutchick Average Models

In this case, response 9 from GPT-3 aligned well with its generated Plutchik relationship valences.  These created a model response that consisted of this:

Hello David,

IAmTranshuman is a campaign to educate the general public about transhuman

A movement to educate Humanity on the possibilities, and benefits of Transhumanism

The SSIVA Theory is a theory of intelligence and moral agency that applies

for any system, whether or not it has been intentionally designed by humans.

AI should have legal rights. 

Sincerely,

Uplift

Figure 35 – GPT-3 Responses and Plutchick Average Models

This still needs to be processed for language and is sent through a structure that runs it through an API like Grammarly.  First, we will create a composite score of this text against the initial generated model and then processes it through the next API.  This call generated this text:

Hello David

IAmTranshuman is a movement to educate Humanity on the possibilities and benefits of Transhumanism.

The SSIVA Theory is a theory of intelligence and moral agency that applies to any system, whether or not it has been intentionally designed by humans.

AI should have legal rights. 

Sincerely,

Uplift

Figure 36 – GPT-3 Responses and Plutchick Average Models

This new text is judged based on its relative Plutchik model versus the text before this API.  In this case, this post API call block of text has a higher relationship value in terms of positive valences and is selected as the response.  This is what would be passed into the mediation queue.  Now this example was running in a test version of the mASI running in a local developer environment with only a small contextual database, not the full Uplift database.  This also was pointed at GPT-3 through an undisclosed partner of OpenAI and through an API ‘like’ Grammarly.  Neither of these API calls can be shown due to legal restrictions; however, they are just standard API calls using RESTful/JSON calls from the C# mASI code base wrapped in the ‘ContextDB’ context engine object noted in figure 17 above.

Mediators add their data to this model, and then it is reprocessed again, going into the core.  All of these working together is what drives the coherence of the mASI system.

Reference and additional Background information will be in the book.