Completely agree, I often find myself wishing that claude could talk back to me. If the model was slightly smarter and had these advance voice features it would be a no brainer for 100$+ subscription
It probably isn’t dangerous at the moment, but you’re giving the AI one of the pieces needed to be truly dangerous: it can operate arbitrary web sites, presumably including web sites that you really, really don’t want a malicious AI to be clicking on. Guess there’s some form of whitelsying here so you can control what it accesses. But I’m thinking ahead, to the smarter malucioys AI that first hacks its way around the white list and then presses the button.
I asked it to use https://youfiles.herokuapp.com/telnetclient/ to access telnet and view some ham radio spots. I told o1 about this and it was like "why would anyone ship a model that can use telnet, are they insane?" hehe
> The Number You Are Calling Is Not Available (In the EU)
Note that, using Operator is against the terms of service of pretty much any website you care to name, and thus illegal in the US as well under CFAA. I would bet at odds this is even true of the demo websites like OpenTable.
If you have to double-check it, it's worse in most cases than doing it yourself, because of the cognitive load associated with establishing trust in the outcome. It's different, much more uncomfortable type of work for the user. It might be laborious to load your grocery shopping cart but cognitive weight of all the decisions are spread over the session and at the end there's little worry that there's a mistake.
If you come in at the end of the process and have to check the presence or absence of every item, whether the quantities are correct, etc etc, it's much more of a head-scratcher kind of work and potentially much less pleasant than the monotony of doing it yourself.
Bixby reached Stage 3 on that list, and it was fucking useless. This stuff has to be better than that kind of "assistant" or it will languish the same. The only agentic thing I think I've liked so far is Gemini's Deep Research, which isn't actually that deep but seems a good place to start on a given topic.
Who wants "a good place to start" on your grocery shopping? There's pain in dropping into a practical process like that midway through or towards the end.
I remain unconvinced. I know its early days; I will watch from the sidelines on this one.
Podcast episode for this post:
https://open.substack.com/pub/dwatvpodcast/p/operator
What are the current bottlenecks for greater improvement?
1. Data?
2. Algorithms?
3. Compute?
What I really want is an Operator with a very chill advanced voice that's just perma-active and that I can ask to do things and converse with.
I don't want to replace my brain, I want to replace my keyboard, mouse and screen.
Completely agree, I often find myself wishing that claude could talk back to me. If the model was slightly smarter and had these advance voice features it would be a no brainer for 100$+ subscription
It probably isn’t dangerous at the moment, but you’re giving the AI one of the pieces needed to be truly dangerous: it can operate arbitrary web sites, presumably including web sites that you really, really don’t want a malicious AI to be clicking on. Guess there’s some form of whitelsying here so you can control what it accesses. But I’m thinking ahead, to the smarter malucioys AI that first hacks its way around the white list and then presses the button.
I asked it to use https://youfiles.herokuapp.com/telnetclient/ to access telnet and view some ham radio spots. I told o1 about this and it was like "why would anyone ship a model that can use telnet, are they insane?" hehe
> The Number You Are Calling Is Not Available (In the EU)
Note that, using Operator is against the terms of service of pretty much any website you care to name, and thus illegal in the US as well under CFAA. I would bet at odds this is even true of the demo websites like OpenTable.
If you have to double-check it, it's worse in most cases than doing it yourself, because of the cognitive load associated with establishing trust in the outcome. It's different, much more uncomfortable type of work for the user. It might be laborious to load your grocery shopping cart but cognitive weight of all the decisions are spread over the session and at the end there's little worry that there's a mistake.
If you come in at the end of the process and have to check the presence or absence of every item, whether the quantities are correct, etc etc, it's much more of a head-scratcher kind of work and potentially much less pleasant than the monotony of doing it yourself.
Bixby reached Stage 3 on that list, and it was fucking useless. This stuff has to be better than that kind of "assistant" or it will languish the same. The only agentic thing I think I've liked so far is Gemini's Deep Research, which isn't actually that deep but seems a good place to start on a given topic.
Who wants "a good place to start" on your grocery shopping? There's pain in dropping into a practical process like that midway through or towards the end.
I remain unconvinced. I know its early days; I will watch from the sidelines on this one.