>it parrots out whatever it learned from scraping the Internet

Most flesh-and-bones people are like this as well.

>>108925785
>it parrots out whatever it learned from scraping the Internet so it doesn't actually understand things
at best half-true years ago, trivially not true now.
>it's actually very bad at math, since it doesn't do the thinking and calculations a human would do
it's (often) bad at maths *if* you force it not to do the thinking/calcs a human would do, it's world class if it uses a few cents of budget to think stuff through and call tools to shortcut some operations (but it can still do those operations expensively by hand if it must, it's just a waste to spend 1000 tokens on long multiplication)
>It's bad for learning because it's biased and makes lots of mistakes
depends on the field, and the intelligence of the user. i certainly worry about the fact that 5 billion midwits are gonna be asking it to play devil's advocate for crystal healing, relationship advice, their favourite brand of left/right politics, whether they're good at business ideas or science ideas, etc. and a world where popular frontier models ignored this issue would be a bad one. still, it can teach you a lot if you're rigorous.
>it's inefficient and le heckin destroys water
water maybe sometimes locally an issue but too small for me to care about. energy more valid as an issue, but not affecting anything you or i might do with thought and intent put into it, more like if you run google you should think carefully about whether it's worth making each of the 10 gajillion queries per day cost 15x as much energy as they currently do. ultimately $ correspond to kJ; I spend $50 on fuel for my car, I spend $50 on tokens, i can assume i raped the rainforest about the same for each spend, and adjust accordingly if i feel that i ought to.

what do you think op?

>>108925785
>Are LLMs/AI legitimately useful?
An LLM helped me convince my girlfriend to indulge my fetish

Yes, I have made a full website with users login, internal whatsapp clone, wordpress page making for admins, pwa, push notifications and many more. Currently around 50 users are using withut any problems. It did took me 3 months to make it since it's very complex and you can't expect any ai to make it in any shorter than that and of course you must know how to program and how to direct the ai to do what you want. It's around 80 000 LOC, so you do the math. Adding some new feature is not that difficult, most time was spent testing, fixing bugs, leaks, hardening, more testing, more hardening.. we trully live in a golden age of programming, it would take me years to make a thing like this on my own. Thank you for reading my blogpost OP pls updoot

>>108925868
well written, i agree completely.
I have to say i find frustrating how it's often hard to casually discuss how ML and LLMs actually work because they're probably one of the least transparent technologies out there.
So often it goes from educated guesses at best and straight up random nonsensical bullshit because it's repeated enough (like the water meme, which is an issue, don't get me wrong. It's just the way it's often discussed that is annoyingly disingenuous).

>>108925785
>Are LLMs/AI legitimately useful?
It can do my job for me

>>108925794
They're not useful though, which I suppose answers OP's question

yes
if you aren't getting a ton of utility out of them you are retarded, it's legit just pathetic to see people who still haven't figured out how to get value out of them.

>>108926111
Show us the value you are getting out of them

I have excellent knowledge of LLM's and the ecosystem around it, plus other varieties of "AI" tools from the last several years. This is not a situation where I am merely ignorant. This is a situation where people like you make claims that are unsupported.

The reckoning of the "AI" conversation is that most people are useless and do useless things at their job. The fact that a glorified python script can do their job is not surprising. They cannot accept this for ego reasons so they cope by deciding that the python script is incredible, alive, sentient, and super-intelligent. This is proven by the fact that it can do what they do, and they are all of those things, after all.

>>108926146
ai "tools" and anything that uses the word "agent" are all scams, there being a bunch of scammers in the market doesn't mean there isn't utility you can get out of it. The companies also being scammers and over-inflated also doesn't mean there isn't utility you can get out of it.

Just some common stuff I do constantly now:
1) Claude code to custom ricing/write custom programs for my linux setup (I have a folder w/ all the relevant dotfile configs in it it can freely modify)
As an example I like using voice to text ( i switched from windows like 6 months ago) and rather than download something for it I just had it write a custom program that loads a local voice to text model into my gpu when I'm talking, when I'm down talking transcribes it and saves it to my clipboard.)
2) Anytime a program crashes I get claude code to look at the crash dump and it figures out what's wrong with it immediately
3) Anytime I update linux and an update breaks something I just have claude figure out what update it was and fix the thing that broke
4) A ton of custom interfaces/scripts for emacs, anytime I get an idea for a script/functionality it can put it together immediately.
5) Have it write those scripts (I use clojure, not python, only retards use python) anytime I need to set something up to automatically do anything/request things.
6) I like using exercism to learn new programming languages, instead of using their stupid website to submit stuff I just had it use their github repos to build me a custom fully local instance of everything I want + some other custom curriculum of things I'm learning.
(A general principle is you can basically just create custom UIs for anything w/ basically 0 cost)
7)

Frankly I basically never have to leave emacs now because I just get it to write me a custom interface to do whatever I want. I have a really convoluted org mode setup where it's integrated and I can trigger it by doing things as well.
at limit but i have like 10x more

>>108925785
The entire LLM discussion is worthless because the pro AI side largely thinks burning through tokens is just great and efficiency doesn't matter while the anti side still in the year of our LORD $1986+40 tries and pathetically fails to argue semantics about what is and isn't AI.
AIs are very helpful in some areas.*
*Assuming you yourself are already capable enough to independently verify the AIs results.
In other words you can pick up a scalpel, it doesn't make you a fucking surgeon.

They are very useful in security research. Reverse engineering is 90% pattern recognition, it just so happens AI is good at that being a literal fuzzy pattern matcher machine.
I am not convinced it will ever take over the software engineer's job; there's too many complex decisions in the process for a bot to take over. And frankly AI-generated code is still quite bad. But it has grown useful enough that I can't imagine security peeps ever dropping it - they will bear the brunt of the token price hikes

>>108925785
>it parrots out whatever it learned from scraping the Internet so it doesn't actually understand things
This is already what 90% of white collar labor is.

It's complicated.
They are useful if you need to generate an existing solution to an existing problem.
For example, creating a personal website is a well understood, widespread problem, with infinite solutions posted online.
But creating something new is a lot more difficult and in my experience has never saved me time. Only when I need to do something trivial that would take let's say a day or 2 max, has it saved time for me. And those results are usually just for myself, for little scripts or tools that aren't important enough to spend a day on, but are worth the 5-10 minutes it takes for the AI to do it.

frankly I can't find much use locally and I'm not about to spend several grand on a machine for this purpose unless I have a clear goal in mind

>>108926238
for the curiosity i dropped my own compiled acoustics library and identified a few core calls, dumped the ghidra decomp (which looked horrible due to aggressive SIMD optimization) and it nailed the flow
i cant say with a straight face it's something utterly useless let alone the fact it can actually be very useful

>>108925785
>Are LLMs/AI legitimately useful?
They are useful in specific use cases

>>108925785
Yes ho le fuck yes.
> 365 /SharePoint.
Microsoft's 365 suite has always had shit search features. Copilot is worth its weight alone for being to search an entire 355 tenant you have access to.
> Automating and processing stupid documents
I have to manipulate documents every now and then and I hate doing it. I tried an agent to do it for me.
> Manipulating websites for me
Same for this, I have a corporate site I have to go to. AI will go in there and enter in mass amounts of data from a templated format I have. The devs who code the platform are all shit and can't seem to figure out how to process an xlsx file. Yet I can give code some shit together and have it dump automatically into their staging APIs.
Basically if you have something at work you hate doing, see if you can have AI do it for you.
>.protip
Not all models are worth using. Copilot has access to chatgpt models and Claude as well as regular old copilot. Claude I've found does a lot of stuff I ask it for data processing and manipulation quickly.
For simple search shit, the simple copilot model is fine.

>>108925785
People are very excited about coding, and that is a big deal, but the most useful thing about LLMs is that they can parse unstructured text into a structured json with a well defined scheme very well, with very little effort.

This seems like a relatively minor thing compared to all of the impressive things LLMs can do, but this is really quite a big deal when you put it in perspective.
This is a new kind of computational primitive that simply didn't exist before.
You could write parsers and try to do regex/pattern matching/scraping, but it was always a pain in the ass and barely worked. There is basically a whole segment of new startups that are simply exploiting this ability of LLMs to create jsons.

>>108925785
use agent and you never have to touch the terminal again

>>108926111
>it's legit just pathetic to see people who still haven't figured out how to get value out of them
Apparently you know something that no CEO has figured out. Why don't you sell your idea for a couple billion dollars?

>>108925875
what fetish was that? What'd you do have AI right you a text to tell her to let you suck her toes?

>>108926111
checked, leaning towards this. some people are just too retarded to make use of great tools or even see the value in them
>>108927181
>Apparently you know something that no CEO has figured out
not really something to brag about

>>108927181
They are valuable on an individual scale for custom tailoring solutions to small problems you find
I went into specifics here
>>108926194
The large scale hiring practices of a company and how they use stuff is contrary to this. AI rewards engagement and understanding, which is the opposite of how large scale companies function.
If you had someone "own" a service/product they could use a bunch of AI stuff to build value for it, however companies work by top down direction and down allow the autonomy that would allow AI to actually be helpful.
Large scale companies are just designed to not get value from AI because they are just based on paper pushing and proceduralist stuff which undercuts any potential value gain from ai.

>>108927132
What's an example of something I could convert into a .json for utility?

>>108927381
>They are valuable on an individual scale for custom tailoring solutions to small problems you find
>I went into specifics here
You might be right on a technical level, but good luck convincing people to pay the real, non-subsidized cost of AI to perform those tasks, because it won't be cheap forever. Eventually the investor money runs out.

>>108927502
The thread is asking if LLMs are useful, they obviously are. Quibbling about the economic stuff or "Hey in some hypothetical future arrangement will they still make sense". Yes if those companies fail and the hardware prices drop and I can just run one locally sure. You could already do most of what I do locally. There is no reason to think this isn't permanent and as I said anyone not extracting a ton of utility out of them is basically just retarded at this point.

File: 91883.jpg (122 KB, 960x720)

122 KB JPG

>>108925785
I'm going to give you my honest opinion and you can take it or leave it.

As a technology, YES. LLMs do have a very specific use as a sort of amateur/intern tech assistant. I would never give it work meant for an intern, but I would give it work that I would find too trivial to ask someone else to do. For instance, the other day I asked GPT (free plan, I don't pay for anything lol) to align data with some bullshit spreadsheet someone gave me that was obviously generated by AI. It took the spreadsheet, aligned the data, and handed it back. Then I handed it to the person who gave me the spreadsheet. It took a job that would probably have taken me an hour of data entry and did it in two minutes.

Or how about recently, since I'm against paying for anything AI related, I started using minimax's free agent online because it offers the most tokens by far out of everything. I made a graph with a data structure, had some rambling engineer talk with myself during a walk, and dumped the talk into Otter.ai for the transcript, and then the chart and the summary into minimax. In five minutes, it gave me a working piece of pseudocode with associated architecture suggestions. Making graphs for that would have taken me several days.

Same with user input. Before, I would only get QA feedback once something hits actual QA. Now, I ask claude or GPT for a fuzzer for user input, and it makes a fuzzer. I then dump the result of that fuzzing to the same LLM and it tells me if all of the results statistically align. Which was something I used to do in Matlab. Not having to write custom matlab files or my own fuzzer has saved me quite a bit of time.

It saves time in such a way that's deceptive, too. If I were a boomer fuck, I could seriously believe it were magic and it could think and listen to me.

>cont

File: Sera pantsu.jpg (161 KB, 1173x1008)

161 KB JPG

>>108925785
>>108927604
>cont.

AI is extremely useful if you remember that it is not human and it excels at getting menial, bite-sized tasks that you can break every single task down into understandable chunks that you can glue together at the end. It's why I am not interested in all of these scam articles saying "I BUILT A BROWSER WITH OPENCLAW!!!" where some boomer fuck gives a jillion Claude tokens and a meager 1 million dollars to a dozen agents to make something that people routinely make in their basements for free.

You just have to remember that it is not alive. It can't understand you. It's matching patterns, and anything that involves pattern matching is what it is best at. It is not best at creative, lateral thinking. You should be doing your own thinking. However, if you're stuck in a rut and you just want to see "What would be the most normal, fastest way to solve something?" as a bit of a reality check, it's great. If you just want to create a devil's advocate version of a program idea in your head, it's phenomenal. And if you're like me and your job has to stuck in the limbo between being a software developer and being a hardware engineer- then you'll ask yourself how you ever lived without it.

But again, never pay for it. At least if you do, get a deepseek subscription or minimax so you get trillions of tokens for the price of a cup of ramen. If you have serious work you would like serious pairs of eyes on, spend that money and hire an intern.

>>108925785
ironically they're probably most useful for very smart and very retarded people who either know exactly how they actually work or don't know how to use a search engine correctly

File: AI scaling fallacy.jpg (452 KB, 1456x819)

452 KB JPG

>>108925868
>it's (often) bad at maths *if* you force it not to do the thinking/calcs a human would do, it's world class if it uses a few cents of budget to think stuff through and call tools to shortcut some operations (but it can still do those operations expensively by hand if it must, it's just a waste to spend 1000 tokens on long multiplication)
Oh right and adding to this: One thing you should be doing and should break the habit of doing is asking LLMs math questions. Some of these LLMs are just taking training material from reddit, and since they just pattern match: You have no idea if more complicated math questions aren't simply hallucinations from redditors. Granted, they've combined most major models with Wolfram to iron out these deficiencies. But a better way to make sure your math always works out is to ask LLMs to create Matlab or Scilab implementations of its own math problems, and then run the implementations and hand the output back to the LLM. Matlab and Scilab, much like most other engineering tools, will always be better at LLMs at solving their particular brand of scientific problem. But the same goes for any simulation software: Don't expect an LLM to be capable of performing any simulation. If you use a program such as Solidworks, don't expect it to solve Solidworks problems. If you use QSpice, you need to be the one to actually put its output into QSpice and test it, which I occasionally do.

It's great to give you a start point but all of these boomer fucks keep failing at making value from it in the long run not because LLMs are bad but because people expect too much from them.

>>108927691
this, it's the chart of the most retarded and most intelligent sharing a common goal: The most retarded people finally get to have middling competency rivaling the most mediocre person they ever met. Intelligent people get to have a mediocre person as their slave for anything beneath them. It's really the most mediocre people who struggle to find a use for it, since they don't need a tool to give them more of what they're already doing.

>>108927695
they are decent at symbolic grind
and llms can do that not because leddit data but things like RLVR

>>108925785
LLMs are legitimately useful. It's not omnipotent or always accurate, but its a great time saver for menial questions or having new subjects explained. I use it all the time, but you will run into issues the second you start giving it harder question, and I have noticed that for some issues it'll almost verbatim spit out the same forum post I saw on page 1 of Google (and friends). It's best when it works will well documented things and you give it small scope questions. The more brain power you try and offload to LLMs the worse of a time you'll have. Even if you want it to write most of the code, sometimes a bit of back and forth with the AI will result in much better choices and design for the program.

Some people are stubborn and only learn if they do things their way, which is perfectly fine but they will likely be frustrated by AI much sooner. Or if you're experience in a subject, get stuck, and hope AI will somehow figure it all out.

Agentic stuff is neat but if you look at what an AI Agent actually does you realize that its not that impressive. It just a bunch of grep's and find's. The more I use it I wonder if I could be ultimately saving time and improving more if I just read the docs and grep'd myself, so I that is something I'm always trying to get better at. Overall, I think I've improved as a programmer since using AI as I am constantly using it to learn and build. Then again I think most people don't do the same thing I do, which is 'never copy paste, everything must be typed by hand' and 'if you read something you don't understand, open a new chat and ask until it makes sense'. You are liable for all the code, so you better have read it all and understand it at least somewhat.

LLMs are great for language learning.

>>108927235
scat and claude

File: 1654659842654.png (49 KB, 535x509)

49 KB PNG

>>108929292

>>108929292
almost spat out my coffee
what'd it say that convinced her?

>>108925868
>at best half-true years ago, trivially not true now.
how is it not true? if you look for some niche topic it will regurgitate what some reddit post says with 3 upvotes.

File: Jack Thompson.jpg (74 KB, 686x386)

74 KB JPG

>>108925785
>Are LLMs/AI legitimately useful?
Most things are not legitimately useful. Stuff like video games might actively be harmful. Nobody talks about banning video games anymore just because they're a huge time sink and corrupt da youth.
People are against or want too control AI because it could possibly connect dots in ways they don't want happening

>>108929292
that's.. impressive

>>108929364
It doesn't actually have real opinions about your stupid anime bullshit

>>108925785
AI is useful. When I was looking for a single frame super 8 to mess with stop motion for fun, I asked my buddy who's a camera snob and he just sent me a manifesto about how people who use film are the worst people on earth and how I should buy a 3CCD camera instead and film him skateboarding instead of doing stop motion. Meanwhile gemini just gave me a list of cameras that fit.

>>108927132
you can't use a llm based parser on arbitrary data because you can do prompt injection quite easily and it can happen by accident

>>108925785
i work in the agentic llm space and basically set up companies large and small with telephony agents. we've replaced the need for call centers at most places we've partnered with. It's not saving them money, but the agents are typically much better than call centres (embedded business logic and CRM integration) so the business owners are generally very happy and eager to keep pushing it (to replace their low end staff, receptionsts. etc)

so yes very useful

>>108929921
Also because you don't need to deal with SAAAR SAAAR REDEEM SAAARS at the call center level