I’m more optimistic about AI than I used to be. I think its applications, and its capacity to change how we get things done, go much deeper than I initially appreciated. Part of me still reaches for “a very good language predictor” as an explanation. That intuition makes the machinery feel familiar, but it does surprisingly little to tell me where the machinery can lead.
Browser use brings that tension into focus. An AI system that can operate an interface has a different relationship with software from one that can only explain it. It can take a step, see what happened, and choose the next step in light of the result. The world gets a turn in the conversation.
The next observation changes the plan
Imagine asking an agent to help rearrange a trip. This is an illustrative scenario, not something I’m claiming to have built or tested. The departure changes, the accommodation needs checking, and a booking page reveals a condition that wasn’t visible in the search results. A useful system would carry that discovery back into its plan before preparing the next action.
A browser gives it a route through software that already exists, including interfaces for which nobody has written a bespoke integration. Anthropic’s 2024 account of developing computer use describes training a model to interpret screenshots and translate instructions into computer actions. The important extension is that model output can become a click or keystroke, followed by another observation. It does not establish that every website is reliably operable.
There is an intellectual precursor in Shunyu Yao and colleagues’ ReAct paper, which interleaves reasoning and actions so that information from an environment can inform what happens next. I find the feedback more consequential than the appearance of a model narrating its thoughts. An elegant plan made before opening the booking page may deserve to be abandoned after reading it.
I would design the travel system to distinguish an attempted action from a confirmed outcome. Clicking a button would record an attempt. A confirmation page would provide evidence about the resulting booking. If the browser lost its connection after the click, the next job would be to inspect the reservation, because repeating the action could create a second booking. A confident sentence saying “all sorted” would settle none of that.
This is part of what makes me optimistic: the next useful step can be an observation. The system needn’t know everything at the start if it can discover what matters and revise its approach. That creates room for useful delegation in the messy middle of a task, where the instructions are understandable but the route through the software isn’t fixed.
The practical consequence reaches beyond travel. Many small tasks fall between products, and someone becomes the person who carries context from screen to screen. Browser agents offer a way to explore automating that connective work without first rebuilding every application around a shared interface. Reliable APIs remain attractive where they exist; the browser expands the set of places a system can reach. How much work looks too small to automate only because connecting the software was the larger job? I think that boundary is worth reconsidering.
Prediction can be a beginning
AlphaFold helps me examine the “predictor” intuition from another direction. In John Jumper and colleagues’ 2021 paper, AlphaFold2 predicts protein structure with an architecture built for that problem. Its Evoformer processes representations of aligned sequences from related proteins and pairs of amino-acid residues; a structure module produces three-dimensional coordinates. It is a specialized structure-prediction system, not a GPT-style language model or a browser agent.
The distinction matters because the lesson depends on the task. A structural prediction can become something a scientist reasons with, with confidence and biological limitations still to assess. A proposed browser action can become something software executes, with permissions and the resulting state still to check. These are different mechanisms. In each case, calling the output a prediction leaves open the consequential question of what someone can do with it.
My optimism sits in those downstream possibilities. A capability becomes more interesting when its output enters a process that can use it, challenge it, or reveal what should happen next. I don’t need AlphaFold to be evidence for general agency to find its implications profound.
That also puts a constraint on what I would build. A travel agent’s ability to read a page should not give that page authority to redefine the trip or instruct it to transmit private information. Simon Willison’s “lethal trifecta” argument explains why combining private data, untrusted content, and external communication creates a prompt-injection risk. It is a useful challenge to any cheerful story about connecting everything. I would enforce booking permissions in the application and restrict where private information can be sent; retrieved text would have no authority to expand either.
The engineering is part of the possibility. Choosing what a system may do, what counts as evidence, and when it must stop gives us a way to make a broad capability useful for a particular purpose.
I still haven’t settled what the language-prediction intuition tells me about intelligence. I’m much more certain that I was underestimating what becomes possible when a model can reach existing software, observe the consequences of its actions, and keep working from what it learns. That is enough to make the space of applications feel considerably larger than it used to.