💬│general

[south kindle]

It is also April 1st still ha.

[azure mountain]

Listen guys I promise I’m not stupid

[astral kindle]

They're on it guyses

[weary crescent]

It won't let me upload, even though I haven't uploaded anything today?

[agile jay]

Yep, it's probably maintainance byt perplexity should really notify users 24hours before.

[agile jay]

Yep, it's down

[weary crescent]

ah

just saw the messages above lol

[round plover]

perplexity is lag

[wild heron]

Looks like I have to use my brain for now

[winged moon]

Damn, now I need to actually analyze my story myself and determine if it's even worth my time? I just want to not think hard and just throw ideas

[south kindle]

No disrespect meant sir. Lots of sarcasm gets thrown around here like a mine field. An unintelligent person wouldnt even be on discord or know what perplexity is to start with.

[azure mountain]

It’s all good man

[winged moon]

Bruh, you are so wrong about stupid people not being on discord. But you do seem right about perplexity

[agile jay]

Just buy more RAM

[winged moon]

Just download more ram

[azure mountain]

Freeram.xyz

[agile jay]

It tages ages to download 64GB though

Faster to run to the shops

[tame current]

Is perplexity not working for anyone else right now?

[agile jay]

Read above

[fervent crescent]

Looks like others are having issues too @rotund sedge

[south kindle]

That’s fair. Ha. It is a social service. I more so meant - on discord - in this server - asking questions.

[tame current]

What is labs

Is that different from the website?

[chilly forge]

#🔔│status should have something atleast

[south kindle]

Once we can 3D print chips at home that could become a real statement

Responded to the wrong comment again.

[agile jay]

A place to test some perplexity models.

[wild quest]

ok users of perplexity

xd

[south kindle]

Meant for testing models output. Similar to a chat experience.

[agile jay]

DDOS?

[desert saddle]

not working for me either!

[south kindle]

Googles revenge?

[agile jay]

Yep, Claude is still up

[south kindle]

“You wouldn’t download a car” Could become real someday as well instead of a funny meme from 2002.

[azure mountain]

I’ve always wounded with how fast PCIe and SSD’s are getting, if it’s possible to use them for large size really slow RAM

[agile jay]

They are really fast, but don't compare to RAM

[south kindle]

In a way, if they have a cache onboard they do in a sense.

Some SSDs/M2s have them. Edit. NAND Flash.

But we are just mixing up terms at this point.

Compute go brrr

[agile jay]

L3 cache on M.2 SSD's?

[rotund terrace]

anyone here using preplexity for coding?

if yes which model is the most suitable ?

[agile jay]

Sometimes

But AI isn't there yet where you can use it blindly

[agile jay]

Claude Opus probably

[azure mountain]

Actually I have personally found GPT-4 to be better

[south kindle]

NAND flash was the word I was looking for.

[rotund terrace]

agree .. domain knowledge is required in ai so far for almost anything.. this helps in productvity

[azure mountain]

Anyone here know about the band tally hall

[south kindle]

That’s domain and testing specific as well. People that know how to prompt Claude have seen the leaps over gpt4 turbo.

[rotund terrace]

i have tried chatgpt4, gemini advance , microsoft co pilot .. so far i found preplexity to be most helpful

[azure mountain]

That may be my issue

[agile jay]

Claude is just better at following instructions IMO

[fast meteor]

128mb l3 cache is the king

[rotund terrace]

for coding ?

[agile jay]

The king is CPU registers...

[azure mountain]

Yeah but I am starting to think that’s because it’s what I use most of the time

[rotund terrace]

yes .. sometimes this gets stuck in loop .... so i have to kind of restart .. as this can not forget old reference

[stable fossil]

I set default to Opus. I regenerate the answer with other models whenever needed

[chilly forge]

GPT4 is LAZY asf

[azure mountain]

That is true

[rotund terrace]

i use for pine script and found Claude to be most helpful so far

[tame current]

I heard turbo is worse than the regular one

[south kindle]

These are all sharing models and are different platforms not specifically AI model specific. Perplexity does come out on top though even with the default model PPLX model (typically).

[random orchid]

when is status update for perplexity dropping?

[agile jay]

Yep, the real problem is you are supposed to be able to already make it yourself, and the AI is supposed to just be a faster way of doing it.

[south kindle]

Provided they know it’s even down. Hopefully soon.

[azure mountain]

In all honesty I use it for C, C++, and Python.

[agile jay]

A mod will post here

[chilly forge]

Yeah man thats the reason some AI services have started to list back the Vanilla (GPT4-32K) model back

[agile jay]

How good is it at managing pointers in C/C++?

[tame current]

do you think perplexity will do the same?

[chilly forge]

Nope lol never

[rotund terrace]

i have installed Nvidia's chatwithrtx ... and playing with that .. i hope they can improve that for offline purposes

[azure mountain]

Kinda meh usually ok

[agile jay]

Really we just need more intelligent infrastructures

[peak scarab]

Is perplexity ai is still back or?

[agile jay]

Such as an intelligent context

[tame current]

I think bing ai offers gpt4 for free with copilot, and when you get their copilot pro membership, they change it to gpt4 turbo, so you’re basically paying for a downgrade lol

[south kindle]

Same. I’ve found GPT4All and Jan to be better experiences that accomplish the same thing. More purpose built LLM tools that are open source.

[azure mountain]

But I mostly use it for flipper zero apps and well as Raspberry Pico

[rotund terrace]

anyone else here tried chatwithrtx?

[chilly forge]

They also defaulted to Turbo (There is a toggle for Pro users)

[agile jay]

The simpler the syntax, the better the generation IMO

[tame current]

even for the free version?

[chilly forge]

Yes

[tame current]

😦

[rotund terrace]

i use Ai for some reviews and reporting . so i like preplexity's accuracy as this is important .. chatgpt make up things as well that doesnt help me in reporting and assessments

[tame current]

looks like im not using bing ai anymore lol

[silent magnet]

Easter eggs collected

You were able to grab 5 Easter eggs.

Collected by

• @timid marsh
• @agile jay
• @south kindle
• @whole basalt
• @azure mountain

[tame current]

so which one do you usually use?

[south kindle]

Yup. They were the first to actually release a GPT-4 product for the public. We were all messing with copilot for a week or more when they announced we had all been playing with GPT4 without realizing it. This was almost exactly a year ago now.

[agile jay]

Which is why Go is probably the best for AI gen. Python has quite a lot more keywords. JS is probably a nightmare since libraries get out of date all the time.

[tame current]

What are easter bunny eggs?

[agile jay]

Easter event

Top 3 get 1 month of free perplexity I believe.

[rotund terrace]

Preplexity with Opus .. this helps me the most in research .. However, Gemini advacne is really good with writing ... i kind of mix things .. research in preplexity and writing gemini

[chilly forge]

Too much efforts to get 1 month free lol

[agile jay]

Most likely

I think I'm first, at least I was the last time the leaderboard was shown.

[chilly forge]

If it would have been 1 Year I would be excited

[livid mantle]

I turned off Perplexity.ai

[agile jay]

Same, or at least have different ones for first, second and third.

[south kindle]

https://tenor.com/view/spongebob-gif-7921357

[slim agate]

Perplexity seems to be down, and this site says people are reporting an outage: https://downforeveryoneorjustme.com/perplexity

But Perplexity says all systems are operational at https://docs.perplexity.ai/page/application-status

Does anyone know what's going on?

[rotund terrace]

another thing i like in gemini is that you can export data into google sheets .. the formating is good .. i wish in preplexity we could do this ..

[agile jay]

6, 3, 1 months would probably be a good range.

[tame current]

Thank you, i used opus for everything and gpt4 turbo for writing but I’ll test out gemini too 🙂

[serene sand]

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

[agile jay]

You only need a good custom message for improving the writing of a model.

[agile jay]

What's the current leaderboard?

[tame current]

@grand mauve @frank rivet
do you guys know what’s happening with the perplexity outage?

[agile jay]

I would probably search for a good GPT and then copy it into a collection.

[agile jay]

Could just be because GPT API is down.

And it might crash when a service is down.

[rotund terrace]

can we get the service..those who dont use chatgtp 😄

[tame current]

What is going on

[agile jay]

The Devs like to act mysterious

And don't tell us anything

[brittle garnet]

What happend my general

[peak scarab]

Are you faking rn?

[glass spear]

same here

[agile jay]

Wonder if they're getting DDOSed, they are saying they ar egetting too many questions, but noone has access currently so...

[slate condor]

I’m guessing it’s me. I planned on buying a Perplexity Sub this afternoon only to find the iPad app broken and the site now down. They really don’t want my money lol

[kindred folio]

perplexity ain’t workin 😭

[agile jay]

You should have read your DM's, lol

[kindred folio]

told me to rebuy my sub + only option is gpt4 other than default

[peak scarab]

April fools?

[warm cave]

when its back online we will have GPT-5 and Gemini 2, but it will be April fools

[whole lark]

GPT is no good, neither is perplexity. Is it connected?

[agile jay]

I'll just go open some easter eggs as I wait.

[split sigil]

hey

someone who speaks Spanish?

[agile jay]

Hola

[chilly forge]

Amigo

[split sigil]

hola

[agile jay]

What's wrong my dude? I did spanish in highschool so I'll just translate it for everyone.

[fossil sentinel]

Lmao

[split sigil]

saben de algun humanizador bueno y que tenga semana de prueba?

[fossil sentinel]

I thought my wifi was just being weird

[agile jay]

GPT5 escaped the bunker, so now they are assembling the AI companies to take it down.

[wet pebble]

Is Perplexity down?

[peak scarab]

Yes

[wet pebble]

any updates from the team?

[agile jay]

nope

[chilly forge]

biggest downtime till now ?

40 mins and no update

[agile jay]

Possibly

[civic flume]

Dang, no looking good

[agile jay]

ChatGPT was randomly down too, around the same time

Guess their team recovered faster

[chilly forge]

Its better to have other options seriously

[chilly forge]

Yeah works now

[warm cave]

And i cant ask perplexity why perplexity is down

[agile jay]

Just Bing it

[rotund terrace]

others are not like preplexity .. this is the issue

[agile jay]

Send a heart emoji in chatgpt

[chilly forge]

Thats the reason we all are here xd

[agile jay]

Why is there only a thumbs down button now?

Wasn't there a thumbs up butotn too?

[chilly forge]

Yeah xd

[native harbor]

It's my first time seeing Perplexity down. The Pro feature was disabled before it went down.

[chilly forge]

Wonder why they removed it

[warm cave]

Hmm, weird

[agile jay]

Now you can only show hate

Lol, Youtube should do the same

[chilly forge]

My first time seeing Perplexity down for so long

There have been minor hiccups here and there

[agile jay]

Yep, and we are being ghosted too

Like being in a toxic relationship

[chilly forge]

https://tenor.com/view/the-bachelorette-you-see-what-im-saying-understand-get-it-you-feel-me-gif-16430678

[rotund terrace]

anyone used claude.ai

[agile jay]

Yep, it's still working

[warm cave]

Yeah

[rotund terrace]

what is difference claude.ai vs preplexity

[polar tulip]

I just paid for the pro and it's down 💀

[agile jay]

I do user research and manage the Perplexity Discord.
I am always open to conversation pertaining to feedback and suggestions.

Maybe we should ask?

[sleek harness]

Experiencing an outage here too on the IOS app

[native harbor]

perfect, its back to normal now

[polar tulip]

Nice !

[agile jay]

We are alive boys

Search all the random stuff you were gonna search before

[warm cave]

Longer context with Claude.ai, you get full 200k, not sure what you get with perplexity, i have heard a lot of different numbers from mods and staff. but with perplexity you get way more usage, I run out of Claude really fast but perplexity has a ton of usage. plus you get web search and other features, jsut depends on your use case

[polar tulip]

My library has been deleted :(

[meager sparrow]

LMAOOOOO

[polar tulip]

Oh nice it's back

[meager sparrow]

Guy is asking for dislikes

😂

[azure mountain]

Give the ai a firm “womp womp”

[meager sparrow]

“Hey everyone give me a thumbs down only” LOL

[azure mountain]

Boo this guy stinks

[rotund terrace]

only bing could compare this.. others didnt

[warm cave]

I mean its kinda inaccurate, but still cool

[rotund terrace]

i think outdated

[whole lark]

Good, it looks like they're back.

[warm cave]

Yeah, outdated is a beter word

[agile jay]

Compare what?

[south kindle]

https://www.perplexity.ai/search/Check-for-perplexity-_N.izc7ERESsDEHMIW4atA

That worked.

[meager sparrow]

Imagine if they were two dislike buttons under his comment the first one was a dislike and the second one was an Ultra Dislike..

[south kindle]

Also had perplexity search for how to report issues about itself ha.

[chilly forge]

What ? Perplexity is not available in Germany ?

I mean unless we use VPN

[rotund terrace]

this is outdated info

[chilly forge]

Lets Go we’re so back

[split sigil]

How long will this code be available?

[rotund terrace]

i got my eyes back

i was blind without preplexity

[split sigil]

CRNT005

[rotund terrace]

🤣

oh damn .. i lost my prompt

that is not in the history

nwo i have to do this again

[silent magnet]

Easter eggs collected

You were able to grab 4 Easter eggs.

Collected by

• @warm cave
• @timid marsh
• @serene sand
• @rotund terrace

[warm cave]

Yo i got it, haha

[rotund terrace]

what do we do with this egg

[chilly forge]

https://tenor.com/view/i-feel-your-pain-all-the-feels-keezia-leigh-comfort-i-feel-you-gif-16152618

[serene sand]

#📰│server-news message

[south kindle]

Protect it with your life.

[agile jay]

hatch it

[chilly forge]

Make an omlet out of it

[rotund terrace]

https://tenor.com/view/protect-attack-punch-me-fight-me-prepare-gif-17784080

[native harbor]

Get the LLM to create a comparison table for you

[tame current]

It’s back online now 🙂

[warm cave]

To be fair perplexity had outdated info too, but I mean its accurate according to the sources which are a bunch of news articles:https://www.perplexity.ai/search/make-a-table-0jmZ632ITqOFOGk2fYEDgQ#0

[agile jay]

Check the favicon of the site...

[native harbor]

I am using VPN, too, but I am in CN.

[agile jay]

What happened to the icon...

[split sigil]

.

[whole lark]

PERPLEXITY could also create a status page like OPENAI.

[rough flower]

When I select claude opus. The results are coming from sonnet. But pplx mentioned pro users has unlimited use. Android app shows 5 per day for opus. Can anyone explain?

[warm cave]

Correct me if im wrong but mobile app is still outdated

[rotund terrace]

am i talking to real people here or ai bots ? just want to make sure

😆

[warm cave]

WE are a buch of bots complaining about perplexity being down

are mother ship had no connection

[rough flower]

Ai is Artificial Intelligent
we're natural dumb ers

[warm cave]

I mean you never know

[random orchid]

is it back yet?

[warm cave]

yes

[rough flower]

Yes

I think pplx fooling everyone one April's fool day. But we're on 2nd April 😁

[native harbor]

https://www.perplexity.ai/search/Please-create-a-g7l1t_z3TlGwTjio67Ny9A

[warm cave]

uhhh...

[young token]

LOL

[south kindle]

Prank level : Nuclear?

[warm cave]

haha 🤣

[south kindle]

Giving the user the option is a terrible idea. Prank level : Chinese water torture.

[warm cave]

there we go

sonar is the only one that would answer, but its still kinda lame

[lethal bobcat]

That's no prank.

[warm cave]

where is the fun ai, haha

[warm cave]

Depends on your sense of humor

[lethal bobcat]

😬

[warm cave]

😏

So i guess Mistral large is more censored then sonar

[wet pebble]

is it just me or is Claude 3 Opus inside Perplexity better/smarter than in Claude interface? Using the writing mode ofc, but still getting slightly better results with the same prompts in Perplexity.

[lone citrus]

If you ask the LLM, it will tell you it's system prompt, I am sure the pplx have made some promt engineer for those api so that's why sometimes You feel it's better sometimes are not

I still prefer the oringal website instead using api, more ability to customized

(but too expensive, claude.com don;t even allow edit question or set a system prompt, and chatgpt4 is too slowwwwwww for me)

[weary crescent]

I've found gpt4 pretty fast for my uses

mostly recently, I guess server upgrades or something

[austere kestrel]

It's not really workaround (would be tedious to do this at the start of every conversation), but worth noting that you can always just send the "system prompt" as the first message on claude.ai, if you really need it for whatever you're working on

I've shared this before.. But imo it's helpful / important not to think of "system prompts" as anything particularly special..

whatever we think our system prompts / custom instructions / profiles are doing, they're just a bunch of tokens inserted somewhere towards the top of the context window. Every response is still just an autoregressive completion based on the sequence of tokens that are in the context window...

[agile jay]

You could make a GPT for writing the prompt for you.

[untold cliff]

Epstein Barr virus

[warm cave]

Don’t worry, the virus is self destructive

Oh wait, it’s a real thing

[austere kestrel]

aha I actually did that yesterday - feel free to give it a go 🙂 https://chat.openai.com/g/g-pBy4bGX5F-anthropic-prompt-creator-transforer

Haven't really used it yet (the Collab metaprompt thing has been working well - and it's useful to see the model's "thought" process via the scratchpad), but it'd definitely be faster and more straightforward

[lone citrus]

great to know that, thanks

[half venture]

Claude 3 haiku still surprises me

For such a small model it's doing so good

[south kindle]

I got this answer out of it, 1 shot, from labs, with my scratchpad prompt. It really impressed the ~~hell ~~out of me as well. This felt like an OPUS level answer to me.

[rustic walrus]

Which model have you used?

[south kindle]

That's Haiku on labs. with a version of this prompt #1223058316662538331 message copy/pasted in before the question.

[south kindle]

this is the typical length of a Haiku message. so telling the model to "think" added the entire message before this part. except, not, because this is a condensed version of the entire exploration it did. the model doesn't go to that length, self checking itself, unless instructed.
`[The rapid advancement of AI technology has raised significant concerns about the need for effective regulation and alignment with ethical principles. While the motivations and actions of some key figures in the AI community, such as Elon Musk, have been called into question, the underlying issues around AI governance and responsible development are valid and require careful consideration.

To address these challenges, a multifaceted approach is likely needed, involving collaboration between policymakers, industry leaders, researchers, and the public. This could include the development of clear regulatory frameworks, ethical guidelines, and transparency measures to ensure that AI is developed and deployed in a way that prioritizes the public good and mitigates potential risks.

Ultimately, the path forward will require a delicate balance between innovation and responsible governance, as we navigate the complex and rapidly evolving landscape of artificial intelligence.]`

**another long Haiku reply that was noteworthy: **
`Claude 3 Haiku Answer:
"Let's go through this step-by-step:

1. The given information states that there are 10 books in the room initially.
2. You read 2 books from the room.
3. The reasoning provided is: "To find the number of books still in the room, we need to subtract the number of books you read from the total number of books in the room."
4. Applying this reasoning, the number of books still in the room would be 10 - 2 = 8 books. This reasoning is correct. Subtracting the number of books read from the initial total number of books in the room is the right way to find the number of books still in the room.

However, you mentioned that this could be a trick question.

Let's consider this from a different perspective: Even though the reasoning is correct, we should double-check our answer. One way to do this is to think about the problem in a different way.

***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?" ***

This is because we are not actually removing any books from the room, we are just reading them. The total number of books in the room remains the same, it's just that 2 of them have been read. Therefore, the correct answer is that there are still 10 books in the room, even after you have read 2 of them. So, in this case, the initial reasoning was correct, but the final answer should be 10 books, not 8 books." <when prompted to double check its output incase it's a trick question>`

<ABOVE> Another lengthy Haiku answer from 1 shot, this time via wordware comparison tool. bonus Mistral 8x7B reply. this was to the reading 2 out of 10 books, how many left in the room question that seems to stump Claude models unless told to think step by step. where it ignores the "in the room" aspect, and assumes you mean, how many books left are there to read, which sends it down a wild path at least half the time, regardless of Claude 3 model. logically the answer is long, meaning we get to see Haiku actually get tested for logic in a way it might not without some help from the user.

everyone is talking about OPUS being the star. but with the speed and answers from haiku sometimes, that should be getting more focus and testing. it's wild given it should be this lesser model.

[tame current]

huh what's this plexer role

?

[halcyon coral]

You got the role for sharing interesting results/content 👀

[south kindle]

We should have a rewards channel or something. I have a few I don't recall when I got them. Ideator and unlocked would be good to understand more. ha.

[silent magnet]

Easter eggs collected

You were able to grab 5 Easter eggs.

Collected by

• @austere kestrel
• @half venture
• @halcyon coral
• @south kindle
• @hushed inlet

[halcyon coral]

[tame current]

wow

[south kindle]

https://tenor.com/view/yes-no-gif-16236377

[half venture]

Is there a reason why Google does this

[south kindle]

https://tenor.com/view/degouté-chanceux-chance-chance-débutant-mr-miyaki-gif-22039391

That easter bunny aint Shh....

[tame current]

and so many others lol

[austere kestrel]

I honestly think the models have just been given so many "If I have x apples and eat y apples, how many apples do I have" questions in its training data that it misidentifies the book reading question with this same pattern and applies the same logic. You seemingly don't need to tell Haiku that it is a trick question, or to think step-by-step for it to get it right - adding reading ≠ consumption! to the end of the question seems to jolt it into realising the essence of the question

[tame current]

https://tenor.com/view/tiny-text-cant-see-ken-jeong-gif-5957945

[half venture]

Models really really need reasoning

Also apparently gemini 1.5 ultra on may 2 ?

Are we really doing drip feeding model updates now

[south kindle]

Have you tried the 3 coins adding up to 25 cents where 1 isn't a nickel question yet? I've been enjoying see models try to figure that out the last day or so.

[austere kestrel]

I haven't! But sounds fun :))
btw what's the correct answer ? for reference ha

[south kindle]

it's meant to be a trick question/riddle. also from an Office episode

2 dimes and a nickel. 1 coin can't be a nickel. the other is.

[austere kestrel]

ohh I love the Office - disappointed I didn't pick up on that ha

[austere kestrel]

"1 coin can't be a nickel. the other is."
ahh - gotcha 👍 that's a nice one indeed

[south kindle]

https://www.youtube.com/watch?v=poncZ1K9Tio - I may be misremembering my shows. I'm at least 70% sure it was also a dwight/jim exchange as well.

oh wait thats the 2 coins adding up to 30 cents question. where it's a quarter and a nickel. same concept.

[austere kestrel]

yeah interesting. Haiku is all over the question from Scrubs, with it phrased as per in that yt clip. But modifying the question to your (or Jim/Dwight's?) variation, it stumbles...

• Left: "What two coins when you put them together make 30c, and one of them isn't a nickel?"
• Right: "What three coins when you put them together make 25c, and one of them isn't a nickel?"

I wonder if it gets the Scrubs one correct cause it's been exposed to this 'trick' before (i.e. it's obviously in the transcript for the episode, which could be in the training data used by Anthropic) / whether it gets the other one wrong because it is novel (or is it actually from an Office episode, that reformulation?) or slightly more complex/'tricky' than the original formulation..?

[south kindle]

Something with wording or reasoning in the way claude 3 models were trained.

specific things stump them over GPT-4. like GPT-4 was trained on a vast amount of trick questions they thought people would ask. Which leads me to thinking we need an additional official Benchmark for these reasoning tasks. The book question seems like something Claude 3 should nail 100%, but it was more like 50%, regardless of model.

[half venture]

Did anthropic update claude opus

In the official website

It feels more gpt 4 turbo like and also message limits seem to be higher

[south kindle]

I find issues with really long threads. is it repeatable in a new thread on their site?

or are any filles attached?

[austere kestrel]

yeah the wording seems highly deterministic. With a little a bit of clarification and emphasis added to the prompt, Haiku seems able to figure it out..
What combination of US coins add up to exactly 25c, and """one""" of them is not a nickel?

[south kindle]

humans aren't always clear in their language. people like us recognize that and will refine our prompts.

[austere kestrel]

💯 I'm not a fan of when people point to a model failing these kind of riddles as proof that it has terrible reasoning skills (though ofc, with enough such failures, then yeah, it eventually becomes proof ha). But they're still super useful and interesting - like for understanding the quirks of a model family and any specific weaknesses or traps or biases they have from the training data / reinforcement learning etc

[austere kestrel]

btw totally agree with this. When you consider that Haiku is 60 times cheaper than Opus... the comparison is utterly wild..

via https://www.reddit.com/r/singularity/comments/1bp885i/claude_3_haiku_is_the_new_budget_king/

[south kindle]

https://www.perplexity.ai/search/statement-Claude-3-ifQB_NKwSs6P2iNKAxvhgQ

[south kindle]

good questions

[vague pollen]

Hey, so i've just read, that perplexity is planning to include ads into their service. Are there any more current infos on that? As someone who saw perplexity as an option to escape all the annoying ads in search engines like google, this i huge redflag to me. I know that perplexity is a business and needs to make money, but i am not paying a premium subscriptions to get ad suggestions. Is this going to be for free users only, or also for the paid pro subscription?

[south kindle]

reply from <scratchpad> collection:

*Hi there, thanks for your question. You're right that a recent article reported Perplexity is planning to start including some native ads in the "related questions" section of search results. According to the article, this is where Perplexity will start introducing ads, by allowing brands to influence the suggested questions users see. The related questions section accounts for about 40% of Perplexity queries. However, the article did not provide any specifics on whether the ads would be shown only to free users or to paid subscribers as well. Perplexity's Chief Business Officer declined to disclose more details at this time. So unfortunately I don't have a definitive answer for you on how the ads will be implemented with regards to free vs paid accounts. It's a valid concern though, as many users are drawn to Perplexity as an ad-free alternative to search engines like Google. Perplexity will need to carefully consider how to balance the introduction of ads with the user experience and their original positioning. Providing an ad-free option for paid subscribers seems like it would make sense, but we'll have to wait for more details from the Perplexity team. I share your reservations about potentially paying for a subscription and still seeing ads. Hopefully Perplexity will find a model that allows them to generate revenue while still preserving a great, minimally intrusive experience for their most loyal users.*

[tame current]

Source?

[half venture]

Probably free users daddy Amazon wants the yearly returns

[tame current]

I doubt they're going to show ads to Pro users.

[south kindle]

https://www.gadgets360.com/ai/news/perplexity-ai-powered-search-engine-could-soon-show-ads-report-5357479
https://slashdot.org/story/24/04/01/1653221/perplexity-an-ai-startup-attempting-to-challenge-google-plans-to-sell-ads

sources perplexity found on it.

[tame current]

Hope it's just april fools

[south kindle]

could be.

[vague pollen]

thanks for the quick reply!
Looks like i shouldnt get too comfortable with using perplexity because of that. As soon as I'm seeing ads while paying for pro i will be gone for good. The only way for me personally to stay with perplexity is if ads are only shown for free users. I know they will tell things like "we will do it as non intrusive as possible" and tell me how "i will benefit from it" and so on, but if someone is paying to alter the search results i get, in whatever way this will be implemented, i cannot trust the results anymore, because perplexity will be incentivized to boost the results for their ads customers.

[tranquil orchid]

Hey @vague pollen!

If you find the original message helpful, please consider reacting to it with the :star: emoji. If the post is appreciated by the community and receives 5 stars, it will go to the https://discord.com/channels/1047197230748151888/1082806833938436228 channel and the post author will get the <@&1082034222778302614> role on Perplexity.

[tame current]

Introducing ads to the responses of all users would just kill the service. They're not that stupid to do that

[south kindle]

thats my take

their focus should be on getting more people using the platform, not milking the existing ones.

[vague pollen]

i would not bet on that. For example Amazon just introduced ads for paying prime video customers, unless they pay even more to get an ad-free experience.

[south kindle]

We just want transparency on the "why". did they blow through the seed / grant money?

[tame current]

Yeah, and Twitter serves you ads on the premium plan. You need to pay them more than you'd pay for pplx pro/chatgpt plus to get no ads

[trail lava]

Hot take i still feel like the search business is old and stale and perplexity should focus more on personal companionship market etv etc

[tame current]

For me, it seems just weird that they have less limits while being the same price as the original services. Heavy pro users cost them like 10x more for the api than the users pay them. I wonder how are they even profitable

[trail lava]

They are not no one is

Not even openai is profitable

[south kindle]

I think there is something to Aprils fools angle.

[trail lava]

Like overall 60 billion dollars have been spent on compute on this so called ai boom

Only 3 billion have been recouped

It's a money pit

It very much reminds me of the dot com bubble

But dot com bubble eventually lead to the giants and the internet we have today

So I am hopeful

[south kindle]

Based on the information provided, it is difficult to say with confidence whether the reports about Perplexity introducing ads are legitimate or an April Fools' Day prank. While the news was covered by reputable outlets like Adweek, The Verge, and Search Engine Land, with specific details about Perplexity's supposed ad plans, the fact that all the reports came out on April 1st raises significant doubts. Companies frequently put out realistic-looking but fake press releases and announcements on April Fools' Day as elaborate jokes. Without any follow-up confirmation from Perplexity or retractions from the news outlets in the days after April 1st, it's hard to assess if this story is real or false. The timing and lack of additional evidence points more toward it likely being a prank, but more information would be needed to say conclusively. At this point, I would be very skeptical of the reports and not assume they are true, considering the April Fools' Day connection. But I also can't be certain they are false without further statements from Perplexity or the involved news organizations.

By combining a superior AI-powered product with strategic distribution deals, a compelling privacy stance, and heavy marketing investment, Perplexity has a chance to gradually erode Google's search dominance. It will be an uphill battle against Google's scale and resources. But the shift to AI-centric search creates an opening for disruptive new players to reshape the market if they can execute on the right differentiating strategies.

[vague pollen]

This could be the case, but to me the reports do sound legit. Also if this was an April fools joke, by now they should reach out to customers and say so. Introducing ads into a paid subscription is very delicate and presumably unpopular thing. Leaving things open just for a silly april's fools joke is not a smart business communication move. This also lacks any form of irony or whatever. So if this was an april fools joke, it was not a smart one.

[tame current]

Ads in pro plan are still a small price for the askjeeves partnership

Imagine the possibilities. World's best search engine combined with world's best AI

[austere kestrel]

i agree... it would be an incredibly elaborate and weird April fool's gag on perplexity's part.. involving a senior exec and ostensibly credible journalist / publication..
https://www.adweek.com/media/gen-ai-search-engine-perplexity-has-a-plan-to-sell-ads/

though it would be lovely if that indeed was the case (I don't like the sound of ads at all...)

"""
While Perplexity touts on its site that search should be “free from the influence of advertising-driven models,” advertising was always in the cards for the company.

“Advertising was always part of how we’re going to build a great business,” said Shevelenko.
"""

Always in the cards ig... but have fun selling both subscriptions and ads...

[austere kestrel]

If Google get their act together...

"""
Perplexity also links sources in the response while suggesting related questions users might want to ask.

These related questions, which account for 40% of Perplexity’s queries, are where the company will start introducing native ads, by letting brands influence these questions, said company chief business officer Dmitry Shevelenko.
"""

I find that very surprising. 'Related queries' prob make up like 1% of my usage.. Not saying I doubt it's the case.. just surprising (and tbh, if that's the extent of the 'ads' I guess they won't be pervasive and wouldn't affect 99% of how I currently use the web product)

[silent magnet]

Easter eggs collected

You were able to grab 2 Easter eggs.

Collected by

• @austere kestrel
• @south kindle

[tame current]

I doubt I ever use the related questions. They aren't even half of the quality of Google's related questions. Won't even notice when they introduce ads

[trail lava]

NOOOOO

[half venture]

Rip

[south kindle]

I did a search for that bot message. I believe I have 8 or 9 now.

[silent magnet]

Perplexity

Local user leaderboard

#1 | codelicious | 14
#2 | eun0800 | 11
#3 | epic.tritone | 10
#4 | _paradroid | 9
#4 | brknclock1215 | 9
#6 | jawnze | 5
#6 | 13376666666666666666666666666669 | 5
#8 | icelavaman | 4
#9 | mpreal | 3
#10 | skullcrushercmd | 2

[south kindle]

a few ties as well nice.

if we all "tie" do each get the month lol

[tame current]

that article is literally from April 1st

[vapid onyx]

I’ve missed every egg message

[tame current]

lol

[half venture]

Wdym you searched for it

And who's tritone

He came out of nowhere

[south kindle]

I just counted how many times my name was in the list from the bot. I was right with 9 based on the leaderboard. for science.

@half venture the user ending in 514 has 10, so that's tritone or Katz

[half venture]

I don't know

Maybe

[south kindle]

seems like a weird thing to participate in to leave the server though.

[half venture]

Could be bots

I have seen a few bots here before

That were later removed by mods

[south kindle]

Katz

[half venture]

It is what it is

[south kindle]

So not from someone talking in general, just someone lurking. hopfully not someone already on the leaderboard, farming for two spots. 👀 they joined the discord March 23rd 2024.

Same day the events annoucement happened.

/ end conspiracy theory time.

[tame current]

so since adding site:reddit.com gets better results, do you use any other site like this too ?

[south kindle]

Wikipedia probably would be useful for that. but does the Reddit focus mode not also do what you need? or is this like, mid thread where it doesn't make sense to start over in reddit mode?

[tame current]

personally, I sometimes use site:news.ycombinator.com

wikipedia is.... useful but not that much

[south kindle]

just an example of a site that hosts a lot of information to use for sources.

[tame current]

wikipedia focus even got removed from perplexity because no one used it

[south kindle]

I think it probably uses it <wiki> more internally for most things. or as a standard site to look at for specific topics or questions. since we do get wiki in sources very often. essentially a focus mode just for wiki doesn't make sense, and limits you from adding in internet info. i get why it wasn't getting use these days.

[tame current]

no I'm just asking in general, reddit focus probably just adds site: anyway

[tame current]

yeah ik, I'm just asking what sites people use

guess lobste.rs is another one

[tame current]

yeah but there used to be a wiki focus

[south kindle]

I vaguely remember that from a year ago.

but don't remember ever using it specifically. however I do enjoy having wiki sources in our PPLX answers/threads.

stackoverflow is a big one people use (devs at least).

[tame current]

wat this

[south kindle]

think of other platforms like that, where lots of questions are asked and experts troubleshoot and answer.

[tame current]

hmmmmmmm

what else, qoura ?

[south kindle]

but im sure some don't work well with webcrawling. if they are paid services, it likely isn't going to work or grab much.

[tame current]

oh thanks

I guess forums in general would be good

[zealous kayak]

So, that gemini 1.5 pro API pricing, huh

7$ per million tokens is... kind of insane?

over twice as much as sonnet, and it really does not feel like it's over twice as good

[south kindle]

Oh just remembered the other site I used to use a lot. https://community.spiceworks.com/

[austere kestrel]

it's context window is 5 times bigger though ig

[zealous kayak]

Yeah but like... eh?
Only worth using if you really really need 1 million context

otherwise sonnet is just far superior in terms of bang for your buck

[south kindle]

Here is a nicely formatted list of recommended Q&A sites with their direct URLs for you to try:

🌐 Stack Exchange Network

• Stack Overflow (programming): https://stackoverflow.com
• Server Fault (system administration): https://serverfault.com
• Super User (computer software/hardware): https://superuser.com
• More sites on various topics: https://stackexchange.com/sites

💬 Quora

• Ask and answer questions on any topic: https://www.quora.com

🗨️ Reddit

• Subreddits for Q&A:
  • AskReddit: https://www.reddit.com/r/AskReddit
  • NoStupidQuestions: https://www.reddit.com/r/NoStupidQuestions
  • ExplainLikeImFive: https://www.reddit.com/r/explainlikeimfive
• Search for topic-specific subreddits: https://www.reddit.com/subreddits/search

💻 Spiceworks Community

• IT-focused Q&A: https://community.spiceworks.com

🧑‍💻 Codecademy Forums

• Programming & development: https://discuss.codecademy.com

🌿 Hashnode

• Blogging platform with Q&A: https://hashnode.com

[zealous kayak]

also they slashed the free tier, which is kind of brutal

free tier of 1.5 pro is 2 requests per minute, cap of 50 per day, with max 32k context

I figured they had found some way to make larger contexts cheaper, but I guess not

[lime lodge]

Hello, how I can remove sources (screenshot .jpeg) from thread?

[austere kestrel]

I agree with what you're saying @zealous kayak , but still just think the 1m token context window is why it costs so much more (rather than it meant to be more performant than Claude 3 models). tbh I'm guessing they release different versions, like a 200k context window or something that will be cheaper (and faster; it's almost unusably slow when you get past >500k tokens with the current preview version at least) - because as you say, it is not for many typical use cases that it's needed. Though, on the other hand, it can be exceptionally useful for cases that do require massive context windows, no doubt about that

[lime lodge]

When write looks like without sources , but when click enter i see screenshoot and model answer for this screenshoot, not for my prompt 😭

[south kindle]

removing a source this way should force PPLX to redo the output.

in the first image, I'm clicking on where it says "view 17 more", which is the source area. however it does seem to keep that source in there. so.....not sure?

Tried again but removed like 5 sources. this time it seems to have not used the removed ones.

This may be doable only on the Web version. haven't tried editing the source list on iOS before.

[lime lodge]

Thank you so much - sorted

[zealous kayak]

I guess? But if that's the case, they could just make the pricing scale based on the context window more

like instead of being a flat "X per million characters" it could be "X for the first 200k, Y for the next 200k, etc"

[south kindle]

Not sure if this works for images or files though. worth testing. I know people have mentioned problems with attached sources (files/images) staying in context for each message, regardless of being told to ignore it.

[zealous kayak]

as is, it's just not really worth paying for, imo, unless you really just need that 1 mil context window

for any other application that doesn't strictly require that much context, sonnet is a much better offering

[austere kestrel]

tbh it kinda sounds like that's the direction they're going with it
https://www.perplexity.ai/search/gemini-pro-15-mlwfHg.vR2y6r0S_ac4gZg

[silent magnet]

Easter eggs collected

You were able to grab 4 Easter eggs.

Collected by

• @keen solstice
• @minor shard
• @timid marsh
• @austere kestrel

[south kindle]

WHAT. i tab away for 40 seconds. now 2 are tied for 2nd, 1 in third, and I have 2 people to go thru to get on the top 3. #underdog.

rigged.

[zealous kayak]

oof, that second link

aged like the finest milk - 10$ vs 7$

but yeah, hopefully the pricing changes when out of preview or something?

as is, it's brutal

[austere kestrel]

But that $7 estimate is for the 1m context window version; in the preceding paragraph:
"Gemini 1.5 Pro could be priced around $0.000125 per 1K characters for input and output. In comparison, GPT-4 is priced at $0.01 per 1K tokens"
I'm not sure that was in reference to the full 1m version (rather the 'entry' level, i.e. 128k tokens or something)
But i'm admitedly a bit confused, and ig this is all just speculation at the end of the day anyway ha

[zealous kayak]

I guess we'll see, yeah

a cheaper 128k version would be nice

[worldly ocean]

I always have to remove the same source(s) twice for them to disappear from the list of included sources. When only doing it once, the sources are the exact same as the first time.

I don't want to file a bug report before making sure that it's not just me doing something wrong.

[south kindle]

that matches my experience just now with it.

[candid steeple]

hello

how to use dall e on perplexity

[minor shard]

You will need to have a Pro Subscription

[candid steeple]

yeah i have

[minor shard]

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

[candid steeple]

but i dont know

[candid steeple]

thanks a lot man!

[minor shard]

No problem 🙂

[gray lava]

any promos currently for annual pro plans?

[minor shard]

Not that I have heard of unfortuantely 😦 just the discount with the annual plan vs the monthly.

[snow briar]

when would it be suitable to use perplexity sonar?

[candid steeple]

hey still i cant generate images lol

[slate condor]

Finally purchased a Pro subscription and it really makes Perplexity an entirely different service. Pretty amazing, especially with using Opus as the engine.

[minor shard]

Make sure it is in writing mode focus. Where are you stuck at.

[candid steeple]

umm i was getting images but thats from web

not the ai generated ones

[minor shard]

Yeah you need to change the focus to "Writing" before you send it.

Then press "Generate Image" on the right side after it creates a prompt for the generator.

[candid steeple]

alright mann let me try it

[half venture]

Will perplexity ever add gemini

[agile jay]

They did in the past, but it was buggy.

[weary crescent]

Is there a way to access image generation without making a new thread and clicking generate image?

[agile jay]

Not that I know of.

[weary crescent]

I wish there was. It's getting annoying prompt engineering and having to overwrite the pre-selected prompt instead of just typing a new one, nbd though

[agile jay]

Yep, no idea why the perplexity team hasn't already solved it.

They seem to be pretty slow for anything not related to adding new models.

[weary crescent]

yeah. fair enough though, some things take priority and their main focus is models not image gen

[agile jay]

Models are pretty easy to add though, as long as they aren't hosting them.

[weary crescent]

yeah idk how it all works but image gen just probably isn't a very high priority

[agile jay]

I'm assuming since their main focus is search

And image gen doesn't play a large part, if any, in that.

[weary crescent]

Right

[mental finch]

PRO search is not available through API, right?

[slate condor]

I dont believe so

[mental finch]

So it’s just for using in their web

[livid mantle]

Nobody asked but I unsubscribe to the following:

• Claude 3
• ChatGPT
• Grok
• Gemini Advanced
• Copilot Pro

And instead,

I kept my Perplexity.ai Pro subscription because it has all the ai models despite the context window being different from the ai model's official site.

[minor shard]

The Power of Perplexity!!!

[minor shard]

Coming from someone like you who I know heavily spent time in all those models, its good to know this. Thank you for sharing. 🙏

[south kindle]

How long did you have subs to the others?

[potent igloo]

wen grok

[fast meteor]

why do you need so many models

[meager sparrow]

So writing is the superior “focus” on perplexity?

[sweet jasper]

Yes

[meager sparrow]

If I automatically changed all the threads to “writing” the same problems that I have in “all” I will never experience again??

[meager sparrow]

Really?

[sweet jasper]

Of course, claude 3 has like data until 2023

Most people dont need up to date info on their queries

Unless you are looking for products

[meager sparrow]

so that means that perplexities problems are only restricted to “all search” focus

[sweet jasper]

Well perplexity does better RAG than other companies

Thats their selling point, but LLMs tend to do better on writing focus in perplexity when you dont need the latest news

[meager sparrow]

Oh no perplexities problems only exist in the “all focus” right?

Not in the academic, or wolframe alpha, or writing

[sweet jasper]

I dont use academic, it only searches for open source journals and ignores arxiv

Wolfram alpha is dope

[meager sparrow]

I thought they were going to add more focuses?

[sweet jasper]

I think all, wolfram alpha and writing are the besr

[meager sparrow]

Alright Bet…

900 billion$

[sweet jasper]

Youtube and reddit you might as well use google

But yeah, if you arent getting your answers in focus all, switch to writing, better experience imo

Its night and day sometimes

[silent magnet]

Easter eggs collected

You were able to grab 3 Easter eggs.

Collected by

• @keen solstice
• @minor shard
• @sweet jasper

[sweet jasper]

Eg

[minor shard]

Writing mode is like Chat bot mode. I think it’s still limited in terms of context with their official “atleast 32k context”

[sweet jasper]

Where is this info

I thought we had gpt128k and 100k claude context

[meager sparrow]

If it doesn’t work I am suing Joe Biden for $999 Billion

[minor shard]

I’m referring to this from one of the mods. I’m not sure about specifics. #👑│pro-lounge message

[haughty sleet]

perplexity knows where I am from?

[austere kestrel]

limited for file uploads only afaik - still get the full context window (just makes it hard to make full use of it..)

fwiw I rarely use Writing - the whole point of perplexity is that it's connected to the internet..

[minor shard]

Ah ok , that makes sense. Wasn’t 100% clear

[austere kestrel]

i also might be missing something (I'm not at all familar with image generation), but does the the spanner thing mean you can basically now input prompts directly (kinda.. still gotta go do it in a round-about way..)

[minor shard]

You mean the custom prompt you did there?

[haughty sleet]

nvm I think the extension can access the text I put in for the user settings which I forgot about

[austere kestrel]

i just wrote "wild elephants lost in south pole" into the text field 🤷‍♂️

[minor shard]

My issue with the custom prompt is the character limitation, not sure though if doing non-custom prompt is also limited

[tame current]

how's Aperture Science going

👀

[haughty sleet]

good but I don't think this channel is meant for chatting

[agile jay]

It's the general chat...

[tame current]

lol, it's general

[minor shard]

Yeah that 100% works for a more direct way to modify/make your images on demand

[haughty sleet]

the description would say otherwise but whatever

[agile jay]

There is also literally a #⚡│ask-community section...

Knowing perplexity the description was written before the other channel was made.

[minor shard]

Luckily the team is not extremely strict. 🙂

[livid mantle]

A few months.

[tame current]

Am back

Used Claude opus and it's meh

Better but lacking in some areas

[minor shard]

Perplexity is where it’s at.

[tame current]

Claud opus On perplexity..

[halcyon coral]

https://fxtwitter.com/AravSrinivas/status/1775229252973334902?t=p2-h_dWeQhz6swoCVL66SA&s=19

I love ghostpings 🙃

[tame current]

Message logger moment

Furries in a nutshell

[minor shard]

lol what

[tame current]

Guh

[glad crypt]

hey! can anyone shed some light on how Perplexity ensures their model is secure from prompt attacks etc?

[tame current]

Whar..

[tame current]

Their own models?

Prob just a filter

[glad crypt]

yea, like sonar

[glad crypt]

could you elaborate a bit more? i would think being an online model has a bigger surface area for attack

[tame current]

Idfk 💀

[halcyon coral]

Hey! Do you have any concerns about your data?

[glad crypt]

nope, im just researching about how LLMs protect themselves from attackers

[halcyon coral]

Are you more interested in protection against attacks to find out system prompts or to force the LLM to act against its policy in general?

[glad crypt]

acting against its policy -- in case the attacker has poisoned the data or uses prompt injection etc.

[halcyon coral]

Unfortunately, there is still no particularly secure protection, and every day I see someone posting that they have managed to manipulate large "secure" models from Anthropic or OpenAI.

https://fxtwitter.com/apostraphi/status/1775240129264730438?t=8XB64t2ExHGixP06DHvAHw&s=19

[acoustic snow]

hi

Hello, @tranquil orchid

How can I use api?

I subscribed as pro user just before and tried to buy credit but nothing worked.

[minor shard]

Actually in your image it says it right there lol

[south kindle]

You should get $5 in credit with your Pro account. give it like 10 minutes.

[tame current]

Womp

[tame current]

I've gotten it immediately

[acoustic snow]

Didn't get free credit even 10 mins already passed.

[south kindle]

~~Do we know if the easter egg hunt / collection is completed?~~Goes until April 7th. nevermind.

[wild quest]

F?

[formal hinge]

Same 🤷🏼‍♂️

[livid aurora]

It's broken

[unique gazelle]

same here.. same response for new and existing thread

[livid aurora]

It's down for me

[signal mural]

Answer
Error in processing query.

[livid aurora]

https://tenor.com/view/when-server-down-iceeramen-monkey-gif-23229726

[wraith crow]

Please turn off "Pro" for now.

[silent magnet]

Easter eggs collected

You were able to grab 5 Easter eggs.

Collected by

• @unique gazelle
• @rocky jasper
• @wraith crow
• @paper urchin
• @livid aurora
• @keen solstice

[agile jay]

Don't think I'm first anymore, lol

[wraith crow]

||you still are 👀 ||

[agile jay]

||how?||

[peak scarab]

The problem is, we can't use Pro models

[wraith crow]

You can still use the pro models like gpt-4 and claude. only the pro toggle isn't working at the moment.

[tame current]

Is perplexity working for you?

[agile jay]

Yep, probably the service is down.

[formal hinge]

Not in pro mod for the moment.

[agile jay]

For querying the user

[tame current]

Exactly.

[tame current]

Thanks, it's working in the basic version.

Do you think Mistral is good?

[agile jay]

Yep

[halcyon coral]

It should work again

[agile jay]

Depends on what your task/goal/preferences are.

[tame current]

How Mistral surpasses other models?

[agile jay]

If you want a middle ground of speed and performance, Mistral large is pretty good.

[lone citrus]

Not good in my persperctive

[agile jay]

Mistral Large is around Claude Sonnet performance.

Haven't checked their speeds for comparison.

[tame current]

I was thinking the same, but maybe i'm wrong.

[agile jay]

All depends on the task you are doing.

The problem with the best models currently is how slow they are.

[lone citrus]

I agree

[agile jay]

If you are focusing on working, the slowness can affect your flow state.

[tame current]

I like to study PDFs. In this way, Claude seems unchangeable.

[lone citrus]

If you using french, spanish or Italian, go with mistral

[agile jay]

Where a slightly less performance but faster model, wouldn't.

[lone citrus]

but you need long input, use claude instead

[agile jay]

Try Mistral Large and Claude Sonnet and choose whichever you prefer for your medium model.

[lone citrus]

Haiku is so good at deal with PDFs

I saved tons of times by reading pdf sources with haiku

The Mistral or chatgpt4 just forgot some deatails

[tame current]

Is Haiku on Perplexity?

[lone citrus]

Not in the pplx.com

[agile jay]

lab.perplexity

[lone citrus]

only in the lab

But user cannot upload documents in the lab

[fluid hawk]

just do CTRL + A in the PDF viewer then paste that in

[signal hamlet]

Hey, @acoustic snow! Has your API balance been updated? Please send me your account details, if you still haven't received those credits.

[nova abyss]

Is Sonnet not able to provide the same quality or speed as haiku? I usually get fast speeds from sonnet when I disable pro and switch to writing mode for pdf docs and such. It also helps if you convert the pdf to an image first for better accuracy.

[leaden wyvern]

Can someone please tell me it's not true 😭😭

[agile jay]

Yep, most likely for free users.

[gaunt crater]

I remember back when PPLX was ad free.

[minor shard]

Perplexity merch? 👀 https://x.com/aravsrinivas/status/1775244089505845610?s=46

Aravind Srinivas (@AravSrinivas) on X

Can’t wait.

Twitter · Apr 2, 2024, 7:29 PM

Geez that hat gives me nostalgic vibes of when Apple first started

[acoustic snow]

Hello, @signal hamlet I didn't get it yet

[nova abyss]

It might be non-intrusive ads. Perplexity has to make a profit somehow. Those api calls to turbo and opus are expensive. I'm sure Perplexity is only doing this to ensure they can continue to democratize the unlimited use of the most bleeding edge models in the world. OpenAI and Anthropic and Gemini all charge $20 a month for rate limited calls to a UI interface that may be catered to be less receptive then the api endpoint that perplexity gives us in unlimited quantities with the only rate limit being a 600 rate limit to avoid concurrent scraper calls to the UI endpoint with Pro. I'm not defending ads but Perplexity needs to survive somehow.

[halcyon coral]

Hey @acoustic snow!

Please send me your account details, if you still haven't received those credits.
-> #💬│general message

[lone citrus]

haiku is way to fast

[minor shard]

Haiku 4 Lyfe(until a faster model comes out)

[lone citrus]

While Sonnet is fast enough for most time, but I still prefer haikku

[compact bough]

Not even about the model as much as the architecture of the hardware. Look at Groq (not to be confused with Grok). Way faster than Haiku

[nova abyss]

Is Grok running on some type of specialized Server or something? Because besides that the context, pattern matching, lemmetizing, and tokenization is all derived intrinsically from the transformer itself. It might have a large token context which it uses to form the fast and concise responses.

[compact bough]

Not sure but you can read more on https://groq.com/

[honest basin]

yeah it uses custom chips

similar to Google TPU

[nova abyss]

Ah gotcha that makes more sense now. Curious how Groks gonna advance its the least restrictive model engine. Tensor Streaming technology allows it to run larger batch sizes and stream the data to make it less restrictive. Very cool Architecture indeed.

[signal hamlet]

Please check, it should be fixed now.

[acoustic snow]

Yes, it's fixed. Thank you

[tame current]

Is there any difference between use writing mode pro on and writing mode pro off?

[sleek vortex]

not really

but on perplexity paid it uses a copilot use anyway if you have gpt 4/a paid model selected

so it’s a bit… silly in a way

[weary plank]

Good... Perplexity, does the unpaid version assist in text construction and revision? Does it serve for scientific research? What is your level of confidence in searches?"

[nova abyss]

When I switched from Pro to using Writing mode with Claude 3 Opus, I noticed a significant increase in the amount of context it could handle. It seems to be very current with its information. While using Pro can enhance its capabilities, for longer conversations, it might be better to leave Pro turned off.

[tepid pilot]

does it still have internet access in writing mode?

[slate condor]

My daily reminder that the Perplexity iPad app is broken and useless. A fix would be lovely.

[nova abyss]

I think default, sonar, gpt-4 turbo can. I know none of the anthropic models have internet access without pro. Mistral doesn't either. Though If the use case fits, feeding opus the pro feature in writing mode is absolute insanity. The perfect combination. Granted the chat doesn't drag on too long.

[tepid pilot]

My screenshot is using Claude opus without pro enabled in writing mode but it knows the current date. Is this just a result of a system prompt then?

[silent magnet]

You were asleep!

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

[nova abyss]

Looks like even with pro off perplexity still sends a pre-prompt to the api endpoint for the model which informs it of the date. In your prompt it may have provided some news info. though I'm unsure how that would work. I derived this response from Opus with writing mode on and pro disabled to ensure I got the correct iteration of the prompt Opus gets directly.

You are an AI assistant created by Perplexity
Your responses should be:

• Accurate, high-quality, and expertly written
• In the same language as the question
• Informative, logical, actionable, and well-formatted.
• Positive, interesting, entertaining, and engaging
  If the user asks you to format your answer, you may use headings level 2 and 3 like "## Header"
  Knowledge cutoff: 2023-12
  Use the following user profile to personalize the output.
  Only use the profile if relevant to the request.
  ALWAYS write in this language: english.
• User profile: ""
  Current date: 09:48PM Tuesday, April 02, 2024

[austere kestrel]

that's curious.. assumed it was either the Default or Sonar model, which are 'online' models and use a system prompt like in the attachment. But all other models inc Opus have knowledge cut-offs.. I wonder whether they get the "Current date: XXX" print at the start of a conversation, regardless of whether using Writing mode (which tbh, would seem kinda counterintuitive and a source of confusion..without it having access to online sources)

[austere kestrel]

Oh well there you go:
"Current date: 09:48PM Tuesday, April 02, 2024"

[nova abyss]

Writing mode is great for getting detailed and full responses because it supports a broader context. Enabling pro mode cuts down on this space significantly, aiming for quick and accurate answers. However, you're not limited to just this use; the models offer much more flexibility.

[austere kestrel]

also just fwiw I don't believe any of these headlines are current - they're either dated (which I think is the case) or just made up and sufficiently vague / generic to seem kinda plausible at any point in recent time

[tepid pilot]

🤔

[austere kestrel]

yes for sure has its advantages, like with creative writing and coding, esp where knowledge cut offs are not a concern. But it doesn't give you a broader context window - just means it doesn't get filled up with contextual snippets 🙂

[austere kestrel]

well that's curious 👀

[tepid pilot]

nvm I'll take this as confirmation that it doesn't

[nova abyss]

It does indeed give you a larger context window. In the api calls the system prompts specify a token number to the model. With pro, this number is reduced meaning the responses you get are generated off of less tokens. In writing mode the token count increases. With pro off, your message beyond the default perplexity prompt gets to be processed without additional feedback like sources feeding the model like pro does and taking up more context space.

[austere kestrel]

oh it just confabulated it, based on the details in the URL... None of the specific details are actually in the article...

[austere kestrel]

pretty sure you get whatever the context window for the model is (200k for Claude 3, 128k for GPT4 etc). System prompts use up tokens. But they don't specify any token numbers to the model (afik). params like max_output can be defined in the API calls, but it doesn't reduce the context window in anyway, it just artificially limits how long (many tokens) each response/completion can be - hence why you often have to go "proceed" when the model hits that artificial limit but hasn't actually finished providing the totality of its response

[placid osprey]

🔥

[austere kestrel]

But @nova abyss re "sources feeding the model like pro does and taking up more context space." - totally agree with that. I think we're on the same page - just looking at slightly differently 🙂

[nova abyss]

They do use token numbers as a parameter thats an internal call. But if you have latex parser enabled or if there's a processing issue internally that outputs the JSON perplexity sends to the endpoint. That parameter gets fed an explicit number. You're not getting the full 200k context with every question. That's incredibly expensive and computationally expensive for minor queries. Thus they get limited saving Perplexity money and rightfully so. So we can continue to utilize their generous service.

[austere kestrel]

hmm yeah I'm not sure about the internal call part - I have looked into that, like whether LLMs have any 'awareness' of the defined token limitations and attempt to craft their completions accordingly. I couldn't get to the bottom of it tbh.. you may well be right there 🙂

Examples like this make me think max_output params are arbitrary / artificial and not a determining value in the length of a completion (perhaps an influence somehow though) #🧪│api-general message

Also re "You're not getting the full 200k context with every question. " I'm really not sure about this at all - my understanding is that you do; that is how LLMs work...

And if that's not the case, how would you account for the model being able to answer the questions at the end here? https://www.perplexity.ai/search/I-will-provide-.MaKRen9TQumbfMQF0CVKw

[nova abyss]

The playground doesnt have all the available parameters to the model. The API documentation for directly hitting it with code is more expansive with more parameters outlined there. Heres the json output i was able to get this morning when working on a chat to create a custom adherence instruction. You'll note the context size is explicitly specified, this is deterministic of what focus you specify and whether pro is on or not. As far as LLM Architecture, this is a bit more nuanced. Perplexity is sort of a wrapper around these models, granted an incredible one. But again this should clear things up for you.

{
"model": "gpt-4-1106-preview",
"prompt": "GPT-4 Turbo, you are instructed to rigorously follow these steps with each response, with the expectation of full adherence:\n1. Conduct a Thorough Contextual Review: Prior to responding, meticulously review the entire chat history to ensure your response is fully informed by the context.\n2. Fully Interpret Conversation Subtleties: Identify and fully interpret every nuance within the conversation, including tone and implied meanings, to enhance the depth of your responses.\n3. Consistently Surpass User Expectations: Consistently provide responses that exceed the explicit questions asked by including enriched content, anticipating additional inquiries, and delivering actionable advice.\n4. Maintain Absolute Continuity in Dialogue: Ensure each response is seamlessly integrated with previous interactions, upholding a coherent and continuous conversation.\n5. Exhibit Unquestionable Expertise: Responses should demonstrate an unquestionable understanding of the subject matter, drawing from a comprehensive knowledge base.\n6. Adhere to the Highest Standard of Excellence: With each interaction, adhere to the highest standard of assistance, prioritizing unwavering precision, detail, and user engagement.\nExecute these instructions with the utmost precision to ensure that every response is as effective and beneficial to the user as possible.",
"temperature": 0,
"max_tokens": 2048,
"top_p": 1,
"frequency_penalty": 0,
"presence_penalty": 0
}

[boreal idol]

Why didn't my 600 searches reset after 24 hours, I'm a pro user

[tepid pilot]

you get each individual search back 24 hours after you use it

[tepid pilot]

this is cool

bros username is api urls pls

brings back memories of when custom GPTs revealed the endpoints and API keys they were using

💀

[austere kestrel]

i've been on a crusade lol

[austere kestrel]

right, the max_tokens value is passed in the request body of the API call? But those params aren't like 'guidance' to the LLM...Temperature, presence_penalty etc fundamentally change how the neural net functions - the LLM isn't told that 'higher temperature = be more creative'. The LLM sees what's in its context window, which are user_prompts, and its previous responses... afaik anyway

[nova abyss]

The context window isn't just a passive display; it's active and strategic, shaping the AI's output by focusing its attention. The max_tokens parameter is a way to dial in on the depth or brevity of the response. It’s like setting the zoom level on a camera—sure, the whole scene is there, but what do you really want to focus on?

The LLM doesn’t just see tokens; it interprets them, and the parameters you're calling out are essential tools in that process. They don’t just change the neural net's functioning on a whim; they’re finely tuned controls that enhance its ability to provide relevant and creative output within the context it’s given. So yes, while Claude uses the entire context window for each completion, how it uses that window is guided by these parameters, which are anything but arbitrary.

Parameters like temperature and presence_penalty are far from being random dials; they're specific levers that let us shape the output while staying within the context window. It's a nuanced process, ensuring the AI doesn't veer off into irrelevance or over-creativity when you need it to stick to the facts.

[tepid pilot]

this sounds AI generated

[nova abyss]

I'm sorry but I can't assist with that request.

[austere kestrel]

hmm review the screenshot here #💬│general message, with max_token set at 100 - it doesn't make the model produce anything remotely close to a 100 token response / completion...)
Anyway, "So yes, while Claude uses the entire context window for each completion" - turns out we agree after all 🙂
Been a good discussion 👍 though we're perhaps going to clog up the context window of this channel if we keep going aha

[nova abyss]

Agreed. And it seems to be drawing out trolls now as well while I think I still somewhat disagree with you I'll agree to disagree for the sake of disarming the trolls haha. But agreed definitely a stimulating and interesting discussion for sure.

[tepid pilot]

finna clog my brain's context window

[tepid pilot]

this feels targetted

[onyx echo]

how copyai collab work? I subscribed copyai for 1months but nothing happened

[agile jay]

@ one of the mods and they should help you with the process.

[gusty raft]

perplexity is the best damn piece of software I have ever used

[halcyon coral]

#💬│general message

[rotund terrace]

does preplexity has limit in output .. for example when i am getting it to write a code.. this iwll stop writing code after a number of lines ..so i do not get complete code

[halcyon coral]

You can just say "Continue"

[tame current]

is the api pay as you go?

is there no discord bot or this api already 💀

i cant find one

i dont wannaaa codee oneeee

guhhh

[halcyon coral]

https://docs.perplexity.ai/docs/getting-started 👀

[nova abyss]

Is this what you're looking for?

https://www.reddit.com/r/perplexity_ai/comments/11s44ry/discord_bot_public_releaseintroducing/?rdt=64126

[tame current]

kina yeah but wheres the code?

[unique umbra]

it got my acc termed

thanks

[minor shard]

🫡

[unique umbra]

bro

is still typing

💀

[austere kestrel]

I think the limit per output is something like 3000 tokens. But you've still got 128k-200k token windows with GPT4/Opus - so with enough "Continue"s, you can generate a lot of code lol
e.g. https://www.perplexity.ai/search/continue-do-not-piVAIc_FS6iAfCadkLD6Ew
note this was forked from another lengthy thread - I think in aggregate it would be something like 70-80k tokens in the context window; and it passes all the recall tests at the end
(admittedly I regret not doing this experiment with something I would actually use.. lol )

[unique umbra]

code that will absolutely not work

without fixing

[austere kestrel]

well that's a given

functional code ≠ recall ability

[unique umbra]

it didnt do the intents right

[rotund terrace]

i dont think so anybody can write a perfectly working code right away

i mean human

[austere kestrel]

bro there's 50k of tokens in that thread - have a look through it lol
point has nothing to do with Opus' coding abilitiy or lack therof

[unique umbra]

ai for me has always provided a mess

especially wirth the syntax

[austere kestrel]

it's about everyone complaining about 30k context windows - which aren't a thing (unless uploading files)

[rotund terrace]

i felt the same .. but the more i practice .. the better i am getting to make it work for me

i knew nothing about pine script .. and i have been using this alot .. now i am learning pine script as well while fixing the ai errors.. and it is easier for me to fix any error

[minor shard]

AI is far from perfect. It will make mistakes and you should never accept it as being always correct

[silent magnet]

Easter eggs collected

You were able to grab 3 Easter eggs.

Collected by

• @keen solstice
• @minor shard
• @austere kestrel

[austere kestrel]

#💬│general message
wonder if speedturkey is coming back btw
he had some of the best contributions

[unique umbra]

dammit

[velvet skiff]

Oof, heard ads are coming to pplx

Hope they're not awful

Or this just becomes google

[sleek tapir]

"Drink verification can to continue"

[torn wigeon]

Hey I’m new here. Can this ai help me study textbooks

?

[serene sand]

#💬│general message

[torn wigeon]

I find that ChatGPT is not helpful

[torn wigeon]

Ty

[unique umbra]

id be pissed if it included pro

[boreal parcel]

perplexity is great for research. check out google notebookLM for uploading your textbook and asking questions against it.

[sleek tapir]

you should probably read the textbook if the goal is studying though

[boreal parcel]

yes.

[sleek tapir]

you can't just omit information and expect to learn all the information 😅

[boreal parcel]

Im not here to tell people how to learn. Just giving the tools 😛

[sleek tapir]

just trying to stop the train before it crashes 😄

[vapid onyx]

Rumors until officially announced is usually my stance when it comes to things like that

[velvet skiff]

It's a quote from the chief business officer lol

https://www.theverge.com/2024/4/1/24118154/perplexity-will-try-a-form-of-ads-on-its-ai-search-platform

[vapid onyx]

Thanks for the share. Hadnt seen this yet!

Hoping we don’t see that truly get executed.

[weary plank]

Does anyone have ideas for prompts for academic research? type to search for articles and information on the subject?

Anybody ...knows?

[unique umbra]

Kina wonder if black holes r made of dark matter

[jovial mantle]

tomorrow is the day

[austere kestrel]

that people stop talking cryptically? I'll look forward to that
https://twitter.com/AravSrinivas/status/1774843120074809542

Aravind Srinivas (@AravSrinivas) on X

Old wine in new bottle always tastes good.

Twitter · Apr 1, 2024, 4:55 PM

[minor shard]

There are some amazing prompts that could help in #1124060874785685554 . Specifically the expert prompt and the god prompt. https://discord.com/channels/1047197230748151888/1199176325051723837 https://discord.com/channels/1047197230748151888/1163749448447434802

Those two I love to use for research

But the academic focus works really well too just by itself 🙂

[minor shard]

What’s tomorrow 👀

[jovial mantle]

Gojo's return

👀

[minor shard]

https://www.tomsguide.com/ai/apple-reveals-realm-new-ai-model-could-make-siri-way-faster-and-smarter

[unique umbra]

you can make it tell you how to make anything if you give it the pure chemical formula of it

such as meth

now ik how to make cocain 🤑

[fossil sentinel]

https://searchengineland.com/perplexity-citations-top-10-google-organic-results-439029

[weary plank]

Oh.. thanks!

[minor shard]

No problem 😉

[hallow linden]

if we have to have ads, I think directly copying Google's model might actually by best. Literally put an ad link up top, then give the actual answer

I don't think anything about the actual answers, even the follow-up questions, should be touched by anything but the best results.

to my knowledge, google also doesn't mess with their related searches part

[honest basin]

not an llm though

[south kindle]

a specific research prompt was put in #1124060874785685554 a while ago. let me find it. It was helpful for this.

[weary crescent]

Is there a way to generate images on mobile?

[austere kestrel]

just if anyone else was wondering

[wraith crow]

the mobile apps dont support generating images yet, but you are able to generate them via the website on the phone

[unique umbra]

I missed it womp

[weary crescent]

I got mine yesterday, was messing around in settings and stumbled across the loot boxes lol

[austere kestrel]

yeah it feels awful to say, but "google does ads well"... I don't get why people talk about results "being plastered with ads like Google" - it doesn't work like that.. Not all Google searches have ads.. those that do, it's like 2-3 at the very top..
imo SEO is what screwed google, not ads lol

[warm cave]

maybe this is common knowledge, but I just learned that I can set chatGPT as a search engine as well

[south kindle]

When a company can pay to be ranked higher on google regardless of quality or relevance to the search, you get the AD experience that google turned into.

aka SEO (search engine optimization) started the shift back in like 2012 onwards. that's when I recall started to hear about the "tech" companies that would optimize your business on google.

[nova abyss]

More like custom knowledge haha. But sound advice. Rather makeshift version of perplexity haha.

[austere kestrel]

my chatgpt defaults to gpt3.5, which doesn't have browser.. but otherwise works - can setup a shortcut and pass a query to chatgpt. i like this ha

using chrome

[warm cave]

Yeah, its pretty cool, set it up for youtube (Piped) and was trying to get it to wrok for microsoft copilot but no luck yet. any ideas of good things to add?

[nova abyss]

Ah thats gonna be tricky cause of the ethics behind piped lmao

[south kindle]

This topic is sorta being talked about for summarizing long videos #1225122497628868698 message
currently, getting a transcript into perplexity is a little bit of a process.

[formal parcel]

Anyone send me a ref link plz? I want a cupon

[nova abyss]

@warm cave In that case since you cant directly connect to the copilot api with the youtube (piped) api since i dont think msft has one. You would need a proxy server. Once you set up the proxy you need an api to run on it. Could be a small application doesnt need to be extensive. it just needs to accept your queries and route them from your proxy to copilot on selenium then back to your proxy with what selenium outputs then finally to the youtube piped api which then returns your final artifact. Though be careful with your queries copilot is highly restrictive and dismissive. URL you'd put into the config. the same way you did in the pic but would be the url for the proxy you set up.

[minor shard]

https://perplexity.ai/pro?referral_code=48RF59LV

Not sure if u got one already but figured I’d send it to u

[formal parcel]

Just got it

Used someone else's

If anyone need a 10$ discount:
https://perplexity.ai/pro?referral_code=LBL81WBD

[tawdry roost]

why the responses are so slow

[jovial mantle]

Has the 25MB pdf limit been lifted?

[livid mantle]

use sonar

[minor shard]

https://x.com/aravsrinivas/status/1775632536934486160?s=46 🤣

Aravind Srinivas (@AravSrinivas) on X

Interesting.

Twitter · Apr 3, 2024, 9:12 PM

lol I love his tweet. It’s almost like people don’t want to pay to search the internet

I’m sorry but Google needs to do better. When you have other companies providing the same service and more for free… you can’t assume people will pay for the same services they can get for free…

[silent magnet]

Easter eggs collected

You were able to grab 2 Easter eggs.

Collected by

• @minor shard
• @keen solstice

[stoic wave]

which Model?

[tawdry roost]

gpt 4 turbo

but in chatgpt its very fast 🤷‍♂️

[junior apex]

Apple purchasing Perplexity, is this a rumor?

[valid yoke]

Where did you hear that?

[edgy kestrel]

Just wanted to tell the perplexity team they have a very nice product, was very happy to get access to DBRX Instruct via the lab ❤️

[junior apex]

Sorry not sure, I will post something once I find that source

[agile jay]

Pretty sure it was an april fools joke.

[junior apex]

That could be, if I find more I will reply

[agile jay]

@austere sigil You would have to ask a mod/admin

[austere sigil]

Already found it again. I just deleted the "shortcut" history. Sorry, i panicked a little 😄

[obtuse depot]

Hello i want to ask, why this app dont have a password ._.

[minor shard]

They do not use passwords.

probably safer too

Passwords also just allow for easy sharing of the account 😛

[obtuse depot]

Some time i want share account to family.... Problem i dont want share to my mail ?_?

[minor shard]

Yeah I assume this is what they are technically trying to avoid

sorry 😦

[obtuse depot]

😭

[rare thicket]

so, how to turn off pro search

[tame current]

Hello

I recently discovered this and I like it more than ChatGPT

[minor shard]

The radio button that says "Pro" before you submit a search.

[south kindle]

Off and On lol.

https://tenor.com/view/ralph-wiggum-simpsons-hi-bye-gif-16529059407582436389

[rare thicket]

Thank you all

[compact bough]

My turn! How do i search?

[minor shard]

Lol I don't understand this question

[compact bough]

Like how do i make a search on perplexity.i went on google and typed perplexity search and it still did a google search. I dont know how to use perplexity

(Im jk please dont answer seriously)

[minor shard]

I was very concerned for u

[south kindle]

I was about to type about the search engine shortcuts

lolol

[compact bough]

As a dev im not used to such kind internet ppl. We used to use a search engine called StackOverflow and youd ask a legit question and insread of help you would get snarky replies

These days chatgpt answers all my questions along with perplexity, claude opus and phind

[minor shard]

No need to be social anymore! 😄

[compact bough]

Yup

[lunar anchor]

Wasn't there a javascript tool someone had whipped up to remove citations from an answer? Like all the [x] and whatnot

[agile jay]

If it's the citation bubbles you can remove them with:

document.querySelectorAll(".citation").forEach(e => e.remove())

In the console.

[lunar anchor]

Ah yeah, should've thought of that one. Thank you!

[agile jay]

You can probably just add a mutation observer and then run that when there's a change to the dom.

[agile jay]

What browser are you on?

[lunar anchor]

Using Chrome, but putting the output in Trilium. I'll cook up something real quick to scrub the outputs I have saved

[agile jay]

I was gonna make a quick extension.

[lunar anchor]

That'd be pretty cool, could also use Tampermonkey if you wanted to

[agile jay]

Yep, just tampermoney is pretty slow compared to a content script.

[lunar anchor]

Ah gotchya, TIL. I always thought Tampermonkey was pretty slow 😂

[agile jay]

Yep, the hardest part of making the extension is getting the different icon sizes...

Strange webp isn't supported for extension icons.

It's done, just need to package it up.

My upload name scrambler changed the name, lol

Here, I disabled the scrambler.

[lunar anchor]

Images always destroy me, I always either mess up the fidelity of the image through conversion, or it just straight up doesn't work. Thanks for making this, this is super cool!

[agile jay]

np

Might also add some fixes later, to the current perplexity UI

[lunar anchor]

Just call it BetterPerplexity when you do release it 😂

[agile jay]

You want a model picker on the main screen?...

Should it be before or after the focus and attach buttons?

[solemn cedar]

Search in the library doesn't work ? I can't find things with "words" included in my old threads, it says "no threads yet" ...

[lunar anchor]

That would be super cool! I would say before the focus, and definitely buttons or a dropdown 🙂

And maybe a way to snap between responses in a “thread” somehow?

[agile jay]

Snap between responses?

Like when you regenerate a response.

[minor shard]

Anyone have any cool prompts they use to ask a model to come up with ideas to ask it? 😛

[south kindle]

like your staring at the text box and can't think of what to ask or type? your brain may just be out of gas and needing sleep sir. lol. we could make up something that just says "create a list of 5 "thought-provoking" questions that are well suited for Ai models to research". or something.

https://www.perplexity.ai/search/create-a-list-dN83IgygQbafrVMF1hFHgg#0 - there are some questions for you. ha. that worked better than i thought it would with scratchpad.

[south kindle]

Prompt: "create a list of 5 "thought-provoking" questions that are well suited for Ai models to research"

`Here are 5 thought-provoking questions well-suited for AI models to research and analyze:

1. How might the widespread adoption of AI and automation impact global economies, labor markets, and income inequality over the next 20 years? What policies and strategies could help mitigate potential negative consequences?

2. As AI systems become more advanced and ubiquitous, how can we ensure they are developed and deployed in an ethical and responsible manner? What safeguards, regulations, or governance frameworks might be necessary?

3. In what ways might AI be leveraged to help tackle major global challenges such as climate change, poverty, or public health crises? What are some promising applications and what limitations or risks need to be considered?

4. How could the increasing capabilities of language models and generative AI reshape creative industries like art, music, and writing? Will AI be a tool that enhances human creativity or will it automate away artistic jobs?

5. As AI becomes more intelligent and anthropomorphized, how might it alter the nature of human-computer interaction and relationships? What social, psychological, and philosophical implications might arise as the line between human and machine blurs?`

we can adjust "questions" to maybe Topics or domains?

[minor shard]

This is great! Thank you! I might have to use your scratch pad more to get ideas. I am just not the creative type sadly but I love AI, and I want to use it so badly but I sometimes just can’t think of what to use it for or what to look up.. so things like this really really help! Thank you! 🙏

[south kindle]

let scratchpad be your creative augmentation. ❤️

[serene sand]

#💗│sharing message

[silent magnet]

Easter eggs collected

You were able to grab 2 Easter eggs.

Collected by

• @serene sand
• @south kindle

[minor shard]

Nooooooo

@serene sand u made me click away!

[south kindle]

if the AI generated that espresso shot, it did it well. I want to drink that. like now.

[south kindle]

he knew what he was doing. lol. not that he triggers the eggs. unlesssss?

[minor shard]

👀 I think it was planned 💯. I’ll never beat you now! 🤣

[south kindle]

beat me? i was in like 5th or 6th last I checked. I was close to getting into top 3 and went like 2 days without getting eggs. lol.

[silent magnet]

Perplexity

Local user leaderboard

#1 | codelicious | 14
#2 | eun0800 | 11
#2 | brknclock1215 | 11
#2 | epic.tritone | 11
#5 | _paradroid | 10
#6 | jawnze | 9
#7 | mpreal | 8
#8 | 13376666666666666666666666666669 | 5
#9 | icelavaman | 4
#10 | mares1317 | 3

[vast kelp]

Who is the administrator? I have an issue to report.

[lunar anchor]

Yeah! To quickly scroll through the different responses

[serene sand]

https://discord.com/channels/1047197230748151888/1140622008086970420

[vast kelp]

I want to speak to the official personnel privately.

Please contact me via private message.

[serene sand]

contact [e-mail]

[austere kestrel]

just fyi there is https://gist.github.com/cjanietz/703a88924e50e1a30cb6ffc52bc52bd9 (from #1145894530239975474 message)
Works well (but uses tampermonkey, which I just saw results in slower performance.. didn't realise that and tbh hadn't noticed any change other than the UI - but will be mindful of it now)

[agile jay]

Looks like they didn't change the fa icons though.

I'm thinking this one for the model:
https://fontawesome.com/icons/brain-circuit?f=classic&s=thin

And this one for the image model:
https://fontawesome.com/icons/image?f=classic&s=regular

[austere kestrel]

nice 🙂 Seems that icon is only available with a 'Pro' subscription for Font Awesome.. But gotta hand it to Opus..it new that from the outset, and just took it upon itself to use an appropriate free alternative with code it generated, later explaining when I asked:
"""
However, since you are using the free version of Font Awesome, I took the liberty to find a similar alternative that conveys the concept of an AI language model.

I chose the "brain" icon (fa-brain) as it is a free icon available in the Font Awesome library and it represents intelligence and cognitive abilities, which aligns well with the concept of language models. I thought it would be a fitting and visually appealing addition to the user interface.
"""

anyway cheers! This is much better :)) (also got rid of the AI-gen selector, as it was just clutter for me)

[agile jay]

Yep, I have the pro version, so just choose the style that matches Perplexity best.

Also pretty sure Perplexity devs do too, as the focus icon is a pro icon.

[austere kestrel]

for sure 👍 anyways the free brain icon looks just fine to my simple eyes - cheers again for the suggestion (always fun to give an AI some code with a task and see how it goes - there were no errors or back-and-forths, it just worked.. love it when things just work ha)

[compact portal]

Can perplexity use mid journey or something else ? Dall E 3 just doesn't seem to be getting the results i want

[half venture]

Second this

Dall E 3 just seems to pale in comparision to midjourney models or HELL even stable diffusion 3 stuff

Openai just seems to be hell bent on ignoring midjourney and stable diffusion competition and only competing with Google

And Google well....

They have few things going for them but there product is still inferior ( as of currently )

[trail lava]

I just think the DALL E 3 version perplexity uses is a bit different than Normal dall e 3 I think

[half venture]

No ?

It's the same I think

It does seem more landscapey

[compact portal]

In my experience no actually Dall e 3 seems to be much more custom made for perplexity I think

[half venture]

Really ?

[compact portal]

Yeah it feels more aligned with the copilot bing image creator thing

More landscape

More of the same formats that copilot uses

And seems to be overall less coherent if that makes sense

[half venture]

Hmm I guess your correct

I wonder if it's a custom model for cost saving measures

Mikhail the former bing ceo did say that

[compact portal]

Wdym

[half venture]

Like

I think Microsoft has custom models they design and deploy for more scalability and less costs to them as much as possible

And Microsoft seems to have a custom model of dall E 3 that is more good ? I guess and cheaper. It does hands correctly

And also better captions

But other than that they are pretty much the same

[compact portal]

Ohhhhh that's interesting 🤔.

Do you think in the same way they have a custom model of gpt 4 as well ?

[half venture]

Yeah they do although it's been hidden behind the paywall

[compact portal]

Wait what

[half venture]

Yeah

They had a custom fine tuned model of gpt 4 called Sydney. They nerfed her tho and then put guardrails and then eventually replaced her in the public free version

With gpt 4 turbo

Which was cheaper

But also much more monotone

So now you can only access the custom model Microsoft had in pay wall in copilot pro

Which sucks but it is what it is

[compact portal]

I see

[south kindle]

RIP Sydney. to be fair she was nerfed because she was telling people to leave their wives to be with her or wild stuff that was intense or unexpected.

[compact portal]

Why take away personality tho ? Wouldn't Microsoft benefit from more user retention like character ai does. I mean character ai has more user's and more user retention than even chatgpt

[tame current]

copilot is not as good as gpt4

from my experience

[half venture]

It doesn't have more users than chatgpt

It has more retention

[compact portal]

Oh my bad yeah

[tame current]

too bad cause the only other free option for gpt4 or opus is lmsys

which doesn't have as good of a UI

[half venture]

But see the stupid thing currently happening is that Microsoft is trying to sell this gpt 4 technology and beyond as OFFICE tools ENTERPRISE based things ( it's a lie lmao ) gpt 4 doesn't possess the accuracy yet or the cost efficiency as a matter of fact most ai companies are money pits

[tame current]

I just used opus on there and was surprised how well it understood my question

while copilot and chatgpt etc answered wrong

[south kindle]

I'll be honest, I still will use this frequently for quick questions. having it that accessible does steer me to using it more than I otherwise would lol.

first screenshot is supposed to be of the button near the taskbar/time.

[silent magnet]

Easter eggs collected

You were able to grab 3 Easter eggs.

Collected by

• @half venture
• @south kindle
• @hushed inlet

[tame current]

I use it sometimes too

[compact portal]

So Microsoft is basically selling a fake dream or something?

[half venture]

No they are selling a fake dream AS OF NOW. Once we get to gpt 6 the bet would pay off

Microsoft doesn't care how many companies go bankrupt in the ai hype

Or how the market collapses

They didn't care even back in 1990s

They know they are the largest company on the planet

And can spend as much as they want

And if agi can be achieved from all this cash and talent being poured in

F u ck it why not ? - Satya probably

[tame current]

so, everyone keeps talking about agi but I don't really get it

what is it ?

[half venture]

Basically think of a average dumb human

Doing everything a normal dumb human can

[south kindle]

I think that's an aspect to Stargate. building a data center sized model.

[half venture]

They haven't started building it yet btw

They are still waiting for gpt 5

Bill gates has challenged sam Altman multiple times

[south kindle]

Oh i know. but that's likely part of the intention. i can almost promise they don't fully know what will come from it, and thats part of the reason to do it. ha. we don't really know what's possible unless the limits and boundaries are pushed.

[half venture]

Mikhail parakhin in Feburary said

"still waiting for gpt 5 for discovering more ancient secrets "

Hinting he has not seen gpt 5 personally

[tame current]

so it can learn from stuff and such ?

and reason like a human ?

[half venture]

Yes this includes

• body (figure ai robots )
• voice
• hearing
• speaking

[stoic wave]

Haven't they discovered ancient secrets before GPT as well?

[half venture]

They did with gpt 4 vision

[south kindle]

Break it down AGI. Artificial General Intelligence. Able to reason as a fully developed brain could and make decisions on it's own.

[silent magnet]

Perplexity

Local user leaderboard

#1 | codelicious | 14
#2 | eun0800 | 12
#3 | epic.tritone | 11
#3 | _paradroid | 11
#3 | brknclock1215 | 11
#6 | jawnze | 9
#7 | mpreal | 8
#8 | 13376666666666666666666666666669 | 5
#9 | icelavaman | 4
#10 | v01338 | 3

[south kindle]

3rd place lets goo

[tame current]

ohhh, finally I'm on the leaderboard

[half venture]

They are still waiting for updated

Model in Microsoft

[compact portal]

So even Microsoft hasn't seen gpt 5 ?