💬│general

[livid mantle]

wasting the claude 3 opus api

[warm cave]

hey, it takes a lot of trash prompts and testing to find the gems, Only benefits

[devout bronze]

how does perplexity make money 💀

im sending crazy af long prompts to Opus that usually bill my Anthropic account like $0.20

[south kindle]

https://tenor.com/view/imagination-spongebob-squarepants-dreams-magic-gif-12725683

[devout bronze]

i guess lol

[south kindle]

evolution is a linear process. can't make useful things without starting from somewhere. i would guess you've built on your skills with prompting with each thing you've "published". using " " because pretty sure we are all just sharing stuff and not actually "publishing" anything. i could be wrong though. lol. I stand corrected, you've been "publishing" to pastebin. I'll count it.

[livid mantle]

Master Perplexity search like a pro!

Let's craft a text to explore every search operator.

👉 #1222013886488645652

[south kindle]

https://pastebin.com/HxBzM6pz - throwing here for now. Grabbed the entire simulation of 65000 characters and dumped it into claude 3 (sonnet, free user) and had it review it. Pretty neat.

i do wish claude had a share function though. come on now Anthropic.

[tulip sonnet]

[wary scroll]

I do

[south kindle]

https://pastebin.com/TZk6svLV - trying to reverse engineer the worldsim prompt. sonnet just spit this out. haven't tested yet.

[wary scroll]

its seems that im using gpt-3 instead of 4 because it can't compile files

[weary crescent]

I'm here from Rabbit, just wondering how well perplexity does math? Here's an example Finite Math, and WolframGPT's response. Would somebody be willing to send it through Perplexity and let me know the response?

[wary scroll]

when i ask it if it's gpt-4, it tells me it is not, and that its perplexity

[south kindle]

#💬│general message - see above.

unless the model was trained to say what it was, it has no actual way of knowing what itself is as far as I understand. so any outputs would be hallucinations.

[wary scroll]

well yeah gpt-4 tells you. but on perplexity it says that it isnt gpt but rather perplexity trained by perplexity

[south kindle]

what is your default model set to in settings?

[wary scroll]

turbo

[south kindle]

at the bottom right of the models reply should show the model used. you should see GPT-4 down there.

I have mine set to OPUS, so I get this on all replies

feel free to share one of your threads having issues. maybe something obvious is goofed for the file aspect?

the reversed engineered "worldsim" prompt "somewhat" works with perplexity/opus. going to attempt to introduce the monolith....

https://tenor.com/view/monkeys-2001aspaceodyssey-stanleykubrick-gif-8729999

https://tenor.com/view/2001a-space-odyssey-2001-bone-scene-bone-to-spaceship-kubrick-gif-21680310

last of the worldsim spam for now. back to the actual site/tool. monolith created and placed on earth by the....Galactic Uplift Agency?

Querying state of Earth 2 million years after monolith deployment...
Assessment:
The monolith intervention has accelerated the development of intelligent life on Earth by approximately 2 million years compared to the original timeline. The species has made significant progress in understanding and manipulating its environment. They stand on the threshold of advanced capabilities such as spaceflight, nuclear power, and digital information technologies.
However, the population remains fragmented by competing nation-states and ideological conflicts. Disparities in wealth and resources are evident. The industrial activity is beginning to destabilize planetary ecosystems. These challenges must be overcome to ensure a unified and sustainable planetary civilization.
Recommendation:
Continue monitoring Earth's progress. Initiate psychic subconscious messaging to promote peace, collaboration, and environmental stewardship. Guide scientific and philosophical development towards cosmic awareness and preparedness for galactic contact. A formal first contact mission may be advisable within the next 2000 years if development proceeds favorably.

EDIT: how did I not realize this entire "worldsim" tool, is just meant to model Asimov's story and writing.

[spice lance]

links broken in perpl now?

[wary scroll]

is there a way to turn pro search off

think i found it

[formal marlin]

Does anybody used the ai for trading ? And what is the script for that

[livid mantle]

ur still playing with that, you've gained my interest

whats so special about it

im backreading ur posts now

so its a kind of story lora simulation

[south kindle]

mostly. yes. but the intention is for universal simulation. managed to get it to spit this out.

[solid moss]

Why should I need to do bot checks if I'm a paying subscriber. Surely the money should be enough of a check? Also, it would be great if requests wouldn't just silently drop. For a long time I thought the service was down before I noticed that every time I refreshed the page, a captcha would come up, and then I could continue

[livid mantle]

u cracked the ode

c

[nimble field]

I wish making money was that easy

[austere kestrel]

lol honestly right

[native robin]

im facing a weird issue , if i attach an image in a to and fro convo and later remove it . still the model takes it into reference each time

tries answering everything i ask using that image or file i sent ago

[tame current]

@exotic spoke

[silent magnet]

Easter eggs collected

You were able to grab 1 Easter eggs.

Collected by

• @gentle pewter

[wary scroll]

does anyone know how to bypass ai detection for writing

[austere kestrel]

same with uploaded files (and pasted text). You can remove it as a source, but then when you ask a follow up question - it's back again! you ofc can then remove it from the sources again; but it's tedious / not ideal imo

[zealous totem]

just gotta do some heavy editing, that’s the way i do it

[whole lark]

To.

Glad to see more icons available.

[wary scroll]

perplexity including my image constantly for each prompt is so annoying

[halcyon coral]

#1221790602895556628 message

[silver epoch]

This website had a good run

Unfortunately I've moved on to Gemini

Anyways

gudbye

[south kindle]

most of us use multiple LLMs/tools at any given time.

[forest oar]

Is this type of things new?

[livid mantle]

Why I vannot see that channel IceLavaMan?

[halcyon coral]

You can, just click on the link. Maybe you have to restart Discord 😉

[abstract steppe]

I just signed up for Perplexity Pro, and still feeling my way around. I also pay for ChatGPT Plus and Poe.com, if that tells you anything. Anyhow, I'm trying to nail down the lingo, so to speak, around Perplexity's interface. Can collections be thought of as the equivalent of GPTs or what Poe calls "bots"? I want to create certain tweaked instances that always remember their function when they start.

The "AI prompt" in collections does seem to be a small window, so there's only so much I can put there.

[serene sand]

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

deep dive ➤ #💬│general message

[lean hamlet]

In perplexity pro tier i am unable to use claude's image model. Whenever I try to ask questions regarding an image it goes through GPT4-turbo. This feels like a bug, Is this same for everyone?

[austere kestrel]

that's the same for me (was wondering whether Claude 3's multimodality is actually being used to process image uploads..)

[lean hamlet]

looking forward to seeing this fixed if any of the perplexity devs are here

[austere kestrel]

Try doing an image upload with Pro toggled off...
There's a bit of work to be done it seems lol

[lean hamlet]

yea shows opus in the response now

[austere kestrel]

I'm getting the correct bade now too

[lean hamlet]

not sure if its actually opus tho since the response is not correct lol

[austere kestrel]

But still, without Pro toggled on, I'm being told the best things to use a plunger for whisking eggs or, now, measuring ingredients in the kitchen lol

[round plover]

nice

[austere kestrel]

yeah - it's curious. I don't get why having Pro on / off could affect the quality of image recognition.. And also it seems that the responses are very similar, regarldess of which model is selected

[tame current]

is this real ?

I'm guessing no ?

[minor shard]

Im waiting for when they start telling you that your computer has viruses because you use Chrome

[fluid mauve]

what you really ewant is a chrome ectension along the lines of a ransomware splash screen saying you have encrypted 'The Internet' and demand payment before you decrypt it. lol. The amount of people that would not cotton on they still had internet access could be a large number.

[hallow linden]

https://www.theverge.com/24111326/ai-search-perplexity-copilot-you-google-review

What do y’all think of this article?

[minor shard]

I read this. Partially correct too. Perplexity obviously is not there yet to be a total replacement. But 100% getting there and totally possible. I think once the models get even faster and the process becomes more streamlined through consistent use by people, it will be the replacement. If google doesn't adapt while they have the advantage they will get replaced. I think Perplexity is betting on the fact that Google isn't going to change fast enough.

It will be like when people were used Internet Explorer and then Chrome came out and provided something way better and faster.

[wary scroll]

anyone knows a fix for when it's providing u code but cuts of midway?

is there a solution/trick

[foggy whale]

I’m perpetually annoyed at the genre of article where a journalist merely tries some of the most basic things, then the thirst of the article doesn’t actually say anything like “hey, this doesn’t hurt, just go try”

But saying that we still need search engines because the browser does a bad job of slapping .com at the end of my dad’s Facebook query isn’t moving the needle for me that’s for sure.

[tame current]

bakock

[weary crescent]

I just claimed my $200 credit from rabbit, it looks cool so far. Do you guys ever find yourself butting up against the 600/day search limit?

[foggy whale]

It’s like “oh, another person who hasn’t seen The Browser Company’s whole thing”

[tame current]

pro search kina sucks so no

[fluid mauve]

It is a decent if not a bit long article. I agree that PPLX and other AI answer engines are not quite there yet. But as @minor shard says, it wont be long. Google at its heart is an ad space agency. Selling placement of ads in yoru search stream. varying the records returned by who payed the most money.

I like that perplexity is clean and ad free. i do wonder how monetization will be baked in, to replace Google search, you cant be chatging 3 billion people $20.00 a month for full access.

As the saying goes "We live in interesting times"

[tame current]

just buy it for the models

[eternal quartz]

idk how you're using it but I love pro search

the questions are pointless. but otherwise it's great

[fluid mauve]

I typicaly just type continuwe in the prompt box and it carries on

[minor shard]

I do not. Maybe doing extensive research all day but I have not hit it.

[weary crescent]

how so?

[tame current]

it pretty much asks you questions about your question

[fluid mauve]

never, lucky id i touch 50 a day

[tame current]

which is kina annoying

[weary crescent]

I saw that, but you can skip it. I think it's helpful if you need to give more context but don't know how

[minor shard]

I just skip the questions Pro Search gives me. The large amount of sources is soooo nice.

[eternal quartz]

It does a lot more than just ask the questions. the search results are much more relevant and up to date with pro search on. I just skip the questions.

[tame current]

you can skip it but it takes an extra 10 or 20 seconds to do all that

[austere kestrel]

that's a great article (I think the way he breaks it down into three categories of why / how people use search engines is spot on) - cheers for sharing

[tame current]

just rapid fire questions about the answer it gives if ur still unsure

[tame current]

should have a dif mode then

[eternal quartz]

yes, it should

[spark flint]

I’m sure there is a place that covered this question or it has been covered many times, but what are the differences between AI models available in pro version? Any advantages or disadvantages for certain ones? I feel the perplexity AI has been the most advanced AI I’ve been able to try out so far but I haven’t given the rest a thorough analysis.

[weary crescent]

You guys run into "File upload failed moderation"? I attached an image of my notion home page...

[frozen wedge]

I am using perplexity for studying finance and law, which model would give me better results, gpt 4 turbo or claude 3 opus?

[eternal quartz]

Honestly there used to be some differences...but now I just use Claude Opus for everything lol. I ❤️ Claude 3

[austere kestrel]

@weary crescent I misread notion as nation

aha

[weary crescent]

hahaha

[eternal quartz]

Depends. Personally I like Opus the most but others disagree. they're close enough that it's not likely to make a big difference.

[tame current]

I found gpt to be more accurate

[frozen wedge]

Thanku

[austere kestrel]

I agree with this 💯

[silent magnet]

Easter eggs collected

You were able to grab 1 Easter eggs.

Collected by

• @austere kestrel

[tame current]

Eggs

[austere kestrel]

would be interesting to do 'blind' tests - aside from stylistic traits, substantively I think it's often splitting hairs

ofc, different topics / niches / tasks, different models will excel where others fall short

[tame current]

There's probably a test designed for ai models

[austere kestrel]

there's all kinds of evals, like the MT-bench in the third column. But the more interesting one is the elo arena - you can contribute yourself https://arena.lmsys.org/

[weary crescent]

Opus just came up with this https://pastebin.com/raw/fVn4xBTM

[austere kestrel]

can see on the elo scores, gpt-4-t and claude Opus are basically tied.
btw it was last updated 13 March.. must be a refresh coming any day now - will be interesting to see

[supple vine]

aren't these always the claims with something new? I mean when google came out I'm sure there were simmilar articles comparing it with Altavista, that didn't think google would take the crown. These things always go this way, until something moves into first place. I personally use search mostly for what the article calls exploratory searches, and AI is just better there. It does need some work on the other types, but I don't see that as an insurmountable obstacle. I could be wrong though.

[stable radish]

Quick question:

Considering "Pro Search" is a separate entity and that LLM is another thing, if I toggle off the button for "Pro Search", does it still use (in my case) Claude Opus 3? I'm on paid plan.

[foggy whale]

I've found that it really depends on my use case too. for true searching, Perplexity search is really great. GPT4 seems more accepting of profiles, at least the way i've tried them so far. I am an attorney and i can definitely tell you that when you're feeding data into the threads for analysis and other work Claude wipes the floor with the rest. swapping between models mid-stream is super helpful too

[austere kestrel]

exploratory search is what works best with AI - for sure. (and it's at least what I'd like think to think I mostly use search engines for). But the point isn't about what you or I use them for, it's what the overwhelming number of people using search engines are using them for...

[supple vine]

no I agree with AI searches not reaching the level of Google for the majority of people. Ai search is also not great for local or map searches. But I also think that's just for now. Not sure about what to do with people that google "facebook" or "google" though.

[austere kestrel]

also is the author of the article a member of this server lol #1213263746143883414 message

[eternal quartz]

Yes. You can always see the model used at the bottom of the output

[austere kestrel]

"Not sure about what to do with people that google google' though."
😂 I know right lol

[stable radish]

Alright, I was aware but I wanted to make sure this wasn't just an UI bug. Thanks

[tame current]

Pastebin

[eternal quartz]

Ever since the Bing Chat fiasco I don't trust any journalists to know anything about AI

LLMs are too new and people are still in "freak out over new tech" mode

[tame current]

?

[eternal quartz]

When Bing Chat first came out it had a real personality and could say some wacky stuff

[tame current]

Real

[eternal quartz]

then a journalist freaked out and MS had to lobotomize their chat bot. It was really disappointing to see.

[tame current]

Never used bing chat bc of bing

[eternal quartz]

when it came out it was the best LLM you could use for free

[tame current]

Tried to use it for the gpt4 on mobile but it was only mixed

[eternal quartz]

since it used GPT-4 before OpenAI released it publicly

Now it's one of the worst, sadly

[tame current]

The picture mode is sometimes useful

Though usually Google vision is fine or whatever the name is

[austere kestrel]

they guy is an established tech journo from what I can tell (though on the other hand so was the NYT guy who freaked out..). but yeah I dunno, read the article and draw your conclusions by what it actually says

not all journalists are fools..

[minor shard]

[eternal quartz]

I never said they were. What I will say is that some are fools, because being a fool gets clicks :P

[supple vine]

Here's an old article looking back on Google as it came out.
https://technologizer.com/2009/05/22/how-long-did-it-take-for-the-world-to-identify-google-as-an-altavista-killer/
"Any lessons here? People are smart enough to recognize good stuff when they see it. But they’re not able to predict the unpredictable–and as good as Google was in 1998 and 1999, it A) wasn’t as good as it would later become; and B) hadn’t figured out how to turn high-quality search results into billions of dollars in advertising revenue."

[tame current]

Exaggeration

More of the time misleading

[austere kestrel]

ok I should have said: "not all journalists [writing about AI] are fools"

[eternal quartz]

Fair enough, I was exaggerating there.

[weary crescent]

Is there a way to automatically scroll with the chat on perplexity? it's annoying how it just stops

[tame current]

Just make sure youre all the way at the bottom of the page when you send ur question

[austere kestrel]

I'm prone to exageration myself aha 👍

[eternal quartz]

Google and other traditional search engines are just so much less functional than Perplexity and similar. Just now I was trying to find some examples of using a python library in a certain way. DuckDuckGo and Google found nothing even relevant to what I wanted. I asked Opus in Perplexity and it wrote the example code for me.

and it's a fairly new library so it's not going to have many examples in it's training data

[tame current]

You mostly gotta rely on forums to get your answer

For programming stuff anyways

[eternal quartz]

Google found no forums related to this

[tame current]

Womp

[supple vine]

which is where Ai search excels

[eternal quartz]

unless you mean make a post asking for help? why would I do that when I can just ask perplexity and have the answer in seconds...

[tame current]

It uses the same engine (chromium prob) so it's either gonna base the answer off what it finds or if it's irrelevant just guess

[eternal quartz]

the move to using Discord for everything actually sucks for finding resources. I wish we went back to the days of independent forums

[tame current]

Forums usually take longer to get an answer back

[supple vine]

maybe, but it also can produce the code. Also if I don't have to look at another microsoft support forum or Apple support forum ever again, I'll be happy.

[tame current]

Based off the libraries/code
it's been trained on...

[jovial mantle]

Nowdays I'm using perplexity in tandem with kagi

[minor shard]

Discord AI not working 😦

[tame current]

Curse words are filtered

For whatever reason

[jovial mantle]

it works well. Even though perplexity sometimes hallucinates, for example it doesn't know he will be back next chapter and defeat Sukuna

[minor shard]

oh lol

[tame current]

: (

[jovial mantle]

https://tenor.com/view/jjk-jujutsu-kaisen-shibuya-gojo-satoru-satoru-gojo-gif-1356799353708080752

[stable radish]

Another question:

How does Perplexity intend to address issues of PRO search completely failing to do what it's supposed to do, a search?

Since it's still debiting search credit, it's quite annoying

[tame current]

By fixing the bug

Idk

[stable radish]

🤓☝️

[tame current]

No one can rll answer what they'll do ab it

[lone citrus]

Ture

I don't use the pro search

[tame current]

ture

[native robin]

im new here , can we report it to the team so they hopefully fix it

[halcyon coral]

It's not a bug, please check my reply.

[silent magnet]

You were asleep!

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

[native robin]

got it

[livid mantle]

any suggestion on what prompt we should make today @warm cave.

something useless

[south kindle]

innovative game design that serves to increase a users brain power (memory/retention/planning/etc). either a prompt to start the project, or just a prompt to summarize ideas. GO!

and if you want "useless" the opposite. design that serves to lower a users brain power.

oh wait that was just WoW back in the day wasn't it. lol. JKJK.

[sweet jasper]

Probably something involving memory palaces

Probably the best game design is not to game at all 🙂

[south kindle]

what, life experiences?

where's the reset button, or the youtube "lets play without commentary"?

full of the snark today. sorry fam. stayed up all night watching the baltimore bridge thing from 2AM est until like 6am...

[sweet jasper]

Should get some zzz

[meager sparrow]

I was going to shoot a video in a school…

But the teacher told me not to

[south kindle]

throwing here instead of sharing because it's just an idea template and hasn't been revised/optimized or anything: https://www.perplexity.ai/search/help-me-brainstorm-4AZ0vlZ2TEycM3DlA24BEQ

[south kindle]

threw the idea into wordware (OPUS Insight), and GPT-4 Turbo spit this out. I'd play this.

[half venture]

I hate how perplexity claude opus just randomly seems to forget things

Like even within like 10k context

[south kindle]

examples? that isn't useful feedback for the devs unfortunately.

[weary crescent]

Anybody remember the old logo?

Used to use it all the time for help on summarizing articles and stuff with the chrome extension

[half venture]

Nevermind it had automatically switched to the mistral large model I am dumb 😭

[tame current]

o.o

Hi

[desert shale]

I love Perplexity. I use Pro version since its very beginning 🤣

[feral marlin]

What’s the context limit on opus

[feral marlin]

I’ve noticed this but idk if it’s a super low context limit because it remembers it if I bring it up again

[devout bronze]

i think it’s full-context, but output is 4096 limit

the 4096 is a limit on Anthropic’s side

[devout bronze]

*with the note that large files will get uploaded as attachments, and i don’t know how perplexity handles that

[frank elm]

There is a play button in the top right when you receive in answer within the app does this allow it to dictate it’s answers?

[serene sand]

#💬│general message

[south kindle]

@livid mantle i have another task. a prompt to perform this better. simulate the computer from the tim and eric sketch.

https://tenor.com/view/tayne-oh-shit-okay-paul-rudd-gif-7396985

[frank elm]

Thanks a lot, I really appreciate it.

[livid mantle]

Want a FREE floppy disk coaster? Of course you do! Click here and maybe I'll send you one...

👉 https://discord.com/channels/1047197230748151888/1222288710062575737

[south kindle]

look at you go sir.

i think this meets your quota for 1 daily useless prompt. congrats. haha.

[livid mantle]

lmaoooo so hilarious response

[warm cave]

Haha, I got no creative juices flowing today, just studying for an exam

[minor shard]

I wonder if there is a way to use a discord bot to store conversations in a channel and then have AI summarize it each day.

[lone citrus]

it's possible, but like big channel have a crazy number of texts everyday, i think it only work in some small channel

[half venture]

Claude 3 haiku is impressive sonnett is impressive opus is even more impressive
Anthropic is clearly ahead of openai now. Wonder why openai is so silent

[livid mantle]

There is a impostor @signal hamlet can we ban them from this server?

1222293400854728750

@o.alex_53

he will do malicious acts with the members giving free perplexity pro

inexchange for credit card

@tame current

Even got the spelling wrong

[south kindle]

WTF

[livid mantle]

[half venture]

There is a imposter among us

[south kindle]

how are you aware of that? did he message you?

[livid mantle]

he just added me but my senses were tingling

[south kindle]

What a sad person.

[brittle merlin]

my bad i just had a think out loud moment and thought id share i saw a view articles that the pro vision was his last main selling point and that they are already looking at people to replace him

[signal hamlet]

Thank you for staying alert and reporting this, @livid mantle! Hope no one gets fooled by such users.

[signal hamlet]

And gone 🔨

[green birch]

I got a friend request from that scam account too! Tried to sell me stock opportunities @signal hamlet

[south kindle]

He is in charge of that decision i'm fairly certain. so it's not so much "people replacing him", it's just that hes in his 60s. The man wants to probably spend some time with family for retirement. I've read they have very extensive succession plans already in place, so likely they are looking at it now. but there's no indication he's leaving just yet. with that in mind, not sure how Woz would be on deck if he's the same age. But that would be my pick, since he litterely helped start the company with Steve in the beginning.

[green birch]

At least your imposter is friendly

[south kindle]

we just need to be aware that he may come back with another spam account. realistically...it's highly unlikely that anyone DMing you for an opportunity is legit or has your best interest in mind. just be smart.

People with real opportunity's to share are in shambles though. lol.

[signal hamlet]

was friendly)

[livid mantle]

STOCK Opportunities HAHAHAHA

i think he made prompt to talk like you sir

You are Alex Romanov you will respond in his perspective.

Here are examples of how he talks

[outer raft]

I figured out how to get past often disappointing prompt suggestions from "related" section. I can just ask AI to give suggestions on what to ask
Then I get 10 instead of just 3, and questions seem more detailed

[silent magnet]

Easter eggs collected

You were able to grab 1 Easter eggs.

Collected by

• @outer raft

[weary crescent]

Which image generator works best in your guys' opinion?

[livid mantle]

first of all

I know it has no correlation to ur question but

i like to utilize notebooks in the cloud

and my favorite would be automatic1111

rather stable diffusion

[weary crescent]

like what, google colab?

[livid mantle]

yeah

never use ur own local specs

not even if its so good

[weary crescent]

I was talking about between SDXL, Dalle 3, and playground

[weary crescent]

why not?

[minor shard]

I use my 4090 for Stable Diffusion all the time xD

[weary crescent]

I've used my 6950xt for this, but I forgot which program. It wasnt SD that I remember, could be wrong though.

[livid mantle]

thats why ur pfp is so perfect

its shining when im lookingat it

[south kindle]

my brain misreading chief as chef

https://tenor.com/view/chef-muppets-gif-13657974759252566916

[leaden oracle]

hello

what is the limit of using Claude 3 Opus ?

i have last 20$ and curious between perplexity and claude pro

any advice ? 💙

[half venture]

There is no limit

[south kindle]

the main difference is the context limit via OPUS on perplexity which is lower.

other wise though, you get access to both OPUS plus all of Perplexity and Pro search.

[leaden oracle]

impossible

[half venture]

Why is bro flabbergasted

[south kindle]

I think it gets set to something around 500 per day? maybe more? i've yet to hit a limit

[half venture]

No it's genuinely limitless

[half venture]

That's pro search

Turn it off

[south kindle]

Fair point.

[half venture]

And you will still chat with opus

[half venture]

This sucks so hard though 😭

[south kindle]

I have done that for writing stuff.
https://www.perplexity.ai/search/increasing-intelligence-of-HLUn3nOzSx6Nc5ecNpe5pA
that entire project was with writing mode. ha. I was just saying that in terms of "limits" via having Pro, thats the only limit I've come across.

[half venture]

Are there any devs here

[leaden oracle]

i saw comment on reddit that perplexity allow just 5 messages daily using opus

and claude pro give like 10 - 20 message daily limit if you use it heavily

[leaden oracle]

so are you sure?

[south kindle]

Do you have Pro?

[leaden oracle]

nope

[south kindle]

@leaden oracle ~~He's talking about having Perplexity Pro. scroll up to see the context.~~I didn't realize you were the original person asking. Without pro, you get 5 PRO searchs a day, with no access to claude or other models. with pro, you get access to gpt-4 turbo/claude sonnet/opus, with no limit on the use.

[south kindle]

#⚡│ask-community would be your best bet. they monitor it. or throw something in #👉│feedback-general

[half venture]

Claude opus 200k is being used probably on the api front the issue is the file analysis thing. It isn't updated to support large than 32k context files since you guys use a private 32k context model can you guys update it for 200k

Ohhh

[hallow bridge]

I can’t select Claude opus on the iOS app can I?

[leaden oracle]

yea i know .. i just want know the excat messages limit per day of using OPUS via Perplexity Pro

[south kindle]

I don't want to tag the main person / dev we work with (because it's midnight) but it's alex R.

[half venture]

Yeah I gave it on feedback section

Hopefully he will see it

[south kindle]

Per the above, with Search off, it's unlimted. just realized we covered this last night. let me find the tweets again. lol.

they annouced it was unlimited

[half venture]

What kind of money does Amazon pay perplexity

😭

[leaden oracle]

If that's really true...you've made my month 😮

i hope 💙

[half venture]

Yeah its unlimited have fun

[leaden oracle]

https://tenor.com/view/minato-gif-15543414

[half venture]

(and may God help perplexity profit margins )

[south kindle]

@leaden oracle https://twitter.com/AravSrinivas/status/1769475725965566167

Aravind Srinivas (@AravSrinivas) on X

We have made the number of daily queries on Claude 3 Opus (the best LLM in the market today) for Perplexity Pro users, unlimited! Enjoy!

Twitter · Mar 17, 2024, 9:27 PM

[leaden oracle]

yup

[leaden oracle]

😍 🥳

thanks bro

[weary crescent]

how do these guys manage

[fossil sentinel]

Manage what?

[south kindle]

I'm assuming the reduced context limit is how they are able to do it.

[fossil sentinel]

Helps that they get juicy investments

[weary crescent]

unlimited opus, $200 credit to 100k+ rabbit buyers, etc

just being the best

[fossil sentinel]

Jeff bezos

[weary crescent]

KEKW

my fake emoji lol

no nitro moment

[silent magnet]

Easter eggs collected

You were able to grab 1 Easter eggs.

Collected by

• @weary crescent

[weary crescent]

what does this do

[livid mantle]

to keep the dophamine flowing its for all of us who have adhd

it helps

it keeps us from crashing

[idle hull]

Is Claude currently better than GPT-4? And in what aspects?

[idle hull]

Well, I know that well, although I am mainly a GPT-4 user on Perplexity I will try to try Claude, despite having tried the site a while ago.

TYSM!

[south kindle]

#1219244097383039007 message - you can try your prompts with multiple models to check outputs with this tool.

if you duplicate that, you can just remove the models/sections you don't want, so it's just GPT-4 turbo and the Claude models.

[idle hull]

Ooh, thank you, I'll try it right now, it helps me a lot to remove the doubt about what is best for me and in what cases. Although since the release of GPT-4 in Perplexity I have gained a lot of confidence in the model, because it served me a lot in my daily life, it is not a bad time to open myself to more models.

[south kindle]

https://app.wordware.ai/r/5ea3e441-33e1-492e-a061-3ffa4591802e - went ahead and just did the thing.

use that instead

nope. i broke it hold on.

[idle hull]

Okay, let me know when you can, it's not urgent anyway.

And I regret my English a little, I'm still learning, actually my native language is Spanish.

Luckily I can understand almost everything you tell me and if I don't, there are translators for that, but sometimes I get stuck writing, sorry. 😅

[south kindle]

should be fixed @idle hull

[fossil sentinel]

Is gemini 1.5 unavailable rn for anyone else

Or is it just me

[south kindle]

possibly just you.

[fossil sentinel]

I'm talking about gemini 1.5 not gemini advanced

[south kindle]

you right. looks like its down.

[fossil sentinel]

Unlucky

Thanks for checking

[south kindle]

looks like it's just their AI studio that's down. nothings working on there.

[jolly cliff]

See #📰│server-news message

[jolly cliff]

Correction on what you wrote: Without Pro, you get 5 Pro Searches every 4 hours (so 30 per day, not 5 per day)

[half venture]

Just saw Sam at the local supermarket and he was looking at his phone with a shocked face. A few seconds later he literally collapsed and started shaking violently. After 2 minutes of shaking and screaming, a crowd was formed around him trying to help him. But surprisingly he stopped shaking and screaming on his own after EXACTLY 2 minutes (I counted), got up, picked up his phone and started dialing a number.

"Prepare for release..."

[silk wolf]

Hello 👋

How are you doing?

[sly cloud]

drop it

[deep pasture]

Hi all, cannot generate images. I thought it is possible with pro subscription.

[halcyon coral]

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

[onyx scaffold]

Hi, I am thinking to switch to perplexity pro. Is it better than claud pro sub? I am a software developer. can anyone suggest would be best for me? I want to keep one sub not multiple, as I don't get enough time to play with multiple stuff,

[mellow yoke]

I just upgraded

so I can use the API within make.com...

[undone thorn]

@halcyon coral Hey,
I have pro version, will I get any role?

[halcyon coral]

You already have the role 🙂

[green thicket]

Is Perplexity Pro worth it over ChatGPT Plus? Does the GPT 4 access via API with Perplexity work juts as good?

What about memory performance. Does it have the same capability to retain information throughout a conversation?

Also, is there something comparable to Custom GPTs recently introduced by OpenAI for ChatGPT Plus subscribers?

[jovial mantle]

As a subscriber

Custom GPT = no, you can create libraries of prompts and they affect the subthreads

Memory performance = the API is good, as a general search engine tool, it's very quick, perplexity models however have a shorter context window ( I believe, my experience with GPT is from Poe subscription)

Is it better than GPT plus at what it does? Yes, it also lets you use different models

[fluid mauve]

Judging by some of the questions I have seen over several months, I think this is one of the major points people seem to overlook. With PPLX you get access to multiple models,

[tame current]

Did you make sure your PDF is not write protected? If it is text cannot be extractec.

[torn ravine]

Yes, but most of the time I use only one model though. I mean Opus is clearly the winner. If nothing wrong, I won't switch to any other.
I have both ChatGPT Plus and Pplx Pro. ChatGPT helps me with some office work like Excel or making plots, while I use Pplx for somewhat more serious research. I have to say ChatGPT is totally transparent, while Pplx uses some tricks to improve the cost efficiency, which often causes trouble. That's the tradeoff.

[vapid onyx]

Its incredible to be honest. Opus is hands down the best for generating in my experience

[fluid mauve]

its only a pity its not OpenSource 😦

[tacit heath]

bro

ai getting real

[minor shard]

Spitting unfortunate facts

[vapid onyx]

This is too deep for me

[brittle swan]

@here who uses vercel to deploy web apps with Python / flask? I'm trying to upload some products here and it's giving me an error in building:

Error: Unable to find any supported Python versions.

This has never happened before
using vercel’s own templates 🥶

https://github.com/orgs/vercel/discussions/6287

[austere kestrel]

well apparently 20,000 people disagree with you lol

[south kindle]

do you have an example prompt to share for us to test out?

not that we don't know how to construct a user adherence specific instruction prompt, but I'm moreso curious the prompts you've used that didn't work as expected.

[dawn ether]

Add gemini 1.5

[austere kestrel]

yeah tbf it is good - beats Haiku here and with some other prompts I gave them

prob prefer Haiku's poem (though Gwen's isn't actually that bad, just a clunky)

[fervent crescent]

@rotund sedge check this out

[silent magnet]

Easter eggs collected

You were able to grab 2 Easter eggs.

Collected by

• @austere kestrel
• @halcyon coral

[halcyon coral]

[tame current]

What's the point of the eggs

Imagine tryna create a copy of gpt4

As a llm

[serene sand]

#📰│server-news message

[tame current]

Where's the leaderboard?

[silent magnet]

Perplexity

Local user leaderboard

#1 | eun0800 | 4
#2 | brknclock1215 | 3
#3 | shaolinmonk3434 | 2
#3 | 13376666666666666666666666666669 | 2
#5 | vardude | 1
#5 | archient | 1
#5 | icelavaman | 1
#5 | winterchai_ | 1
#5 | _paradroid | 1
#5 | drewk312 | 1

[tame current]

These things must be rare 💀

[austere kestrel]

I got an AI to work out the pattern

[livid mantle]

So I have quite some eggs, what are they for again?

[halcyon coral]

#📰│server-news message

[warm cave]

Interesting seeing Haiku above GPT-4-0613 on the leaderboard

I have been using Haiku a bit and its good, but idk, seems a but inflated

[agile jay]

Yep, you have to remember that it's the battle arena leaderboard which isn't that accurate.

Since most people in that benchmark are asking pretty simple questions.

So it ends up being a benchmark more on preferred style than how capable the models are.

[warm cave]

Yeah, I don’t know how we are meant to have accurate comparisons, Benchmarks are trash and i guess arena is only so good. I guess the best way is still trying it for yourself on your use cases

[agile jay]

Yep, I normally just test the models on stuff I normally use them for, and compare the results.

[agile jay]

You mean CoT (Chain of Thought)?

[wise granite]

which ai model do you guys use that gives the best accuracy

[south kindle]

that entirely depends on the task you want to do. and if it has access to the internet to find new information / if that's even needed for your usecase.

a lot of us here are still toying with the new claude models pretty heavily.

[wise granite]

yea i just seen those. which combo would you say does best for mathematical, business and accounting type task? I know AI isn't the best with numbers but would you suggest a certain model + focus + pro search combo?

[south kindle]

someone that has actually used it for all of those tasks can chime in. I've mostly used LLMs lately for creative writing/searching for new information and formatting it in a way that makes sense to me.

[west sage]

no codes for a free month?

[south kindle]

I swear we need to sticky this post now. lolol.
https://app.wordware.ai/r/5ea3e441-33e1-492e-a061-3ffa4591802e @wise granite - this tool will let you throw some test prompts at gpt-4 turbo / claude 3 models to see the output for yourself.

https://app.wordware.ai/r/b0f0a2c9-da4f-4524-b662-3584ac0fdbc2 - this one tests more models (gemini, openchat, mixtral, mistral, etc) but otherwise is the same tool. the gpt-4 vs claude one is just a reduced version of this.

it should work without a free account. but if you do make an account, you can duplicate projects and adjust them to your liking.

[grizzled echo]

I recently discovered that the Usage Dashboard of the Perplexity API looks very different to how it looked yesterday and doesnt provide as much information and is very limited in its functionality compared to the dashboard that I saw before. Do you experience the same? How can I change it back? Thanks in advance for your help!

[hazy ingot]

Hello Wondering if anyone can tell me how to see more of my threads--like those dating back weeks? Write now I can only see my last 8 threads and can't figure out how to see more thanks

[tame current]

you can always add your threads to collections to keep them organized like i do 🙂

[vapid onyx]

The best way to manage your threads. Especially if you believe they are useful to you

[south kindle]

i haven't looked in a bit but is there not a "next page" button in the library?

nope. just infinite scroll. most of the ones I come back to were shared, and I have that link saved in a personal discord. long way around the problem though.

[south kindle]

does using the search help? or is this an issue of not specifically remembering the threads from that time period and wanting to go back thru them? also Mobile or PC/web browser for use?

[serene sand]

Opus Chronicles

[fossil sentinel]

https://chat.lmsys.org/

Wait Claude 3 opus taking the lead? 😂

[halcyon coral]

It's a new dashboard provider, you can't change it 🙂

[finite badger]

I have noticed more hallucinations from Experimental compared to Mistral Large but it is definitely blazing fast

[south kindle]

Sam is sweating even more now.

[grizzled echo]

Thank you for your answer! Do you think they will change it back? Because the new one curretnly doesnt support many of the things the old one did and all of my usage data from the past is also gone.

[fossil sentinel]

Gotta build up the gpt 5 hype even more now 😂

[south kindle]

there gonna just bench 5 and get 6 out instead. lolol.

[fossil sentinel]

I expect great things to come

[agile jay]

The way you write the prompt will largely affect the quality for the answer. For Maths/Stem related tasks, ask to always show its working out. There is also a wolfram alpha mode, if you click focus in the search bar.

[ripe marsh]

Hey, has anyone using the android app had problems to subscribe?

(I tested with 2 different credit cards in 2 different cellphones

[agile jay]

That might be a google services issue.

I recommend trying to sub using the browser.

[lime lodge]

I have a questions about perplexity model. I mean exactly Claude 3 Opus
Is this perplexity model modified or original from Claude?
I wonder if the model I use on perplexity is 100% the same model I use on claude.ai

[agile jay]

Also helps, sine the devs get a better cut...

[agile jay]

Yes and No. It's Claudes API they are using, but the context length might be different.

But the search results makes perplexity more accurate IMO.

Also I assume that Claude has a different System Message for their Client.

[lime lodge]

What are the main differences in practice and which are in favor of claude/perplexity?

[foggy whale]

I currently still have claude pro and perplexity pro. still working out when i choose to use one over the other, in part because of those things that Code mentions. I've noted some quirks that I work around when using Perplexity via Claude that i dont think i'd have to do with regular Claude

[agile jay]

I believe that Claude in their client has a longer context length.

But for most tasks, that isn't that relevant.

[lime lodge]

And @foggy whale writed that regular claude is more effective and has fewer 'mistakes'.

[foggy whale]

agreed, if it's a transactional kind of thing then I'll come to Perplexity. If it's a very long running project or conversation then it goes into Claude for me right now.

[foggy whale]

haven't seen that in practice with writing mode, but would say that Claude's not my preferred default model for searching. Flopping between them to find the result that's best for what you need at the moment makes the choice a click instead of getting into a long term relationship

so what may be right for me might not be right for you, play around. it's a good time

[south kindle]

im gonna get downvoted but its "wrote". writed isn't a word. writen maybe?

[lime lodge]

Sorry for my mistake, English is not my first language 😉
I meant 'wrote'

[south kindle]

Sorry for being one of those people. I try not to be. ha. you are doing well friend!

[hoary python]

so whenever I am using, I am getting cut off like I am making a CSS file, is their a way to continue from where it left off?

[lime lodge]

Thanks 😉

[hoary python]

I tried 70B and I tried sonnet and opus

[silent magnet]

Easter eggs collected

You were able to grab 2 Easter eggs.

Collected by

• @agile jay
• @south kindle

[agile jay]

Also most people are not using 32_000 context, lol. That's almost hald the first harry potter book

[lime lodge]

I have one more question, what does this mean in practice? I have Perplexity Premium, of course

[agile jay]

1. More sources used
2. Asks questions to refine the search

[south kindle]

the best way to know is to test it yourself honestly. that adds another layer that first understands your question first and asks for clarification.

[agile jay]

Might also do a life crawl for more upto date sources, but not sure.

[lime lodge]

Sources, you mean internet search engine or other model's ?

[agile jay]

What do you mean by Internet search engine?

[lime lodge]

If 'PRO' is checked, the model uses more sources.
What are the sources in this case?
The source is a different AI model or ability to search the Internet.

[south kindle]

search the internet

only one model in use at a time but you can have it "rewrite" outputs with another model mid chat/thread.

[lime lodge]

thank you so much 😉

[south kindle]

?

what was that

lol.

I saw a man in a cosplay suit just now

[agile jay]

A troll

[south kindle]

was that the dva outfit?

[livid mantle]

i got something to show u in the dms, i wanted ur opinion on something

[agile jay]

@livid mantle is a troll, so better not...

Also don't trust and leak screenshots from him, since he edits the html of the sites, lol.

[south kindle]

think of it like doing a deep dive on google, but instead of actually going thru pages of links, perplexity will review them and output the relevant info for you.

[minor shard]

Happy Hump Day 😄 (Wenesday or may not if ur timezone is ahead

[serene sand]

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

[west sage]

Question, how many tokens can be used with CLAUDE 3 OPUS on perplexity?

[halcyon coral]

#💬│general message

[solid whale]

Lets say that I have a long session in perplexity... is there an easy way to extract the prompts and output in JSON format?

[solid whale]

I recently measured an upper bound of 65k, where it had forgot information that had been ~65k tokens in the past. Someone on another discord server claimed to have measured a 59k least upper bound. It is not clear if that is correct, since there are reportedly remarks that long session context is converted into RAG format, although I have not seen any evidence of recall occurring from that...

[junior apex]

Does Perplexity have an app or is it just web based, could not find anything in the iOS realm?

[agile jay]

Yep, there is both an android and iOS app

https://apps.apple.com/us/app/perplexity-ask-anything/id1668000334

[junior apex]

Sorry wrong question, nothing for the Mac OS and Windows?

[rain owl]

I have no generate image button with pro subscription

[agile jay]

Nope, just the web browser.

And no PWA yet, as far as I know.

[junior apex]

Thanks

[agile jay]

You need to send a prompt first, then it appears at the top right.

[junior apex]

Do you know about the integration with Rabbit r1,anyone?

[rain owl]

thanks -- help needs to be updated

[agile jay]

Maybe ask perplexity

[fossil sentinel]

another gemini 1.5 problem

[junior apex]

New to Discord, is there a way to find the person or person on this board that are from Perplexity?

[fossil sentinel]

Seems like they didn't properly implement rate limits 😂

[agile jay]

Look at the right, green people are perplexity employees, yellow are the server mods.

[junior apex]

I have that normally closed, I'll check it out thank you

[south kindle]

you can also see via #📢│announcements - or rather, i think i'ts just alex that posts the annoucements.

[cinder axle]

hi all, first day here. if there's an existing discussion that answers my question, I'd appreciate a link: "ask follow-up" never seems to know the context of my first question+answer. It acts like i'm starting a new thread. is this typical?

[pearl sky]

guys i have perplexity pro, i don't understand how to make him generate images

[halcyon coral]

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

[south kindle]

someone is gonna have to thorw a tutorial video toghether of image generation . have the perplexity bot post it when every anyone asks about image generation. ha.

[pearl sky]

is there any tutorial i can find that can help me understand better how to use perplexity efficiently?

[halcyon coral]

There are a lot on YT

[south kindle]

Lots of youtube videos of AI content creators showcases it from the past year. that's probably the most direct way to see with a use case in mind.

[pearl sky]

thanks

over all would you say it is better than chat gpt 4? i mainly have to use it to help me with my studies liek to write accademic papers, so maybe you guys know what can be better in my case, i currently have both but maybe it has no sense since perplexity has gpt 4 or am i missing something?

[south kindle]

it replaces crawling thru google result pages for what your looking for, and then grabs those sources and content and breaks it down for you. the entire platform is built on the idea of "exploring curiosty". so part of the value is being able to ask good questions? if that makes sense? you get out of it what you put into it essentially.

Perplexity might honestly handle that task better. that's what it's built for. research.

[pearl sky]

yeah totally

i know i am the real issue not the bot 😅

[south kindle]

with only perplexity pro, you wont have access to ChatGPT's other features like the GPT store and what not. but if you have pro, you'll have access to GPT-4 turbo, latest claude models/plus perplexitys own model that is used by default.

[pearl sky]

i heard that claude 3 is the best at the moment is that right?

[south kindle]

leaderboard says so as of like a few hours ago

lol.

[pearl sky]

do you know what actually changes from these models like why should i choose one instead of another?

[south kindle]

yes and no. they are trained differently. they are black boxes though that we as users are just sending prompts to.

https://huggingface.co/spaces/lmsys/chatbot-arena-leaderboard - lets go for posting this a 3rd time today.

there's a tool posted to community projects that lets you send 1 prompt to multiple models so you can compare and contrast. not about to post it again and spam the channel. but its the wordware links.

[pearl sky]

okay thanks

how limited is the usage of claude 3 on perplexity?

[south kindle]

trial and error is how we all got to the current point. there aren't really shortcuts though.

not limited anymore as of a few weeks ago. or whenever that was..

i gotta eat dinner. i keep replying lol. someonme else can pick up from here. im starving people and my hyperfocus keeps sending me down rabbit holes lolol.

[pearl sky]

perplexity seems too good to be true with all these bots 😅

[silent magnet]

You were asleep!

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

[pearl sky]

no worries thanks a lot i am going to do my own research from now on

[serene sand]

https://www.perplexity.ai/hub/faq

deep dive ➤ #💬│general message

[pearl sky]

thx mate

[signal hamlet]

🐰 #👉│feedback-general message

[torn zinc]

When using voice in perplexity is it normal for it not to send the prompt until you physically tap the submit arrow button?

[south kindle]

normal

[torn zinc]

Ahhh damn ok thanks

[blissful kestrel]

.

[brittle merlin]

he wants nothing to do with apple in that regard when he walked away from it back in the day he hasnt worked for apple for so long and sold most of his stock as well i dont think they would be try for that

[foggy whale]

Apple's got serious succession planning in place at the executive level because of obvious historical reasons. The thing about tim cook is he's so damn good with the supply chain

[brittle merlin]

For real, he's been a huge part of apple i read some articles that we have been using Tim Cook's products and the last Steve job product was the iPhone 4s I believe

makes me wanna buy the 4s just so i have one last thing by him lol

[half venture]

Just tried this 7b model it's very impressive for its tiny size

Although it's not better than mixtral or claude 2 as the rankings might make it seem

It's still impressive nonetheless

Openai is not that far ahead of the competition it seems

[livid mantle]

is it just me or the ai model performs better when in Writing mode via prpelrpelpxplxpxlx?

[half venture]

Yes I do think the ai performs better when writing in perpleperpeprpeprlex due to less garbage from the internet same with bing copilot.

[livid mantle]

yeah thats what I thought

[half venture]

Turn search off take the based pill that internet sucks now

[half venture]

Wait claude 3 has 100 billion. Para

Or is it hallucinating

[half venture]

We need a small agent searching through the net and finding things

Agents can't come soon enough

[foggy whale]

Depends on how you read the history of the product developemnts. Steve passed away around when the 4S was released but had been fighting cancer hard for years and years. Tim had been running operations before, so when he got the nod for intierim ceo while steve was away people scratched their heads thinking the Phil Shiller or Jony Ive would be the best choice. Turns out the boring guy with the spreadsheets who knows how to hire the right people's worked out better than anyone could have thought. The mix we got was probably the right one. Ive's desin ideas forgot about the fact that yes, depsite how garish it looks an HDMI plug is handy and a keyboard should be thick enough to not break.

[sly cloud]

You cant ask the llm how many parameters it has lmao

[eternal quartz]

Yeah, that's like asking a human how many brain cells they have.

[austere kestrel]

surely not lol. Claude 2 was estimated to have like 200bn iirc

I'd have thought Opus would be closer to something like 1.5-2t parameters

[austere kestrel]

That answer seems a bit dubious tbh
https://docs.anthropic.com/claude/docs/glossary#rag-retrieval-augmented-generation

[honest basin]

definitely

depends if they found anything new or just trained the same model a bunch as experts for MoE like gpt 4

[half venture]

Does anyone know about any prompts that creates newer prompt. Kind of like a prompt engineer but it's the prompt that's the engineer

[austere kestrel]

Haven't used it, but there's this - looks useful / interesting https://docs.anthropic.com/claude/docs/helper-metaprompt-experimental

https://anthropic.com/metaprompt-notebook/

[oblique ore]

woem

[south kindle]

Something I've been doing - just asking Claude / GPT-4 to review my prompt and then optimize it. i'm sure there are more official elegant ways of doing it.

[toxic ibex]

guys is it just me or the academic focus mode is not working on the perplexity app or website it tries to do a search but ends up getting 0 sources p.s im a premium user but dont have access to the premium channels also here is the perplexity thread https://www.perplexity.ai/search/I-am-a-XNNmFl83Rry_HEhBb3k8Tw

same thing ive tried with a different focus mode and it seems to be working like a charm

https://www.perplexity.ai/search/I-am-a-XcDpUIl8Sc.p5XDANbU.Dw

[sweet jasper]

Its not just you

Its not very good

Use focus all and ell it to search arxiv

[toxic ibex]

hmm okay

any ideas when the academic is gonna be back up tho

there seems to be some sort of outage from semantic since yesterday

[vague thistle]

G'day

I'm kind of lost here, the documentation says I can generate image prompts but on the website I can't seem to find that feature. Even though I'm a pro user.

[toxic ibex]

there will be a generate image button on the corner

[serene sand]

https://discord.com/channels/1047197230748151888/1194794305362071552
#⚡│ask-community message

[tight wedge]

DBRX coming to Perplexity?Sweet! 💪🏻

[fluid mauve]

what is DBRX

[tight wedge]

It is new open source model by Databricks, which is FAST and beating most other open source models. Bit large tho.
Source: https://www.databricks.com/blog/introducing-dbrx-new-state-art-open-llm

[junior apex]

How do I know Copilot is turned on in Perplexity I have the Pro version thru the Rabbit r1 purchase?

[fluid mauve]

If you are using teh web interface you just flick the pro button to the right.#

no idea in teh app

[half venture]

Meh too large to run locally

Does anyone even make any profits running these huge trillion parameters models ?

[tight wedge]

Nvidia?

[half venture]

Except them

Anyone else ?

[tame current]

it is the same in the app

[whole lark]

Here is a table listing the HTML tags that appear in the attached file, along with their English descriptions:

| Tag | Description | | --- | --- | | <Task> | Indicates the start and end of a task description | | <Inputs> | Indicates the start and end of the inputs for a task | | <Instructions> | Indicates the start and end of the instructions for completing a task | | <example> | Indicates the start and end of an example demonstrating the task | | <functions> | Indicates the start and end of a list of functions available to complete the task | | <function> | Indicates the start and end of details about a specific function | | <function_name> | Contains the name of a function | | <function_description> | Contains a description of what a function does | | <required_argument> | Describes a required argument that needs to be passed to a function | | <returns> | Describes what a function returns | | <raises> | Describes the exceptions a function may raise | | <example_call> | Shows an example of how to call a function | | <question> | Contains a question that needs to be answered | | <scratchpad> | Indicates a section where the assistant can work out its thought process | | <function_call> | Indicates where the assistant should call a specific function | | <function_result> | Contains the result returned from calling a function | | <error> | Contains any error message returned when calling a function | | <answer> | Indicates where the assistant should provide its final answer to the question |
The tags are used to structure task descriptions, inputs, instructions and examples for an AI assistant to follow in order to complete various tasks. The assistant is able to call certain functions to get additional information to help it answer questions.

[green flame]

Hi everyone. Not sure if it was written anywhere but if I’m a pro user, how many messages can I send to heavy model ? (Opus, gpt 4,etc)
Thanks 🙌

[serene sand]

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

[half venture]

Quick question is it only me or

Is every ai startup now jumping to the ai AGENT trend

I have yet to see a proper agent in action

[whole lark]

https://www.perplexity.ai/search/AI-Agent-Trend-dVIcjEvlRUC1uhmABi_Sew

It sounds like a growing market.

[half venture]

I am sorry but it feels like such a grift. There is yet to be a good agent

[silent magnet]

You were asleep!

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

[undone thorn]

just saw

[halcyon coral]

I don't quite understand...how can I help you?

[undone thorn]

*saw

Thanks for help

[tranquil orchid]

Hey @undone thorn!

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.

[granite quartz]

What happened to citations and links in academic focus?

[half venture]

NOOO

Godamnit

[minor shard]

U missed it by 1 min 😦

[finite rock]

Can I still preorder the Rabbit r1 and get a year of perplexity pro?

[wraith crow]

the Rabbit R1 crossed 100K sales, so not anymore.

[tender terrace]

What is the best model to use on perpexity pro? New guy here

or combination

[minor shard]

Claude Opus or Experimental (for speed)

[livid mantle]

nobody asked but what i like about perplexity experimental model is that the safeguard is very minimal compared to all the others gpt-4, claude 3 opus, sonnet, malastral. when you need to input something controversial or explicit you know.

[sleek tapir]

use expiremental for smut

noted

[oblique ore]

tbf personally, i've asked a few probing questions around israel-palestine to gpt-4 and it answered completely fine

[south kindle]

Progress is linear. Agents wont improve unless they are being developed and worked on. It's a natural progression here if you ask me. the entire AI space is in a place of "rennasance", every startup trying their hand at the AI boom.

[half venture]

I knowwwww. But all the big companies that should BE doing something about agents are just loafing around

[south kindle]

True true.

[supple hare]

Wow, Opus became lightning fast.. not sure why, but I hope it stays like that ;d

[vast kelp]

not opus

what happen

[supple hare]

well, it says "Opus" and is very fast

not sure how to determine

[vast kelp]

not opus now

[foggy whale]

context going into that can change how sensitive the content filters are. I ran into that when using Claude to summarize some of the epstein documents that were released a couple months ago. Eventually i figured out that working around the content filters just made that a bad set of public legal docs to work with for researching

[tame current]

noticed this but output quality seems to have deteriorated

doesn't feel like native API opus

reminds me of when bing AI was completely murdered, faster does not equal better because now the model feels much worse

[supple hare]

yeah, there's no way it's that fast...it's probably sonnet... but it says Opus in the end. Probably the guys are doing something

[vast kelp]

Has it stopped being the opus model now? Or has the output quality declined? What exactly is going on?

[tame current]

yeah it says its opus but it definitely doesn't feel like it

and also it seems experimental is now sonar

maybe they messed up and sonar is responding for c3o requests? who knows

[weary crescent]

you guys hear about sora?

got released

https://trysora.io/

[tame current]

that's not sora

💀

[weary crescent]

yes it is wym

gotta try it

its crazy

[tame current]

if it was sora it would be on openai.com 🙄

also rick astley is a hottie

[tame current]

oh boy

[south kindle]

https://tenor.com/view/older-meme-checks-out-gif-14849207

IT's been YEARS since i've waltzed directly into Mr. Astley. well done.

https://www.anthropic.com/news/prompting-long-context - was reading through this and realized they utlize an internal "scratchpad" for the model to use for thoughts/notes. Has anyone tried using this technique? ill try and find more direct examples of how they use it.

[half venture]

What happened to claude opus ?

It feels so bad now

[tame current]

yeah

[south kindle]

this is obvious to most of us, but i like seeing things explained clearly. although - i just want a guide on how to properly use the scratchpad function. pulled from their KB: https://docs.anthropic.com/claude/docs/let-claude-think
(OG paper I was reading about the scratchpad function: https://arxiv.org/html/2401.05566v3)

[half venture]

I am convinced claude opus is now just sonnett

In perplexity

What is going on

[tame current]

doesnt even feel like sonnet

feels like gpt-3.5 😭

or maybe haiku

[half venture]

Yeah more like haiku

[eternal quartz]

https://status.anthropic.com/

known issue

[minor shard]

I think the Easter bunny is late on the egg delivery 👀

Maybe it’s busy fixing Claude 🤣

Pretty interesting open source option: https://www.youtube.com/watch?v=GanTUWLUUWQ

i don't think its a replacement but thats my opinion

[south kindle]

https://github.com/developersdigest/llm-answer-engine. I may try this out at some point today/this weekend just for the sake of testing it out.

[serene sand]

https://x.com/perplexity_ai/status/1773418423726305529?s=20

Perplexity (@perplexity_ai) on X

DBRX, the state-of-the-art open LLM from @databricks is now available on Perplexity Labs. Outperforming GPT-3.5 and competitive with Gemini 1.0 Pro, DBRX excels at math and coding tasks and sets new benchmarks for open models. Try it free on https://t.co/2yYQB9jln7

Twitter · Mar 28, 2024, 6:34 PM

[minor shard]

Very interesting! Lol is Grok going to be canceled now since DBRX is better?

Love that its in labs tho

fast too 🙂

[silent magnet]

Easter eggs collected

You were able to grab 1 Easter eggs.

Collected by

• @minor shard

[halcyon coral]

switched the channel for two seconds...

[lone citrus]

Tried, love it

[minor shard]

Is this legit from OpenAI?

Lmao

[weary crescent]

Kinda insane, I thought it would be a lot later that they released it

[minor shard]

u guys are mean lol 😢

[weary crescent]

You have to do the tutorial first

It's a YouTube video

[minor shard]

Its kinda great that I don't even have to set up these open source models to try anymore cause PPLX does it for me for free lol

[south kindle]

https://tenor.com/view/rickroll-meme-internet-never-gonna-gif-26474110

[empty siren]

hi when is grok coming out

im cheesed

[minor shard]

I don't think there has been any update yet on an ETA

[dull narwhal]

@halcyon coral @signal hamlet What should we do to claim the 6 month pro with copy.ai What should we do?

[halcyon coral]

You will certainly receive an e-mail about this 😉

[dull narwhal]

What do you mean?

[minor shard]

Is it only for existing memberships with Copy.Ai or new ones also? 🙂

[halcyon coral]

You will soon receive an email if you are a copy.ai user 🙂

[livid mantle]

grok fun mode

[distant moon]

i didnt get any mail from copy.ai

do we need to be a pro user?

[halcyon coral]

For all new and existing subscribers

The emails will be sent shortly, please be patient 🙏🏻

[dull narwhal]

Subscribers you mean paid users? Because copy.ai offers also a free plan for individuals

But with less words

[halcyon coral]

Yes, paid users.

[dull narwhal]

OK I want to make another question about Perplexity. If my current plans expires before I renew, will I lose all my current pro threads while I revert back to the free plan?

[halcyon coral]

No, the threads will remain 🙂

[raw elm]

My antivirus (Avast One) blocks https://labs.perplexity.ai/

[halcyon coral]

Yes, thank you, avast is informed, they'll probably whitelist pplx in the next 48h. Please disable Avast in the meantime.

[minor shard]

@halcyon coral Thank you for everything you do

Lol 2024 already looking like the year of Models https://www.ai21.com/blog/announcing-jamba

[halcyon coral]

I'm slowly losing track...too many models

[teal tulip]

Can the main perplexity labs domain not be added as an exception in Avast’s settings instead of disabling the entire antivirus or the web guard feature entirely?

[halcyon coral]

That's also possible.

[arctic dune]

huggingface keeps it all in one place (ish)

[livid mantle]

all the new llm models are chatgpt 3.5 level

[rotund terrace]

has anyone tried to save the output of direclty in google docs ? or is there anyway like in gemini you can export into google docs as excel or docx

[turbid iris]

Hello, I am a CopyAI user. How can I get the free 6 months of Perplexity Pro?

[minor shard]

#💬│general message

You will get an email when it is sent out 🙂

[hallow linden]

I actually find sonnet to be the best combo of speed and accuracy, rather than opus or GPT4-turbo

[minor shard]

So many people trying to copy PPLX. https://www.reddit.com/r/OpenAI/comments/1bq5aw1/working_on_opensource_alternative_to_perplexityai/

I don't see it being as capable as PPLX at the cheap cost as well.

[agile jay]

Yep, not even hiding it, lol

Even uses the exact same sources icon etc...

He needs to learn to type faster though

[minor shard]

Yeah they said they custom made them. I guess it can be fun to try to code a copy.

[agile jay]

Yep, not that hard to copy the UI with tailwind.

But if you can only copy something, then it's not worth much.

And the sidebar doesn't even match the rest of the UI lol

Feels like multiple apps squished together

[minor shard]

Yeah its definitely not ready haha. They quickly built it from what it looks like. People definitely should not assume though that this can easily replace what PPLX is doing

[agile jay]

Yep, to me it just looks like they made the crud endpoints for making posts etc, wrapped GPT4Turbo api and then encased them in perplexity and and another apps sidebar.

No idea what indexer etc they are using, since they don't give any further details.

Okay, the attach files doesn't even work in the live preview...

[minor shard]

yeah definitely not ready

[agile jay]

I'm looking forward to more agent based models though.

So they can carry out more complex tasks.

[snow briar]

what AI model is good for summarizing a lecture and formatting it into point form for notes?

[minor shard]

Probably Sonnet or Opus. 🙂

[honest basin]

why not do something cool and original

if I want something like perplexity I'll use perplexity

[minor shard]

I feel the same way 🙂

[winged bough]

/imagine a car

[south kindle]

sorry not sorry. That guy sounds like a tool. also it's supposed to be sound not should. obviously edited now.

only went off on him because of his previous comment thread

stepping away now. lol.

little too heated over nothing.

[minor shard]

Yeah idk why he started making comments like that

[south kindle]

I just don't get that level of disrespect. More and more people just forego any kind of compassion or kindness. Like respect is some alien concept... Is it that hard to NOT be a insufferable person with a terrible personality? lol. <end rant>

[minor shard]

lmao xD people just have a hard time just accepting things. They tend to blame things that have nothing to do with it because they have no argument.

Cool kid syndrome

[fossil sentinel]

Blud is a yapper god

You'd think he'd be making bank if he can do stuff so easily

[livid mantle]

no way jose, mr. alex romanov would say

[orchid magnet]

So with perplexity we get copyao too?

[south kindle]

otherway around.

[south kindle]

that was my take. go do something productive instead of just being an assh@t for some kind of spit. his attitude is insufferable. skills or not. Still not really confirmed what he has because he hasn't posted to github yet. his personality alone leads me to think hes actually really good at using other peoples stuff and mashing other existing things together. its a skill but don't act like your some deity. lol.

[livid mantle]

I asked it twice.

idk its a joke

its on fun mode:

[narrow spoke]

any way of getting in text citations?

[minor shard]

Have you tried requesting it as part of your prompt?

[foggy whale]

@south kindle so check out Notion, but the second platform i saw it in was more of a help desk KB type system. The interesting thing that I've found in working with Notion is that just by pumping your KB full of that marginal at best documentation from the vendors and then asking your questions via the langage model that it kicks out substantially better kinds of working documenttation like checklists, etc. It also eliminates the need to curate them

[tranquil orchid]

Hey @foggy whale!

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.

[foggy whale]

you dont' need to curate the checklist because you can create it on the fly, paste it into the Notion page where you're doing the work, check off the boxes, then you've done your work...according to the documentation...and you documented the work with really not a heck of a lot more effort than you would have done just winging it

[south kindle]

I've used Notion (barely, but I've copy/pasted stuff into it to make nice looking documents) so that's what I was curious about the helpdesk aspect of the 2nd one.

[foggy whale]

yeah it was just the same kind of implementation of the RAG model that you can see in Notion but for a..you know...boring kind of enterprise help desk KB system

[south kindle]

verrrry curious to see how this scratchpad prompt would work in that context, with access to dedicated KBs/documents. #1223058316662538331 message

Still boring, but having done the work, having it do all the annoying low hanging fruit clean up tasks sounds like a dream.

[foggy whale]

this is no bs, a few weeks ago i had to troubleshoot a corner of some software that i'd never gone before. I dumped the documentation into a claude thread and it talked me through where to click as if i was talking to a tier 3 help desk person via chat

to the degree that i was like "i dont see that" and the response was "oh, they must have configured that thing this way then, it's over here" and there it was

[south kindle]

Wait

that's almost worrying as someone trying to transition away from doing that style of work. do I need to move faster here....

[foggy whale]

oh...you haven't had the exestential crisis yet?

[south kindle]

oh i have i just didn't think it was this soon.

https://tenor.com/view/lost-go-back-jack-we-have-to-go-back-gif-5469258

lol

[foggy whale]

yeah there it is. wait until you see it returned in a prompt to you

[orchid magnet]

So if I get copy ai, I also get perplexity, which means I also get chat gpt, and opus

[orchid magnet]

Should I get grok? I have been thinking about it

[south kindle]

just going off what they've posted here. They want paid copy.ai users to have access to the PPLX stuff, so they get the free Perplexity Pro. It wasn't for us specifically. ha.

[orchid magnet]

Might sign up for that then ngl.

[livid mantle]

I don't endorse it, for some reason the persona looks like its worshipping gpt-4, maybe sam altman and elon musk have a deal idk.

stick with claude, gemini, perplexity

[silent magnet]

Easter eggs collected

You were able to grab 3 Easter eggs.

Collected by

• @foggy whale
• @minor shard
• @livid mantle

[orchid magnet]

I have gpt plus and perplexity

[south kindle]

You should try to recreate that in a little video. Would be really neat to see that process in a real world situation.

[orchid magnet]

$40 a month in ai tools 💀

[south kindle]

really depends on your usecase, but it seems wild to me to being paying upwards of 60 or more a month. not that it isn't powerful and useful, but it's kinda like someone subbing to every video service when they only really use 1 or 2. Savings are nice to have sir. you don't know when you may need that extra $20 or whatever. <end preaching sorrry>

[orchid magnet]

From my experiences, perplexity doesn’t have the gpt cap like gpt plus does?

[foggy whale]

I'm trying to find the thread in my claude history

[orchid magnet]

Ik ik, I love both of them but the custom gpts are so nice to have

[south kindle]

I do miss the few I worked on for like weeks. but so it goes. I'm enjoying other stuff. they'll probably see my money again when the next big thing comes out. or if SORA is that good. or GPT-5 or Qstar or whatever.

Don't pigeon hole yourself with a tool. use many. but also don't have too many that you'll never really learn to utilize them fully. connecting those later dots with a project and knowing exactly what tool fits where takes using it and pushing it to its limits. this is advice i wish reflected my Steam library. i'd say i've only played 1/3 of what I've purchased. lol.

[orchid magnet]

Yeah, perplexity is the best of all of the ai tools. Gonna be a sad day coming here

Canceling gpt plus

[livid mantle]

got is a good start for people who only found chatgpt to get the grasp and idea on what they are missing on

chatglt*

chTglt

chatgpt

[south kindle]

As a package it's very complete (images/voice/gpts/plugins) and it all works very well. it's just not more than any of that at any point in time.

[smoky oriole]

I can't wait for Sora to drop

It's going to be wild

[south kindle]

Youtube in shambles.

[livid mantle]

youtube will be saturated with garbage

[south kindle]

Ha

you can amazon me a coke sir.

[livid mantle]

which garbage will stand out

[smoky oriole]

That's how I felt when xbox added in the feature to record and upload game play

I'm sure it'll work out

I can't wait to start making content again

[south kindle]

They just added it after PCs had it. it just spread something that already existed. the worry for me is that has extended to just people holding a phone in selfie mode and putting out 100 of hours of content a month. accesible leads to over saturation.

[livid mantle]

everyone going to be doing the same thing goodluck

[smoky oriole]

🫡

[livid mantle]

im not trying to end the quota of ur hopes, im just telling u the truth but i believe that if u persevere and become consistent. you can surpass many

[smoky oriole]

Don't worry I was already aware of what you brought up

I appreciate the input though

Time will tell if I will be able to stand out

If I finish my MBA that'll add some authority to my name

[livid mantle]

[idle hull]

How can I get 6 months of Pro from Copy.ai?

[idle hull]

I also use it to communicate with a colleague.

[smoky oriole]

What's your go to use case for AI?

[livid mantle]

Gemini Advanced Ultra 1.0

I am subscribed to chatgpt plus, copilot bing gpt 4 turbo, claude 3 opus and grok but i wont be renewing them for next cycle.

[smoky oriole]

Damnn

I'm subscribed to Claude 3 and Perplexity

God bless the Rabbit R1

[jagged pike]

why? theres any diference than just use opus model at perplexity?

[smoky oriole]

I subscribed to claude 3 first

After it's up I won't resubscribe

Isn't the context window different though?

I'm also Canadian so it's really my American friend who is subscribed

[jagged pike]

claude limit files to 10mb, a shame. Lovely, perplexity loads biger files (20 or 25mb i think)

[smoky oriole]

I mean input in terms of tokens

That is pretty dope ^

I did see that difference

Either way I'll only be subscribed to Perplexity going forward.

My only complaint is that Perplexity does hallucinate a decent amount if I have it write papers

If I use it how it's intended to be used it's fine though

[weary crescent]

is there a way to get pplx to read back its response?

[chilly fox]

Has this been shared today?

https://x.ai/blog/grok-1.5

[minor shard]

not open source 😦

As much as I love AI I won't pay for Grok

[chilly fox]

Apparently Elon Musk said he's going to expand access to all premium members so I think i might try it for a month. It would be 3 bucks so I wont be losing much by trying it out

I definitely won't be keeping it though

[minor shard]

That makes sense

[warm cave]

So I'm guessing we will see Grok become open-source right before each update, ultimately a positive for open-source, tho I wish he would open-source as they advance,, but still better than (Closed) OpenAI

[late bay]

would be nice if we can test Grok locally, say via Ollama and some consumer hardware (say 128 RAM and 12gb VRAM).

[warm cave]

Yeah, would be nice

I want a good model to use with open interpreter, something more consistent than most open llms

I don't know know much about quantizing, but I was hoping that something like that could make grok good for OI

And would run locally

[late bay]

QQ: So I just subbed to PX. Can I use it like I would ChatGPT+ ? like ask it to write code and stuff? My understanding is that PX is mainly for advanced web search

[minor shard]

Yes in writing focus with pro turned off you can do coding. Writing focus is like a chatbot.

[tame current]

"Excels in math"

[silent magnet]

Easter eggs collected

You were able to grab 1 Easter eggs.

Collected by

• @keen solstice

[south kindle]

Anyone else with Spectrum in the US been down for a while?

[feral marlin]

tbf models are non deterministic and struggle in things such as math where there's only one correct answer

[south kindle]

Looks like a big one with that many reports in such a short time

[feral marlin]

its not actually doing the math its just predicting the text you'd want to hear

[tame current]

So no "excels en math" as the notification said

[feral marlin]

advertising for smaller models usually kinda misleading tbf

ai advertising in general

is odd

only like recent thing that ive seen that was actually as advertised was claude opus

[south kindle]

Try the scratch pad prompt I posted and see if it can reason thru the math before the output.

At least in last like 6 months or so.

#1223058316662538331 message

We could adjust it to say “you excel at working through math problems to find the solution”

My internet is out though so I cant really test anything that’s not on my phone.

[feral marlin]

i think opus usually does pretty good with math i use it to help with calculus stuff

i'll try that out if it ever struggles with any particular questions though

the main models that struggle with math are the ones that are trying to compete with like gpt 3.5

[south kindle]

For research and content creation it’s been a huge improvement in my limited testing so far.

[south kindle]

And they would strive to be on the same level of something that struggles with math. Ha.

[feral marlin]

have you tried it for programming at all, opus does pretty solid with programming usually no matter what but maybe since it works everything out on the scratchpad it'd have more elegant final answers

[south kindle]

I know they trained it using that method specifically for CoT but not sure if it uses it by default. In the training I saw it would state “a hidden scratchpad for the model to use for notes and reasoning”

So without seeing it physically, zero idea. I mean you can take the same prompt with/without the added custom prompt I have and compare I guess. I’ll have to try it now. Ha.

Not now but when spectrum deals with an outage across half the country. Ha.

[feral marlin]

pain

[hard wind]

hello, can i generate image from perplexity? if i can, how ?

[halcyon coral]

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

[brittle merlin]

what is labs vs perplexity?

[halcyon coral]

#👑│pro-lounge message
Perplexity is the main product

[brittle merlin]

ahh and the lab is just a playground? DBRX will it come to perplexity?

[halcyon coral]

Yes it's a playground. There is no guarantee that labs models will come to the main product, only if the tests are good.

[brittle merlin]

ahh understandable

[devout cargo]

Hello Have you noticed improvements in the default Perplexity model? Highlight the text, give longer answers, better structured text.

[serene sand]

#💗│sharing message

[sinful smelt]

Having carried out my own tests for Claude Pro (via Perplexity) last week, it was maxing out at 32k words. I've adjusted my expectations and have also signed up for Claude Pro via Anthropic. I use Perplexity for anything below 32k word document scanning and Claude Pro via Anthropic for anything above.
I still believe Perplexity is tremendous value for money and will continue to pay for their service, but I do agree that a little transparency with the context window is warranted.

UPDATE: Having carried out the same test today, Claude Pro (via Perplexity) is now maxing out at 3k words. The test was simply to output the final paragraph of the document I had attached. Now I'm getting a little upset and would like an official announcement from Perplexity as I've been carrying out work on large documents for several days believing at least 32k words were being scanned. I now have to redo days of work!

[halcyon coral]

Please check #👑│pro-lounge message and #👑│pro-lounge message
Furthermore, a "show me the last section" test is not reliable and falsifies the result because RAG is used as mentioned in my message linked above.

[sinful smelt]

Here is what I know; when asking Claude Pro (via Anthropic) to output a full chapter list and the last paragraph of a PDF it correctly outputs 12 chapter headings and the last paragraph of a 70k word PDF. Via Perplexity, It spits out a truncated paragraph at the 3000 word mark and only three chapter headings. I would happily pay three, even four times as much to get the full context window. So at the very least, a little transparency on the limits is warranted here.

[austere kestrel]

you can get Claude to that with Perplexity. All you need to do is circumvent the file upload process. https://www.perplexity.ai/search/I-will-provide-.MaKRen9TQumbfMQF0CVKw

(btw inputting a document manually bit by bit is not an actual workaround to the limitation you – and many others – have been describing... It is just to demonstrate that the model's full context window is there... and that this is whole issue is about how Perplexity handles file uploads...It uses gpt-4-32k to process them...which doesn't make sense but anyway)

[austere kestrel]

In a not entirely unserious way, another way of circumventing this file upload issue would be to use images of the pdf, rather than pdf files (as I think @fast meteor has mentioned before)

that way i'm pretty Claude 3 is using its mulitmodality to just do what is effetively OCR, so you at least bypass the gpt-4-32k thing ha

[halcyon coral]

Yes, I definitely understood that! 👍🏻

I have already linked you to a message where the context limit, 32k, was mentioned and as I said RAG is used to get information from a text (this is more efficient) and not to repeat text parts (Perplexity is not designed for this), for this you can use text editing programs like Word or simply copy text from a PDF 🙂
Please let me know if you have any further questions ☺️

[sinful smelt]

I really appreciate you taking the time to break this down for me. So as long as document is broken down in chunks to less that 32k tokens it will scan the sum total? It's a workaround, but I would be more than happy with that.

[tame current]

Perplexity needs to be more transparent about Opus context limits

numbers being thrown around left and right and it isn't even consistent because the context depends on a gazillion things

[halcyon coral]

#💬│general message 👀

[glad crypt]

hi does pplx-api already support claude 3 haiku?

i dont see any information on the website about it

https://docs.perplexity.ai/docs/model-cards

[halcyon coral]

No, the pplx-api will never support closed-source models.

[glad crypt]

oh man

[halcyon coral]

You can just use the API via Anthropic 😉

[austere kestrel]

No problem! But no unfortunately it’s a bit more complex.. Basically, if you wanted to manually do the chunk-by-chunk thing I referred to (which is not practical), you could in theory give 199k tokens of text for Claude 3 to analyse to using Opus with Perplexity. But.. If a file is uploaded, the to understand is that it is not actually uploaded to Claude (i.e. the file content/tokens are not just not inserted into its context window, like would happen if using claude.ai)...

[glad crypt]

from testing, is mistral-7b-instruct or sonar-small-chat better?

[austere kestrel]

Instead, the file is processed by a different model (GPT-4-32k I suspect), and what actually happens is that when you ask “Claude” a question about that uploaded file, it basically says to the other model “I have this query, give me whatever you can find in the file relevant to it”. Claude then uses those snippets as a “source” (much in the same it does with websites; they're both Retrieval Augmented Generation / RAG systems). So basically, it doesn't have access to the actual 'file', because it was actually uploaded to and processed by another model (which apparently has a context window of ~32k tokens); but Claude can retrieve snippets from the file and use them to try answer your queries about it (it works fine for general stuff, just to be clear, and as has been mentioned, Perplexity is not built to be a long document analyser / parser, it's meant to be good at answering questions by drawing on the internet for additional context)

[austere kestrel]

I find it curious that Haiku is on labs
(without any input restrictions btw - can literally paste 150k token document into the prompt field and get it to anlyse it, for free)

[glad crypt]

haha yea, i also wonder whats in it for them to host it

[austere kestrel]

But also, labs is meant to be their 'API playground'. And all the other models there open source (or pplx in house)

so it's a real outlier ha

[glad crypt]

btw, does anyone know if Perplexity online models are not allowed to scrap google reviews from online? im noticing it omits information hosted by google

also ive realised the generated output by the model is very inconsistent with the exact same input. sometimes it elaborates way more, sometimes the output length is half of what i was expecting

is it common in LLMs? any resource i can use to make it more predictable>

[onyx echo]

same question here

[tame current]

i change my mind about claude 3 opus

supremely better at citing evidence

[stable radish]

The discrepancies in information regarding context, rag architecture... Would be nice to have one official document released by Perplexity regarding all this lol

[onyx echo]

Would the same happen if one split the file into smaller files?

[austere kestrel]

unfortunately yeah more or less afaik - the only way around is to manually input it as text, avoiding any file uploads (including 'pasted'). It's not practical.. You are better off working within the 30k limit, or using one of the third party providers directly, if parsing and analysing long documents - in ways that leverage the full extent of the LLM's context window – is central to your use case (and internet connectivity is more peripheral)

[minor shard]

Would love to see what happens when you write the same paper with got4 turbo and run it through a test

[austere kestrel]

this was meant to a response to @onyx echo

[silent magnet]

Easter eggs collected

You were able to grab 3 Easter eggs.

Collected by

• @austere kestrel
• @hushed inlet
• @minor shard

[tame current]

hm

what does collecting easter eggs even do

[minor shard]

The first 3 people to collect the most get a free month of pplx

#📰│server-news

[tame current]

ohh

yeah I'm not on discord 24/7 so unlikely I'll get it

👀

[minor shard]

Same lol xD trying anyways

[austere kestrel]

One whole month...They're really going all out lol

[minor shard]

lol

Let me code a bot with pplx to click it for me

[half venture]

How do we check the leaderboard for ehgsb

Eggs

[tame current]

btw, I have a question about chatgpt

it says this in chatgpt app, but isn't gpt 3.5 turbo faster ?

[half venture]

It says 3.5 but it's actually turbo

[tame current]

oh but isn't turbo only available in the subscription? did they make it free

Edit: Perplexity says no... maybe they just forgot to update it or don't care to

[leaden hare]

Hey guys, I am pretty excited with the perplexity online model, it looks like it performed much better that GPT in websearch

Any ideas how much does it charge per normal API call for sonar-medium

[cold osprey]

i am tired of secondary followup questions it keeps asking every time i try to search for something. i have tried custom prompt to say it strictly that dont ask me any followup questions and do your best, but that has not worked. i dont think i can keep using it this like this. gets very annoying.

so just to confirm, there is no way to stop this ?

[halcyon coral]

Please check https://docs.perplexity.ai/docs/pricing

[leaden hare]

ok got it it actually conviniently tells the number of tokens

great

so i guess it should be about may be 10 dollars per 1000 requests

[halcyon coral]

I assume you have a pro subscription and are using "Pro Search"?

[cold osprey]

that is right. yes.

[halcyon coral]

It is the purpose of "Pro search" to ask questions, if you do not want this you can simply turn it off.

[cold osprey]

alright. thanks. but all the models settings etc would still apply. right?

[grand sand]

in normal every day stuff opus and sonnet perform about the same

[halcyon coral]

Yes, "Pro search" is just to understand your question better and to dig deeper into your question but it's not always recommended.

[tame current]

already did

[cold osprey]

awesome. thanks,.

[tame current]

like always it doesnt include like any evidence

although i did find an issue with the essay by opus

it was some grammer mistake

[half venture]

There should be a voice transcription feature in the website with whisper

[livid mantle]

Goodmorning Perplexity people, have a good friday.