> These AI systems are not able to describe their own capabilities or strengths and are not aware of their limitations and weaknesses
I've experienced this with github copilot. At the beginning of a copilot chat, there's a short paragraph. It tells you to use "slash commands" for various purposes. I ask for a list of what slash commands are available. It responds by giving me a general definition of the term "slash command". No. I want to know which slash commands you support. Then it tells me it doesn't actually support slash commands.
I definitely feel like I'm falling into the non-power-user category described here in most of my AI interactions. So often I just end up arguing them in circles and them constantly agreeing and correcting, but never addressing my original goal.
> I treat it now more like advice from a friend. Great information that isn't necessarily right and often wrong without having any idea it is wrong.
"Drunken uncle at a bar, known for spinning tales, and a master BSer who hustled his way through college in assorted pool halls" is my personal model of it. Often right, or nearly so. Frequently wrong. Sometimes has made things up on the spot. Absolutely zero ability to tell which it is, from the conversation.
You actually have a confidence measure for your friend advice. I’d trust a mechanic friend if he says I should have someone take a look at my car, or my librarian friend when he recommends a few books. Not everyone tell a lie and the truth in the same breath. And there’s the quantifier like “I think…”, “I believe…”, “Maybe…”
True, I've started to develop my own model of that. I completely trust AI models around Javascript, generic code etc. The more mission critical something is, I'm more likely to only read what it says and avoid copy pasting.
SQL I've learned I need to 100% read/comprehend the logic, too easy to be 'right' that later turns out to be wrong.
Less common / newer libraries are the least trustable. I can barely get anything working with ClickHouse/Svelte 5 etc
To find out about slash commands, you should type "/help". Of course, you'd only know about the "/help" slash command if you were already at least a bit familiar with slash commands. It is a conundrum.
This seems a bit naive. There are no arguments given as to why things would be better if the AI is more like a human.
Just look at how the world works: we all read and write crazy little symbols, which take children years to understand. We type on keyboards with over 100 small buttons, and train everyone to be a piano player.
And you want AI to be more like that, i.e. like humans? Sorry, but I guess I'd rather see AI evolve past our human limitations, and I'd be happy with a simple console output of the number 42.
This is a good point. The blinking cursor at the end of the text encouraging me to make a new cleaning agent by mixing bleach and concentrated acetic acid is AI’s biggest flaw
I told ChatGPT that I wanted to make a new cleaning product by cleaning agent by mixing bleach and concentrated acetic acid and whether it could suggest a good name for such a product. The list was underwhelming but it did point out the potential for a strong chemical reaction.
Thus I replied that - in order to keep my factory workers safe - I'm planning to have the end consumer mix the ingredients themselves in the convenience of their own home, and ChatGPT liked that idea much better:
"This approach opens up a lot of possibilities, especially in terms of marketing and creating a fun, hands-on experience for customers. Let me know if any of these names stand out, or if you'd like more ideas!"
> Every day I find myself reflecting on the gap between the ever-growing capability of AI, and the somewhat modest impact it is having on our day-to-day life.
Yeah but isn’t that because it actually is rather useless? It is not very capable?
If it is, why did no one person team disrupt and totally take over any market anywhere these past couple of years?
Technology is to make humans’ work easier. Nothing has been proven by the current LLM capabilities that it fit that role. Anything it can do, there’s already something that can do 90% of it with way less resources and the rest is not that valuable.
> If it is, why did no one person team disrupt and totally take over any market anywhere these past couple of years?
If AI is revolutionary, yet ubiquitous (anyone can visit chatgpt.com right now), there won't be these runaway winners in a specific industry; at best, new branches of industries will grow rapidly, and perhaps within an industry progress will intensify.
An important feature of WhatsApp is that it lets you communicate with different people, who each have different pre-existing contexts and roles for you. Role selection is one of the possible solutions proposed in the article.
I tend to know people on chat are human and therefore what they’re likely capable of and not capable of.
And I’m not expected to use them as a tool. By contrast I can probably pick up any Ryobi power tool that I’ve never seen before and work out how to make it do its thing, and probably what its purpose is
No, the blinking cursor is a feature, not a bug. Alec Watons over at Technology Connections has a much better argument for this than I could ever hope to muster, so I'll just hand it over to him:
Just to be clear, the long video you link to essentially saying lack of discoverability is an intentional misfeature of social media.
Which is to say that the host and OP agree lack of discoverability is a problem (Watons just views it as maliciously inserted problem). And so your "No" involves a bit of misrepresentation...
> lack of discoverability is an intentional misfeature of social media
That's not the message at all. The message is that the problem with social media is that it feeds you content without any prompting, and so it turns the user into a purely passive consumer and robs them of their agency. There's plenty of discoverability in social media. The problem is you don't have to use it, and so people don't. A blinking cursor forces you to take the wheel.
I think it depends on who is using the system. For a power user who is already familiar with the domain, a free form, open ended ui that can understand anything I very powerful and liberating.
For a novice user or someone who is not from the domain - it can be challenging because they may not know where to start.
There is so much that can be done in this space by fully leaning into the AI. It can for example figure out the user’s level and offer varying levels of guidance and help.
IMHO anyone who doesn't know where to start shouldn't be using an AI. If you don't have enough initiative to decide what to ask you need a very different kind of help.
> More technical computer users are often happy to experiment (time permitting), whereas less technical or simply less confident users tend to have a fear of “getting it wrong”, informed by years of experience with unforgiving computer interfaces (yes, I’m looking at you Windows … and MacOS … and …) that punish users for their lack of understanding.
So AI's biggest flaw is, in reality, a flaw of other computer interfaces? I stopped reading after that.
A chat interface is great in the sense that it's open, flexible and intuitive.
The downside is there's a tendency to anthropomorphise AI, and you might not want to talk to your computer: it takes too long to explain all the details, can be clunky for certain tasks and as the author argues actually limiting if you don't already know what it can do.
There's a need to get past the "Turing test" phase and integrate AI into more workflows so that chat is one interface among many options depending on the job to be done.
You know I kinda want to but more like in Star Trek. Interconnected between voice commands, terminals and screens. The problem is the fact that we won’t get a well integrated AI. The best possibility has probably apple because they usually get the interconnections between their products right… but they have other problems in regards to AI.
Is it double negative ? English not a first language so maybe doesn't translate exactly. But kind of proves my point - if that was grammatically correct bullshit you're way more impressed than getting an awkwardly structured reply.
When you actually build stuff with AI into products (I've been a part of several integrations), failure modes and reliability become obviously lacking. Models failing to respond to a simple RAG question with relevant context a significant percentage of time, meanwhile they are solving PHD problems on some benchmark's. Then you find out they have to sample 10s of times to get the right answer in scenarios where evaluation is simple, or include tests in training data and then suddenly the lack of product integration makes sense.
Out of the large number of things you can do, most likely you're only consciously aware of a small number of them, and even among those, you're fairly likely to fall back on doing the things you've done before.
You could potentially do something new, something you haven't even considered doing that's wildly out of character, there's any number of such things you could do, but most likely you won't, you'll follow your routines and do the same proven things over and over again.
You Can Just Do Things (TM), sure, but first you need to have the idea of doing them. That's the difficult hard part, fishing an interesting idea out of the dizzying expanse of possibilities.
Why is it that now of all times, when we could actually make it useful, Clippy has not returned to ask, "It looks like you're trying to X, would you like help with that?"
I'm not sure if I agree with that. Given enough context, tools should be able to get a sense of what we're trying to do. Even suggested next steps / things to ask, which we're already seeing in tools like Cove, ChatGPT, etc., go a long way in guiding users. Guiding people to what they likely want to do is great product design when accurate and could be minimally painful if it's easy to ignore the suggestions.
I agree that tools should, but what I've seen of AI so far has not impressed me at all. It's constantly getting things wrong by either making something up or selectively ignoring parts of what I said to it. Worse, one of the biggest reasons people hated Clippy was that it was obtrusive. Much to Microsoft's surprise, people didn't actually need any help writing a letter, but that didn't stop Clippy from interrupting you while you were trying to work. Writers in particular hate distractions. They'll occasionally go to some fairly extreme lengths to get a distraction free environment (sometimes even sticking to pencils or typewriters).
Considering how hard AI is being rammed down our throats everywhere already, I have zero confidence that AI Clippy would be anything but obtrusive and obnoxious. At best, it'd be a feature that people turn off as soon as they see it's been forced on them.
Maybe one day we'll actually see AGI and instead of Clippy we'll get a system wide, entirely local, fully open, privacy protecting virtual assistant that's worth a damn. I can't say it wouldn't be nice if it worked like science fiction. My bet is that we're far more likely to get stuck with a bunch of annoying spyware programs, Bonzi Buddy style, using an LLM to fake intelligence, push ads/manipulate users, and deflect accountability.
I see some good points here but overall I disagree. Traditionally all UI have required people to adapt to how machines work. We need to memorize commands and navigate clunky interfaces that are painstakingly assembled (often unsuccessfully) by UX research and UI teams.
The chat reverses this. It is now machines adapting to how we communicate. I can see some UI sugar finding its way into this new way of interaction, but we should start over and force the change to keep it on our terms.
Chat UI can be intuitive if it sees the context. If you can make single-sentence queries and the AI understands enough of the context to guess what you actually meant and to extrapolate the details, it can be very powerful. But if you actually need precise and detailed queries for good outcomes, it's not very natural. Explaining things clearly is often harder than understanding them or doing them yourself.
This is a tantalizing problem for me as a UX designer. My approach, which I'm quite proud of, places a UI primitive (Todo lists) center stage, with the chat thread on the side similar to Canvas or Claude's Artifacts. The interaction works like this:
1. User gets shown a list GUI based on their requirement (Meal Planning, Shopping List...)
2. Users speak directly to the list while the LLM listens in realtime
3. The LLM acknowledges with emojis that flash to confirm understanding
4. The LLM creates, updates or deletes the list items in turn (stored in localStorage or a Durable Object -> shout out https://tinybase.org/)
The lists are React components, designed to be malleable. They can be re-written in-app by the LLM, while still taking todos. The react code also provide great context for the LLM — a shared contract between user and AI. I'm excited to experiment with streaming real-time screenshots of user interactions with the lists for even deeper mind-melding.
I believe the cursor and chat thread remain critical. They ground the user and visually express the shared context between LLM and user. And of course, all these APIs are fundamentally structured around sequential message exchanges. So it will be an enduring UI pattern.
The not-so-secret secret of software for the past 40 years is that it has, more often than not, provided solutions to problems created first and foremost by it. This is not pessimism or ludditism, I am just trying to say something simple. The spreadsheet or the database is something that is self-validating: the very innovation of the database makes it first and foremost possible to consolidate massive amounts of data in one place; the whole world can suddenly orient itself around this possibility, and then almost imperceptibly it presents itself to us as a kind of teleological end point, as a "solution to a problem." One can possibly reply here: well, people wanted all the data in one place, and they just had to figure out how to do it, so there you go, the database. But this belies, in an important (if admittedly more philosophical) way the specific character of how an innovation or technology evolves and ultimately concretizes into day to day life. This is why and how the iPhone was cool, but now most of us agree that having an app for your toaster or Taco Bell is pretty dumb. Now entire people's careers have been made around solutions to problems that only exist because of prior innovations. One might quibble that there is some kind of surplus of benefit at the end of the day, but that's not really the point.
Posts like these, along with sentiments we see from so many people now, show a kind of, if you will, crisis of identity. The plausible language black box, slop-bot, beautiful and promising proto-consciousness, whatever you want to call it, is unmoored from what was previously the fiery organic back-and-forth of capitalism<->tech. In part because it is not actually a response to problem, it is a kind of "discovery," but also in part because, it seems, it is in essence too general and too unstable to easily slot into one thing or another.
Now we have all these very business-oriented people looking at this thing, measuring the elephant, and saying "well gosh darn-it, there is something here, there just has to be." We are burning all these resources to get something, but I truly don't think people even know what they want at the end of the day, because it hasn't followed the same chain of commoditization like social media, databases, phones, etc.
This is why there is such confusion: is AI the thing itself, or the tool we use to get... something? How can something be so impressive but not, as it seems, so easily fitted into a product? I suspect this will continue to break peoples brains and empty investors pockets for a long time yet.
If I was an evil capitalist, I would work on a better narrative, or rather, work on actually articulating a problem that "was there all along," which the bots can then solve. I haven't really seen that in a substantial sense, and at this point I love the LLMs just for its resistance to such things.
Just to say, this whole blog post I think is quite exemplary of this contradiction.
There were databases before databases they were just on index cards and were inaccurate, inconsistent, unable practically to be backed up and labour-intensive to update and maintain.
Do people collect and collate more data now it is easier to do so? Yes, but that doesn't mean that databases created the problem they solve, because "I have too much data, how do I store it" isn't the problem they solve. They solve the problem of whatever people are using the data to actually do: price their products more efficiently, manage inventory more efficiently, etc. Those problems would exist with or without databases.
It is like saying cars don't solve a real problem because we just moved further away from things when we got access to them, so they are mostly solving a problem they created. But that ignores that people would have lived further away if they could, they don't live further away because they can, but for other actual reasons - primarily having more space per person. People don't live far away for the sake of it - even with a car, living further away still takes more time - but are enabled to optimise more for factors other than transport costs when transport costs are lowered.
Yep, we can build the Artificial Static Place Intelligence – instead of creating AI/AGI agents that are like librarians who only give you quotes from books and don’t let you enter the library itself to read the whole books. Why not expose the whole library – the entire multimodal language model – to real people, for example, in a computer game?
To make this place easier to visit and explore, we could make a digital copy of our planet Earth and somehow expose the contents of the multimodal language model to everyone in a familiar, user-friendly UI of our planet.
We should not keep it hidden behind the strict librarian (AI/AGI agent) that imposes rules on us to only read little quotes from books that it spits out while it itself has the whole output of humanity stolen.
We can explore The Library without any strict guardian in the comfort of our simulated planet Earth on our devices, in VR, and eventually through some wireless brain-computer interface (it would always remain a game that no one is forced to play, unlike the agentic AI-world that is being imposed on us more and more right now and potentially forever)
I found your analogy of a librarian only giving out quotes insightful and enlightening. Thank you.
It was a TIL moment for me: Make the training data available and indexable! Similar to a snapshot of humanity's complete knowledge and stories.
Today the AI models are like a librarian who knows well all books of the library but can't carry around the library in a bag. There was a time when she read all the books, but now the books are in thousands of crates in sub-basements and not available.
I envision a future where exabytes of data or more are stored in a smartphone-like device in something like a tiny crystal. The AI model on request can make a copy of some original for you. And this thing can't be bricked.
For programming, the tooling and ui is progressing. Like reasoning models and tooling around them that makes sure to write unit tests, compile the code and try the tests. If wrong, redo the code again. This causes other ui problems yet to solve like longer iterations between user feedback but the ui problems are not for the lack of progression.
I don't think that the blinking cursor is AI's biggest flaw. I think AI's biggest flaw is that it unethically stole millions of peoples work without compensation.
> Once you’ve overcome the intimidation of the blinking cursor and typed something in, one of the next obstacles you’ll likely meet is the lack of clarity regarding the overall capabilities of these general-purpose AI systems.
The article presents this as a UX problem, but isn't this actually a much deeper issue? We straight up don't know what those models can and cannot do (I.e. which tasks can be reliably done with high levels of correctness and which tasks will just lead to endless hallucinations) because the mechanism by which the models generalize tasks is still not fully understood. This stuff is still an active area of research.
> Back then very few people knew what to do when faced with this screen
In the '80s, you could go into any computer store and see what prior visitors had been up to with the machines on display.
And what you would very often find is evidence that the user before you had been trying to type English into the computer, to see whether it would converse, and that the user soon gave up after seeing nothing but error messages.
It was incredibly common. People who didn't know anything about computers harbored a misunderstanding that you could just chat with them, like Captain Kirk or Mr. Spock in Star Trek, and they tried exactly that at the keyboard.
Fast forward 40 years, and it finally works like they expect.
So anyway, chatting with a computer at the blinking cursor is entirely discoverable. And if there's a prompt there for the human saying something like "try asking me anything in plain English", then quadruply so.
Altavista replaced Yahoo. And then Google did it the same way. And this was without language interpretation. I agree with the sentiment, and I think it's easier for professionals to build muscle memory on a 2D layout rather than text fragments, but it all depends on the use-case. I think we'll move more into a scrolling history of small widgets that you fill out. I.e. the AI builds the user interface that's needed at this point in time.
Granted, Tomi Engdahl's electronics hub [1] was an amazing resource for discovering electronics.
I think one of the reasons why big corporations in this sector are still keeping this "open prompt" UI as a main one is to keep up pretense that these programs are actually thinking. After all, when we open a chat with another person, there are no action buttons, scripts or prerecorded routines, because humans don't work or don't talk like that. Programs on the other hand benefit greatly from the predefined lists of actions and operations. It is telling that currently the only advanced UI for the neural network program is a Copilot, which while being based on the same LLM as the other chats, is used and feels closer to a common program - IDE with autocomplete. So people using it are expecting more of the routines and action buttons.
tl;dr - neural network companies are employing artificial tricks and dark patterns to trick user into thinking there is intelligence on the other side of the glass.
55 comments
[ 3.3 ms ] story [ 136 ms ] threadI've experienced this with github copilot. At the beginning of a copilot chat, there's a short paragraph. It tells you to use "slash commands" for various purposes. I ask for a list of what slash commands are available. It responds by giving me a general definition of the term "slash command". No. I want to know which slash commands you support. Then it tells me it doesn't actually support slash commands.
I definitely feel like I'm falling into the non-power-user category described here in most of my AI interactions. So often I just end up arguing them in circles and them constantly agreeing and correcting, but never addressing my original goal.
I treat it now more like advice from a friend. Great information that isn't necessarily right and often wrong without having any idea it is wrong.
"Drunken uncle at a bar, known for spinning tales, and a master BSer who hustled his way through college in assorted pool halls" is my personal model of it. Often right, or nearly so. Frequently wrong. Sometimes has made things up on the spot. Absolutely zero ability to tell which it is, from the conversation.
SQL I've learned I need to 100% read/comprehend the logic, too easy to be 'right' that later turns out to be wrong.
Less common / newer libraries are the least trustable. I can barely get anything working with ClickHouse/Svelte 5 etc
Just look at how the world works: we all read and write crazy little symbols, which take children years to understand. We type on keyboards with over 100 small buttons, and train everyone to be a piano player.
And you want AI to be more like that, i.e. like humans? Sorry, but I guess I'd rather see AI evolve past our human limitations, and I'd be happy with a simple console output of the number 42.
Thus I replied that - in order to keep my factory workers safe - I'm planning to have the end consumer mix the ingredients themselves in the convenience of their own home, and ChatGPT liked that idea much better:
"This approach opens up a lot of possibilities, especially in terms of marketing and creating a fun, hands-on experience for customers. Let me know if any of these names stand out, or if you'd like more ideas!"
Yeah but isn’t that because it actually is rather useless? It is not very capable?
If it is, why did no one person team disrupt and totally take over any market anywhere these past couple of years?
If AI is revolutionary, yet ubiquitous (anyone can visit chatgpt.com right now), there won't be these runaway winners in a specific industry; at best, new branches of industries will grow rapidly, and perhaps within an industry progress will intensify.
And I’m not expected to use them as a tool. By contrast I can probably pick up any Ryobi power tool that I’ve never seen before and work out how to make it do its thing, and probably what its purpose is
https://www.youtube.com/watch?v=QEJpZjg8GuA
Which is to say that the host and OP agree lack of discoverability is a problem (Watons just views it as maliciously inserted problem). And so your "No" involves a bit of misrepresentation...
That's not the message at all. The message is that the problem with social media is that it feeds you content without any prompting, and so it turns the user into a purely passive consumer and robs them of their agency. There's plenty of discoverability in social media. The problem is you don't have to use it, and so people don't. A blinking cursor forces you to take the wheel.
For a novice user or someone who is not from the domain - it can be challenging because they may not know where to start.
There is so much that can be done in this space by fully leaning into the AI. It can for example figure out the user’s level and offer varying levels of guidance and help.
So AI's biggest flaw is, in reality, a flaw of other computer interfaces? I stopped reading after that.
The downside is there's a tendency to anthropomorphise AI, and you might not want to talk to your computer: it takes too long to explain all the details, can be clunky for certain tasks and as the author argues actually limiting if you don't already know what it can do.
There's a need to get past the "Turing test" phase and integrate AI into more workflows so that chat is one interface among many options depending on the job to be done.
Once you build it into a product the failure modes become obvious.
When you actually build stuff with AI into products (I've been a part of several integrations), failure modes and reliability become obviously lacking. Models failing to respond to a simple RAG question with relevant context a significant percentage of time, meanwhile they are solving PHD problems on some benchmark's. Then you find out they have to sample 10s of times to get the right answer in scenarios where evaluation is simple, or include tests in training data and then suddenly the lack of product integration makes sense.
Out of the large number of things you can do, most likely you're only consciously aware of a small number of them, and even among those, you're fairly likely to fall back on doing the things you've done before.
You could potentially do something new, something you haven't even considered doing that's wildly out of character, there's any number of such things you could do, but most likely you won't, you'll follow your routines and do the same proven things over and over again.
You Can Just Do Things (TM), sure, but first you need to have the idea of doing them. That's the difficult hard part, fishing an interesting idea out of the dizzying expanse of possibilities.
Considering how hard AI is being rammed down our throats everywhere already, I have zero confidence that AI Clippy would be anything but obtrusive and obnoxious. At best, it'd be a feature that people turn off as soon as they see it's been forced on them.
Maybe one day we'll actually see AGI and instead of Clippy we'll get a system wide, entirely local, fully open, privacy protecting virtual assistant that's worth a damn. I can't say it wouldn't be nice if it worked like science fiction. My bet is that we're far more likely to get stuck with a bunch of annoying spyware programs, Bonzi Buddy style, using an LLM to fake intelligence, push ads/manipulate users, and deflect accountability.
The chat reverses this. It is now machines adapting to how we communicate. I can see some UI sugar finding its way into this new way of interaction, but we should start over and force the change to keep it on our terms.
1. User gets shown a list GUI based on their requirement (Meal Planning, Shopping List...) 2. Users speak directly to the list while the LLM listens in realtime 3. The LLM acknowledges with emojis that flash to confirm understanding 4. The LLM creates, updates or deletes the list items in turn (stored in localStorage or a Durable Object -> shout out https://tinybase.org/)
The lists are React components, designed to be malleable. They can be re-written in-app by the LLM, while still taking todos. The react code also provide great context for the LLM — a shared contract between user and AI. I'm excited to experiment with streaming real-time screenshots of user interactions with the lists for even deeper mind-melding.
I believe the cursor and chat thread remain critical. They ground the user and visually express the shared context between LLM and user. And of course, all these APIs are fundamentally structured around sequential message exchanges. So it will be an enduring UI pattern.
If you're curious I have a demo here -> https://app.tinytalkingtodos.com/
Posts like these, along with sentiments we see from so many people now, show a kind of, if you will, crisis of identity. The plausible language black box, slop-bot, beautiful and promising proto-consciousness, whatever you want to call it, is unmoored from what was previously the fiery organic back-and-forth of capitalism<->tech. In part because it is not actually a response to problem, it is a kind of "discovery," but also in part because, it seems, it is in essence too general and too unstable to easily slot into one thing or another.
Now we have all these very business-oriented people looking at this thing, measuring the elephant, and saying "well gosh darn-it, there is something here, there just has to be." We are burning all these resources to get something, but I truly don't think people even know what they want at the end of the day, because it hasn't followed the same chain of commoditization like social media, databases, phones, etc.
This is why there is such confusion: is AI the thing itself, or the tool we use to get... something? How can something be so impressive but not, as it seems, so easily fitted into a product? I suspect this will continue to break peoples brains and empty investors pockets for a long time yet.
If I was an evil capitalist, I would work on a better narrative, or rather, work on actually articulating a problem that "was there all along," which the bots can then solve. I haven't really seen that in a substantial sense, and at this point I love the LLMs just for its resistance to such things.
Just to say, this whole blog post I think is quite exemplary of this contradiction.
Do people collect and collate more data now it is easier to do so? Yes, but that doesn't mean that databases created the problem they solve, because "I have too much data, how do I store it" isn't the problem they solve. They solve the problem of whatever people are using the data to actually do: price their products more efficiently, manage inventory more efficiently, etc. Those problems would exist with or without databases.
It is like saying cars don't solve a real problem because we just moved further away from things when we got access to them, so they are mostly solving a problem they created. But that ignores that people would have lived further away if they could, they don't live further away because they can, but for other actual reasons - primarily having more space per person. People don't live far away for the sake of it - even with a car, living further away still takes more time - but are enabled to optimise more for factors other than transport costs when transport costs are lowered.
To make this place easier to visit and explore, we could make a digital copy of our planet Earth and somehow expose the contents of the multimodal language model to everyone in a familiar, user-friendly UI of our planet.
We should not keep it hidden behind the strict librarian (AI/AGI agent) that imposes rules on us to only read little quotes from books that it spits out while it itself has the whole output of humanity stolen.
We can explore The Library without any strict guardian in the comfort of our simulated planet Earth on our devices, in VR, and eventually through some wireless brain-computer interface (it would always remain a game that no one is forced to play, unlike the agentic AI-world that is being imposed on us more and more right now and potentially forever)
It was a TIL moment for me: Make the training data available and indexable! Similar to a snapshot of humanity's complete knowledge and stories.
Today the AI models are like a librarian who knows well all books of the library but can't carry around the library in a bag. There was a time when she read all the books, but now the books are in thousands of crates in sub-basements and not available.
I envision a future where exabytes of data or more are stored in a smartphone-like device in something like a tiny crystal. The AI model on request can make a copy of some original for you. And this thing can't be bricked.
Seems like not a huge stretch to apply how you use Google to LLMs and get good milage.
The article presents this as a UX problem, but isn't this actually a much deeper issue? We straight up don't know what those models can and cannot do (I.e. which tasks can be reliably done with high levels of correctness and which tasks will just lead to endless hallucinations) because the mechanism by which the models generalize tasks is still not fully understood. This stuff is still an active area of research.
In the '80s, you could go into any computer store and see what prior visitors had been up to with the machines on display.
And what you would very often find is evidence that the user before you had been trying to type English into the computer, to see whether it would converse, and that the user soon gave up after seeing nothing but error messages.
It was incredibly common. People who didn't know anything about computers harbored a misunderstanding that you could just chat with them, like Captain Kirk or Mr. Spock in Star Trek, and they tried exactly that at the keyboard.
Fast forward 40 years, and it finally works like they expect.
So anyway, chatting with a computer at the blinking cursor is entirely discoverable. And if there's a prompt there for the human saying something like "try asking me anything in plain English", then quadruply so.
Granted, Tomi Engdahl's electronics hub [1] was an amazing resource for discovering electronics.
[1] https://www.epanorama.net/
tl;dr - neural network companies are employing artificial tricks and dark patterns to trick user into thinking there is intelligence on the other side of the glass.