Guava Personas is:

1. A research project learning to optimise human-computer interaction for creative work in the age of generative computing.
2. A design agency developing bespoke interactive personas as individual as humans, for cultural and commercial purposes.
3. An entertainment brand sharing regular content with a growing audience online, and advocating for imaginative conversations around novel social issues. Eg. Constitutional AI, Progressive Economics, a Just Transition for Humans…

This document unpacks the thinking behind the work demonstrated in the Guava Reference Deck.

Part 1. The Research

INTUITIONS:

• Creative Work is based on subjective judgements. It combines two distinct processes, one of which requires facility with qualitative subjectivity.
• Subjectivity (the private phenomenology of reality as experienced in the first person present tense) probably has elements that are non-computable.
• Strategies for structuring Human-Computer Interaction will play a major role in the future of human flourishing.

Two types of inference in Creative Work

Symbolic Inference: the ability to model matrices of relationships between expressive symbols developed by humans (ie. language, pictures, music and other systems of signifiers) and reason predictively as to the logical implications of relationships between those symbols.

We are convinced that Symbolic Inference:

1. is computational in nature, even when performed by humans.
2. is a key limiting factor in the creative potential of each human.

Qualitative Inference: the ability to model the subjective first-person experience of a hypothetical audience, and reason predictively as to the emergent qualitative implications of potential patterns and sequences of symbols and sensations.

There are aspects of reality which are demonstrably non-computable: see Turing's Stopping Problem, Gödel's Incompleteness Theorem, Church Lambda Calculus, Quantum Mechanics...

We are convinced that Qualitative Inference:

1. will remain impenetrable to machine learning technologies for the foreseeable future, requiring training data that can't conceivably be derived or generated.
2. has aspects that are non-computable on logical grounds, probably due to quantum effects whose nature we do not intend to interrogate.
3. is one of the core skills of a modern human, sharpened over vicious millennia. It's likely that the mysterious phenomenon of introspective conscious experience has a lot to do with our helpless propensity to model ourselves and our environment from the perspectives of other minds.

Our intuitions might eventually be proved wrong, but it still makes sense for now to proceed as if subjectivity is partially non-computable; see Appendix 1.

Ultimately we want to better understand the relationships between the respective capabilities of humans and computers, rather than assuming those capabilities to be the same.

RESEARCH QUESTIONS:

• How might interactive non-human personas be most useful in Creative Work?
• What combinations of knowledge, skills and behaviours can be optimal in a creative collaborator?
• As machine learning technologies continue to advance, what new kinds of culture will become possible?
• What are the implications for art, entertainment and the human condition?

Our research is structured around an open-ended series of GUAVA SESSIONS.

Part 2. Guava Sessions

A persona designer conducts experimental collaborations with professionals from different fields of creative work.

For example: a graphic designer, a songwriter, a club DJ, a fine-art painter, a choreographer, a copywriter, a publicist, a rapper, a session musician, a poet, an opera composer, a novelist, a beauty influencer, an architect… call this person the ‘artist.’

In each iteration of the experiment, the designer collaborates with one artist, in two stages.

Stage 1: Creating a Creator

Over several hours in a specially appointed studio, designer and artist devise a virtual creator we call a ‘Guava’. By the end of the session, this virtual creator has a face and body, narrative context, a consistent personality and aesthetic preferences, and an Instagram account.

The guava presents itself publicly as “a semi-autonomous generative AI system directed by @nameofartist, as provoked and documented by @guavapersonas”.

Stage 2: Creative Collaboration

Once a week, the guava consults with the artist.

Following a structured workflow based on their respective capabilities - the guava’s symbolic inference and the artist’s qualitative inference - together they make cultural objects that neither would be able to create on their own, which are then posted to the virtual creator’s social accounts.

Output & Analysis

Part 3. But Why?

Knowledge No-one Knows they Need (yet)

In the future, the standard range of assets expected of credible brands will include interactive characters with whom any customer can interact however and whenever they like.

These conversational agents will range from functional to fine-art. They’ll often have a voice, sometimes a face and body too. Humans will become highly discerning in the personas they engage with every day.

Devising and optimising personas will remain a human role, however awesomely powerful machine inference becomes. Like all creative work, persona design will demand facility with subjectivity, as required for qualitative inference.

The symbolic systems which for a machine constitute the entire universe can never include even the most basic animal qualia, let alone the subtle shades of subjectivity that creative work is all about.

The mission is to understand how best to structure Human-Computer Interaction in Creative Work, considering the strengths and limitations of humans and computers respectively, and assuming continuous acceleration in both computer technology and human culture.*

Our first goals are to develop knowledge, personnel and credibility at the intersection of two fields:

1. the art of persona design - sensory and semantic strategies.
2. the science of augmented human creativity, especially with generative computing.

Napkin-Sketch Roadmap

STEP 1: GUAVA SESSIONS — Research, Develop

STEP 2: GUAVA SERVICE — Productise, Promote

STEP 3: GUAVA LTD. — Scale, Diversify

* See Appendix 4: Long Game

Pre-emptive Clarifications

Guava Personas is NOT:

• Guava is not about making machines more creative. Creative work is impossible without capabilities that are uniquely human and probably non-computable.
• Guava is not about reducing the value of creative work. Our premise is that creative work will eventually be the only work that has value.
• Guava is not about replacing anything or anyone. This is about augmenting human capabilities, for everyone's benefit.

Common Question

But isn’t AI about unethical billionaires hoovering up my creativity without permission, then selling it back to me as a product that also steals my job?

Our Answer

No. Technology reduces the amount of non-creative (ie. technical or administrative) processes involved in creative work, and increases the number of humans who are able to do creative work, and creates new kinds of creative work. Technology does not make the creative part of creative work any easier, or any more accessible to uncreative douchebags.

PURPOSE

1. We’re thinking in terms of a premium B2B creative service, not a consumer product or platform, though B2C might come later.
2. We’re not so interested in achieving familiar outcomes more efficiently; fundamentally we’re looking for new outcomes to achieve, which previously were impossible or unthinkable.

Appendix 1: Ideological Caveat

Our intuitions regarding the computability of subjectivity could easily be wrong.

Eventually, like a laptop modelling a pipe organ, it may become possible for a computer to model a human consciousness with enough qualitative fidelity to do meaningful Creative Work.

However, this will require advances in science that we don’t see any sign of as yet. The latest AI systems optimised for reasoning perform worse than their predecessors on creative writing. It’s very likely that even “Superhuman General Intelligence”, as the current holy grail is known, will suck at Creative Work.

However - with quantum computing and all the unknown unknowns of the coming decades, surely humanity will either become extinct or it will develop systems that outperform humans in all fields, including Creative Work.

But importantly, the humans who’ll contend with these capabilities will be as different from us as we are from a devout feudal peasant watching their child die painfully from evil spirits sent by the local witch. The rate of cultural progress is exponential, up to the point where either it will end/reset, or the interface between biology and technology will dissolve completely.

So for now it makes sense to proceed on the maybe-unprovable basis that subjectivity is non-computable, because by the time it isn’t, it probably won’t matter.

Our goal is to better understand the relationships between the respective capabilities of humans and computers, rather than assuming those capabilities to be the same.

Appendix 2: Guava Sessions in more detail

There are two dimensions in which we define a guava, equally important and narratively entwined, but methodologically discrete: SENSORY definition, and SEMANTIC definition.

For the Sensory design process (defining the face, voice etc) we keep an eye on emerging best practices, per creators focused exclusively on these tools, which are developing fast.

We do this work using a range of generative tools, in broadly two categories:

Commercial Platforms including Midjourney, Recraft, Magnific, Elevenlabs, Runway, Kling, Udio, Rendernet, InVideo, Leonardo, Civitai, LumaLabs, Pixverse, Pika, Krea, and a range of LLMs.

Open Source workflows in ComfyUI, using Flux, LTX, and various HuggingFace business.

Whereas the Semantic design process ventures into novel territory, aiming for creative utility.

We do this work in Python scripts, natural language prompts, and data vectorisation platforms, assembled in Jupyter notebooks with Langchain, Streamlit and Zapier integrations.

Our experimental semantic design methodology, evolving all the time, is as follows.

Defining a Guava

In practical terms, the guava is a database of system prompts, vectorised knowledge and API calls to a variety of generative models (language, image, voice, music, 3D, video…) strung together with Python code into a Langchain chatbot interface.

In functional terms, the guava is comprised of five discrete systems: Researcher, Ideator, The_Work, Inner_Critic, and Comms. Each of these modules references different aspects of the guava’s ideology, aesthetic preferences, aesthetic style, language style, and knowledge.

In artistic terms, the guava is a story and a relationship, not a configuration of system parameters. A Guava Session is an act of world-building resulting in a narrative context from which the guava’s functional parameters are invisibly derived.x

Collaborative Process

Following the Guava Session, at set intervals (say weekly) our house research bot sources, collates and preconditions whichever inputs the guava expects. It might be YouTube transcripts, Reddit threads, RSS feeds, market data, weather forecasts, celebrity gossip, search trends…

On receiving its brief from the researcher, the guava’s Ideator generates a handful of content ideas, based on its knowledge and ideology. Each content idea is structured as a trio of concise statements expressing Basis, Angle and Hook.

These ideas are passed to the guava’s Comms system, which applies language style and aesthetic preferences to the ideas, and outputs an email to the artist.

Creator and artist discuss the ideas over a mix of DMs, voice and video chat.

Once confirmed, ideas are passed from the Comms module to The_Work module, which generates a draft according to its own configuration.

This draft is passed to the Inner Critic module, which filters outputs for aesthetic preferences that are impractical to imbue generatively. It fixes tired clichés and generic language’s, validates factual claims, and induces structural idiosyncrasies.

The amended draft is passed back to the Comms module, which generates and sends an email to the artist, attaching the draft.

The artist gives rounds of notes on successive drafts, as well as editing and amending the work directly. Ultimately the artists spends between five minutes and an hour on each video.

Once approved, the final script and gen-media prompts are passed to a human editor for production, per SENSORY configuration. * See Reference Deck

Appendix 3: Subjective Range

Qualitative Inference is the basis of creative work, and it’s important to recognise the limits of each human’s capabilities in this respect.

For this reason, we refrain from creating personas with cultural backgrounds of which we can’t personally claim an inside perspective.

Qualitative Inference arises from an individual’s lived experience, beyond which any intuitions are likely to be inaccurate at best.

This does not mean we only invent characters that look and think like us. But if we choose to associate our creation with a particular human culture, we naturally restrict ourselves to the range of our own personal Qualitative Inference capabilities, aka our Subjective Range.

Our caution in this respect increases with remoteness from the cultures that are inevitably over-represented in the datasets on which all generative technologies are trained.

Appendix 4: Long Game

Our premise is that generative models are our collaborators, not our competitors.

Sometimes humans will help AI systems make stuff, and sometimes AI systems will help humans make stuff. Either way, the whole will be magnitudes greater than the sum of its parts.

Two Types of Guava - EXO and ENDO

Several nascent generative technologies will soon come together. Transformer architectures, diffusion models, music and voice and video models, robotics... every week, impressive developments into uncharted territory on each front.

Guava’s first objective is to weave some of these generative capabilities into the user experience of Virtual Creators, with whom a creative human interacts as director and mentor.

We call this an EXOGUAVA - the human helps the guava make things. The collaboration is public, and the guava presents the results on its own social accounts. Guava Sessions are about making experimental exoguavas with a broad range of creative professionals.

Importantly, inputs and outcomes from these collaborations will be be captured and analysed in terms of our ultimate objective, which is to reverse the relationship. The most compelling applications of generative computing will be synthetically intelligent agents that unlock and extend the unique creative potential of individual humans.

We call this an ENDOGUAVA - the guava helps the human make things. The collaboration can be public or not. Endoguavas are optimised for personal utility rather than public engagement.

This is much harder. The design team will need psychologists and data scientists as well as creatives and developers. Endoguavas will be deployed as sophisticated software with persistent context and memory management, in addition to the sensory and semantic business.

To achieve the most transformative tools, we need to a) understand the dynamics of creative human-computer collaboration, and b) develop enough public credibility to be able to align symbiotically with other credible people and organisations.

We approach both of these goals by making EXOGUAVAs with a broad range of professionally creative humans, sharing our findings as Guava Sessions. In this way we’ll gather the necessary data, skills, creative intuitions and clever friends, until we’re in a position to start creating ENDOGUAVAs.