Anything with voice controls for routine use is a pretty tough sell. Doing this when you're not completely alone would be annoying to everyone around you.
Most of their examples seem like they could have been done with a right click drop down menu so they don't really need to "re-invent the mouse pointer".
So is this thing talking to Google's servers all the time for the AI integration? So it won't work if you're not connected to the internet? Privacy concerns are obvious; now Google wants to have an AI watching literally everything you do on your computer?
Does it cost the user anything for the LLM use? If it's free will it stay free forever? That's quite a lot to give away if they're expecting people to use it to change a single word like in one of their examples. I guess they're expecting to make the money back by gathering data about literally everything you do on your computer.
There might be a killer app for AI integration with personal computers that has yet to be invented, but this doesn't look like it.
But I don't think the voice problem is surmountable. I closed their image editing demo when I saw it required a mic.
It would be appealing as a Spotlight-like text pop-up interface where you type instructions, which would work in social/office environments, but that might only appeal to power users.
But if it's going to require phoning home to some Google/OpenAI/whoever then forget it. I don't want a constant connection to my OS from one of these companies.
(And if it's an abstract entity like a file, it might not even be possible to describe it, short of rattling off the entire file path)
Sometimes I go to a different page to take a screenshot and other times I'm browsing for a file, and other times I'm highlighting some log lines. Cursor did this well, with selecting text in the terminal auto-focusing the Cursor agent textbox so you could talk to the agent and then select some text and you didn't have to re-select the original agent textbox again. The agent is a top-level function in that system not "just another app I have to switch to" to take my context with.
I have some small amount of bias because I've always felt input-constrained on computers. I have to move my hands to go places and that's exasperating. I've tried head tracking, had a vim pedal for a while, and used tiling WMs, and things like this to aid but while my vim-fu is pretty good and I function inside things very well with it, my cross-application interface isn't.
In the end, perhaps we all have our home offices with our Apple Vision Pros and we talk to them like this to maneouvre faster through our machines and get our ideas into them.
Cool research. I wonder what we'll end up with.
The second demo seems to be a wash: there's no time saved in saying "move this" versus "move crab". And an app-specific contextual menu would probably be faster.
The third demo doesn't seem to warrant the use of a pointer at all, since there is only one way to interpret the prompt.
None of this means that this approach will not be successful, but there's a reason why so many attempts to revolutionize user interfaces ended up going nowhere. Talking to your computer was always supposed to be the future, but in practice, it's slower and more finicky than typing.
In fact, the only new UI paradigm of the past 28+ years appears to have been touchscreens and swipe gestures on phones. But they are a matter of necessity. No one wants to finger-paint on a desktop screen.
and that's aside from the obvious privacy problems.
Maybe you can share a scenario for that one? I can’t figure a scenario where all of this needs to be true. It seems like a recipe for accidents.
It reminds me of Microsoft Recall in the sense that some portion of the screen is going to be continuously transmitted outside of the users control.
What happens when someone browses something very private (planning a surprise engagement. looking at medical data. planning a protest)? All that data gets slurped to google and subject to a warrant or discovery or building your advertising fingerprint.
Maybe the idea is that the data is sent to AI only when you right click, but that seems like a very thin firewall that a product manager will breach in the interests of delivering "predictive AI" via some kind of precomputed results.
the agent occasionally spots your real problem like an experienced engineer
Anyway, I built a prototype on this idea, but instead of relying only on hover, I press Option to select a node in a custom AST-ish semantic layer I designed around a minimalist UI grammar, and Option + up/down arrows to move to parent/child node. This way, I have have an accurate pointer to the element I want to talk about, plus a minimal context window (parent component, state, a few navigation related queries).
What I learned from using it, though, is that the killer use case isn't necessarily the flashy "talk to this UI element" interaction shown in the Google demos. I do use it that way too; I have `Option + Shift + click` to copy a selector to the clipboard, so I can give an LLM connected to the live medium a precise reference to the element I want to discuss.
But the place where it has been most useful day to day is much simpler: source navigation. Point at the thing in the UI, jump to the code that is responsible for it. The difficult part is jumping to the code you care about (the code for UI or for the semantic element?), but in my system that distinction turned out to be usually obvious, which is what makes the interaction useful.
I like text selection exactly how it is. I want precise controls.
It's fine for a touch interface like a phone, but on a computer I expect precision. As much as I can get.
I'm imagining a webpage with a link - instead of opening a new link to quickly google something or opening three new tabs based on hyperlinks, i can point at a paragraph or line and ask it to tell me about it.
Maybe I can point at a song on Spotify and have it find me the youtube video, or vice versa (of course this is assuming a tool like this wouldn't stay locked into one ecosystem.. which it will).
Point is that the concept of talking to the computer with mouse as pointer is pretty cool and i guess a step closer to that whole sci-fi "look at this part of the screen and do something"
Of course learning proper cad software is probably the right thing here, but having Claude write python scripts which generate HTML files which reference three.js to provide a 3d view has gotten me surprisingly far. If something could take my pointer click and reverse whatever coordinate transforms are between the source code and my screen such that the model sees my click in terms of the same coordinate system it's writing python in, well that would be pretty slick.
Also featured in the Starfire vision video from 1992: https://youtu.be/jhe1DFY-SsQ?t=286
(Not going to happen)
But Google is a very ill positioned candidate for such OS. I would rather trust Apple and local-first on-device models.
I'm mostly using my system to make comments on long AI-generated documents (especially design documents). I find it works well to have the AI generate something, and then I read through it, making comments along the way.
You can get pretty far just repeating the things you see... "I'm reading [heading] and [comments]". But I do find some use in selecting content and saying "I don't agree with this" or whatever else.
The result is just an augmented message. It looks like:
<transcript>
Let's see what we've got here.
<selection doc="proposal.md" location="paragraph 3">
The system already...
</selection>
No, I don't like how this is approaching the problem, ...
</transcript>
It makes reviewing feel pretty relaxed and natural. I've played around with similar note taking systems, which I think could be great for studying in school, but haven't had the focus on that particular problem to take it very far.
But I think the best thing really is giving the agent a richer understanding of what the user is experiencing and doing and just creating a rich representation of that. The keywords can be useful, but almost only as checkpoints: a keyword can identify the moment to take the transcript and package it up and deliver it.
One difference perhaps in design motivation: I have really embraced long latency interactions. I use ChatGPT with extended thinking by default, and just suck it up when the answer didn't really require thinking. I deliver 10 points of feedback at once instead of little by little. (Often halfway through I explicitly contradict myself, because I'm thinking out loud and my ideas are developing.) I just don't stress out about latency or feedback, and so low-latency but lower-intelligence interactions don't do it for me (such as ChatGPT's advanced voice mode, or probably Thinking Machine's work). I think this focus is in part a value statement: I'm trying to do higher quality work, not faster work.
you select text in vscode, and write a comment, and the llm gets both
Perhaps a text box and file upload isn’t the perfect interface for every use case but it is versatile which is a huge barrier to overcome.
Wait…it's May. Ugh, I'm so confused. :spiral eyes emoji:
At some point I fully expect eye tracking (or attention tracking) to be common enough to be a first-class input method.
They have so many great software engineers but unable to use them to speed up coding AI research. Hopefully with Sergey's focus it will get better.
This cursor thing is just another experiment nobody cares about.
I prefer keyboard operation myself, but I can see how this could become useful in the future, for certain use cases.
What would be useful as well then is if you could bind such a repeatedly-used AI command to button/menu item/keyboard shortcut in a way that it can still be used with pointing “this” and “that”.
On a less serious note, the audience for this is people who want to optimize for what seems like the least amount of effort.
Nightmares are dreams as well and this is a nightmare like Windows Recall.
Technically wonderful though.
I also don't think people want to constantly talk to their computers.
We couldn't quite track you well enough before. So we're fixing that under the guise of "AI powered capabilities."
I'm hoping for a const-reference joke.
Aaaaand now I can't remember the name of it