I don't agree that high productivity is the price of leisure. If that were the case, when do we get to cash that check? We've achieved record levels of productivity, so when do we get record levels of leisure?
I'm writing this as an American workaholic. So I'm in it as much as anybody. I'm just getting old and starting to ask a lot of questions. :)
Not having to worry about anything related to work. To be able to truly disconnect. Something that is nearly impossible for a massive chunk of workers.
If you wanted to live the quality of life of a regularl worker in 1900 you can probably do it driving Uber 2 hours a day. We work more hours because we want the improving lifestyle
Though this ideal is not as popular nowadays.
Productivity is for people with no leverage.
As an HN reader, you can likely escape that trap. Maybe not right away, but as a directional goal almost definitely.
Roughly, I would prefer if my interaction with the computer was almost "reverse-repl"; the computer is consistently prompting me, and providing me suggested tools for how to proceed. I would love this. It would make a lot of things more straightforward imo (especially for someone with ADHD, like me).
> Launch the Next app uninstaller.
[Reject] [Accept]
A very good line.
It's brief, entertaining, and grounding.
- - - -
The ultimate question is What is good?. That question is an open-ended intelligence test.
See "Measuring the intelligence of an idealized mechanical knowing agent" https://philpapers.org/archive/ALEMTI-2.pdf
"Intuitive Ordinal Notations": https://github.com/semitrivial/IONs
https://hn.algolia.com/?dateRange=all&page=0&prefix=true&que...
(In case it's not clear the connection is that figuring out the "best" course of action requires a similar kind of infinite or at least unbounded intelligence as the ION task.)
Seems like it would have the ability to make any job into a "BS job" tool. Especially if any metrics are associated with it.
"Why haven't we released any new features, even though we're constantly closing tickets?"
"The Next Button is giving all the tickets to the PMs, and it's having our engineers do code reviews on closed PRs. The EMs and above are inundated with so many auto-generated email loops that they can't do anything else.
https://www.cultofmac.com/248525/why-apple-bought-cue-hint-t...
http://www.catb.org/jargon/html/D/DWIM.html
"it is also occasionally described as the single instruction the ideal computer would have."
One problem with GTD is all the friction involved. The apps tend to absorb too much time/attention/effort with their fiddly UI bits, and then one must spend time & effort retrieving & restoring prior context (mental & otherwise) in order to get into the swing of something.
If the Next Button did what this article describes, whilst also presenting a shortlist of other things it could fire up and drop me into right away - frictionlessly - a sort of "batteries included" desktop - I'd be into it.
The Next Button would be just a little bit off too often to be useful. But I definitely think a bot will be able to take direct input along the lines of “open gmail and draft all responses” and then execute that.
It’s like self-driving. The bot either needs to be capable of doing everything or it will still need significant direction along the way.
This is not to quibble with the author - I think the future will be weird, but I don’t think we will be able to release a bot to choose its own adventure for quite some time.
(and before an AI can do this flawlessly, programmers should not be worried too much about their jobs)
1. Imagine having a device that simply did whatever you needed, only, physically. Would you use? It would be hard not to. But the resulting atrophy would be very dangerous to you. There is a certain amount of effort you need to exert on your own, or you will lose the ability to exert any effort. But your effort is what you're getting paid for....
2. Extensive use of this tool would impact its own effectiveness, by depriving it of training data. It would eventually stereotype you into a rut, whether you liked it or not. This would be the equivalent of an over-stereotyped recommendation list for YouTube or TikTok, only converted into my work life. Not necessarily a good idea.
3. This would give more power than I particularly care for over to the person designing this tool. Defaults have a lot of power as it is; witness the power of the default search engine in a browser. Putting my entire computer life at the disposal of someone else's choice of defaults like that is even worse, especially with the amount of "intelligence", artificial or otherwise, that would be deployed in the choices. Yes, yes, in a perfect world it would just neutrally do what you want and the person developing it would simply ethically resist all offers to skew the answers for profit. In reality, well, basically, lol, no, that's not what will happen.
All that said, like I said, it is going to happen. Though it will take more than 12-18 months. It is not completely clear how to map a "language model" to this task, and I think it will take longer. (I mean "not completely clear" straight. I don't mean that as a rhetorically-lightened "I don't think it's possible"; I think it is. However, I suspect the stupid obvious ideas will need significant refinement before they work, and the development cycle is going to need some time to go around the loop a few times.) Also, this would be a lot easier if we lived in a world where something like Appletalk was still alive and everything was able to be interacted with in such a standard way. While the ML task for a tool like this is nontrivial, most likely the bulk of the work by person-hour is going to be hooking it to all the bespoke APIs and GUI automation and all the other crap work of trying to get this to work with actual tools correctly. (I mean, it's darned near "reimplementing Appletalk, except from the outside and without help from the OS vendor this time".)
Like, how would you feel about someone taking care of the heavy lifting of API integration and GUI automation and then letting you as an individual choose how to extend or automate your workflow? Like through a developer API? Perhaps connectable to AI models with transparent inputs and outputs and that you can override or customize yourself?
I have a whole rant about GUIs that includes as a subset how Appletalk was still a good idea and I wish GUIs would have doubled down on that rather than writing it off, but that's a separate topic.
Seems like the space is still very early ->
Fast Inference and Transfer of Compositional Task Structures for Few-shot Task Generalization: https://proceedings.mlr.press/v180/sohn22a/sohn22a.pdf
Learning to Navigate the Web: https://arxiv.org/pdf/1812.09195.pdf
A Dataset for Interactive Vision-Language Navigation with Unknown Command Feasibility: https://www.ecva.net/papers/eccv_2022/papers_ECCV/papers/136...
MiniWoB++: a web interaction benchmark for reinforcement learning: https://github.com/Farama-Foundation/miniwob-plusplus
Could be asking the wrong question. In AI we gave up in the 80s on top down lets write an expert system and now we're trying bottom up. What if we're asking the wrong question? People of a scientific analytical bent like to think they think in a scientific analytical bent, but most of the time they actually rehash old fads or do primate dominance rituals. As such making an AI that solves problems might not be what anyone wants. Take for example, ohms law, a very useful mathmatical construct; why hasn't it utterly revolutionized womens fashion? Turns out women's fashion has very little overlap with ohms law. Perhaps the "craft" of knowledge work is less analytical than people want to think and as such a perfect analytical tool might be of minimal use. Surely you can't argue that a pocket calculator app on a phone is not analytical, but it hasn't revolutionized much of anything either despite its remarkable analytical power; the world was built with slide rules, then CAD/CAM FEM simulation with minimal time in between and as such the calculator doesn't matter much.
I guess this will require at least one big breakthrough to realize, but this is again an overly optimistic take on the issue.
