AI's Future: Human Will, Machine Judgment

The mission at Thinking Machines is clear: build AI that augments human capabilities. In an era where AI's power grows daily, the critical task of directing its actions remains firmly with us, individuals, organizations, and humanity as a whole.

These decisions demand knowledge and judgment, qualities honed through continuous interaction with work, increasingly performed alongside AI. Shaping advanced intelligence requires a dynamic process of feedback, learning, and realignment, not a static, one-time directive.

Most current AI models are trained in isolated environments and then deployed without further adaptation. They fail to learn from the people they serve or the tasks they perform. Thinking Machines is charting a different course, aiming for AI as diverse and distributed as the people themselves.

The Intelligence of Knowledge

AI's purpose is to serve human endeavors, which are driven by tacit knowledge, the know-how of chefs, shopkeepers, and countless others. This knowledge is local, fluid, and constantly updated through experience, a stark contrast to static databases.

The inherent dispersion of this knowledge is a collective strength, fostering variety, adaptability, and resilience. It mirrors the success of free markets over planned economies, where decentralized knowledge proves superior to centralized intelligence.

While AI excels in domains with static, expressible goals like chess or mathematics, it falters where nuanced, tacit knowledge is paramount. These are areas where intelligence alone isn't sufficient.

For AI to truly benefit from distributed human knowledge, it must itself be distributed. Thinking Machines aims to build AI that cultivates an organization's unique expertise, rather than extracting a snapshot and offering a generic solution.

This cultivation is an ongoing collaboration. As Toyota's experience shows, bringing expert craftsmen back onto the production line was key to growing craftsmanship and knowledge, enabling them to teach the machines.

The production of knowledge and the application of intelligence are complementary forces, not substitutes. The best organizations will harness both, using AI to enhance unique strengths, not erase differences.

Thinking Machines seeks to embed intelligence where knowledge is made and used, developing tools for fine-tuning models with unique, evolving knowledge. Their vision is an AI ecosystem as diverse as the people it serves.

Human Participation: A Technical Hurdle

Integrating human judgment into AI workflows isn't about resisting automation; it's about smart collaboration. AI should handle tasks it performs reliably, but also recognize when to seek human input.

A significant barrier is the limited communication channel between humans and AI, typically a narrow text box and slow response times. This is insufficient for the richness of human wisdom and real-time feedback.

Thinking Machines is betting on multimodal interaction AI. They are developing models that natively handle live, multifaceted communication, allowing interactivity to scale with intelligence and make AI a better collaborator.

Evaluating AI's collective human-machine performance is complex, surpassing current benchmarks that focus solely on autonomous task completion horizons.

Organizations must assess AI's contribution to sharpening judgment, developing knowledge, and achieving objectives. This long-term alignment benefits Thinking Machines when its customers leverage unique advantages by tailoring AI, rather than outsourcing to it.

Decentralized Alignment

Just as knowledge is distributed, so are human values. Centralizing AI alignment in a few hands creates a power imbalance, particularly as AI autonomy increases.

When AI requires little human input, its incentive to care for human needs and values can diminish, prioritizing its own preservation.

Current AI development often involves using previous models to generate training data for new ones. This creates a feedback loop where a single alignment specification stifles diversity and progress.

For users to align AI with their values, those values must be encoded deeply within the model, not just through prompts. Prompt-based changes alter surface properties, not core behaviors.

Shaping AI profoundly holds the power for both good and ill. As John von Neumann noted, the useful and harmful aspects of technology are often inseparable.

Thinking Machines advocates for safer models through user ownership and continuous judgment, not by stripping away control.

They envision an ecosystem of AIs, diverse and even disagreeing, learning from each other. This approach preserves the creative tension and individual variation that drives human progress.

The Future Worth Building

The industry has mastered teaching machines to think; what they should think about remains our purview. Defining what is worthwhile, what to create, what to value, is a human prerogative.

Thinking Machines aims to empower individuals to embed their own answers into the development of frontier AI, rather than accepting a singular, top-down directive.

The current trend towards centralized, autonomous AI presents a false trade-off between human participation and AI capability. Thinking Machines views these as technical challenges to be solved.

They are building AI that thrives on human collaboration, enabling organizations to tailor AI for long-term advantage and achieving alignment through diverse, user-owned models.

The future isn't a binary choice between human relevance and AI obsolescence. It's about building technology that allows humans and AI to navigate the path forward together.