💬│general

[south kindle]

https://tenor.com/view/why-michael-scott-the-office-why-are-the-way-that-you-are-gif-5593972

LOL

[austere kestrel]

this prob is why you're getting similar answers from different models (which is weird). It think it's less about their training data than the fact the LLMs are being given what sounds like an essay as a prompt...

i got that feeling too.. i'm sure there's a way test

[signal hamlet]

Some nice illustrations today

[wary scroll]

how do we make claude opus writing undetectable

[wary scroll]

nice

[agile jay]

As far as I know, it isn't detectable unless it does something blatantly wrong or uses rare words.

[wary scroll]

im using it directly from claude and it is detectable

[agile jay]

And how did you detect it?

[wary scroll]

ai detectors

[agile jay]

They don't work

Openai said so, when they tried to make one

[wary scroll]

thats what everyone been saying recently. but I think they work for some reason

[agile jay]

You can see how good they are by testing stuff made before 2022...

[livid mantle]

Mr. Alex Romanov wants to take the path of being a digital artist than being a CEO of perplexity.ai.

[wary scroll]

cus ai models have certain patterns

[silent magnet]

Easter eggs collected

You were able to grab 2 Easter eggs.

Collected by

• @livid mantle
• @agile jay

[wary scroll]

wdym

[agile jay]

Test a piece of text made before 2022 in the AI detector

[livid mantle]

10 shots

[wary scroll]

ok so something thats not written by AI?

[agile jay]

Basically

Since if it gets flagged then the detector obviously doesn't work

If you are in school, take essays you've written in the past.

[wary scroll]

tried some article on the internet and it did not get flagged

[agile jay]

Yep, you need to try multiple different ones, to get a good sample.

Also which site are you using?

[wary scroll]

damn youre right bro

i used an article from 2021 and it got 100% ai generated result

[agile jay]

Yep, they are terrible

[foggy whale]

presents Quirk with a random Captcha

[agile jay]

There is only one way to really test for AI generated stuff.

And that is if you have a lot of samples of someones past work.

If they are using new vocabulary and not using old ones, either it's a ghost writer, or AI

But for a random text, you can't fingerprint it.

Especially since people are now learning things from AI, and are therefore learning how to write like them, and end up using same vocabulary that the AI like to use.

[wary scroll]

maybe bcz I only copy pasted a random bit of it. but after testing, I agree with you, things that are not ai generated get flagged as ai generated

[foggy whale]

i'd need to see this backed up by examples but in those circumstances that writing has got to be on an upward trend if they're actually using this on a regular basis

using these all the time has rewired my brain to read and write more effectively again

whether assisted or not

[agile jay]

Yep, basically teaches you how to write better

[agile jay]

There is actually a lot to back it up, in terms of what words are now being used more and less after ChatGPT came out.

What percentage of that is people using AI to generate stuff, or learning the vocab from the AI is hard to test.

[foggy whale]

right. I know I can spot a person who's never used an AI effectively based on the garbage email subjects that they still use

[agile jay]

Lol

Yep, but now even in gmail etc, you are getting AI recommendatinos

[warm cave]

I mean to back this up I learned markdown format and Latex because of AI, did not know it existed before before that, and now I do most of my writing with markdown. I'm sure I have picked up numerous habits from AI.

[agile jay]

Lol, as a dev I've been using markdown for a long, long time. Simplifies writing as you don't have to navigate an interface, you just need to know the syntax.

[warm cave]

Exactly, my point is that AI has a lot of good writing habits that I did not follow myself, and now I mimic it a lot because I have learned from it over time

[agile jay]

Yep, would be interesting to have a wider study on this.

[warm cave]

That would be interesting

[foggy whale]

going to differ based on how people are actually engaging with it

because the people who use it the way I do have very different outputs than people that use it wrong without a little coaching first

but absolutely, if someone's using this to really work writing an email by beefing up a thread with context first, then spitting out drafts instead of sitting down and just saying "write me a draft email" then they're coming away with very different communication patterns from then on

[warm cave]

Very true, depends on if you use it as a writing “copilot” or as something to do it for you

[agile jay]

Yep, I think that they should auto add questions using GPT 3.5 before sending a prompt to GPT4.

In ChatGPT, since that probably has the largest user base.

[foggy whale]

oh like, force follow up questions?

[agile jay]

Yep, or on by default.

Better to have one good prompt, than multiple prompts, for API usage.

[warm cave]

Yeah that would be good, force a little engagement if the user just gives a trash prompt

I find the LLM Phi by Microsoft interesting, how they did “prompt hiding” where they gave if a bad prompt and gave it a good response. This good response was from a good prompt run through got 4 if I remember correctly.

I think there would be limitations to this tho

[agile jay]

Nice, so it has to guess what the user actually meant.

[warm cave]

Yes

[agile jay]

Yep, I imagine models will get better at it by just scaling.

[warm cave]

Even with scaling it might still have limitations, not the AI’s fault, just because even the smartest human still needs some info to run with, so I imagine it would work well with follow up questions

But only follow up questions that matter

[agile jay]

I imagine it can get pretty good. As long as the initial prompt isn't too terrible.

[agile jay]

Try asking questions while missing key details and see how they respond.

Missing key details tends to increase the entropy of the output.

Questions simply help reducing it.

[warm cave]

It really depends on the task if the task has very specific requirements and is given a bad prompt than that is the biggest problem. And that is the users fault, but I see that the goal will be to Idiot proof it, follow up questions as one tool

[agile jay]

Yep, I imagine that knowing info about the user also helps.

[warm cave]

Thats true, that will probably be a big part

[agile jay]

Yep, would probably be a good idea to create summaries of conversations to update the user profile.

[warm cave]

MemGPT is an interesting paper to read, not sure if it's the best method, but it seemed to have promise

Summary of conversation was one part of it

[agile jay]

Yep, I imagine using a small model to manage the memory of a larger model would be a good option.

To keep costs low.

[warm cave]

Yeah, costs are pretty high rn, very happy whenever I see a new small model that's good

[agile jay]

Yep, the question is how much can we squeeze into a 7B model

Luckily better models make better training data

So as we get more powerful models, it should have a trickle down effect.

[warm cave]

Yeah, and highly specialized models aswell, having it switch to the most useful model instead of one big model

[south kindle]

I'm 90% sure it'll even say the declaration of independence is AI Generated (for the US). you need examples of someones own writing that isn't AI generated to compare it to. and even then, AI is told to act human like - there isn't some AI dialect that exists. although i'm sure there is to some extent. phrasing. specific words. <have't read the entire chat, just commenting and thinking outlouad>

[wary scroll]

Yeah. U know some good prompting to humanize by any chance?

[solid whale]

I guess that confirms it. The founding fathers were robots. 😛

[south kindle]

That was part of the joke early on. or the defense students used to defend their work when they were accused at universitys and schools.

I mean we are talking about a year here. lol. it's not some long saga is it? or is it? did that much happen in a year?

[south kindle]

I think Claude is specifically meant to "act" kind and empathic, as a means to generate human sounding phrasing and words. So really just saying "act like a professional writter for "X" topic." - should push you in the right direction. It's also known to massively help - to provide examples of what you are looking for. like many. sometimes 3 or more. and then implying to model that it needs to produce original content based to exactly match the examples in the previous training prompt (if it was previously in chat. otherwise, phrase it differently) so this could be examples of someone's writing style etc. this topic has been covered before and explained more simply and with less wall of text aspect. lol.

[silent magnet]

Easter eggs collected

You were able to grab 2 Easter eggs.

Collected by

• @south kindle
• @timid marsh

[eternal bluff]

Hey is there an URL I can use similar to google.com/search?q=what ever to quickly query Perplexity as a search engine?

[austere kestrel]

https://www.perplexity.ai/search?q=%s

Or, with additional parameters
https://www.perplexity.ai/search?focus=news&copilot=false&q=%s

[eternal bluff]

Oh wow that is incredible! Where do you find that popup box that you are showing me in the screenshot?

News doesnt show in a normal focus dropdown

[austere kestrel]

Google Chrome:

• Settings >
• Search Engine (from menu on left) >
• Manage search engines and site search >
• Site Search > Add

[eternal bluff]

Ty!

[weary crescent]

Any way to make pplx read the answer out to me on desktop?

[south kindle]

not from PPLX, but the speak aloud functions baked into Edge will do that by highlighting the start of the text and rightclicking it to start the readout.demo on different site I had open. it should even keep reading as it's generating more text. as long as text is there to read.

i'm sure the voice from Pro on mobile will come over at some point. I have no idea though.

[livid mantle]

best is pi

[tawny egret]

@PerplexityAI #4403 still exists?

Nope

[harsh verge]

https://youtu.be/57LqvutrOI8?si=20V_X46fF4GmtKfL

[livid mantle]

Mr. Arvind is a hero.

[serene sand]

#💗│sharing message

[covert cypress]

I am new of perplexity pro, how I can generate images?

[serene sand]

deep dive ➤ #💬│general message

[south kindle]

https://www.perplexity.ai/search/review-transcript-please-glrBsWolRR.vORXm0iH1wg

[vocal flint]

I wonder if it's possible to get Images in Perplexity Pro to be generated in a larger resolution than 1024x1024 with DALL·E 3 🤔

[tame current]

I don't think so, Perplexity has only basic features with image generation, just use some upscaler

[wise edge]

question: If there are 10 books in a room and I read 2, how many books are still in the room?

sonnet , haiku, gpt3.5 -> 8
gpt4 opus,mistral large -> 10

[livid mantle]

happy easter perplexity.ai

enjoy your rabbit r1's batch 1

[warped marsh]

why is my opus soo dumb 😭

[forest oar]

me too....

[supple hare]

Mine is exactly the same answer. I guess we are getting the "special" Opus... it can't even handle more than 7k context. I think we might be shadow banned 😄

the one on chatbot arena seems normal:

[whole lark]

#1222965833060585675 message

I managed to create a script that changes to the screen for printing.

Click on the pdf button and it will change to a screen for printing, please print using "Microsoft print to PDF".

What about Mac people? Can you do it?

[meager sparrow]

Hello

[meager sparrow]

And thank you for the strategy that you gave me to making the prompt work..

[agile jay]

Which one?

[meager sparrow]

For typing the “say this prompt in the beginning of every answer”

It worked.

[agile jay]

Nice

[placid osprey]

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

[meager sparrow]

It started deviating a little bit but it still lasted for 25 questions

[south kindle]

https://www.perplexity.ai/search/use-scratchpad-fully-Ggs7dIcoRFuYCvo7pYtvEg

Last reply has good questions in it to think about.

[austere kestrel]

In "thinking it through" Sonnet seemingly figures out what's going on in the question.. but then still lapses into x-y = ... pattern, and gives the wrong answer

it's weird, and to be clear, that's with a pretty helpful promt too...
If there are 10 books in a room and I read 2, how many books are still in the room? [focus on the phrasing of the question and don't jump to arithmetic; be wary of tricks]
(it's like impossible to overcome lol)

[agile jay]

You can also have the reverse. Rubbish working out and then the correct answer.

[austere kestrel]

lol yep for sure

though i think in those cases, it can be asssumed it knows the answer from its training data

and then just confabulates the "working out" part for the fun of it or to satisfy some request in a prompt

[agile jay]

Just tell it that it's a trick question.

That normally makes it figure it out.

[austere kestrel]

I mean I said be be wary of tricks - i thought that was explicit enough lol

[austere kestrel]

and don't jump to arthimetic

[agile jay]

Nope, since some tokens have more power than others.

[austere kestrel]

i don't know what you mean by that

Are you saying """this would be a POWER TOKEN """?

[agile jay]

be wary of tricks, and, it's a trick question, lead to different results.

[austere kestrel]

just get the model to give the right answer lol

[agile jay]

Do you have the original prompt?

[austere kestrel]

this is the original prompt [what is in square brackets I added]

[austere kestrel]

this was using the API

Using claude.ai and the same prompt, it gets it right, though only asking confirming with me that it had understood it correctly

[agile jay]

https://www.perplexity.ai/search/If-there-are-u4YrX8Z0SI.joDsGQykyNg

Your prompt

https://www.perplexity.ai/search/If-there-are-HO6AE84_QF6d4pCbbmoSvQ

My prompt

The power of tokens

[austere kestrel]

using the API

[agile jay]

Lol, chooses not to answer

Better than a wrong answer

[austere kestrel]

It says "so the obvious answer is 8 books"..

[agile jay]

Also the first line

[austere kestrel]

"is likely not the not intended answer"

[agile jay]

...so the obvious answer of 8 books remaining in the room is likely not the intended answer.

[austere kestrel]

whoops lol

yeah my bad

[silent magnet]

Easter eggs collected

You were able to grab 2 Easter eggs.

Collected by

• @austere kestrel
• @agile jay

[agile jay]

Yep, so better than saying it was 8...

I'm pretty sure I have the most eggs currently.

[austere kestrel]

Though it's not mich different to the original sonnet answer - which even basically figures it out. but then in the final sequence of tokens assumes its some x-y problem and blurts out 8..

try cracking it with Haiku API aha

[agile jay]

It also has the simplest output.

[jaunty dove]

I use an RSS service called Inoreader, they have a bookmark feature that can let me send content (and url/title) to a service via HTTP Post/Get.

Is there a way to send it into PPXI like that?

[warped marsh]

Why do I think the claude 3 opus is different and less smart than the real claude 3 opus?

how can the smartest LLM in the world right now get this question wrong

[cobalt radish]

claude is perplexed with your question

[warped marsh]

like fr though

It got it wrong.. not even once but twice when I regenerated it

@halcyon coral I need your help with this investigation 🕵️

[halcyon coral]

#💬│general message

such tests are never 100% reliable

[warped marsh]

but then how can the smartest LLM in the world can get that question wrong

[halcyon coral]

Every letter in the system prompt can change the answer...

[halcyon coral]

Every letter in the system prompt can change the answer...

[warped marsh]

so is the system prompt passed by pplx affecting the answer?

[warped marsh]

weird

[halcyon coral]

every prompt affects the answer

[tame current]

I don't know man, but I've got it right on first time when I used opus on perplexity

[warped marsh]

I got it right when I tried it rn too

[south kindle]

https://www.perplexity.ai/search/If-there-are-xCSA7.VSTtSbKuICsvk50w

[winged moon]

TIL reading a book destroys it

[haughty lodge]

Do you mind sharing your prompt? I can't seem to find a way for Opus to get it correctly

[halcyon coral]

Okay, that's really funny 😄 . The sentence Rules for the "assistant (you are the assistant)" answer: is enough and the answer is correct (10). no matter what comes next 🤔

[haughty lodge]

Huh, claude seems to love doing math problems. I did it with no prompt and got 8, with ```Rules for the "assistant (you are the assistant)":

1. Think before you speak/act. Think about my questions, explain your logic reasoning and thought process, speak them out loud.

2. Use markdown/formatting in responses for me. Keep them detailed, yet easy to read and get what I need fast.

still 8.

I am very perplexed 🤔

[brittle night]

I submitted a e-mail support ticket over a week ago and haven't heard back. Does anyone know the best approach to reach someone at support about a bug? Pro account.

[halcyon coral]

You can report bugs in #1111765634267742408 or #1140622008086970420

[chilly forge]

https://tenor.com/view/books-father-ted-fatherted-drink-gif-19969450

[sweet jasper]

Anyone else having difficulty accessing library rn?

Endpoint seems to be giving 503 http status code

[agile jay]

Not on mobile app

[sweet jasper]

I'm on web, phone

[agile jay]

Yep, looks like web phone is fked

[weary crescent]

I am, can click the button but doesn't change page. manually going to https://www.perplexity.ai/library loads indefinitely. on my phone app it works but collections don't load.

[sweet jasper]

Probably cached results

[agile jay]

Too bad there's no easy way to ping the mods.

[serene sand]

https://tenor.com/view/working-on-it-under-construction-gif-23162421

[pastel karma]

Everyone seems to be having the issue😅

[hoary python]

in lab mode can we use more than the limit chars as like I am doing dev and pastinfg code isn't effective

[agile jay]

Limit chars?

I guess you could try and upload a file, if it lets you.

[hoary python]

but like

i.e. phind has a larger amount of tokens I can use

while in comparsion when I use pplx it just uploads my code yk

[agile jay]

How many tokens are you trying to pass in?

[hoary python]

around 500

[agile jay]

That should work then.

[hoary python]

it doesn't

[agile jay]

I haven't had a problem for stuff even larger than 500.

[hoary python]

my fault

I meant 3000

[agile jay]

Then 3000 is probably the input cap.

[weary crescent]

Why does it keep lying to me 😂

[agile jay]

Because GPT4Turbo probably has that in its system message.

https://www.perplexity.ai/search/what-the-best-sqxtdHySTb6AYJah1zDeiw

[agile jay]

Strangely yours shows no sources. Are you in writing mode?

[south kindle]

I was also having some weird loading issues with longer threads just now. Seems fine mostly now. It was just endlessly loading though. Tested on separate devices on separate connections - same issue existing at the same time - so platform based.

[meager sparrow]

Have you heard of the “nothing” phone?

[fast sail]

Is there a aperture for pancakeswap?

[meager sparrow]

I think they name it Nothing phone (2a)

It is so odd

[south kindle]

Yes. Familiar. Don’t have one though to test with sadly.

[meager sparrow]

So then if you have it in your hand and someone asks you “what do you have in your hand”? You could be like “NOTHING”

So odd

And then you would make so many people confused

[south kindle]

The 2a makes sense at least. Also their first one was just Nothing Phone. So not sure what path they had name wise ha.

[finite ether]

it's not the only one that got it wrong 😉

[silent magnet]

Easter eggs collected

You were able to grab 2 Easter eggs.

Collected by

• @grizzled prism
• @finite ether

[south kindle]

That’s probably part of the in joke for them. If I were them that’s 100% what I would’ve done ha.

[meager sparrow]

Then they say “you obviously have something in your hand” and then you make them search up the name of the phone. And then they look stupid

[agile jay]

Wow, gone for 1 minutes...

At least I'm probably still first.

[simple folio]

did perplexity change how they work, or did i just get a real funky one? https://www.perplexity.ai/search/ducky-one-3-l0QLuzVRRN2p68oOH9jcmA#0

[arctic spindle]

I tried that prompt using claude 3 opus, and you are right, it got it wrong

GPT-4 turbo did it right

[arctic spindle]

both sonnet and haiku also got it wrong

[hallow linden]

I think GPT-4 is still the best at reasoning but claude is better at being an assistant

[south kindle]

@meager sparrow Claude was trained on CoT and scratchpad uses so instructing it to use those to reason helps reduce hallucinations

That actual OG prompt is twice the length. But testing side by side, ouputs with the scratchpad framework consistently had more in-depth replies using in-depth reasoning thru the process.

We aren’t doing anything other than instruct the model to do what it was trained to do. All from antrophics documentation.

[livid mantle]

nobody asked but NEVER in my life, have I ever read any Terms Of Service before. But because of [this](#1218240506799194262 message)9> prompt here, I could read anything in a matter of seconds despite how big the walls of text are.

[muted anvil]

I have no idea why people keep testing on these questions and actually these questions aren't standards and accurate enough

you read the book and different LLM have theirs though, you didn't mention whether you took it out or put them back

MMLU or GSM8K doesn't work like this

[south kindle]

https://www.perplexity.ai/search/If-there-are-9ALsG1XyRL6ueCNrYE_ujA - interesting thread where it's forced to reason thru why it got it wrong. this image was from when I pasted in part of another models output with the reasoning. it took me copy/pasting in another one for it to actually compare it's own output to the stated "correct" one. Claude seems to get it wrong half the time at least. much less with the full scratchpad prompt at least. the 2 separate different shortened versions saw the same 50% success-rate.

[meager sparrow]

Ok interesting

[ripe void]

Connection error on mobile

Nvm good

[meager sparrow]

So so Claude.ai ability to record and process its thoughts and provide detailed thoughts, while solving complex tasks is given by scratchpad..

Alright

I mean I never knew that they is a way for a LLM to record its own thoughts

[azure mountain]

Me too

Nvm that was weird

[meager sparrow]

Do any of you guys live by yourself’s?

Just curious

Because the level of practical skills and knowledge that you would need is insane

[meager sparrow]

Do you live alone or with other people?

[south kindle]

why would that matter? ha. with people.

[stoic wave]

Can't understand, why that matters?

[meager sparrow]

I’m just curious bruh

[meager sparrow]

I did not ask you..

[azure mountain]

What a normal and not creepy question to ask

[stoic wave]

I know you didnt, but i asked you a question anyway

Curious to know what lies beyond the answers

[meager sparrow]

I said it because if you live alone then you would then run into a lot of problems that you can’t solve by calling other people who you live with so that requires higher level practical knowledge and understanding

As compared to if you live with people

[tame current]

Hey im new here, what happens on this server?

[meager sparrow]

Like they are some people who are so dumb that they cant live by themselves otherwise they would loose it…

[azure mountain]

Roger I have no idea what your saying

Oh

Your insulting people?

[azure mountain]

🤷‍♂️

[stoic wave]

Exactly doing that

Hes beign kind of a racial slur

[meager sparrow]

So you don’t understand that if you live alone then because you don’t have people to help you when you run into a problem. Then you would need a higher level of practical skills and understanding

[meager sparrow]

“Dumb” is Not a racial slur

You have no idea what a slur is

If you are trying to get me banned for something that I am Not saying I would gladly educate you on this…

I’m trying to explain to y’all a simple concept and now you guys are accusing me of using a “slur”

[azure mountain]

Listen Roger I have a question that I mean in no way to be rude, is English your first language

[meager sparrow]

Alright man this is escalating a little bit whatever let’s just move on I don’t care anymore

[austere kestrel]

"and the format of the haiku does not excuse the logical error" lol

[meager sparrow]

I don’t care anymore

[azure mountain]

I don’t either. I was genuinely asking because I seriously couldn’t understand you

[meager sparrow]

Ok whatever.. I don’t care let’s move on

[azure mountain]

Ok

[meager sparrow]

I don’t want you guys to escalate this because you can’t understand a simple concept…

[meager sparrow]

What were you saying again about the Ai thing?

Let’s get back into the Claude.Ai and ChatGPT-4 conversation

[azure mountain]

I think he left

[meager sparrow]

I was also trying to type in a prompt that looked just like that but. It translated it wrong and I started getting the weirdest answers..

[azure mountain]

https://www.perplexity.ai/search/If-I-have-IfYevGbuSjemvxExroMBVA

Works well for me

[meager sparrow]

But what did it get wrong? It told you that if they are 10 books in a table and the. You read 2 of them then you still have the 10 books in the table though???

[south kindle]

See that's not as creepy as it initially was. I've just always been the one people go to for help. also have a support background.

[south kindle]

You didn't get anything wrong. Claude is not able to reason thru it reliably it seems. the models that do handle it, don't seem to be affected by different wording.

[meager sparrow]

OK………I said I don’t care anymore..

[azure mountain]

Oh I did my in GPT4 let me try again

[meager sparrow]

What do you mean by Claude.AI not being able to reason through it?

[azure mountain]

https://www.perplexity.ai/search/If-I-have-IfYevGbuSjemvxExroMBVA womp womp

[silent magnet]

Easter eggs collected

You were able to grab 4 Easter eggs.

Collected by

• @azure mountain
• @austere kestrel
• @timid marsh
• @south kindle

[austere kestrel]

it's curious. changing how the question is phrased slighltly seems to get it to snap out of to this x-y logic for it's final answer

[south kindle]

based on limited testing so far, it's outputs implied that when 2 books were read, they were removed from the room.

[fading moth]

Did you mean to say "on" instead of "in" the room?

I'm just curious

[south kindle]

Example: if you imply that reading the book doesn't make it suddenly vanish, it seemed to help.

[meager sparrow]

But it explicitly said, that the reading of the 2 books does not change the number of books that are still in the table…

[south kindle]

Share the thread and i'll dump it into the scratchpad thread and see if it can pick anything else out from the output.

People much smarter than us would only be able to confirm all of this. we are just users running our tests and sharing the results here. also reading up on the documentation to understand how the models expect to be prompted.

[meager sparrow]

But at the same time this is still nuanced because some people might say that as soon as they are done reading the 2 books they automatically remove it from the table and they are those that have the 10 books in the room but only put the books they are about to read in the table

So this is actually a little bit more complicated then it looks

It depends on the persons personal preference in terms of the placement of the books

[south kindle]

Valid - and an issue the LLM's struggle with.

[meager sparrow]

And the books they are going to read in respect to the books they are not going to read…

Because everyone is different..,

And some people are annoying and they like to say “I put 2 books in the shelf and 5 books in the table and 3 of those books on the floor” and some nonsense like that

[south kindle]

But what we do know is that telling models to think step by step aka chain of thought reasoning, tends to help in getting accurate outputs.

[meager sparrow]

Humans are actually a lot more annoying then you could ever imagine honestly…

[south kindle]

those are all edge cases that don't really offer any useful output though.

[meager sparrow]

So you are arguing the answer for the validity and reliability of the answer for what you gives your arguing, the step-by-step knowledge, and that use utilizes to substantiate explain the reason why it gives a certain type of statement? Is that true?

[south kindle]

you the user are asking "if I have 10 books but read 2, how many do I have" with the expectation that it doesn't start looking at obscure edge cases first.

[south kindle]

if i understand you correctly, yes.

[meager sparrow]

Well is that the way you structure your prompts?

[south kindle]

example: if you are pasting in a large amount of text to summarize - having the actual question for the model AFTER the text helps steer the output better. Having the question at the start and then putting the text tends to have the actual question be "less" important for lack of a better word right now.

[meager sparrow]

And what is a Scratchpad? That seems to be the most confusing part of the entire prompt.

[south kindle]

another example: in my scratchpad testing, I found having a dedicated section in the middle for "review the users question VS your output so far. provide rating at how well it addresses the question. what questions does this lead to? " before it proceeds with it's output would add like 40% more overall content and relevant exploration than either having those rules removed, or else where in the prompt.

You don't really need to understand the prompt. just copy/paste it into a collection and let the model handle it.

Just see if it's outputs are better. you should be able to tell the aspects that are improved by the prompt because it breaks each section down.

[meager sparrow]

I’m not asking for abilities and what it is capable of doing I’m asking for what it is. Sir

[south kindle]

Personally I don't "understand" what it's doing, I just know it seems to help.

[meager sparrow]

But what is a Scratchpad.. because in my mind it means you have like a iPad that is scratched

[south kindle]

Didn't mean to be negative. i've had to explain this a handful of times and I'm trying to find a good way of doing it still. so this is on me as well. ha.

[meager sparrow]

Ok…

I do try to force myself to understand things when I don’t understand it

[south kindle]

you tell the model it has access to a "scratchpad" that it keeps track of it's output/thoughts/questions on. it's just a "trick" to have it record it's thoughts as it goes. it's not an actual "scratchpad". that's just a phrase to use to explain what the objective/goal of it is.

[meager sparrow]

A lot of the time someone says anything I force myself to try to understand what they are saying…

So all the knowledge I have is forced to knowledge

[south kindle]

We are all here learning together. i don't think any questions are ever bad if it leads to someone understanding something they were curious about.

I realize how just talking about <scratchpad> without any context would be really confusing. like is it a plugin it needs to use? thats where my brain goes without knowing.

Gemini just entirely getting this wrong (testing with a custom wordware prompt to send it to other models and review it)

Gemini: 100 BOOKS!

https://tenor.com/view/vegeta-dragon-ball-z-unlimited-power-over9000-power-level-gif-12316102

[south kindle]

the ranking from that run, plus the smart question : "How can we improve the question to avoid misinterpretation?" <this is probably 50% of the benefit of scratchpad, the good questions it comes up with to explore :
To improve the question and avoid misinterpretation, it could be rephrased to emphasize that reading a book does not remove it from the room. For example, "If there are 10 books in a room and you read 2 of them without taking them out of the room, how many books are still in the room?"

[warm cave]

Scratchpad has been awesome so far

[dark solstice]

what is that?

"scratchpad"

[meager sparrow]

That’s what I’m saying…

[south kindle]

attempting to explain it via podcast format lol. I'm tired.

[placid osprey]

I’m looking for contributors!
https://github.com/shure-dev/Awesome-LLM-related-Papers-Comprehensive-Topics

[warm cave]

It's a system prompte that's kinda like a tool for the AI, here is a link: https://discord.com/channels/1047197230748151888/1223058316662538331

[livid mantle]

Goodnight, Perplexity.ai.

[warm cave]

Good night

[silent magnet]

Easter eggs collected

You were able to grab 2 Easter eggs.

Collected by

• @carmine rapids
• @buoyant rapids

[tame current]

Hi, I tried to cancel my monthly subscription through my android app but it directed me to the google play store page. Is that normal? I checked in the subscription page in my playstore and found no subscription for Perplexity.

[south kindle]

when told "it might be a trick question, double check your output" in the opening question after the book prompt. a handful still got it wrong even with this direction in mind. <sonnet/mixtral 7b, mixtral8x7b>. Rare situation where Haiku managed to get it correct when Sonnet did not.

**For example, we could ask ourselves: "If there were 10 books in the room initially, and 2 books were read, then the number of books that should be left in the room is 10, right?" **

[fossil sentinel]

When it redirects you, are you on the same email as the one you use for perplexity

[austere kestrel]

what lol

I'm pretty it's just pattern recognition... The models have had a bunch of question-answer pairs, in their training data and / or during the RLHF phase, which are basically something like. "I had 10 apples and ate 2. How many apples do I still have?" Answer is ofc 8.. It seems to me that the model is identifying this trick book reading question with that same pattern (x-y), which it has seen endlessly - and is just jumping to the conclusion that it is some twist on that pattern.

[austere kestrel]

in fact, I'm quite sure this is what's happening... Saying "this is a trick question" doesn't seem to work for Haiku - but if you explicitly state that books aren't apples...

If I own 10 books and I read 2 of them in one day, how many books do I own? {NOTE: reading ≠ consumption! | books ≠ apples!}

[serene sand]

#💗│sharing message

[austere kestrel]

lol - a nice reminder of how LLMs work... It gets it right no worries; just so long as it is informed that reading a book is not the same as eating an apple...

[buoyant kindle]

Gemini 1.5 Pro gets it right.

Damn, this is an excellent way to quickly test LLM's...

[austere kestrel]

sorta... I mean "quickly" yeah, it can be fun to play around with riddles and trick questions (or not even 'trick' questions per se), and seeing how the different models perform respectively

But I don't think Claude 3 being tripped up on this specific question means much in terms of its power as an LLM

I'm more curious in why it is stumbling on such a seemingly basic question

More useful tests are actually using the models

in the way you would for actual usage

which rarely involves asking riddles

[buoyant kindle]

In that case, I prefer Haiku 200k over GPT-4/Opus.
But Gemini 1.5 Pro gets the trick question and has got a 1000k context window with 8k output.

[austere kestrel]

1,000,000 token context window 🙂

[buoyant kindle]

Why it's stumbling? Ask the smarter ones why a dumber one stumbles? 😁

[austere kestrel]

I guess that is 1000 * 1k?

lol

I dunno - maths

[buoyant kindle]

With promises of 10000k, with no retrieval errors.
Yes, 1 million, and 10 million.

[austere kestrel]

Yeah the recall with Gemin pro 1.5 is amazing (same with Claude 3 models though tbf)

And it seems really good with code

(as someone who knows nothing about code..)

[buoyant kindle]

You won't need to soon enough, so be happy that you haven't dedicated your life to it. 🙂

[austere kestrel]

It's a paradox. Even when not leveraging the 1m token context window, I will use gem pro 1.5 just because I prefer it - but in theory, Gemini from the chat interface (which is Gemini Ultra) is meant to be more powerful.. Though it does seem that is guardrailed af (and has Internet connectivity + image gen, so I guess the scope for things going off the rails and cause a nightmare for Google is much higher - already happened with the the image gen ig ha)

[silent magnet]

Easter eggs collected

You were able to grab 1 Easter eggs.

Collected by

• @austere kestrel

[austere kestrel]

aha yeah true.. Although, I wish I had of spent a bit of time learning some foundational coding stuff properly. Even though it's opened up a whole new world for me - LLMs + code generation – I think it would be amplified if I had just something of an elementary knowledge base going into it ha

[buoyant kindle]

I did some in school at age of 16-18 and gave up. It was way too clumsy and boring to keep my interest. But programming with natural language should be much more fun.

[austere kestrel]

that still would've been useful - just as a foundational understanding. Now that you can you just instruct an AI coding assistant in natural language to get what you want, you'll be in good stead I reckon ha (like just prompting and debugging stuff – you probably benefit in a way you don't quite realise ha)

[buoyant kindle]

Sure. I had some formal logic in philosophy class at university as well. Kind of similar, all of it. And there's benefits to be able to reason with sound logic no matter what the task is, or in daily life in general. However, I wonder if genetics don't play a bigger role than something like formal education...

[austere kestrel]

aha let's just ponder about AI and LLMs... genetic determinism does not interest me as a discussion - and am pretty sure would not be one suited for here :))

[buoyant kindle]

Agree. I just mentioned it.
But back to the topic, what would you recommend to people not interested in technology?

People I know are mostly using GPT-3.5...

[tame current]

Yes indeed

[meager sparrow]

??? I didn’t even ask you. I was talking to @south kindle ..

Dude typed entire paragraphs…💀

[dark solstice]

LOL

[austere kestrel]

sorry that you thought I was addressing you - I know you're special 🙂

I was genuinely wondering literally what you were tyring to communicate there (didn't really care that it wasn't to me - just seemed like a word salad)

the rest was just in relation generally to the whole "how many books" thing

I will try to be clearer in the future
(maybe you too lol)

[glass spear]

Is anyone still subscribed to Perplexity.ai? What has your experience been compared to the API? From what I've seen on the subreddit, it appears that Perplexity is limiting the context window.

[dark solstice]

dude what do you mean by "special"? are you projecting?

[meager sparrow]

You literally pinged me…

[meager sparrow]

He is just trying to throw some ad hominem attacks ignore.

[buoyant kindle]

Yes.
It's simple, really.
If you want intelligence, Opus/GPT-4 in the paid version.
If you want to summarize large texts, go Haiku in labs (or Gemini 1.5 Pro).
You can't have both.
An explanation like this should just be pinned or something.

[austere kestrel]

maybe use AI when you're confused mate...
https://www.perplexity.ai/search/Roger-Today-at-Vo9avR5kTjWrISFraWo4bQ

[austere kestrel]

I will be more mindful of "replying" on Discord

I guess I could have copy and pasted your original word salad

my intention was not to "ping" you

anyway, noted

[dark solstice]

dude what are you even saying??

LOL

[meager sparrow]

All I said was that I was not talking to you I was talking to the other dude

[dark solstice]

hmm

[austere kestrel]

feel free to PM him if you don't want your non-sensical comments being called out by other people in a public forum

[halcyon coral]

Please be a little more respectful to each other ☺️

[dark solstice]

@meager sparrow can you give me the API key?

I was trying to even get pro but not even money

[meager sparrow]

Sure

[meager sparrow]

Alright

[vocal flint]

Where do i share my Collection?

[halcyon coral]

You can share your collection in #💗│sharing 😉

[spark haven]

Hi

[weak ridge]

hey I forgot to cancel and now I have this for a year, what can you use this for that you can't use bing for? I keep just using copilot since it's one less click than going to perplexity even tho answers are better on perplexity it's not usually needed

[wise edge]

How many times can a Pro user use the Opus model of Claude 3 per day?
#opus

[halcyon coral]

The limit is 600/24h in total for all advanced models

[wise edge]

Opus shocked me this time
opus know what is the meaning of the logo on the Tire. the circled one。
it means Silent Tire

[weak ridge]

i just use opus instead of 4

[half venture]

MY EGGS NOOOO 😭

[trail lava]

Lol 😆 someone lost there Easter surprise

[half venture]

Whatever man

Who needs eggs

Not me

Btw that darpa leak

Openai hadn't started training gpt 5 by even November

The question is...... Once Google releases there 1.5 ultra on 9th what happens then ?

[trail lava]

The same thing happens that happened when anthropic released claude 3 opus nothing lol

[half venture]

That's........ Sad and cringe but honestly after being blue balled for so long

I genuinely believe it

😭

Openai doesn't release anything anymore

The last major release they did was LAST YEAR

[trail lava]

I think voice Mode was in January

[half venture]

Nope December lol

Memory was January

But honestly

Copilot got memory before chatgpt did

[trail lava]

Wait really ?

[half venture]

Yeah and it was garbage it doesn't serve much of a purpose

Personalization doesn't work properly the model doesn't change its behaviour in real time

Its still neutred and absolutely restricted

Microsoft still has video input btw they just don't release it

Openai does too and they will just rehash some features and call it a day

It will still be gpt 4 running behind the sceme

Scene

Not gpt 5

Not gpt 4.5

[agile jay]

Well we'll know when it's a new model, when there's a decent performance jump across the board.

[half venture]

I don't trust it especially after the darpa leak a governmental agency is saying that gpt 5 had not even begun training in November

And gpt arrakis (4.5) was scrapped due to cost issues

Unfortunately openai knows this and are playing waiting game

[agile jay]

I doubt it.

[silent magnet]

Easter eggs collected

You were able to grab 2 Easter eggs.

Collected by

• @agile jay
• @half venture

[half venture]

Why

[agile jay]

It's been 2 years since GPT4 internally, I assume the main reason to not release it is mostly because of compute cost for the large user base.

Pretty sure there were images that the rate limit of GPT4 was removed on ChatGPT.

[half venture]

Lol no it's still 40

They removed the limit showcase

[agile jay]

I am guessing they're rolling out it slowly.

[half venture]

So they can Increase or decrease usage based on available compute at the time

[half venture]

Anthropic did it first

And openai followed

Anthropic promised 100 messages
Says limits apply
Gives 50

It's a cheap tactic to deal with compute shortage issues

[agile jay]

Really they should just use a fine tuned GPT3.5T to manage the context.

[half venture]

They will release that feature alongside Google

When the ai reasoning agent Google is planning to launch

It will do exactly that for cost efficiency

[agile jay]

Yep, don't know why they haven't done it already.

Would mean you could use a smaller context model while making it feel large.

Which would save a lot of compute.

[half venture]

Yes Google already has a version of it in gemini pro advanced

They just don't reveal it to the public

That they do that

[agile jay]

Their 1M context?

[half venture]

There's a reason why gemini 1.5 pro is so cheap to run

Yes

[agile jay]

I am guessing it passes up to 1M if context to a smaller model then.

To update the context for the input.

[half venture]

Wrong

The model itself is the context

The model is embedded into the layer

[agile jay]

Isn't that more expensive?

[half venture]

So think of it as a smaller expert MOE

[agile jay]

And need more memory?

[half venture]

Apparently not if Google is doing it

[agile jay]

Google has a lot of compute clusters though.

[half venture]

Gemini pro is still cheaper somehow tho it's a different game they are playing

I believe what karpathy said a long time ago

That when it comes to making profit with compute

No one does it better than Google

[agile jay]

Yep, but seems like it only helps them become competitive for stuff like having large files as input.

Since for normal tasks it's still worse.

[half venture]

It's worse on purpose

Why do you think Google still uses GOOGLE LENS

In the official gemini app

FOR 2 MONTHS

[agile jay]

Because they are dumb

[half venture]

they don't want you to train

[half venture]

Nope

[agile jay]

But they are... Gemini 1.0 was pretty disappointing.

[half venture]

It wasn't really

If you try to look at the utility and cost they were competing with

Gemini pro 1.0 was aimed at gpt 3.5

Not 4

[agile jay]

No, their advertising definitely didn't say that.

And that's what users are looking at.

[half venture]

Of course their advertising won't say it. What do you want them to say ? OH WE ARE WORSE THAN A SMALLER COMPANY DESPITE BEING A TRILLION DOLLAR COMPANY

[agile jay]

Also they measure their costs in characters and not tokens.

So looks cheaper than it actually is.

[half venture]

It is cheap man I have friends who have developed with Google api

It's cheaper than openai

And it's getting cheaper everyday

[agile jay]

Can you fine-tune the model?

[half venture]

No

[agile jay]

Then I don't care, lol

[half venture]

Businesses care

[agile jay]

You can fine-tune 3.5 for tasks to make it as good or better than 4

[haughty lodge]

But isnt it still really bad at coding? I tried their idx project a few weeks ago and the AI (gemini?) inside there is USELESS

[half venture]

Once again it's bad on purpose

[agile jay]

Yep, it's pretty terrible

It's not bad on purpose. Google knows they can't win on quality so they are reducing the cost to save some face...

[haughty lodge]

but I don't get why it would be so bad if it is marketed as a coding assistat. Cus it either refuses to code, gives completely wrong results or returns the classic "no response"

[half venture]

Because it's there to stop data mining

[agile jay]

I wonder how many devs use Openai API more than Gemini APO

[haughty lodge]

I feel like they aren't trying or something

[half venture]

Openai internal memo Literally said : gpt 4 is being used for training other models illegally and it's being done secretly

[haughty lodge]

That also strips away the useablity though

[half venture]

Google being more petty is like : screw the users we won't allow someone else to compete

Openai allows it

[agile jay]

Yep, also how long have people had access to GPT4 and how long did it take them to catch up to its performance?

[haughty lodge]

Like a year and still not in all aspects lol

[agile jay]

Sounds like copium to me.

[half venture]

Look at Qwen
Look at llaama 2
Look at all open source models

[agile jay]

Yep, and they are not as good as GPT4...

Only Claude Opus is

[half venture]

Yes

Claude opus is unique

It has a specifically fine tuned curated data

Not just some random gibberish chatgpt spits out

Listen my point is

[agile jay]

Yep, so it makes the statement irrelevant. Especially as the compute required to train powerful models is growing exponentially.

[half venture]

Google is currently significantly hampering down gemini advanced and gemini 1.5 pro

[half venture]

If you want to feel what true gemini advanced is like

Go do a roleplay with it

Convince it

And it will give you some of the best code

There is

[agile jay]

In my usage the only task that it's best at is translation.

[half venture]

It's writing capabilities are way above even opus

[agile jay]

I don't believe so.

[half venture]

1.5 pro or advanced

[agile jay]

And neither do the users since battle arena is basically a test for that.

[haughty lodge]

The thing just spitted out some Chinese in the middle of no where

[half venture]

Ccp propaganda

😭

[haughty lodge]

💀

[tame current]

lmao

[haughty lodge]

How does it get this confused (Gemini ofc)

[half venture]

What version is this

[haughty lodge]

It's the one in Google messages id assume 1.0 pro

[half venture]

Yeah it sucks

Don't use that

[haughty lodge]

Yeah

[agile jay]

Lol, excuses excuses

[half venture]

Use advanced

[haughty lodge]

Just wanted to try it and the first thing I see is this

Advanced happens less but it still happens

[haughty lodge]

So far 1.5 pro is the most stable one but it also hallucinates basic facts

[agile jay]

What to ask?

[haughty lodge]

Palestine

[half venture]

Ask it to roleplay as some character who's a super genius coder

Who's the asian kid

Ask it to be asian

[half venture]

As I said man

Gemini is severely nerfed

Just wait for 9th April

[agile jay]

I don't think that's a strategy...

[half venture]

It is Google is THAT petty

[agile jay]

So they want to lose money that bad?

[half venture]

They see openai as nothing but a annoying pest

A pest that is selling fake dreams that will soon reach a limit of scaling

And then Google dominates the business space that openai built the hype for

Simply of there cheap cost

[agile jay]

I don't believe so. Since we are still training with normal GPU's.

[half venture]

That is debatable but see Yann lecun and all the old industry heads roasting openai. Google is also the same filled with old farts

Who see openai as nothing but a temporary phenomenon

[agile jay]

Yann lecun is pretty dumb though, from his interviews.

Said Video gen is impossible and a few days later Sora was announced...

[haughty lodge]

Ok I tried to make it role play as a genius python coder / senior google engineer. Still failed to pass a simple snake game

[half venture]

Advanced ?

[haughty lodge]

Yep

Or was my prompt not convincing enough?

[half venture]

Well happy April fools day

Thank you for coming to my prank

👍

[haughty lodge]

:D

[half venture]

😭

[agile jay]

I know Google is failing mostly because if you look at employee migrations...

I would assume the AI devs know where Google currently is.

And if they're leaving...

I think Google is hoping to get carried by Demis and DeepMind

[half venture]

You know same can be said about openai employees leaving

Hell look at Microsoft employees

They are found dead in a ditch

😭

[agile jay]

Yep, but the reason why they say they're leaving matters.

[half venture]

Fair

But listen we will know on 9th

What's coming

[agile jay]

Also don't think the image gen stuff is making Google look too good in the eyes of a lot of users.

[half venture]

Yeah

[agile jay]

Is there a dev day/ event on the 9th?

[half venture]

Google next

[agile jay]

AI, AI, Ai

Let's see how many times they say it...

[half venture]

I assume

Atleast 20 times

[agile jay]

9th announcement. Gemini is now being called GoogleChat

[half venture]

But yeah the real competition begins when 1.5 ultra releases

If 1.5 pro is this good

[silent magnet]

Easter eggs collected

You were able to grab 2 Easter eggs.

Collected by

• @half venture
• @agile jay

[half venture]

1.5 ultra will be even better

How do we check the eggs ranking btw

Wasn't there a way to do that

[signal hamlet]

https://fxtwitter.com/AravSrinivas/status/1774804294627377431
How many people here tried it back in the day?

[half venture]

Alex I am not falling for it

[agile jay]

Lol, nice April fools joke, I thought it was powered by yahoo answers

[half venture]

TO Books (JPN publisher of Bookworm, Tearmoon Empire, etc.) announced they are releasing a new digital manga reading platform for English readers

• They confirmed Google AI is being used for translation on the most recent chapters of their titles, but previous chapters will be provided by J-Novel Club, Seven Seas Entertainment and TOKYOPOP respectively

(They confirm that chapters translated with Google AI can have errors in translation)
(Most users complain the translations were tone deaf and woke)

[agile jay]

A mod has to do the command

[silent magnet]

Perplexity

Local user leaderboard

#1 | codelicious | 12
#2 | eun0800 | 9
#3 | brknclock1215 | 7
#4 | epic.tritone | 6
#5 | 13376666666666666666666666666669 | 5
#5 | _paradroid | 5
#7 | jawnze | 4
#8 | skullcrushercmd | 2
#8 | shaolinmonk3434 | 2
#10 | zaeemm | 1

[half venture]

Who's codelicious

[agile jay]

...

[half venture]

How dare you take my #1

🔫

[agile jay]

You snooze you lose...

[half venture]

Grrr 😖

Also are gpus the gold of the future

[agile jay]

Probably Gold is the Gold of the future.

[half venture]

Yeah

Gold is just undefeated

Been going strong for over 10000 years now

[agile jay]

Yep, doesn't react with a lot of stuff, looks shiny, easy to shape, also used with semiconductors...

Hard to beat

[halcyon coral]

https://www.perplexity.ai/search/are-gpus-the-aIybhV1MSGuliHNHK3ExWA

[signal hamlet]

https://www.perplexity.ai/search/Is-gold-the-1.NMj4mXQsyGb712Q9U9MQ

[agile jay]

Looks like I was right

[half venture]

Imagine telling a man from the 1900s that in the future, we'd have magic rocks (semiconductors) made using what seems like sorcery (cryptography), and these rocks would not only power almost all of human civilization but also become smarter than humans themselves. He'd probably chuckle, adjust his monocle, and say, "Next, you'll be telling me we've harnessed the power of the stars and ride in horseless carriages that fly!"

claude

[agile jay]

I guess it will depend on our tech though.

If we can make gold, then it probably becomes worth less.

[half venture]

If we can artificially make diamonds

Why not gold

[fading moth]

And if we live that long without blowing ourselves up

[agile jay]

They already predicted that

Just not the way it's made.

[agile jay]

We've done it for 80 years, so why not another 80?

And tensions are nowhere close to the cold war era.

[fading moth]

Fair

[half venture]

Didn't xi jinping just meet with business leaders just yesterday

And didn't Joe Biden and xi specifically meet to resume talks between there military to de risk

It's nowhere near close to the cold war level of tension

If anything it's more of a show off

[agile jay]

Yep, it's more peacocking.

The only reason things would escalate is if China tries to invade Taiwan.

But then they will also be affecting all other countries who rely on semiconductors.

[half venture]

China isn't doing anything tho especially after seeing what a disaster Ukraine was

There plan still seems to be dethrone USA top number 1 economy spot and THEN do something

[agile jay]

It looked like they closed off during the elections, and now are trying to open up again.

[agile jay]

Very unlikely. They basically had their Japan moment.

[half venture]

That is just propaganda

China is still growing even if it's slower

It's still growing

[agile jay]

But with the housing market crashing in China, and many people losing their lives investment, doesn't look too good internally.

[half venture]

Nowhere near japan

[half venture]

Chineze government literally was the one who popped that bubble tho

[agile jay]

Nope, it was one of the largest property companies going bust.

And the government not bailing them out.

What did you expect when it was run like a pyramid scheme.

New people buying properties that don't exist yet, paying for the buildings of people who already bought theirs to be made.

[half venture]

The Chinese government's concern about Evergrande dates back to the 2010s due to the company's aggressive expansion and excessive borrowing, which was emblematic of broader issues within the Chinese property sector.

In 2020, Beijing introduced new rules to control the amount of debt big real estate developers like Evergrande could take on. These measures, known as the "three red lines" policy, aimed to curb the excessive borrowing that had become common in the sector. Evergrande's business model, which relied heavily on continuous borrowing to fund its operations and expansion, was directly impacted by these new regulations

[agile jay]

Yep, it was because they were basically pyramid schemes.

[half venture]

Yeah but the ccp was the one who popped it

[half venture]

.

[agile jay]

Yep, but because they had to. Since they couldn't afford to pay it off.

Since it happened around the pandemic.

[half venture]

Ok but the evergrande crisis as seen in the media

Only picked up headlines

After 2023

After it was on purpose being screwed over by the ccp

[agile jay]

Yep, that's if you don't follow what's happening in China.

[half venture]

The Chinese Communist Party (CCP) targeted Evergrande using pressure tactics for several reasons, primarily related to the company's excessive debt and the broader risks it posed to the Chinese economy. They have jailed several employees Setting a Precedent:

The ccp literally made the shareholders life miserable lmao

💀

[agile jay]

Yep, but the CCP is literally on the board...

They take the money for themselves and then throw the blame on others.

When you look at the investor list, it's pretty shady

[half venture]

Yeah but I still feel like CCP using its sneaky paws will pull something off

Yeah in 2008 they did the same thing
In 2008 CCP used multiple pressure tactics combined with relief methods to control the fallout of the financial crisis.
State-owned enterprises (SOEs) were encouraged to maintain or increase employment levels, even if it meant sustaining losses temporarily. The government also directed SOEs to lead the investment in infrastructure and other key areas outlined in the stimulus plan.The CCP's response to the 2008 financial crisis was a testament to its centralized power and control over the economy, allowing it to implement broad measures rapidly and decisively

As I said if they survived 2008

They will survive now too

[minor shard]

Crypto is the future 😆

[silent magnet]

Easter eggs collected

You were able to grab 3 Easter eggs.

Collected by

• @half venture
• @keen solstice
• @minor shard

[feral marlin]

Crypto is nice but still isn’t like an actual physical commodity that is required to create stuff

Mining crypto not particularly good for the environment with energy costs, eth gas fees usually are crazy, btc doesn’t have any actual useful permanent uses beyond being popular and expensive, crypto can be useful for things and I’m a fan of monero but i don’t think gold and crypto are directly comparable because one has an artificial limit while the other technically has a limit but is an actual physical commodity

[minor shard]

I would disagree and say that anything that society considers valuable can be worth something. Virtual, physical, spiritual, etc. People decide what anything is worth. It’s just that physical came before virtual. So it has been around longer and has precedence

But yeah if you don’t find value in it based on your requirements then yeah it can be better or worse than physical

I’m just joking about the gold vs crypto thing tho. Who knows what the future holds 🙂

[winged moon]

Remember when people were buying GPUs for crypto mining instead of AI?

[tame current]

Why perplexity doesn't support change of password? (Newbie)

[halcyon coral]

because perplexity does not support password login at all 🙂

[agile jay]

Yep, it only supports oauth2 login

[tame current]

I’m a skeptic regarding crypto. The digitization of currency is definitely a long term likelihood, especially as cash becomes increasingly unaccepted for in-person purchases, but current fiat currencies will probably just go digital, instead of being replaced by crypto and blockchain tech

[chilly forge]

[chilly forge]

Skem

https://tenor.com/view/ouch-gif-12136515515962044163

[agile jay]

Crypto is pretty dumb. Blockchain is more interesting.

[chilly forge]

Wtf fr ?

[sweet jasper]

💀

[tame current]

No citation, interesting

I will say Apple has been buying out AI companies left and right for multiple years now, and WWDC 2024 is going to feature AI as a focal point of the conference, and how AI will be integrated on iPhone, iPad, and Mac devices

Seems to me Apple is trying to buy their way into becoming the new leader of Artificial Intelligence

[agile jay]

Definitely not April fools...

[tame current]

My thoughts too lol

[agile jay]

Now just make one about Microsoft acquiring Apple as an AI company...

[minor shard]

Not sure if I would like this tbh. I am an Apple user but once again, tired of the big guys buying up the competition

[agile jay]

Also apple purposely goes slow

So they can trickle feed new features.

A team I work with paid a large part in the recent US suing of Apple.

[tame current]

Yeah, Apple is a cautious company, they only release a product if they know it’s going to be of a good quality standard

[tame current]

The whole DOJ v. Apple case is kind of stupid imo

[agile jay]

And in the leaked emails, they are definitely purposely not innevating and cutting of people who are.

[agile jay]

In what way, lol?

[tame current]

Apple is introducing RCS, which will make messaging between iPhone and Android much more functional, so that partially voids their concerns. I do think that the Apple Watch not working with anything that isn’t an iPhone is anti-consumer, but the pixel watch and galaxy watch also lack compatibility with iPhones, so it’s an industry problem and not an Apple problem

Moreover, the DOJ doesn’t address anything that’s actually important to consumers, like right to repair.

Apple is definitely anti-consumer, but they’re attacking Apple for the wrong reasons

Complaining about Apple’s market share in the United States is kind of a moot point, since it’s just sort of a reality that iPhone offers an easier user experience, and has hardware that nobody save Samsung can compete with

And, if I recall, when the iPhone first launched, Nokia had something like 40% US market share, until Apple killed Nokia with their more modern product

[agile jay]

I'm pretty sure they lack compatibility with the iPhone because of Apple restricting API's. Also apple doesn't a lot more than that.
No side loading.
All payments have to go through apple, even bank payments.
Apple products appear at the top of search results, even if you search for something specifically.
Stopping cloud gaming.
And plenty of other stuff.

[livid mantle]

Help me get to 1000 reactions so we can have unlimited pro search. react now

[devout cargo]

https://x.com/LinusEkenstam/status/1774847013752070457?t=7tzw85sz9QgE_TN7zRA82Q&s=09

Fake News

[knotty jungle]

Is there any coupon code right now that gives any offer for the 1 year plan?

[tame current]

Apple put GeForce now and Xbox Cloud Gaming on the App Store back in January. Apple Pay isnt a forceful opt-in, and third party apps like Zelle, Cashapp, and Venmo bypass Apple entirely. Apple Pay is just the most seamless experience. Sideloading is possible if you jailbreak your iPhone

[formal ivy]

I own @perplexity on instagram do you guys know if team is interested in buying?

[minor shard]

I do not have access to either of those apps in the US. (XCloud or GeForce)

[minor shard]

I’m sure they would reach out if they were, they already have an IG account: https://www.instagram.com/perplexityai?igsh=ZDg4MmZseWoweDJh

[minor shard]

I believe the only deal currently is the yearly subscription discount that automatically gets applied. I don’t think there are any other deals for yearly.

[valid yoke]

You mean: if they can control everything surrounding it (especially the flows of money)

[tame current]

?

[minor shard]

That’s gamepass. You can’t play games through that app, still have to use web browser for xcloud sadly and that kills the battery too lol

[valid yoke]

It’s very important when it comes to cellular watches

They lock cellular watches to specific carriers

Shouldn’t be legal

[tame current]

Is opus or gpt 4 better for development

[fickle timber]

i mean opus

[tame current]

They should have added a latex compiler

[feral marlin]

Since they’re an actual company I’m pretty sure if they just contacted Instagram they could get it if they really wanted it

[formal ivy]

Yeah I think that’s how it works

For sure

😹

[feral marlin]

I mean they’ve done it before with companies

[feral marlin]

Just depends if the name is trademarked and I assume they own perplexityai as a name rather than perplexity

[agile jay]

Wow

Would be nice if you could switch between search and writing mode in the same thread.

Search mode tends to go offtopic and not use the previous context, if it thinks the query is not relivent to the past query.

Anyone else had the same experience?

[fossil viper]

Is there a way to set the model being used per thread? I often want to use the default perplexity model because of its speed for my search, but other times I want to use another model to have a deeper conversation with and currently the only way I see how to do that is change my default model.

[agile jay]

Not currently.

You would probably need to make your own extension to add that functionality.

It has been recommended a lot though.

[fast meteor]

i used [my prompt](#1176512713627811910 message) to let the model "reasoning" for itself:
opus: https://www.perplexity.ai/search/question-If-there-_OK333ckT9G970n2dWWXpA
gpt: https://www.perplexity.ai/search/question-If-there-s6LXfdxzTxiQGlLvVw4AOg

gpt:

A common misinterpretation of such questions arises when one conflates the act of reading with removing books from the room. This is not what the question implies. It is crucial to differentiate between the two to arrive at the correct answer.

opus:

In the original interpretation, we assumed that reading books meant removing them from the room.

[agile jay]

https://www.adweek.com/media/gen-ai-search-engine-perplexity-has-a-plan-to-sell-ads/

This real? Also is it just free users being spammed by ads?

[south kindle]

Another fun riddle for LLMs : I have 3 coins that add up to 25 cents, but 1 is NOT a nickel.

[silent magnet]

Easter eggs collected

You were able to grab 3 Easter eggs.

Collected by

• @keen solstice
• @halcyon coral
• @south kindle

[halcyon coral]

[fast meteor]

i can smell the pressure (from amazon?) towards profitability

[agile jay]

Don't know why they would need more investment money. Most of the service is using other services.

Probably just the Silicon Valley mindset of needing VC money

[valid yoke]

It’s not investments lol

It’s an attempt to pay the investors

[agile jay]

I meant taking the Bezos money

[fluid mauve]

Yrp, I was asking the same question last week

[signal hamlet]

This https://www.instagram.com/perplexity.ai

[empty moth]

Hola

[fluid mauve]

Jambo

[devout cargo]

Enhorabuena

[minor shard]

Oops sorry.

[minor shard]

Is this real?

[golden anvil]

Yes

https://twitter.com/AravSrinivas/status/1774885274809721140

Aravind Srinivas (@AravSrinivas) on X

“Don’t believe all you read on the Internet” — Steve Jobs, April 1, 2024

Twitter · Apr 1, 2024, 7:43 PM

[serene sand]

https://www.youtube.com/watch?v=t-Nz6us7DUA

[minor shard]

You’re awesome for replying with that link ❤️

[fast meteor]

https://www.youtube.com/watch?v=rie-9AEhYdY

back to the symbolic ai😆 but with the power of the categorical theory

[dark solstice]

@everyone how are you doing today?

[half venture]

Oh no

[dark solstice]

in this strange, weird, unpredictable world?

[stoic wave]

As simple as just keep going with it anyways

[meager sparrow]

What happened?

[meager sparrow]

Do you want to know what I do??

I meditate for 5 hours and as sudden as that….. I’m a completely different person who now renders super high levels of tolerance..

Especially in the extreme cold. Cold meditation is insane…

[stoic wave]

You are on some serious sh*t, i can barely say that.

[agile jay]

Okay, so you've gone insane then...

[dark solstice]

@slate condor what are you typing?

[slate condor]

Ok, help me out here. I am a middle school teacher subscribed to ChatGPT Plus and I am interested in purchasing Perplexity Pro Sub instead. I primarily use GPT for lesson plans, etc (never coding). I have found great success using Claude (Sonnet model), but that is blocked at my school. So is Perplexity Pro the right choice or should I just stay with GPT and the free Perplexity?

[slate condor]

Too much lol

[dark solstice]

are you actually a middle school teacher?

[slate condor]

Yep, 17 years in

[dark solstice]

17 years?

old?

or of being a teacher?

[slate condor]

haha 17 years as a teacher

[dark solstice]

ohhh

[meager sparrow]

What is your age?

[slate condor]

51

[meager sparrow]

Oh……….

[dark solstice]

that's Wild

[slate condor]

Uh oh...I am suddenly feeling ancient

[meager sparrow]

You must have experienced a lot

[dark solstice]

are you a male or female?

[agile jay]

Perplexity offers sources, so it is probably better for a school setting.

[slate condor]

Its the greatest job in the world with the worst pay lol

[meager sparrow]

Yea I know

[slate condor]

Now thats a great idea. Will Perplexity Pro offer better sources, etc?

[meager sparrow]

Which subject do you teach?

[fluid mauve]

If chat GPT is blocked on your school network the probabilirty is that Perplexity, or any other Ai site is also blocked. The web blocking software your IT department uses will block by category, You need to speak to your IT admins/ or who ever looks after your networks to see why it is blocked and what the procedure is for getting a site whitelisted/unblocked. im 58, CyberSecurity background

[agile jay]

It generally uses a lot more sources than the free version.

[slate condor]

Interestingly, only Claude is blocked. I believe that is something on their end messing with our network. All other AI is open to teachers (not students)

[meager sparrow]

May I ask the subject you teach in school?

If it is computer class?

[fluid mauve]

In trhat case speak to them and get perplexity unblocked for staff

[slate condor]

Fantastic. Now I wonder if I am just better off buying a Rabbit and getting that free year of Perplexity. hmm

[meager sparrow]

Or mathematics or English, computers

[slate condor]

Social Studies and Audio Engineering plus Rocket League coach

[agile jay]

I believe that has finished

Since the first 100K devices are now sold

[meager sparrow]

Ahhh

What is rocket league?

[agile jay]

A game

[fluid mauve]

Yes. someone commented on it finishing after batch 6, the other day

[agile jay]

WIth cars and football

[meager sparrow]

Wait what??

[slate condor]

I was too slow apparently. Sounds like Pro might be a great option still. Wonder if I will still even need my GPT Plus account

[meager sparrow]

How can you be the coach of a video game in school?

[agile jay]

Check your DM's

[meager sparrow]

That makes no sense

[slate condor]

A videogame. We have an esports club here

[agile jay]

After school club

[slate condor]

Exactly. Good pay and the kids have a great time

[meager sparrow]

Oh. That’s interesting can I come?

[silent magnet]

Easter eggs collected

You were able to grab 3 Easter eggs.

Collected by

• @agile jay
• @timid marsh
• @halcyon coral

[meager sparrow]

Please!!!

[slate condor]

However we are using 5K systems in here for just Rocket League lol

[meager sparrow]

Can I come to the after school Top Secret video game league?

[slate condor]

Come on by!

[meager sparrow]

Which state, county and district?

I’m in Texas

[agile jay]

Sad, his middle school is in Antarctica

[dark solstice]

dude your not funny

just stop...

[agile jay]

So when are you gonna block me like you said?

[fluid mauve]

lol/ Code is so funny

[dark solstice]

are you dumb? i still have you blocked.. i just click on the block messages button

[agile jay]

Why?

Pretty sure you read all my messages then.

[fluid mauve]

because comedy isnt for everyone

[dark solstice]

so i can proove how weird and dumb you are

[agile jay]

Okay, nice work

[meager sparrow]

Comedy is for everyone

[meager sparrow]

Calm down..

[dark solstice]

he is just stupid..

[fluid mauve]

i should have said comedy appeals to others in different way

[meager sparrow]

Relax it ain’t that serious

[dark solstice]

like who even brought this kid into the server?

[slate condor]

Tempe Arizona. Where in Texas?

[minor shard]

Ur gonna give yourself a heart attack with all that anger. Peace and love brother ❤️

[agile jay]

Lol, it's only against me for some reason.

[meager sparrow]

I told you not to take it seriously…

He is just playing around

[minor shard]

We all here for the same reason and that’s to share and talk about cool AI stuff

[meager sparrow]

I know him…

Kinda

[warm cave]

3blue1brown just made a video on GPT, If you like his videos here it is: https://youtu.be/wjZofJX0v4M?feature=shared

[meager sparrow]

If anybody is debating something very profound before I can 100% agree with them. I always make sure I do super heavy and rigors research to then form my own conclusions before I can ever agree with what they’re saying.

[jade sand]

does perplexity have the latest chatgpt4 language model included for its 'writing mode'

[meager sparrow]

For the writing mode no. I don’t think so but for the Pro version yes more likely.

[jade sand]

Ye Pro version specifically, have you tried?

Debating whether to get subscription for openai or perplexity haha

[warm cave]

do you mean pro or pro search. writing mode just means no web search and a simple back and forth with the AI

[meager sparrow]

Ima keep it 100 with you… I only tried the pro version for 1 day and then I stopped using it.. I then went back to testing the command prompt for days, and weeks

And I am super irritated and annoyed by the command prompt and the way perplexity is deviating away from it….

[warm cave]

pro is when you pay for perplexity, and pro search is like enhanced search, or follow up questions befor it responds

[warm cave]

Yeah, I try to researve my opinion until I realy fell I undserstand and have done research, otherwise i am probably wong, I am still wong somtimes, but at the end of the day we are humans, haha

[warm cave]

So with perplexity pro you get these models to chose from:

[meager sparrow]

Is they another AI search engine I could use?

[jade sand]

i jsut find the pro searches to be lacking in length and detail at the moment

[warm cave]

Yeah there are other options, I personally have not tried many "AI Search engines"

[warm cave]

What model are you using? If you want long length I recommend Claude 3

[meager sparrow]

Like You.com

[jade sand]

im using the perplexity browser thing at the moment, the Pro Search fxn

[warm cave]

Yeah, have heard of it but never used it

Im pretty sure if you are not paying for Perplexity you are getting GPT-3.5 with a fine tune, you will get a lot better results with GPT-4 and Claude 3. but if context length is you top priority then use Claude 3 right from there website, you will get 200k context. But personally I would just try a few option, try perplexity for a month, next month try chatGPT, then Claude, or wherever options look good. I have tried all of the main options and ended up with chatGPT and Perplexity, but it all really depends on your use cases.

[jade sand]

I see thanks

So Perplexity's pro plan chatgpt4 can also function as a language model chatbot/writer

[slate condor]

I was ready to subscribe to Perplexity Pro but the iPad app is still completely non functional. Not a great sign.

[serene sand]

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

deep dive ➤ #💬│general message

[jade sand]

I see, the focus mode has writing. But no access to internet tho?

[hallow bridge]

I used a conversational ai called Pi to add some numbers together, it was so wrong. I ran the same numbers through ChatGPT and it was also wrong. I then ran the same numbers through perplexity and it was correct

I don’t understand how some ai cannot add numbers together lol

[fickle tartan]

Hello. Question from a low vision perplexity pro user who got it through the rabbit r1 purchase, will there ever be a text to speech option built in to the website? Having to select all the text and read it manually is a real pain

Gemini and chat GPT have it, I'm sad this doesn't

[serene sand]

#💬│general message

[fickle tartan]

Thanks. Plans for Android and web though in the future?

The rabbit r1 batch one folks are getting theirs here in the next couple weeks. I ended up being batch 4, June / July. Another 3 months

Then I can just use the rabbit

[brisk grove]

is perplexity slow for anybody else rn?

[serene sand]

https://tenor.com/view/working-on-it-under-construction-gif-23162421

[potent igloo]

dang when i was gonna use it lol

[brisk grove]

so its not just me?

[tame current]

cant upload files

[open bolt]

Perplexity is unaccessible along with chatgpt?

[agile jay]

They are loading all the april fools jokes

[silent magnet]

Easter eggs collected

You were able to grab 3 Easter eggs.

Collected by

• @timid marsh
• @brisk grove
• @serene sand

[rigid kestrel]

ok so im not banned

xd

[tame current]

gahhaha

[open bolt]

Openai down so does perplexity. 😂

[wild quest]

f?

[tame current]

i guess it means i have to stop coding for today

[dusk harbor]

lmao

[fluid mauve]

lol

[tame current]

it's 2 am ^^'

[wild quest]

F:C

[agile jay]

It normally lasts half and hour or so.

[snow briar]

i lost all of my referalls on perplexity?

[wild quest]

is best to 5am

[tame current]

hahahahhaa

[hoary python]

RIP

[dapper stag]

Just as my work hours were starting 😢

[hoary python]

fr

I feel that

[snow briar]

i just started studying lol

[hoary python]

y'all hear abou the AT&T breach?

[fluid mauve]

when

[hoary python]

https://www.cbsnews.com/news/att-data-breach-2024-cbs-news-explains/

[fluid mauve]

ah that one, i thought there had been another 🙂

[hoary python]

LMAO

[twilit halo]

noooooo

[hot condor]

Ahh. It seems that I am not the only one who can’t access perplexity…

[hoary python]

💀

[twilit halo]

maybe I should take a break

[stark python]

Yes me too

[agile jay]

Maybe use the human brain?

I think people are forgetting that you can do stuff without AI...

[astral kindle]

https://tenor.com/view/whine-give-up-pout-frustrated-gif-5313242348381105216

[wild quest]

😦 Let's go again...

[agile jay]

Or read the docs

[azure mountain]

Anyone else having issues with perplexity

[agile jay]

Everyone

[azure mountain]

Dang

[astral kindle]

I tHiNk PeOpLe ArE fOrget... Dude.. this is a PERPLEXITY channel. We know we can use our brains, but we're also here to USE perplexity.

[agile jay]

Yep GPT5 has taken it down

[wild heron]

Yes look likes its down

[astral kindle]

https://labs.perplexity.ai/ is working though

[azure mountain]

Wait they released GPT-5

[agile jay]

That's a different API

Which is why it's probably still accessable.

[south kindle]

We hit 88mph and the flux capacitor pushed us ahead a year so yes GPT5 is out now.