Winning the Next AI Battle

Key Concepts

• Local Language LLMs: Large Language Models tailored for specific languages and dialects, addressing limitations of general models in cultural nuance and grammar.
• AI Sovereignty: A nation's ability to control its AI development, deployment, and data, encompassing chip fabrication, infrastructure, and model layers.
• Attribution and Compensation: Mechanisms to track content origin and ensure creators are compensated when their work is used in AI models.
• Regulation and Governance: The framework of laws and policies governing AI development and deployment, balancing innovation with protection.
• Open Source Ecosystem: The collaborative development and sharing of AI models, impacting geopolitical competition and innovation.
• Disintermediation: The removal of intermediaries in a supply chain, particularly relevant to content creators losing direct customer relationships due to AI.
• Reasoning Models and Reinforcement Learning: Advanced AI paradigms enabling AI to tackle mission-critical problems with high reliability, opening new investment opportunities.

Arabic Language Models and Localized AI

Norah highlights the critical need for localized Arabic language Large Language Models (LLMs). She explains that Arabic is not a monolithic language but comprises numerous dialects. While Arabic is the fourth most spoken language online, it constitutes only about 4% of internet data. High-quality Arabic datasets are scarce and often proprietary, residing within companies. Existing global LLMs, while advanced, struggle with cultural nuances and specific grammatical structures in Arabic, necessitating human intervention for correction. This underscores the importance of building and structuring high-quality datasets, often involving human-in-the-loop processes, to improve these models. Tarama's focus is on solving complex Arabic-specific problems, including data extraction from archived sources using Optical Character Recognition (OCR), which itself requires custom data creation for model training.

AI Sovereignty: Achievability and Layers

The discussion broadens to AI sovereignty, with Ana and Bill weighing in on its feasibility. Ana breaks down AI sovereignty into layers: the chip stack, infrastructure, and the model layer, with agents built on top. She argues that achieving sovereignty in frontier pre-training (developing cutting-edge models from scratch) is extremely difficult due to the dominance of a few hyperscalers and model companies. However, the open-source ecosystem, significantly boosted by China's DeepSeek, has made the model layer more accessible. Nations can now leverage base models and perform last-mile customization for local values and cultural norms.

The chip stack, however, remains a significant hurdle. Countries lacking the capability to develop advanced chip fabrication supply chains (like 2nmter chips, currently dominated by only two countries) are better off partnering. Ana emphasizes that sovereignty is achievable at different layers, with the foundation model layer showing the most significant recent progress. Countries investing in their own compute infrastructure, owning and hosting chips, and running models in ways that align with their values are moving towards a more sovereign AI future, which she deems "non-negotiable."

Bill concurs that AI sovereignty is achievable, driven by the strong desire to preserve culture. He views the current era as a civilizational turning point, akin to BC/AD, where humanity's knowledge is being absorbed. Preserving cultural components within sovereign AI is crucial. His personal goal is to ensure fairness and proper attribution in AI.

Attribution and Compensation in AI

Bill elaborates on his work at Pera, addressing the pressing issue of content attribution and compensation for creators whose work is used to train AI models. He envisions a system where any individual can contribute content (text, music, designs) to AI, and when that content is used, proper attribution and compensation are provided. This was previously impossible but is now achievable with AI's ability to trace output components back to their origins.

Pera has developed technology to "unscramble the egg" of AI output and provide attribution. The challenge lies in convincing AI companies to adopt this. While small AI companies are on board, progress with larger ones is ongoing. Lawsuits, like the $1.5 billion settlement between Anthropic and book publishers, are driving change. However, Bill believes that the ultimate driver will be competition: AI systems that use content fairly will produce better answers and thus gain a competitive advantage. He clarifies that this process does not require access to model weights or activations, as it can be determined by analyzing the output alone. This is critical because AI is becoming the primary interface for search, shopping, and discovery, making fairness paramount.

Regulation and Governance of AI

The conversation shifts to regulation and governance. Norah suggests two regulatory approaches: the US/Europe model, which she feels might be "overprotected" and could stall innovation (citing GDPR as an example), and a balanced approach that fosters innovation while providing some regulation. Key areas for protection include copyright for publishers, writers, and contributors, and governance of API usage to ensure data protection. She expresses concern that larger players might "distill off" proprietary models, claiming capabilities they didn't develop. Deploying models on-premises for clients concerned about data sovereignty also raises IP protection issues, as the model itself could be reverse-engineered, a concern not yet adequately governed.

Ana discusses the shift in global AI governance, referencing the Paris AI Action Summit. She notes a move from a primary focus on AI safety to regulation that favors innovation and growth. Previously, the US faced a chaotic landscape of state-by-state legislation, making it difficult for entrepreneurs. The Paris summit, occurring after the release of DeepSeek, marked a turning point. The US administration's AI Action Plan has provided a unified federal framework, which is also influencing Europe's approach to unifying its own regulations.

However, Ana expresses concern about the geopolitical implications of the accelerated open-source ecosystem, largely driven by China. She states that currently, the most powerful open-source models outside of Mistral (France) and a few specialized US models are from China, which is not an encouraging picture for the US and its allies. She acknowledges that China's actions have spurred competition, leading to US competitors "upping their game." She believes that entrepreneurs and researchers should focus on pushing the frontier of capabilities rather than navigating complex legislation. The current environment, with a more unified federal approach, is freeing up this "mind space," and she anticipates a wave of open-weight models from Western labs in the coming months to "level set."

The Future of Content Creation and Commerce in AI

Bill addresses the fundamental challenge for publishers and content creators: disintermediation and the loss of direct customer relationships, even with potential compensation from AI usage. He draws parallels to Spotify, YouTube, and the App Store, which share revenue and pay billions to creators. Apple News also pays billions to publishers, demonstrating that a revenue-sharing model can be successful. Spotify, with $20 billion in revenue and $12 billion paid out to artists, is a profitable example of this model.

Regarding disintermediation, Bill points to the emerging trend of direct commerce within AI products, citing OpenAI's integration with Walmart. While Walmart wants to be present where conversations happen, Amazon is hesitant due to its need to maintain its existing experience. Bill believes AI will become the "storefront" for human exploration and knowledge, making it crucial for businesses to be present to avoid missing out.

Investment Opportunities in the AI Stack

Ana identifies a "golden age" and an "explosion of new frontier teams" in AI investment. She refutes the earlier notion that only a few large labs would dominate AI training. The advent of reasoning models and the success of reinforcement learning have changed the game. A year ago, frontier generative AI excelled at creative writing and companionship, where hallucination was acceptable. However, for mission-critical problems in defense, healthcare, and enterprise, where high reliability is essential, these models failed.

Now, with reinforcement learning, if the reward model is correctly defined, startups can build entirely new multi-billion dollar companies by embedding themselves deeply within industries, going vertical, and understanding customer problems end-to-end. This represents a significant new area for investment.