It bothers me how the conversation about artificial intelligence is filled with certainty about things that are still being investigated. An advance becomes an announcement of sweeping change, a warning becomes a prediction of catastrophe, and anyone trying to keep up is left wondering how much has already happened and how much depends on a chain of possibilities.
A concern raised by someone who works on developing these systems deserves attention. That person may know about problems that have not yet reached the public. But when an account starts to circulate, we also need to be able to distinguish what they observed from what they believe could happen as a result. Their experience matters, as do the limits of their conclusions.
There are concrete advances in this field. Anthropic itself describes results from AI helping to develop new systems, while acknowledging that it has not yet reached the point where an AI can develop its successor entirely on its own, and that this trajectory is not inevitable. Even within a company involved in this race, there are differences between the results presented and the future envisioned.
What I find curious is how the expectation of an extraordinary leap can serve arguments that seem to disagree completely. The possibility of much more capable systems supports concerns about losing control, but also promises of enormous benefits. In Machines of Loving Grace, Dario Amodei imagines profound changes across several fields, acknowledging that he is working with projections. The future being described may inspire hope or fear, while how close it might be remains an important question.
I also find it difficult to leave money out of this conversation. For people using these tools, there are subscription fees, credits, and the cost of repeated attempts to get a useful result. It is reasonable to ask how much value is already being delivered and how much of the expectation depends on future versions. That question deserves space even when the tools available today are useful.
I worry that all of this, mixed together, ends up creating an aversion to anything related to AI. Different applications come to carry the weight of the same debate, as though trying a tool meant agreeing with every promise made by the companies developing it.
Anticipating risks is necessary, including when there is still uncertainty. At the same time, I would like to see that uncertainty preserved when the subject reaches the public. When someone makes such a far-reaching prediction, I want to understand how they arrived at it and what would still need to happen for it to be borne out.
