Well, at least for me: no thanks.
You not envisioning use for it is just a past bias. You can't know that. You can't because we haven't yet reached the point where the OS is fully useful when controlled with AI.
What people want has not changed for millennia and it is unlikely to change soon.
Most of the things that are useful have already been imagined millennia ago, even if at that time nobody had any idea about how one could develop any technology for building such things in reality. For instance in the Ancient Greek literature there are descriptions of artificial robots for doing the hard work, means for flying etc.
The past bias can block indeed one to envision the usefulness of some things, but only when those things are not a goal in themselves, but they are only intermediates for achieving things that are already known to be useful and the past bias prevents the user to realize that there exists an alternate path to the useful goals, instead of the known traditional path.
LLMs are indeed tools that can be used to achieved some useful goals, so in some cases a user may not realize how they can be used, due to past bias.
There is no doubt that there are a few applications for which LLMs are very useful, but for experienced people, even if they have never used LLMs yet, it is easy to recognize with certainty that some of the proposed applications for LLMs will never be useful for them.
For example, I would never use an LLM for searching the Web or for summarizing documents. What I recognize as important in a Web search or in a document differs too much from what typical humans would recognize, for an LLM to have any chance to generate equivalent results.
The only reason why I may find useful to put some questions to a big LLM is because it is likely that it may have had access during training to documents to which I do not have access. Thus the answer might provide some clues about other sources than those known to me. Instead of this, I would very much prefer to use a traditional search tool on the training set, but the LLM may be a poor substitute for its training set, which is better than nothing.
For now, the most lucrative application for LLMs is as coding assistants. Here there is no past bias, because since the earliest times of automatic computers, people have hoped for methods that would allow the generation of computer programs with minimum input from a human.
I do not think that there is anyone who would dispute that LLMs have allowed a much greater progress than before in this direction. Here what are frequently disputed are only the correct strategies of using LLMs for this purpose, because it is obvious that they are frequently misused and those who do not understand programming, like most managers, have completely unrealistic ideas about what can be done and what should be done with LLMs.
We don't see the same obvious applications of AI because nobody has developed a proper user interface for it. We're stuck with voice, chat, and dumping documents onto it. The current pro-AI stance is basically "fuck the user and fuck interfaces".
Depends who 'we' is - I've seen plenty of non-tech people in the real world begin to use ChatGPT as a primary information source rather than the web (rightfully or not!)
I suspect that 'we' might not be the true early adopters here, similar to how quite a lot of the most technical users in the 80's thought GUI's were a waste of time.
I don't think that's really what people are talking about when they talk about 'agentic' PCs.
To be fair, I find the term to be as contrived as “performant”
Scandalous!
The most likely outcome is the world in the children’s cartoon “Thundarr the Barbarian”. People living in the collapsed ruins of the past society, belief in magic, etc.
A post-apocalyptic hellscape, essentially.
I fully expect that local models models that are comparable to current frontier models in performance will appear in the near future. Additionally, a lot more can be done with the harness as well, which in my opinion is an under-explored territory right now. For example, ATLAS does some clever tricks in this area https://github.com/itigges22/ATLAS
I started working on my own harness and also notice a significant improvement in model capability with it https://dirge-code.github.io
Apple seems to be one of the few companies to have realized that the future is likely local, and they've been focusing on optimizing hardware for that while everybody else seems to still be stuck in a model as a service paradigm.
> I started working on my own harness and also notice a significant improvement in model capability with it https://dirge-code.github.io
You should mine your session logs for examples of scenarios that demonstrate this improvement. If you can characterize it in a time series metric, like tokens/feature, as you applied improvements, then you're offering a receipt.
He also mentioned that the idea of agentic computing was already 30 years old, and that he was busying himself with the topic for 15 years by then (1990). So... five years from taking interest (mid-70s) to his first practical implementations (1980).
But for marketing, “artificial intelligence” is fine. And better than LLMs being called “AI”