I get that you want to save the world by reducing processing, and I agree that using an LLM to develop deterministic and efficient code is just a better idea overall, but “stop using natural language interfaces” is overly restrictive.
Interactive fiction / text-adventures written in the 20th century used a deterministic natural language interface with low load as an intentional flexible puzzle to solve, so the problem today is efficiency.
You could just as well argue to stop using modern bloated operating systems, websites, and apps. I understand that the processing required for LLMs can be much higher. But the side-effect of additional power needs will be a global push for more energy, which will result in more power stations being available for future industries once LLMs become more efficient.
If you want to reduce complexity overall and have simple, flexible interfaces and applications that use fewer of the worlds resources, I’m all for it. But don’t single out LLMs assuming they will always be less efficient. Cost will drive them to be more efficient over time.
>As a clear obvious example: interactive fiction / text-adventures use a deterministic natural language interface with low load as an intentional flexible puzzle to solve.
Even though games can technically do this, should they? Do consumers actually find it fun and engaging? Considering there has never been a AAA game of that genre I don't think there is true consumer demand for games with such an interface.
Are there any successful examples of LLM text adventures? Last time I heard someone here said it's hard to develop robust puzzles and interactions, because it's hard to control and predict what the LLM will do in a dialogue setting. E.g. the user can submit reasonable but unintended solutions to a puzzle, which breaks the game.
I think text interface sucks, but at the same time I like how Claude code solve that with questionnaires, I think that’s the most elegant solution to get a lot of valuable context from users in a fast way
You can still have “chat interface” but if you use it for specialized applications you can do better than that.
If I can do some actions with a press of a button that runs code or even some LLM interaction without me having to type that’s so much better.
Feedback interface with plain text is awful, would be much better if there is anything that I have to repeat or fix on my end standing out - or any problem that LLM is looping over quickly discoverable.
Love this, this is what I have been envisioning as a LLM first OS! Feels like truly organic computing. Maybe Minority Report figured it out way back then.
The idea of having the elements anticipated and lowering the cognitive load of searching a giant drop down list scratches a good place in my brain. Instantly recognize it as such a better experience than what we have on the web.
I think something like this is the long term future for personal computing, maybe I'm way off, but this the type of computing I want to be doing, highly customized to my exact flow, highly malleable to improvement and feedback.
Veering offtopic a bit... Google lost its (search) way years ago. See the "The Man who Killed Google Search" [1], and the room they left for alternatives like DuckDuckGo.
At work, we have full access to Claude, and I find that I now use that instead of doing a search. Sure it's not 100% reliable, but neither is search anyhow, and at least I save time from sifting through a dozen crappy content farms.
The same, I suppose, as using Wikipedia to get an overview of a topic, a surface understanding, before following the citations to dig deeper and fully validate the summary.
Unless I am wildly misreading this, this is actually worse that both GUIs and LLMs combined.
LLMs offer a level of flexibility and non-determinism that allow them to adapt to different situations.
GUIs offer precision and predictability - they are the same every time. Which means people can learn them and navigate them quickly. If you've ever seen a bank teller or rental car agent navigate a GUI or TUI they tab through and type so quickly because they have expert familliarity.
But this - with a non-determinstic user interface generated by AI, every time a user engages with a UI its different. So they a more rigid UI but also a non-deterministic set of options every time. Which means instead of memorising what is in every drop down and tabbing through quickly, they need to re-learn the interface every time.
It’s intended for conversations that are probably different every time too. It’s like a more expressive form of what Claude Code already does with the “AskUserQuestion” interface.
I don't think you have to use this if it's not working in your case. I think the idea is to try to anticipate the next few turns of the conversation, so you can pick the tree you want to go down in a fast way. If the prediction is accurate, I could see that being effective.
Human abstract language, particularly the English language, is a pretty low-fidelity way to represent reality and in countless instances it can fail to represent the system to any useful or actionable degree.
Interfaces are hard, abstraction is hard. Computer science has been working on making these concerns easier to reason about, and the industry has put a lot of time and effort into building heuristics (software / dev mgmt / etc frameworks) to make achieving an appropriate abstraction (qua ontology) feasible to implement without a philosophy degree. We, like biological systems, have settled on certain useful abstraction layers (OOP, microservice arch, TDD, etc.) that have broad appeal for balancing ease of use with productivity.
So it should be with any generative system, particularly any that are tasked with being productive toward tangible goals. Often the right interface with the problem domain is not natural language. Constraining the "information channels" (concepts/entities and the related semantics, in the language of ontology) to the best of your ability to align with the inherent degrees of freedom, disambiguated as best as possible into orthogonal dimensions (leaning too hard on the geometric analogy now). For generating code, that means interacting with tokens on ASTs, not 1D sequences of tokens. For comprehending 3D scenes, a crude text translation from an inherently 2D viewpoint will not have physics, even folk physics, much in mind except by what it can infer from the dataset. For storing, recalling, and reciting facts per se, the architecture shall not permit generating text from nonverifiable sources of information such as those vector clouds we find between the layers of any NN.
These considerations early in the project massively reduce the resource requirements for training at the expense of SME time and wages to build a system that constrains where there are constraints and learns where there are variables.
The latency argument is terrible. Of course frontier LLMs are slow and costly. But you don't need Claude to drive a natural language interface, and an LLM with less than 5B parameters (or even <1B) is going it be much faster than this.
I get that you want to save the world by reducing processing, and I agree that using an LLM to develop deterministic and efficient code is just a better idea overall, but “stop using natural language interfaces” is overly restrictive.
Interactive fiction / text-adventures written in the 20th century used a deterministic natural language interface with low load as an intentional flexible puzzle to solve, so the problem today is efficiency.
You could just as well argue to stop using modern bloated operating systems, websites, and apps. I understand that the processing required for LLMs can be much higher. But the side-effect of additional power needs will be a global push for more energy, which will result in more power stations being available for future industries once LLMs become more efficient.
If you want to reduce complexity overall and have simple, flexible interfaces and applications that use fewer of the worlds resources, I’m all for it. But don’t single out LLMs assuming they will always be less efficient. Cost will drive them to be more efficient over time.
>As a clear obvious example: interactive fiction / text-adventures use a deterministic natural language interface with low load as an intentional flexible puzzle to solve.
Even though games can technically do this, should they? Do consumers actually find it fun and engaging? Considering there has never been a AAA game of that genre I don't think there is true consumer demand for games with such an interface.
> never been a AAA game
Infocom sold 450k copies of Zork I and 250k copies of The Hitchhiker's Guide among their many other titles.
Beam Software sold over 1M copies of The Hobbit.
Sierra On-Line sold ~400k copies of King’s Quest VI in a week.
Are there any successful examples of LLM text adventures? Last time I heard someone here said it's hard to develop robust puzzles and interactions, because it's hard to control and predict what the LLM will do in a dialogue setting. E.g. the user can submit reasonable but unintended solutions to a puzzle, which breaks the game.
stop using natural language interfaces is just a hyperbole.
I think text interface sucks, but at the same time I like how Claude code solve that with questionnaires, I think that’s the most elegant solution to get a lot of valuable context from users in a fast way
You can still have “chat interface” but if you use it for specialized applications you can do better than that.
If I can do some actions with a press of a button that runs code or even some LLM interaction without me having to type that’s so much better.
Feedback interface with plain text is awful, would be much better if there is anything that I have to repeat or fix on my end standing out - or any problem that LLM is looping over quickly discoverable.
Love this, this is what I have been envisioning as a LLM first OS! Feels like truly organic computing. Maybe Minority Report figured it out way back then.
The idea of having the elements anticipated and lowering the cognitive load of searching a giant drop down list scratches a good place in my brain. Instantly recognize it as such a better experience than what we have on the web.
I think something like this is the long term future for personal computing, maybe I'm way off, but this the type of computing I want to be doing, highly customized to my exact flow, highly malleable to improvement and feedback.
My boss used to say: "there is an easy way and there is the cool way".
We no longer have StackOverflow. We no longer have Google, effectively.
I used to be able to copy pasta code with incredible speed - now all of that is gone.
Chatbots is all we have. And they are not that bad at search, with no sponsored results to weed through. For now.
> We no longer have Google, effectively
Veering offtopic a bit... Google lost its (search) way years ago. See the "The Man who Killed Google Search" [1], and the room they left for alternatives like DuckDuckGo.
At work, we have full access to Claude, and I find that I now use that instead of doing a search. Sure it's not 100% reliable, but neither is search anyhow, and at least I save time from sifting through a dozen crappy content farms.
[1] https://www.wheresyoured.at/the-men-who-killed-google/
When you use a search engine, you can evalute the "trustness" of the source (webpage), this essentially disappears when using a chatbot
Only if you're dumb about it. Asking for source links is one thing I do all the time, and chatgpt gives citations by default.
What's the benefit of using a chatbot if you still have to go and read all of it's sources on your own?
The same, I suppose, as using Wikipedia to get an overview of a topic, a surface understanding, before following the citations to dig deeper and fully validate the summary.
> I used to be able to copy pasta code with incredible speed - now all of that is gone.
What do you mean by that?
Meaning I would get an answer from SO pretty quickly and it would most often work.
Of course not. Users love the chatbot. It's fast and easier to use than manually searching for answers or sticking together reports and graphs.
There is no latency, because the inference is done locally. On a server at the customer with a big GPU
> There is no latency
Every chat bot I was ever forced to use has built-in latency, together with animated … to simulate a real user typing. It’s the worst of all worlds.
Because they are all using some cloud service and external LLM for that. We not.
We sell our users a strong server, where he has all his data and all his services. The LLM is local, and trained by us.
The post suggests how to optimize the LLM text with UI elements that reduce the usage of pure/direct prompts.
And that’s perfectly fine.
Though the title in that sense is more of a click-bait.
Is this a bad bait or is it a bad post? I can't decide.
Let's go further. Why not have a well specified prompt programming language for LLMs then?
Unless I am wildly misreading this, this is actually worse that both GUIs and LLMs combined.
LLMs offer a level of flexibility and non-determinism that allow them to adapt to different situations.
GUIs offer precision and predictability - they are the same every time. Which means people can learn them and navigate them quickly. If you've ever seen a bank teller or rental car agent navigate a GUI or TUI they tab through and type so quickly because they have expert familliarity.
But this - with a non-determinstic user interface generated by AI, every time a user engages with a UI its different. So they a more rigid UI but also a non-deterministic set of options every time. Which means instead of memorising what is in every drop down and tabbing through quickly, they need to re-learn the interface every time.
> GUIs offer precision and predictability - they are the same every time.
Except after an update everything is in a different place.
It’s intended for conversations that are probably different every time too. It’s like a more expressive form of what Claude Code already does with the “AskUserQuestion” interface.
I don't think you have to use this if it's not working in your case. I think the idea is to try to anticipate the next few turns of the conversation, so you can pick the tree you want to go down in a fast way. If the prediction is accurate, I could see that being effective.
This is something I agree with.Will be interesting to see if more and more people take this philosophy up.
Human abstract language, particularly the English language, is a pretty low-fidelity way to represent reality and in countless instances it can fail to represent the system to any useful or actionable degree.
Interfaces are hard, abstraction is hard. Computer science has been working on making these concerns easier to reason about, and the industry has put a lot of time and effort into building heuristics (software / dev mgmt / etc frameworks) to make achieving an appropriate abstraction (qua ontology) feasible to implement without a philosophy degree. We, like biological systems, have settled on certain useful abstraction layers (OOP, microservice arch, TDD, etc.) that have broad appeal for balancing ease of use with productivity.
So it should be with any generative system, particularly any that are tasked with being productive toward tangible goals. Often the right interface with the problem domain is not natural language. Constraining the "information channels" (concepts/entities and the related semantics, in the language of ontology) to the best of your ability to align with the inherent degrees of freedom, disambiguated as best as possible into orthogonal dimensions (leaning too hard on the geometric analogy now). For generating code, that means interacting with tokens on ASTs, not 1D sequences of tokens. For comprehending 3D scenes, a crude text translation from an inherently 2D viewpoint will not have physics, even folk physics, much in mind except by what it can infer from the dataset. For storing, recalling, and reciting facts per se, the architecture shall not permit generating text from nonverifiable sources of information such as those vector clouds we find between the layers of any NN.
These considerations early in the project massively reduce the resource requirements for training at the expense of SME time and wages to build a system that constrains where there are constraints and learns where there are variables.
> just because we suddenly can doesn't mean we always should
Author should take his own advice.
The latency argument is terrible. Of course frontier LLMs are slow and costly. But you don't need Claude to drive a natural language interface, and an LLM with less than 5B parameters (or even <1B) is going it be much faster than this.
And it's highly circumstancial, as LLM efficiency keeps improving as the tech matures.
Yeah … no. It’s really nice interface. It’s here to stay.