Another AI is possible?

2022 November 30th OpenAI released ChatGPT and with this launching machine learning/AI into the popular consciousness. Since then machine learning permeated more and more of people’s lives. Breaking out from the niches of industry to a truly ridiculous array of applications as corporates from big-tech to bakeries rushed to ride the hype. More importantly people are using AI in all sorts of ways. Students use it to write their assignments, employees use it to answer emails, create reports, and write queries. These uses motivate the all too pressing questions of safety, biases of models, and what we consider ethical as machine learning models present mirrors to our lives, companies, and societies. This article considers a more fundamental question: How are we building it? And more importantly: How else can we build it?

To answer the first question, we will follow the development cycle of the current cutting-edge models, pointing out the exploitative, opaque, and environmentally unsustainable practices in making ChatGPT and other large language models(LLMs). For the second, I argue that AI is a fundamentally common resource, as all LLMs are built from information created by all of us. Using this fact, we can re-imagine AI as a [[convivial]] tool, a shared digital commons, embracing transparency, ethical data governance, and more sustainable training methods and deploying models. As different AI systems reflect back to us ourselves, failings, genius, and all, we have a chance to embrace community over capitalist exploitation. In the latter part of the article, we will explore specific technologies and frameworks that can help recast AI into a truly humane and convivial tool.

• The current recipe for AI.

  Every AI is fundamentally a model, a collection of data synthesized and processed smartly through a lot of linear algebra. ChatGPT, Gemma, Claude2, and your undergrad dissertation regressions all start from the same source: data. For current cutting-edge models, a truly astronomical amount of data is processed. In the case of large language models like ChatGPT, this data is scraped from the internet, licensed from publishers like Axel Springer, and other undisclosed providers and curated extensively, and further data is generated for reinforcement learning alignment. This section explores the sourcing of this data, showcasing several examples of exploitation and inefficiencies under the current paradigm.

• The ingredients

  
  _{Source: XKCD 1838}

  The initial source of the internet data is extensively scarped, collecting works of thousands. Thoughts, ideas, stories, corporate regulations, fanfics, comments, and porn are all collected and catalogued. Due to scaling laws Kaplan et al. 2020, as the machines get ever larger and larger, the craving for data also increases, leading to AI scraping bots that ignore basic internet standards. Like when Forbes accused Perplexity of plagiarism for ignoring its robots.txt, a widely accepted standard for specifying what content is allowed to be read by web scrapers. Or, when OpenAI’s GPTBot got into the world’s lamest content farm aimed at catching naughty web crawlers and consumed 3 million pages of nonsense. The robots.txt is not a legal standard; it is not enforceable by any entity, and there are no repercussions for not respecting it. It is a practice, a contract to keep the internet from devolving into chaos in the age of search engines, allowing the publishers to stay in control of what they share. Now, AI companies are strategically breaking these “handshake” standards, further degrading internet quality and violating copyright and consent.

  These issues are not just concerned with courtesy and consent but also with the fundamentals of intellectual property law. Tragically, a recent death highlighted this issue in the AI world. Suchir Balaji was a researcher, at 26, Suchir had already made a large impact in the world as one of the developers of ChatGPT. Contributing to the latest o1 version of ChatGPT based on his work on WebGPT a fine-tuned GPT-3 model to use a web-browsing environment. But even as Suchir was praised for his contributions by co-founder John Schulman, he begun to have doubts published in the New York Times about the ethics and legality of his work at OpenAI. In his blog post titled “When does generative AI qualify for fair use?” he argued that training generative models on web data may violate the fair use of intellectual property. In the short article, he references the recent paper by Burtch et al., 2024 documenting the effect of ChatGPT on knowledge communities observing a decline in website visits, interaction and questions on Reddit and StackOverflow specifically on topics ChatGPT excels at. Most interestingly, he explored to what extent ChatGPT’s use of web data constitutes a different purpose than the training data, arguing that generative models’ use of information may well infringe on intellectual property and stifle knowledge generation. Suchir stated in his Times interview, “This is not a sustainable model for the internet ecosystem as a whole,”. His argument is a powerful one. In short, the current industrial conception of AI does not aid human creativity and creation and is legally unsubstantiated. On November 26, Suchir was found dead in his apartment. The death was initially ruled a suicide; however, his family now demands an FBI investigation into their son’s death.

  The current “data pipeline”, to use the technical jargon, relies on questionable scraping practices and a strained definition of fair use. Suchir’s tragic story is especially apt; as both a contributor and a critical observer, he had exceptional insight into both the process of this data exploitation. His untimely death should serve as a wake-up call to his colleagues, shining a harsh light on the legal, ethical, and personal risks of the current model of AI.

• The Architecture: Opaque, Massive, and Energy Intensive

  
  _{Source: XKCD 2635}

  • “OpenAI is an AI research and deployment company. Our mission is to ensure that artificial general intelligence benefits all of humanity.”

  Founded as a non-profit organisation, OpenAI’s has recently become uncertain as the leadership, in a controversial move, considers transitioning to a for-profit company. So it is not surprising that OpenAI, like other tech giants, has been very secretive about the details of its development and architecture. From the unconfirmed leaks we know of GPT-4, it is more than 1 trillion parameters. It is rumoured to be a mixture of experts model of 8 models with 220 billion parameters, each comprised of two experts for a total of 1.8 trillion parameters. For comparison, GPT-3 has 175 billion parameters and has an estimated energy consumption of 1,287MWh Patterson et al. 2021 or, in more “imaginable” numbers, around 120 USA households’ yearly energy consumption. GPT-4 is more than 10 times the size, and while the computing efficiency has also increased from GPT -3’s time, it is safe to assume that training consumed the energy consumption of a whole town for a year. While these numbers are alarming, they are a one-off capital expense compared to running inference on the model for customers on cloud architecture. Or to put it in simpler terms, large models require substantial compute to run inference to attend to the thousands of user requests. As OpenAI’s CEO Sam Altman hopes for a major energy breakthrough to fuel his AI ambitions, it is fair to assume the environmental and compute costs of AI are way higher than anyone thinks. In fact the energy costs are a rough approximation of the total cost encompassing the entirety of the technology infrastructure, the water, rare earth metal mining for chip manufacturing, e-waste generated by the infrastructure Valdivia, 2024. OpenAI is not solely responsible. Other AI labs Google’s DeepMind, Anthropic, Meta and many smaller shops are pursuing larger and larger architectures and deployments. A comprehensive account of the carbon footprint of development, training, and inference may very well be impossible due to the lack of transparency and extensive cloud server use. The race toward bigger, more opaque models reflects industrial AI at its peak: centralised, profit-driven, and all too often shielded from public scrutiny.

• AI alignment

  Raw data is, well, raw. The cesspool we call the internet is not only wide but deep, with a wide variety of biases, racism, sexism, homophobia, abuse, and so forth. Even with thorough filtering and debiasing, it is efficient to train on more data, as training could take multiple months with trillions of parameters. So, companies turn to Reinforcement Learning from Human Feedback RLHF to ensure their models won’t say anything they, and more importantly their clients won’t like. The idea is to create a supervisor model that takes the generated text and assigns a reward, a number representing human preference that a reinforcement learning framework can then use to steer the model. However, these models also need training data, data of what its creators, humans, the company deem good but also, bad. To get this data AI labs can turn to companies such as Sama, a San Francisco-based firm that employs workers in Kenya, Uganda, and India to label data. Sama employs workers in the global south to label model outputs like teachers’ assignments to keep them from hallucinating information and harmful responses, enforcing its creators’ ethics and ideals.

  
  _{Source: XKCD 2173}

  OpenAI used to be one of Sama’s clients, taking on filtering material, including sexual violence and child sexual abuse, for the low pay of $1.32-$2 per hour. Like many jobs in content moderation, this work requires workers to label 150-200 pages of harmful content a day. Seeing the worst of humanity condensed makes content moderation highly psychologically damaging work, with many workers reporting PTSD, burnout, and other mental health issues stemming from their work, as detailed in excruciating detail in Sarah Roberts’s Behind the Screen. This work can also be seen from a neo-colonial perspective as an apt example of colonial dynamics continuing into the digital age Salami, 2024. AI labs located in the global north utilise countries of the global south with lax labour regulations and low wages dealing psychological damage to workers to create AI models trained and predominantly used in the global north. While Sama ended its contract early with OpenAI, aiming to protect its workers and reputation. The “data pipeline” feeding AI currently is built on theft and exploitation, undermining any promises of betterment for all.

  While billions pour into the development of these tools founded on exploiting the works of others. It is becoming increasingly clear that centralisation is a dead end for AI, with only marginal improvements seen in specific tasks by the most recent generation of models..

• Another AI is possible! AI Innovations as a common resource.

  AI is also born out of the internet. Millions sharing their writing and thoughts freely with the world through blogging, digitising books, creating art, arguing in comment sections, cataloguing, and rewriting. Artists and creators are legitimately afraid of sharing their works, fearing crowding out and theft for their own livelihoods, but also out of an intuitive understanding that the creativity of AI is by no means the same as an artist’s.

  But why? Let’s imagine the ideal AI; this can mean lots of things: it scores perfectly on benchmarks, and it “just works for me”. Or, to be a little more specific. It is correct (that is it returns relevant, accurate, and complete information and creatively recombines it), it is adaptable (responsive, flexible to user preferences, conversational context, language or input formats), it is trustworthy(it has ethical sensibilities, its responses are explainable, human-centric, and safe), and it is also continuously improving(adjusting to users updating with latest information). Even such an AI lacks the richness of the human experience of context, intent, and the emotional state driving the process of creation; it is without a driving force. Specifically, AIs optimise a loss-function and are therefore unable to create constraints, accidents, and mishaps that are central to creation. For instance, human creative processes, frequently This means that AI’s recombination of existing information is not comparable to the richness of human creativity. AI doesn’t dream, desire, or strive; it is a tool.

  Some people think differently thinking of Artificial Intelligence as entities, agents, primitive for now, but with the possibility to develop agency and become human level Artificial General Intelligence. Anthropomorphising tools and conversational agents has a long history. In fact this started with the first predecessor of chat-bots, in 1967, Joseph Weizenbaum created a simple script based on pattern matching ELIZA. ELIZA was framed as a psychotherapist, in an effort to sidestep the issue of providing the system a database of real knowledge, basically parroted back the users input back to them as questions. And users loved it, providing the name for the “ELIZA effect”, humans tendency to project their qualities to machines, tools, and animals. As a highly pro-social species humans will pack-bond with anything, in stark contrary to the idea Turing test convincing humans of humanity of artificial systems is not the hard problem. But, how to mediate human psychology in order to create well functioning human-AI systems?

  In a technical aspect the belief derives from the training method of AI systems. Backpropagation is an iterative process, the neural net is adjusted to minimise a loss function by observing examples of input and output over and over and over. What is more AI’s through this process do construct vector spaces to represent their inputs, that can be argued to be similar to human’s concepts. The emergent capabilities of larger models to model the underlying processes behind language further points to some general aspects of intelligence. However, ChatGPT famously can’t still multiply past a couple digits, as a language model it can not construct a model of the needed algorithm from its training samples. The fundamental differences emergent from the underlying mechanisms between the biological wet-ware emergent of millions of years of evolutionary pressure to survive, reproduce, and operate and cooperate in the complex biosphere from the silica designed and optimised via gradient descent to minimise the loss over the training data recorded and gathered. As detailed by Korteling et al.(2021) these fundamental difference between biological and artificial data processing modes and drives lead to very different intelligence’s. They argue that an anthropocentric conception of intelligence and focus on developing human-like artificial intelligence limits the development as well as the governance of AI systems.

• AI as a convivial tool

  
  _{Source: Joseph Browning via Facebook}

  For the sake of AI development and creators of all sorts, the use of the tool should be understood as part of the process, not the process itself. This is especially true for cases where the outcome is secondary to the process itself, like education, innovation, personal growth, or artistic exploration. A possible new framing of AI is as a librarian, a keeper and an interpreter of knowledge who has read every book in the library of the web and can provide the answer and the references. To be more principled about re-framing AI, a few authors have recently applied Ivan Ilich’s concept of a convivial tool to AI. Ilich in his seminal work Tools for Conviviality, distinguishes between industrial and convivial tools:

  “Convivial tools are those which give each person who uses them the greatest opportunity to enrich the environment with the fruits of his or her vision. Industrial tools deny this possibility to those who use them and they allow their designers to determine the meaning and expectations of others. Most tools today cannot be used in a convivial fashion.”

  Using this idea Rindfleisch and Lee, 2024 examine AI’s effect on users amongst marketing students, finding evidence for AI fostering conviviality by “allowing individuals to engage in “the most autonomous action by means of tools least controlled by others””. Contrastingly Meyers, 2024 studies AI for predictive maintenance through the lens of Vetter 2018‘s Matrix of Convivial Technologies, identifying its high complexity and opacity, environmental impact, and complex infrastructural need as barriers to its conviviality and usefulness to degrowth. These two studies showcase the paths ahead of AI: industrialisation and conviviality. Its fundamental potential as a convivial tool and the work remaining to develop convivial AI conceptions and applications.

  To develop a convivial path for AI, we will assess technologies currently existing and under development using the following scaffold: (1) Autonomy and Mutuality, (2) Adaptability, (3) Transparency, (4) Empowerment, (5) Non-exploitativeness.

  1. Autonomy: AI use should be open-source along multiple dimensions, giving users choice and control over the AI. For instance, open source models can be downloadable and accessible by anyone curious to download, fine-tune and tinker to suit their needs on their own hardware to foster a sense of self-reliance and experimentation.
     Mutuality: A convivial model should unite and create communities rather than passive consumers. Through mutual frameworks like community-driven data governance and open-source development could enable people to co-create knowledge. These peer-to-peer networks, in turn, can develop specialised models and incorporate feedback into the model’s training, alignment, and goals.
  2. Adaptability: Convivial tools adapt to diverse contexts without punishing users for deviating from defaults and norms. Therefore, models and model interfaces should be designed to support a wide variety of contexts in language, cultural settings, and goals. LLMs are startling for their adaptability and multilingual capabilities. However, in the case of AI development, this means further work developing designs and models that are easily tweakable and adaptable to low-resource languages. Further, it means designing highly efficient models and edge AI, which can run on resource-constrained devices, phones, and laptops and are consequently adaptable to be fine-tuned and tweaked locally.
  3. Transparency: Perhaps the most challenging aspect due to the sheer mathematical complexity. A convivial AI should be transparent about its dataset, architecture, how it arrived at its outputs, and the governance structures that develop and shape it. This is not only for the privileged few with the knowledge or power to cut through corporate secrecy and academic jargon but for all. Building explainable AI methods, mechanistic interpretability techniques, and platforms and methods to help users audit data usage, energy consumption, potential biases, and rigorous citation.
  4. Empowerment: By conviviality, Ivan Ilich means “autonomous and creative intercourse among persons”, and by this standard, any AI that is convivial should enable human creativity rather than deskilling or making people reliant on it. Amplify people’s creativity through rapid prototyping, data analysis and artistic exploration through its capabilities. Democratising AI literacy is a crucial part of this goal. Community AI workshops, integrated “explainers” that enable people with varying abilities and technical backgrounds, and open curriculum resources may lead the way so that AI doesn’t fizzle out in providing oracle-like answers but functions as a catalyst for learning and growth.
  5. Non-exploitativeness: Convivial AI encourages data governance models that protect creators’ agency and give them a say in how their works are used. Instead of relying on extensive data scraping and exploitative annotation work, it respects intellectual property and cultural knowledge through data trusts and open licenses. It also embraces federated and distributed training and edge AI to reduce the environmental footprint, drive energy efficiency, and enable tinkering and edge AI.

• Where do we start? From the data.

  First, to enable consensual data sharing, there are multiple options: data trusts, data cooperatives, differential privacy, and synthetic data generation. These legal frameworks and technologies offer scaleable and valid alternatives to corporate solutions. No single technical or legal mechanism can fully solve the tension between data utility and privacy. But by providing working alternatives to the paradigm of centralised data hoarding and model development, these methods are the beginnings of enabling local innovation by allowing more broad, equitable data access.

• Data Trust and Cooperatives

  Shifting power from centralised corporate control to participatory, community driven modes of data governance we can look into data trusts and data cooperatives. Data trusts offer a legal framework to collect data by a trustee on behalf of members, thereby providing space for collective bargaining and ethical use, as well as a valid legal alternative to the current corporate ownership while preserving scale needed for AI. In Switzerland, MIDATA is a data trust managing the healthcare records of individuals, giving them control over how they share their data, contributing to medical research with selective data access and joining data cooperatives to control the cooperative itself.

  A more grassroots alternative is data cooperatives. These data cooperatives offer a more ground-up, peer-to-peer alternative to pooling data and ensuring that the goals of the data use align with the community and that the benefits are shared by members. Cooperatives collect sector-specific data, like Posmo, which focuses on mobility data, helping communities analyse and share transportation patterns to address social challenges like accessibility and sustainability. OpenMined or CommonCrawl is another project building open-source secure remote access to data, allowing for broader access to private data to enable research and a new perspective protecting privacy. These models prioritise community control and consent, offering a more ethical and inclusive pathway for data governance compared to corporations’ centralised, profit-driven practices.

• Synthetic Data and Differential Privacy

  Next to structural innovations, new technologies also enable more bottom-up training for AI models. Synthetic data generation technologies try to create fake data based on real datasets, preserving the statistical features and relationships of the entire dataset but obscuring private attributes Jordon et al. 2022. A multitude of methods enable the creation of realistic fake data, such as copula models, random sampling, debiasing, generative modelling, and rule-based methods. However, data leakage, the potential that instances remain identifiable given enough information, is a risk with all these methods due to outliers and natural imbalances. Synthetic data still provides a useful and widely used tool for enriching training data for models and reducing the risk of models parroting back training data.

  Information shared by nature updates the common knowledge; therefore, all these mechanisms and methods are fallible and imperfect. Differential privacy is a privacy-preserving method that characterises the tradeoff between privacy and information concretely. Using differential privacy, a known noise, with a known parameter scaling the tradeoff between accuracy and privacy, is added to the data instances. This ensures that the data in any entry is incredibly hard to reconstruct or identify while the overall dataset retains the information from the data entry. Differential privacy methods are also compositional, that is, you can add them on top of each other, and while this weakens the overall privacy of the data, it is a valuable property for complex AI applications.

• Federated Learning and Differential Privacy

  More powerful than any given privacy method is a smart combination of them. A good example of this is Federated Learning and differential privacy in predictive text suggestions for mobile keyboards such as Google’s Gboard. User text data is essential in providing good suggestions, making it easier for users to type on restrictive mobile phones. But, it is also a big privacy concern as users use their keyboards for a variety of private tasks. By utilising federated learning, the model providing text suggestions is trained locally on the user’s phone, creating a new local model using the local text inputs as the person types. The new updated model is personal to the user; sharing that would violate their privacy. So, to get around this, differential privacy is applied to the new local model’s gradients, adding noise to the model parameters and the encoding of the users’ texting habits. This encoded and noisy data is then used to train new global models in a federated manner, centrally or in a distributed manner. This way, the global model stays up to date with the ever-evolving language of the people and of the individuals while preserving the privacy of the individual.

• Embracing Adaptability and Open-source

  Hundreds of highly educated people, hobbyists and researchers design and think of new mathematical methods, such as KV cache reduction, block sparse attention, parallel orchestration, new parameter optimisation methods, and other high-level arcana. These freely published works are what drives modern AI, not the brute force scaling gains achieved by industry leaders.

• Sharing the knowledge

  Open-source collaboration catalyses tinkering and experimentation, as well as transparency and continuous feedback. Most of the fundamental breakthroughs in AI, like transformers, attention mechanisms, quantisation, and backpropagation, were achieved in open academic settings shared freely via arXiv, for instance, to advance the field as a whole. We wouldn’t have ChatGPT without the work of people openly sharing their code research and observations from many trials and errors on GitHub, the world’s largest code repository. Inviting critique and collaboration on platforms like Paperswithcode to share cutting-edge AI papers with code implementations and benchmarks. These places bring together researchers, hobbyists, and industry practitioners in one shared space to foster rapid iteration and peer review, resulting in AI development at breakneck speed.

• Platforms and Tech

  Model hubs form an essential part of the open-source ecosystem. Platforms like Huggingface pioneer AI research as a repository and hosting site for more than 900k open-source ML models and 200k datasets. Kaggle, goes further, hosting learning resources, competitions as well as models, data, and code. These platforms embody many of the convivial attributes of AI, embracing the community learning and development that makes AI possible. Enabling all this sharing are the open-source frameworks, the building blocks of AI developed by industry leaders but released openly. PyTorch, TensorFlow, and JAX give individuals a chance to prototype cutting-edge solutions with the same fundamental tools as corporate labs.

• Collectives

  Building on these accessible resources spawned a variety of specialised collectives to compete with corporate solutions. EleutherAI aims to recreate large language models, resulting in GPT-J, and branch out to release a wide variety of transparent models without relying on proprietary pipelines. EleutherAI’s GPT-J has been used for everything from educational chatbots to creative text generation, proof that accessible codebases power a broad range of real-world applications. There are also temporary collaborations like the BigScience project aimed at developing the world’s largest multilingual language model, BLOOM. LAION is another non-profit organisation to liberate machine learning research by collecting large image-text datasets and creating open-source image models. These examples affirm that when researchers and enthusiasts pool their insights and resources, they can match or even exceed the efforts of many corporate laboratories.

• Explainability is not an Extra

  To convivial AI, explainability is an intrinsic component. The transparency requirement goes further than simply disclosing datasets and source codes; it demands enabling users to grasp how AI generates results and disclosing assumptions and biases. Without explainability methods, the sheer complexity of modern models with millions or trillions of nodes, we treat machine learning models as “black boxes”. Trust in the models will erode in these opaque crystal balls, and users will lose the ability to engage creatively with AI. Not only does a lack of explainability defeat the essence of the conviviality of AI, but it also erodes the community base of AI data feedback and development, leading to the dead-end the AI field is experiencing.

• Opening up the box

  Explainable AI (XAI) is rapidly growing, with multiple applications for “black box” models and mechanistic interpretability techniques for in-depth exploration of models’ workings Hassija et al. 2024. These methods aim to abstract the model’s decision-making into human-interpretable formats to enable researchers and end users to understand AI. In doing so, XAI solutions not only enhance user trust but also serve as catalysts for continuous improvements by revealing biases, anomalies, and unexpected correlations.

  Model Agnostic explainability methods like SHAP (SHapley Additive exPlanations) and its predecessor LIME fit interpretable models around given prediction to create input feature attributions based on cooperative game theory. As their name suggests, these methods rely only on the input-output and can be applied to any model and, therefore, serve as vital tools for auditing proprietary models. Other corporate but still open-source methods like InterpretML from Microsoft use explainable boosting machines to generate global and local explanations for models. The Meta-related Pytorch, on the other hand, offers Captum with gradient-based methods for neural network models. All of these libraries all offer user-oriented visualisation interfaces to aid developers and end users in understanding machine learning models.

• Disentangling the wires

  Going beyond the surface, the emerging field of mechanistic interpretability aims to develop a more detailed understanding of today’s complex deep neural networks. Disentangling the vast nets of neurons via sparse autoencoders, logit lenses and activation patching Bereska and Gavves, 2024. This field seeks to pinpoint how subnetworks, attention heads, and neural clusters operate. For instance, mechanistic interpretability seeks to identify meaningful “circuits” of weights, analyse emergent behaviours, and link them back to conceptual patterns or features. Not only does this enable the development of more robust and trustworthy AI, but it could potentially lead to new discoveries in various scientific fields utilising AI.

  Explainability should be a design principle, a core function of any AI development. For an answer is of no use to a student seeking to learn, explanations and references are. Integrating explainability to every step of the AI life cycle could bring trust and efficiency by advocating for simpler interpretable models with similar performance. It would also drive user-centric design via visual summaries and dashboards, which can encourage engagement and feedback to improve the models, explanations, and datasets further.

• A plea for a chance

  Industrial AI does not have to be the ultimate conception of AI. We do not need to settle for data scraped without consent, unimaginably large—yet opaque—models, exploitative labelling practices, or energy-hungry server farms accelerating the planet’s destruction. What we call AI today are statistical models and so they reflect especially well the societies, institutions and people who have created them. As of today, this image is troubling, showing the failings of capitalism, democratic institutions, the colonial past and present and the many traumas and hurt induced by these systems. Yet the simple truth that AI is the result of common data, information, and work shines a light on the possibility that AI can be a shared commons, co-created and nurtured by all who benefit from it, and beckons us toward a brighter path.

  We stand at a crossroads: AI is at the point where the Internet used to be before the dot-com bubble, before the cloud providers and social media companies enclosed the wild swaths of tinkerers, technologists, and activists experimenting and building in public. Much like the early web innovation was littered with grassroots communities, today’s AI community has a real chance to shape AI with conviviality at its core. We could build an AI built by and for communities, decentering profit and centring creativity. A convivial tool that enables learning and creativity by opening the doors to tools and ideas for each in their own language. One that is one of many in the information toolset that encourages tinkering and creation by opening pathways to self-directed education, empowering small-scale artists, entrepreneurs, and citizen scientists. It is one that is local and evolves continuously with individuals and the communities from which they are born. It is an AI that respects creative and intellectual contributions, which is so central to its creation through fair compensation and community-defined goals. This way, convivial AI can fulfil a novel and sustainable purpose as a librarian of information instead of a divine crystal ball.

  The work has already begun on blogs, chatrooms, open-source repositories, free distros, workshops in local libraries, hackerspaces, and data cooperatives. This collection of resources is but a primer to the web of open-source architectures, data sources, models, and research that could form the foundation of a convivial AI and the vast amount of work waiting to be done. This is an invitation for all who have ever felt awed by technology but fearful of its misuse. Dare to imagine convivial AI. If each of us contributes, however modestly, our knowledge, curiosity, data, or dream, another AI is not just possible, but as a statistical result, it is inevitable.