AI Shell is an interactive shell that provides a chat interface with language models. The shell provides agents that connect to different AI models and other assistance providers. Users can interact with the agents in a conversational manner.
↫ Microsoft Learn
Basically, what Microsoft means with this is a split-view terminal where one of the two views is a prompt where you can ask questions to an “AI”, like OpenAI or whatever. The “AI” features are not actually integrated into your shell, which instead lives in the other view and acts like a completely normal, standard shell. Instead of opening up an “AI” chatbot in a browser window or whatever, you now have it in a split view in your terminal – that’s really all there’s to it here.
I’m going to blow your mind here and say that in theory, this could be an actually useful addition to terminals and shells, as a properly vetted and configured “AI” that has been trained on properly obtained source material could indeed be a great help in determining the right terminal commands and options. Tons of people already blindly copy and paste terminal commands from websites even though they really shouldn’t anyway, so it’s not like this introduces anything new here in terms of dangers.
Hell, tutorial writers still add -y to dnf or apt-get commands, so it can really only go up from here.
> I’m going to blow your mind here and say that in theory, this could be an actually useful addition to terminals and shells, as a properly vetted and configured “AI” that has been trained on properly obtained source material could indeed be a great help in determining the right terminal commands and options.
This is basically the premise of warp (https://www.warp.dev/)
I’ve been using it for ages without really touching the AI stuff as it’s just one of the better options for terminal emulators on macos, but a couple of times i’ve asked the AI when web search was fruitless and it’s been good
What should happen next is for market regulator to forbid Microsoft to pre-install “AI” in Windows. Instead of allowing it and then a couple of years down the road accusing Microsoft they abused their market position to squeeze out “AI” competition. Microsoft obviously can continue to develop and provide “AI” as a separate product and not as being part of Windows.
>”a properly vetted and configured “AI” that has been trained on properly obtained source material could indeed be a great help in determining the right terminal commands and options”
Is it worth the enormous environmental impact to give developers little more than they would get by keeping a well written coding bible at hand? Or to give them code snippets they could easily look up themselves on stackoverflow?
The suggestion on how this AI should be set up shows a fundamental lack of understanding of how LLMs work. It’s not a database. It’s a mind.
An LLM is a set of statistics and rules that are tuned for a purpose. Some of them appear to generate replies to questions, others create videos that appear real.
Use of words like intelligence, training, etc are very unfortunate because they hint at a similarity with humans which just isn’t there.
> Use of words like intelligence, training, etc are very unfortunate because they hint at a similarity with humans which just isn’t there…
… yet. What exactly will it take to turn “statistics and rules” into conscious minds? What is the “secret ingredient” or the spark to ignite true intelligence? A human baby is actually born “blank”, starting to collect data and training by endless “trial’n error” until (at least a small percentage of the specimen) becoming intelligent.
How sure are you that “intelligence” is more than just a function of the “statistics and rules”?
LLMs are more statistics than rules. A baby is probably the opposite. It requires far less data to generate a given level of intelligence. A baby is also self aware. We need to solve for awareness to derive true AGI. I used to think that it would require Penrose style quantum tubules or whatever. Post LLMs I am not so sure. In any case, creating truly situational and contextually aware AI remains a significant challenge. This is why we are probably a breakthrough of three away from AGI. I also think it will be a multi-sensory robot, not a command line. Getting back on topic, it is ironic that the current pinnacle of our computational achievement has come back to a command line in a shell. Stephenson will have to update his book: “In the Beginning Was the Command Line”.
LLMs are far from “minds”. I use the LLM “AI Chat” facility in duckduckgo.com [other search engines are available] when simple search doesn’t help. Even then, I find LLMs to be hard work drilling down until they either help or don’t help. Conversational Search does not mean LLMs are even remotely “minds”. Ask an LLM if you don’t believe me.