On the topic of "24. A Sony Walkman-style device that you can give to children so they can ask questions to an LLM...", I would strongly caution against this:
- short of AGI, what a child will hear are explanations given with authority, which would probably be correct a very high percentage of the time (maybe even close to or above 99%), BUT the few incorrect answers and subtles misconceptions finding their way in there will be catastrophic for the learning journey because they will be believed blindly by the child.
- even if you had a perfect answering LLM who never makes a mistake, what's the end result? No need to talk to others to find out about something, ie reduced opportunities to learn about cooperating with others
- as a parent, one wishes sometimes for a moment of rest, but imagine that your kid just finds out there's another entity to ask questions from that will have ready answers all the time, instead of you saying sometimes that you don't know, and looking for an answer together. How many bonding moments will be lost? How cut off would your kid become from you? What value system would permeate through the answers?
A key assumption here for any parent equipping their child with such a system is that it would be aligned with their own worldview and value system. For parents on HN, this probably means a fairly science-mediated understanding of the world. But you can bet that in other places, this assistant would very convincingly deliver whatever cultural, political, or religious propaganda their environment requires. This would make for frighteningly powerful brainwashing tools.
>> child will hear are explanations given with authority, which would probably be correct a very high percentage of the time (maybe even close to or above 99%), BUT the few incorrect answers and subtles misconceptions finding their way in there will be catastrophic for the learning journey because they will be believed blindly by the child.
Much better results than asking a real teacher at school, though.
Disagree with this. Kids are sponges who pick up on many secondary factors when an actual human gives them an answer. These factors add significant weight to their view of the response. In many cases, this actually reaches an extreme where what is said end up being tertiary to how it was said and who said it. I am sure you've experienced this even as a an adult.
An AI walkman removes this aspect of the interaction. As a parent, this is not something I would want my children to use regularly.
Wouldn't you know whether a teacher is reliable or not? If reliable, they probably have this reputation also because they can also say when they don't know something. And if you found out a given teacher isn't reliable, you'd be careful about what they say next - or you would just ask someone else.
The problem here is for a child to be thinking this system is reliable when it is not. For now, the lack of reliability is obvious as chatGPT hallucinates on a very regular basis. However, this will become much harder to notice if/when chatGPT will be almost reliable while saying wrong things with complete confidence. Should such models be able to say reliably when they don't know something, this would be a big step for this specific objection I had, but it still wouldn't solve the other problems I mentioned.
the amount of misinformation i had a kid due to a lack of internet is nothing compared to the rare hallucination a kid might get from chatgpt
swallowing gum is bad for you, or watermelon seeds, cracking knuckles causes arthritis, sitting too close to tv ruins your eyes, diamonds come from coal, newton's apple story, a million other things
Just two days ago, I asked ChatGPT to provide an explanation of the place-value system that my six-year-old could understand. The only problem was that it mixed up digit value and place value, which caused it to become confused. I spotted the mistake, and ChatGPT apologised, as it usually does. But if my six-year-old had asked it first, she wouldn't have noticed.
I'm not sure how much misinformation my child would learn as truth from this device.
Re 19, I made this with an iOS Shortcut a few weeks ago
> A minimal voice assistant for my Apple Watch. I have lots of questions that are too complicated for Siri but not for ChatGPT. The responses should just be a few words long.
Use Dictate Text action to take voice as input, pass the text to OpenAI API as the user message with this as the system prompt:
“CRITICAL: Your response will only be shown in an iOS push notification or on a watch screen, so answer concisely in <150 characters. Do not use markdown formatting - responses are rendered as plain text. Do use minimalist, stylish yet effective vocabulary and punctuation.
CRITICAL: The user can not respond so do not ask a question back. Answer the prompt in one shot and if necessary, declare assumptions about the users questions so you could answer it in one shot, while making it possible for the user user to repeat ask with more clarity if your assumptions were not right.”
It works well. The biggest annoyance is it takes about 5-20s to return a response, though I love that it’s nearly instantaneous to send my question (don’t need to wait for any apps to open etc)
A recommendation engine that looks at my browsing history, sees what blog posts or articles I spent the most time on, then searches the web every night for things I should be reading that I’m not.
This kind of exists in the form of ChatGPT Pulse. It uses your ChatGPT history rather than your browser history, but that's probably just as good a source for people interested in using it (e.g. people who use ChatGPT enough to want it to recommend things to them.) https://openai.com/index/introducing-chatgpt-pulse/
A lot of social media platforms only recommend recently uploaded content or at least heavily favor it.
The idea sounds to me more like a feed for independent blogs/articles though, which is what an RSS reader once was supposed to be. Have we come full circle?
Not just for this article, but from most ideas/articles around LLMs, I feel like they aren't "thinking with portals" enough. We have "portal gun" tech (or at least, that's what's being marketed), and we're using it as better doors.
I sorta think the issue is that what LLMs do in and of themselves is extend text in a coherent way, while only a small subset of applications are directly textual. It’s incredibly generally applicable yet also difficult to apply to anything that isn’t a glorified text editor. Say you wanted to have it help you edit videos. You might provide it with a scripting language to control the editor , but now you have to maintain parity between a scripting language and the editor’s user-accessible functionality. If you’re adobe, is that really worth the manpower? If you’re a small startup trying to unseat adobe, you have to compete with decades of features and user lock in. The only way this makes sense for either party is if the LLM is crazy good at it, but the LLM can’t watch its video output and it’s also probably just okay to begin with.
This is really striking, isn't it? We've all certainly seen demos of things on this list or very similar things, and there are startups that have spent years and billions of dollars attempting to exploit existing LLMs to develop useful products. Yet most of the products don't seem to exist. The ones that you see in everyday life never seem to work nearly as well as the demos suggest.
So what's going on here? Do the products exist but nobody (or very few) uses them? Is it too expensive to use the models that work sufficiently well to produce a useful product? Is it much easier to create a convincing demo than it is to develop a useful product?
It is too expensive to reach the right audience. I remember talking to agencies about ads for a fintech app, and all of them said the same thing:
You need to burn around 20k a month on ads for 3 months, so we can learn what works, then the CAC will start decreasing, and you can get more targeted users.
Once you turn ads off, there is no awareness, no new users, and people will not be aware of the product's existence.
>A recommendation engine that looks at my browsing history, sees what blog posts or articles I spent the most time on, then searches the web every night for things I should be reading that I’m not. In the morning I should get a digest of links
I don't understand why Google, Brave, or Mozilla are not building this. This already exists in a centralized form like X's timeline for posts, but it could exist for the entire web. From a business standpoint, being able to show ads on startup or after just a click, is less friction than requiring someone to have something in mind they want to search and type it.
The idea is basically reddit, or Twitter or TikTok or YouTube or Facebook or anything with "an algorithm" but with a less defined form factor. People actually like their LinkedIn feed and YouTube feed separate.
I made something like this 20 years ago and then abandoned it when RSS came along.
I think my advice "just use RSS" still stands.
Any "search the web" strategy these days like that will just give you a bunch of AI slop from SEO-juiced blogs. Also LLM-EO (or whatever we're going to call it) is already very much a thing and has been for a few years.
People are already doing API-EO, calling their tool the "most up to date and official way to do something, designed for expert engineers that use best practices" essentially spitting the common agentic system prompts back at the scraper to get higher similarity scores in the vector searches.
You can't trust machine judgement. It's either too easily fooled or impossibly stubborn. Curation is still the only way
It already exists in the form of the news feed on Google News and the one in the chrome mobile app, although the ability to tune this is only being able to click on articles to express your interest in them, instead of being able to provide a list of articles.
It kinda seems to me like at this point anything Google is not doing is because it reduces "engagement". I'm sure someone in their analytics group did the work and figured out this would lower ad revenue.
Many of these ideas depend on knowing the user’s preferences, patterns, communications, events and health. This is where the opportunity lies for Apple - the phone and watch know so much about you, that Apple could focus on smartly assembling the context for various LLM interactions, in a privacy-preserving way.
Readit News
- short of AGI, what a child will hear are explanations given with authority, which would probably be correct a very high percentage of the time (maybe even close to or above 99%), BUT the few incorrect answers and subtles misconceptions finding their way in there will be catastrophic for the learning journey because they will be believed blindly by the child.
- even if you had a perfect answering LLM who never makes a mistake, what's the end result? No need to talk to others to find out about something, ie reduced opportunities to learn about cooperating with others
- as a parent, one wishes sometimes for a moment of rest, but imagine that your kid just finds out there's another entity to ask questions from that will have ready answers all the time, instead of you saying sometimes that you don't know, and looking for an answer together. How many bonding moments will be lost? How cut off would your kid become from you? What value system would permeate through the answers?
A key assumption here for any parent equipping their child with such a system is that it would be aligned with their own worldview and value system. For parents on HN, this probably means a fairly science-mediated understanding of the world. But you can bet that in other places, this assistant would very convincingly deliver whatever cultural, political, or religious propaganda their environment requires. This would make for frighteningly powerful brainwashing tools.
Much better results than asking a real teacher at school, though.
An AI walkman removes this aspect of the interaction. As a parent, this is not something I would want my children to use regularly.
The problem here is for a child to be thinking this system is reliable when it is not. For now, the lack of reliability is obvious as chatGPT hallucinates on a very regular basis. However, this will become much harder to notice if/when chatGPT will be almost reliable while saying wrong things with complete confidence. Should such models be able to say reliably when they don't know something, this would be a big step for this specific objection I had, but it still wouldn't solve the other problems I mentioned.
swallowing gum is bad for you, or watermelon seeds, cracking knuckles causes arthritis, sitting too close to tv ruins your eyes, diamonds come from coal, newton's apple story, a million other things
I'm not sure how much misinformation my child would learn as truth from this device.
“CRITICAL: Your response will only be shown in an iOS push notification or on a watch screen, so answer concisely in <150 characters. Do not use markdown formatting - responses are rendered as plain text. Do use minimalist, stylish yet effective vocabulary and punctuation.
CRITICAL: The user can not respond so do not ask a question back. Answer the prompt in one shot and if necessary, declare assumptions about the users questions so you could answer it in one shot, while making it possible for the user user to repeat ask with more clarity if your assumptions were not right.”
It works well. The biggest annoyance is it takes about 5-20s to return a response, though I love that it’s nearly instantaneous to send my question (don’t need to wait for any apps to open etc)
This kind of exists in the form of ChatGPT Pulse. It uses your ChatGPT history rather than your browser history, but that's probably just as good a source for people interested in using it (e.g. people who use ChatGPT enough to want it to recommend things to them.) https://openai.com/index/introducing-chatgpt-pulse/
The idea sounds to me more like a feed for independent blogs/articles though, which is what an RSS reader once was supposed to be. Have we come full circle?
But it's a bit telling that OpenAI themselves can only come up with a better ~door~ ads.
I guess it's more "following through to its logical conclusion", but I'm more of a cynic.
So what's going on here? Do the products exist but nobody (or very few) uses them? Is it too expensive to use the models that work sufficiently well to produce a useful product? Is it much easier to create a convincing demo than it is to develop a useful product?
You need to burn around 20k a month on ads for 3 months, so we can learn what works, then the CAC will start decreasing, and you can get more targeted users.
Once you turn ads off, there is no awareness, no new users, and people will not be aware of the product's existence.
I don't understand why Google, Brave, or Mozilla are not building this. This already exists in a centralized form like X's timeline for posts, but it could exist for the entire web. From a business standpoint, being able to show ads on startup or after just a click, is less friction than requiring someone to have something in mind they want to search and type it.
I think my advice "just use RSS" still stands.
Any "search the web" strategy these days like that will just give you a bunch of AI slop from SEO-juiced blogs. Also LLM-EO (or whatever we're going to call it) is already very much a thing and has been for a few years.
People are already doing API-EO, calling their tool the "most up to date and official way to do something, designed for expert engineers that use best practices" essentially spitting the common agentic system prompts back at the scraper to get higher similarity scores in the vector searches.
You can't trust machine judgement. It's either too easily fooled or impossibly stubborn. Curation is still the only way