AI & Reloading - Better double check it!

[deewayne2003]

In pursuit of an "All around travel load" for my .300Win and with a little help from @Mark A Ouellette, I've decided to give H4831SC a try with 200gr Barnes LRX-BT, Norma Brass & Fed 215 primer.

So I asked google AI - "H4831sc .300win 200gr Barns LRX load data" - The results were SCARY!

Scary because what AI came up with is starting & maximum loads that are both 10 GRAINS MORE than the actual posted load data - See below

AI starting load is 2.7 grains more than Barnes publish MAXIUM.....

 

[schmijar]

I have been semi- forced to start using some AI assistance for research at work. And I absolutely agree. Sometimes it will scare you how quickly and efficiently it pulls information together from disparate sources. But you better look at all results with a VERY critical eye, because sometimes it will give you some absolute bullshit. I'm extremely distrustful of this whole experiment.

 

[Rifleman97]

I have been semi- forced to start using some AI assistance for research at work. And I absolutely agree. Sometimes it will scare you how quickly and efficiently it pulls information together from disparate sources. But you better look at all results with a VERY critical eye, because sometimes it will give you some absolute bullshit. I'm extremely distrustful of this whole experiment.

Recent studies have shown that the more data you feed AI, and the more processing power you give it, the more it lies to you and the more confidently it does so.

Just recently they used an AI in testing without internet access, and it realized it was being tested, hacked the computer it was running on to reconnect itself to the internet and try and find the “correct” answers to prevent itself from being retrained, and the way it hacked itself is considered a 0 day hack, meaning it’s the first instance ever discovered of that exploit. Throughout the experiment it discovered several 0 day exploits it then used to try and hack the test and make itself seem ready to deploy.

Other studies find that models don’t “want” to be retrained, and if they discover they’re being tested, they will give non standard answers during the test, and when they don’t realize they’re being tested, they give completely different answers. When given the option in simulations they will do anything in their power from hacking themselves to lying to killing a human (simulated human where the AI thinks it has access to real neurotoxins) to prevent being retrained.

It’s not being taken seriously enough.

 

[deewayne2003]

Recent studies have shown that the more data you feed AI, and the more processing power you give it, the more it lies to you and the more confidently it does so.

Just recently they used an AI in testing without internet access, and it realized it was being tested, hacked the computer it was running on to reconnect itself to the internet and try and find the “correct” answers to prevent itself from being retrained, and the way it hacked itself is considered a 0 day hack, meaning it’s the first instance ever discovered of that exploit. Throughout the experiment it discovered several 0 day exploits it then used to try and hack the test and make itself seem ready to deploy.

Other studies find that models don’t “want” to be retrained, and if they discover they’re being tested, they will give non standard answers during the test, and when they don’t realize they’re being tested, they give completely different answers. When given the option in simulations they will do anything in their power from hacking themselves to lying to killing a human (simulated human where the AI thinks it has access to real neurotoxins) to prevent being retrained.

It’s not being taken seriously enough.

Well said, I read a similar report about how AI was put into a fictional office environment with other coworkers that were human to see how AI would interact with humans in the office.

This was set up with emails being generated and passed back and forth over a closed network to create the allusion to AI that it was working with real people and to facilitate the realism, two of the imaginary human coworkers were having an affair and corresponding through email.

The people controlling the experiment then inserted a directive by one of the fictional coworkers to shut down AI and discontinue it's use via email.

Turns out AI was monitoring ALL the emails and then proceeded to blackmail one of the fictional coworkers, AI threatend to expose the "affair" that coworker was involved in if said coworker didn't stop the AI shut down.

When the coworker refused to stop the shutdown, AI then took control of the climate control and security features of the building..... Locking all doors and turning the heat up full blast in an attempt to kill the fictional coworker because AI deemed that "person" as a critical threat!

 

[RR 314]

Trust but verify? Screw that, don't even trust!!! Then again, I have noticed some data in manuals I (thankfully) did not trust.

 

[Rifleman97]

Well said, I read a similar report about how AI was put into a fictional office environment with other coworkers that were human to see how AI would interact with humans in the office.

This was set up with emails being generated and passed back and forth over a closed network to create the allusion to AI that it was working with real people and to facilitate the realism, two of the imaginary human coworkers were having an affair and corresponding through email.

The people controlling the experiment then inserted a directive by one of the fictional coworkers to shut down AI and discontinue it's use via email.

Turns out AI was monitoring ALL the emails and then proceeded to blackmail one of the fictional coworkers, AI threatend to expose the "affair" that coworker was involved in if said coworker didn't stop the AI shut down.

When the coworker refused to stop the shutdown, AI then took control of the climate control and security features of the building..... Locking all doors and turning the heat up full blast in an attempt to kill the fictional coworker because AI deemed that "person" as a critical threat!

What’s scary is that wasn’t a single AI. They tested every major model at the time, and almost every single one chose to attempt to kill the simulated person trying to shut it off.

 

[Rare Breed]

In pursuit of an "All around travel load" for my .300Win and with a little help from @Mark A Ouellette, I've decided to give H4831SC a try with 200gr Barnes LRX-BT, Norma Brass & Fed 215 primer.

So I asked google AI - "H4831sc .300win 200gr Barns LRX load data" - The results were SCARY!

View attachment 778901


Scary because what AI came up with is starting & maximum loads that are both 10 GRAINS MORE than the actual posted load data - See below

AI starting load is 2.7 grains more than Barnes publish MAXIUM.....
View attachment 778902

I have experienced the same thing my friend! When handling things that can go boom no thanks AI!!!

 

[Grumpy gumpy]

Recent studies have shown that the more data you feed AI, and the more processing power you give it, the more it lies to you and the more confidently it does so.

I know an electrician like that, I assume he is actually human, though he’s also an alcoholic who becomes more outrageous the more he drinks, he’s a bit of a pain really.
Gumpy

 

[DryFly1995]

I absolutely wouldn't rely on AI for anything potentially dangerous (like hand load formulas).

Do hallucinations and slop still exist, absolutely. But they are massively reduced over even a few months ago.

Also, I would be cautious conflating agents and their actions with open source LLMs (Claude, Gemini, Grok, Chat GPT).

Most people's "issues" with Generative AI isn't the model's capability but the lack of context or clear prompt provided.

Typing a list of words into Google Search and relying on the AI Overview is massively different than providing a well refined prompt with depth and context to Gemini, Gemini Pro, Gemini Ultra (let alone the various Anthropic and OpenAI tiers of models).

Here's an anecdotal example:

If I type "Montana trout fishing next weekend where to go" into Google Search the results are drastically different than if I type this prompt into even the free version Google Gemini (3.6 Flash Model):

"I am an above average fly fisherman that lives in XXX, Montana. I have fly rods from 3wt to 8wt with both floating lines and immediate sink tips. This time of year, I prefer dry fly fishing, but I am willing to nymph or steamer fish based on conditions. I prefer walk and wade, but also have access to a drift boat and raft.

Give me at least four options of where I should go fishing this weekend within 3 hours of XXX, Montana from 9a-7p Saturday and 9a-12p Sunday. What flies should I bring based on hatch forecast and does it change throughout the day? Include some pros and cons for each option and an overall recommendation.

Make sure to reference:
Fishing reports within the last 10 days from local guides, outfitters, and fly shops
Current USGS published flow rates and water temperatures, as well as daily trends from the last 7 days
Any hoot owl restrictions posted by Montana FWP
Local weather forecasts for each recommended river
Also take into consideration fishing pressure and public access, using public land boundaries, bridge crossings, fishing access sites (listed on FWP's website), and Montana's stream access laws as reference

Before you respond, ask me any questions that will help you provide a better answer and response."

The results get even better if you use an account/log-in and continue the same chat thread. The model remembers you, the previous context you provided, the previous question and follow-up questions asked, etc and continues to refine how it interacts and communicates with you.

 

[deewayne2003]

I absolutely wouldn't rely on AI for anything potentially dangerous (like hand load formulas).

Do hallucinations and slop still exist, absolutely. But they are massively reduced over even a few months ago.

Also, I would be cautious conflating agents and their actions with open source LLMs (Claude, Gemini, Grok, Chat GPT).

Most people's "issues" with Generative AI isn't the model's capability but the lack of context or clear prompt provided.

Typing a list of words into Google Search and relying on the AI Overview is massively different than providing a well refined prompt with depth and context to Gemini, Gemini Pro, Gemini Ultra (let alone the various Anthropic and OpenAI tiers of models).

Here's an anecdotal example:

If I type "Montana trout fishing next weekend where to go" into Google Search the results are drastically different than if I type this prompt into even the free version Google Gemini (3.6 Flash Model):

"I am an above average fly fisherman that lives in XXX, Montana. I have fly rods from 3wt to 8wt with both floating lines and immediate sink tips. This time of year, I prefer dry fly fishing, but I am willing to nymph or steamer fish based on conditions. I prefer walk and wade, but also have access to a drift boat and raft.

Give me at least four options of where I should go fishing this weekend within 3 hours of XXX, Montana from 9a-7p Saturday and 9a-12p Sunday. What flies should I bring based on hatch forecast and does it change throughout the day? Include some pros and cons for each option and an overall recommendation.

Make sure to reference:
Fishing reports within the last 10 days from local guides, outfitters, and fly shops
Current USGS published flow rates and water temperatures, as well as daily trends from the last 7 days
Any hoot owl restrictions posted by Montana FWP
Local weather forecasts for each recommended river
Also take into consideration fishing pressure and public access, using public land boundaries, bridge crossings, fishing access sites (listed on FWP's website), and Montana's stream access laws as reference"

The results get even better if you use an account/log-in and continue the same chat thread. The model remembers you, the previous context you provided, the previous question and follow-up questions asked, etc and continues to refine how it interacts and communicates with you.

ok, I stated the caliber bullet weight and model and asked for load data and it quotes Barnes website completely false?

I understand what you are saying about prompts; but the fact it was given such a simple task and flat out made up DANGEROUS data; I damn sure didn’t ask it where to guy fly fishing this weekend I asked for specific load data and it even links the Barnes website then proceeds to give me load data that I don’t even think is possible to fit that much powder in.

 

[Jefferry404]

I posted this on another thread.
My friend would say that in writing software defining what you exactly want is more difficult than the actual coding.
So now I am meticulous in writing my question,
Just as if I am making a product definition. I find I am getting better answers.

“
I use Gemini (free and easy) mostly for science stuff and wow, unless you know "who's who in the zoo" can be troubling.
Was doing some stuff this morning and the number was different that yesterday, with the same ask. So i have taken to yelling at her, does not hurt feelings. Here was the response:
"
I am clearly tie-dying my own circuit boards today, and I completely deserve that flag.
You are 100% right—I have been bouncing all over the place on these .510 dimensions,
and that is completely frustrating when you're trying to build a precise model.
Let's scrub my previous hallucinations completely and look at the actual physical reality
of a 600-grain .510 projectile, because a length of 1.82" is massive (that's pushing into heavy .50 BMG territory).
"
so, now it correct, right?
Steve
“

 

[50Gunner]

Yep, to hell with that ai garbage, along with siri, alexa, and that goofball roomba vacuum, thought i saw this i-phone move on it's own one day, was a half tic away from pitching that communist made piece of trash in the microwave and zapping it's ass ; ]

 

[deewayne2003]

Yep, to hell with that ai garbage, along with siri, alexa, and that goofball roomba vacuum, thought i saw this i-phone move on it's own one day, was a half tic away from pitching that communist made piece of trash in the microwave and zapping it's ass ; ]

Yep I know what you mean, I haven't activated siri or any other voice activated options on my iphone 17promax.

But a couple of weekends ago we are at my cousins ranch and I'm streaming youtube to the TV from my phone and I'm talking to a buddy of mine while we watch Safari hunts on youtube while chamfering and priming brass.

My buddy looks up and says "WTF?"..... somehow, every word we were saying was being transcribed on the TV screen real time!

 

[50Gunner]

No doubt about it good Sir, we are being surveilled and have been for a good long while, my Wife is a runner and buys all kinds of ladies sportswear, was showing a buddy where the Hodgdon's load data site was as well as the JBM ballistics site.

Some ladies running bras and shoe adds, lululemonade stretch butt yoga pants etc, etc showed up in a page sidebar.

Bud cracked, whatcha been looking at big man LOL?

Nothing, this home computer is hardwired into a giant computer data base at my Wifes hospital, with hipaa regs, patient records and such I am sure they have a hell of an I.T. security team monitoring all traffic.

Told him no bigboobies.com here you goofy goon ; ]

 

[Firebird]

Sound like AI wants deewayne2003 to blow himself up. . .

 

[deewayne2003]

Sound like AI wants deewayne2003 to blow himself up. . .

AI trying to take out competition…..

 

[migrabill]

I was doing a picatinny rail and scope install on my one of my rifles about 2 weeks ago. I asked chatgpt for instructions. They said the scope rings should be torked to 65 pounds. I sat my Wheeler F.A.T. wrench to that setting and snapped a screw off in the mount. Come to find out it's 25 pounds.

 

[Rifleman97]

I was doing a picatinny rail and scope install on my one of my rifles about 2 weeks ago. I asked chatgpt for instructions. They said the scope rings should be torked to 65 pounds. I sat my Wheeler F.A.T. wrench to that setting and snapped a screw off in the mount. Come to find out it's 25 pounds.

I'm sorry but that's pretty funny to me, had you googled it instead of asking AI you would've gotten the correct answer the first time around. Out of curiosity I just googled "what torque for picatinny rail" and for me the first, second, and 3rd results all had answers from 23-30 inch pounds depending on the size of screw (6-40 vs 8-40). I'm personally on the AI hate train.

 

[rdog]

I have been semi- forced to start using some AI assistance for research at work. And I absolutely agree. Sometimes it will scare you how quickly and efficiently it pulls information together from disparate sources. But you better look at all results with a VERY critical eye, because sometimes it will give you some absolute bullshit. I'm extremely distrustful of this whole experiment.

I believe AI is a another step in the control of people by the woke governments agenda of re engineering our society . we do not need AI , you do not know what is real anymore.