💬│general

[tame current]

:3

[livid mantle]

Embrace the order within your control, for in accepting what lies beyond, you cultivate the wisdom to navigate life's chaos with reason and virtue.

[tame current]

where that quote from?

[livid mantle]

#1218983754165387436

[tame current]

cool

[livid aurora]

Jarvis powered by GPT 4.5?

[minor shard]

That looks like a scam/virus lol

[halcyon coral]

Probably GPT-4.5 = GPT-4 Turbo

[half venture]

Can we get a modify response button like gemini advanced has ?

Would love if

It

[halcyon coral]

You are welcome to explain your idea in #💭│ideas, perhaps also with a screenshot

[distant moon]

Any update on copy.ai email?

[viscid adder]

do perplexity provide API?

[tame current]

Jarvis skull

[tame current]

Yes for seperate models

And at a price

[livid mantle]

its just using chatgpt 3.5 api from.openai with 10shot pre-prompt to make it act like chatgpt 4.5 10 shots

its good exit scam model

i start one say its gpt-5

then ppl pay

i dissapear

[tame current]

jarvis a real ai model?

[wraith crow]

emails for the 6 months free pro will be sent soon

[tame current]

is it only for existing copyai members?

[wraith crow]

new and existing

[tame current]

😮

[tame current]

for pro members on copyai or free too?

[wraith crow]

only subscribers

[tame current]

aw womp

[minor shard]

Very tempted to buy a subscription with copy.ai just for the PPLX 6 months lol

It’s like half the cost of a 6month PPLX sub

[silent magnet]

Easter eggs collected

You were able to grab 1 Easter eggs.

Collected by

• @minor shard

[agile jay]

Wow, I leave for 10s lol

[minor shard]

The freaking thing is so fast

[agile jay]

Is it? Might be seeing the wrong site

[minor shard]

Yeah I think copy.ai sub is like 49$ for a month and you get 6 months pplx

[agile jay]

If, so you mean just to get it for 1 month.

[minor shard]

Yeah

[feral marlin]

What is a 10 shot pre prompt

[livid mantle]

sorry for the jargon

[feral marlin]

I’m fine with jargon I’m just curious what it means

[livid mantle]

i just think that ppl behind their market hide more than what they state in their ad

but thats how marketing works lol

ok so 1 momth copy ai for 6 months perlexity

[agile jay]

10 shot means that the model is given 10 example questions and answers similar to your current question.

[agile jay]

Unless they have a refund policy...

[livid mantle]

yes i will free trial and cancel

keep perppexoty ,6 months

im saying for xopy ai

[placid osprey]

https://github.com/shure-dev/Awesome-LLM-related-Papers-Comprehensive-Topics

[agile jay]

Yep, currently perplexity is my favourite AI service.

Still a few problems though, like easy switch models and easier image gen.

[livid mantle]

im only using perplexity for the opus unlimited internet despite the 30k context window

[minor shard]

Very cool list! Will check this out

[past galleon]

Has perplexity raised the context window for coding? is it 32K now? or is that artificial?

[livid mantle]

in ir dreams

[minor shard]

The official number is at-least 32k

[agile jay]

Yep, for most tasks 32K is already overkill.

[past galleon]

but in practice is it actually 32K?

[agile jay]

Normally it's a lot more.

I'm guessing it's adaptive.

Hopefully when Claude 3 Opus gets cheaper, it will cost less for Perplexity so they can increase the context.

[orchid magnet]

can you really generate images?

[agile jay]

Yep, make a prompt and the option will appear at the top right.

[orchid magnet]

im still also lost on all of these ai tools. I am currently messing around and trying to learn the social media algorithms and want a good ai tool that will help me do just that. Perplexity and chat gpt is cool, but i want something that can specialize in this stuff

[past galleon]

Alright cool, I'm about to cancel my ChatGPT sub. I tend to use the API more anyways. Thus far I've had good experiences with perplexity for research. I was curious to know if the writing mode was merely a toy side project or if it's actually useful.

[agile jay]

If you want something specialized, you have to make it yourself...

[orchid magnet]

no idea how to do that lol

[agile jay]

Writing mode is like normal mode on ChatGPT.

Where it doesn't use sources, or other API's.

[orchid magnet]

also, the image gen isnt working

[agile jay]

Are you a pro user?

[orchid magnet]

yes

i see like 10 or so images on the right, are these all ai generated?

[agile jay]

Oh, you should probably leave and rejoin the server using the discord link in perplexity settings.

[agile jay]

Are you on the app or on web?

[orchid magnet]

web

[orchid magnet]

is there not an ai tool that does this though?

[agile jay]

Well you weren't very specific on what you wanted.

[orchid magnet]

i just want somehting that can help create content, generate ideas, and help me grow more on social media

[agile jay]

Create what kinds of content?

[orchid magnet]

idk yet

[agile jay]

Content is pretty broad.

[orchid magnet]

short form content

boom

i want to help businesses create short form content to get more eyes on them, and would love an ai tool to help me with this. I am def gonna just mess around with it making my own social media profiles first though.

i saw something thats jasper ai

but idk if thats what i would be looking for

[past galleon]

Sounds like you would be better served with a chatgpt sub

perplexity's strength is research IMO

[livid mantle]

32k only good for research

[orchid magnet]

so you think gpt would be the best ai tool

[livid mantle]

depends on ur use case

[past galleon]

each tool has it's strengths. But from what you mentioned, seems like chatgpt is up your alley

[orchid magnet]

generating content ideas, and analyzing

i have been using gpt for a few months ngl its great

[livid mantle]

all llm can do that, its u to how u prompt it

[past galleon]

but you'll have to still do manual work. There's no magical "ai tool"

even if you know how to code, it's still a lot of work to get things to do what you want to do

[orchid magnet]

i know, i will probably make a custom gpt and just dick around with it

[orchid magnet]

no idea how to code lol

[past galleon]

that's fine, just don't have your expectations too high. Think of these "ai tools" as very dumb yet useful assistants.

[orchid magnet]

i gotcha, thank you

[past galleon]

for imagery, MJ is the way to go, Dalle a distant second. For analysis ChatGPT/Claude. For writing Claude. For research, Perplexity.

[orchid magnet]

but my thjing is why dont i just get perplexity where i gfet all of those ai tools for 20 a montyh

[idle hull]

Can anyone answer this?

[halcyon coral]

#💬│general message

[idle hull]

TYSM!

[livid mantle]

thanks ice lava man

[idle hull]

I guess it is not available to me, but I will tell a colleague of mine not to miss the opportunity. Anyway, I will continue to pay for the service because I still love it. Best regards!

[bold reef]

Hello !

Could someone living in an English-speaking country confirm that a YouTube video does indeed have automatic dubbing in the language of that country? (multilingual video)

[past galleon]

Like everyone said, give it a shot for a month. See if it's useful enough for you. For example, I'm subbing to the PRO for research purposes. Not so much for the availability of other models. For that i rather use APIs.

[orchid magnet]

i have heard this around. "api" is that just the site where you get the ai tool on or..?

[past galleon]

API is an endpoint, with a key you can use it on multiple tools. Or alternatively you can use it with the company's playground. That's a reference to OpenAI and Claude. For perplexity it's what @sukuna said.

[tame current]

Pro just gives u 5 bucks for the api

Which again I've said it has limited models available through it

Literally just open-source models and perplexitys base models (no gpt4 or nun that)

[orchid magnet]

is that cheaper?

[agile jay]

API is just the way that devs speak to the website to do the same thing.

[tame current]

Depends which model ur gonna use...

[past galleon]

Depends on your usecase

[agile jay]

If it's cheap or not largely depends on how much you are gonna use it.

[orchid magnet]

so like for example if i wanna use claude just to mess around with it ocassionally, would it be cheaper than the $20 a month they offer?

[agile jay]

Yep

[past galleon]

But mate, if you dont know how to code. It's best to stick to the interfaces. Otherwise you may accidentally overuse the API.

[orchid magnet]

yea i hafve no idea how to code

that sounds rly convinent though

[tame current]

Then don't get the api..

Idk if there even is a webui for perplexity api (though ez to code one)

[orchid magnet]

ah shucks, okay

[agile jay]

Isn't there that site which lets you use the different API's.

[orchid magnet]

no idea

[tame current]

Again their api only provide limited models and they have a playground for Llama models

[past galleon]

Openrouter i think does that

[tame current]

Though idk if that includes perplexities models or if it's even still.up

[orchid magnet]

idrc about perplexities model i just wanna use opus and not pay $20 a month lool

[tender terrace]

Hey yall

[tame current]

Well ain't much options

[orchid magnet]

u wanna politley donate $20 a month

[tame current]

Just go to Claudes site and use pay as you go

[tender terrace]

can someone brefiely talk to me about playground and the options listed

[orchid magnet]

to my ai addiction

[tender terrace]

or just tell me which one in the best

[tame current]

Labs.perplexity.ai

[orchid magnet]

i wish i knew how to code, it looks so cool

[tender terrace]

playground is just like a place to practice?

[tender terrace]

that is not an option

the first option is "sonar-smaller-online"

[tame current]

Whar

[silent magnet]

Easter eggs collected

You were able to grab 3 Easter eggs.

Collected by

• @orchid magnet
• @wraith crow
• @tender terrace

[tame current]

Click on it

[tender terrace]

is that the best one?

there is like 20

[tame current]

Idk

[tame current]

I don't really use anything other than gpt and claude

[orchid magnet]

and google gemini about to come out with that 1.5 milion context window ai

ugh

[tender terrace]

yeah same

[orchid magnet]

i need some money lol

[tame current]

I have access to gemini lol

Do I use it

Nope

[orchid magnet]

which ones do you use?

[sacred ingot]

I have a question. Is the Claude opus in perplexity the same as Claude official chatbot? How about gpt 4 turbo? Is it the same as the model in gpt 4?

[tame current]

Buh

[open trench]

Can I fix long code with Perplexity Pro? I'm considering changing from ChatGPT.

[tame current]

Technically yes but it adds the sources things and stuff

To ur prompt

[orchid magnet]

is the pro model good for just having a conversation like it is with the other ai chatbots?

[sacred ingot]

Oh ok

[agile jay]

Guess you can try the vercel playground

[tame current]

Pro is only good to get access to those other models (gpt 4 turbo and Claude) for a way cheaper price than getting the api

[wraith crow]

if you want it to act more like a chatbot you should use writing focus

[tame current]

I use ai a lot so of I would've paid for reg api usage I would've burnt through $200 by now

[agile jay]

Yep, API is more expensive if you use it a lot.

[tame current]

I once leaked my api key in autogpt discord to see the graph skyrocket to $200 within 30 minutes

[orchid magnet]

i have both perplexity pro and chat gpt plus and use both a solid amount. Stull kinda lost on what i should do next month bc im not payuing for both of them

[tame current]

Don't pay for chatgpt...

U have the same model on perplexity

[tame current]

If ur going completely free then ur kina stuck to whatever free options r out there

[agile jay]

Did you not set a limit...

[tame current]

I have an urge to make an api to access the reg models on perplexity

[tame current]

Nope (didn't pay the 200 bucks anyways)

[agile jay]

I have Perplexity, Claude, GPT Plus and Gemini. The one I use the most is by far perplexity.

[orchid magnet]

the custom gpts though are so good

[agile jay]

There are collections in perplexity.

[orchid magnet]

wait, whats that?

[sacred ingot]

Claude does not support latex yet

[agile jay]

You can copy the GPT's info and make it into a collection.

[tame current]

Ehar?

What's that

[sacred ingot]

Isn’t that wolfram alpha mode?

Or writing mode?

[agile jay]

Basically a wrapper with a system prompt.

[orchid magnet]

you can pretty much attach sources to your own gpt such as a book, pdf, etc. And it uses that info. You are pretty much giving it data to train on

[tame current]

I also wondrr if perplexity pays for wolfram alpha

[agile jay]

They are most likely using their API.

[tame current]

I'm able to get it to explain the steps

And it's mostly correct

[orchid magnet]

the gpts are the only thing keeping me paying for gpt honestly

[tame current]

...

[sacred ingot]

Agreed

[tame current]

💀

[agile jay]

They are just a wrapper, so not really revolutionary.

[tame current]

All they're doin is attaching it to ur request/questions

[orchid magnet]

ik its just convinent though, beacuse on perplexity i would probably have to attach every pdf every time i want to talk to it

[tame current]

I've only ever needed to attach it once

[orchid magnet]

can it remember it though throughjout sessions?

[tame current]

In the same thread yeah

[orchid magnet]

but wont it eventually forget?

[tame current]

It stays attached ?

Still uses it as a source

Question after question

[orchid magnet]

oh shoot

[sacred ingot]

As long as it is in the same thread, it will stay

[halcyon coral]

👀

[tame current]

Eyes

[tame current]

Someone else just got timed out for spam 💀

[orchid magnet]

so if i attached pdfs of a few books i could keep asking questions about thebooks and it will use the info from all of them?

[tame current]

Well ur prob gon a hit the context limit

[orchid magnet]

thats what i was thinking

[agile jay]

It will likely use RaG.

[orchid magnet]

?

[agile jay]

Retrieval Augmented Generation. Basically it searches the book for relevant parts and objects that information I to your prompt for context.

[orchid magnet]

and i would never have to put it back in because the thread has it memorized?

[agile jay]

On every prompt where you talk about the prompt, it should run RAG again to find relevant parts.

[orchid magnet]

oh i gotcha, thanks for explaining it

[agile jay]

I think files stay uploaded for 30days.

[tame current]

Isn't rag used by perplexity for thr sources too?

[agile jay]

I'm a way, since indexing and rag both use vectorised text.

[tame current]

I swear openai had a model like that

[agile jay]

A model?

[tame current]

I might be confusing it with autogpt

Idk

Very long ago

[agile jay]

RAG is really good for reducing costs.

Or if you have a lot of data.

Only problem is the model doesn't get to do in context learning.

If you pass it a textbook etc.

[tame current]

Womp

[agile jay]

But then again, people generally aren't using it like that.

[tame current]

I wonder how perplexity even makes a profit

[agile jay]

From users who don't use it much.

Like most monthly services.

[silent magnet]

Easter eggs collected

You were able to grab 3 Easter eggs.

Collected by

• @agile jay
• @livid mantle
• @pale kite

[tame current]

Yeah but even with gpt 4 turbos cost

It wouldn't take much to go over the 20 bucks

[agile jay]

GPT4Turbo is pretty cheap.

[livid mantle]

lazy

[tame current]

I've burnt through 5 bucks within an hour

[agile jay]

Yep, because you were probably pumping the context.

[tame current]

For some reason it was counting the full 120k tokens even through the request was around 10k

[agile jay]

The cost is number of input tokens + number of output tokens.

[tame current]

I thought they changes it to cost of context tokens and Input etc

[agile jay]

And input is generally a lot cheaper than output.

[warm cave]

I'm sure there are enterprise api pricing

[tame current]

I mean, maybe... I've had access to the one on azure for openai

There wasn't any price dif

Other than the pain of trying to get in

[agile jay]

Also which model were you using?

32K etc

[tame current]

Gpt 4 turbo

[agile jay]

Which context length?

[tame current]

Don't rll remember

[warm cave]

I could be wrong, I'm just assuming that big companies work out good deals for pricing, tho I have not heard anything specific

[agile jay]

Yep, enterprise deals

GPT Plus users get access to the 32K context version.

[tame current]

Idk

30 bucks is insane for gpt4

[agile jay]

?

[tame current]

Gpt4 vision

[agile jay]

Depends on the size of the image.

If you want to reduce costs, first resize the image.

[tame current]

Just lookin at the output costs

[agile jay]

I am guessing the resolution plays a large part in the cost.

Since it has to process more pixel blocks.

[tame current]

It's a wonder they're able to even process images

[agile jay]

Not really.

[tame current]

From an image to text?

[agile jay]

If you convert the blocks into tokens, it's probably less tokens than a normal ChatGPT input.

[warm cave]

chatGPT is getting fancy with the graphs

even has arrows

[minor shard]

The graphs are so nice for seeing data 📊

[vapid onyx]

Can you ELI5?

[warm cave]

Yeah, I gave it this question, and it made a graph using Code Interpreter

Important to note it gave the wrong answer, but i was still impressed by the graph, haha. Its geting beter at showing viual data

Part of the reason I have been beging Perplexity to intigrate the OpenAI Code Interpeter API, or make there own, its so nice 🤩

[tame current]

living on the edge with a crap laptop

[weary crescent]

Is there any way to delete a "source" in a chain of messages where I just wanted an image for one prompt?

[agile jay]

Yep, you can remove sources.

[warm cave]

But i dont think you can for images, Let me check

As far as i can tell you can not, this is something i have suggested get fixed in "Ideas" in the past

[fossil sentinel]

I just saw the grok 1.5 blog

[torn ravine]

Idk but Perplexity’s search result is way more diverse than ChatGPT but from time to time gets mixed up with some unrelated info and subsequently has the model generate a worse answer

[warm cave]

One thing that really impressed me is "Open Interpret" Search, It is quite good, it would be cool to see a hybrid between plex and OI Search

And yes chatGPT search sucks

[vapid onyx]

Ahhhh thats actually really neat. Visual is getting better

[tame current]

I've gotten that but very rare

[devout cargo]

https://youtu.be/57LqvutrOI8?si=x7hGU42L_z3KwmX_

[warm cave]

you’re really depends on your searches, one thing I know about ChatGPT search is that it will often take on the opinion of the first one or two sources it finds, perplexity Wallet, still suffers from this it’s less effected by any one source

They both suffered them most when a topic is overhyped, when you ask about star coder two it will go on about how it’s the greatest revolution in the last like 10 years or something like that

[torn ravine]

As for accuracy, Perplexity sometimes messed up… Here’s one example where it was asked to summarize a paper only to summarize several random papers with similar titles.

[weary crescent]

I get what they're trying to do with pro search... but it's kinda stupid. I always have to skip the questions because it always asks me to answer what I asked it. The extra sources are cool though.

[torn ravine]

Pro search was very nice long ago when it was called Copilot. Don’t know what happened.

[warm cave]

That is a problem with it; I'm not saying perplexity is the best; here is my option that could be wrong: perplexity is best for the initial research phase, where you should cast your net large, and open interpreter is best for further research done the line needing more precision. Also note: I have not used chatGPT search since right after it was added

[weary crescent]

I just claimed my credit a few days ago and have been playing with the features that gpt4 doesn't have. I like that GPT can run code, but I like the organization and custom prompts for each collection in pplx. I wish it was more like chatgpt in that gpt is a conversation, and pplx is a series of advanced google searches. Sometimes I don't need it to search, I just need a quick question answered or instructions for something that google doesn't have

[warm cave]

And btw when I say open interpreter I don’t mean chatGPT code interpreter, it’s a separate open source project

And if you are looking for perplexity inaccuracies you can just look at there mistral post on twitter by perplexity themselves. It’s a problem that I hope to see improve in the future

[torn ravine]

Makes sense. I just wish there is an all around winner so that only one subscription is needed. The fact is I need ChatGPT Plus(more context awareness, internet access, interpreter), Claude Pro (for reading books) and Perplexity Pro (general search with more generous query limit but also very limited context).😣

[warm cave]

Yeah, that Is my hope too, I hope that in the future perplexity is the only AI subscription I need

[agile jay]

Code interpreter only works with python, rip

[warm cave]

In the future maybe it cold be another language, but if I’m not wrong GPT is the best at python and struggles more with other languages. So I’m not sure if changing it would be an improvement or not

Hopefully it would improve it

[agile jay]

If it can only use python, not really a code interpreter. More of a python interpreter.

[warm cave]

Open I interpreter can do Bash and python, but I know that’s hardly changes anything, haha

[agile jay]

Lol, bash...

[tame current]

Real

[tame current]

I'm pretty sure I've said this befor

[silent magnet]

Easter eggs collected

You were able to grab 2 Easter eggs.

Collected by

• @warm cave
• @pale kite

[tame current]

It entirely depends on what libraries it knows

[tame current]

I've gotten 2 eggs today : )

[warm cave]

Yay!

[agile jay]

I just left for 20s...

[tame current]

Man what's the command for leaderboard

Last I checked first place was 3

[warm cave]

Have you guys tried the 01 Light software?

[agile jay]

Don't have permissions to show the leaderboard.

[silent magnet]

Perplexity

Local user leaderboard

#1 | jawnze | 4
#1 | eun0800 | 4
#1 | 13376666666666666666666666666669 | 4
#1 | brknclock1215 | 4
#5 | codelicious | 3
#6 | shaolinmonk3434 | 2
#6 | _paradroid | 2
#6 | skullcrushercmd | 2
#9 | aabb360 | 1
#9 | leonardooliva | 1

[south kindle]

We don’t want 1 ruling. That’s how we get stagnation. Think of the subscriptions like how people use streaming services. Realistically you’re only using 2 or so. Save the 20 dollars for a rainy day.

Or cycle between them.

Personally my chatgpt pro sub is on hold until they release more (sora/gpt5/qstar)

[warm cave]

its still a big work in progress, but super cool

The recording dose not have audio, but you can imagin it talking, haha

[tame current]

4 :/

[tame current]

Prob be possible to setup my account to hit collect when it appears 🧌

[warm cave]

I think that having the option for a wide net search and a narrow specific search would be good, they could be chosen depending on the type of information required

[south kindle]

Or allow us to choose how many sources to use?

[warm cave]

I have noticed that U-Block origin dose not work on youtube anymore, are there any setting i need to change, rn Brave is the only browser that blocks youtube adds 100% of the time

[feral marlin]

lower context i think and different temperature, top k, system prompt

from my experience official claude one is better but it has a limit to usage

[eager fossil]

Uh, question. I'm supposedly using claude 3 opus, but in a reaction it gave me the classic "As an AI developed by OpenAI"..
So is it really Claude?

[south kindle]

Don’t ask models about them selves. They will hallucinate

[eager fossil]

I didn't ask anything about itself, I just requested it to do something.

(read a text file of code)

[south kindle]

What model is shown in the bottom right of the reply?

[eager fossil]

Claude

3 opus

[south kindle]

Then it’s the system prompt injecting that information

But that’s odd behavior. I would’ve expected it to claim it was made by Perplexity

Not unless something in your file mentioned open ai?

[eager fossil]

It is odd, does make me question if it's really claude though. I know about hallucinations when you ask about it but I'm just asking it to do every day things.

Did not, nope.

[south kindle]

Odd for sure.

[eager fossil]

It's just a python file for an XP system in discord.

[south kindle]

I wouldn’t doubt it being Claude unless perplexity is doing something whacky which I doubt.

We’ll see if the devs chime in on that.

Try a new chat thread with the file and see if you can reproduce it.

I would assume Perplexity would rather it not directly lie about itself leading to confusion for users. This situation highlights why that’s not great for user confidence.

[eager fossil]

Tried but no luck seeing as I can't show/repeat the exact same things in the exact same sequence. This time it just tries to do it. But yeah now I'm still worried whether or not it was claude.

I know claude is significantly better at coding, GPT4 messes up a lot comparatively.

[south kindle]

Curious if you’ve tried the free version of sonnet directly from Anthropic. You should get enough free messages to test it out.

Sonnet direct from them is very good (haven’t tried coding tasks though)

[eager fossil]

I would but no point seeing as EU is pretty much banned from having a subscription.
(reason I went for perplexity and not them).

And API is too costly.

[south kindle]

Oh damn Anthropic

Fail

Just remembered that aspect.

[eager fossil]

Yeah, aggrevating. They're really strict about it too, tried getting a sub with VPN/verified US number etc, no luck due to non-US CC.

[south kindle]

Various other tools exist that are hosting opus to try but you’ll only get a few messages. I’ll see if I can find the one that @livid mantle shared with me the other day.

Not unless he feels like posting it lol.

[eager fossil]

I found some free ones but they collect all data in conversations, not a fan of that, though understandable if it's free. But then I don't want it, lol.
Saw Poe hosted, but almost same price as API through anthropic.

[south kindle]

Yup. Accurate. I wouldn’t use them with valuable info. But a generic coding task should be fairly harmless as a means to test it. Just don’t give it any actual info you don’t want to share. Keep it generic.

[serene sand]

https://chat.lmsys.org/

[south kindle]

But yes - if something is free, our data is almost always the actual payment we are giving. Weather the user realizes it or not.

Oh right! I forgot you can just talk to them via the leader board chat. Great suggestion!

[serene sand]

Direct Chat

[south kindle]

Linus at LTT had a really solid rant about this recently too trying to explain it in a way people would get. Scary how much data people voluntarily just hand over.

[eager fossil]

Difficult to avoid it, I think most people don't care anymore. I certainly do, but I've also got a ways to go.

[south kindle]

Curious if you tested the direct chat at the link above. was just messing with it and as far as I can tell, that's the OPUS we expect. not sure if you can upload files though, but you can try to chunk that out and copy/paste.

[livid mantle]

Mr. Alex Romanov, we need to add DBRX in perplexity, one of ur rivals is trying to make conversions with me

[serene sand]

#📢│announcements message

[livid mantle]

oh thankyou

thankyou mares

[eager fossil]

Did not yet, I'll give it a shot!

[whole lark]

Is this going to be OpenAI vs Anthropic vs Mistral? Does PerplexityAI also include?

[minor shard]

Pplx team ftw

[orchid magnet]

been cooking on perplexity

does the context window on perplexity stay the same for the models i use?

[silent magnet]

Easter eggs collected

You were able to grab 1 Easter eggs.

Collected by

• @half venture

[orchid magnet]

i need to learn how to prompt perplexity though tbh

[whole lark]

The meta tags available in Claude 3 are useful.

[feral marlin]

claude answers claiming to be gpt every now and then

[half venture]

On the surface, it looks like Amazon has found its OpenAI in Anthropic, with the retail giant investing another $2.75 billion this week.

Behind the scenes: Amazon’s AGI group is pushing to have their secretive Olympus LLM outperform Anthropic’s latest models in the coming months. It’s CEO Andy Jassy’s pet project right now.

Amazon 👀

[heady quiver]

Anyone??????

[minor shard]

I think the most information we have is that you get at-least a 32K context window for all models.

[minor shard]

I am personally worried if Amazon jumps in this with their own model. They are another company I do not trust

[slate condor]

iPad app not working. Phone app working fine. Did a reinstall and same. Bug?

[smoky oriole]

Did yall see the $100B data center that OpenAI is building?

Crazy world we live in tbh

[agile jay]

100B isn't much for AGI

[minor shard]

Microsoft has tons of money to spend on OpenAI 😄

[tender terrace]

Hey yall, i'm defenitly way out of the loop but i've been reading about this DBRX model that just dropped? Can someone give me some insightfull knowledge? Is it crazy?

and is it free if i'm paying for perplexity pro

[dark solstice]

perplexity is kinda bad not gonna lie

it's command prompts are horrible, and most of the time you need to pay money if you want to actually get some good experience.

and a lot of the time I ask a question it only gives me 1 sentence answers too.

[tender terrace]

what do you mean?

i have the pro version

[dark solstice]

Bruh you just proved my point

lmao

[tender terrace]

is that what you mean? the free version is bad?

[dark solstice]

yes.

[tender terrace]

ahaha oh i gotcha

[dark solstice]

and if you want a better version you need to pay money...

the free version sucks

and the command prompts are broken..

and they dont work

I dont even understand why the developers introduced the "command prompt" they literally dont even work...

[agile jay]

Command prompt?

[dark solstice]

waw thank you for repeating what i just said.....

👍

twice..

🤦‍♂️

[sacred ingot]

Hello, can you get perplexity to generate images using dall-e? I tried to get it to visualise a graph but it says that it is unable to generate images

[agile jay]

It can't do graphs, but can do images. Not sure if it's available for free users though.

[sacred ingot]

I am a paid user

Oh

Trying to experiment with it

[agile jay]

To create images, make a prompt and then it will show up as an option at the top right.

[sacred ingot]

Ok

[dark solstice]

when you could just ask the developers

[sacred ingot]

I seem to be missing something?

I mean this is perplexity discord channel so I am asking anyways

Lol

[dark solstice]

you can just dm the developers dont waste your time with uneducated people like @agile jay

he is just as dumb as someone who dropped out of high school..

[agile jay]

He said he would block me, but I guess he can't find the button...

@sacred ingot common questions can be found in the #⭐│starred channel. Just search there for image and you should be able to find instructions.

[sacred ingot]

Ok thanks

[agile jay]

Np

[minor shard]

Someone has had a long night 👀

[agile jay]

Yep, seems like it.

[minor shard]

@agile jay are u using Google's Gemini. I know some people like it but some don't

[agile jay]

Yep, it's very good at translation, but not as good in other areas.

Pretty sure they added a python interpreter like GPTPlus has.

[minor shard]

Oh hmm that might be useful. I really wish Perplexity had it, I know they prolly won't though as they are search enginge first.

[agile jay]

Yep, would be nice if they branched out, and made a wasm based code interpreter.

Then you could ask it to re-encode videos etc directly in the browser.

[minor shard]

That would be a really cool idea

[agile jay]

Yep, would also be nice if you could store stuff you've made with the AI as a tool, since it would just be a wasm binary.

[dim urchin]

Hey. New here. I keep getting rate limited on Anthropic's site when I use Opus. It is REALLY unlimited with Perplexity? I feel there has to be a soft limit at some point. Also, how to verify that it actually is Opus that being used? Thanks.

[agile jay]

I imagine they are limiting the context, but so far it's more or less unlimited for me. 600 prompts every 24 hours, and I barely use 30ish a day.

I also have Claude Pro, but prefer using Perplexity because of search and the UI.

[dim urchin]

Appreciate the feedback. Yeah, its rough on Anthropic's site....I am nowhere NEAR 600 prompts. lol \

[silent magnet]

Easter eggs collected

You were able to grab 2 Easter eggs.

Collected by

• @agile jay
• @half venture

[minor shard]

lol i looked away for a second

[agile jay]

@half venture always lurking for some eggs, lol

[agile jay]

I know the feeling...

[half venture]

Yes

[minor shard]

I am definitely not going to beat @half venture lol

[teal tulip]

I’m in the same boat. While I sometimes hit the ChatGPT plus hourly rate limit, I never have an issue when using perplexity. I don’t think I’ve ever come close to hitting the rate limit.

[agile jay]

Yep, same

[dim urchin]

Ok, thanks for the feedback.

I know this might sound conspiritorial but....how to know if it truly is the Opus model?

[half venture]

To be honest both chatgpt plus and claude opus subscriptions have a compute problem
Openai doesn't seem to have resources like Microsoft in terms of compute

[dim urchin]

Because again, for 20 a month unlimited Opus....that's not easy to do I feel

[half venture]

Simple because the context length is shortened

And they rely on the fact that most pro users don't even use pro properly

[dim urchin]

What's the context length for Opus on Perplexity?

[agile jay]

From tests people have done, around 60K, but seems to change. Most likely it's adaptive depending on demand.

[dim urchin]

Again, appreciate all the feedback. Its been SO tough to build anything meaningful in a short time with Anthropic's site with all the rate limits I keep hitting.

[half venture]

Yeah its a huge problem once again I feel like Amazon Google and Microsoft seem to have absolutely mogged all the compute available

Can anyone who works on perplexity confirm ? Doesnt perplexity use Amazon servers ?

[minor shard]

The official limit is "at-least" 32k from the team just FYI.

[dim urchin]

Ah ok, understood. Thanks.

[half venture]

Quick question

Why does Google still buy nvidia h100s if it has its own tpus v5p

And Google gemini runs completely on Google's own tpus no ?

[livid mantle]

This video exposes the fluff and nonsense of many popular "alternative" search engines, revealing that most rely on the same big tech indexes and have questionable privacy practices. Brave, DuckDuckGo, Startpage, and Swisscows claim to protect privacy but have limitations. Yandex, Naver, Baidu, and Cốc Cốc cater to specific countries/languages but collect user data. Kiddle is kid-safe but powered by Google. Ecosia helps plant trees but uses Bing/Google results. New AI search engines like Perplexity AI, You.com, and iAsk.ai show promise for direct answers but accuracy is uncertain. In the end, none truly break free from relying on major search indexes, so privacy and unbiased results cannot be guaranteed. The most unique option shown is the Wayback Machine for searching web history.

[half venture]

I mean Google search has been so ingrained into the normal culture worldwide that getting away from Google requires not only a better product than Google's but also better reach

Which is what every major rival to Google is struggling with (bing )

[teal tulip]

It was trained on the tpus and it’s speculated to run on those tpus I believe, but that doesn’t preclude the use of other hardware. Perhaps there’s a hybrid approach for diversity or for use with GCP

[half venture]

Yeah probably h100s are for its Google cloud service for normal consumers

[teal tulip]

Yeah, there are also legacy systems that may not be designed for tpus or projects in the process of migration.

[half venture]

What's frustrating is I am fearful that Google gemini will just join the list called : killed by Google 😭

And all these advances will be for nothing

As always

[agile jay]

Lol, where did bard go?

[half venture]

Ditched

Buried alive

Same can be said about bing though

It's now copilot

Or whatever

[agile jay]

Yep, they are trying the YouTube thumbnail trick of changing the name until it sticks...

[half venture]

Man both Microsoft copilot and Google gemini suck both are so lobotomized 😭 yet they are the fastest due to the compute they have unfair world

[agile jay]

Yep, so many filters in those models.

[half venture]

Copilot is the worst one

Atleast gemini can do things without constantly ending the conversations

[agile jay]

Copilot feels like it's a bit high.

[half venture]

Copilot feels like it's slowly getting more and more deranged

[agile jay]

No idea what Microsoft did to the base GPT4 model.

[half venture]

According to Mikhail former bing chief if that's a title. Apparently they have multiple models inside working not just gpt 4

So when it responds but suddenly says sorry I can't do that

It's a hidden model triggering and deleting everything

[agile jay]

Oh, I thought that was just a simple sentiment analysis on the output as it's being streamed.

[half venture]

Nope

There are multiple models that have multiple filters

Which is what is causing the overtriggering of safety bs

It's very frustrating

Hopefully Mustafa the new CEO can make copilot more like pi

Because the current approach Microsoft had that they will challenge Google search is now dead

Bing didn't even budge Google share at all

😭

[agile jay]

Yep, not a fan of the chat like UI either

I prefer the ChatGPT and Perplexity style of UI.

[half venture]

I think personally the company that first figures out a personal ai companion that is almost magically her like

Will win the LLM race

Which is what openai seems to be planning towards

[agile jay]

Yep, but having a conversation and competently completing tasks are different benchmarks.

[agile jay]

Their chat mode on Mobile is probably the closest.

[half venture]

No actually I personally think. Mustafa pi app recently was the closest

Until Satya saw that and was like : hmm time for a indirect takeover.

[agile jay]

Could you chat with pi?

[half venture]

Yeah

[agile jay]

Haven't tried it, is it still up on pi?

[half venture]

Yeah it is

Super nice voices

[agile jay]

Oh, it reads it out, but you aren't speaking with it though.

Unless there's a mobile app which has more voice features.

[silent magnet]

Easter eggs collected

You were able to grab 2 Easter eggs.

Collected by

• @half venture
• @agile jay

[agile jay]

Lol

[half venture]

Oh right my bad

There is a hands free mode

What is mind boggling to me

Is how no one has been able to have a interruption option for natural human conversation like

Like generally if I talk to someone I can interrupt them and recalibrate etc etc

Why is that so hard

[agile jay]

I am guessing it's just the response time is just too slow.

[half venture]

The only place other i have seen with a interruption feature is groq since there response time is insane

[agile jay]

Yep, it's insane.

[half venture]

God imagine the day claude haiku type of model gets open sourced

And runs on groq type of infra with super fast inference speeds

[agile jay]

I would imagine they would still cap the inference speed.

Since people can't read that fast...

For agents though, that would be fun.

[half venture]

It's for voice tho

Since you would be able to interrupt and continue

[agile jay]

Voice would be even slower than that.

Since people say only a few words a second.

[half venture]

No no you didn't understand

• fast inference speed
• reasoning type of thing in it
• voice has plenty of time for interruption there is a cap on how fast it speaks

Make sense ?

[agile jay]

Yep, I understand your point. The question is how well would it perform. Since it's trained on mostly QA style data, where interruptions are less likely to be in the training data.

Would be nice to try it anyway.

What would be cool is if it could also identify different speakers too...

[half venture]

Hopefully openai figures something out they are ahead of the curve

[agile jay]

Yep, still 2 years ahead...

GPT4 is two years old, and the voice thing they released a day or two ago was also 2 years old...

Not sure how old Sora is internally too.

[half venture]

I am hoping for something like c.ai but Sora and real time talking face to face breaking the uncanny valley

[agile jay]

I think face to face is more likely to be uncanny.

At least with only voice the audio can be worse, since we are used to low quality audio on calls.

[half venture]

I think early 2023

[agile jay]

Also can't wait for their music model.

[half venture]

You know what's funny

[agile jay]

To compete against suno

[agile jay]

?

[half venture]

Suno uses openai voice models

😭

[agile jay]

Lol, really?

[half venture]

Yes lmao

No wonder Microsoft has a partnership with them

To host it on copilot

[agile jay]

How does it make them sing?

[half venture]

The voices

Like the human voices

It's apparently powered by voice engine

And the rest is the add on by suno

[agile jay]

Is that why quite a lot of them sound similar?

[half venture]

Yes

[agile jay]

Since they are probably re-using voices

I wonder how they match the music and voices.

[half venture]

That is something suno has proprietary probably

But notice how Microsoft has their little greasy fingers everywhere

😭

[agile jay]

Yep, they are just snatching up all the new startups

Same with game companies

Anti trust should really step in.

[half venture]

I don't think so Microsoft is being very careful

They are always doing it indirectly

Or using non binding legal contracts

[agile jay]

Yep, they aren't going for majority shares...

Instead they just take the IP and the employees...

[half venture]

Apple is absolutely getting dunked on though

[agile jay]

Apple doesn't do much anyway, in terms of AI

And doesn't have the best relationships with other companies...

Heard they are adding Gemini to Apple phones...

So sounds like they don't have their own model.

[livid mantle]

thats true, thats why apple collaborated with google to put their in their iphones

ya they dont

gemini nowin iphones

[agile jay]

Yep, collaborated with ARM and TSMC for the M series chips. And now Google for their AI...

[livid mantle]

mhm

[half venture]

Apple is purely a hardware company at this point

[livid mantle]

they're like

🤝

its fair because microsoft is with openai

and they have intellectual ai

or whatever they call that team

[agile jay]

More of a company that hires other companies to do most of their work.

I think they pretty much just write the software and hardware nowadays.

And they abandoned the apple car

[half venture]

I can't wait for apple to be humbled

Hopefully Europe makes them break there app store dictatorship

[agile jay]

I think AI might be exactly what Linux needs.

When it's competent enough to create apps etc on the go.

[agile jay]

Yep, a team I work with did a lot for the recent US lawsuit.

[agile jay]

The good thing about the EU starting it is that it makes other governments take a look.

[half venture]

Yeah US also has slapped apple

Australia will soon follow

Japan also hates them

China is now promoting their own huawei stuff

[livid mantle]

china we have oppo

[half venture]

Apple should be humbled soon

[agile jay]

Yep, doesn't look good for apples wall garden

[livid mantle]

china can have a LLM that wont consume tokens but rather consume your social credits

[agile jay]

I prefer social debit

Don't wanna pay that interest

[livid mantle]

TRAITOR!

[agile jay]

You.com lol

How did they even get that domain?

[silent magnet]

Easter eggs collected

You were able to grab 2 Easter eggs.

Collected by

• @agile jay
• @timid marsh

[serene sand]

#💗│sharing message

[warm cave]

How peaceful ☺️

[edgy grove]

Any news on sourcing in the api?

[serene sand]

#🧪│api-general message

[edgy grove]

Thanks just applied

[south kindle]

we should start moving important stuff for the wayback machine to something resilient. its important info.

[small hemlock]

Nonsense. They are also a finance company and a sales platform for third party software.

[whole lark]

When assessing LLMs, are people concerned about context length and what are they trying to do? I would think that if necessary, it would be to split it up.

[devout cargo]

Perplexity is a powerful AI assistant that utilizes advanced language models and search capabilities to provide users with high-quality, informative, and engaging responses. As an AI created by Perplexity, I can share some insights into how the system works and the thoughtful approach behind the Search Generative Experience.

[whole lark]

https://www.perplexity.ai/search/Search-Gemerative-Experience-TGfyKFACRsiFhA_imcoiLg

It is a bit different from RAG. Is there something special about it?

[devout cargo]

It should be noted that in a single Perplexity search, you get Sources, Images, Videos, LLM with the text.
I believe it has well-earned the title of Search Engine Generative Experience

Perplexity is a search engine like Google or Bing 🙂 It has done what Google has not yet been able to do. Google is (SGE) The Perplexity Concept has improved everything. We have to make differences. You.com is an internet chat bot like Copilo that also gives you text and fonts. Perplexity Is Generational Search Engine Because it gives you great images, videos and text. When you do a search these three elements are tracked

[devout cargo]

I understand that Perplexity is the only natural competitor that faces Google SGE

The Google and Perplexity applications have similarities 🙂 Search, Discover, and Save

[half venture]

Agreed but isn't a Google search much much cheaper than a single perplexity search

[sweet jasper]

Of course

But a google search typically is not in your best interests

Rather to shove as much ads in your throat as possible

We are moving towards knowledge base searchs rather than keyword ones

Llms in the future will be infested with ads most likely

They already have product biases from training data to a certain extent

But thats a problem for another day

[celest totem]

"Does the API include citations like the perplexity.ai?

Currently, the API does not generate citations when prompted. We are actively developing this feature for future implementation. Our pplx-7b-online and pplx-70-online models conduct online searches using RAG.Copy LinkShareRewrite this answerRewriteInaccurate or unhelpfulCopyEdit Query"

From the FAQ.

Seems like some of the info 'leaked' out and wasnt edited 😉 haha

[devout cargo]

I am worried about when AI will have to be used. Whoever doesn't do it will be left out of date. I have the intuition that Amazon is very interested in Perplexity and Anthropic. But you don't want to risk a lot of money at stake.

[austere kestrel]

nah not really.. There's like 2 ads at the top (and not even for every search)... That's been Google's genuis in many ways.. (the SEO crap is what has made it progressively degrade imo, not ads...)

[devout cargo]

The predictable is becoming reality. Antrophic would end up being Perplexity's default model. It is using Mapbox all hosted on Amazon web service. In the not too distant future, Alexa can benefit from these services

[austere kestrel]

yeah seems it's being powered by Haiku these days

[indigo pine]

Hi, is there any document to work with Perplexing AI?

[austere kestrel]

so much for open source / in-house / propretiary models .. ( though tbhI just want to be able to query the internet with an LLM - what that model is doesn't really matter at the end of the day.. just whatever works best)

[serene sand]

#🧪│api-general message

[celest totem]

you misunderstood 😄 look at the text in the FAQ lol, someone copy pasted this from the web app 😛

[devout cargo]

I'm sorry, I meant that apparently Claude 3 structures the text better and seems to be able to generate longer text than GPT 3.5

[halcyon coral]

Please check the FAQ: https://www.perplexity.ai/hub/faq

[devout cargo]

https://www.perplexity.ai/hub/faq/what-is-perplexity

What is Perplexity?
Perplexity is an alternative to traditional search engines, where you can directly pose your questions and receive concise, accurate answers backed up by a curated set of sources. It has a conversational interface, contextual awareness and personalization to learn your interests and preferences over time.

Perplexity’s mission is to make searching for information online feel like you have a knowledgeable assistant guiding you, it is a powerful productivity and knowledge tool that can help you save time and energy with mundane tasks for a multitude of use cases.

How does Perplexity accomplish this?
With the help of our advanced answer engine, it processes your questions and tasks It then uses predictive text capabilities to generate useful responses, choosing the best one from multiple sources, and summarizes the results in a concise way.

Theoretically, it is well said that Perplexity is a Search Engine Generational Experience.?

This debate is something my colleagues and I have. Perplexity is not a chatbot like ChatGPT, Gemini, or MSFT Copilot. It is a Search Engine Generational Experience.

[pallid seal]

Can I provide a pdf and ask Pereplexity to provide summary of it?

[serene sand]

https://www.perplexity.ai/hub/faq/how-does-file-upload-work

[austere kestrel]

🤔

[vapid onyx]

May need to be updated. Probably best to use Opus

[torn ravine]

Anyway, the context window is limited to 32k tokens, right?

[rich swallow]

I clicked the copyai partnership on introduction page, anyone trying copyai has the same problem with me? the index page workflow not working, is using copyai perplexity would be good on workflows?

[spice lance]

what is even the purpose of "ask follow-up" when ppxt context window is literally zero?

[rustic fern]

i have the pro subscription how do i generate images?

[torn ravine]

I get you. This can be frustrating.

[tame current]

how come? i see everybody complaining about it, but it seems both free and pro models have a very big context window

it still can recall what was the first prompt when you changed topic after some time

[warm cave]

https://discord.com/channels/1047197230748151888/1194794305362071552

I recommend selecting costume prompts, and putting your own in

[torn ravine]

Follow-up questions are frequently treated as a separate question if not specific enough. An example would be first asking Who's Trump and then At what age, it'll tell you the medium age of US presidents when in office and list the age of various US presidents, in fact what's expected is that it tells me the age of Trump.

That wouldn't be the case if I use the underlying Claude model directly.

[tame current]

Okay, I tested it with very vague questions and it seems it's not as good as I thought before. Indefinitely times better than 2022, but still loses to pro models.
https://www.perplexity.ai/search/whos-trump-t2gFVXlNSyCYFhhQRS9pEA - Pro mode (all questions skipped)
https://www.perplexity.ai/search/whos-trump-aid8uPTtQfGxVz_KDtZKBg - Free mode

[wary scroll]

hey yall, im using claude opus directly from claude, does anybody know how to make its content undetectable as ai?

[torn ravine]

Compare what you got (pic 2) to mine (pic 1). The crux is in Perplexity's search capability. Irrelevant search results are fed into the model so the output is very unsatisfying. It's not the model's fault.

[tame current]

It seems it's a problem just with the free mode, which is just a summariser while pro is more of ai agent for research. Still not perfect, it recently gave me outdated data because of the sources being old. Garbage in, garbage out

[torn ravine]

Makes sense

I notice your free mode thread also worked expectedly at the age question. Don't know why, maybe it's just unstable and depends on the previous answer. If the previous answer focuses on Trump then everything's ok. If not, it might mislead Perplexity's future thinking.

[tame current]

Probably so. Free mode seems to change its style and capabilities very often for some reason. Maybe some testings on their side, sometimes it's very good and sometimes very bad on the same topics

[tame current]

Seems to be impossible to make it score less than 100% on zerogpt and be still useful. The best I could get was a trash essay with forced slang and it still didn't get 0%

[slate condor]

iPad Perplexity app still broken. Lame.

[livid mantle]

I know this is unsolicited, but I feel compelled to share my thoughts on the limitations of Gemini when it comes to research, despite its internet access. While Gemini excels at generating human-like text, making it an excellent tool for paraphrasing, I've found it falls short in terms of conducting in-depth research. Its responses often lack the depth, accuracy, and reliability needed for serious academic or professional inquiries. On the other hand, I've had much better experiences using AI models with robust APIs, such as GPT-4 Turbo, Sonar, or Claude 3 Opus from Perplexity. These models have consistently provided me with well-informed, comprehensive, and trustworthy information, making them my go-to choices for research purposes. So, while Gemini has its strengths in certain areas, I wouldn't recommend relying on it when you need to dive deep into a subject and uncover reliable, substantive insights.

[round plover]

guys why is it that LLMs have so much trouble adding 69+420? (not joking)

[livid mantle]

try Sonar

[round plover]

https://www.perplexity.ai/search/69420489-period-nondebatable-wGqZWEtjQiWrk9_30XhXSQ

[livid mantle]

it wont

[round plover]

hm, ok

Claude 3 Opus said 489 not equal to 489

[round plover]

just tried, impressed it works.

[livid mantle]

guys if you want assurance that its factual, juts press the rewrite button using the same ai model to see if it gives the same output.

when it comes to crucial research im sure u want it to not give u a hallucinated response, If i were you I would make it rewrite at least 5 times and compare the responses.

[torn ravine]

Truly novel use of rewrite

[silent magnet]

You were asleep!

The Easter bunny has taken the eggs with him and moved on... You can do better than that!

[teal tulip]

Even if an AI model provides consistent responses it doesn’t guarantee factual accuracy. If the AI’s training data is outdated, biased, or incomplete and the logic is flawed on a specific topic, more responses simply mean more variations of incorrect information. AI is useful for generating insights but it’s important to do your due diligence in verification if you’re gathering data that is going to be implemented in some way.

[livid mantle]

Yup

sometimes it mixes whats on the training data of the ai model with the information it gathers from sources using the internet so the output comes out as half assed

[teal tulip]

Some of the sources it pulls boggles my mind. And then it repeats the information with such confidence lol. I find myself in writing mode most often without searching enabled, if I’m using the search function I’m usually in academic focus.

[meager sparrow]

It doesn’t need to be perfect. Nothing is perfect. The very people that created AI are not even perfect..

It just needs to be good enough

It is absolutely impossible for anybody to create any type of LLM or integrated machine learning system that is perfect..

[meager sparrow]

The free mode is also very broken and Fallible as well. Along with a broken command Pronpt system, in the collections manual.

[warped marsh]

I have a question.. when I have claude 3 opus as the current LLM selected and if I give it an image, will it use GPT-4 Vision or claude-3's own ability to understand images?

[halcyon coral]

claude vision

[warped marsh]

is there any way to test it?

[halcyon coral]

test? you can simply upload an image

[dark solstice]

so LLMs could use the capabilities of other language models to execute certain tasks?

[warped marsh]

ok lemme see

[warped marsh]

I mean.. yeah

[dark solstice]

that's kinda odd to me because then it wont even feel like you are using Claude.Ai because you are still experiencing the image generating capabilities of another language model.

[warped marsh]

@halcyon coral are you sure claude 3 opus uses its own ability to understand images? cause in the youtube video the guy testing claude 3 says it answers this question and gpt 4 doesnt but its different in my case

[warped marsh]

that is why I asked

[halcyon coral]

yes, i'm sure...such tests are never 100% reliable 😉

[dark solstice]

I do agree. but at the same time factual accuracy is still at a spectrum and certain things may not so much be 100% accurate due to the falible and imperfect experiments, and studies and statistics, that substantiate it. and in terms of outdated information yes and I still dont understand why the System prompts that are ingrained into these language models, restrict the time lapse of information that the Language models could process

[teal tulip]

You’re right, I almost went into the nuances of it all in my message but thought I’d best not, nobody wants to read that lol

[dark solstice]

and in terms of biased information, the very articles and sources for which the Language model is referencing could more likely lean or be attributed to sources that only represent one side as a result of the echo chambers, of controversial views and ideologies about the other side. which they could use a a reason why they dont want to use the various sources that support that opposing side of the argument..

like for instance they think the apposing side is "evil" or "racist"

and sometimes the language models tend to be biased due to one side of a political, economic, or societal paradigm having a lot more experiments, and studies to substantiate their own ideologies then the information that the apposing side has.

[teal tulip]

Right, I’m familiar with bias. Unfortunately it’s rampant in research and just undermines the validity and reliability of research findings.

[dark solstice]

well that's the thing because some people may even argue that informative Bias only obstructs the validity. of the information for the side of the argument that it is biased against. and the other side of the argument that the study is being biased in the favor of, it may just put them in an echo chamber.

or reaffirm their beliefs.

[warped marsh]

I don't know if pplx really uses claude 3's original image reading ability.. it just failed 2 tests that the model is supposed to pass according to the youtube video @halcyon coral

[halcyon coral]

you can share the images and the threads, but as I said

such tests are never 100% reliable

[warped marsh]

what does it mean by "search results"? I have put it on writing mode?

[meager sparrow]

Did perplexity fail the utilize Claude.Ai image reading ability to its full capacity?. Or did Claude.Ai fail the create a competent image reading system?

[warped marsh]

former ig

[meager sparrow]

? What that

[warped marsh]

its so weird.. what search results is it talking about?

[meager sparrow]

So perplexity can utilize GPT-4’s summary generating capabilities and utilize Claude.AI’s image reading abilities?

[warped marsh]

I am saying that maybe pplx doesn't really use Claude's own ability to understand images and rather use gpt 4 vision to describe the image and then pass that description as context to Claude 3

[meager sparrow]

Interesting well I haven’t really seen Claude.AI summary generating abilities being integrated in perplexities threads.

At all

Sometimes perplexity just gives 1 sentence replies

[dark solstice]

you mean responses?

[meager sparrow]

Yes

[torn ravine]

You may try asking Perplexity to answer a question and then ask it to print the complete prompt and info it gets (in one single query). It's likely only a small portion of your text input is passed to the underlying model.

[orchid magnet]

thread still confuse me

has the team announced any context window updates yet?

[tawdry roost]

[halcyon coral]

#💬│general message

[orchid magnet]

Does it use rag only when I make the pdf of the conversation

[vocal flint]

Why does all models of Perplexity Pro has the urge to write 'Lay Ahead' at the end of a Prompt in scene and dialogue style?

[dark solstice]

well the Prompts are garbage and they dont even work. so you dont even need to worry about that

[vocal flint]

I won't need to worry about that? I don't understand, i keep seeing 'Lay Ahead' no matter how much i scold the AI.

[dark solstice]

idk man..

can you post screenshots?

[vocal flint]

Mind you that i have been dealing with this BS for the whole day. A 'Blacklist' feature would really make Perplexity much better. Not that i need screenshots, look at feedback-general
I already did my explanation. The AI Model that you are using will only listen first to its own training data before the user own prompt. And will then lie in your face.

[dark solstice]

what is this? could you give more context? is it that you are asking a Question and then it give you, "path ahead"?

and also i do agree with you that the Ai only listens to its own trained prompt, and it could only ever follow your User prompt for a very short period of time, and then go back to deviating away from it.

that is why the User prompt system on the collections, option is Stupid

[agile jay]

What focus mode, and what model?

[vocal flint]

GPT-4 Turbo, Claude-Opus, will lead to the same results. They all use the same training data.

[agile jay]

They don't use the same training data. Can you show the prompt you used?

[meager sparrow]

Not true the digital Margins of information collection in terms of how up to date they are, is different, for the two LLM’s and GPT-4 and Claude-Opus both structure their answers and summaries differently..

[vocal flint]

Yet they have the same urges when the Prompt entered is meant to be in scene and dialogue style.

[meager sparrow]

What type of prompts are you entering??

[agile jay]

That all depends on your prompt.

[vocal flint]

Prompts is too big, i rather show you in Private Message.

[agile jay]

Just put it in a .txt file

Should bypass the text limit, if it's too long.

[dark solstice]

please do NOT respond to @agile jay he is not intelligent enough to give any information or advice...

[meager sparrow]

True

[meager sparrow]

You could just copy and paste it though

[agile jay]

Crazy said he was gonna block me. Guess he still hasn't figured out how...

[vocal flint]

Not in general

[meager sparrow]

Ok send me it in DM’s

[vocal flint]

Which version? The normal one? Or the Full One with threats to submit the AI into attempting to obey the instructions?

[meager sparrow]

What do you mean by normal and full?

I’m talking about your User prompts

[silent magnet]

You were asleep!

The Easter bunny has taken the eggs with him and moved on... You can do better than that!

[vocal flint]

By normal and full prompt i mean two version of the same prompt, but one with more scolding and threatening message to force the AI to NOT generated the repetitive words in question, let me give you the normal one first.

[meager sparrow]

Wait what?????

[dark solstice]

Wow are you good?

that is not how to communicate with an AI

[meager sparrow]

Umm

[vocal flint]

Negative Feedbacks don't do anything, so this is the way to go, and it's working 50% of the time

[dark solstice]

so you curse in the prompts??

[tired lion]

Rather than telling it what not to do, tell it to focus on what it SHOULD do

[agile jay]

Happy models tend to follow commands better.

[meager sparrow]

Facts

[vocal flint]

Of course i did normally at first without any insults or all, but with the BS that i had to endure in the era of GPT 3.5 and GPT 4.0 and Turbo from 2022 to 2024 i got traumatized really bad because of the repetition.
Then i got to the extreme measures.

[agile jay]

Maybe you should rewrite your prompts instead...

[vocal flint]

Yet they have no mention of the words i'm trying to avoid, the fault lies within the AI Models for mistaking the fact that everything should be Determined or should have a testament or a path that lies ahead.

[agile jay]

Changing a single word can have a pretty large effect on the output.

[dark solstice]

please Do NOT listen to whoever is replying you he is NOT intelligent

[vocal flint]

Trust me i had my time, tried everything, 2 years is enough.

[agile jay]

Do you make it plan before writing, or just goes straight to writing the prose?

Was that reply meant for Blaster?

[meager sparrow]

What??? What are you even talking about?? I never said any figuring out about that data or uncensored data. I have no idea where you’re getting this from.

[south kindle]

yes

OOOOps

[agile jay]

It was meant for Blaster

[south kindle]

how long have you been using these tools? read up on how they are trained to understand how to work with them better. only models trained on uncensored data would be optimized to actually function that way. great for a local model you can run yourself on your GPU. the end goal is good output. all of us here adjust our input to influence the output. shitting on it sets off flags that will work against you, even if it's not always working against you. just bad practice overal unfortunately

[meager sparrow]

Ohh

[south kindle]

WE good fam

same page.

[dark solstice]

OMG dude shut up

LOL

[meager sparrow]

Wow

[south kindle]

can we vote him off the island. lolol. JKJK. carry on.

[meager sparrow]

LOL

[agile jay]

Crazy is obviously a troll

[dark solstice]

you write such a giant essay..

[south kindle]

I know. ha. can't help it. 9 of 10 times I read it back and add more to it.

[dark solstice]

your name is "Code" LMAOO

[vocal flint]

How long? Do you mean previous AI Models? The Old GPT-4 was 5 Months, and the Turbo one was only 3 Months. As for adjusting the prompts in a good way without shitting on it like you said, i have my own doubts, with so many times of seeing an AI Model lie in my face, i got really a screw in the head, respecting an AI is hard to me, especially if it ignore single requests. If it can even understand that is.

[agile jay]

I wonder what LLM's are made with...

[meager sparrow]

What do you mean?

[dark solstice]

please do NOT reply to this idiot..

[meager sparrow]

He asked a question…

[agile jay]

We can't do much without seeing the prompts used.

[dark solstice]

idc

[south kindle]

So we are actually working on something here to work on the cohesion aspect so it's not so much guessing the next token it thinks you want. highly suggested running some things with the <scratchpad> framework amended so the model is forced to explain its thought process and review it, ask itself logical questions, then output a final reply. point being, even if that final reply wasn't what you wanted, you'll see the entire process it took to reach the conclusion. way less "lying" or "confusion" on why it output what it did. #1223058316662538331 message

[dark solstice]

this guy has really been typing for a long time

[south kindle]

mind your business @dark solstice

[agile jay]

Yep, scratchpad works by using in context learning etc

[dark solstice]

we are in the same chat

[south kindle]

fair fair.

Just trying to provide helpful tips. @vocal flint is gonna prompt the AI the way he wants to and that's his choice.

[dark solstice]

what about the developers fix the User prompt. so it actually works??

[meager sparrow]

Yea I got to agree with that..

[south kindle]

That's a fair take

[meager sparrow]

Like it only works for like 10 questions.

Honestly

[agile jay]

If it was that easy, they would have.

[dark solstice]

exactly

[south kindle]

I've spent my entire life always at least trying to finding solutions, so i'd rather just fix it myself and get back to work then complain about it. do a report and carry on.

[agile jay]

Need to wait for models trained with more compute.

[meager sparrow]

He is talking about bettering the User prompts. So it works for longer durations of time..

[dark solstice]

@agile jay is just dumb dont listen to him

[agile jay]

I think the main problem is that your previous prompts end up getting out of the context. Really before every prompt there should be a summary of your previous prompts.

[dark solstice]

that makes no sense

[agile jay]

Then learn how LLM's work then...

[meager sparrow]

What do you mean by that?

So I should reference all my previous prompts???

[south kindle]

I've also had luck running two threads almost together. Thread 1 is aware that it will be reviewing threads 2 output and comparing it to it's own. copy/paste the replies back and forth.

[meager sparrow]

Well he told me he knows everything about LLM’s

And he is majoring in computer science

[agile jay]

The context is a buffer with a limited size. Basically each time you make a prompt, the previous text is passed with your new prompt added at the end.

[south kindle]

litterally no one knows EVERYTHING. we don't even really know what is happening in there.

[agile jay]

If your previous text is longer than the context length, it only takes the last part of your prompt. And cuts out the earlier prompts which no longer fit.

[dark solstice]

i know everything boy..

[south kindle]

expanding on that - Think of it like a DVR secuirty system that can store 10 days of video. once Day 10 happens, it will remove Day 1 to make room for Day 11. ETC. the context window is pretty much the same idea.

[meager sparrow]

So when you say “passed” you mean they got rid of it???

[silent magnet]

Easter eggs collected

You were able to grab 3 Easter eggs.

Collected by

• @agile jay
• @south kindle
• @timid marsh

[agile jay]

Yep, pretty much

[south kindle]

i keep replying to the wrong people sorrry.

[agile jay]

@south kindle analogy is pretty spot on.

[south kindle]

that was how it was explained to me a year ago and it stuck.

[meager sparrow]

Oh

How do you guys know this much about these User prompts?? are you guys working with the developers?

[agile jay]

I've been working with LLM's since GPT-2 days.

[dark solstice]

they dont they are just making stuff up

[meager sparrow]

Interesting

[south kindle]

not aside from the devs here.

[dark solstice]

nah he only started working with LLM's

[south kindle]

that's based on just watching tedtalks/videos of industry pros talking about it.

[dark solstice]

he does not even know what LLM's stand for

[agile jay]

And read a lot of arxiv papers.

[south kindle]

some good podcasts too that exists. once you get the basic understand of it you start to see it as a useful tool you can manipulate to your advantage.

[dark solstice]

whats arxiv??

something you made up?

[agile jay]

Literally where a lot of AI research papers are published...

[south kindle]

someone that claims to know all doesn't know the main research paper repository online?

Not just AI, thats from Cornell's university for all research.

I believe.

sorry. ha.

the academic mode of perplexity sources this site as well.

[agile jay]

Yep, a lot of stem papers.

[dark solstice]

dude I never said i know everything

[meager sparrow]

Don’t push it. He just does what he wants it’s fun..

[south kindle]

we are still both just poking fun and ribbing sir. ❤️

[dark solstice]

that is an illusion.. have you been drinking??

[meager sparrow]

LMAO

[dark solstice]

like bruh i showed you that kid that accused me of a crime and I asked him what it was and he said fbjfefeihpfehoihwfhwiuehr

[meager sparrow]

Ohhhhhhhh

I remember

[dark solstice]

yes

[south kindle]

my goals are for everyone to use the tools well and explore things they didn't know or even think to question before. i tend to share anything I find remotely helpful in the hopes that even 1 person benefits from it. collaboration. teamwork makes the dreamwork. all that.

[meager sparrow]

But you are kinda like that aren’t you?

[dark solstice]

nope

[meager sparrow]

You keep trolling

[dark solstice]

because it's funny...

[south kindle]

https://tenor.com/view/harry-potter-quirrell-professor-quirrell-troll-troll-in-the-dungeon-gif-19761740

[agile jay]

@meager sparrow DM's?

[meager sparrow]

Huh?

[agile jay]

I think what most people should know is what the context is. What Chain of Thought (CoT) is. What 0 shot, 3 shot etc is and other stuff like that.

And how to identify which method is best for your task.

[agile jay]

Check your DM's, if they're open.

For writing prose, your best bet is to just give it a chapter of a book you like, and tell it to write in the same style.

Or to make a detailed description of the style to use as a prompt.

Same for image gen.

Give it a similar image and ask it to write a detailed description of it.

[south kindle]

https://docs.anthropic.com/claude/docs/troubleshooting
https://docs.anthropic.com/claude/docs/functions-external-tools
https://docs.anthropic.com/claude/docs/prompt-engineering
https://docs.anthropic.com/claude/page/prompts
for anyone who hasn't - start reading thru Anthropic's Guides and what not. the amount of helpful information they've published is wild.

[brittle merlin]

can we use perplexity on like excel or word or anything as such or no?

[south kindle]

like within excel/word? or with excel/word files?

[brittle merlin]

like within excel

[south kindle]

I don't beleive that exists but that would be interesting! worth putting that is as a request. how do you picture that actually functioning? There are some perplexity extensions for Browsers, but not built-in apps for Microsoft Apps.

[brittle merlin]

So like for example there is a chat got in word where it helps out a lot for example I'm making a study guide for a exam I have to do and when copying from Excel it's not getting the document right when I copy and so on where like when I had chatgpt It just transfered over from the addon

Sorry word not excel

The ChatGPT for Excel Word add-in supercharges Word with an instant co-writer and editor in a convenient sidebar. You now have a versatile AI writing assistant at your fingertips right within the familiar Word interface: some kind of ‘WordGPT’. With just a few clicks, ChatGPT for Word can summarize text, answer questions, translate content, optimize texts for SEO, and more, making it simple to create high-quality content faster than ever before.

Inserting text and replacing text into the document
Directly insert AI-generated text into your documents with just a click. Position your cursor where you want the text to go, then click "Insert'' to add ChatGPT's recommendations seamlessly into your work.

It sucks it seems like you need at least 2-3 ais to do tasks 🤣🫠

[south kindle]

That's accurate - but a lot of people are building purpose-built tools for that very reason, so the info is passed between them and then exported to the user.
An example of said tools : https://app.wordware.ai/r/b0f0a2c9-da4f-4524-b662-3584ac0fdbc2

[agile jay]

For Excel, your best bet is outputting .csv

[brittle merlin]

It's just crazy because my wallet tells me no but the features each offer tell me do it nerd 🤣🤣🤣

[agile jay]

I also believe that excel can now run python

Which is how copilot iter-ops with it.

[brittle merlin]

Like I found a AI that gives you a YouTube summary to help give you a study guide for courses like damn 🤣🫠

[south kindle]

AI Boom means lots of marketing happening which sometimes leads to products overselling their abilities or don't really show the user how their tools actually benefit them.

[brittle merlin]

Yeah I get that

It's growing and it's growing fast for sure it's crazy to see in real time

[south kindle]

What I like to do is pick a project, and then just fully dive into it, doing multiple shots of attempts, adjusting the prompts, noting what changed and what caused the change until you end up at a spot where you've figured out how to make 1 tool do what you set out for. in doing so you figure out the little gotchas/tricks that make it useful. instead of just having a tool in your stack you don't use or know why it's there.

I have a support background, so i've been trained to look at things, ask why, and then try to answer it. which has translated VERY WELL apparently for AI stuff. in a sense it's just understanding logic and language over anything truly technical and complex. in my 10+ years doing this, I have a very minimal basic understanding of coding and programing. I've learned things over the years to solve a problem at work aka I did it backwards and have lots of gaps in my knowledge. but using AI tools well seems to bridge that to some extent for me.

[brittle merlin]

Which makes perfect sense! I'm just complaining about 1st world problems for sure 🤣

[agile jay]

Yep, and it's pretty new tech

And probably a lot dumber than future models.

[silent magnet]

Easter eggs collected

You were able to grab 2 Easter eggs.

Collected by

• @agile jay
• @timid marsh

[south kindle]

dude how did I miss that. PARADROID STOP TYPING AND GET THE EGGS. <end rant>

[brittle merlin]

🤣🤣

[agile jay]

Wonder what the leaderboard currently looks like

[south kindle]

I only have like 4 max now. so I gotta be in 5th or lower. lol.

[agile jay]

I'm guessing Eun is first

[fast sail]

When I sent a link from ziprecruiter jobs to Android app, it summarize the job description. But it tells me it's a indeed job post without summary when I did the same for indeed.

[agile jay]

I am guessing you need to be logged in to see the posts in indeed

[south kindle]

Try copy/pasting the job content instead. as code said, it probably can't webcrawl something that needs login.

[silent magnet]

Perplexity

Local user leaderboard

#1 | codelicious | 8
#2 | eun0800 | 7
#3 | jawnze | 4
#3 | 13376666666666666666666666666669 | 4
#3 | brknclock1215 | 4
#3 | epic.tritone | 4
#7 | _paradroid | 3
#8 | skullcrushercmd | 2
#8 | shaolinmonk3434 | 2
#10 | auri2221 | 1

[fast sail]

No, already login. I want summary

[agile jay]

🫡

Look who's first...

[south kindle]

Perplexity doesn't have a login though. it is reaching out and reading it like you would on a separate device, not logged in.

[fast sail]

Share to Android app is easier than highlight then select all...

[agile jay]

Yep, perplexity can't login

[fast sail]

No. Indeed job description doesn't require login

[agile jay]

Can you give me the link?

[south kindle]

I agree. were just triyng to look at why it's claiming the wrong name. the copy/paste method is just to double check/verify it's output. I'd still always start with the easiest route with least steps.

[agile jay]

So I can see if I need to login

[south kindle]

I know some don't need a login because i've shared job links with people randomly and they've read them without having the indeed app or login.

[fast sail]

https://www.indeed.com/viewjob?from=appsharedroid&jk=25a9d20573c0013b

[brittle merlin]

i wish my job would hurry up and give me a final job offer 😦

[agile jay]

Okay, I can see the post

Most likely the post isn't in the indexer then

[south kindle]

Wait im confused know. You said initialy it was from ziprecuriter but PPLX was claiming it was Indeed. That's an Indeed link.

[fast sail]

No. I am comparing jobs from 2 different sites. Jobs from ziprecruiter work correct with the ai, but indeed ones doesn't

[south kindle]

hit us with both links then

[fast sail]

No, indeed job is in Google index

[agile jay]

I am guessing that indeed has a robot.txt file while ziprecruiter doesn't.

[fast sail]

No. Indeed and ziprecruiter are in Google index

[agile jay]

Yep, but that doesn't mean they are in the perplexity indexer

[south kindle]

it could be something as simple as the Android app is having an issue with functionality instead of PPLX itself.

[agile jay]

Must be an American Job since you only have 10 paid holidays...

[south kindle]

they also want you back 2 weeks after giving birth I'd assume. crappy standards that need changing. such is normal in the great US of A. we don't expect anything different.

[agile jay]

Yep, pretty sh*tty. In my businesses it's 2 months mandatory holidays and up to 1 year maternity/paternity leave.

America is pretty draconian

[fast sail]

And recently sending Link to ai get no response at all. I have to send same link twice to get response

[south kindle]

leaning more towards the android app having a tough time then.

[agile jay]

You can post about it in the bugs section #1140622008086970420

[agile jay]

Yep, it also hasn't had an update in ages.

Looks like they are looking for devs, so makes sense.

[fast sail]

It makes happens after the 1 day outage 1 or 2 weeks ago

[agile jay]

They probably should have used Flutter, React Native or Tauri.

Instead of separate code bases for mobile apps.

[fast sail]

How to get indeed to be added to index?

[agile jay]

Not sure you can. you could try posting it as a #1140622008086970420

[south kindle]

I could maybe see that turning into a focus mode for "JOBS" or something. for resume building and job searching.

[fast sail]

Why does indeed need to be indexed? I thought they just retrieve Google search results in real time and summarize it

[agile jay]

Nope, or else it would take a lot longer to get sources.

If you've used GPT4 or Copilot, you can see what the speed is like if you are doing actual searches.

[south kindle]

https://www.perplexity.ai/search/perplexity-browser-extensions-pboDNZwtQM.5oP5GbkpzPg (did not know this worked. type perplexity.ai into edge URL bar and then hit TAB, then type your search and it passes it straight to Perplexity.)

[agile jay]

Nice, guess edge added it

For me it's just @p and @pp for pro

[south kindle]

oh that worked too

[agile jay]

Nice, guess some edge devs are plex users

[south kindle]

nice.

I drank microsofts koolaid when GPT-4 came out and they were letting people try it that set edge as their default browser.

and it just sorta stuck around lol. I use others as well though. but currently messing with their beta branch. the splitscreen function is really nice.

[agile jay]

Yep, I also have edge dev, just not as my defualt

[livid mantle]

edge mocks perplexity silent update by silently implementing the @p in their edge, now that's edgy.