A week ago, much of the attention was on @AnthropicAI and @OpenAI pulling ahead. Now @Google is claiming frontier-level performance with Gemini 4 Argon, while Meta says its Muse Wonder Full model can compete with the leading systems. Google and Meta have advantages that AI-first companies cannot easily reproduce. They already have enormous computing resources, established distribution and products used by billions of people. They can put new AI capabilities in front of existing users without having to build an entirely new audience first.

We tend to measure intelligence by looking at what one person or one AI model can do. But some of humanity's greatest achievements depend on knowledge distributed across thousands of people, institutions and generations. Nobody needs to understand every part of an aircraft engine, a satellite or a communications network to contribute to building one. Progress becomes possible because people specialize, share what they learn and build on work that came before them. AI development is beginning to reflect this. One model might be good at planning, another at coding, and another at checking the results. Connecting these capabilities can produce something more useful than asking a single model to do everything. I think this has implications for how we approach AI infrastructure. We shouldn't focus exclusively on building ever larger individual systems. We should also pay attention to how different models, tools and people can work together reliably. The value of intelligence often depends on how well it is organized and made available to others. Humanity has been demonstrating that for centuries. AI gives us another way to put it into practice.

I take the possibility of losing control of advanced AI seriously, but I don't think we should talk about it as though the outcome has already been decided. There is a difference between recognizing that a technology could become difficult to control and concluding that humanity is inevitably approaching its end as the dominant intelligence on Earth. The first is a reason to invest in research, testing and safeguards. The second risks turning a difficult engineering and governance problem into a story about an unavoidable future. We are building systems that can perform complex tasks, and we still have important gaps in our ability to understand their behavior in every situation. That deserves sustained attention from researchers, companies and governments. It also means the decisions made during development matter. What capabilities are released, what safeguards are required, who can independently evaluate a system, and who is accountable when something goes wrong are all practical questions we can address. I don't believe we need certainty about the long term future to act responsibly today. The more capable these systems become, the more important it is that the people developing and deploying them can demonstrate that their capabilities are matched by appropriate oversight.

AI can give you a very convincing answer and still miss the most important detail about a real-world situation. That doesn't make experience infallible, and it doesn't mean AI cannot develop better practical judgement. It means we need to evaluate these systems on how they perform in real situations, not just how well they explain things. I see the strongest use of AI in combining its ability to process information with the experience of people who understand the environment where decisions have to be made. The goal should be to make better decisions, not only to produce more confident answers.

AI play a more direct role in how scientific discoveries are made, and the partnership between CERN and Mistral AI is an interesting example. The two organizations are developing Flare, a foundation model trained on billions of simulated particle collisions. The goal is to improve how physicists reconstruct and identify particles from the enormous volumes of data produced by experiments at the Large Hadron Collider. CERN already uses multiple AI tools to process experimental data, but the volume of information is expected to grow significantly as the collider is upgraded.

The ability to choose the right problem will become more valuable as AI makes it easier to execute ideas and solve technical problems. This is where experience, curiosity and an understanding of people's needs matter. The ability to recognize problems, understand why existing solutions fall short and decide what a better outcome should look like. I see this as a major opportunity for people who understand a particular industry or community deeply. They may not have been able to turn their knowledge into software before, but AI is lowering that barrier. Technical execution is becoming more accessible. Good judgement about what deserves to exist remains difficult, connecting a real human need with the capabilities of AI will have a meaningful advantage.

Some businesses will turn better models into sustainable revenue. Others may struggle to defend their margins as competitors gain access to similar capabilities. Infrastructure providers, application developers and established software companies will face very different economics. I remain interested in what AI can make possible, but I wouldn't assume every AI company is a good investment simply because the technology keeps advancing. More useful is where that progress creates lasting value.

The prospect of AI systems that can improve themselves and adapt closely to individual users raises questions that go well beyond model performance. Imagine an AI system that learns how you work, understands your priorities and can take on increasingly complex responsibilities without needing detailed instructions every time. That could make advanced capabilities useful to people who don't have the technical knowledge or resources to build such systems themselves. But personalization also brings a difficult question: whose interests does the system ultimately serve? An AI that understands you exceptionally well could help you make better decisions, or it could become very effective at influencing those decisions. The difference depends on how the system is designed, what objectives it follows and how much control you retain over its behavior. I believe people should have meaningful control over the AI systems they rely on, including how those systems use their information, what actions they can take and whose interests they are designed to serve. As AI becomes more autonomous, these choices will become harder to postpone. Building more capable intelligence is one challenge. Making sure that intelligence remains accountable to the people it serves is another.