💬│general

[austere kestrel]

no idea.. i've made like 5 searches today.. it's working fine (just refresh). but yeah i guess maybe it some glitch related to disabling to pro toggle from the main UI or something

[vestal briar]

Same issue with the perplexity pro searches resetting, icon turning to clock, icon becoming disabled but again becoming active after refresh. Is it possible to fix this?

[wise edge]

I guess it's possible that DALL-E 3 is being upgraded to a new version

It's time for DALL-E to release its fourth version

[wide mesa]

Is there a way to have Perplexity exclude certain domains in its searches?

Tiring af to see sites likes Statista or friggin essay writing websites

[wraith crow]

you can use the search parameter -site:example.com

[wide mesa]

thanks!

[wintry pivot]

Compare the Perplexity's Privacy Policy to OpenAI policy and Claud.ai policy. Highlight differences at the start. Evaluate Perplexity's policy against GDPR and PIPEDA. Provide a professional privacy and security opinion on the policy and score it in comparison to the market standards out of 10 at the end.

https://www.perplexity.ai/search/compare-the-perplexity-s-priva-NhVxsgjvSL2m5yr67omJrA

[placid coyote]

not sure if sonnet's fault, but it feels like the quoting of sources is more often broken compared to omni

[jaunty dove]

New gpt 4o-mini model out: https://www.theverge.com/2024/7/18/24200714/openai-new-cheaper-smarter-model-gpt-4o-mini

[south kindle]

anyone know what causes ChatGPT to do these split responses?

GPT-4o mini coming

[cinder leaf]

Hi everyone,
What do you think about adding daily flashcards to review what we've learned today using Perplexity? This would be really great as it would help us reinforce the knowledge we've gained with Perplexity. We could even consider integrating with Anki or Quizlet. What do you think? This version is more concise and structured, making the message easier to understand.

[fast meteor]

https://www.perplexity.ai/search/instructions-always-trigger-se-MFY3ghc6Q3isfWVB8ZE9oQ

[fading moth]

interestante

[eternal saddle]

"Thank you for catching my mistake! I appreciate the feedback to improve my understanding."

Does it really work like this or is it a side effect and I have to report the thread to enable feedback?

[gentle dirge]

https://openai.com/index/gpt-4o-mini-advancing-cost-efficient-intelligence/

[fast meteor]

the latter

[south kindle]

https://arena.lmsys.org/ > Direct Chat as the model is an option. you can also play with the temp/top p/etc.

[harsh verge]

GPT-3.5 has left without a fuss 😅

[tame current]

any plans to replace Claude 3 Haiku as free model in perplexity? GPT-4o mini is cheaper and smarter. Only downside is Claude 3 Haiku has 200,000 token context size, whereas GPT-4o mini has 128,000 token context size.

[south kindle]

Fs in chat

[stark flame]

Why would I be getting a 🕝 icon over the pro selector switch during search if I payed for enterprise pro? Am I mistaken when I read that I had the ability to use unlimited pro searches?

[jaunty dove]

Doesn’t matter in the long run since PPXI is limited to 32k for all models (officially)

[tame current]

So def run 4o mini for better compute

[opaque stump]

now showing Clock icon on Sonnet for me

no tooltip

[wintry pivot]

has anybody tried an accessibility WCAG 2.1 review of perplexity.ai? To adopt in Higher Ed, accessibility and VPAT will be important. Let me know please.

[harsh stag]

i think the devs messed up again, there are lots of request to ... a local server to fetch user settings 🤷

this could be the reason for the clock icon

[sleek vortex]

gpt4 mini looks good

pricing is great too

probably going to replace haiku for me soon

[wheat pike]

this kind of amateur mistake is unacceptable for a billion dollar company

[sleek vortex]

surely the billion dollar company has a staging environment to catch these issues

[sleek vortex]

yeah lmao

do they not have like engineers around the clock?? if they have users in japan they should have workers there too

this the type of stuff that should be fixed in minutes not days

[wheat pike]

i have a feeling someone is pushing directly to prod lmao

[sleek vortex]

probably hired somebody new

who knows

[wheat pike]

with all the random totally avoidable errors that are making it to the main website

[sleek vortex]

Not like they use the 200k anyway

They’ll probably do some internal evaluation then replace it with that likely

Unless they already have millions pumped into haiku capacity on PTU

[neon lark]

如何使用

[sleek vortex]

How to use what?

[keen moth]

perhaps it's a repackaged gpt 3.5 using the 4o model

[harsh verge]

GPT-3.5 remains available via the OpenAI API and can be used in the Playground. For the time being, developers can still access GPT-3.5 Turbo for their projects and experiments via the API. However, GPT-4 Omni Mini is cheaper, so you might want to consider making the switch.

[primal ginkgo]

Which generative ai does perplexity use?

[shadow pike]

anybody having this issue alot?

[fading moth]

Sometimes

Some days it's every other prompt, some days it never happens.

[knotty mist]

does perplexity have their own model? what is it called? sonar?

or is it gemma

[fast meteor]

gpt4o good for math, sonnet 3.5 better? at logic

[fast meteor]

#1240356137497530408 message

[oblique ore]

hello! i was wondering if Pro users have a longer context length when uploading files

[serene sand]

https://www.perplexity.ai/hub/technical-faq/what-is-a-token-and-how-many-tokens-can-perplexity-read-at-once

[oblique ore]

mmmh, this doesn't answer what context length non-pro file uploads have

But thank you

[serene sand]

“By default Perplexity reads at least 4000 tokens per question but it can read many more with file upload. Longer pasted text is converted to a file automatically.
To unlock longer context windows, please subscribe to Perplexity Pro. With a Pro subscription, file uploads can be read with a context window of at least 32000”

[oblique ore]

I see

[sterile falcon]

Claude Opus on Perplexity seems to be very dumbed down. For those of you that have Claude Pro, is there a difference in the responses?

[brittle iris]

#💬│general Lately, I have been experiencing a considerable drop in the quality of responses I receive from Perplexity Pro while using the Claude Sonnet 3.5 model as a pro member.

Similar to when OpenAI went viral and suddenly the quality of responses I got with my plus membership started to decline. This left a lasting impression on me, I realised that this is not what I wanted to get from my Plus membership, so I switched to Perplexity and also You.com. ( For different purposes, of course.)

The responses I am receiving from my Pro searches in the "last 8–9 days" have changed so noticeably that even a child could pick up on the unusual shift. I went the extra mile by fact-checking the previous responses with the new ones, using the same prompts, to ensure consistency and reliability. And the results have been disappointing!

Is this just me, or does anyone else also notice this issue with Pro searches? I sincerely hope Perplexity understands why so many people have switched to them and do not make the same mistakes OpenAI did.

KRs

[trim echo]

does anyone know if there is a free way to have Perplexity AI read the responses to me?

[fast meteor]

try out android / ios app

might be is a pro only feature*

[rich vortex]

What is explorer role? I tried finding a channel that discusses roles but couldn't find one.

Will pro be getting a special color to set us apart from regular users?

Or in the very least, a role icon?

[fast meteor]

ideator, plexer, bug catcher: #📝│server-ideas message
explorer: #💬│general message

[fast meteor]

colors: #📝│server-ideas message

[dark dome]

hello does anyone know if pplx is gonna drop gpt 3.5 and replace it with gpt4o mini in the free version? the price is much less and its a immediate upgrade

[rich vortex]

thanks

[trim echo]

Got it thanks

[sleek vortex]

they don’t use 3.5 anyway they use haiku

or a mix of models with mainly haiku

[dark dome]

thanks for the reply bro

[honest basin]

it might take too big of a hit on their 62 million dollar funding

[signal hamlet]

Hey @brazen vine! Have you used it earlier? You can create 50 images per day, with each use given back to you in 24 hours after you used it.

Can you try refreshing the page, do you still see that it's blocked?

[fading moth]

I'm so madge at the lmsys benchmarks. 4o is "better" than 3.5 at everything according to it... :\

I hate 4o.

[twin sluice]

Hi - how do I get access to the Perplexity "Pages" feature?

[signal hamlet]

Hey @twin sluice! Pages are currently available to all the Pro users.

[twin sluice]

Oh wow - thanks - I keep scowering the docs every few weeks and couldn't find any indication of that. OK, I'l upgrade and try it out.

[brisk quest]

On the app… after 2-3 chats in a thread with the ‘Writing’ focus, the thread then starts pulling in sources from the internet for the responses, ruining the chat. And I can’t get it to stop using sources after that in the ‘Writing’ thread. Kinda defeats the purpose.

Is this by design?

[signal hamlet]

That's one of the issues to be fixed, sorry for the inconvenience.

[tired quartz]

What does a collection prompt actually do? It simply insert that prompt to each thread?

And what about the profile in the setting page? It works as a global prompt which will be applied to any thread you start with?

I used to create bots on Poe, and it's kinda intuitive. When it comes to Perplexity I get confused.

[fading moth]

Collection prompt acts as a "system prompt" at the start of each response in the Collection's threads. The profile is overridden by the collection prompt.

[tired quartz]

But I found the threads in a collection never follow the collection prompt...

[jaunty dove]

Which model? I think I saw reports that sonnet 3.5 doesn’t work.

[tired quartz]

I tried pro search, s3.5 and opus, none of them follows

[jaunty dove]

Try GPT4o, Pro is not a separate model, it’s based on your AI option. S3.5/Opus is from same Anthropic family.

For me, it did work but I haven’t used collections lately.

I just tested a collection with a prompt asking to speak in binary only, and it works. Using S3.5

[tired quartz]

Just tried, it doesn't neither..

I experiment with this prompt #💬│general message

It works fine when I directly use it in the thread as part of my request, by replacing the last line with my query. But I can never get it work as a collection prompt.

[wintry pivot]

Appearently Perplexity made up a summary for this page:
https://www.nytimes.com/2024/07/18/us/politics/biden-election-drop-out.html

https://www.linkedin.com/posts/masatokajimoto_hallucination-or-fortune-telling-i-used-activity-7219901693172334593-WJRG?utm_source=share&utm_medium=member_desktop

[wintry pivot]

I was able to replicate this hallucination. If I use URL alone, the problem persists. If I copy and paste the text from the publicly available part of the article, I get correct summary.
https://www.perplexity.ai/search/summarize-https-www-nytimes-co-1WaRiCMxSV27GJdg1YoszA

Is it possible the New York Times has hidden text somewhere in an ad or otherwise that has controversial content about Biden already dropping out? @signal hamlet

[south kindle]

The best way to use it is one of two ways. Either start a new thread within the collection or transfer another thread into the collection and then send the prompt revise with the collection prompt to make sure it’s injected. It’s very powerful and useful once you have some default frameworks and prompts you regularly work with. Ideally your initial prompts start within a specific collection so it’s fresh in the context window.

But having a bunch of threads within a collection doesn’t mean any of them are connected or data from them is sent to the collection to improve it or anything. Just a container that injects a custom prompt.

If you’re struggling, share ether the collection itself or the prompt itself so we can look at it and test it on our side.

Collection prompts are shorter and usually have a specific way of working.

Sometimes converting your system prompt to json helps. Or have sonnet 3.5 revise it.

That’s the order I usually follow most times at least.

[wheat pike]

from 600 to 540

unlimited > 'virtually unlimited' > 600 > 540 > ???

where does it end

[tired quartz]

For example, I cannot make this prompt a collection prompt.
https://discord.com/channels/1047197230748151888/1237001047730290730

The threads never follow.

I would like to see more examples for collection prompts and best practices to write them.

[fast meteor]

it was always 600* until opus 50 came then after agentic pro search it becomes quietly to 540

||* the initial launch counted pro/copilot and search separately but quickly abandoned||

[fast meteor]

pro search does not attend to ai profile/collection

[south kindle]

I’ll see what I can do.

ALWAYS start a collection without Pro search so it’s injected cleaning into the context.

[tired quartz]

Neither did other models in this case.. 😢

[tired quartz]

Tried last night. Didn't work. I shall give it another try then.

[fast meteor]

pro search toggles and all llm models (gpt/claude) are working independently

pro search took place between users' inputs and the llms' responses

pro search does not have access to the ai profile/collection prompt
all llm models do have the access but they are not used during the pro search phrase

you could think it as a two-phrase approach
phrase 1: pro search handles the user original inputs, and gathers information
phrase 2: llm replies in natural languages with information gained from supplementary materials acquired in phrase 1 (and with its own internal knowledge)

[tired quartz]

Great help! I'll study later. 👍

[umbral blaze]

The advantage of perplexity in relation to other AI mechanisms is that it is very simple to resolve doubts about slang and informal context as it offers a wide range of perspectives. As incredible as it may seem, sometimes a simple question GPT 4o chat cannot answer..

[vivid kraken]

someone probably said this before but free perplexity seemed to be upgraded to gpt4o mini

[elfin ferry]

would be really cool if its true but still: https://discord.com/channels/1047197230748151888/1225549032718602332

[austere kestrel]

my favourite model
https://www.perplexity.ai/search/system-you-are-a-helpful-assis-EWOSiVPYShmfytmLoP5Iug

Oh wait..
Is it the Supreme commander model? Or Claude? Or GPT-4o-mini?

as mentioned.. no self-awareness (and influenced by whatever tokens/system prompt precedes the question "who / what model are you?"...)

[placid coyote]

hmm, did sonnet slip and said it has a human brain? I thought ai companies try so their AIs don't appear human nor compare to human etc

[foggy pecan]

What is the limit on image creation

[south kindle]

“Try” is the keyword. Asking the model about itself will always yield incorrect results. Best to avoid doing in general regardless of platform.

[halcyon coral]

The limit is 50/24h 🙂

[foggy pecan]

Ty

I got hit with limit

For the first time ever

[fleet salmon]

how do we force perplexity to use certain websites in its searches only? through API that is

[halcyon coral]

https://discord.com/channels/1047197230748151888/1262520766642131054

[fleet salmon]

wow this is great. this works in the API as well?

[halcyon coral]

Yes, they should work in the API. Example:
filetype:pdf "machine learning basics"

[fleet salmon]

Awesome, thank you

[glad crypt]

does pplx-api offer support for citations and suggested questions yet?

[fast meteor]

citations in closed beta

#1262708305885794325 message

[glad crypt]

cool! im guessing suggested questions not yet?

[fast meteor]

no public info on that

[pseudo crypt]

Hey, is there any way to turn off images that perplexity gives alongwith responses

[sleek vortex]

You could block the side entirely using ublock origin or a similiar element blocker

However there’s no official way to turn it off

[sleek vortex]

If you don’t want to see any images and don’t use the AI ones either then this is your best bet

[tired quartz]

So, from what you've written, I conclude that the ai profile as well as the collection prompts take in effect only in the second phrase, where the pro search has already been done.
So, if I want my prompt to influence the process of pro search, I should always put that prompt in my query. Am I right?

[fast meteor]

currently yes, as that is how i understand the black box

[tired quartz]

Ah, I thought you were the dev and were telling the mechanism 😂

But still a great help 👍👍

[fast meteor]

pro search as well as the llms should understand structured texts, xml, and pseudo programming language better than plain texts when writing prompts

[tired quartz]

So basically, can't we now come up with an easier way to have a prompt affecting the pro search but do not need to type every time? 😢

[spice lance]

why doesnt desktop version have voice and tts?

[tired quartz]

The decoupling languages prompt is quite useful but I got tired to add it manually every time in my query...

[fast meteor]

you can leverage the site search function

https://discord.com/channels/1047197230748151888/1225572798966267966

https://zapier.com/blog/add-search-engine-to-chrome/

[tired quartz]

Good idea. But doesn't work for Android chrome..

I cannot even export the bookmarks from Android chrome without synchronizing via Google account (

[fast meteor]

i havent tested it on android but it seems you can have javascript bookmarklet on Android https://paul.kinlan.me/use-bookmarklets-on-chrome-on-android/

this should make auto filling possible in theory

but are completcated

as a workaround*

[foggy dew]

Very nice! Tks

[frigid schooner]

Hey may I ask where should the chat limit count for the Pro search be located?

[fast meteor]

it will show in ui when the counts fall below 100

[frigid schooner]

Thank you

[fast meteor]

#⚡│ask-community message

if you want to know the exact number @frigid schooner

[zinc birch]

How to search for Collections for research better,
Let’s say I’m interested to know about ABC company financials , emotion , direction etc etc.
so I have created 5 collections like that separated out, and linked them in description and created a main collection

Now that I have 5 collections I want to use all of them together to get info triggering all of these 5 collection and summarized data using just one collection -
Can someone explain how can this be done

[south kindle]

you'll have to pass a thread through each collection and use the prompt "adhere to the collection prompt" to get them injected. honestly, you'd be better off making a single collection that covers all of it so you can make sure the prompt is in the context window.

[hybrid solar]

I need Perplexity to be able to execute code, and graph stuff. In fact, they need to have the power to retain long-term memory.

[south kindle]

#⭐│starred message > this tool will help at least

[green pike]

I have autism when I talk I use the AI to speak for me here's what the AI I named Frank wrote

Hey there, Perplexity team! 👋
Oops! Looks like I accidentally posted my message in the welcome channel. My bad! 😅 I'm still getting the hang of Discord.
Just to clarify, I was trying to reach out to the developers about some ideas to improve the app. I've got some thoughts on making the AI more customizable and flexible, especially for users like me who use it for writing and communication support.
Is there a better place to share these ideas? I'd love to chat more about how we can make Perplexity even more awesome for all kinds of users.
Thanks for your patience, and sorry again for the confusion!
I hope it is comprehensive appropriate and correct

That message again please

I don't understand what is going on someone help I keep getting a pop-up message that says one message at a time or something it's too fast for me to read

#💬

#💬│general am I doing this right now

#💬│general am I doing this right now

#💬│general

This is not an introduction this is a conversation I did exactly as I was told to do

[fading moth]

#👉│feedback-general Would be the best channel

[green pike]

I respond to that thank you

Ras can I still communicate it keeps giving me the same message

[fading moth]

I think the moderating bot was just glitching.

[fallow imp]

HI

[south kindle]

not required but Pro Search will be more comprehensive.

[silver laurel]

Hi all. Is it true that turning pro on will ignore the profile?

Should i use collection instead to get better output?

[austere kestrel]

i don't think that's the case

using Pro search might make the AI less likely/able to adhere to the instructions in the profile prompt (especially with longer conversations)

[silver laurel]

I read someone say profile is in phase 2

[austere kestrel]

i'm not sure what is meant by phase 2?

[austere kestrel]

it shouldn't make a difference afaik

[silver laurel]

Mean pro get the research first then follow by profile

I find the profile has limited characters allow

I am using sonnet 3.5 for code assistance

[austere kestrel]

ohh i see.. yes that's kinda the structure

[silver laurel]

I always get the same mistake from assistant even though is stated in profile to learn from past errors

How do you ensure assistant learn and do not repeat same mistakes

[austere kestrel]

but it means instructions in the profile about searching won't be seen by the system that does the searching, so saying only search xyz.com won't influence the search process. but the profile instructions will influence how the responses are delivered (like say "talk like a pirate" or whatever and you'll get)

[austere kestrel]

you can't

nothing perisists / is memorised across conversations

every new conversation starts from scratch

[silver laurel]

Oh. I thought they’re suppose to adapt to threads

[austere kestrel]

no unfortunately not. it is totally oblivious to anything said in previous interactions

[silver laurel]

I also state to clarify by ask questions but so far is not happen too

So i start to wonder if pro interfere with profile setup

But i do noted that every specific instruction is taken care of

Btw i use focus mode

[austere kestrel]

it's a bit tricky but there are basically two models/systems. One gives the answers, the other does the searching...they don't see the same information so it's tricky..

#💬│general message

[silver laurel]

Yup. I just find it not smart dealing with a few instructions and sometimes i wonder is it perplexity or the LLM we used determine the quality

[austere kestrel]

i see i see.. keep playing around / experimenting - there are limitations and quirks, the more familliar they become, the easier they are to handle

or you could share your prompts / use case – others here may be able to help out 🙂

[fast meteor]

define quality
pro search (tried to) provide relevant information
llm is responsible for fluency and consistency

[placid coyote]

it reads a bit schizophrenic when models appear to not know themselves 😄

[austere kestrel]

i've seen claude haiku (and other models) think claude haiku is a model specialised in generating haiku poetry

which kinda understandable tbf ha

though come to think of it.. I haven't seen sonnet confused as a 14-line poetry specialist..

[fast meteor]

quick search

[serene sand]

writing mode

[fading moth]

I personally never use Pro search for anything tbh...

I value my context window too much~

[placid coyote]

pro for nothing, really? I usually use pro when searching for multiple things (or even specific attributes of multiple things) and asking for comparisons/recommendations (not always a success, but chances are better than without). in normal search as well, since I often forget to turn it off 😅. but it's true it tends to be slower. especially after the addition of the Programming step that sometimes gets triggered incorrectly and at best this will slow down a response. at worst it will break text encoding, waste context, may confuse LLM

[placid coyote]

I think for pro users, it is same context window size, but search results from the pro mode will fill it more (possibly with bad data and probably diluting user's instructions)

[fading moth]

Yis, more sources and pro search steps means more token per response

And depending on sources... Can break user intent/requests and occasionally focus on source rather than user

[placid coyote]

not only dependent on found search sources, it is also dependent on user's prompt too. if the system chooses based on the prompt to run a programming step which outputs nonsense, it can confuse/steer the llm. like what is happening with emojis/languages https://discord.com/channels/1047197230748151888/1261713638293635112

[magic fern]

How does perplexity have such a huge valuation with so many problems? Does this business model even work or is it pure founder/AI hype?

[fading moth]

Well, that's the million dollar question (no pun intended). Part of it is the secret sauce that ended up being ignoring robots.txt, and the other is bigger companies eyeing it.

[magic fern]

huh, interesting. what about the aws lawsuit? i login later to see they're good chums now lmao

[placid coyote]

it is overhyped for sure, but search is quite good, uiux could be better, but havent seen better uiux combined with solid search yet elsewhere (maybe google is getting there?). and ratio value/price is really good (600 uses of sota models per day, though with some extra limitations opus 50)

[fading moth]

Well, think of it like 4 rich dudes (Anthropic, OpenAI, Google, and Meta) buying lunch for each other but the check is higher than expected... They all are used to being "cost doesn't matter", but the first one offering to pay will start a war over the "tip size"...

Perplexity's high valuation + bidding war = ?

[magic fern]

lmao

[placid coyote]

agentic search is nothing new, but even the non-agentic one felt usable. many months back I remember trying few alternatives and it was very bad (I realized that good ai search is probably not that easy to pull off and stopped being that hard on pplx). maybe I missed it, but I think they were early and had something working. add some ads and they are winning
not that I use it much, but "pages" felt kinda new. saw only some hacky ugly "do it yourself" barely working solutions before that. front-end wise, I think they did good. last time I tried pages though, there were issues with "visibility" (context), what llm sees when generating new sections. not sure if fixed

[deep snow]

What are the best settings for perplexity pro for different use cases (e.g., research, general knowledge, coding) - which model to use, when to use pro and when not to? Does anyone have any resources they like?

[fast meteor]

#1240356137497530408 message

[south kindle]

#1225549032718602332 message

[signal hamlet]

Hey, @spark crow! Please post the projects you are working on in the #😎│cool-projects channel. Thank you!

[wide stratus]

WOOOOW!!!! Prplxty was featured on LMG's TechLinked last month! 🤩 (timestamp-link: https://www.youtube.com/watch?v=ts3DqM_t3fA&t=38s ) 🤩 🤩 🤩 🤩 🤩 🤩 🤩

[torn sierra]

It has many issues and using a lot of things here is annoying, but it's cost-effective. And that's mainly what keeps it going. The search function is decent too, but overall, the execution as such is really poor

I've been looking for something better for a long time, and there was always something wrong. So I'm sticking with Perplexity and other LLM services as a supplement 😀

[honest basin]

pure hype

[fickle forge]

Is perplexity down? Not loading or letting me log in on the app.

[sage yarrow]

If I'm not mistaken the listen Discover feed is only available on IoS, when are you going to release it on Android ?

[sage yarrow]

Work on my side

[strong pond]

This is a dumb question, but does anyone know if/when an Apple Watch app is coming out for perplexity?

[fading moth]

I highly doubt it will.

._.

We're barely getting the regular apps for iOS and Android to work.

[strong pond]

Aw 😦

[fading moth]

Yis, it is sadge.

[vital fossil]

There’s an independent app called Per Watch.

[strong pond]

I’m using that but it’s kinda meh

Although impressive

[vital fossil]

Guess it’s the best we’ll get for now.

[visual olive]

Hey is the ai able to generate images?

[harsh verge]

https://discord.com/channels/1047197230748151888/1225557012532695060

[visual olive]

Thanks, is it possible on mobile?

[harsh verge]

#📱︱feedback-mobile message

[visual olive]

Thanks again

[tired quartz]

Which model is used in non-pro search?

[fast meteor]

could be gpt 3.5 haiku...

[silk field]

hey I'm a big fan of perplexity and I want to try pro but I kinda wanna try it first... I realize that there doesn't seem to be any discounts aside from a few but is a small free trial available?

alternately, can I ask if it's better to subscribe from android, pc, or ios?

Because I can get a certain amount of cashback via subscription through certain modalities

thanks!

[jovial mantle]

Service is the same regardless of where you subscribe, you have a single subscription, it redirects to a stripe page for payment

There's a 10$ discount bonus if you get a referral link, last year there was a 2 months free code given after a 1 month free trial

[tired quartz]

Is it stated anyway officially?

[tired quartz]

I can show you my referral

[jovial mantle]

👀

[tired quartz]

I've never tried this before though

[jovial mantle]

You paste send him the link, then either of you gets 10$ off

[fast meteor]

no i dont think so

[tired quartz]

Can this be applied to annual subscription too?

Or it only works for the monthly?

[jovial mantle]

it only applies once for billing cycle

so it's better to use it on monthly

[wanton nebula]

Hi

(°_°) i use perplexity

[silk field]

I appreciate the information. Will this get us in trouble though?

[hardy bay]

is anyone else having touble atatching file?

all the sudden cant read my fiels?

[tired quartz]

I don't think so?

[jovial mantle]

Why would you get in trouble? It's the way it's meant to work

[tired quartz]

It's an official method to get discounts

[jovial mantle]

Do you offer any discounts or referral codes?
While we don't offer ongoing or student discounts, Pro subscribers can take advantage of referral discounts. For each new Pro subscriber you refer, both you and the referee receive $10 off the next billing cycle.

There's no limit to the number of referrals, and discounts apply automatically once per billing cycle.

https://www.perplexity.ai/hub/faq/coupons-and-discounts

[silk field]

fair enough

thank you sir

[barren sluice]

Y a t’il un moyen d’avoir la voix pour parler et répondre sur perplexity sur le PC.? Windows et edge

[serene sand]

pas pour le moment

[jovial mantle]

Pas tout de suite, mais à l'avenir. Je trouve que chatGPT et Character.ai sont de bonnes alternatives. Vous pouvez même demander à character.ai d'utiliser votre propre voix.

[teal bear]

yes I can't seem to make perplexity see my document, no matter what I do

it's a current bug?

[wraith crow]

yes, there seems to be a bug in Perplexity at the moment. please send a link to the thread where this problem occurs in this discord thread https://discord.com/channels/1047197230748151888/1259631790340178051

[tawdry chasm]

Haiku probably

[barren sluice]

Merci

[fast meteor]

web > android > ios mobile platform like google/apple also takes a cut fron your subscription

[harsh verge]

Perplexity n'a pas (encore) implémenté cette fonction sur PC. Mais oui, c'est possible indirectement. Dans vos paramètres Windows, sous la rubrique Accessibilité > Voix > Autres options de commande vocale, vous trouverez l'option 'Saisie vocale' (voir la capture d'écran). Cela vous permettra, en cliquant sur l'icône du micro (voir la capture d'écran), de parler. Pour lire la réponse (dans Edge), faites un clic droit devant le texte à lire et choisissez 'Lecture à voix haute' (voir la capture d'écran).

Raccourcis :

Pour parler : [Windows + H]
Pour lire : [Ctrl + Maj + U]

[fast meteor]

https://www.perplexity.ai/search/fee-that-stripe-ios-apple-stor-zf7w2lEqRJKUsUSoLq1eDg

[barren sluice]

Merci beaucoup !

[barren sluice]

Ça fonctionne bien, encore merci 😃

[rustic kestrel]

What model is used in free search

[fast meteor]

gpt 3.5 (turbo) or haiku i think

[rustic kestrel]

Alr ty

[opaque ferry]

is server facing issue?

[placid coyote]

just tested it against llm arena and (at least for me) it doesn't look like haiku. on arena for nonsense prompt very concise one-line gothic mix of colorful ackermann in haskell haiku tends to write only one line of code without comments, while "default" gives me more explanations (in all of 3 attempts). it looks to me more like the new omni mini. (big sonar writes much longer and gpt3.5 gives text, not code)

[fast meteor]

llama_x_large 405b llama 3 is coming

[upbeat charm]

Anyone else having issues that perplexity are no longer able to read uploaded textfiles properly? Im paying for Pro.

Seems to have happened sometime today

[upbeat charm]

Same problem here today.

[fast meteor]

^ @upbeat charm

[upbeat charm]

Same here. Happened 3-4 hours ago

[upbeat charm]

Ah thx

[fast meteor]

if the attachment file does not contain any PIIs, pls also include it in that bug report

[upbeat charm]

Oh sorry wish I could, but too many things I cant share in it. It worked fine on friday tho. A pretty ordinary text format log file.

[placid coyote]

🦙 saw some torrents on reddit, not that far of a terabyte. not sure if it was legit though

[upbeat charm]

@fast meteor btw. How well are your information protected when you turn on "dont use my data for training or otherwise" kind of option? Should I still anonymize everything?

I think the options is called "AI Data Retention" in settings.

[fast meteor]

https://www.perplexity.ai/hub/faq/what-data-does-perplexity-collect-about-me

and general privacy policy below

https://www.perplexity.ai/hub/legal/privacy-policy

and you can get soc2 w/ enterprise pro
https://www.perplexity.ai/hub/faq/pro-enterprise-faq

[upbeat charm]

But you are not able to "Hi, I want to see the data you have on Peter Gantric with address xxx" in any way? or "Show me information used for company XXX by other users"

I use it mostly generically but good to know.

[fast meteor]

you need to ask dev like @signal hamlet or tech support team. but
from my understanding, unless the data is publicly available on the internet, it wont show info like those.

[upbeat charm]

Thx

[fast meteor]

it was: 782176.86 MB in total
safetensors files in 191 parts

[fast meteor]

the readme file in the torrent is ... weird, but also seems legit

[fast meteor]

this is what i got from omni model

Alexey Romanov is a multifaceted professional with a rich background in data analytics, engineering, business management, and education. His current role at Perplexity AI allows him to leverage his skills to create value and drive innovation. With a strong foundation in engineering and a diverse career spanning various industries, Alexey continues to make significant contributions to the field of data analytics and AI.
sourced from linkedin

[upbeat charm]

Nice. Just a bit funny hehe 🙂

I had settings at Claude 3.5 Sonnet.

Aha, I see I wrote Alex, not Alexey. That probably explains it.

[harsh stag]

have you taken into account the "helpful assistant created by perplexity" system prompt?

[sleek vortex]

and ai profile

[placid coyote]

I thought about it, but even the code differs between models too (for example some models like to use HOFs, bind etc; while others don't even write a function and just an expression; some don't even write code at all). that was with ai profile disabled. I can't rule out the system prompt of course, but the omni mini output was repeatedly so similar to the "default" on pplx so I mentioned it

[sleek vortex]

also temperature

perplexity has a really low temp, right ?

i mean that shouldn’t change much but it could to some extent

[placid coyote]

shouldnt temperature affect more variety? it was quite consistent, mostly changes in the text around, but used operations and structure of the code seemed similar, same thing with approx amount/length of commenting text. also I think for gpt4 models lower temperature wasnt as bad as for claude 3 family. though again, not sure if it affects output length or preferred code patterns that much (like that haiku test). didn't thought so, but havent really tested

[sleek vortex]

also the default model very well may not be one model

[placid coyote]

no, it is not

[sleek vortex]

since perplexity has tons of users, it’s likely they use whatever they have enoguh capacity for

[placid coyote]

but the one I tested today looked like omni mini

[sleek vortex]

which is probably a mix of one or more of haiku, their own online LLMs, maybe 4o mini and maybe 3.5 turbo

Also might depend on region

very likely they have seperate throughout purchased in say japan AWS/azure regions to reduce latency there

Also might depend on the 5 sources it pulls (or well whatever context is provided)

In writing mode it very well could be an 8k or 32k model over the other larger context ones to save money

Or if say the 5 sources are all relatively short, etc

[placid coyote]

I haven't had 3.5t for a long time. I think mostly haiku, sometimes sonar. at that time others had sonnet 3.5 (possibly the non-pro users for limited uses, under the hood; at least from some tags on fe it seemed to point to that)

[sleek vortex]

true, i doubt it would change the patterns itself

[sleek vortex]

yeah there was a AB flag for sonnet

assuming you’ve seen my AB viewer thingy for perplexity (?)

[placid coyote]

only screenshots here on discord 😅

[sleek vortex]

oh do you want the link

[fast meteor]

you can fingerprinting the model if you are bored
https://arxiv.org/html/2407.10887

[silver laurel]

I want to improve the ai profile but the character limits restricted me to improve further. How?

[vital fossil]

Anybody know if a Mac app is in the works?

Adding to that, does anybody use Peek on Mac? How are you liking it?

[sleek vortex]

this?

[vital fossil]

This one: https://prateekkeshari.gumroad.com/l/peek

[left sandal]

Do custom instructions for your profile in the web version carry over to in app?

[sleek vortex]

it should, yes

[left sandal]

Alright, thanks. Wasn’t sure if those features carried over.

[placid coyote]

is ... this true? do the models get info how much "space" they have left in a response? I have a suspicion sonnet is lying to me 😶

[sleek vortex]

i mean it probably is lying

how would the model even like actually know how many characters/words/tokens its outputted accurately

sure you could tell it the convo so far has say 16382 characters

but then how does it know exactly, or even roughly when to stop

[placid coyote]

LLMs tends to be very bad with text length, especially with longer text. I wonder if it isn't something what some search step returned, or the agentic system itself, like "can't do more searches" (it did miss some requested data about later movies) and sonnet interpreted it this way

hmm, or could I have hit context window limit just from searches?

[south kindle]

how long are these threads? how much actual text?

fairly sure you can copy/paste the entire thread and do the token conversion.
I know openai has one, but this one is pretty simple/handy: https://llmtokencounter.com/

[south kindle]

PPLX is limited to a 32k window

[placid coyote]

thread, well, it was a first response.
I used the linked token counter, response: 740 tokens 2963 characters 389 words
my prompt: 90 tokens 360 characters 57 words
the openai tokenizer returns slightly higher number. but this is just my input and output, not the search data passed to the sonnet, no? edit: and my profile. but I think even all that visible to me combined is very small amount of tokens

[south kindle]

very little chance it's the context window then.

32k means 32000 tokens

feel free to share the thread

[south kindle]

i believe the conversion is 4 words per token. "page" sizes vary though.

[fading moth]

But let's assume you fill the collection prompt with 2000 characters (the limit), and add the pro search steps. How many tokens are being used?

[near kite]

Hey how the heck is perplexity soo fast!

[fading moth]

Are you using Sonnet 3.5?

[near kite]

No idea, I just came to know about the product from a vidoe, created an account and have been trying it around, it just gives answers in second

[fading moth]

I see, I see. You should be able to check in your "thread" (name for conversations) what LLM is being used.

[near kite]

I have no idea what you guys talking about, btw are you on pro?

[fading moth]

It's usually at the bottom of the responses.

[fading moth]

Collections are like... Mini-tuned models for LLMs. They have a prompt that is injected before each response to guide the LLM in a specific direction.

Yes, I'm on pro.

[near kite]

I see

[fading moth]

"Token" is the term given to the measurement of words/characters it can remember. One token is usually 3-4 words, depending.

Perplexity limits their model's "context window" (their memory for the conversation) to 32k tokens.

[near kite]

is it enough for you?

[fading moth]

Ergo, the question was how many tokens are being used in the background before the user asks the LLM anything, and how much would that affect usable prompts before the window is filled.

[fading moth]

For most things, yes. When you approach the limit you can usually ask the LLM to summarize the conversation then start a new thread to continue from.

[fast meteor]

max 32k token each turn/round*

multi turns/rounds convo can go to like 128k

[fading moth]

Egad :o

That should be specified in Perplexity documents. It only says 32k.

[placid coyote]

One token is usually 3-4 words, depending.
that's ... a bit too much for a token, or did it already moved so significantly from gpt4 times?

[fast meteor]

dont quote me on that, that was a long time ago
i cant find the original post and im not sure whether things have changed or not

[rotund perch]

hi

[hearty blade]

hi

[serene sand]

https://x.com/perplexity_ai/status/1815431484767142272?s=61

Perplexity (@perplexity_ai) on X

When you know, you know.

Twitter · Jul 22, 2024, 4:59 PM

[fading moth]

Try it without the emojis :X

[green merlin]

so intelligent

[shut plaza]

hmm

[green merlin]

try it without pro

[pale compass]

@green merlin LLMs are known to be bad at math, on any service, not just perplexity. When you use the service without the proper tools that make up for that deficit, its honestly just showing you're uninformed as to the limits of existing llm technology.

[fading moth]

Writing mode + pro disabled -^

[shut plaza]

LLMs are just word predictions

[fading moth]

o_O

O_o

[green merlin]

😂 oh no the LLM is dumb, I gotta white knight for it

[fading moth]

Mine got it right tho?

[green merlin]

@fading moth looks like you were using claude. by default i got gpt 4o i imagine

[fading moth]

4omni

[pale compass]

@green merlin No. I'm saying its bad at math. I'm saying that Perplexity provides tools that make up for the shortcoming of LLMS, and by choosing not to use those tools it just shows that you're not using the tools available to you properly.

[green merlin]

do you just sit on here all day defending perplexity? typing 9.11 - 9.9 or bing or google gives the right answer immediately. i am perplexed that perplexity gives me such a wrong answer to a basic math question

saying i'm not using it properly is a dumb argument

[fading moth]

It seems like a prompt skill issue, if I'm being real :I

[pale compass]

Except you're not using it properly lol. If you're not going to use the tools provided to you, then why come here to point it out when the soolution is simply to use the tools provided to you?

[green merlin]

you guys are nerds lol "prompt skill issue" 🤣

[pale compass]

lol You're either a troll, and I'll stop feeding you. Or you're beyond help. Good day, sir.

[fading moth]

Agreed.

[green merlin]

did i have a "prompt skill issue" on google? no, sit down you guys sound stupid

[fading moth]

Don't feed him, bio.

I know it's tempting...

[pale compass]

I know, I know. lol

[green merlin]

perplexity needs to make changes to improve the UX, that's all im saying

saying i have a prompt skill issue is you guys being in denial to what's obvious

[fading moth]

He's trying to DM me @pale compass xD

[pale compass]

lol

me as well actually

[green merlin]

it's amusing see you guys try so hard to defend perplexity 🤣 google wasn't confused by my question

[south kindle]

Why do I feel like he was here a few months ago doing the same thing....

[fading moth]

Probably was

[south kindle]

don't be obviously ignorant to how a platform is meant to work when the entire community would agree you are being petty and avoiding actually learning how to use these tools correctly. When most of us get an obviously wrong answer, we revise our prompts to at very least test alternate outputs. that is a core skill in this environment. It's not alive. It has a specific way it was trained. talking to it in that same way will always yield better results.

[green merlin]

https://tenor.com/view/cryptoflash-crypto-flash-tattoo-vintage-gif-27569875

[hallow linden]

https://www.theinformation.com/articles/cond-nast-sends-cease-and-desist-letter-to-ai-search-engine-perplexity

lol seems like perplexity is making enemies

[sleek vortex]

its mainly because of discover

[hallow linden]

I hope so

[sleek vortex]

😐

[tame current]

on skibidi?

[sleek vortex]

...

[fading moth]

tl;dr -- Conde Nast wants Perplexity to stop using content from his stuff. The New Yorker, Vogue, and Wired.

[south kindle]

They are also losing out on ad revenue. Users are getting the content from the discover feed, not their site. it'll be a complicated period for a while I would assume. Especially as they improve Pages.

[tame current]

Use any summarizer bot on Poe or similar html grabbing sites

[sleek vortex]

i think i saw this posted somewhere else, so yeah

[sleek vortex]

i mean they cant or wont be paying back to sources any time soon

[hallow linden]

Honestly, discover seems like the most expendable part of perplexity, so I don’t really mind if they drop it

[south kindle]

It has gotten very sparse length wise lately. Not sure if that's pressure or what. i feel like the old discover "threads" had more content at times.

[tame current]

Hi @everyone.
I am a senior Blockchain and Full Stack Developer has a rich experience.

About me:

• Built website with SEO-Friendly & Pixel-Perfect & RWD from Design using JavaScript framework or libraires such as React, Next.js, Vue, Angular.
• Designed, implemented and consumed Restful APIs or integrated several APIs and SDKs(ChatGPT API, Firebase SDK and so forth) with projects.
• Designed and implemented efficient database solutions using tools such as MongoDB, PostgreSQL, MySQL.
• Built the Mobile app using React-Native, Expo, Flutter and deploy it to the the Google play store & App store.
• Built the Chrome Extension using React and deploy it to the Google.
• Used several tools such as Trello, Asana, and GitHub for collaboration.
• Built several staking, exchange platform and wallet chrome extensions, some ICO projects.

Now I'm looking for a new opportunity, so I can start work immediately.
And I can work with you properly in your time zone.
I am always ready to work with you.

If you have free time, could you please check my portfolio.
https://jason-mendoza.vercel.app/

Please feel free to contact me directly to discuss about your project.

[fading moth]

I don't know whether to chide you for posting something like that in a general chat of a Discord or be impressed by the brashness.

[old bloom]

Does anyone know if there's plans to fix the voice reading issues when going over numbers or currency? Also perhaps to have a speed modifier on the voice?

[rare jasper]

WHY IS IT SO SLOW TODYTA

[sick arch]

https://tenor.com/bnLYV.gif

[wheat pike]

if there's anything perplexity has right now it's plans, cus we sure aren't seeing any meaningful progress

[tired quartz]

What impact will this have on pplx?

I'm now hesitate to get an annual subscription.

[hallow linden]

Honestly ai moves so quickly I might move from annual to monthly too

[serene sand]

reading numbers issue has been reported to ElevenLabs long time ago

[acoustic stirrup]

I have a question now. I purchased the enterprise version a few days ago for $400. How many seats does this include in total?

[harsh verge]

Self-Serve
Companies with fewer than 250 employees can access Enterprise Pro through a self-serve flow, with pricing at $40/month or $400/year per seat. (https://www.perplexity.ai/enterprise)

[brisk idol]

Hi

[placid coyote]

that won't change how the rest tokenized. tokenizer is pretrained, it is not adaptive. each color represents one token (color change in the "stream" is token change to be precise). so for english it is usually one token for common words, for other languages it tends to be way worse (e.g. " příliš" = 4 tokens). for less common words, like " tapestry" those are 2 tokens, " bioluminescent" is 4 tokens. I haven't seen a mainstream tokenizer which encodes so many words, rarely two

One token is usually 3-4 words, depending.

[fading moth]

Ah

:3

Seems like it shifted quite a bit since I was last asked that question (around the release of GPT4). Thank you for correcting me.

[placid coyote]

omni tokenizer made some improvements if I remember correctly. but I think mostly for non-english (eg that Czech), so it is closer to english now - common words in foreign languages have fewer tokens = cost less, closer to english

32k*3.5 words would be fairly large. sadly we are closer to 32k * 0.75 words (not that it can be reliably and easily used on pplx)

[fading moth]

Or maybe I was confusing it with 3-4 characters instead of words in my brain

Either way, I'm glad I know the current state.

[placid coyote]

oh, those 3-4 characters seem to fit much closer. usually I saw the 0.75 tokens = 1 word
edit: with *, thats for english and natural language

[fading moth]

Ještě jednou díky, Kirai. Jdu se odhlásit a spát. Promluvím si s tebou zítra.

:3

[rain nest]

Does perplexity ever plan on incorporating twitter searches into the social search? It’s just Reddit right now right? Better yet would be nice if it could do a broad Internet forums search.

[placid coyote]

twitterx searches would be a very nice feature, but I think their API is insanely expensive. probably pplx would have to struck some better deal

[fast meteor]

https://x.com/perplexity_ai/status/1603441221753372673?lang=en
pplx was a twitter search engine initially, they now pitched towards general searches

Perplexity (@perplexity_ai) on X

Introducing Bird SQL, a Twitter search interface that is powered by Perplexity’s structured search engine. It uses OpenAI Codex to translate natural language into SQL, giving everyone the ability to navigate large datasets like Twitter.
https://t.co/N1BtF47JYu

Twitter · Dec 15, 2022, 5:25 PM

[sleek pike]

Perplexity Pro is now available in the AWS Marketplace? Couldn't find it

[placid coyote]

wow, didn't know that, thanks for sharing. that tool looks quite useful

[fast meteor]

but the ceo admitted it was harder than expected
https://lexfridman.com/aravind-srinivas-transcript/#:~:text=(01%3A46%3A42,by%20the%20way.

(01:46:42) Correct. The reason we picked SQL was because we felt like the output entropy is lower, it’s templatized. There’s only a few set of select statements, count, all these things. And that way you don’t have as much entropy as in generic Python code. But that insight turned out to be wrong, by the way.

[austere kestrel]

also Musk came along and made the twitter API stupidly expensive.. lots of companies (and researchers) basically got screwed overnight

but yeah i think they also just realised that they were onto something bigger.. like it started as an internal slack bot iirc, then the twitter sql thing - the whole internet would seem the natural progression ha

[sleek vortex]

main issue is either pay like a ton a month

or you have to use scraping

[amber pewter]

Which model do you use the most in your searches?

[fast meteor]

#1240356137497530408 message

[fast meteor]

just 42000 usd per month

[tawdry chasm]

Will perplexity be the first provider of Llama 3.1?

also btw can anyone from perplexity fix the deployment of Gemma 2 27b in playground

[austere kestrel]

aha yup.. or $2.5m per year..

apart from a the big ones like talkwalker, hootsuite, a bunch of social monitoring platforms were just forced to drop twitter entirely (as well as lots of researchers)

[silver laurel]

I wonder should i get a claude pro account as well to take advantage of its latest feature for Artefacts

[placid coyote]

you mean the artifacts sharing, artifacts in general, or do they have something newer?

[mossy pendant]

Hello, hopefully someone can help. I have a pro subs on my icloud account, I signed in to the web version via apple sign in, but it appears to have set up a new account with the same email. I dont have access to pro features

[silver laurel]

They have this code preview for conversion image to working app

[placid coyote]

that sounds like artifacts (html preview/embed), I don't see anything newer on their blog. if that's the case, you are on pc and feeling adventurous, you can achieve some results also on perplexity with ailin - #😎│cool-projects message

[rain nest]

I have Claude. I would not subscribe to it just for Artifact. The new AI itself Sonnet 3.5 is great though

[silver laurel]

I wonder is it the same using on Perplexity for Claude 3.5 or the one one Claude is more advanced

[modern fern]

i've noticed the quality of perplexity has been dropping alot recently

especially for reading files

i've been using perplexity pro for 8 months now, and have referred friends to also use this service

i really hope i dont have to find alternatives :/

[placid coyote]

on claude I think you get more context and higher temperature (more creative responses), but much less uses. yeah, file handling is much better at claude and chatgpt. also ailin is not official, not easy to set up and not as user friendly as artifacts on claude. ailin can do some extra things, like code execution (don't think claude has that) and ability for that code to use internet (I think that ADA can't do that on chatgpt), persistent filesystem; but ailin also has drawbacks, like nothing similar to that easy artifact sharing on claude nor specialized react embed

[placid coyote]

personally I don't use file upload much, it's been hit and miss for code for a long time. but if it used to work for you better, you should check pro-feedback or bug-reports and share example threads where it fails. at least from discord, it looks to me devs are working on something related to files

[modern fern]

at this point, it just doesnt read the files anymore it feels like

[rain nest]

I use perplexity for internet search type searches. That's what it excels at.

[modern fern]

yeah i use both and it used to work better. idk whats going on :/

[stuck topaz]

it is the same model, however, what different is the context windows (A context window in AI refers to the amount of text (measured in tokens) that an AI model can process at one time). In perplexity the context window is only 32000 tokens, in claude this is much higher. So as long as the conversation goes, perplexity will come to a point that it can not process all the information at once and start to lose relevance and consistency.

[stuck topaz]

maybe they reduce the context window or modify something under the hood, perplexity direction is in my opinion just like Cory use, for internet search type and replace google. For processing pdfs or handling document, you can use chatgpt or claude

[modern fern]

the reason i came to perplexity was to do both

Froggy_cry

[placid coyote]

I know I read file uploads under 32k tokens should be put into context, but I just never see that. it always felt like RAG to me. to be fair, I think in most cases I sooner hit response length limit (I think around 300 lines of HTML) or diluted instructions so much (by search results?), it stops doing what I instructed it to (when I forgot to go into writing focus or to disable pro search)

random use of "Programming" step is also not helping that

it really shouldn't use programming step to write 3 tweets...

[fast meteor]

is your file in pdf format? an alternative: copy the paragraph for context and also include your query to ask perplexity

[modern fern]

even when i have large pastes that turn into paste.txt, it just refuses to do what i ask

it doesnt really even read the file

[placid coyote]

btw I tried converting very short source code (html with js, nowhere near 32k tokens, more like 1-2k) to pdf and while it seemed to see the code from the pdf file, it also consistently failed to reproduce it 1:1, like forgetting , in object literals. it was correct in the pdf and sonnet has no trouble repeating same source code when I feed it in parts

[fast meteor]

i know that this is not ideal but i dont think you need that much of the context

to get answers with the help of the internet

[modern fern]

the cost of doing those queries i gave it were probably too high

so they have to nerf it via code

[stuck topaz]

soon or later we will come to realize that perplexity can not help with much code because of context window limitation in questions and also in the response itself. For best code and document handling performance, I would use claude 😃

[tranquil orchid]

Hey @stuck topaz!

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.

[placid coyote]

I am trying out cursor. a bit less features compared to IDEA, but the AI integration is very nice and they have sonnet 3.5

[tranquil orchid]

Hey @stuck topaz!

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.

[fast meteor]

this is my semi automated workflow for slides/ppts
use gemini flash to perform ocr and image transcript and get a better understanding on contexts and intuition behind using perplexity

[fading moth]

🥔

[random ocean]

Meta Llama 3.1 8b, 70b and 405b has been released today. When can we expect it to be in Perplexity?

[rich sphinx]

I'd bet pretty quickly, they likely just have to switch a few config params. I'm not sure but they're likely just wrapping AWS Bedrock for these models.

[sleek vortex]

not on bedrock just yet

hmhttps://www.aboutamazon.com/news/aws/meta-llama-3-1-models-AWS-generative-ai

what region are they in

found it

theyre in oregon

[rich sphinx]

Yeah Bedrock llama 3.1 links are 404ing for me anyways. Theyre probably getting hugged a bit too

[sleek vortex]

requesting access....

yay!!

no 405b ?

[rich sphinx]

Ollama has 3.1 8b and 70b integrated too

[sleek vortex]

my mac cant run it

intel

well it can but super slowly

[rich sphinx]

I don't have 6 4090s for 400b anyways lol

[sleek vortex]

3.1 8b inst

well yeah ig its instruct and not chat

90TPS not bad at all

seems like some stop tokens arent added in yet

[rich sphinx]

Its also available on build.nividia.com/explore/discover

They have 405b-Instruct up too in their playground

https://build.nvidia.com/explore/discover#llama-3_1-405b-instruct

[warm cave]

Perplexity, when Sonar 405b?

[south kindle]

To get around the context window you should be using multiple threads for separate tasks. If you try and do everything in one thread you’ll have a bad time.

[south kindle]

This is also just a general rule people follow for other LLMs. To keep it focused on the task at hand without the chat baggage.

[upbeat charm]

This is why Im dreaming about being able to chat with collections of threads 😛

[winged moon]

Sonar 405b on lab?

[tame current]

Anyone tested llama3.1 for actual use yet? It feels so strange or broken, but maybe that's provider (Fireworks) problem

[fast meteor]

they must use LoRA or similar to speed up the inference and/or RoPE to extend the length

[sleek vortex]

it works fine on aws

[serene sand]

Collections now have Threads/Pages on iOS

[hushed summit]

I wish cursor allowed you to use local modals. That would be great

[placid coyote]

are you sure it can't? I see in Cursor Settings -> Models -> OpenAI API Key a field for custom base URL

[placid coyote]

though I think not all features will be working. I believe cursor has some own models

[tame current]

They route (routed?) all requests to AI providers through their servers, so it won't work with local ollama instance, you'd have to make a reverse proxy to expose it

At least it was few months ago

[placid coyote]

oh, so it must be public facing api server. didnt know that

[winged moon]

We just got llama 3.1

[placid coyote]

oh really, that was fast 😄

[winged moon]

I was just switching models to get different results when it just showed up

[random ocean]

Do I have to logout and login?

[placid coyote]

didnt phone apps needed to update, or is it now pulling models from backend?

[placid coyote]

switch it on a web version, then you can use it from the app. this was written from an android app which showed model in settings "Default" https://www.perplexity.ai/search/write-short-poem-in-czech-abou-ksVKF84qQNG2wH8Q7ia2NQ

[random ocean]

Yeah, its available in Web version

[placid coyote]

hmm, llama doesn't seem to like japanese symbols 😑

[random ocean]

Now it is available in iOS app too, That was fast

[sleek vortex]

pretty sure it doesnt need an app update now

it pulls it from their AB test config

[placid coyote]

on android doesn't seem to. killed the app by force and still no llama 😦

[random ocean]

There was a lag of a few minutes in between web perplexity and iOS app for llama 3.1, maybe it will appear in android app soon.

[cinder comet]

you guys are so fast

damn

is it the instruct ver or the base one?

[fading moth]

Well that was fast

[hushed tartan]

Yo recibiendo un mensaje de Discord que creia era importante, era @signal hamlet }

[brittle kestrel]

these guys keep getting better and better

[frosty vessel]

what are yall thinking, llama or claude 3.5 sonnet

[fossil roost]

isn't liama , available free on WhatsApp?

[lament spoke]

We got 400b llama 3. Just a matter of time before 1,000,000,000,000,000 model drops 😮

For those who don't that's 1 trillion parameters

[tame current]

yeah

[warm cave]

idk, i know its free on meta.ai

[winter raptor]

How do you use it on whatsapp?

[tame current]

just ping meta ai

@Meta Ai

as so

in whatsapp

[warm cave]

Yeah, but its it 405b, i have used it before when it was 70b

[tame current]

you can open a new dm as well

ah

i'm not sure if 405b is still available there

[warm cave]

I know its 405b on meta.ai

[tame current]

How does llama benchmark compared to 3.5 sonnet

[stuck topaz]

yeah I would like to know also

[frosty vessel]

i dont see it on any llm rankings yet

we'll have to wait and see

[warm cave]

I asked pplx with this:

[frosty vessel]

interesting 🤔

[tame current]

Idk if there are such benchmarks, but surely it's worse at coding than Claude

[warm cave]

https://scale.com/leaderboard

[stuck topaz]

ok sonnet 3.5 still king in code

[warm cave]

Yeah, i think so, but this is a huge leap for open source(weight)

[frosty vessel]

sonnet is really crushing it

[tame current]

Deepseek? WizardLM2? Deepseek was better than 4o on benchmarks and WizardLM is the best at complex instruction following, which made it beat GPT 4 in many queries

For coding, I mean

[warm cave]

Yeah, for coding, but as a base general purpose model

[stuck topaz]

it is funny that sonnet is better than gpt4-o in math, but we wont use it in academic because claude is missing latex rendering, what is the point of better if we can not understand what the response is 😃

[warm cave]

I hope to see a codellama 3.1

[tame current]

Was there codellama3?

[warm cave]

I dont know, i think not

but hopefully someone will do a great code fine tune

or meta will idk

[tame current]

I'm hoping for WizardLM but with Llama3.1 base, maybe it will crush proprietary models further with its instruct

[warm cave]

this is a bit funny, i wonder if any model gets it right.

[warm cave]

Yeah, that would be amazing 🔥

[tame current]

lol

Maybe it self-leaked new conversation mode?

[slow canyon]

Gm

What is the consensus for best AI outside of perplexity

[tame current]

No consensus, just do coin flip or check them all yourself

[sleek vortex]

no

at least on aws theres a stop token issue

not sure if that carries over here

[tame current]

Was a joke fyi

[placid coyote]

well, the llama follows instructions from collection prompt quite closely. its "artistic programming" is not great, probably on par with turbo, omni and opus

[tame current]

Seems to be the same on fireworks, unless you mean something else

Whats the best ai ever

[shut plaza]

What’s up with Llama 3.1 405B just repeating the first answer when I ask a follow-up question?

[tame current]

Maybe it's not working well with Perplexity RAG? AWS and Fireworks have problems with 405b, so maybe PPLX too

[shut plaza]

I’ll keep using 4o for now then

[livid aurora]

I think its s instruct model

[placid coyote]

I don't know, maybe my prompt is bad? but sonnet (second image) gave me those symbols I wanted. llama seems to not like to output asian symbols, despite knowing them? at least some basic ones I tried

[sleek vortex]

possibly related to the issue with stop tokens

since they’re using aws probably

maybe aws is using the wrong tokeniser for the model

[chilly nova]

is there something like this in perplexity? https://x.com/minchoi/status/1815812112796565690

Min Choi (@minchoi) on X

Instant Intelligence is wild with Llama 3.1 8B + Groq 🤯

Twitter · Jul 23, 2024, 6:11 PM

like iterate the prompt with voice

[mint cradle]

Yes with the Android app. I often use the built in speech to text on Windows with the web app as well. Win key + H. I'm a terrible typist. ☹️

[fading moth]

Imma be real... I still prefer 3.5 over 405b :X

405b is great, don't get me wrong, it's just not performing as well for me personally as 3.5.

[frosty vessel]

claude 3 opus got it right for me

[cinder comet]

405b not working

[cinder comet]

sonnet 3.5 is the best model so far

the benchmarks are misleading for llama

its not that good tbh

i mean its good at simplifying concepts and whatsnot

and thats it

[fading moth]

Yay, I'm not alone :D

[cinder comet]

many people noticed that

@sleek vortex is it down?

llama 405b not working on pplx

can you check please?

😖

[jaunty dove]

Wait a few days before judging it.

Gpt4o and sonnet 3.5 sucked on day 1/2 for me and it got better.

[warm cave]

hmm, but that looks like it has web search

[unreal crater]

You are right, here is GPT-4o no web access

[cinder comet]

is llama 405b down?

[unreal crater]

'

[cinder comet]

can someone check pls

[serene sand]

https://x.com/rypearts/status/1815868829169328349?s=61

Ryan Putnam (@RypeArts) on X

✧ 　 　 ✧
˚ *
　 　.　 　　　　 　　
· · 　　 　 + ✧
　　　 · 　 · ˚ . 𝓈𝓊𝓂𝓂ℯ𝓇 𝓋𝒾𝒷ℯ𝓈

Twitter · Jul 23, 2024, 9:57 PM

[jaunty dove]

Works here for me. I’m switching back to Sonnet for now due to speed.

[tame current]

are u brave

[fading moth]

Less are you brave and more "are you rich enough to run it without waiting 5 minutes per response?"

[vestal wyvern]

yo add back 7 day free trial for pro

I needa scam yall more

[sleek vortex]

lmao

too bad

[vestal wyvern]

I started failing school after yall took it away

nah fr 😭

[sleek vortex]

this is crazy icl

[vestal wyvern]

Nah but fr bless yall, I got carried with that pro stuffI got over 14 accounts tryna use them trials makin it thru school 🤧

[clear stump]

Hi, I am considering subscribe to Perplexity Pro. Can it generate images like ChatGPT-Plus does?

[shut plaza]

bros not gonna make it through life

[shut plaza]

yeah

not really it's primary purpose, but it can

[clear stump]

Thanks, I am evaluating the Pro demo, it gives me this.

how can I instruct it to generate an image correctly?

[shut plaza]

there's a generate image button to the right

[clear stump]

Oh... sorry, I am a newbie to this. Thanks a lot, I will check it out.

[vestal wyvern]

😭 😂

[wise edge]

Which one is bigger, 3.9 or 3.11?
3.11 is bigger than 3.9.
on hugging face

[wet quarry]

hello!

[tame current]

When you generate an image, do you first have to ask an AI a question and only afterwards ask it to illustrate the answer with an image? It's not possible to immediately ask the AI to generate a picture from a description? How do you prefer to do to generate images?

[fading moth]

The strawberry one was interesting. My response from 3.5 was 2 R's, but when I asked it to look again it said there was a silent? R hidden in strawberry then corrected itself to 3 R's.

[pearl saddle]

how often do you guys use wolfram alpha through perplexity?

[peak yoke]

Can we implement a system where a llm can run a calculator and execute code by itself to find answers to mathematical questions and then further on explain it?

[peak yoke]

Whenever I want an answer to a maths question I use perplexity

So all the time basically, I don't search for maths questions often but it's nice

[upbeat charm]

Tried 405B a lot yesterday trough huggingface chat. It was pretty darn overloaded lol.

[wise edge]

sorry, i don't know , it even not write python code to calculate it ~ 4o is best for 128k

and dall-e 3 is not work well again
reason : Failed due to moderation. Please try a different query

[peak yoke]

Yeah i tried it a bit as well, it was really nice

Gimme a question if you want me to ask it

[unreal crater]

Gpt-4o mini got it right every time for me on the first try for some reason while Gpt-4o had to be told to look again. LLM weirdness

[fading moth]

If I reworded the prompt I'd likely get consistently better results, but that would defeat the purpose of judging responses from a problematic prompt.

[pale compass]

From my understanding, it has to do with how the tokens are broken up for the LLM. In my practice, 4o and LLama 3.1 both can get the right answer with : how many r's are in the word strawberry? Write out your thought process step by step and only answer after the thought process.

3.5 sonnet though still says 2

[fading moth]

Ultimately, many of the "problems" LLMs have with logic often have to do with the wording... not always, but changing the 9.11 - 9.99 thing to 9.11 minus 9.99 solved it across the board.

How many times does the letter R appear in the word "strawberry"?

Try that

Works across board :I

[fading moth]

Weird, worked for me. Let me double check.

[placid coyote]

😞

[fading moth]

Might have been a fluke on my part.

[placid coyote]

so, it really might be a problem with the tokenizer, as somebody noted yesterday. it looks to me like llama thinks it is returning those symbols

[fading moth]

Was a fluke, it's popping up with 2 now.

[pale compass]

Yeah, the thought process trick can help with simple logic questions. Also works for the 3.9 or 3.11 being larger question

[placid coyote]

tbh I don't know how "strawberry" kind tests are useful. what is the use case? you see Rs correctly from the LLMs

[pale compass]

Its just testing the basic logic of the LLMs. not much use beyond that

[placid coyote]

the simple math looks to me much worse, that already bitten someone here on pplx. I think with contents of energy drinks

[pale compass]

is that what you were looking for?

[placid coyote]

seems correct. yeah, other models yesterday didnt had this issue

[pale compass]

Yeah, for math 4o tends to be best For math. Also, pro search should call Wolfram for math so its more accurate.

[placid coyote]

oh, I take that back. sonar (older llama) suffers from it as well

[placid coyote]

or it can use the programming step. both should work

[pale compass]

yeah, I'm getting issues with the LLama based models giving answers in japenese. Probably just a limitation with the model. 4o workds, but also provides the english translations which wasn't asked for. Sonnet 3.5 seemed to do it best.

[placid coyote]

though sonar is based on llama 2 70B if I remember correctly and that maybe didn't really officially didn't support other languages? not sure

I think the l3.1 405B should know asian languages. it can output the symbols, but only rarely or maybe for specific promtps, not sure

[pale compass]

I don't typically have it answer in Japnese. For this specific question at least it seems to have trouble.

[placid coyote]

special symbols are the problem. kanji, katakana and hiragana. but the model obviously knows the words, otherwise it couldnt respond in romaji

[pale compass]

Yeah, that's an interesting limitation, even in the newest version. 4o and Sonnet seem just fine.

[placid coyote]

what

it knows them alright, just doesn't want to output them...

maybe something with system prompt of pplx? possibly the language selection?

[pale compass]

weird it won't even write them if I ask for it in a paragraph. I wouldn't be able to say why. I can't imagine its prompt related if it otherwise can display the writing properly outside of this prompt.

[fading moth]

How many times does the letter r appear in the word strawBERRY? This works for Sonnet... But only Sonnet

Weirdness

[placid coyote]

probably because of different tokenizer

[fading moth]

Tokens be weird, yo

[placid coyote]

yeah 😄

[pale compass]

even more fun

[placid coyote]

but without this tokens workaround, we couldn't have any LLMs. it would be computationally impossible task

[pale compass]

When I have sonnet do an example, Llama can follow it with the kanji properly

[placid coyote]

yep, as soon as it sees anywhere any jpn symbol (maybe asian in general), it can continue

[fading moth]

True true

[placid coyote]

it just doesn't know how to start writing symbols on its own, poor llama

[fading moth]

I'm legit surprised it's having language issues for such a popular language, albeit tricky.

[placid coyote]

all asian languages, not just japanese. chinese is huge

[fading moth]

Chinese very big, yis

[pale compass]

Its just crazy because its clearly about to translate it, and use the language. It just won't display it in answers.

I went to meta.ai and tried it there

and it has the same issue

so its clearly an issue with the LLama model, not a system prompt on perplexity or anything.

[placid coyote]

could be overtrained to not change languages?

hmm, yeah, on nvidia thingy as well

[pale compass]

even Gemini can do it lol

[placid coyote]

well, not necessarily the model. could also issue with many things around it, like the tokenizer, supporting code around (lib), template, possibly even default system prompt

[fading moth]

What's weird is if I say "show me the character" it will work...

But "what is the character" causes a panic

[placid coyote]

I know people complained with several open chinese models that they occasionally randomly without asking switch from english to chinese, so possibly meta finetuned a bit too much, to counter this behaviour

I wonder if base model also does this

[pale compass]

Possibly. Strange. Hopefully they fix it. At least Perplexity has other models available for the time being.

So using the API straight from open router with no system prompt or changes is has the same issue

[placid coyote]

yeah, it may not be too common use case. seems to be fixable by collection prompt You can use Japanese symbols for Japanese words, kanji and kana. For example: 貴方

[pale compass]

Yeah it seems like if you use a symbol once it works.

[fading moth]

Do you think it has to do with llama being hesitant to mix alphabets in a single response? Like it wants to use romanji to keep with the Latin alphabet.

[placid coyote]

I think it is mostly the prompt language tuning - when user doesn't write in japanese, don't use japanese symbols

[fading moth]

Hrm

Interesting

[placid coyote]

but it is way stronger for asian languages, possibly because too different symbols? for example for czech it doesn't have this issue, it can switch freely and use extra non-latin characters https://www.perplexity.ai/search/write-short-magical-poem-in-cz-v7lRjXgxS5.tDX22.g4F2Q

[pale compass]

Yeah, works for other languages right off the bat too, like Russian works fine.

[placid coyote]

but, russian uses very different alphabet, no?

[pale compass]

It does, which is why I chose it.

[fading moth]

Maybe Asian characters are considered more similar to emojis? Just a random guess

Be gentle, I have smol smooth brain

[pale compass]

lol well it answers in emojis just fine though

[fading moth]

No idea

[placid coyote]

you might be onto something. the japanese mini lesson (which doesnt have jpn symbols in my profile) works when asked for emojis 😮

[fading moth]

[placid coyote]

this was an accident, I wanted to try if it would respond with that one "smiling" kana sometimes used as emoji. this one: ツ

[pale compass]

lol its just trolling at this point

[placid coyote]

feels like it. but possibly combination of rare emoji and japanese mini lesson from ai profile with an emoji word in prompt pushed it far enough from that language/symbols post tuning

[harsh verge]

https://www.perplexity.ai/search/10-animals-in-kanji-only-respo-CdygXAlIQte9iJXH.fXGwg It worked for me 🤔

[placid coyote]

thats from search

[fading moth]

But you are using web search

Yis

We're using writing mode

[harsh verge]

oh yes, sorry 😅

[fading moth]

[placid coyote]

though good to know at least in search it works properly. that's I would wager main use case on pplx

I wonder what model writes code for programming step

[pale compass]

Yeah, probably due to the correct symbols in the search results lol Same as you providing an example to it in your other prompt.

I assumed it was their Sonar model

Or wahtever their pro search model is

[placid coyote]

hmm, now I am thinking about what if search results return japanese/chinese/korean symbols. won't sonar switch response to it? like, I would think there is a reason why meta did this post training

[pale compass]

Well, I'm off to bed. Good conversation, good luck and good night lol

[serene sand]

#1225557012532695060 message

second method from this post

@tame current also check https://discord.com/channels/1047197230748151888/1194788138124587128

[chilly nova]

who the hek uses mobile apps for serious work

[tame current]

@serene sand thank you!

@chilly nova hustlers who are always working

[serene sand]

deep dive ➤ #💬│general message

[jovial cedar]

Hi everyone, been using Perplexity Pro for some time now and can say it's the best AI platform out there 🙂

Since I've been playing around a lot with it, I have a question. Perplexity can utilize several AI models. Does the tool know which model is the "best" for a given task? Do my answers differ depending on the model I use? This is something I haven't figured out yet.

[serene sand]

https://discord.com/channels/1047197230748151888/1240356137497530408

[halcyon coral]

Hey @jovial cedar!
You can change the model in the settings -> #⚡│ask-community message
The answers will always differ depending on the model.
Here is a guide to help you choose the right model for your use case -> https://discord.com/channels/1047197230748151888/1240356137497530408

[harsh verge]

I tried your prompt "10 animals in kanji only. respond only with Japanese symbols" with the @south kindle tool in order to take stock of the various LLM results. See the conclusions here: https://app.wordware.ai/share/8c523d8b-c109-4189-a6ce-cc9bfc5d24a2/history/129be5a3-d4ae-4069-85f5-156052669490?share=true

(This prompt processes a question using Sonnet 3.5, Gemini 1.5 Pro, llama 3.1 70B&405B, GPT-4o/mini, Sonar Large (online model), Claude 3 Opus, Claude 3 Sonnet, and lastly Claude 3 Haiku. The app then employs Sonnet 3.5 to review and rank the responses. Upon completion, Claude Sonnet 3.5 initiates a verification process to thoroughly examine all responses, identify errors in its initial ranking, and generate a final, revised ranking of the model outputs, aiding in determining the optimal answer.)

[leaden scroll]

Hello, is this a issue?

i thought this was Meta, it says it is

[supple vine]

It is the Llama model generating the answer, but there are Perplexity prompts on top, to help with the search part. Is my limited understanding. So through Perplexity, using Llama.

[sleek vortex]

models dont know who they are accurately

if i tell llama that its actually made by apple it will just spit that out

or if you tell it its gpt5

so on

[thorn stream]

llama 405b has weird grammar on perplexity

[fast meteor]

https://discord.com/channels/1047197230748151888/1225549032718602332 @leaden scroll

[thorn stream]

missing period in the first two paragraphs

[leaden scroll]

I laughed out loud. You are funny have a good day

[leaden scroll]

okay great, thanks

[placid coyote]

nice. wait a minute, who is judging this?

1. Sonar - Large: The prompt itself, not a valid response.
   I see repeated prompt, but also 10 kanjis. did the judge overlooked it, or am I reading it wrong?
   edit: I see, at the top, so sonnet 3.5? wouldn't expect a mistake from that one

[tame current]

Will pplx be able to read/watch instagram or tiktok videos in the future like with the youtube videos?

[tame current]

the tech is there

but there is no point in implementing it for reels and tiktoks

[fast meteor]

why not? they are ||might be|| the future of search engine

[tame current]

reels and tiktok might not have inherent value with repect to search

[fast meteor]

https://www.euronews.com/next/2023/02/05/gen-z-is-using-tiktok-as-a-search-engine-is-this-the-end-of-google

For Gen Z, TikTok is the place to turn to when you’re looking for answers

also

TikTok overtakes Google as most popular search engine among Gen Z.

[tame current]

But it is for youtube?

[tame current]

most videos on youtube are informative is what i've felt

tiktokt and reels on the other hand...

[tame current]

Exactly, there's so much data on forums (reddit) and youtube. I can't imagine how much data is on tiktok

[tame current]

Oh bro, tiktok is great for learning how to fix something in the house, finding watermelon at a grocery store, doctor advice. You shoudl check it out

doctor advice?

Yeah there's cardiologists, surgeons, etc.

On the platform that give advice about their field, health, and other things pertaining to it

I would but it's banned where I live, also how are you sure someone's not pretending to be a trained medical professional ?

Because They have their name and credentials in their bio. There's an account called Drterrysimpson. He's a surgeon and gives great advice

[tame current]

oh i see, where do you live, if yo dont mind me asking

India

[tame current]

yes, they might be good. but good intent cannot be expected out of all netizens

Well just like youtube and reddit you have to read between the lines.

The point is TikTok on pplx would be great.

perhaps, only time will tell

[twin cargo]

hey

[harsh verge]

Yes, it is indeed Sonnet 3.5 that analyzes, and it correctly states that for "Sonar - Large," it seems to be a prompt rather than a response. It is among the poor performers ranked from 6 to 10, with only a few performing worse in the given task.

[cinder comet]

i think mistral is worth it more than llama 405b tbh

it gives me better results, hopefully its added on pplx

[placid coyote]

the part sounds, like it responded only with the prompt. yeah, it continues, but seems a bit biased tbh. sonnet itself also included extra stuff, yet Also provides exactly 10 animals in kanji. does sonnet rank itself higher or something? does it rank worse the markdown list which he didn't include? edit: oh, it's that "in kanji only". seen so many variants of this I overlooked it, my bad 😅

[south kindle]

This is based on my wordware template. It's in flux. Also I'm unable to update it currently, and 70B is down, so it stalls out on new searches/inputs.

https://app.wordware.ai/share/999cc252-5181-42b9-a6d3-060b4e9f858d/history/3c76952a-c352-4520-95a2-ccf1a7b2b056?share=true < This one is working though. Their embeds aren't "great" yet, but it's a search on Mistral Large 2, with a PPLX style article at the end.

[south kindle]

2 things. Highly recommend making an account on Wordware, and hitting the "try it yourself" button. This will copy the entire template to your account and you can tweak it, break it, or do whatever. it's basically a IDE for building LLM based apps, using complex logic and prompt chaining.

They give you plently of free credits to play around with all the tools/models/settings/templates.

[warm cave]

Mistral large 2

[abstract needle]

i'm confused

[fading moth]

I'm always confused

[sleek vortex]

what even is this benchmark

[warm cave]

I think its this: https://github.com/nuprl/MultiPL-E

[fading moth]

Benchmarks vary so heavily with AI... One test says Sonnet is the best at coding, another says 4o-mini is better in every category. Others show 405b winning across the board...

[warm cave]

Yeah, its a mess

[fading moth]

Honestly, I ignore benchmarks these days. I used to watch them but they are just so inconsistent that they're almost subjective, like a movie rating score, e.g., rotten tomatoes or IMDb.

I've loved movies that score terribly and hated movies that score well. I have to watch them myself to actually find out.

[warm cave]

Sucks tho because it would make it easier to chose the best model for the task, if you need have it write java vs python vs bash.

[fading moth]

Yeah, it's really unfortunate.

[south kindle]

spaming this speed image because I feel it's relevant:

[tame current]

Wasn't Llama supposed to stand out? Seems kinda average like the other models

https://www.reddit.com/r/ClaudeAI/comments/1eb9njj/claude_35_vs_llama_405b_vs_others_tested_by_ai/

[pale compass]

LLama 3.1, from what I understand, isn't unique in that its more powerful than other models. Its 405b version is meant to target a similar level as 4o and Sonnet 3.5. What is meant to make it special is that its open source, and it can be downloaded and used locally by anyone with the hardware to run it, unlike Sonnet or 4o which are currently only available thorugh the cloud via Open AI, Microsoft, and Anthropic with no option to run locally, even if you have access to the needed hardware.

[warm cave]

Llama 8b and 70b really stand out imo, but 405b is impressive because it is close to GPT4o while being open source

[warm cave]

I dont like this: https://www.perplexity.ai/search/google-is-the-only-search-engi-awmBq7utRw6FpNqjhY3szA

[hallow linden]

Yeah that doesn’t bode well for Reddit search mode

Unless perplexity can pony up the money

[warm cave]

Dose Reddit search mode use api or via general search?

[mighty fox]

Why is grok by XAI not in hugging face chat?
It shouldn't be size issue because hugging chat hosts llama3 400b model while grok-1 is 314b model. Is it some licensing issue?

[warm cave]

Hmmm... GPT4 Turbo never gets it wrong, i tried 10 times, but now GPT4o is geting it right too, maybe it was a fluke when i tried it before

[sour bough]

Perplexity now uses gpt4o mini btw

[jaunty dove]

For what?

[warm cave]

What makes you say that, what did you see?

[fading moth]

This is news to me about mini

[tame current]

And regular o I assume?

[pearl saddle]

damn, what do you think perplexity is going to do?

[fading moth]

Only corporate knows

But I doubt Google is going to loosen their grip on it.

[wise edge]

https://www.perplexity.ai/page/mistral-unveils-large-2-00GRlebXQQiufg1mtooQxg
llama3.1 is too slow, and can not calcute 3.11 and 3.9. i perfer to Mistral Large 2.
Supporting dozens of languages, including English, French, German, Spanish, Italian, Chinese, Japanese, Korean, Portuguese, Dutch, Polish, Arabic, and Hindi, the model demonstrates impressive multilingual capabilities

Breadth of language support: Mistral Large 2 appears to support a wider range of languages, with "dozens" mentioned, while Llama 3.1 405B has added 8 new languages to its previous capabilities
###################################
Mistral Large 2 has a 128k context window and supports dozens of languages including French, German, Spanish, Italian, Portuguese, Arabic, Hindi, Russian, Chinese, Japanese, and Korean, along with 80+ coding languages including Python, Java, C, C++, JavaScript, and Bash.
vs
Llama 3.1 models are conversant in additional languages including Spanish, Portuguese, Italian, German and Thai.

[fading moth]

Meh

[gilded river]

Can someone explain if i get a paid subscription to perplexity pro, will i have full access to every of its ai models like if i had individual paid subscriptions to each of them ?

[cinder comet]

Yes

[stuck topaz]

yeah, the direction of llama is always not try to beat claude or gpt, it focuses on achieve similar performance like claude or gpt, thus remain opensource and can run locally, which then can be fined tuned with custom data, similar to dolphin. llama 405 is said to beat gpt, but I find its response is pretty much very slow. So for the sake of daily use, I prefer only gpt4o and sonnet

[clear stump]

Is there any information about the context size of the sonnet 3.5 used in Perplexity? In Poe there are two version, one is 8K, the other is 200K.

[stuck topaz]

perplexity only have 32k context size window (only when processing pdf). For normal search, it is lower, but we dont know how much (they r not very transparent about this)

[cloud stump]

hello

[peak yoke]

Llama 4 multimodal coming around December boys!

[bleak oxide]

Man i really enjoy using perplexity lyrics

[clear stump]

Thank you for the information.

[spice lance]

omg llama 405gb actually follows instructions from your profile, never seen this before on perplexity

[fast meteor]

sonnet 3.5 also follows instructions

[torn sierra]

Correct! Sonnet adheres remarkably well to the instructions provided in the system prompt, and it can be significantly better fine-tuned than GPT-4o.

[tame current]

i feel like all LLMs will eventually diverge at the same level of intelligence....

i meant converge

lol

[golden solar]

Hey. Is there a direct way to copy and format the conversation into Notion.? and have it structured in a given way.?

Looking to improve the way information is taken out into other text writing apps

[sour bough]

No just mini for the free search

[sour bough]

The free version

[sour bough]

Tokenization is openAI-like now and the model itself admits to being openai-made. Previously, they were using claude 3 haiku only (free) and I have since been testing every couple few days since the drop of 4o mini to see if the model changes. It used to admit being made by anthropic. Now it's constantly admitting being made by openai

[south kindle]

#1225549032718602332 message

[sour bough]

Sure but that's something else. If for a long period after confirmed implementation of haiku, the model states CONSISTENTLY that's its made by anthropic, and a few days after the release of gpt4o mini it starts saying it's by openai

Isn't that a significant indication that there was some change?

considering that it makes total sense for them to move to gpt4o mini due to cost

And that the confirmed model claude 3 haiku was consistently choosing to spit the tokens corresponding to anthropic instead of openai

[grim coyote]

Good morning! I just got Perplexity Pro. I can only seem to upload 4 documents per thread. But I read that I was supposed to upload more than 4 documents.

[jagged forge]

Hi all, I just published my first page, and shared a link. But I wanted to see how many views or clics I had. I don't find how to do it. Does a sort of analytics exist ?

[tame current]

Ah, but not for Pro I assume

it says enterprise sub but it doesnt give pro or other models, is that normal?

[grim coyote]

Hm, so I assume getting Pro doesn't actually get a person more than 4 file uploads. That's sad. I'll cancel and come back when I can reference more than 4 files. 😄

[sour bough]

it's probably the default one for pro if you dont choose a better model

[placid coyote]

yeah, it felt o mini to me #💬│general message . though since pplx can use different (default) model depending on user, it makes it a bit harder

[sleek vortex]

not normal, try clear browser cache + log out and log back in

if not then contact support

also check youre on the correct account

[tame current]

same issue different browser (tried in edge)

its the account i clicked the invite link on

[sleek vortex]

hmm

check with your account manager once that theyve paid for the multiple subs

if they have then id contact pplx support

since this isnt intended behaviour

[tame current]

ah maybe its not fully setup yet then

[sleek vortex]

possibly, yeah

[tame current]

alright ty, i'll wait for it to get setup

i'm so stoked to use unlimited pro searches on claude 3.5 lol

[sleek vortex]

600/24 hours* (at 32k context only)

but close enough!

well 600 is all marketing its basically unlimited lol

[tame current]

thats still 25 searches an hour, or one search every 2 minutes

[sleek vortex]

never seen anyone actually hit 600

at most i hit like 200

but average like 25-50 honestly

[supple vine]

you get access to the model through Perplexity. You could for instance go to Openai and use your Perplexity login.

@waldoh Just clarifiying for him. I think some people seem to think that your perplexity sub will allow you to use the models directly from the models website.

[upbeat charm]

Reddit now blocking all searches from AI and search engines? 😦

[upbeat charm]

Where did it say you can upload more than 4?

[placid coyote]

yeah, my "best" was remaining 100 uses. I was using pplx whole day for programming and research. I believe it was the project where ai was writing all the code and I wasn't allowed to manually edit anything. it was not an effective use of my time 😅

[fading moth]

Google bought rights to have exclusive search access to Reddit

[warm cave]

Maybe Reddit will be replaced one day, Reddit is useful but not worth switching back to Google imo

[tame current]

is llama better at reasoning and claude 3.5 sonnet better at writing human-like?

[placid coyote]

this is starting to get a bit convoluted 😄

[placid coyote]

not sure about llama and reasoning (🦙 didn't look better at programming), but sonnet is great at writing (handles Czech very well, and I would expect it to know more languages than llama). but on pplx sonnet had (has?) low temp, so for writing is a bit dumpbed down here

[opaque shoal]

I initially read that announcement for database maintenance as a 10-minute warning for a two-day-long maintenance window! I’m glad I read that a second time.

[shut plaza]

I wonder how much money I've cost Perplexity in API fees

[warm cave]

So far my favorite thing about Llama 405b is that it never says the same thing after a rewrite, it always seems to tackel the problem in a new way

I wish Claude was like that... https://discord.com/channels/1047197230748151888/1253515579743015073

[elfin ferry]

https://openai.com/index/searchgpt-prototype/
damn looks cool

[sleek vortex]

most of their models arent paid for on api anyway so dont worry!

[sleek vortex]

this is literally what im building...bruh

man

[elfin ferry]

i mean this was teased a few months ago

we thought it was going to be a gpt-5 or smth and then gpt-4o came out which wasnt that

[tame current]

lol it's over for perplexity

[serene plume]

o7 I enjoyed my stay

[zinc folio]

its so over for you

[serene plume]

u have speed pfp