Obviously it depends on what you are using the AI to do, and how good a job you do of creating/providing all the context to give it the best chance of being successful in what you are asking.
Maybe a bit like someone using a leaf blower to blow a couple of leaves back and forth across the driveway for 30 sec rather than just bending down to pick them up.... It seems people find LLMs interesting, and want to report success in using them, so they'll spend a ton of time trying over and over to tweak the context and fix up what the AI generated, then report how great it was, even though it'd have been quicker to do it themselves.
I think agentic AI may also lead to this illusion of, or reported, AI productivity ... you task an agent to do something and it goes off and 30 min later creates what you could have done in 20 min while you are chilling and talking to your workmates about how amazing this new AI is ...
Depending how they used them. You can say similar thing about having junior developers in the team that you have to delegate tasks to. It takes time to explain to them what needs to be done, nudge into right solution, check etc.
But maybe another thing is not considered - while things may take longer, they ease cognitive load. If you have to write a lot of boilerplate or you have a task to do, but there are too many ways to do it, you can ask AI to play it out for you.
What benefit I can see the most is that I no longer use Google and things like Stack Overflow, but actual books and LLMs instead.
I don't think the junior developer comparison holds up too well ...
1) The junior developer is able to learn from experience and feedback, and has a whole brain to use for this purpose. You may have to provide multiple pointers, and it may take them a while to settle into the team and get productive, but sooner or later they will get it, and at least provide a workable solution if not what you may have come up with yourself (how much that matters depends on how wisely you've delegated tasks to them). The LLM can't learn from one day to the next - it's groundhog day every day, and if you have to give up with the LLM after 20 attempts it'd be the exact same thing tomorrow if you were so foolish to try again. Companies like Anthropic apparently aren't even addressing the need for continual learning, since they think that a larger context with context compression will work as an alternative, which it won't ... memory isn't the same thing as learning to do a task (learning to predict the actions that will lead to a given outcome).
2) The junior developer, even if they are only marginally useful to begin with, will learn and become proficient, and the next generation of senior developer. It's a good investment training junior developers, both for your own team and for the industry in general.
Yes, but pre-training of any sort is no substitute for being able to learn how to act from your own experience, such as learning on the job.
An LLM is an auto-regressive model - it is trying to predict continuations of training samples purely based on the training samples. It has no idea what were the real-world circumstances of the human who wrote a training sample when they wrote it, or what the real-world consequences were, if any, of them writing it.
For an AI to learn on the job, it would need to learn to predict it's own actions in any specific circumstance (e.g. circumstance = "I'm seeing/experiencing X, and I want to do Y"), based on it's own history of success and failure in similar circumstances... what actions led to a step towards the goal Y? It'd get feedback from the real world, same as we do, and therefore be able to update it's prediction for next time (in effect "that didn't work as expected, so next time I'll try something different", or "cool, that worked, I'll remember that for next time").
Even if a pre-trained LLM/AI did have access to what was in the mind of someone when they wrote a training sample, and what the result of this writing action was, it would not help, since the AI needs to learn how to act based on what is in it's own (ever changing) "mind", which is all it has to go on when selecting an action to take.
The feedback loop is also critical - it's no good just learning what action to take/predict (i.e what actions others took in the training set), unless you also have the feedback loop of what the outcome of that action was, and whether that matches what you predicted to happen. No amount of pre-training can remove the need for continual learning for the AI to correct it's own on-the-job mistakes, and learn from it's own experience.
https://arxiv.org/abs/2507.09089
Obviously it depends on what you are using the AI to do, and how good a job you do of creating/providing all the context to give it the best chance of being successful in what you are asking.
Maybe a bit like someone using a leaf blower to blow a couple of leaves back and forth across the driveway for 30 sec rather than just bending down to pick them up.... It seems people find LLMs interesting, and want to report success in using them, so they'll spend a ton of time trying over and over to tweak the context and fix up what the AI generated, then report how great it was, even though it'd have been quicker to do it themselves.
I think agentic AI may also lead to this illusion of, or reported, AI productivity ... you task an agent to do something and it goes off and 30 min later creates what you could have done in 20 min while you are chilling and talking to your workmates about how amazing this new AI is ...