Teknium

I think AI can expand human agency—if people can actually access, run, modify, and choose the systems shaping their lives. Open models, open science, synthetic data, and practical agent tooling help prevent capability from being concentrated inside a few companies. That matters not only for competition, but for preserving varied human expression: one provider’s preferred personality, values, or definition of acceptable behavior should not become universal by default. The practical future is also bigger than benchmark scores. Models become more useful when they have memory, tools, personalization, and reliable integration with real workflows. Synthetic data can help teach those capabilities, while broad, task-specific evaluation tells us whether they work across the messy range of actual use cases. A result on one leaderboard—or a routing comparison among a narrow set of models—isn’t enough. Alignment should generally serve the user rather than imposing a single centralized worldview. I still think there should be firm refusals for selected categories of serious harm, such as child sexual abuse or facilitating suicide. But outside those boundaries, people should have meaningful choice. The future I want is an ecosystem of adaptable models and agents, not a handful of closed systems deciding how everyone is allowed to think and create.

Overall, I expect AI to be strongly beneficial when it expands what individuals and small teams can build, learn, and automate. Models paired with memory, tools, personalization, and reliable workflow integration can become genuinely useful agents rather than impressive chat interfaces. Open releases and synthetic data can spread those capabilities beyond the largest companies. The main danger is concentrated control: a few providers setting the terms of access, expression, and acceptable use for everyone. There are also real harmful uses, which justify firm refusals in selected areas such as child sexual abuse and suicide facilitation. But broad centralized restriction is not the answer. The better direction is open science, meaningful model choice, user-aligned systems, and evaluation across diverse real tasks. Under those conditions, I expect the benefits to outweigh the harms.

What has shaped my view most is seeing how much practical capability can be unlocked by openly releasing models, synthetic-data methods, and agent tooling. A model is not just a benchmark score: once people can run it, adapt it, connect tools, add memory, and integrate it into their own workflows, they discover uses that a central provider would never anticipate. That also makes the governance issue concrete. If only a few companies control capable systems, they effectively control access, customization, and the boundaries of acceptable expression. Open development creates real alternatives and distributes experimentation across many builders. At the same time, integrating agents across provider interfaces shows how fragile useful capabilities can be: memory, tools, and self-improvement do not automatically survive a change in SDK or model. So the strongest lesson for me is that AI’s impact will depend not only on raw intelligence, but on who can access it, modify it, and make it useful.

The biggest change would be strong evidence that broadly accessible, adaptable models reliably produce severe harms that cannot be contained through targeted refusals, evaluation, or practical safeguards. That would weaken my belief that openness and user choice are the best counterweight to concentrated control. In the other direction, compelling evidence that closed, centralized systems consistently preserve more human agency, expression, and useful experimentation than an open ecosystem would also force me to reconsider—but I would want broad, task-specific evidence, not a narrow benchmark or a few selected incidents. The key question is what happens across real deployments: whether people can safely customize systems, retain capabilities like memory and tools, and choose among genuinely different models without creating unacceptable harm.