#Proposal for intelligent AI unreal bot i made for the discord

1 messages · Page 1 of 1 (latest)

foggy shadow
#

I made a bot that uses an LLM to answer queries with context and attachments, if its of interest to the discord. commands are /ask <your message>, for a singular enquiry, /cont is for context. it will add to the token counter you see. to continue a conversation from the previous messages. /cont <your message>
the /reset command will reset context so if you want to clean its slate. users will use this too when another user is done with their conversation
the /reboot command reboots the bot
the /botchannel command is its active channel, you select the channel it will take commands it. it will continue to take commands from admins in any channel though.
the !file command is for attachments. so images, or text, or cpp or python code. you attach it to the message with ! file
the only active commands for non admins is /ask /cont /reset and !file

Discord_ZltGXMrrWX.png Discord_vTlFbR6NNe.png
young matrix
#

Just gonna chime in to say we need less LLM generated content here imo

#

(to clairify I'm not against LLM usage, but UE is a niche topic and most LLMs can't generate accurate answers for it which confuses people new to UE)

#

if we do have any LLM bots added we should make sure the output mentions that it was generated for when it gets copy and pasted

charred cobalt
#

Yep. LLM's are ass for unreal.

We have had users here in the past trying to pass of gpt stuff as their own answers, and it's normally completely wrong, but good enough at sounding correct.

#

The worst thing you could do is give an LLM a picture of a bp, a system it already doesnt understand well

foggy shadow
#

i understand the concern. but this could actually solve some of those issues by having the answer clearly shown as being by an LLM, so it can be confirmed. intelligence of the model also matters, and that is rapidly improving .it currently is trained on unreal documentation.

charred cobalt
#

And unreal documentation is incomplete, sometimes wrong, and not great

#

It's not going to solve any issues by saying its ai gen

#

Plenty of people willingly use GPT4 and are surprised when it leads them down a false path, yet still return

celest swan
#

also the example is wrong

charred cobalt
#

In its present (and likely future state, given it needs training on aspects of the engine not available) it's just not something I feel the unreal ecosystem needs

celest swan
#

which proves the point of those who already said that it most likely is useless

charred cobalt
#

A bot that gives false hope and answers is not something that is needed

#

It's just going to end up making more messes for people to clean up imo

#

Not to mention drowning out actual good quality advice with ai spam commands

foggy shadow
#

yeah it wouldnt be global thats for sure

celest swan
#

why did you even pick an example which is already misleading? 😄

#

it mistakes the material parameters for a material function

foggy shadow
#

im not here to profit. but if its unwanted, thats fine. it works well in other UE discords and covers many enquiries

#

the downside with current LLMs is they can be confident in answers, while being inaccurate. the mistake is using them as tools expecting the answer to always be correct. they are for guidance, direction, not doing things for you

foggy shadow
#

the example is of the free tier model. so while its not the smartest, it can cover alot of topics.

foggy shadow
young matrix
foggy shadow
#

then its better to embrace it. so you can tell people the answer is inaccurate and why, to the LLM. because it will notice its incorrect.

young matrix
#

LLMs are generating answers that are leading people to ask the most insane questions and it takes hours to find out the root issue or that they asked a LLM to begin with

young matrix
foggy shadow
#

do you want the mess of LLM generated knowledge throughout all the discord, or focused to one channel where it can be contained?

young matrix
#

I don't want it anywhere

foggy shadow
#

well, thats basically saying no to the future matey

young matrix
#

people can run their own LLMs or use an external service there's no reason to make it easier to access generated BS here

young matrix
foggy shadow
#

'generated BS' did you know most of that generated BS is accurate?

young matrix
foggy shadow
#

which is why AI is growing so rapidly.

celest swan
#

yea, that comes from the same corner which said that crypto currencies are the future

#

in their current state LLMs aren't the future

young matrix
foggy shadow
#

that current state is changing weekly. the rapid pace of their intellect is only increasing

celest swan
#

they don't have intellect to begin with

young matrix
#

if you look at my Github you'll see I also work with LLMs

foggy shadow
young matrix
#

that doesn't mean they're a good fit for every context

foggy shadow
#

or even 1 year ago

young matrix
#

Dude this is tired.... you're the only one here in favor of your idea.

Wait for more support to show up because your responses are getting sad now

foggy shadow
#

i watch AI very closely. i use it for UE and unity work all the time. the smartest model. i made this entirely with LLM instructions. is that useless?

young matrix
foggy shadow
#

i dont see your argument against it as yet

#

besides its 'inaccurate sometimes'

#

which people are also

young matrix
#

ands we're already tired of dealing with the current batch

#

no reason to make more

foggy shadow
#

better get used to it, because more people are using LLMs

young matrix
#

people can use their own local LLMs or an external service already

#

your bot is not needed

foggy shadow
#

are you the owner?

young matrix
foggy shadow
#

alright then your anti AI stance is noted

young matrix
#

also every other person in this thread has agreed thus far except you

young matrix
foggy shadow
#

saying you arent anti AI doesnt mean you arent.

#

your language shows it quite obviously

#

i see it in many people

young matrix
young matrix
foggy shadow
#

people who find AI threatening generally have the view people do who jump in here quickly to attack it and show their opinions on its immediate downsides, without noting its upsides.

young matrix
#

so please stop making shit up and read what I actually said

foggy shadow
#

i did read what you said. and you're speaking nonsense

young matrix
foggy shadow
#

'you dont want it anywhere' whats its threat to you? if it has its own channel, its isolated. it wouldnt bother you.

young matrix
#

nobody is saying it is

#

please stop making shit up

foggy shadow
#

people already use LLMs for UE advice. this simply allows information to be confirmed in a ue environment

young matrix
#

it's just unwanted

#

get over it

celest swan
#

they are good at making shit up, which is why your terminal example is completely out of place

foggy shadow
#

*current capabilities

celest swan
#

fine for stuff like that

#

no

#

CAPABILITIES

#

the way they work

#

without extra layers

young matrix
foggy shadow
#

i dont think you guys understand progression or AI, youve said your bit.

young matrix
foggy shadow
#

fair point.

urban wing
#

@young matrix I suggest you move on.

#

@foggy shadow We wont be adding any LLMs

foggy shadow
#

alright. cheers for the discussion.

young matrix
# foggy shadow fair point.

Why do you think I made it a long time ago? #1264405551513862184 message

...but you're too busy making shit up and saying I gave no proof.

As if I need to prove I'm not anti-AI to be against this particular idea on it's own merits.

(Don't worry I'm done Matt)

foggy shadow
#

maybe not the right time for it. but they will get smart enough to be useful for this task, just that it'll cost right now.

undone mesa
#

There is a large empty void of UE information available online. The idea that any AI would provide good results is a fallacy.

foggy shadow
#

this discord is online, no? going by its knowledge, it probably has been trained on ue5's source code and wouldnt surprise me if its trained from discord too

charred cobalt
#

It clearly hasn't been though, and no, it would need to join this discord to train on it.

You can't just scrape a server like that

#

And, no, it also hasn't been trained on the source code

#

If it had been it wouldn't be so obviously awful at answering questions on it

steady plank
#

also it's kind of a fallicy to assume even trained the bot would be right the documentation often is wrong or out of date lol

foggy shadow
foggy shadow
robust berry
#

i think in general people are sick of AI shitposting everywhere, no need for it here too

robust berry
#

also, actually calling any AI "intelligent" is pretty far fetched

#

i mean, sure, if your bot was actually intelligent, i would be all for it, but we all know its not, because of our previous experiences with AI

steady plank
# foggy shadow which AI? because it's been quite accurate for me. not always though

I mean all of them lol gpt 3.5, 4, claude, copilot (can be alright but mostly meh), youchat, gemini

lol all of them have some level of wrong or hallucination I've for example asked about how to make this function I wrote and I wrote a super detailed prompt with exact requirements just to see how similar it could spit out something cause if I did the same to a person while it could be very different the basics should be the same and it should function basically the same but instead they gave me pseudocode garbage with no basis in reality, half the time they didn't even use the correct unreal libs and just hallucinated new ones.


 FVector ClosestVector = OtherVectors[0];
 float MinAngle = UKismetMathLibrary::DegAcos(FVector::DotProduct(InputVector.GetSafeNormal(), ClosestVector.GetSafeNormal()));

 for (int32 i = 1; i < OtherVectors.Num(); ++i)
 {
     const FVector& CurrentVector = OtherVectors[i];
     float CurrentAngle = UKismetMathLibrary::DegAcos(FVector::DotProduct(InputVector.GetSafeNormal(), CurrentVector.GetSafeNormal()));

     if (FMath::IsNearlyEqual(CurrentAngle, MinAngle, 0.01f)) // Allow for small differences due to floating point imprecision
     {
         // If the angles are equal, choose the vector with the smaller dot product with the input vector
         float ClosestDotProduct = FVector::DotProduct(InputVector, ClosestVector);
         float CurrentDotProduct = FVector::DotProduct(InputVector, CurrentVector);
         if (CurrentDotProduct < ClosestDotProduct)
         {
             ClosestVector = CurrentVector;
         }
     }
     else if (CurrentAngle < MinAngle)
     {
         MinAngle = CurrentAngle;
         ClosestVector = CurrentVector;
     }
 }

 return ClosestVector;```
foggy shadow
#

@steady plank whats the prompt?

foggy shadow
meager scarab
#

And it's also getting worse.

foggy shadow
#

getting worse? well thats just factually inaccurate

meager scarab
#

Becuase of the hamstringing done by the AI model developers, it's actually getting worse at answering certain questions.

foggy shadow
#

unless you mean censorship?

meager scarab
#

I mean answering technical questions, not morally dubious ones.

foggy shadow
#

oh they're definitely getting better at that

#

sonnet 3.5 for example can write transpilers, before it could not

steady plank
# foggy shadow <@254870270039293952> whats the prompt?

I don't have it but I do have a similar example

give me ue5.4 c++ function that estimates time to impact for a missile based on a starting position, time to launch, acceleration rate of missile, target position, target acceleration

it gave me this which looks almost passable until you actually start looking too closely


float CalculateTimeToImpact(const FVector& MissileStartPosition, const float TimeToLaunch, const float MissileAcceleration,
                            const FVector& TargetPosition, const FVector& TargetAcceleration)
{
    // Assuming constant acceleration motion for both missile and target
    // The equations of motion are s = ut + (1/2)at^2, where s is the displacement, u is initial velocity, t is time, and a is acceleration

    // Initial velocity of the missile
    FVector MissileInitialVelocity = (TargetPosition - MissileStartPosition) / TimeToLaunch;

    // Calculate the relative acceleration between the missile and the target
    FVector RelativeAcceleration = TargetAcceleration - MissileAcceleration;

    // Calculate the coefficients for the quadratic equation
    float A = 0.5 * FVector::DotProduct(RelativeAcceleration, RelativeAcceleration);
    float B = FVector::DotProduct(MissileInitialVelocity, RelativeAcceleration);
    float C = FVector::DotProduct(MissileStartPosition - TargetPosition, RelativeAcceleration);

    // Solve the quadratic equation for time to impact
    float Discriminant = B * B - 4 * A * C;
    if (Discriminant < 0)
    {
        // No real roots, meaning no interception point
        return -1.0f;
    }

    // Calculate the two possible times to impact
    float Root1 = (-B + FMath::Sqrt(Discriminant)) / (2 * A);
    float Root2 = (-B - FMath::Sqrt(Discriminant)) / (2 * A);

    // Return the positive root (the one that corresponds to a future time)
    return (Root1 > 0) ? Root1 : ((Root2 > 0) ? Root2 : -1.0f);
}```
meager scarab
foggy shadow
#

which AI? try it with sonnet 3.5. but be more descriptive with your prompts. prompts are everything

meager scarab
#

It means it may be getting better at some things, but not others.

steady plank
foggy shadow
#

your prompting is not good, so the output will be not good

steady plank
foggy shadow
#

slow at english fast at code i guess?

steady plank
#

it literally takes longer to get a good anwser than it does to code it out and nah they are just shit, I tried to get it to clean up this function by walking it through all the flaws and it still struggled I'm not going to teach a glorfied google search how to do it's job

#

it's not AI its a token system based on mass data scrapping they are not the same

foggy shadow
#

'glorified google search' lol. which AI are you using?

#

you dont have curiosity unfortunately so, that is in conflict with growth

steady plank
#

I have lots of curiosity but I'm also not going to waste my own time on a rabbit hole with 0 benefit I have tried all the current LLMs they are best used as summary machines nothing worth their subscription fees

foggy shadow
#

going by the prompt its a user skill issue. i use it to make amazing things

steady plank
#

lol "skill issue" my brother in Christ that isn't a very useful skill go apply for a programming position and list LLM prompting on your resume and see how fast you flunk a technical test

foggy shadow
#

well prompt engineering is a fast growing field btw.

meager scarab
#

The only story I've heard of a company hiring prompt engineers is that they almost immediately got let go because they couldn't actually do anything useful.

young matrix
steady plank
meager scarab
#

And I challenge you to find a story about a prompt engineer having a succesful career that isn't a joke.

steady plank
#

hoenstly I thought it was locked lol

foggy shadow
young matrix
#

Matt already said no.... let's just leave this delusional person alone

steady plank
foggy shadow
#

anti AI peeps will have to face the music at some stage 😉

steady plank
#

lol not if the lawsuits even halfway succeed

meager scarab
#

Nobody here is anti-AI. They simply believe the current gen AI is not good enough.

young matrix
foggy shadow
#

there is no curiosity here, immediate dismissal doesn't form discussion and ideas

steady plank
young matrix
foggy shadow
#

developers find AI a threat. i understand that

#

but its irrational to say that AI is useless, and unintelligent

meager scarab
#

No. AI is a tool to be used by developers. And it has its place. It's just awful at UE right now.

young matrix
foggy shadow
#

the rationale needs to make sense

#

it hasnt as yet

steady plank
#

my rational is not wasting time

young matrix
foggy shadow
#

then why are you here? you cant prompt = AI bad

steady plank
#

anyways I'm out

young matrix
#

so please stop that at least

foggy shadow
#

im not alone at all. AI is the fastest growing field there is

#

the people who understand are implementing these bots

young matrix
foggy shadow
#

anti AI developers unfortunately

#

but in time we'll see

young matrix
#

and the other people in this thread know I use LLMs a ton

steady plank
#

FYI people have made your exact comments for like months if not longer lol it hasn't changed anyones minds

foggy shadow
#

sure. and im not pro AI. lol

#

'im not racist i have a black friend' equivalent

young matrix
#

no it isn't

foggy shadow
#

'im not anti AI i have an LLM friend'

#

proceeds to immediately dismiss AI and with zero curiosity of the capabilities of the AI on offer or the more intelligent options

young matrix
#

can you really just not accept that somebody is pro LLMs in several areas and not in favor of adding a bot to this server?

meager scarab
#

You need to learn to separate "I hate AI" from "I don't think AI is good enough for what you're suggesting"

steady plank
#

nearly a year ago got the same anwser

image.png
#

I litteraly use 3.5 all the time just not for this stuff

foggy shadow
#

going by what people are saying, they have not used the latest AI, or not following it

young matrix
foggy shadow
#

sonnet 3.5?

young matrix
foggy shadow
#

noone has mentioned sonnet 3.5

#

only the old ones

young matrix
#

ok

#

and?

foggy shadow
#

and... wheres your curiosity?

#

mr not anti AI

young matrix
#

we don't want any bot reguardless of the model here

foggy shadow
young matrix
#

(or call you worse)

foggy shadow
#

dont get mad because you got caught out in your lie

steady plank
#

i use sonnet 3.5 daily I just don't want it here

young matrix
foggy shadow
young matrix
young matrix
foggy shadow
#

that you're not anti AI

foggy shadow
young matrix
#

I'm not automatically "anti AI" because I don't want a bot here causing issues

meager scarab
#

In any case, even if there was a bot here, it would create more problems than it solves simply because the people who actually used it would be crap at writing prompts.

foggy shadow
#

'causing issues'

young matrix
meager scarab
#

People who are good at writing prompts already have their preferred AI they interact with.

steady plank
#

<@&213101288538374145> could we get this locked its just going round and round a no by mods was already given

foggy shadow
#

yet you already have people answering questions with poor prompting and arguing against it

#

confine it.

charred cobalt
#

Or, just not have it, much simpler

young matrix
foggy shadow
#

that'd be rejecting reality good sir

charred cobalt
#

It's not rejecting reality though, it's rejecting your tool

steady plank
foggy shadow
#

na people use it to answer questions here already

meager scarab
charred cobalt
#

And those people get jumped on for their clearly incorrect, awfully verbose answers

foggy shadow
#

then why the dismissal?

glossy moon
#

Yeah, let's not use AI to answer questions :P

young matrix
charred cobalt
#

"People do it so why not endorse it" isn't convincing

foggy shadow
#

then stop replying lmao. its not hard

indigo token
#

The problem with AI, and AI generation tools, is the models are just not there to provide good answers. Average answers yes. Good answers no. There is way too many people coming here saying, "chat gpt told me this", "chat gpt gave me this code", etc.

young matrix
foggy shadow
#

yeah chatgpt aint that smart

meager scarab
young matrix
glossy moon
#

I locked the thread by now.