Tiles Research

Tiles ResearchDiscordGithub

We are Tiles Research, a product-driven research lab, and our mission is to advance the communication of human intent with machines to design a more natural way of working. We are doing so by starting with building an intelligence-age operating system using Rust, WebAssembly, and WebGPU. At its core is an on-device intent router, crafted to seamlessly translate human intent into machine action.

What follows is an early draft of our launch post, a six part essay intended to explain the thesis and methodology of our efforts at Tiles Research, which will be published by Q4 2024.

Introducing Tiles Research

On this page

  1. Homo Faber
  2. Software 2.0
  3. Hybrid AI
  4. Intent Router
  5. Our Blueprint
  6. Conclusion
  7. References


The essay draws on Vilém Flusser's vision of the future factory as a space where the boundaries between human creativity and machine intelligence blur. It explores the growing need for tools that allow us to understand and interact with increasingly complex and abstract intelligent systems. The rapid progress of AI is traced through eras of pre-deep learning, deep learning, and large-scale models, highlighting the seminal "Attention is All You Need" paper that introduced the transformative Transformer architecture. The game-changing release of ChatGPT in 2022 marked a major milestone, and efforts are now underway to make large language models more interpretable and controllable.

The AI landscape is bifurcating between narrow on-device AI and massive cloud models, but a new hybrid AI paradigm is emerging that combines edge and cloud computing, exemplified by Apple's recent AI initiatives. With AI driving the cost of software creation towards zero, a shift is coming—from attention-based software that competes for our focus to ephemeral, intent-driven software that assembles itself on-demand to fulfill user goals. Local AI "intent router" agents will be key to this new software paradigm.

Tiles Research aims to create Tiles, an AI-native operating system that enables intelligent cross-platform experiences. Drawing inspiration from Alan Kay's approach of "inventing the future" by pulling compelling ideas from decades ahead into the present, Tiles Research adopts a three-stage strategy—short-term prototyping, developing a "killer app" disguising the infrastructure, and gradually expanding the app's role—potentially replacing or significantly altering the role of the traditional operating system from within the application itself.

Tiles Research will focus on efficient on-device ML, shifting OS functions to the compiler for hardware independence, and using local-first techniques for collaboration and software longevity by default. It will leverage web technologies and APIs as a bootstrapping strategy, taking advantage of the existing web ecosystem to rapidly develop and deploy the OS.

The mission is to advance the communication of human intent with machines to design a more natural way of working. Tiles Research embraces uncertainty and is guided by real-world use cases in pursuit of this ambitious goal. Currently, the top priority is developing an ML compiler for fast inference to overcome the primary bottleneck in realizing the envisioned system. The essay concludes with a reflection on the challenges and potential of inventing the future, and an invitation to follow along as Tiles Research works to make a dent in the universe of human-computer interaction.

1. Homo Faber

In his 1991 essay The Factory[1], Vilém Flusser envisions a future where the factory evolves from a place of alienation to a center of learning and creativity. Flusser traces the development of human manufacturing from hands to tools to machines, and finally to robots, arguing that each stage redefines our relationship with our environment and ourselves. As we transition into the age of robots, Flusser predicts that factories will increasingly resemble schools or academies, places where humans learn to work with and understand complex systems. This transformation reflects the growing abstraction and sophistication of our tools, requiring ever more theoretical knowledge to operate effectively.

This conception of the future factory as a space for learning and insight aligns with more recent ideas about understanding complex systems, such as Bret Victor's notion of Seeing Spaces[2] from his 2014 talk. Just as Flusser envisions factories becoming places where humans learn to interact with robots, Victor argues for the need for specialized tools and environments that allow us to visualize and comprehend the internal workings of intelligent systems. As our creations become more advanced and abstract, our ability to "see" inside them and understand their processes becomes crucial, not just within individual systems but also across interconnected systems. We are often blind to the complexity that emerges from the interactions between different systems, and it is essential to develop tools that illuminate these intricate relationships.

A truly remarkable body of work reveals its brilliance even when examining just a fragment, often sparking innovative ideas beyond its original context. This is certainly true for both Flusser and Victor, whose insights continue to resonate in our rapidly evolving technological landscape. Their ideas find a striking parallel in Richard Ngo's "Tinker"[3], a speculative fiction that explores the symbiotic relationship between humans and advanced AI. In Ngo's narrative, we see a realization of Flusser's vision of the factory as a place of learning, where the AI protagonist not only designs new chips but also deepens its understanding of nanoscale physics. However, it is crucial to note that the AI in "Tinker" goes beyond merely extending human capabilities in nanophysics research; it augments our understanding by developing entirely new ways of perceiving and interacting with the nanoscale world, as described in Andy Clark's book "Extending Ourselves"[4].

The story illustrates Victor's concept of "seeing spaces" as the AI uses advanced simulations to visualize and manipulate molecular structures. Moreover, "Tinker" takes these ideas further, suggesting a future where the boundaries between human creativity and machine intelligence merge, with each driving the other to new heights of innovation and understanding, transforming the seeing space into the ultimate expression of homo faber - not just a maker of things, but a maker of understanding, where technology ceases to be a black box and instead becomes a transparent partner in the dance of creation, allowing us to see not just the product of our labor, but the very process of thought itself. As we navigate the increasing complexity of our technological landscape, it is essential to develop tools and environments that not only extend our capabilities but also augment our understanding, enabling us to see the intricate connections and emergent properties that arise from the interaction of multiple complex systems.

2. Software 2.0

The history of artificial intelligence stretches back decades, with progress marked by distinct eras of computational power and algorithmic advancement. Based on an analysis of training compute trends, we can identify three major eras in the development of machine learning:

  1. The Pre-Deep Learning Era (1950s to 2010): This period saw the birth of AI and its early development. Starting with Claude Shannon's Theseus in 1950, a remote-controlled mouse that could navigate a labyrinth, AI systems gradually improved in specialized tasks. During this era, training compute approximately followed Moore's Law, with a doubling time of about 20 months. Notable achievements included IBM's Deep Blue defeating world chess champion Garry Kasparov in 1997 and IBM Watson winning Jeopardy! in 2011.
  2. The Deep Learning Era (2010 to 2015/2016): Beginning around 2010-2012, this era marked a significant acceleration in AI development. The doubling time for training compute shortened dramatically to approximately 6 months. This period saw rapid advances in areas like image and speech recognition. By 2015, AI systems began outperforming humans on specific visual recognition tasks, and by 2017, they surpassed human-level performance in speech recognition. A major milestone was reached in 2016 when DeepMind's AlphaGo[5] system defeated Lee Sedol, the world champion Go player.
  3. The Large-Scale Era (2015/2016 onwards): This era is characterized by the emergence of large-scale models developed by major corporations. These systems use training compute 2-3 orders of magnitude larger than those following the Deep Learning Era trend in the same year. The growth of compute in these Large-Scale models appears slower than in the Deep Learning Era, with a doubling time of about 10 months. This period has seen the development of increasingly powerful language models and generative AI systems.

The emergence of ChatGPT in late 2022[6] marked a watershed moment in the field of artificial intelligence, dramatically contrasting with and building upon earlier AI developments. While pre-ChatGPT systems showed impressive but narrow capabilities, ChatGPT and subsequent Large Language Models (LLMs) demonstrated a level of general intelligence and versatility previously unseen. This breakthrough not only brought LLMs into the mainstream but also effectively democratized access to advanced AI capabilities. The impressive abilities of these models to engage in human-like conversations, generate creative content, and assist with complex tasks across various domains captivated the public imagination and sparked a surge of interest and investment in AI technologies.

It is important to acknowledge the landmark research that paved the way for these advancements in natural language processing (NLP). The seminal paper "Attention is All You Need"[7] by Vaswani et al. (2017) introduced the Transformer architecture, which has become the foundation for most state-of-the-art language models. This groundbreaking work, along with other significant contributions in NLP research, laid the groundwork for the development of powerful language models like ChatGPT.

Tech giants and startups alike rushed to develop and deploy their own LLMs, while investors poured billions into AI ventures, recognizing the transformative potential of these technologies. This AI boom not only accelerated the pace of innovation but also intensified the need to better understand the inner workings of these powerful yet opaque systems. As LLMs became more integral to various aspects of society, from business operations to creative endeavors, the urgency to demystify their functioning and ensure their responsible development became increasingly apparent. This rapid progress has brought AI from the realm of specialized research into everyday life, fundamentally changing how we interact with technology and raising new questions about the future of human-AI interaction.

Large Language Models (LLMs) are often compared to black boxes due to their opaque inner workings and decision-making processes, which pose challenges for researchers, developers, and users who wish to understand, control, and steer these systems towards desired goals. To facilitate human creativity and collaboration with LLMs, it is essential to develop intuitive "dials and knobs" that allow users to fine-tune and direct the behavior of these models, adapting them to specific tasks, domains, and preferences. These controls could include adjusting the level of creativity, specificity, or formality in generated text, as well as setting constraints on the output to align with user intentions, enhancing the usability and versatility of LLMs while fostering a more collaborative and interactive relationship between humans and AI systems.

Research efforts by both OpenAI and Anthropic aim to make LLMs less of a black box through their work on AI interpretability[8] with techniques like sparse autoencoders and feature extraction. OpenAI's work on extracting concepts[9] from GPT-4 and Anthropic's efforts to scale monosemanticity[10] in Claude 3 Sonnet represent significant steps towards understanding the internal representations of these models. Both approaches aim to decompose the complex neural activity within language models into interpretable patterns or features, potentially unlocking insights into how these models process and generate language.

As we continue to explore and refine these systems, we may uncover new insights not only about artificial intelligence but also about the nature of information processing and representation in complex systems. This evolution in software development aligns closely with Andrej Karpathy's concept of Software 2.0[11], where neural networks represent a fundamental shift in how we create and understand software, moving from explicitly programmed instructions to learned representations that can adapt and improve with more data and compute. By making LLMs more transparent and controllable, we can harness their power to augment human creativity and problem-solving abilities, ushering in a new era of collaborative intelligence.

3. Hybrid AI

The artificial intelligence landscape is undergoing a significant transformation, driven by rapid advancements in hardware, software, and the increasing demand for AI-powered applications. This evolution is characterized by a bifurcation between on-device narrow AI and cloud-scale general-purpose AI. However, a new paradigm is emerging that bridges these two worlds: hybrid AI architectures. On the edge side, the compute performance of on-device AI accelerators has grown exponentially. Apple's Neural Engine, for example, has increased its performance by 60x between 2017 and 2024[12]. This growth, coupled with algorithmic progress, is enabling smaller AI models to achieve impressive capabilities on narrow tasks while running directly on edge devices like smartphones and laptops. There is a trend of the smallest commercially relevant models shrinking over time. Models with just 1-10 billion parameters[13] are becoming viable for a wide range of on-device AI use cases, such as transcription, translation, and image generation. This compression of models is democratizing access to AI capabilities.

Meanwhile, on the cloud side, the performance of GPUs has been steadily doubling every 2.3 years[14], with a further 10x boost[15] from the adoption of new number formats like 16-bit floating point. Memory bandwidth and capacity are also increasing, enabling the training of massive AI models with hundreds of billions of parameters. These cloud-scale models are pushing the boundaries of artificial general intelligence (AGI). While these two tracks of AI development have been largely separate, the concept of hybrid AI is bringing them together. As outlined in Qualcomm's paper[16], hybrid AI architectures leverage a combination of edge and cloud computing resources to deliver optimal performance, efficiency, and user experiences.

Apple's recently announced Apple Intelligence[17] initiative is a prime example of a hybrid AI architecture in action. Instead of relying solely on cloud-based models, Apple is integrating its own foundation models directly into various features across its devices and services. This approach treats AI as a technology rather than a standalone product. By running smaller, specialized AI models on-device, Apple can deliver features like email prioritization, document summarization, and Siri enhancements with low latency and high privacy. At the same time, more complex tasks are seamlessly offloaded to large cloud models when necessary. This hybrid approach allows Apple to offer deeply integrated, personalized AI experiences while leveraging the collective processing power of millions of edge devices. Apple Intelligence showcases the benefits of hybrid AI architectures that Qualcomm and others envision. By intelligently distributing workloads between edge and cloud, hybrid AI reduces strain on cloud infrastructure, improves energy efficiency, and enables functionality even with limited connectivity. Sensitive user data can be processed locally, enhancing privacy.

Moreover, hybrid AI allows for continuous improvement of models based on real-world usage, benefiting from both local and global insights through lightweight training techniques like LoRA adapters.[18] The interplay between edge and cloud creates opportunities for adaptive, personalized AI experiences across industries, from smartphones and IoT devices to vehicles. As more companies adopt hybrid AI strategies, following the path paved by Apple Intelligence, we can expect a proliferation of powerful, efficient, and user-centric AI applications. However, this rise also presents challenges, including managing distributed AI systems' complexity, ensuring data privacy and security, and developing edge-cloud interoperability standards. As the technology matures, regulatory frameworks and industry best practices must evolve accordingly. In conclusion, the convergence of edge and cloud computing heralds a new era of hybrid AI architectures, enabling more powerful, efficient, and personalized intelligent experiences. Navigating this shift will require ongoing innovation, collaboration, and thoughtful governance to realize the benefits of hybrid AI while mitigating risks and challenges.

4. Intent Router

The software industry is witnessing a paradigm shift that will fundamentally change the nature of software development and user interaction. With the advent of large language models (LLMs) and AI agents, the cost of creating software is rapidly approaching zero, just as the internet drove the cost of creating and distributing content to zero. This shift will lead to the rise of ephemeral, intent-driven software that focuses on fulfilling user intent efficiently rather than capturing and monetizing user attention.

In this new world, users will express their high-level goals or intents, and lightweight, disposable software will dynamically assemble itself to achieve those goals. This approach aligns closely with the concept of home-cooked software – small-scale, personalized applications created by individuals for their own use or for their immediate community. The rise of AI-powered development tools will democratize the creation of home-cooked software, making it accessible to a wider range of people, including "barefoot developers"[19] – a term coined by Maggie Appleton to describe individuals with some technical savvy but not necessarily professional programming skills.

The development of local-first[20] AI technologies will further accelerate this shift by enabling software to adapt to users' needs and contexts in real-time, without the need for constant server communication. Interestingly, the concept of breaking down app silos has historical precedent. The Apple Newton, released nearly 30 years ago, featured a design that allowed entire apps to be nested within one another, with the nested app able to access properties of the surrounding app. This approach facilitated app compositionality and data sharing. Looking forward, the future architecture of intent-driven, home-cooked software will need to address the siloed nature of data across different apps by enabling cross-app data access, potentially facilitated through a new kind of operating system that uses local AI to responsibly manage and utilize data from multiple applications while respecting user privacy and control.

For consumers, this shift will lead to a significant reduction in the cost of digital experiences and increased personalization. Users will have access to a fluid, AI-driven interface that can pull together the exact functionality they need at any given moment, saving money, reducing cognitive load, and increasing productivity. More importantly, it empowers users to create and customize their own software solutions, tailored precisely to their needs and preferences.

Businesses can leverage the AI's ability to access and interpret data across multiple apps (with user permission) to offer hyper-personalized services that truly understand and anticipate user needs. This could lead to more loyal customers and new revenue streams based on the value provided rather than attention captured. However, businesses will need to adapt to a world where users have more control and where generic, one-size-fits-all solutions may no longer suffice.

The democratization of software creation through AI aligns with the long-standing goal of end-user programming[21], empowering all computer users to modify and create their own software tools. In this new paradigm, users won't just be consumers of pre-packaged software; they'll be active participants in shaping their digital tools. The barrier between using and creating software will blur, allowing users to naturally extend and customize their digital environments.

The core idea, as articulated by industry experts like Ben Thompson of Stratechery in an interview[22] with Daniel Gross, is that "The most important agent in AI is going to be the local agent that decides where to dispatch jobs. It doesn't need to be big, it doesn't need to be complex, but it is the linchpin and will control all the value." This local agent — which we can call an intent router — would serve as the user's primary interface to AI capabilities, understanding their needs and routing requests to the most appropriate AI models and tools.

Steven Sinofsky, former President of the Windows Division at Microsoft, has emphasized that AI may indeed require a new OS[23]. This new paradigm is not just about creating smarter AI; it's about making that intelligence accessible, intuitive, and seamlessly woven into existing workflows. The true challenge lies in bridging the gap between raw AI capabilities and practical, user-friendly applications. This insight underscores the importance of rethinking our approach to operating systems in the age of AI, aligning closely with the concept of an intent router as a core component of future OS design.

A critical aspect of this new paradigm is the relationship between the user's device and cloud services. In a user-centric AI ecosystem, the cloud should feel like an extension of the user's computer, not vice versa. Apple's Private Cloud Compute (PCC)[24] exemplifies this approach by implementing a revolutionary security architecture for cloud AI processing. In the PCC model, the user's device encrypts requests directly to cryptographically certified PCC nodes, ensuring end-to-end encryption and data privacy. The intent router, as the user's personal agent, remains the authoritative source of user intent and permissions. Cloud services should respond to and extend the capabilities of the local intent router, rather than the other way around. This principle is embodied in PCC's design, which maintains strict privacy guarantees through technologies such as stateless computation, non-targetability, and verifiable transparency. By adhering to this model, even when leveraging powerful cloud-based AI models, user data remains protected and under the user's control, with cloud services enhancing rather than superseding local capabilities.

Apple's recent unveiling of their AI strategy offers an early glimpse of this paradigm, with a local LLM kernel on the device that listens to requests and decides how to handle them. This approach allows Apple to provide a unified AI experience while maintaining flexibility on the backend. As local AI models improve, more tasks can be handled on-device. Complex requests can be routed to cloud services. And third-party AI providers can be integrated when needed.

Anthropic's introduction of Artifacts[25] represents a significant step towards making AI a true collaborative partner in the workplace. This feature allows users to manipulate, refine, publish, share, and remix AI-generated content[26] in real-time, fostering a collaborative environment where users can iterate on AI-created content. This concept aligns with James Addison's vision of future AI interfaces[27] that allow users to create and consume media in whatever way feels most natural, including seamless semantic transformations[28] between different media types and complex manipulations.

The remixing capability of Artifacts addresses the lack of creative control in AI-generated content by providing a middle ground between fully automated content creation and manual work. It allows users to think about media at a higher level, filling in gaps and simplifying tasks without reducing their ability to express themselves.

The Paradigm's article on intent-based architectures in cryptographic systems[29] highlights potential risks and challenges in adopting intents, while emphasizing their shift towards a declarative paradigm. Intents allow users to specify "what" they want to achieve rather than "how" to achieve it, offering improved flexibility and efficiency. For example, instead of specifying exact steps like "do A then B, pay exactly C to get X back," an intent might simply state "I want X and I'm willing to pay up to C." This declarative approach enables users to outsource complex execution details to sophisticated third parties, potentially improving user experience and reducing inefficiencies. However, the article also warns of risks such as centralization, trust issues, and opacity in intent-based systems.

The age of software designed for user attention is coming to an end, and in its place, we are entering the era of software summoned on demand to serve user intent – an era where home-cooked software, crafted by and for individuals and communities, will flourish. This shift, powered by AI and embracing principles of end-user programming and local-first design, will put more power and agency back into the hands of users. It will create a more democratic, personalized, and efficient digital landscape, where the boundaries between users and creators blur, and technology becomes an even more seamless extension of human capability.

As we move into this new era, we have the opportunity to build a digital world that is more responsive to individual needs, more respectful of user privacy and agency, and more aligned with the values of local communities. The end of software as we know it marks the beginning of a more empowering, personalized, and human-centered computing experience.

5. Our Blueprint

We are building an intelligence-age operating system using Rust, WebAssembly, and WebGPU. We believe that the future of computing is not limited to your physical hardware. To realize this vision, our core technology includes a compiler that manages the execution of on-device ML models and WebAssembly modules, all packaged within a local-first layer. This system is powered by a local agent that functions as an intent router driven by action models.

The intent router is designed with several key properties:

  1. Multimodality for understanding, allowing it to process and interpret various types of input such as text, voice, and gestures natively. This capability enables the system to comprehend and model complex human-computer interactions across different applications and interfaces, similar to how the Toolformer paper's[30] or Adept's[31] action models learns from demonstrations to replicate user actions.
  2. On-device operation to minimize latency, ensuring quick response times and improved user experience.
  3. On-demand code generation and execution capabilities, similar to OpenAI's code interpreter[32], enabling dynamic problem-solving and task completion.
  4. Utilization of LoRA (Low-Rank Adaptation) adapters[18] for fine-tuning on specific tasks, allowing for efficient customization and specialization of the system.
  5. Advanced tool use capabilities and understanding, with the ability to merge outputs across different mediums. This allows the system to leverage a wide array of tools and seamlessly integrate their results, whether they're in the form of text, images, audio, or other data types.

The fact that it's built on web infrastructure means the OS is hardware-agnostic and inherently cross-platform, enhancing flexibility and compatibility. By exposing these capabilities through developer-friendly Tilefile, CLI, and Runtime APIs, we empower developers to build intelligent cross-platform user experiences that open up new possibilities for human-machine collaboration.

Tiles could leverage action models to interpret user requests and utilize web APIs to perform tasks across different applications. For instance, a user might ask Tiles to "generate our monthly compliance report," and the system would understand how to use existing software like Notion for organizing information, Figma for creating visualizations, and Dropbox for storing and sharing the final document. However, given the complexity and emergent nature of AI systems, it is difficult to accurately predict how these tools will evolve and interact with Tiles in the future. They may need to adapt to the device, operating system, user preferences, or a combination of these factors. As a result, Tiles must be designed with flexibility and adaptability in mind, allowing it to accommodate and integrate with the ever-changing landscape of AI-driven applications and services, while still enabling it to become a universal collaborator for knowledge workers, working hand-in-hand with users and utilizing the same tools they already use.

As Alan Kay detailed in his roadmap for inventing the future[33], Tiles Research is taking a "woowoo" intuition about a future technology, identifying favorable exponentials like Huang's law[34] and LLM improvements, and imagining a concrete version 30 years out. By finding something that would be ridiculous not to have in 30 years, bringing the idea back to 10 years, and building it now, Tiles Research aims to "live in the future" and invent the future from within.

As technology evolves rapidly, people's minds change slowly. To drive adoption and make a significant impact, Tiles Research recognizes that there are two realistic paths: either a huge platform imposes the change and others follow, or a killer app paves the way[35]. Given Tiles Research's position, the focus is on developing a killer app that disguises the infrastructure within.

However, figuring out the infrastructure pieces for this killer app requires deep systems research, which often leads to second-order effects. As highlighted in the "Introducing Speculative Technologies"[36] piece, the history of technology is dominated by such effects. Problems are often solved not by tackling them directly, but by creating powerful technologies that open up new frontiers, break zero-sum games, and create human agency. New frontiers require new tools, and Tiles aims to be that tool for the AI-native era.

To navigate this challenge, Tiles Research is adopting a three-stage approach:

  1. Short-term: Create prototypes and quick demos using existing tools for immediate feedback loops. This stage focuses on cutting back scope to close the feedback loop with reality as fast as possible, recognizing that things that don't get done quickly often don't get done at all.
  2. Medium-term: Develop the killer app with infrastructure disguised inside, built from scratch. This app will incorporate findings from research and prototypes, aiming to drive widespread adoption and showcase Tiles Research's innovative technology. Drawing inspiration from bridge interfaces like the transition from telephones to the internet, as discussed in the "Answer AI Blog Post: Lessons from History's Greatest R&D Labs[37], the app will serve as a bridge to new technological frontiers, rather than a final destination.
  3. Long-term: Leverage the success and user base of the killer app to gradually expand its capabilities, potentially replacing or significantly altering the role of the traditional operating system from within the application itself.

Tiles Research is particularly interested in the following research domains:

Currently, our top priority is developing the ML compiler for fast inference, as we've identified this as the primary bottleneck in realizing our envisioned system.

We draw inspiration for deploying these ideas from projects like Ratchet (opens in a new tab), Luminal (opens in a new tab), CubeCL (opens in a new tab), Ollama (opens in a new tab), Extism (opens in a new tab), Playbit (opens in a new tab), Mistral.rs (opens in a new tab), sqlite-vec (opens in a new tab), Theseus OS (opens in a new tab), Automerge (opens in a new tab), and Fuyu (opens in a new tab). Building upon this rich ecosystem of innovation, we aim to synthesize and extend these breakthroughs in AI and systems engineering. These projects serve as both a foundation and a catalyst, informing our approach as we forge new paths in AI infrastructure and applications.

Throughout this journey, Tiles Research acknowledges that the artifacts we produce may initially seem suboptimal, as systems often experience performance setbacks when escaping local optima. Supporting systems research requires embracing uncertainty and persevering through discouraging benchmarks.

By staying true to its vision, leveraging second-order effects, and learning from historical R&D labs, Tiles Research is poised to revolutionize personal computing in the AI-native era.

6. Conclusion

As we embark on this journey to build Tiles, we find ourselves constantly returning to Vilém Flusser's vision of the future factory—a space where the boundaries between human creativity and machine intelligence blur, where our tools become our collaborators in the pursuit of new possibilities. It's this vision that guides us, that reminds us of the potential for technology to transform not just how we work, but how we think and create.

Our mission is to advance the communication of human intent with machines, to design more natural ways of working. We envision an era of AI-native computing, where powerful yet accessible tools adapt to our needs and augment our potential. Tiles is our moonshot attempt to create the operating system for this new paradigm.

However, we approach this ambitious project with the humility of beginners. We're reminded of John Cleese's insight: "The most dangerous thought you can have as a creative person is to think you know what you're doing." At Tiles Research, we embrace uncertainty. We cultivate an openness to new possibilities and unconventional approaches. This willingness to question our assumptions and explore uncharted territories is what enables us to perceive what others might overlook. Our strength lies not in presumed knowledge, but in our readiness to learn, adapt, and reimagine.

There is so much more to share about the technical details of our work, but that will have to wait for upcoming demos and blog posts. For now, we simply want to say thank you for joining us at the start of this journey into uncharted territory.

We are under no illusions that this will be an easy path. Inventing the future is not a task for the faint of heart. There will be setbacks, unexpected challenges, and moments of doubt. But that is the price of attempting something truly new. Our strategy is to move fast, to try many things, to let the reality of use cases guide us. Gradually, bit by bit, we will find our way.

If, through our efforts, we can make even the smallest dent in the universe of human-computer interaction, we will consider it a resounding success. Because even the smallest changes, when compounded through network effects and the passage of time, can transform the world. As Gifford Pinchot wisely noted, "The vast possibilities of our great future will become realities only if we make ourselves responsible for that future." We take this responsibility seriously, understanding that our work today shapes the technological landscape of tomorrow.

We hope you'll stay tuned, through our blog posts and open-source repos. The factory of the future is under construction, and we're all building it together.

7. References

[1] The Factory. (1991). Flusser, V. Third Rail Quarterly. Link (opens in a new tab)

[2] Seeing Spaces. (2014). Victor, B. Vimeo. Link (opens in a new tab)

[3] Tinker. (2024). Ngo, R. Asimov Press. Link (opens in a new tab)

[4] Extending Ourselves: Computational Science, Empiricism, and Scientific Method. (2004). Humphreys, P. Oxford University Press. Link (opens in a new tab)

[5] AlphaGo - The Movie. (2020). Google DeepMind, YouTube. Link (opens in a new tab)

[6] Introducing ChatGPT. (2022). OpenAI. Link (opens in a new tab)

[7] Attention Is All You Need. (2017). Vaswani, A., et al. Advances in Neural Information Processing Systems, 30. Link (opens in a new tab)

[8] What is interpretability? (2024). Anthropic, YouTube. Link (opens in a new tab)

[9] Extracting Concepts from GPT-4. (2024). OpenAI. Link (opens in a new tab)

[10] Scaling Monosemanticity. (2024). Anthropic. Link (opens in a new tab)

[11] Software 2.0. (2017). Karpathy, A. Medium. Link (opens in a new tab)

[12] Apple introduces M4 chip. (2024). Apple. Link (opens in a new tab)

[13] OpenELM: An Efficient Language Model Family with Open Training and Inference Framework. (2024). Apple Machine Learning Research. Link (opens in a new tab)

[14] Predicting GPU Performance. (2022). EpochAI. Link (opens in a new tab)

[15] Trends in Machine Learning Hardware. (2023). EpochAI. Link (opens in a new tab)

[16] The future of AI is hybrid - Part I: Unlocking the generative AI future with on-device and hybrid AI. (2023). Qualcomm. Link (opens in a new tab)

[17] Apple Intelligence. (2024). Evans, B. Benedict Evans. Link (opens in a new tab)

[18] LoRA: Low-Rank Adaptation of Large Language Models. Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., & Chen, W. (2021). arXiv. Link (opens in a new tab)

[19] Home-Cooked Software. (2024). Appleton, M. Maggie Appleton. Link (opens in a new tab)

[20] Local-First Software: You Own Your Data, in Spite of the Cloud. (2019). Kleppmann, M., Beresford, A. R., & Svensso, S. Ink & Switch. Link (opens in a new tab)

[21] End-User Programming. (2023). Ink & Switch. Link (opens in a new tab)

[22] An Interview with Daniel Gross and Nat Friedman about Apple and AI. (2024). Thompson, B. Stratechery. Link (opens in a new tab)

[23] On AI Requiring a New OS. (2024). Sinofsky, S. Hardcore Software. Link (opens in a new tab)

[24] Private Cloud Compute. (2024). Apple Security Research. Link (opens in a new tab)

[25] Why Anthropic's Artifacts may be this year's most important AI feature: Unveiling the interface battle. (2024). Venturebeat. Link (opens in a new tab)

[26] You and your friends can now share and remix your favorite conversations with the Claude AI chatbot. (2024). Techradar. Link (opens in a new tab)

[27] AI Interfaces. (2023). Silicon Jungle. Origami Party. Link (opens in a new tab)

[28] Folk Practices & Collaboration. (2024). Shapeshift Labs. Link (opens in a new tab)

[29] Intent-Based Architectures and Their Risks. (2023). Paradigm. Link (opens in a new tab)

[30] Toolformer: Language Models Can Teach Themselves to Use Tools. (2023). arXiv. Link (opens in a new tab)

[31] ACT-1: Transformer for Actions. (2022). Adept. Link (opens in a new tab)

[32] OpenAI launches API that lets developers build 'assistants' into their apps. (2023). TechCrunch. Link (opens in a new tab)

[33] Alan Kay - Inventing The Future. (2014). Kay, A. YouTube. Link (opens in a new tab)

[34] Huang's law. (2018). Wikipedia. Link (opens in a new tab)

[35] The Myth of The Infrastructure Phase. (2018). Grant, D., & Grossman, N. Union Square Ventures. Link (opens in a new tab)

[36] Introducing Speculative Technologies. (2023). Speculative Technologies. Link (opens in a new tab)

[37] Lessons from History's Greatest R&D Labs. (2024). Answer AI. Link (opens in a new tab)

© Ankesh Bharti and contributors. Except as otherwise noted, the content of this website is licensed under the CC BY 4.0 License, and code samples are licensed under the Apache 2.0 License.