I am being 100% genuine here, I struggle to understand how the most useful things I've ever encountered are thought of this way and would like to better understand your perspective.
I'm already a pretty fast writer and programmer without LLMs. If I hadn't already learned how to write and program quickly, perhaps I would get more use out of LLMs. But the LLMs would be saving me the effort of learning which, ultimately, is an O(1) cost for O(n) benefit. Not super compelling. And what would I even do with a larger volume of text output? I already write more than most folks are willing to read...
So, sure, it's not strictly zero utility, but it's far less utility than a long series of other things.
On the other hand, trains are fucking amazing. I don't drive, and having real passenger rail is a big chunk of why I want to move to Europe one day. Being able to get places without needing to learn and then operate a big, dangerous machine—one that is statistically much more dangerous for folks with ADHD like me—makes a massive difference in my day-to-day life. Having a language model... doesn't.
And that's living in the Bay Area where the trains aren't great. Bart, Caltrain and Amtrak disappearing would have an orders of magnitude larger effect on my life than if LLMs stopped working.
And I'm totally ignoring the indirect but substantial value I get out of freight rail. Sure, ships and trucks could probably get us there, but the net increase in costs and pollution should not be underestimated.
So for some professionals, mental math really is faster.
Make of that what you will.
The math that isn't mathing is even more basic tho. This is a Concorde situation all over again. Yes, supersonic passenger jets would be amazing. And they did reach production. But the economics were not there.
Yeah, using GPU farms delivers some conveniences that are real. But after 1.6 trillion dollars it's not clear at all that they are a net gain.
Mathematicians are not calculators. Programmers are not typists.
Sure. They don't meaningfully improve anything in my life personally.
They don't improve my search experience, they don't improve my work experience, they don't improve the quality of my online interactions, and I don't think they improve the quality of the society I live in either
At this point I am somewhat of a conscientious objector though
Mostly from a stance of "these are not actually as good as people say and we will regret automating away jobs held by competent people in favor of these low quality automations"
I have the same feeling with AI.
It clearly cannot produce the quality of code, architecture, features which I require from myself. And I also want to understand what’s written, and not saying “it works, it’s fine <inserting dog with coffee image here>”, and not copy-pasting a terrible StackOverflow answer which doesn’t need half of the code in reality, and clearly nobody who answered sat down and tried to understand it.
Of course, not everybody wants these, and I’ve seen several people who were fine with not understanding what they were doing. Even before AI. Now they are happy AI users. But it clears to me that it’s not beneficial salary, promotion, and political power wise.
So what’s left is that it types faster… but that was never an issue.
It can be better however. There was the first case just about a month ago when one of them could answer better to a problem than anything else which I knew or could find via Kagi/Google. But generally speaking it’s not there at all. Yet.
Unfortunately yes I do, because it is placed in a way to immediately hijack my attention
Most of the time it is just regurgitating the text of the first link anyways, so I don't think it saves a substantial amount of time or effort. I would genuinely turn it off if they let me
> That's a feeling, not a fact
So? I'm allowed to navigate my life by how I feel
No.
Because I cannot trust it. (Especially when it gives no attributions).
Anyway, that about sums up my experience with AI. It may save some time here and there, but on net, you’re better off without it.
>This implies that each hour spent using genAI increases the worker’s productivity for that hour by 33%. This is similar in magnitude to the average productivity gain of 27% from several randomized experiments of genAI usage (Cui et al., 2024; Dell’Acqua et al., 2023; Noy and Zhang, 2023; Peng et al., 2023)
Our estimated aggregate productivity gain from genAI (1.1%) exceeds the 0.7% estimate by Acemoglu (2024) based on a similar framework.
To be clear, they are surmising that GenAI is already having a productivity gain.
As for the quote, I can’t find it in the article. Can you point me to it? I did click on one of the studies and it indicated productivity gains specifically on writing tasks. Which reminded me of this recent BBC article about a copywriter making bank fixing expensive mistakes caused by AI: https://www.bbc.com/news/articles/cyvm1dyp9v2o
It's actually based on the results of three surveys conducted by two different parties. While surveys are subject to all kinds of biases and the gains are self-reported, their findings of 25% - 33% producitivity do match the gains shown by at least 3 other randomized studies, one of which was specifically about programming. Those studies are worth looking at as well.
However, what doesn't get discussed enough about the METR study is that there was a spike in overall idle time as they waited for the AI to finish. I haven't run the numbers so I don't know how much of the increased completion time it accounts for, but if your cognitive load drops almost to 0, it will of course feel like your work is sped up, even though calendar time has increased. I wonder if that is the more important finding of that paper.
I use AI in my personal life to learn about things I never would have without it because it makes the cost of finding any basic knowledge basically 0. Diet improvement ideas based on several quick questions about gut functioning, etc, recently learning how to gauge tsunami severity, and tons of other things. Once you have several fundamental terms and phrases for new topics it's easy to then validate the information with some quick googling too.
How much have you actually tried using LLMs and did you just use normal chat or some big grand complex tool? I mostly just use chat and prefer to enter my code in artisanally.
If I need information, I can just keyword search wikipedia, then follow the chain there and then validate the sources along with outside information. An LLM would actually cost me time because I would still need to do all of the above anyways, making it a meaningless step.
If you don't do the above then it's 'cheaper' but you're implicitly trusting the lying machine to not lie to you.
> Once you have several fundamental terms and phrases for new topics it's easy to then validate the information with some quick googling too.
You're practically saying that looking at an index in the back of a book is a meaningless step.
It is significantly faster, so much so that I am able to ask it things that would have taken an indeterminate amount of time to research before, for just simple information, not deep understanding.
Edit:
Also I can truly validate literally any piece of information it gives me. Like I said previously, it makes it very easy to validate via Wikipedia or other places with the right terms, which I may not have known ahead of time.
You're using the machine that ingests and regurgitates stuff like Wikipedia to you. Why not skip the middleman entirely?
The same reasons you use Wikipedia instead of reading all the citations on Wikipedia.
How do you KNOW it doesn't lie/hallucinate? In order to know that, you have to verify what it says. And in order to verify what it says, you need to check other outside sources, like Wikipedia. So what I'm saying is: Why bother wasting time with the middle man? 'Vague queries' can be distilled into simple keyword searches: If I want to know what a 'Tsunami' is I can simply just plug that keyword into a Wikipedia search and skim through the page or ctrl-f for the information I want instantly.
If you assume that it doesn't lie/hallucinate because it was right on previous requests then you fall into the exact trap that blows your foot off eventually, because sometimes it can and will hallucinate over even benign things.
For most questions it is so much faster to validate a correct answer than to figure out the answer to begin with. Vague queries CANNOT be distilled to simple keyword searches when you don't know where to start without significant time investment. Ctrl-f relies on you and the article having the exact same preferred vocabulary for the exact same concepts.
I do not assume that LLMs don't lie or hallucinate, I start with the assumption that they will be wrong. Which for the record is the same assumption I take with both websites and human beings.
1. To work through a question I'm not sure how to ask yet 2. To give me a starting point/framework when I have zero experience with an issue 3. To automate incredibly stupid monkey-level tasks that I have to do but are not particularly valuable
It's a remarkable accomplishment that has the potential to change a lot of things very quickly but, right now, it's (by which I mean publicly available models) only revolutionary for people who (a) have a vested interest in its success, (b) are easily swayed by salespeople, (c) have quite simple needs (which, incidentally, can relate to incredible work!), or (d) never really bothered to check their work anyway.
That is pretty significant in my book.
How do you quantify such things? How can you say with a straight face that this magic box gives you more relevant information (which may be wrong!) and that will revolutionize the workforce?
I still do 10-20x regular Kagi searches for every LLM search, which seems about right in terms of the utility I'm personally getting out of this.
Analyze it this way: Are LLMs enabling something that was impossible before? My answer would be No.
Whatever I'm asking of the LLM, I'd have figured it out from googling and RTFMing anyway, and probably have done a better job at it. And guess what, after letting the LLM do it, I probably still need to google and RTFM anyway.
You might say "it's enabling the impossible because you can now do things in less time", to which I would say, I don't really think you can do it in less time. It's more like cruise control where it takes the same time to get to your destination but you just need to expend less mental effort.
Other elephants in the room:
- where is the missing explosion of (non-AI) software startups that should've been enabled by LLM dev efficiency improvements?
- why is adoption among big tech SWEs near zero despite intense push from management? You'd think, of all people, you wouldn't have to ask them twice.
The emperor has no clothes.
That said, I recently saw a colleague use a LLM to make a non-trivial UI for electron in HTML/CSS/JS, despite knowing nothing about any of those technologies, in less time than it would have taken me to do it. We had been in the process of devising a set of requirements, he fed his version of them into the LLM, did some back and forth with the LLM, showed me the result, got feedback, fed my feedback back into the LLM and got a good solution. I had suggested that he make a mockup (a drawing in kolourpaint for example) for further discussion, but he had surprised me by using a LLM to make a functional prototype in place of the mockup. It was a huge time saver.
Consider something like Shopify - someone with zero knowledge of programming can wow you with an incredible ecommerce site built through Shopify. It's probably like a 1000x efficiency improvement versus building one from scratch (or even using the popular lowcode tools of the era like Magento and Drupal). But it won't help you build Amazon.com, or even Nike.com. It won't even get you part of the way there.
And LLMs, while more general/expressive than Shopify, are inferior to Shopify at doing what Shopify does i.e. you're still better off using Shopify instead of trying to vibe-code an e-commerce website. I would say the same line of thinking extends to general software engineering.
Shopify is tangential to this. I will add that having had experience with similar platforms in the past (for building websites, not e-commerce), I can say that you must be either naive or a masochist to use them. They tend to be mediocre compared to what you can get from self hosted solutions and the vendor lock-in always will be used to bite those foolish enough to use them in the end.
I would say yes when the LLM is combined with function calling to allow it to do web searches and read web pages. It was previously impossible for me to research a subject within 5 minutes when it required doing several searches and reviewing dozens of search results (not just reading the list entries, but reading the actual HTML pages). I simply cannot read that fast. A LLM with function calling can do this.
The other day, I asked it to check the Linux kernel sources to tell me which TCP connection states for a closing connection would not return an error to send() with MSG_NOSIGNAL. It not only gave me the answer, but made citations that I could use to verify the answer. This happened in less than 2 minutes. Very few developers could find the answer that fast, unless they happen to already know it. I doubt very many know it offhand.
Beyond that, I am better informed than I have ever been since I have been offloading previously manual research to LLMs to do for me, allowing me to ask questions that I previously would not ask due to the amount of time it took to do the background research. What previously would be a rabbit hole that took hours can be done in minutes with minimal mental effort on my part. Note that I am careful to ask for citations so I can verify what the LLM says. Most of the time, the citations vouch for what the LLM said, but there are some instances where the LLM will provide citations that do not.
In fact much automation, code or otherwise, benefits from or even requires explicit, concise rules.
It is far quicker for me to already know, and write, an SQL statement, than it is to explain what I need to an LLM.
It is also quite difficult to get LLMs into a lot of processes, and I think big enterprises are going to really struggle with this. I would absolutely love AI to manage some Windows servers that are in my care, but they are three VMs deep in a remote desktop stack that gets me into a DMZ/intranet. There's no interface, and how would an LLM help anyway. What I need is concise, discreet automations. Not a chat bot interface to try and instruct every day.
To be clear I do try to use AI most days, I have Claude and I am a software developer so ideally it could be very helpful, but I have far less use for it than say people in the strategy or marketing departments for example. I do a lot of things, but not really all that much writing.
Nope; cloning a bundle created from a depth-limited clone results in error messages about missing commit objects.
So I tell the parrot that, and it comes back with: of course, it is well-known that it doesn't work, blah blah. (Then why wasn't it well known one prompt ago, when it was suggested as the definitive answer?)
Obviously, I wasn't in the "the right mindset" today.
This mindset is one of two things:
- the mindset of a complete n00b asking a n00b question that it will nail every time, predicting it out of its training data richly replete with n00b material.
- the mindset of a patient data miner, willing to expend all they keystrokes. needed to build up enough context to in effect create a query which zeroes in on the right nugget of information that made an appearance in the training data.
It was interesting to go down this #2 rabbit hole when this stuff was new, which it isn't any more. Basically do most of the work, while it looks as if it solved the problem.
I had the right mindset for AI, but most of it has worn off. If I don't get something useful in one query with at most one follow up, I quit.
The only shills who continue to hype AI are either completely dishonest assholes, or genuine bros bearing weapons-grade confirmation bias.
Let's try something else:
Q: "What modes of C major are their own reflection?"
A: "The Lydian and Phrygian modes are reflections of each other, as are the Ionian and Aeolian modes, and the Dorian and Mixolydian modes. The Locrian mode is its own reflection."
Very nice sounding and grammatical, but gapingly wrong in every point. The only mode that is its own reflection is Dorian. Furthermore, Lydian and Phrygian are not mutual reflections. Phrygian reflected around is root is Ionian. The reflection of Lydian is Locrian; and of Aeolian, Mixolydian.
I once loaded a NotebookLM with materials about George Russel's concept of the Lydian Chromatic, and Tonal Gravity. It made an incomprehensible mess of explaining the stuff, worse than the original sources.
AI performs well on whatever is the focus of its purveyors. When they want to shake down entry-level coding, they beef it up on entry-level coding and let it loose, leaving it unable to tell Mixolydian from mixing console.
"Among the seven modes of C major, only Dorian is its own reflection.
Understanding Mode Reflections When we reflect a mode, we reverse its interval pattern. The modes of C major and their interval patterns are:
Ionian: W-W-H-W-W-W-H
Dorian: W-H-W-W-W-H-W
Phrygian: H-W-W-W-H-W-W
Lydian: W-W-W-H-W-W-H
Mixolydian: W-W-H-W-W-H-W
Aeolian: W-H-W-W-H-W-W
Locrian: H-W-W-H-W-W-W
The Palindromic Nature of Dorian Dorian mode is palindromic, meaning it produces the same scale whether you read its interval pattern forwards or backwards. When you reverse the Dorian interval pattern W-H-W-W-W-H-W, you get exactly the same sequence: W-H-W-W-W-H-W.
Mirror Pairs Among the Other Modes The remaining modes form mirror pairs with each other:
Ionian-Phrygian: Mirror pair
Lydian-Locrian: Mirror pair
Mixolydian-Aeolian: Mirror pair
For example, when you reflect the C major scale (Ionian), which has the interval pattern W-W-H-W-W-W-H, you get H-W-W-W-H-W-W, which corresponds to the Phrygian mode.
This symmetrical relationship exists because the whole diatonic scale system can be symmetrically inverted, creating these natural mirror relationships between the modes"
Are you hoping to disprove my point by cherry picking the AI that gets the answer?
I used Gemini 2.5 Flash.
Where can I get an exact list of stuff that Gemini 2.5 Flash does not know that Claude Sonnet does, and vice versa?
Then before deciding to consult with AI, I can consult the list?
What would make 2.5 Pro (or anything else) categorically better would be if it could say "I don't know".
There will be things that Claude 3.7 or Gemini Pro will not know, and the interpolations they come up with will not make sense.
You must rely on your own internal model in your head to verify the answers it gives.
On hallucination: it is a problem but again, it reduces as you use heavier models.
This is what significantly reduces the utility, if it can only be trusted to answer things I know the answer to, why would I ask it anything?
I have written about it here: https://news.ycombinator.com/item?id=44712300
Do you build computers by ordering random parts off Alibaba and complaining when they are deficient? You are complaining that you need to RTFM for a piece of high tech?
If they are about something you're not sure about, and you're making decisions based on them ... maybe it would actually help, so yes?
> Do you build computers by ordering random parts off Alibaba and complaining when they are deficient?
We build computers using parts which are carefully documented by data sheets, which tell you exactly for what ranges of parameters their operation is defined and in what ways. (temperatures, voltages, currents, frequencies, loads, timings, typical circuits, circuit board layouts, programming details ...)
Lately, I have been using Grok 4 and I have had very good results from it.
Spam emails are not any worse for being verbose, I don't recognize the sender, I send it straight to spam. The volume seems to be the same.
You don't want an AI therapist? Go get a normal therapist.
I have not heard of any AI product displacing industrial design, but if anything it'll make it easier to make/design stuff if/when it gets there.
Like are these real things you are personally experiencing?
That depends on the quality of the end product and the willingness to invest the resources necessary to achieve a given quality of result. If average quality goes up in practice then I'd chalk that up as a net win. Low quality replacing high quality is categorically different than low quality filling a previously empty void.
Therapy in particular is interesting not just because of average quality in practice (therapists are expensive experts) but also because of user behavior. There will be users who exhibit both increased and decreased willingness to share with an LLM versus a human.
There's also a very strong privacy angle. Querying a local LLM affords me an expectation of privacy that I don't have when it comes to Google or even Wikipedia. (In the latter case I could maintain a local mirror but that's similar to maintaining a local LLM from a technical perspective making it a moot point.)
I honestly don't see technology that stumbles over trivial problems like these as something that will replace my job, or any job that is not already automatable within ten thousand lines of Python, anytime soon. The gap between hype and actual capabilities is insane. The more I've tried to apply LLMs to real problems, the more disillusioned I've become. There is nothing, absolutely nothing, no matter how small the task, that I can trust LLMs to do correctly.
For me, LLMs are also the most useful thing ever but I was a C student in all my classes. My programming is a joke. I have always been intellectually curious but I am quite lazy. I have always had tons of ideas to explore though and LLMS let me explore these ideas that I either wouldn't be able to otherwise or would be too lazy to bother.
Are you saying that LLMs are most useful if you're not intellectually curious, and therefore most interested in immediate answers, but also that they're very useful if you're a really great programmer for an unstated reason?
Especially people on the left need to realize how important their vision is to the future if AI. Right now you can see the current US admin having zero concern for AI safety or carbon use. If you keep your head in the dirt saying “bubble!” that’s no problem. But if this is here to stay then you need to get involved.