That's pretty much my take. LLMs aren't a bad idea, they are useful, in certain fields, but they aren't living up to the sales pitch and they are to expensive to run.
My personal take is that the whole chat based interface is a mistake. I have no better solution, but for anything beyond a halluciating search engine, it's not really the way we need to interact with AI.
In my mind we're 6 months away from one of the biggest crashes in tech history.
Agreed. Text is used for a lot of things. A fantastic text parser/generator that doesn't need regex and can extract /meaning/ would have been a sci-fi fever dream even a decade ago. So, LLMs will definitely have their use and will probably disrupt several industries.
But this hype-storm just reminds me of the fever-dream blogs about the brave new world of the Internet back when hypertext became widely used in '93 or so (direct democracy, infinite commerce, etc, etc). Yes, of course, the brave new world came along, but it needed 3G and multi-touch screens as well and that was 15 years later and a whole different set of companies made money and ruled the world than those that bet on hypertext.
The AI hype also made voice to text and text to voice a lot better.
It's not just chat. But I really enjoy the Claude or chatgpt chat UI.
I'm not sure what your problem with that UI is tbh?
I don't think there will be any crash soon. I'm hoping for lower prices but I can't imagine not having these tools available anymore.
I'm also waiting for better and cheaper support for local development and inferencing.
I'm waiting left and right for a lot of smaller things. Like cheaper Claude, Hardware LLM chips, secure integration of my bank account and other personal data. More capacity for the beta projects of coding agents.
It would feel like a regression to not have access to such tools, models like Claude are really useful on a day-to-day basis (not only for programming, but for any question you have in life where you would typically use Google and read yourself a lot of information).
We don't need AGI to have increased productivity and benefits.
If it would be _that_ useless, half a billion people wouldn't use such tools every week (and this is not even counting Chinese users).
I think many hate LLMs because they realize that it starts to eat their bread and butter and are already taking their job and find it difficult to accept.
Investors are over-paying, because they are betting on companies that can be outpaced by others, but the technology itself is here to stay and grow, until it can do more and more tasks that are done by humans.
It doesn't mean that LLMs and other AI tools are going to run completely alone anytime soon (maybe with cars first ?), but eventually it could be.
What will be the world in 10 or 20 years from now ?
> I'm not sure what your problem with that UI is tbh?
It's not a particularly efficient way to interact with the LLMs. Despite how people hate AI being integrated into everything, I do think that's the only way it can be done effectively. My take is that the AI/LLM isn't the product, it's just a feature of a product. Like when there's "AI" in the camera apps on your phone, or some sort of machine learning in medical imaging. To get the full benefit it needs to be a technology built into your product, not a product in it self.
The voice to text, text to voice is a good example. Voice to text is a feature of a transcription or video conference system. Text to voice is an accessibility feature. It doesn't matter that it's powered by an LLM.
If Claude, ChatGPT, whatever do become cheaper the market will crash. My thinking is: Prices go down (they will eventually), if that happens to fast, OpenAI will tank, their valuation is to high to support a $20 a month AI offering, or worse a world where ChatGPT isn't a product, but a feature. If OpenAI goes down the entire industry will have to re-align to a world where we now know that there's a limit to the value an LLM can command. Most of the current AI companies will be bought up at high discounts (as compared to their current valuations) and people are going to be laid off.
Not saying people should read the article, but it's a short one and hides some nuanced points behind an ambiguous title. Probably intentional. I liked how it showed plenty of examples of scaleups getting egg-faced after overpromising AI capabilities. LLMs are useful, but still too costly to run and too error-prone to be let loose on anything mission-critical. I use these tools daily, but that doesn't mean I want AI rammed down my throat at every turn.
A few of my favorite tech authors and OSS contributors have gone full AI-parrot mode lately. Everything they post is either "how they use AI," "how amazing it is," or "how much it sucks." I think a bunch of us have had enough and just want to move on. But try saying anything even slightly critical about AI, or about how the big companies are forcing it into everything in the name of “progress,” and suddenly you're a Luddite.
I'm just glad more folks in the trenches are tired of the noise too and are starting to call it out.
> For a good 12 hours, over the course of 1 1/2 days, I tried to prompt it such that it yields what we needed. Eventually, I noticed that my prompts converged more and more to be almost the code I wanted. After still not getting a working result, I ended up implementing it myself in less than 30 minutes.
This is very common. We need a name for this phenomenon. Any analogies/metaphors?
I generated (huh!) some suggestions and the funniest one seems to be:
> Just One More Prompt Syndrome (JOMPS)
But I assume there's a better one that describes or hints at the irony that describing precisely what a program should be doing _is_ programming.
Could be that some people are not familiar with the limitations of LLMs and what to expect from them, and could benefit from learning how to properly do prompting, before outright rejecting them.
The LLMs are sometimes frustrating, but over time you learn how to drive them properly, exactly like when you learn how to use the right keywords on Google.
Also, considering he is a free user, the author might be using one the default models ("auto" -> o4-mini or gpt-4o, rarely Claude Sonnet/Opus 3) in Cursor
-> Both of them are rather weak, especially with the automatic context splitting that Cursor does to save money
Claude Code + Opus 4 would likely yield better results here.
> Also, considering he is a free user, the author might be using one the default models ("auto" -> o4-mini or gpt-4o, rarely Claude Sonnet/Opus 3) in Cursor
Revontulet is my favourite myth. The fire fox of the northern lights. There's lots of great stories and animations of the myth, like this one: https://www.youtube.com/watch?v=sN5goxeTfjc
I gave copilot a screenshot of my work schedule 3 times. The first time it extracted the data flawlessly and organized it the way I asked. The second time it did everything wrong until I talked it though the process in tiny steps. The 3rd time I've asked it to repeat the same steps for the new schedule and it got it wrong just like the initial second attempt. I then tried to talk it though the process step by step and it got everything wrong!? With each step it messed up things it got right the step before. It eventually asked me to upload the image again.
I suppose I could upload the image to a table to json website and provide copilot with the json but the point was to make things easier. In my mind there is nothing complicated about the structure of a table but if I ask copilot to merely extract the text from a row starting with some text it goes insane. Optical character recognition is an idea from 1870 and they had working implementations 50 years ago. I read a bunch of comments online about models getting progressively worse but my experience was over the span of a few weeks.
Would they be doctoring with the quality to make newer models look even better?
I just thought "AGI will be coming in 6 months every month from now on, until the peddlers run out of suckers^H^H^H investors. And of course I skipped the article.
Not saying that you should read the article, but for some context the actual title as ";-)" at the end. Which for some reason is removed here, that should give you a clue that it actually isn't making such claims.
Did you drop that ";-)" from the title because you want to showcase how few people actually read articles posted on HN and just respond to titles. Or did you not read further yourself and didn't even notice it? :P
> because you want to showcase how few people actually read articles posted on HN and just respond to titles.
Making a misleading title in order to bait responses would be clickbait. People responding to a title that assertively states something is not some amazing gotcha.
Although in this case it must be what the sibling comment said about HN stripping the smiley.
The smiley very much implies that the title shouldn't be taken seriously. I'd say it is right on the mark.
To be clear, bullish has a few meanings. But they all boil down to assertively communicating. From your comment I gather you might think it's etymology traces back to "bullying", but it actual relates back to "bull".
Either way, it removes an important bit of context. Without the it, it seems exactly like one of those over the top hyped claims the article is critical about. It's going to be interesting what sort of comments this will draw in then.
Readit News
My personal take is that the whole chat based interface is a mistake. I have no better solution, but for anything beyond a halluciating search engine, it's not really the way we need to interact with AI.
In my mind we're 6 months away from one of the biggest crashes in tech history.
But this hype-storm just reminds me of the fever-dream blogs about the brave new world of the Internet back when hypertext became widely used in '93 or so (direct democracy, infinite commerce, etc, etc). Yes, of course, the brave new world came along, but it needed 3G and multi-touch screens as well and that was 15 years later and a whole different set of companies made money and ruled the world than those that bet on hypertext.
It's not just chat. But I really enjoy the Claude or chatgpt chat UI.
I'm not sure what your problem with that UI is tbh?
I don't think there will be any crash soon. I'm hoping for lower prices but I can't imagine not having these tools available anymore.
I'm also waiting for better and cheaper support for local development and inferencing.
I'm waiting left and right for a lot of smaller things. Like cheaper Claude, Hardware LLM chips, secure integration of my bank account and other personal data. More capacity for the beta projects of coding agents.
We don't need AGI to have increased productivity and benefits.
If it would be _that_ useless, half a billion people wouldn't use such tools every week (and this is not even counting Chinese users).
I think many hate LLMs because they realize that it starts to eat their bread and butter and are already taking their job and find it difficult to accept.
Investors are over-paying, because they are betting on companies that can be outpaced by others, but the technology itself is here to stay and grow, until it can do more and more tasks that are done by humans.
It doesn't mean that LLMs and other AI tools are going to run completely alone anytime soon (maybe with cars first ?), but eventually it could be.
What will be the world in 10 or 20 years from now ?
It's not a particularly efficient way to interact with the LLMs. Despite how people hate AI being integrated into everything, I do think that's the only way it can be done effectively. My take is that the AI/LLM isn't the product, it's just a feature of a product. Like when there's "AI" in the camera apps on your phone, or some sort of machine learning in medical imaging. To get the full benefit it needs to be a technology built into your product, not a product in it self.
The voice to text, text to voice is a good example. Voice to text is a feature of a transcription or video conference system. Text to voice is an accessibility feature. It doesn't matter that it's powered by an LLM.
If Claude, ChatGPT, whatever do become cheaper the market will crash. My thinking is: Prices go down (they will eventually), if that happens to fast, OpenAI will tank, their valuation is to high to support a $20 a month AI offering, or worse a world where ChatGPT isn't a product, but a feature. If OpenAI goes down the entire industry will have to re-align to a world where we now know that there's a limit to the value an LLM can command. Most of the current AI companies will be bought up at high discounts (as compared to their current valuations) and people are going to be laid off.
A few of my favorite tech authors and OSS contributors have gone full AI-parrot mode lately. Everything they post is either "how they use AI," "how amazing it is," or "how much it sucks." I think a bunch of us have had enough and just want to move on. But try saying anything even slightly critical about AI, or about how the big companies are forcing it into everything in the name of “progress,” and suddenly you're a Luddite.
I'm just glad more folks in the trenches are tired of the noise too and are starting to call it out.
This is very common. We need a name for this phenomenon. Any analogies/metaphors?
I generated (huh!) some suggestions and the funniest one seems to be:
> Just One More Prompt Syndrome (JOMPS)
But I assume there's a better one that describes or hints at the irony that describing precisely what a program should be doing _is_ programming.
The LLMs are sometimes frustrating, but over time you learn how to drive them properly, exactly like when you learn how to use the right keywords on Google.
Also, considering he is a free user, the author might be using one the default models ("auto" -> o4-mini or gpt-4o, rarely Claude Sonnet/Opus 3) in Cursor
-> Both of them are rather weak, especially with the automatic context splitting that Cursor does to save money
Claude Code + Opus 4 would likely yield better results here.
I wonder where you got the free user info.
There's a classical one: Sunk cost fallacy
One effective use of these tools is to get past the blank page dread. Let it write _something_ to get going, then turn it off and start working.
Recognizing where this cutoff should be can prevent "doom prompting"!
I suppose I could upload the image to a table to json website and provide copilot with the json but the point was to make things easier. In my mind there is nothing complicated about the structure of a table but if I ask copilot to merely extract the text from a row starting with some text it goes insane. Optical character recognition is an idea from 1870 and they had working implementations 50 years ago. I read a bunch of comments online about models getting progressively worse but my experience was over the span of a few weeks.
Would they be doctoring with the quality to make newer models look even better?
Deleted Comment
Making a misleading title in order to bait responses would be clickbait. People responding to a title that assertively states something is not some amazing gotcha.
Although in this case it must be what the sibling comment said about HN stripping the smiley.
To be clear, bullish has a few meanings. But they all boil down to assertively communicating. From your comment I gather you might think it's etymology traces back to "bullying", but it actual relates back to "bull".
I don't think anyone expected it to make all books obsolete, but you'd struggle to buy an encyclopaedia that isn't for kids these days.