Is AI actually helpful?

[XerXesXu]

Here's a useful recent paper, including a literature review, on 'hallucinations', how they're caused, and how they can be prevented.

Why Language Models Hallucinate

Further to which, apparently requiring COT models to show their working and rewarding them for the correct procedure, even if they give an incorrect result reduces hallucinations as well as providing other benefits.

For the full 90 pages:

Stress Testing Deliberative Alignment

 

[NaughteeDragon]

While AI did not boost my productivity it did help me overcome writer's block.
I have a story I've been working off-and-on for almost a year. I've hit this point in the story when I need to write escalating tension between two female characters. One is strictly het, one is falling in love with the other. The two have been friends more than half their lives and are both knowingly and encouragingly seeing the same man. It started as a polygamist relationship and is drifting toward a polyamorous triangle. I know where it ends (het girl finally confronts the issue that is driving her fear) but was having issues with how to build the tension.

Enter AI: ChatGPT, I need a 5k word story about two female best friends in a polygamist relationship. One is strictly heterosexual, the second is sexually attracted to the first. Make it a dramatic story in alternative 1st person view.

Do this a couple of times with minor tweaks (het girl had a traumatic lesbian relationship earlier in life, het girl's motivations are not religiously motivated, the man is actively encouraging them, the man is subtly encouraging them, the man is staying out of their relationship until they solve the problem, there is no man they are just best friends).

Great jumping off points for what I needed. I'm over half done with clearing the log jam just using those stories as inspiration for where I wanted to go.

 

[AlexBailey]

AI is a tool.

Feel free to pound nails with your forehead if that makes you feel superior.

I used it many times today to find the most up to date information about a wildfire and road closures where I was traveling out of state.

ChatGPT searched all available bulletins and announcements in real time. I was able to find out what roads were still open and what areas were being evacuated as well as the most recent progress on containing the fire.

It scoured sources, returning source links and travel possibilities in seconds while I was driving in evacuation traffic when I had no other idea where to look for the information.

 

[EmilyMiller]

Any answer you ever want can be found in a slice of 𝝅

It’s pedantic Emily time (like every day, right?).

The property you refer to is - broadly speaking (there is actually a tighter requirement, but it’s not relevant beyond experts) being a Normal Number (the word ‘normal’ is used for all sorts of different shit in math). A statistical interpretation of a Normal Number is one whose expansion (strictly in any base, not just decimal) is essentially random and uniform (I’m speaking informally here).

Rational Numbers (like 0.33333333… or 0.142857142857…) can’t be Normal, only non-Rational Real Numbers (Irrational Numbers).

It has been rigorously proven that most (again speaking loosely) Irrational Numbers are Normal Numbers [Borel’s Theorem]. But proving a given Irrational Number is Normal is tricky (aka fucking impossible often). In particular, though most mathematicians might be quite surprised if someone shows that π is not Normal, no one - to my knowledge - has yet proven that it is.

As an aside, the decimal expansion of π is no more remarkable than the vast majority of other Irrational Numbers. So √2 is also probably Normal, but this is equally unproven.

Any Normal Number, not just π, would have the property of having the entire works of Shakespeare encoded in it (and a version of Hamlet where he is a centaur).

 

[EmilyMiller]

It’s pedantic Emily time (like every day, right?).

The property you refer to is - broadly speaking (there is actually a tighter requirement, but it’s not relevant beyond experts) is being a Normal Number (the word ‘normal’ is used for all sorts of different shit in math). A statistical interpretation of a Normal Number is one whose expansion (strictly in any base, not just decimal) is essentially random and uniform (I’m speaking informally here).

Rational Numbers (like 0.33333333… or 0.142857142857…) can’t be Normal, only non-Rational Real Numbers (Irrational Numbers).

It has been rigorously proven that most (again speaking loosely) Irrational Numbers are Normal Numbers [Borel’s Theorem]. But proving a given Irrational Number is Normal is tricky (aka fucking impossible often). In particular, though most mathematicians might be quite surprised if someone shows that π is not Normal, no one - to my knowledge - has yet proven that it is.

As an aside, the decimal expansion of π is no more remarkable than the vast majority of other Irrational Numbers. So √2 is also probably Normal, but this is equally unproven.

Any Normal Number, not just π, would have the property of having the entire works of Shakespeare encoded in it (and a version of Hamlet where he is a centaur).

Have you all cum yet?

 

[onehitwanda]

Have you all cum yet?

asymptotes are a bitch

 

[ShelbyDawn57]

asymptotes are a bitch

I almost got that one.

 

[LargoKitt]

AI is booming mostly because an unprecedented amount of money has been invested in it and a non-trivial portion of that has gone to media blitzes.

There has been enormous work done on how to get programs to improve themselves. (If you care about it, first name that pops into my head is work Josh Bongard was doing 20 years ago) Almost none of the effective stuff involves LLM's. The point of LLM's is it takes little human effort (in theory), just lots of data. But it turns out that observation alone with no generalization or logical processing has limited effectiveness. It will asymptotically approach a solution, but never reach there, requiring ever more enormous amounts of data.

Imagine trying to learn a moderately complicated board game, like Settlers of Catan, by only watching other people play. You cannot ask for any explanations. You cannot read the rules. How many games would you have to watch before you mostly got it right. How many before you got every nuance right? Now make that game a million times more complicated. There will never be enough games played to fully learn it. Without reasoning, it is hopeless.

LLMs have already reached this point, investing many billions of dollars in training (even when stealing all the training data) to get minimal improvements. And in many evaluations, regressions. This generation of the hype cycle should have already crashed but too much money is betting on it to let it fall that easily. They are just waiting for the new marks to take the white elephant off their hands, so it's more hype hype hype.

Back in the day I played a game called Wff'n proof; which is essentially using mathematical logic to 'win' the best logical sentences. It taught me something about how programming works; 'if-then' vs 'if-only' statements and so on. Playing with Basic programming taught a bit about how to tell a computer to create something. But I confess I can't see how a device with no experience and no sense of esthetics can do anything but parrot a famous text or picture and substitute 'a different face for Van Gogh's self portrait.' And my art app still consistently 'paints' people with three legs and seven fingers. Duck.ai time and again has given me 'fabulated' or 'repeat back' answers to the most basic questions. Often the old Google 'search path' algorithm would be more accurate and less deceptive. I can't trust A.I. And it is being used to solve 'problems' like 'find me ten compounds more lethal than sarin gas.'

 

[EmilyMiller]

AI is booming mostly because an unprecedented amount of money has been invested in it and a non-trivial portion of that has gone to media blitzes.

It’s classic bubblenomics. This is such an eye-wateringly expensive area that it just has to be successful. It’s not like all that money could possibly be wrong and the only actual use case is guarding against FOMO.

 

[Duleigh]

AI is only what you make of it. I use AI lot for art and illustrations, historical recreation, a pool of story data for series novels, amateur radio license data lookup, math conversions (F to C, kilometers to inches, etc) and other help with math. My favorite is this: "My story has a score of 4.75 from 395 votes. The next day the score is 472 from 397 votes. Votes can only be whole numbers between 1 and 5, what were they?" AI will show you the answers and, if it's a reliable AI (Because there are various AIs out there) show you the math it used.

There are many things that AI is exceedingly bad at - Map making is one. I had two different AIs brag that they can make a beautiful map of the Georgetown Loop railroad. Somehow the moved the famous loop from Georgetown to Silver Plume, and added the Baker Tank (which is on the other side of Mt. Evans) it labeled Clear Creek as Corkscrew Curve and put the Lebanon Mine out on a siding. Would you know any of these inconsistencies? These are items that are clearly displayed on dozens of websites for tourists, how could AI screw it up that bad?

That's the user's fault. I told the AI "Make me a map of the Georgetown Loop railroad." So, it made a map. I didn't stress accuracy. Think of a genius level 4-year-old with a search engine. I've done some amazing (to me) things with AI, I've used it to recreate images of life in my hometown in 1905, I've used it to colorize and animate the Buffalo Waterfront in 1910. I can go to AI now and say "I need cover art for my latest book I want to show Josh and Veronica at their cabin in the woods. It's autumn and they're cutting firewood. Veronica is on the chain saw and Josh has an axe splitting firewood. They are wearing work clothes. Because of the data I've entered, AI knows who Josh and Veronica are and what they look like, and what their cabin looks like and what kind of trees are in their forest. Since I said cover art, it knows the size and how much space top and bottom is needed for titles. But I had to tell it what they're wearing because the default is what they wear to work.

As Greg Gutfeld said, the next big thing isn't building AI, it's writing input prompts. Like with the map, I didn't stress accuracy, my input gave AI free rein to improvise when it made a map. It knows what I want my characters to look like, but I also have to add how to draw them. I normally use words like "Photorealistic" and "Hyperdetailed" but I could also use words like "2.5d anime" and "Studio Ghibli" and Josh and Veronica would come out looking like they were staring in Spirited Away. Like the original Luddites, AI isn't something to panic about. It is new and it has a lot of overhyped capabilities, but it won't change your spark plugs, it won't cut your lawn and it can't write a story to save its life.

 

[GuiltyCowboy]

It’s classic bubblenomics. This is such an eye-wateringly expensive area that it just has to be successful. It’s not like all that money could possibly be wrong and the only actual use case is guarding against FOMO.

I'm sure the point has been made repeatedly but it's worth emphasizing that the business plan of AI is to make money by persuading your boss to lay you off - your boss saves half your salary, and the AI company takes the other half in return for the AI software that will supposedly replace you.

Is the bet that AI will actually be able to do your job? It is not - AI's bet is that it can successfully persuade your boss to make that saving on your salary. They will probably succeed.

So, the AI boom is very much an American thing because AI is a capitalist's tool.

If you think that's me being a hippy, you should look at the example of decidedly un-hippy China: it's really notable how China is putting far, far less into AI ($100bn in 2024 compared to the USA's $350bn in 2025) while they put a staggering $940bn into green infrastructure.

American capital wants your boss - sorry, wants you - you to believe that the future is AI. The state that's responsible for 1.4bn people seems to think the future is still going to be based on molecules.

Which is why 70% of the world’s EVs were manufactured in China, 80-85% of global solar photovoltaic manufacturing, and more than 75% of all global battery production too.

 

[EmilyMiller]

The fact that these companies are trying to IPO and skirt the rules about getting on the big indexes tells me that we're just entering that last phase (and that it may be brief).

The claimed rationale is that they need even more cash to build even more data centers. But the X crap about space data centers tells you this is all hyperbole. The IPOs are about getting the money out before anyone notices how narrow the actual use cases are.

And X is allocating like 30% of shares for gullible fanboys individual investors. This tells you all you need to know.

At my work, access to genAI is being rationed in IT as it’s become eye-wateringly expensive.

Why am I visualizing tulips?

 

[GuiltyCowboy]

At my work, access to genAI is being rationed in IT as it’s become eye-wateringly expensive.

Quite. Any company that becomes reliant on AI is going to be handing over more and more of its revenue to the AI companies that provide its software.

Someone's got to provide the profits on all that investment...

 

[onehitwanda]

The claimed rationale is that they need even more cash to build even more data centers. But the X crap about space data centers tells you this is all hyperbole. The IPOs are about getting the money out before anyone notices how narrow the actual use cases are.

It's even more insidious - the play is to conduct some wide-scale looting, then pass the bag full of "100% authentic golden eggs, guv" to the pension funds and institutional investors, and leave them holding the bag of shit and empty promises when the rug gets yanked out from under the share price as the world wakes up.

 

[EmilyMiller]

It's even more insidious - the play is to conduct some wide-scale looting, then pass the bag full of "100% authentic golden eggs, guv" to the pension funds and institutional investors, and leave them holding the bag of shit and empty promises when the rug gets yanked out from under the share price as the world wakes up.

This is why they are gaming the rules about being included immediately in indices, so passive funds have to buy them. And our investment funds and pensions get raped, while they buy a third Carribean island.

 

[nice90sguy]

Interesting that this thread was started so long ago that the information in the O.P. is now completely out of date.
I started using a copilot in September of 2025, and it was frustrating and didn't help me more than it hindered me. Today, it's unrecognizably more powerful and "correct" -- it's extremely rare for Claude Opus 4.7 to produce thousands of lines of code from a single prompt that doesn't work the first time -- and created the code the prompter intended.

Without exaggeration, my productivity has increased about 100-fold over the last few months.

Even stuff I don't really care about, like imagegen, has improved considerably. Yesterday, I tried giving Gemini a set of image examples (part of an icon set) and asked it to create a new one based on a prompt - it did a very good job, color/style matching, getting transparency right. That did not happen 6 months ago when I last tried it.

I still maintain that the main reason why people complain about it is that they're not using it right.

It's okay for brainstorming, but not for generating ideas on its own
It's amazing for coding, but only when it has a really clear understanding of the task (the human-created spec).
It's good at style-matching, but poor at reasoning.
It's extremely good at comprehension, but not great at recall.

 

[liliput1]

The claimed rationale is that they need even more cash to build even more data centers. But the X crap about space data centers tells you this is all hyperbole. The IPOs are about getting the money out before anyone notices how narrow the actual use cases are.

And X is allocating like 30% of shares for gullible fanboys individual investors. This tells you all you need to know.

At my work, access to genAI is being rationed in IT as it’s become eye-wateringly expensive.

Why am I visualizing tulips?

The weirdest thing about the whole tulip thing is that it turns out that the most valuable tulips were infected. The patterns on the flowers were a disease, not genetics. I wonder what the modern parallel is. Elon?

 

[liliput1]

YouTube keeps showing me testimonial videos for some weight loss/health product where the "people" keep saying "I lost weight" or I that, and really that should be straight up illegal. I thought we had laws about that in the US but either apparently not, or this is another bubble thing where "this time it's different" like all the other times.

Vonnegut has a good one about being a true believer. It goes something like: at 30 you realize you were an idiot at 20. At 40 you realize you were an idiot at 30. At some point you realize that at 50 you're going to think you're an idiot now. So why invest absolutely in anything you believe now when it's going to be wrong in ten years?

 

[liliput1]

I played around a lot with AI over the last few months (I used to work in tech before AI, and if I ever want to go back I need to be up-to-speed about what it can do and what it can't).

My takeaways:

Critique - This one's interesting. It's okay at this although it's got a very literary lens since it was trained on MFA texts and writings. I'd never in a million years trade away a human beta reader for AI critique, but AI doesn't get bored critiquing the same thing 25 times in a day. The sycophantic nature of AI makes most of its critique suspect, but once you get used to working with it you learn to tease out the real feedback. Your mileage may vary, and I think this use case is questionable although I find it valuable.

This is the only one I use. Half the time it's wrong, but some of those wrong critiques trigger a train of thought that helps me find something to improve anyway. So it's right-ish upwards of 2/3rds of the time.

Brainstorming - If you're stuck and just looking for randomly generated text to see if anything connects with you, randomly generated text is kind of AI's thing.

This is a life skill people should have already mastered and I press junior employees into flexing, the way moms push you to eat your vegetables. This is literally the vegetables of creative work. I worry that the ease of AI will now kill it. You should not use AI for brainstorming. The paths in your brain that you use increase their myelinization. Myelin speeds up communication along those paths, and so your brain will reach for those first when thinking (that's some of the truth in the 'two wolves and the one you feed' metaphor). If you lose this then you're useless without your computer. Which means no more epiphanies over lunch, or brushing your teeth. And those are usually the best ones.

 

[Lumiere_Amie]

Before I retired, my colleagues were using AI to accomplish tasks that humans are incapable of; pattern recognition was the focus. When you're trying to troubleshoot microelectronics from images obtained from various non-invasive/non-destructive technologies, and the feature sizes are measured in nanometers, no human has the capacity to undertake the task. So yes, some types of AI are useful, but we were using Machine Learning algorithms, not LLMs.
Now that I'm retired, I primarily use AI as my personal sommelier.

 

[XerXesXu]

Interesting that this thread was started so long ago that the information in the O.P. is now completely out of date.
I started using a copilot in September of 2025, and it was frustrating and didn't help me more than it hindered me. Today, it's unrecognizably more powerful and "correct" -- it's extremely rare for Claude Opus 4.7 to produce thousands of lines of code from a single prompt that doesn't work the first time -- and created the code the prompter intended.

Without exaggeration, my productivity has increased about 100-fold over the last few months.

Even stuff I don't really care about, like imagegen, has improved considerably. Yesterday, I tried giving Gemini a set of image examples (part of an icon set) and asked it to create a new one based on a prompt - it did a very good job, color/style matching, getting transparency right. That did not happen 6 months ago when I last tried it.

I still maintain that the main reason why people complain about it is that they're not using it right.

It's okay for brainstorming, but not for generating ideas on its own
It's amazing for coding, but only when it has a really clear understanding of the task (the human-created spec).
It's good at style-matching, but poor at reasoning.
It's extremely good at comprehension, but not great at recall.AIs moving at amazing speed.

AI is improving at amazing speed.
It's great at syntax but, unless prompted into the correct universe, poor at semantics. It lacks 'tacit' knowledge, the expertise acquired by practical experience which never gets written down.

I've played about with it to make my own AI Agent but that's been overtaken by the production of so many propriety/open AI Agents.

My favourite usage is to restore old photos of family. I've also scanned Polaroid boudoir shoots of my former SO. Not only is she restored to her youthful beauty, but she moves as she strips.

 

[SkyBubble]

In other writing outside what I do here, I've used AI when I've done the writing to suggest good outlets to submit the work to. I'll tell it what it's about, give it roughly a two-sentence summery, and say give me the five best places to send it. It will name them and tell me why. So it can be a useful tool for things like that.

 

[Rob_Royale]

At work I have a few thermal test chambers that I know absolutely nothing about because the company refuses to cough up the $$ to send me to the school the manufacturer offers. So, I put the model number and symptoms into Copilot, and damned if that thing didn't troubleshoot my problem correctly. I got an answer - if it's doing A, replace this. If it's doing B, replace that. I replaced the thermostat and it was fixed. It saved me easily two thousand or more on what would have been at least a pair of visits from a techrep. Pretty fucking handy if you ask me.

Do I feel bad about the techrep not getting a service call from me? Not even a little bit. Those bastards know that there is a limited pool of qualified reps for that equipment and charge like lawyers. And I saved all the data from Copilot to a Word doc, so I can build a troubleshooting guide of my own.

For my writing, I use Gemini widely for research as I've been doing a lot of historical pieces. Super easy during my commute. "What model pistols were common in the summer of 1873 in Newton, Kansas? Where did Newton get water for the town? Where the Santa Fe railroad, go after Newton?"
I also use OpenArt for my book covers and advertisements which I then modify in PaintShopPro.
Berate me if you want, but it's been very helpful.

 

[8letters]

There are many problems with AI: (Edit: Large Language Model AI)
1. ChatGPT is coming up to its fourth anniversary of its release. In that time, there have been no successful startups whose business is not AI but is enabled by AI. I'm talking about something akin to ID software or eBay during the Internet age
2. Many of the uses of AI are nice, but not something people will pay big bucks for. They fall into the "It's a nice feature" category. Yesterday, I listened to an AI investor/advocate, and one of the success stories he mentioned for AI is transcripting video calls. Having a generated transcript of a Zoom call is nice, but I'm not sure I'd ever pay money for it. It's a nice feature of Zoom calls
3. The vast, vast majority of AI users aren't paying for it. That's not sustainable. At some point, people will have to pay for AI to generate images, write papers, etc. Once that happens, the use of AI will collapse. Someone mentioned NetFlix as an example of something like AI. Netflix was never free for years. Yes, you might have gotten a free month, but that was it.
4. For people who are actually paying for AI, they are paying heavily subsidized rates. Companies like OpenAI are basing their rates on what they hope their costs will be sometime in the future and not now. In the meantime, they are burning cash like crazy. They could easily go into a death spiral if they had to raise their rates to something closer to their costs
5. New AI companies seem to catch up quickly to established AI companies, and there doesn't seem to be much of a cost of switching AI providers. Chinese Open Source LLM models are almost as good as ChatGPT and are much cheaper. So pioneers like OpenAI are unlikely to ever recover their costs of developing the technology
6. The only way the unprecedented investments make sense is if there are huge layoffs because AI has taken over people's jobs
7. People hate AI. People hate data centers that power AI. People are only beginning to organize to move against data centers and AI in general, but it'll probably get ugly once they do. (Edit: In general, I'm referring to Large Language Model AI, but I'd say it'd also include robots that replace human workers)

 

[EmilyMiller]

There are many problems with AI:
1. ChatGPT is coming up to its fourth anniversary of its release. In that time, there have been no successful startups whose business is not AI but is enabled by AI. I'm talking about something akin to ID software or eBay during the Internet age
2. Many of the uses of AI are nice, but not something people will pay big bucks for. They fall into the "It's a nice feature" category. Yesterday, I listened to an AI investor/advocate, and one of the success stories he mentioned for AI is transcripting video calls. Having a generated transcript of a Zoom call is nice, but I'm not sure I'd ever pay money for it. It's a nice feature of Zoom calls
3. The vast, vast majority of AI users aren't paying for it. That's not sustainable. At some point, people will have to pay for AI to generate images, write papers, etc. Once that happens, the use of AI will collapse. Someone mentioned NetFlix as an example of something like AI. Netflix was never free for years. Yes, you might have gotten a free month, but that was it.
4. For people who are actually paying for AI, they are paying heavily subsidized rates. Companies like OpenAI are basing their rates on what they hope their costs will be sometime in the future and not now. In the meantime, they are burning cash like crazy. They could easily go into a death spiral if they had to raise their rates to something closer to their costs
5. New AI companies seem to catch up quickly to established AI companies, and there doesn't seem to be much of a cost of switching AI providers. Chinese Open Source LLM models are almost as good as ChatGPT and are much cheaper. So pioneers like OpenAI are unlikely to ever recover their costs of developing the technology
6. The only way the unprecedented investments make sense is if there are huge layoffs because AI has taken over people's jobs
7. People hate AI. People hate data centers that power AI. People are only beginning to organize to move against data centers and AI in general, but it'll probably get ugly once they do

I agree with all of the above.

The most established use case - developers using genAI to do some of the heavy-lifting of coding - is one where the costs are already becoming punitive for corporate clients (not speculating on my part, this is happening where I work, though I’m not in IT I work closely with them often). Access to genAI is being strictly rationed due to costs spiralling.

It doesn’t save money, it at best does a few things that make life easier for a narrow set of people. And none of this is stuff that corporations will pay a lot of money for. They have been promised a Ferrari and get a beat up ten-year-old Corolla.

It’s a house of cards.

The technology is useful at the margins, but examples of it failing to add promised value are piling up. The IPOs are the billionaires trying to get their money out before too many people notice the emperor has no clothes.