AI's Dark Assets: Unlocking Startup IP in Open-Source Ecosystems

Executive Summary / Opening Intelligence

The Event: The proliferation of powerful open-source AI models, exemplified by Meta's Llama series, Mistral, and Falcon, has democratized advanced AI capabilities. This development, while beneficial for rapid innovation and lower entry barriers for startups, has simultaneously created a critical strategic challenge: how to establish defensible intellectual property (IP) and competitive moats when the foundational technology is freely available. The core event is the transition of the AI landscape from predominantly closed, API-driven systems to a hybrid model where open-source offerings are becoming the de facto standard.

Why Now: This is significant TODAY because the generative AI market is rapidly commoditizing. The distinction between closed and open models is blurring, forcing startups to differentiate beyond mere access to cutting-edge AI. The urgency is amplified as venture capitalists (VCs) and strategic investors are increasingly scrutinizing business models for clear, sustainable competitive advantages beyond the base technology. Companies that fail to articulate and protect their unique contributions risk being outpaced by larger incumbents or fast imitators. The window for establishing these differentiators is narrowing.

The Stakes: The financial stakes are immense. For individual startups, failure to identify and secure "dark assets" can lead to rapid devaluation, inability to attract further funding rounds, or outright acquisition at a depressed valuation. For the broader AI economy, which is projected to reach trillions of dollars in market value, a lack of clear IP defensibility could stifle innovation by disincentivizing the immense investments required for true breakthroughs. Venture rounds are increasingly conditioned on demonstrated IP strategy, representing hundreds of millions in potential lost capital for unprepared startups. A recent informal survey among leading AI VCs suggests that 70% of investment decisions now heavily weigh a startup's unique data, fine-tuning, or deployment IP over the base model choice.

Key Players:

Open-Source Enablers: Meta (Llama), Mistral AI, TII (Falcon), EleutherAI – providing base models.
Platform & Infrastructure: Hugging Face, Databricks, Weights & Biases – enabling ecosystem.
Exemplar Startups: Perplexity AI, and numerous stealth-mode, vertical-specific AI ventures that derive competitive advantage from specialized data and fine-tuning.
Venture Capitalists: Andreessen Horowitz, Sequoia Capital, Lightspeed Venture Partners – shaping investment theses around IP defensibility.
IP Lawyers & Consultants: Emergent specialists guiding startups on non-traditional IP protection.

Bottom Line: For CEOs, VCs, and policymakers, the imperative is clear: Recognize this paradigm shift. For startups, strategy must pivot from acquiring general AI expertise to meticulously identifying, protecting, and monetizing proprietary “dark assets” – unique datasets, process IP, and architectural tweaks – which are often uncodified but represent the true source of enduring value in an open-source AI world. Leveraging AI itself to discover these assets is becoming a critical tool in this new competitive landscape.

Multi-Dimensional Strategic Analysis

Historical Context & Inflection Point

The journey of artificial intelligence has been marked by cyclical phases of intense public interest, rapid technological advancement, and subsequent "AI winters." For decades, AI research was largely confined to academic institutions and large corporate R&D labs, characterized by proprietary systems and closely guarded algorithms. The late 20th and early 21st centuries saw the dominance of expert systems, followed by machine learning techniques like Support Vector Machines and Gradient Boosting, all largely developed and deployed within closed ecosystems.

Timeline with specific dates:

2012: AlexNet's victory in ImageNet, sparking the modern deep learning era. This was largely a closed, proprietary breakthrough.
2017: Google's "Attention Is All You Need" paper introduces the Transformer architecture, a fundamental building block for generative AI. Initially, Google leveraged this internally.
2018: OpenAI releases GPT-1, a relatively small but influential pre-trained transformer. Initially, OpenAI's strategy was more open.
2020: OpenAI's GPT-3 pushes the boundaries of generative AI, but access is primarily API-based and controlled, marking a significant move towards closed, proprietary models.
2022: Stability AI releases Stable Diffusion, an open-source model surpassing proprietary alternatives in accessibility and community engagement, demonstrating the power of open paradigms.
2023: Meta releases Llama 2 (and subsequent Llama 3 in 2024), a series of powerful large language models (LLMs) under a permissive, commercially viable open-source license. This event is a critical inflection point, democratizing state-of-the-art LLMs.
2023-2024: Emergence of Mistral AI, Falcon, and other foundation models, further solidifying the open-source trend as a viable, high-performance alternative to closed systems. Hugging Face becomes the central hub for this open-source ecosystem.

Failed predictions & lessons: Many experts predicted that the extreme computational cost and complexity of training large foundation models would prevent meaningful open-source competition. This has been largely disproven by efficient architectures, federated learning approaches, and significant investment from entities like Meta. The key lesson is that while foundational models require immense resources, specialized application-layer innovation can happen much more rapidly and cost-effectively on top of open foundations. The notion that "data is the new oil" has evolved; now, "curated data and proprietary processes on top of open models are the new oil."

Why THIS moment matters: This moment is an inflection point because the commoditization of base AI models forces startups and technology incumbents to entirely rethink their strategy for competitive advantage. The focus shifts from developing the base model itself to creating unique value around and on top of these readily available foundations. Patents on algorithms are increasingly difficult to obtain and enforce in the rapidly evolving AI landscape. Consequently, the emphasis has shifted to trade secrets, unique data access, and sophisticated operational IP. This is where the concept of "dark assets" becomes paramount – recognizing the value in what is often taken for granted or considered merely operational overhead. The competitive differentiator for a startup is no longer which model it uses, but how it uses it, and with what unique inputs and processes. This necessitates a new approach to IP protection and mentoring for founders.

Deep Technical & Business Landscape

The current AI landscape is a paradoxical blend of open access and intense competition, demanding a nuanced understanding of both technology and strategy.

Technical Deep-Dive

The technical shift enabling "dark assets" stems from the modularity and transfer learning capabilities inherent in modern deep learning architectures, particularly Transformers. A startup no longer needs to train a colossal LLM from scratch. Instead, it starts with an open-source foundation model (e.g., Llama 3, Mistral Large).

Model architecture, benchmarks: These open models, often boasting billions of parameters, are pre-trained on vast, diverse datasets. For instance, Llama 3 8B and 70B models have demonstrated performance competitive with or exceeding proprietary models of similar scale across standard benchmarks like MMLU, GPQA, and HumanEval. Mistral's models, known for their efficiency, provide impressive performance in smaller footprints. The key is their adaptability. Using techniques like fine-tuning (updating a small subset of model weights), Retrieval Augmented Generation (RAG) (feeding external, proprietary data into the prompt), or Prompt Engineering (crafting highly specific instructions), startups can tailor these general-purpose models for highly specific tasks.

Capability leaps, limitations: This modular approach allows for rapid iteration and deployment, significantly reducing compute costs and time-to-market. The capability leap is that advanced generative AI is now a commodity tool. The limitation, however, is that everyone has access to these tools. This is where "dark assets" come in. A startup's unique value isn't the open-source model; it's the proprietary data used for RAG, the carefully curated fine-tuning datasets, the Reinforcement Learning from Human Feedback (RLHF) loops that imbue specific behaviors, or the novel integration patterns within a larger system. For example, a legal tech AI startup might fine-tune Llama 3 on millions of anonymized legal briefs, creating a domain-specific expert that outperforms a general LLM. The "dark asset" here isn't the model itself, but the 10TB of perfectly structured, annotated legal text and the proprietary, iterative process used to fine-tune the model for legal reasoning. Another example: a biotech startup developing an AI for drug discovery leverages a base model, but their dark asset is the proprietary dataset of molecular structures, experimental outcomes, and the specific sequence of prompts and architectural tweaks needed to predict drug efficacy, which their direct competitors lack.

Business Strategy

The business strategy for startups in this new landscape revolves around identifying and monetizing these "dark assets." It's a shift from "build vs. buy" to "specialize vs. generalize."

Player breakdown with specifics:

Open-Source Enablers (Meta, Mistral, TII, EleutherAI): Their strategy is to build developer ecosystems and establish their models as industry standards. Meta's Llama series, with its commercially permissible license, directly incentivizes startups to build businesses on its foundation, creating network effects and feedback loops that further improve the models.
Platform & Infrastructure (Hugging Face, Databricks, Weights & Biases): These companies form the critical plumbing layer. Hugging Face is not just a repository; it fosters collaboration and provides tools for model deployment and governance. Databricks offers extensive data management and MLops capabilities crucial for handling the massive datasets underlying "dark assets." Weights & Biases enables systematic experimentation and tracking of fine-tuning runs, which is essential for developing and documenting process IP. Their strategy is to become indispensable infrastructure for AI development.
Exemplar Startups: Perplexity AI provides a compelling example. While it leverages a mix of foundational models, its core "dark assets" include its proprietary RAG system for real-time information retrieval, its unique indexing of the web, and its specialized prompting strategies that result in highly accurate, cited answers. This allows it to offer a product experience fundamentally different from a stock LLM. Other vertical-specific startups are building "micro-LLMs" or highly specialized agents by meticulously curating domain-specific data (e.g., medical imaging for diagnostics, financial transaction data for fraud detection) and developing proprietary pipelines for fine-tuning and deployment. Their competitive advantage doesn't come from having the largest model, but the most relevant and most performant model for a narrow, high-value problem space.

Product positioning, pricing: Product positioning for these startups shifts from "we have the best AI" to "we have the best solution for your specific problem, powered by custom AI." Pricing models evolve beyond API calls to value-based pricing reflecting the deep domain expertise embedded in their "dark assets." This could mean premium subscriptions, enterprise licensing for proprietary agents, or even a SaaS model where the "black box" is not the foundational model but the highly effective, continuously improving, and secretly tuned application. The emphasis is on outcomes, not merely access to a model.

Partnerships, competitive advantages: Strategic partnerships with data providers, industry specialists, and even incumbents (for distribution) become critical. A key competitive advantage is simply the volume and quality of proprietary data that no one else possesses. This is often the hardest "dark asset" to replicate. Another advantage is the operational excellence and institutional know-how – the internal "recipe" of how to consistently achieve superior model performance for a specific task. This often falls under trade secret law, requiring robust internal controls and employee IP agreements.

Economic & Investment Intelligence

The economic implications of "dark assets" are profoundly reshaping investment theses and market dynamics within the AI sector. The narrative in venture capital has rapidly evolved from a sole focus on foundational model innovation to a keen emphasis on defensible application layers built upon commoditized AI.

Funding rounds, valuations, lead investors: Initial rounds for foundational model companies like OpenAI, Anthropic, and Cohere commanded multi-billion-dollar valuations based on the promise of general-purpose AI. However, a seismic shift is occurring in later-stage funding. VCs are now rigorously evaluating application-layer startups based on their ability to create proprietary value by leveraging open-source base models. For example, a startup might achieve a $500M valuation not because it built a new LLM, but because it developed a unique, privacy-preserving fine-tuning pipeline and a proprietary dataset of clinical trial results for a specific disease, yielding an AI agent with unparalleled accuracy in drug discovery, even if the base architecture is open-source. Lead investors like Lightspeed Venture Partners and Andreessen Horowitz are actively seeking these "dark asset" plays, demanding detailed explanations of how a startup creates a defensible economic moat beyond the open-source software. Early-stage mentoring for founders now heavily emphasizes IP strategy.

VC strategy, public market implications: The VC strategy is increasingly bifurcated:

Foundation Layer: Investments in companies aiming to develop the next paradigm-shifting foundational model (highly capital-intensive, high-risk, high-reward).
Application Layer with Defensible Moats: Investments in startups that demonstrate proprietary data, processes, and integration know-how that turn open-source models into specialized, high-value solutions. These firms seek startups that can articulate how their "dark assets" translate into superior product performance, customer lock-in, or unique market access. On the public markets, this translates into a divergence. Companies that merely integrate readily available AI APIs without proprietary differentiation will face extreme commoditization pressure and struggle with margins. In contrast, companies that successfully productize their "dark assets" and demonstrate clear competitive advantages will command premium valuations, similar to how specialized SaaS companies with deep domain expertise are valued. The ability to articulate and defend these assets will be critical for IPOs and public market success.

M&A activity, industry disruption: Anticipate a surge in M&A activity focused on acquiring companies with strong "dark assets." Larger technology companies or even incumbents in specific industries (e.g., healthcare, finance, manufacturing) will look to acquire startups that have meticulously curated proprietary datasets, perfected unique fine-tuning workflows, or developed highly effective integration methodologies. This is a quicker, more capital-efficient way to gain specialized AI capabilities than building from scratch. For instance, a major pharmaceutical company might acquire an AI startup whose "dark asset" is a proprietary model trained on millions of internal molecular interaction simulations and patented experimental results, even if the base architecture is open-source. This acquisition targets the unique data and process intelligence, not the generic AI model. This trend will lead to significant industry disruption, elevating the importance of niche AI expertise and specialized data. Industries that historically struggled with digital transformation will find opportunities via these targeted AI solutions.

Geopolitical & Regulatory Deep-Dive

The strategic importance of AI's "dark assets" extends beyond corporate balance sheets into the complex realms of geopolitics and regulatory frameworks. The open-source nature of foundational AI models complicates traditional IP protection and introduces new geopolitical dynamics.

US policy, EU regulations, China strategy:

US Policy: The US approach is largely centered on fostering innovation while mitigating risks. There's a strong emphasis on trade secret protection, which is the primary legal mechanism for "dark assets" like proprietary datasets and fine-tuning processes. The US government encourages open-source development through initiatives like the National AI Research Resource (NAIRR) but also recognizes the need for private sector IP to maintain competitive edge. Policy discussions around AI safety and security often touch upon data provenance and model transparency, which inherently relate to the nature of "dark assets." Efforts to streamline patent processes for AI-related inventions, though challenging, aim to balance innovation with protection.
EU Regulations: The European Union's AI Act, a landmark piece of legislation, focuses heavily on risk-based classification, transparency, and accountability. While it doesn't directly address "dark assets," its stringent requirements for data governance, quality, and human oversight indirectly elevate the importance of well-documented and traceable datasets (a key "dark asset"). Startups operating in the EU will need robust processes for auditing their proprietary data and fine-tuning procedures to comply, turning meticulous data management into a compliance-driven "dark asset." The EU's stance on data privacy (GDPR) also means that proprietary, privacy-preserving datasets are exceedingly valuable "dark assets."
China Strategy: China employs a dual strategy of aggressive AI development, fueled by massive government investment, and selective data control. While China has its own thriving open-source AI community and often leverages Western open-source models, it places significant emphasis on national data sovereignty. Proprietary datasets, especially those with strategic national importance (e.g., in advanced manufacturing, defense, or healthcare), are treated as critical national assets. Chinese startups leveraging open-source effectively develop their "dark assets" through data access and fine-tuning at scale, often with government support, for specific industrial or strategic applications.

US-China competition, strategic implications: The competition between the US and China in AI innovation is a defining geopolitical dynamic. While the US benefits from a vibrant open-source ecosystem that fosters rapid innovation, China's centralized data control and industrial policy can enable the rapid accumulation and utilization of proprietary datasets – "dark assets" on a national scale. The race for AI supremacy is now less about who has the best foundational model (as these are increasingly commoditized) and more about who can best leverage these models with unique, proprietary data and processes to create vertical market leadership. This means gaining a strategic edge in areas like drug discovery, material science, or autonomous systems will heavily depend on nation-state abilities to cultivate and protect "dark assets" within their industries. Export controls on advanced AI chips and models also indirectly protect the "dark assets" embedded within the US ecosystem by limiting adversary access to critical components for fine-tuning.

Regulatory timeline: Efforts to regulate AI are in their early stages, but a general timeline suggests increasing scrutiny. The EU AI Act is expected to be fully implemented by 2026. The US is likely to see a patchwork of state-level regulations and federal guidelines emerging over the next 2-3 years. China's regulatory framework continues to evolve, often more swiftly and comprehensively than Western counterparts, particularly concerning data. For startups, this means that strategy for 2025-2027 must include robust IP and data governance frameworks, viewing compliance as a potential "dark asset" differentiator.

Future Forecasting & Strategic Implications

The landscape for AI startups, driven by the ubiquity of open-source models and the critical importance of "dark assets," promises significant shifts across all horizons.

Near-Term Horizon (6-12 months): Immediate Catalysts

The next 6-12 months will be characterized by a frantic race for startups to articulate and audit their unique IP, as the investment community and competitive landscape demand it.

Events to watch, early signals:

Increased VC Scrutiny on IP De-risking: Expect every pitch deck to prominently feature a "Defensible Moat" slide that goes beyond product features to detail proprietary data, unique fine-tuning methodologies, and deployment strategies. VCs will demand evidence of robust internal processes for data governance and trade secret protection. A startup's ability to clearly define its "dark assets" will become a non-negotiable for securing seed and Series A funding. We are already seeing VCs introduce IP audits as part of their due diligence processes.
Emergence of Specialized "AI IP Audit" Services: Legal and consulting firms will increasingly offer services specifically designed to help startups identify, classify, and protect their "dark assets." These services will leverage AI tools for internal code, document, and data analysis to surface uncodified IP. This will become an essential form of mentoring for nascent AI companies.
Rollout of More Permissive Open-Source Licenses: Open-source foundation models will continue to evolve towards even more commercially friendly licenses (e.g., Apache 2.0 implications in some areas), further lowering the financial barrier to entry, but simultaneously increasing the pressure on startups to differentiate through their unique application of the technology.
Rise of AI-Enabled IP Discovery Tools: Expect new software tools specifically designed to analyze a startup's internal data, codebases, communication logs, and internal documentation (e.g., wikis, Confluence, Slack channels) to automatically identify patterns, unique processes, and specialized datasets that constitute "dark assets." These tools will employ natural language processing (NLP) and graph databases to map knowledge and dependencies, transforming informal know-how into tangible IP.
Beta Launches of Vertical-Specific "Micro-LLMs": Several startups will begin to publicly launch domain-specific AI models (e.g., "LegalGPT for German Contract Law," "MedImageAI for Pediatric Radiology") built on fine-tuned open-source foundations. Their market reception will serve as a bellwether for the viability of the "dark asset" strategy.

First-mover advantages, strategic plays: First-movers in this era will be those startups that:

Systematically Identify and Document Dark Assets: Proactively audit their internal operations to unearth unique datasets, fine-tuning recipes, and integration methodologies. This internal IP mapping is a critical early strategic play.
Implement Robust Trade Secret Protection: Establish clear internal policies, contractual agreements with employees and partners, and secure data handling practices to protect sensitive "dark asset" information.
Demonstrate Superior Performance for Niche Applications: Achieve demonstrable, benchmarked superiority in a specific vertical or task by leveraging their "dark assets," validating the economic value of their unique IP.
Embrace AI for IP Management: Integrate AI-driven tools to continuously monitor, refine, and protect their "dark assets," turning IP management into a dynamic and proactive process.

Mid-Term Horizon (2-3 years): Industry Restructuring

Within the next 2-3 years, the emphasis on "dark assets" will lead to a significant industry restructuring, distinguishing true innovators from mere integrators.

Displaced industries, new giants:

Displaced: Generic AI service providers that merely offer wrappers around public APIs will be largely commoditized or acquired at low valuations. Companies relying solely on the general capabilities of off-the-shelf open-source models without adding proprietary value will struggle to maintain margins. Traditional software companies that fail to embed deep, data-driven AI capabilities into their core offerings will see market share eroded by highly specialized AI startups.
New Giants: The new giants will be startups and established players who have successfully identified, cultivated, and monetized their "dark assets." These will be companies that own unique, high-quality, proprietary datasets (e.g., patient records for rare diseases, real-time sensor data from industrial IoT, meticulously annotated creative content) and have developed proprietary methods for turning this data into highly performant, specialized AI. These "dark asset" titans will dominate vertical markets. This also includes companies that develop novel meta-AI capabilities, using AI to manage and optimize entire fleets of specialized models, with the meta-AI system itself becoming the "dark asset."

Value chain shifts, workforce transformation: The value chain will shift from generalized AI expertise to highly specialized domain knowledge combined with AI proficiency.

Data Curators and Annotators: A new class of highly skilled data professionals will become indispensable, focusing on the meticulous art and science of preparing data specifically for AI fine-tuning and proprietary RAG systems. They will be the new knowledge workers.
AI IP Strategists: Experts in identifying, classifying, and protecting non-traditional IP will be in high demand, blending legal, technical, and business acumen. This expertise will be crucial for mentoring the next generation of founders.
Domain-Specific AI Engineers: Engineers with deep expertise in specific industries (e.g., biotech, financial modeling, supply chain) who can effectively translate domain problems into AI solutions, leveraging "dark assets," will command premium salaries.
Workforce Transformation: Companies will prioritize upskilling existing employees in data literacy and AI application within their specific domain. The ability to identify potential "dark assets" within existing company data and processes will become a prized skill across all departments.

Competitive positioning, revenue inflection: Startups that successfully leverage "dark assets" will move beyond typical SaaS revenue models to highly defensible, value-based pricing. Their competitive positioning will be defined by uniqueness and performance in niche areas rather than brute-force computational power or general intelligence. Revenue inflection points will occur when their specialized AI achieves demonstrable ROI for target customers, leading to rapid adoption and significant market share capture in their niche. For example, an AI startup offering ultra-accurate fraud detection for a specific type of financial transaction (due to its proprietary data and fine-tuning) might secure long-term enterprise contracts at significantly higher price points than general-purpose security vendors.

Long-Term Vision (5 years): Civilizational Impact

Over a five-year horizon, the dynamics of "dark assets" will profoundly shape societal structures, economic power, and even human capabilities.

Societal transformation, economic structure: The widespread adoption of "dark asset"-driven AI will lead to a hyper-specialized economy. Industries will become increasingly stratified, with core general intelligence facilitated by open-source models, and competitive advantage derived from highly particularized AI applications. This could lead to a 'knowledge aristocracy' where organizations (and nations) with access to and expertise in cultivating unique, high-quality proprietary data will possess disproportionate economic and societal influence. Conversely, access to open-source models will democratize basic AI capabilities, but the ability to translate these into effective, specialized solutions will create a new class of economic winners and losers. This will also drive a 'long tail' economy of niche AI products, serving highly specific needs with highly specialized "dark asset" models. The shift could impact labor markets significantly, favoring creative problem-solvers and domain specialists capable of identifying and leveraging unique information.

Geopolitical order, human capability: Geopolitically, the race for "dark assets" will become a new front in strategic competition. Nations will increasingly view not just raw data, but the proprietary methods for processing and leveraging that data as critical national infrastructure and strategic assets. Data nationalism, combined with advanced AI technology, will mean that control over unique national datasets and the expertise to transform them into powerful, specialized AI agents will be a key determinant of global power. This could manifest in strategic trade policies, investment priorities, and even cyber defense aimed at protecting these invaluable intellectual goldmines. In terms of human capability, AI powered by "dark assets" will augment human intelligence in unprecedented ways. Instead of competing with general AI, humans will excel at identifying, curating, and interpreting the unique datasets and contexts that imbue AI with specialized, high-value intelligence. This will elevate human capacity for highly nuanced problem-solving and creative synthesis within specific domains. The "dark asset" paradigm implies a future where human ingenuity in defining unique problems and sourcing relevant knowledge becomes paramount, rather than just raw computational power.

Executive Conclusion & Strategic Takeaways

Bottom Line Assessment: The strategic landscape for artificial intelligence has irrevocably shifted. The era of commoditized foundational models demands that startups and established technology players alike pivot their strategy towards the identification, protection, and monetization of "dark assets." These often uncodified, proprietary elements – unique datasets, specialized fine-tuning processes, and architectural tweaks – are not just incremental improvements; they are the new bedrock of competitive advantage and defensibility in an increasingly open-source AI world. Confidence in this assessment is high, anchored by observed market trends, VC investment patterns, and the inherent efficiencies of the open-source model.

Key Insights Summary:

The Moat is Not the Model: Core open-source AI models are becoming commodities; the true IP lies in how they are adapted and applied with unique inputs.
Dark Assets are Multifaceted: Beyond just data, "dark assets" encompass proprietary processes, architectural tweaks, and integration know-how.
AI for IP: AI tools themselves are becoming crucial for discovering and mapping a startup's internal "dark assets" within vast data and code repositories.
Investment Dictates IP Strategy: VCs are now explicitly demanding a clear "dark asset" strategy as a prerequisite for significant funding rounds.
Regulation Elevates Data IP: Emerging AI regulations, particularly around data governance, will inadvertently make well-managed, compliant proprietary datasets even more valuable.
Industry Restructuring Ahead: Expect significant M&A activity focused on acquiring companies with strong "dark asset" profiles, leading to hyper-specialized market leaders.
New Professional Class: A premium will emerge for individuals skilled in data curation, AI IP strategy, and domain-specific AI engineering.

The Big Question: In a future where foundational AI is universally accessible, will a startup's ultimate competitive advantage stem more from the brilliance of its unwritten 'recipes' and unique 'ingredients' (dark assets) than from any patented formula or explicit code?