448
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
this post was submitted on 24 Sep 2026
448 points (98.7% liked)
Showerthoughts
43813 readers
1049 users here now
A "Showerthought" is a simple term used to describe the thoughts that pop into your head while you're doing everyday things like taking a shower, driving, or just daydreaming. The most popular seem to be lighthearted clever little truths, hidden in daily life.
Here are some examples to inspire your own showerthoughts:
- Both “200” and “160” are 2 minutes in microwave math
- When you’re a kid, you don’t realize you’re also watching your mom and dad grow up.
- More dreams have been destroyed by alarm clocks than anything else
Rules
- All posts must be showerthoughts
- The entire showerthought must be in the title
- No politics
- If your topic is in a grey area, please phrase it to emphasize the fascinating aspects, not the dramatic aspects. You can do this by avoiding overly politicized terms such as "capitalism" and "communism". If you must make comparisons, you can say something is different without saying something is better/worse.
- A good place for politics is c/politicaldiscussion
- Posts must be original/unique
- Adhere to Lemmy's Code of Conduct and the TOS
If you made it this far, showerthoughts is accepting new mods. This community is generally tame so its not a lot of work, but having a few more mods would help reports get addressed a little sooner.
Whats it like to be a mod? Reports just show up as messages in your Lemmy inbox, and if a different mod has already addressed the report, the message goes away and you never worry about it.
founded 3 years ago
MODERATORS
I'd wager it's simply not possible to be profitable using AI for most use-cases. You either burn way too many tokens on your queries or risk your business falling victim to an under-tokenized AI hallucination.
Or any AI hallucination
I love the assumption that if you spend enough tokens, it won’t hallucinate, like it isn’t a mathematical certainty or anything.
I agree with you, but I liken to to how companies tend to replace competence with bureaucracy. With enough layers, a significant amount of mistakes get caught. It just slows everything down and costs a bunch of money. There are still problems that get through, just less of them.
Right like. The more tokens you spend, the more hallucinations you get.
You can cut hallucinations by 90% just by prompting the LLM to.not guess API and function names and to provide it with full docs or MCP to find docs.
Yeah but it’s not guessing in the first place. It has zero reasoning skills. It just outputs the most likely next word. Even when you ‘tell it not to’
I know I am getting down voted for this because people haven't tried it. And let me differ, it has reasoning skills, it now gets beyond predicting the next word. And even reducing it to next word is some kind of primitive reasoning.
By experience, by telling the AI to not guess APIs and function names and instead provide it with a way of looking through docs, and inspecting the code and other tools to get it do some tests, the hallucinations disappear. I've been able to have better results with this setup.
This is very different than the "no mistakes" joke.
I’ll back you up on this. I’ve had the experience of seeing Claude fabricate stuff, and then when I challenge it, it says something like “I guessed, let me look deeper” and then it indeed does check deeper with docs or a configured LSP etc and then discovers the real answer.
It doesn’t matter if this was the result of reasoning or just more pattern prediction, the result is that if you tell it not to guess, it responds appropriately.
"it has reasoning skills" Nope. It has larger context windows and someone coded in the ability to reprompt itself for "thinking" steps that are literally just a new prompt with the same context that gets fed back in which merely increases the chances of it guessing the next output token correctly.
It still doesn't "reason". It still doesn't know what the words mean. The modified algorithms are just attempting to generate the corrective prompts for you. If you've ever messed with them long enough, you'd understand that these, "no wait fix it" prompts have rapidly diminishing returns at the best of times. Their "thinking" is a bandaid, and a piss poor one at that.
People here have tried it, LLM use is basically enforced at a ton of companies and the tech industry crowd is a huge portion of the population here.
You're likely being downvoted as you seem to have a fundamental misunderstanding of how LLMs function. LLMs don't reason. LLMs don't think. Telling an LLM to not guess might help ever so slightly, but LLMs can and do hallucate search results regularly.
It's a mistake to trust that LLMs aren't hallucinating just because you told the LLM to not guess.
I juste don't want to talk about this any more becausefor me adding those instructions in AGENTS.MD and giving it the full API and language and platform docs dropped significantly the hallucinations. You just disregarding my experience and want me to believe in your assumptions.
It's not an assumption. You can give an LLM all of the instructions in the world and it can still hallucinate. This is the fundamental flaw with LLMs, they're always hallucinating. They just have guidelines which make it so they get things right sometimes. They are not to be trusted just because you tell it to search or use agents.md.
You still have this fundamental misunderstanding. I encourage you to actually look into how LLMs work. Do not use an LLM to do this research for you.
It helps, sure. But it still does stupid shit CONSTANTLY.
So do humans. Constantly.
It doing sgupid stuff is another issue separate from hallucinations. Preventing LLM from doing dumb shit is more difficult, you need more guards and harnesses.
I don’t agree it’s a separate issue. Sure you can add in more guardrails, and spend more time making the best prompt. But no matter what you do, at some point it WILL do something you either didn’t ask for, or explicitly told it not to do.
Agree, That is why I said 90% reduction of hallucinations.
How do you know it's 90% reduction of hallucinations?
This logically leads to a similar question to OP’s again: if it’s so easy to cut down on hallucinations, then why haven’t the designers of this software already integrated that in?
This is kinda fundamental to economics: the moment AI learns to perform some task, that task becomes less valuable.
But then aren't there just different tasks to do? VisiCalc and Excel didn't make accountants irrelevant.
That's the economists angle, yes, but that's not what I'm getting at.
You can't make a business with AI because anyone can use an LLM to do anything an LLM can do.
There are exceptions in the short term, but they may not last. For example, we're paying a hefty subscription fee for an AI service that scrapes financial data from scanned pages. It's much more accurate than traditional OCR.
Presently its better than its competitors so commands a premium price. However, I don't think it will enjoy that position for long.