5 ms·

I don't. Chat interface sucks; for most of these things, a more direct interface could be much more ergonomic, and easier to operate and integrate. The only rea

I don't. Chat interface sucks; for most of these things, a more direct interface could be much more ergonomic, and easier to operate and integrate. The only reason we don't have those interfaces is because neither restaurants, nor airlines, nor online stores, nor any other businesses actually want us to have them. To a business, the user interface isn't there to help the user achieve their goals - it's a platform for milking the users as much as possible. To a lesser or greater extent, almost every site actively defeats attempts at interoperability. Denying interoperability is so culturally ingrained at this point, that it got pretty much baked into entire web stack. The only force currently countering this is accessibility - screen readers are pretty much an interoperability backdoor with legal backing in some situations, so not every company gets to ignore it. No, we'll have to settle for "chat agents" powered by multimodal LLMs working as general-purpose web scrappers, because those models are the ultimate form of adversarial interoperability, and chat agents are the cheapest, least-effort way to let users operate them.

gordon_freeman 2y ago
I also do not like Chat interface. What I meant by above comment was actually talking and having natural conversations with Operator agent while driving car or just going for a walk or whenever and wherever something comes to my mind which requires me to go to browser and fill out forms etc. That would get us closer to using chatGPT as a universal AI agent to get those things done. (This is what Siri was supposed to be one day when Steve Jobs introduced it on that stage but unfortunately that day never arrived.)
- TeMPOraL 2y ago
  > This is what Siri was supposed to be one day when Steve Jobs introduced it on that stage but unfortunately that day never arrived. The irony is, the reason neither Siri nor Alexa nor Google Assistant/Now/${whatever they call it these days} nor Cortana achieved this isn't the voice side of the equation. That one sucks too, when you realize that 20 years ago Microsoft Speech API could do better, fully locally, on cheap consumer hardware, but the real problem is the integration approach. Doing interop by agreements between vendors only ever led to commercial entities exposing minimal, trivial functionality of their services, which were activated by voice commands in the form of "{Brand Wake word}, {verb} {Brand 1} to {verb} {Brand 2}" etc. This is not an ergonomic user interface, it's merely making people constantly read ads themselves. "Okay Google, play some Taylor Swift on Spotify" is literally three brand ads in eight words you just spoke out loud. No, all the magical voice experience you describe is enabled[0] by having multimodal LLMs that can be sicced on any website and beat it into submission, whether the website vendor likes it or not. Hopefully they won't screw it up (again[1]) trying to commercialize it by offering third parties control over what LLMs can do. If, in this new reality, I have to utter the word "Spotify" to have my phone start playing music, this is going to be a double regression relative to MS Speech API in the mid 2000s. -- [0] - Actually, it was possible ever since OpenAI added function calling, which was like over a good year ago - if you exposed stuff you care about as functions on your own. As it is, currently the smartphone voice assistant that's closest to Star Trek experience is actually free and easy to set up - it's Home Assistant with its mobile app (for the phone assistant side) and server-side integrations (mostly, but not limited to, IoT hardware). [1] - Like OpenAI did with "GPTs". They've tried to package a system prompt and function call configuration into a digital product and build a marketplace around it. This delayed their release of the functionality to the official ChatGPT app/website for about half a year, leading to an absurd situation where, for those 6+ months, anyone with API access could use a much better implementation of "GPTs" via third-party frontends like TypingMind.
sky2224 2y ago
I think the chat interface is bad, but for certain things it could honestly streamline a lot of mundane things as the poster you're replying two stated. For example, McDonald's has heavily shifted away from cashiers taking orders and instead is using the kiosks to have customers order. The downside of this is 1) it's incredibly unsanitary and 2) customers are so goddamn slow at tapping on that god awful screen. An AI agent could actually take orders with surprisingly good accuracy. Now, whether we want that in the world is a whole different debate.
- krapp 2y ago
  McDonald's already tried having AI take orders and stopped when the AI did things like randomly add $250 of McNuggets or mistake ketchup for butter. Note - because this is something which needs to be pointed out in any discussion of AI now - even though human beings also make mistakes this is still markedly less accurate than the average human employee.
  - ItsMattyG 2y ago
    For now
    - TeMPOraL 2y ago
      Indeed. I think a GPT-4o class model, properly prompted, would work just fine today. The trick is, unlike a human, the computer is free to just say "no" without consequences. The model could be aggressively prompted to detect and refuse weird orders. Having to escalate to a human supervisor (who conveniently is always busy doing other things and will come to you in a minute or three) should be sufficient at discouraging pranksters and fraudsters, while not annoying enough to deter normal customer. (I say model, but for this problem I'd consider a pipeline where the powerful model is just parsing orders and formulating replies, while being sanity-checked by a cheaper model and some old-school logic to detect excessive amounts or unusual combinations. I'd also consider using "open source" model in place of GPT-4o, as open models allow doing "alignment" shenanigans in the latent space, instead of just in the prompts.)
- 2y ago
csharpminor 2y ago
Why assume that chat will be the interface? Multimodal and dynamically generated seems more likely.
- TeMPOraL 2y ago
  Voice chat with LLMs is a complete interface, and it's one that already works and can be slotted right into the product. You can prototype voice chat-based ordering app via no-code tools today, and without much effort going into it. Dynamically generated interactive UIs are something people are barely beginning to experiment with; we don't know if current models can do them reliably for realistic problems, and how effort has to go into setting them up for any particular product. At this point, they're an expensive, conceptual solution, that doesn't scale.
thatjoeoverthr 2y ago
Yes. Chat is absolutely bad, because it is opaque. It perfectly reproduces what used to be called "hunt the verb" in gaming, for the same reason. The simple truth is you're interacting with a piece of software, with features and subroutines. GUIs are great at surfacing features, affordances, changing with context. A chat interface invites you to guess.

gordon_freeman 2y ago

I also do not like Chat interface. What I meant by above comment was actually talking and having natural conversations with Operator agent while driving car or just going for a walk or whenever and wherever something comes to my mind which requires me to go to browser and fill out forms etc. That would get us closer to using chatGPT as a universal AI agent to get those things done. (This is what Siri was supposed to be one day when Steve Jobs introduced it on that stage but unfortunately that day never arrived.)

TeMPOraL 2y ago
> This is what Siri was supposed to be one day when Steve Jobs introduced it on that stage but unfortunately that day never arrived. The irony is, the reason neither Siri nor Alexa nor Google Assistant/Now/${whatever they call it these days} nor Cortana achieved this isn't the voice side of the equation. That one sucks too, when you realize that 20 years ago Microsoft Speech API could do better, fully locally, on cheap consumer hardware, but the real problem is the integration approach. Doing interop by agreements between vendors only ever led to commercial entities exposing minimal, trivial functionality of their services, which were activated by voice commands in the form of "{Brand Wake word}, {verb} {Brand 1} to {verb} {Brand 2}" etc. This is not an ergonomic user interface, it's merely making people constantly read ads themselves. "Okay Google, play some Taylor Swift on Spotify" is literally three brand ads in eight words you just spoke out loud. No, all the magical voice experience you describe is enabled[0] by having multimodal LLMs that can be sicced on any website and beat it into submission, whether the website vendor likes it or not. Hopefully they won't screw it up (again[1]) trying to commercialize it by offering third parties control over what LLMs can do. If, in this new reality, I have to utter the word "Spotify" to have my phone start playing music, this is going to be a double regression relative to MS Speech API in the mid 2000s. -- [0] - Actually, it was possible ever since OpenAI added function calling, which was like over a good year ago - if you exposed stuff you care about as functions on your own. As it is, currently the smartphone voice assistant that's closest to Star Trek experience is actually free and easy to set up - it's Home Assistant with its mobile app (for the phone assistant side) and server-side integrations (mostly, but not limited to, IoT hardware). [1] - Like OpenAI did with "GPTs". They've tried to package a system prompt and function call configuration into a digital product and build a marketplace around it. This delayed their release of the functionality to the official ChatGPT app/website for about half a year, leading to an absurd situation where, for those 6+ months, anyone with API access could use a much better implementation of "GPTs" via third-party frontends like TypingMind.

sky2224 2y ago

I think the chat interface is bad, but for certain things it could honestly streamline a lot of mundane things as the poster you're replying two stated. For example, McDonald's has heavily shifted away from cashiers taking orders and instead is using the kiosks to have customers order. The downside of this is 1) it's incredibly unsanitary and 2) customers are so goddamn slow at tapping on that god awful screen. An AI agent could actually take orders with surprisingly good accuracy. Now, whether we want that in the world is a whole different debate.

krapp 2y ago
McDonald's already tried having AI take orders and stopped when the AI did things like randomly add $250 of McNuggets or mistake ketchup for butter. Note - because this is something which needs to be pointed out in any discussion of AI now - even though human beings also make mistakes this is still markedly less accurate than the average human employee.
- ItsMattyG 2y ago
  For now
  - TeMPOraL 2y ago
    Indeed. I think a GPT-4o class model, properly prompted, would work just fine today. The trick is, unlike a human, the computer is free to just say "no" without consequences. The model could be aggressively prompted to detect and refuse weird orders. Having to escalate to a human supervisor (who conveniently is always busy doing other things and will come to you in a minute or three) should be sufficient at discouraging pranksters and fraudsters, while not annoying enough to deter normal customer. (I say model, but for this problem I'd consider a pipeline where the powerful model is just parsing orders and formulating replies, while being sanity-checked by a cheaper model and some old-school logic to detect excessive amounts or unusual combinations. I'd also consider using "open source" model in place of GPT-4o, as open models allow doing "alignment" shenanigans in the latent space, instead of just in the prompts.)
2y ago

csharpminor 2y ago

Why assume that chat will be the interface? Multimodal and dynamically generated seems more likely.

TeMPOraL 2y ago
Voice chat with LLMs is a complete interface, and it's one that already works and can be slotted right into the product. You can prototype voice chat-based ordering app via no-code tools today, and without much effort going into it. Dynamically generated interactive UIs are something people are barely beginning to experiment with; we don't know if current models can do them reliably for realistic problems, and how effort has to go into setting them up for any particular product. At this point, they're an expensive, conceptual solution, that doesn't scale.

thatjoeoverthr 2y ago

Yes. Chat is absolutely bad, because it is opaque. It perfectly reproduces what used to be called "hunt the verb" in gaming, for the same reason. The simple truth is you're interacting with a piece of software, with features and subroutines. GUIs are great at surfacing features, affordances, changing with context. A chat interface invites you to guess.