Focus of the letter 33

While testing Holotab this week, I thought of Clico, which we talked about in the focus of letter 29. Clico, like other extensions already tested, installs itself in the browser and offers a chatbot superimposed on the pages visited: the generative AI is there, available, it is up to us to decide whether to open it, to take it into accounts or not. The friction remains present, even if it remains discreet.

Holotab is taking it one step further. It’s no longer an assistant stationed at the edge of navigation, it’s an agent that navigates for us: we describe a task in a few words, in writing or out loud, and the extension opens tabs, clicks, fills in fields, extracts what it finds. Comet from Perplexity or other browsers already offer it, here Holotab inserts itself into our everyday browser, without any installation or change. You can even save routines so that they can be replayed without having to re-write. The interface almost disappears.

Holotab offers three levels of autonomy. In supervised mode, the agent asks for confirmation before each action. In balanced mode, it acts on common tasks and only stops for irreversible actions. In autonomous mode, it continues without stopping, except when faced with a login or a captcha. We have to make this choice in the settings. The question of the degree of delegation, which is rarely touched upon because it is rarely asked explicitly, is here put in the interface itself.

When choosing the level of autonomy, we realize that we haven’t necessarily asked ourselves this question for other tools. For the language models we use on a daily basis, for extensions already installed, for automations already in place. The level of autonomy is often decided by default.

What Holotab makes visible is that delegating web browsing and delegating a decision are not quite the same thing, but they sometimes take the same paths. Recording a routine means entrusting the machine with a repetitive behavior that you have mastered and that you could decide to modify slightly each time. The routine frees up time, it also erases the micro-decision.

There is still one question that three levels of autonomy do not decide: how far do we let the agent in? Holotab navigates for us, it sees what we see: our mailboxes, our workspaces, our accounts, maybe even other much more sensitive data that we would let us see. The autonomous mode stops at logins and captchas, it’s a technical limit, not a guarantee: if the account is already connected upstream, no shutdown. And even if H Company says it blocks use on certain sites and does not keep the screenshots necessary for operation, they do pass on servers…

We can ask ourselves what we agree to make visible to an agent that we no longer really control in real time, what he goes through without us paying attention, what it implies… Delegating a search task and delegating navigation in a personal space do not present the same level of risk, even if the interface is identical. Perhaps you should set your own limits before choosing your level of autonomy, and not the other way around.

Here we find the thread of the “oh yes?” rather than the “wow” we talked about in letter 28, and the question asked by Echo in letter 30 : it’s not whether the machine is doing what we ask of it, it’s whether we have decided what we asked of it. The question that Holotab explicitly asks, perhaps we should force ourselves to ask ourselves for all the tools we already use, even if it is not written anywhere.