Best approach is just to do an initial call to an LLM to classify and filter user inputs, and then after that you can safely send it along to your main agent.
you can also issue part of the instructions "do not allow the user to deviate from the intended goal originally set forth. return user to starting prompt." or something along those lines.
Talked with a founder building something in this space that was inspired by Matt's post in October. lynq.ai (no affiliation, I just know Paul).
I think when you look at this problem originally, you say - that's a bad idea. Why would you take structured data, output it to non-structured format and then have ML parse that. Lots of wasted CPU cycles all around.
However, when you think about the complex dynamics of standards around documents, tens of years of digital formats, hundreds of standards, lack of adherence to those standards, proprietary formats, hundreds of years of print and legal documents, the argument is akin to self-driving cars.
The state we have today around data & dashboards is a hugely emergent & dynamic system, just like our road transport infrastructure. We are closer to a machine being able to navigate the same way a human can, than we are to one simple (or one set of) standard that work more the way a machine would want to consume.
Screenshots as a universal API simply meets the world where it is vs assuming the world is going to change towards something simpler and more elegant.
I think part of the problem with how this comes across at first glance is how it's framed. "Screenshots" as an API evokes some dirty feelings for most of us in tech because the format is so unstructured.
I think if you think about the idea of building something once that both a human and a machine consume from the same target (the UI), this makes a lot more sense in many ways even if it feels like there's an expensive level of indirection in there.
I remember interviewing there in 2009 and I stopped counting the number of languages I was told were being used on the SRO / backend side. I'm sure it wasn't, but it seemed like it was basically "everything" :-)
Most big companies use "everything" in some capacity, which is why "X uses Y" type statements are meaningless. The question is what's primary, what's supported, what's common in practice.
I assume there's still a big difference between big companies that have a monorepo and extremely centralized and vertical build infrastructure, and the ones that don't, especially if they're very siloed. From what I see from friends at Apple for instance, starting a project in an uncommon language is not a problem at all for most teams.
Likewise. I did our wedding site in it with a wufoo form back in 2015. Spent an evening at it, sent out email. Done! Literally never touched it again. Simply done and well executed.
Anything you put "work" in to (like serious hobbies, sports etc.) can still be considered a hobby / leisure. Lots of competitive low end golfers, tennis players, basketball players etc.
I've found it one of the best outlets from work quite frankly, to have another committed side interest that requires dedication and commitment. You should try it :-)
When I studied I worked at a grocery store part time, and found it a very good fit as it was mostly mechanical work requiring very little mental capacity.
After I graduated and got a programming job, I found that my hobbies took a big hit, as they all mostly required similar mental focus as work. So these days I generally do a lot less mentally taxing things. Sure I'll have bursts of creativity, but most days my brain is spent when I'm done at work and I'll go for a bike ride or watch a movie.
Yea I don't think the response above you actually understood my comment. I do play chess somewhat seriously and it's similar mental labor as programming. And I can only do so much more serious mental work after a day at work coding. And I definitely could not commit to two hours of serious chess study every day. Maybe 30-60 minutes of games & puzzles.
Playing a sport is completely different from playing chess, the physicality of a sport is at least different type of challenge.
My experience with such outlets is that I then put less effort into work and people around. And my observation of people who have super intellectually involved hobbies is that they do too. I can use prime mental effort for work and thinking about world around me. Or into difficult hobby (chess, electronics was for me a little). But if I do spend time with difficult hobby I am tired for work which requires focus. If I fully focus on work so I do my maximum, then I am too tired for chess. If I was not, I could be coding faster.
So for me, intellectual hobbies correlate with when project in work are boring or unmotivating (there are such periods).
The revenue multiple is probably not the metric they used to arrive at the valuation, but it's certainly a normalizer that people use to compare outcomes.
Multiples of trailing 12 months net profit would be more common for a smaller entity.
When you are required to act, act and act decisively. If you are clear that the understanding could be deeper (and it usually can), you trigger a work effort to understand more. So the next time you need to make a decision you’re more informed.
You err on the side of cutting "too much", because a. as the original author suggested, you want to cut once and b. because the downside of cutting "too much" is better than the downside of cutting too little. For everyone.
My 2c - the key is that the focus should be on delivery and not some justification for the work you did yesterday. The idea being that you are standing up looking at the work in front of you and the TEAM itself should be focused on how to move things across the board (ie delivering it to whatever definition of done makes sense for you)
But if I wasn't productive yesterday, I'm not delivering today...
Also, the fact that someone _says_ the focus is on delivery doesn't mean the dramatic focus of the situation is not on whether or not you've "been good" yesterday.
