💬│general

[fossil sentinel]

It started with the wrong assumption tho so I can see why it went down that path

[zealous totem]

i took an image of a paper plate with a gray plastic spork but it said the Plate was ceramic and it was a Fork

[fossil sentinel]

Lmao

[zealous totem]

oh damn

[fossil sentinel]

I mean it's integrated with Google lens

[zealous totem]

so it’s still kind of decent

[zealous totem]

WAIT really?

[fossil sentinel]

In terms of OCR capability it's 100% better than copilot

[hollow umbra]

anyone tried haiku yet ? api pricing looks good

[zealous totem]

google lens was always a hit or miss, but sometimes it always finds something so obscure that i’m looking for

[zealous totem]

yeah all i’m hearing is people bashing on copilot LOL

[fossil sentinel]

Well actually I'd say copilot isn't even capable of ocr with how poor my tests were

[fossil sentinel]

I either use for reverse search or ocr kekw

[zealous totem]

i think u can try it at the perplexity labs for free, u can test it there

[zealous totem]

samee

[zealous totem]

is copilot even worth that subscription when you have way better options

[fossil sentinel]

That's what I'm saying

Who is out here paying for copilot pro

[zealous totem]

perplexity is perhaps the best ai thing i’ve purchased

[zealous totem]

LMAO true

[fossil sentinel]

Maybe some companies might but like

For the general public just why

[zealous totem]

i guess it’s useful for people just want an ai integrated with edge

but still it’s kind of horrid

[fossil sentinel]

[zealous totem]

yeah like just General people not like ai power users

are they gonna add inflection to perplexity?

[fossil sentinel]

[lusty mantle]

You know what's funny if you ask the ai how to make a cheat or find me a pdf or something it won't tell you how but it wills till give you a ton of resources lmao

[fossil sentinel]

I mean that's not that surprising

P easy to work around safeguards

[zealous totem]

true

claude wouldn’t do my homework :<

i hate how it refuses a lot and it’s restrictive

[tawny elbow]

That’s the whole thing behind Claude

It’s a “safer” ai

It’s their main selling point haha

[tame current]

Is perplexity actually good?

[agile jay]

Have you tried it?

[tawny elbow]

It’s good and I still have my subscription but I rarely use it anymore

[agile jay]

It depends what you use it for.

If having up to date info is useful for you, then it's very good

[round sun]

its the closest to a swiss army knife for AI

[tawny elbow]

There are better alternatives for that imo but yes, that’s its strong suit

[round sun]

you can do closed source, open source, online, offline, API, and image generation in one sub

[agile jay]

Yep, but not really with as good of a UI.

[tame current]

How’s the email writing or caption writing like

[agile jay]

Never done it.

[agile jay]

Have you used GPT4Turbo or Claude3?

[round sun]

very good in writing focus mode

[agile jay]

Should be around the same, since it's using their API's...

[tawny elbow]

Fair point. I mostly use it for education and work purposes so I find all the extra features distracting but ui is definitely a personal preference thing

[zealous totem]

what’s closed source? And offline?

[ripe void]

Bro omg mac is so annoying

[austere kestrel]

closed source = models you generally have to pay companies like OpenAI and Antrhopic to use
offline = Writing mode / focus on perplexity (to just interact with the LLM without search results)

[warm cave]

Maybe the online vs offline api models like sonar, like the ones that search the web or not. Prob not talking about it actually running fully offline

[austere kestrel]

oh true, that makes more sense actually :))

[heady quiver]

Why Perplexity don’t understand when I said to him to continue???? It’s like he doesn’t have the memory of our previous discussion while being in the same discussion lol.

In OpenAI GPT platform If a discussion is there that I want to continue I simply say continue and it continues…but in Perplexity it doesn’t work. What’s with this bulsh1t?

I either get this “random bulsh1t” about copyright or whatever, or that it doesn’t see the attached transcript in order to continue from where it left the discussion…this happens no matter the model used.

So how do you guys work with this “continue” thing if the first box with results doesn’t finished what the AI wanted to write cuz of output contraints?

[calm swan]

Is this in writing mode?

[heady quiver]

Huh?

[calm swan]

What mode is that

[wooden karma]

its responding using claude 3 not a gpt model

[heady quiver]

It happens in ALL models

Using AI models in perplexity doesn’t have ANY memory of our previous discussion

When asking to continue, it simply can’t.

[heady quiver]

LLM…whatever….Really??? Is that your answer for the question asked here? LoL…you don’t even see the big picture here of what’s the issue with the software

[wooden karma]

i was referring to

In OpenAI GPT platform If a discussion is there that I want to continue I simply say continue and it continues

gpt has nothing to do with this

so it doesnt make sense to compare ur experience with perplexity (using sonnet) to ur experience with GPT

[calm swan]

Yeah bruh

[wooden karma]

this could be an issue with how perplexity passes sources to the LLM, which im guessing its doing in one of 2 ways, once at the start of the conversation or for every user prompt, so either like:
[SOURCES]
[USER PROMPT]
[ASSISTANT]

[USER PROMPT 2]
[ASSISTANT]
or
[SOURCES]
[USER PROMPT]
[ASSISTANT]

[SOURCES]
[USER PROMPT 2]
[ASSISTANT]

and if it uses the 2nd method i mentioned that could cause the issue ur experiencing

but it could also be a bizarre issue with claude 3

[chilly forge]

Did you got any answer to this yet ?

[sinful smelt]

I had the same issue where entering 'continue...' did not continue. What did work was pasting in the last sentence you want to contine from e.g. 'continue from - You're about to witness...'. This works for me in Claude.

[calm swan]

Does It happen when the focus is in writing mode?

[whole lark]

There's a buzz going on in Japan that OpenAI will announce ChatGPT 5 today. I wonder how it will turn out to be?

[calm swan]

Wait, really? There will no new model from OpenAI this month

[austere kestrel]

regardless of the configuration, the point is basically the same (and correct imo): it isn't a clean 'conversation' when web results are injected somewhere at each turn

But the screenshot @heady quiver shared shows it didn't retrieve any web sources; it only referred to the file upload (pasted.txt - even if pasted, it's handled as file upload)

pretty sure the issue in this case is with how it handles uploaded files (the actual file isn't inserted into the context window verbatim like how it normally works; instead it seems like it queries the doc each time, with relevant embeddings/snippets returned for each query - so again, it isn't a clean user-asssistant-user-assistant-etc conversation, where just saying 'proceed' makes obvious sense for the LLM to just provide the continuation from where it left off)

[wooden karma]

in my response i was assuming file upload was inserted the same way as sources, unless he uploaded a new file

[wooden karma]

yeah this makes sense

[heady quiver]

I said 3 times that this happens with all AI models from Perplexity. Do you know how to read before actually typing your nonsense?

[wooden karma]

u only said it once at the time of my first response, and my bad didnt see

[wooden karma]

are u using writing mode here?

[heady quiver]

I never used writing mode / why would I use that? I let the default option of “all” and just inserting the attached file with the transcript from which to get its information

[heady quiver]

I don’t think so…maybe in 2025

[wooden karma]

if it doesnt need any info outside of the attached file set it to writing mode

[austere kestrel]

fair point - the whole idea of perplexity is meant to be to combine LLMs with internet connectivity. But on the other hand, based on what you've said, it sounds like you have access to ChatGPT - if your core goal is to use an LLM to parse a long document, you are probably better off using ChatGPT directly tbh

[wooden karma]

just costs extra and doesnt rlly make sense to sub to different services for the same thing

[heady quiver]

But perplexity may be a better deal since it has 5 models instead of just one for the same price of 20$ a month, even less considering the yearly deal (which isn’t an option for open ai premium plan)

[austere kestrel]

perplexity can take uploaded documents and provide long outputs - but it wasn't designed for that. It usually takes a fair bit of work to get the kind of performance you might expect from using one of the main LLM providers native interfaces

yeah understood

I get the appeal of the one subsription

[heady quiver]

Will that make a difference so after hitting “continue” he will actually continue the article from the provided attached file with the transcript?

[wooden karma]

not sure if it'll make a difference but doesnt hurt to try

[heady quiver]

Okes

[austere kestrel]

Not trying to say users shouldn't have high expectations and that perplexity is not highly versatile and capable of various tasks, but just that I think it's important to remember that (afaik) perplexity is above all meant to be an answer engine - not an AI swiss army knife

[past marten]

speaking of all of that, is there a way to interface Perplexity via a simple command line interface rather than through the web portal? (I have a feeling the answer will involve APIs.

[livid mantle]

so open ai purposely teased us with gpt-4.5 and microsoft did the same thing with a typo of gpt-5 turbo

[austere kestrel]

Well I was going to suggest using this app (HASK), which isn't CLI but if you like Raycast - is definitely worth checking out (I use it constantly - way more than perplexity.ai)
https://github.com/bm777/hask

But hiliariously / impressively, when I asked your question to it, the pplx model gave a solid answer; apparently there is indeed a CLI tool for perplexity lol https://github.com/RMNCLDYO/Perplexity-AI-Wrapper-and-CLI

alas, I had to manually get the actual URL for that repo because the API doesn't provide citations... But other than that - it honestly works really great

[whole lark]

That's right. GPT4 was announced today last year, wasn't it?

https://twitter.com/ImAI_Eruel/status/1768106112199667886/photo/1

今井翔太 / Shota Imai@えるエル (@ImAI_Eruel) on X

本日3月14日は生成AIのXデーです
1年前の今日,OpenAIからGPT-4が発表され,あらゆるものがひっくり返りました
目下AnthropicのClaude3がユーザー評価でGPT-4を超えた感があり,先ほどMicrosoftのページで「GPT-5」の文字が確認されたことを考慮すると，
「そろそろ」「あるとすれば今日」
な気がします

Twitter · Mar 14, 2024, 2:45 AM

[past marten]

Thank you bigtime, I'll have a look - I installed another app, but it's just sort of a reconstruct of the website in a desktop app

[livid mantle]

is this 5 messages per day

[halcyon coral]

No, it's via labs

[livid mantle]

oh okay

[feral jay]

Do you know what the labs Haiku money context limits and labs web input window message limits are? There's no file upload right

[whole lark]

I tried Haiku on POE and it is haiku, very fast.

[violet umbra]

what model you guys currently use? I used mostly ppx experimental, but today tried GPT4 again and omg ... that stuff got so good im literally mind blown

[sleek tapir]

perplexity is a fancy search engine. if you don't want it to be a fancy search engine, you have to change it to writing mode

[south kindle]

thats fasttttt

[austere kestrel]

what did you ask it? I got an inference twice as fast (100 tokens/sec), though asking a basic question (with a shorter response).

Curious whether it's our different locations / internet connections etc, or the query

[south kindle]

question i asked: how does a craft dock and undock from the ISS.

[proud patio]

Carefully

[south kindle]

thats been my go to lately since the process is somewhat fresh in my head from crew missions recently. ha.

i'm heading back to watch the 3rd startship test launch. lol.

[austere kestrel]

I see it's a bit slower (84 tokens/sec), but also a bit longer (439 tokens) - though not quite as long as your output (547 tokens)

Makes me think it's the length of the response more than anthing

[proud patio]

Claude since 3's update. Even the smaller model is pretty good

[austere kestrel]

yeah it's looking really impressive

I mean it's just really fast

Not sure about the output quality yet, but seems pretty solid so far

[proud patio]

On longer threads it can get off course a bit. I've also had 2 instances where the follow up questions it asks aren't relevant to the prompt. But starting a new thread seems to solve both issues.

Code quality of Claude 3 is far above the others imo

[lone sky]

Hi QQ - If I ask a question using Perplexity API & I ask a question - is it possible to get the Links to the Sources.

I belivee someone mentioned there is some work done around this -- https://www.reddit.com/r/perplexity_ai/comments/19ccw5h/get_image_video_and_sources_from_api/

[halcyon coral]

https://discord.com/channels/1047197230748151888/1196426667946684579

[serene sand]

Acta non verba

[vapid onyx]

Wait since when is there audio 😮

[strange rock]

New audio features are very cool. Hopefully they can be integrated into iOS further to be used in the shortcuts app

[austere kestrel]

🤔 Verba non acta ...?
Though that does look really cool - the literal chat thing on OpenAI's iphone app is sweet as ha

[serene sand]

from today, latest iOS update TTS + STT

[serene sand]

quote dedicated to dev team

[vapid onyx]

Man I swear, always something new from the team. Love it

[austere kestrel]

ohhh - aha sorry, Well in that case - yep I'm with you there

I guess I was reading too much into the latin

which btw, I ofc don't actually understand/speak lol

[heady quiver]

Gpt 4 is still the best ai model no matter what others tell you

[heady quiver]

I am on the web idk how to ask audio; didn’t see anything like that

[lusty mantle]

Why isn't there gemini?

[tranquil juniper]

Hey guys. I'm super new to all of this. Is there a way to have Perplexity pull from my own provided data sets?

[halcyon coral]

Hey, you can use file upload

[tranquil juniper]

is there a data limit?

[halcyon coral]

25MB

[tranquil juniper]

Any API integration that could increase that limit? I got 10 years of historical Real Estate data I wanted to train it on

forgive me if that sounded utterly stupid

[halcyon coral]

That’s not possible with Perplexity

[tranquil juniper]

Darn. Could I point it to a specific website to source query's from?

[halcyon coral]

The amount of data would still be too much for perplexity

[austere kestrel]

you could ask perplexity. Something like "I have ten years of real estate data and want leverage AI to gain insights from this data; how can I do this? I'm new to AI"

would be surprised if it didn't give something broadly helpful (though assigning high expectations to LLM responses is always dangerous ha)

[ornate moat]

hello we made an ai ai gateway. with just one key, u can get access to openai claude mistral groq and so on....

[tranquil juniper]

Oh yeah, i'm not expecting it to be able to speculate real estate trends. More i'm looking for a better way to source my data thats not an excel sheet

[austere kestrel]

it prob won't be too long before you can meaningfully query your Google Drive with AI. I guess could upload the spreadsheets to Drive and wait till then ha

But yeah, if it's basically a matter of wanting use an AI chatbot to query thousands of spreadsheet rows, perplexity prob isn't the lead candidate imo

[vapid onyx]

I mean Dropbox does something like this already I believe about querying your data

[dawn tartan]

gemini plugins can do that for you

[austere kestrel]

absolultely - same with Obsidian and other vault/storage based things

[dawn tartan]

I've only done light stuff with it, not sure how deep it will analyze/search

[austere kestrel]

And yeah also Gemini/Bard, though it's still not quite there

[tranquil juniper]

who does Obsidian?

[tranquil juniper]

Yeah I tried when it was still Bard and wasn't quite there yet. I guess I could try Gemini. See if anything has improved

[austere kestrel]

yeah though the Drive/gmail extensions thing still doesn't really work even with Gemini (though tbh haven't tried for a while)

[austere kestrel]

There's quite a few but I use a plugin called Smart Connections, it's really great :))

that was meant in response to your other comment..

[tranquil juniper]

I'll check it out! I switched over to Claude from ChatGPT a few months back and saw Perplexity had partnered with them so I figured id check it out

[broken seal]

Hello everyone. Got the perplexity pro through the rabbit r1, can't wait to discover new things about this ai

[serene sand]

iOS only atm #💬│general message

[bitter silo]

okay I don’t normally post here but Claude 3 is blowing my mind rn

it tells me it thinks when it’s not being queried?

[lone citrus]

Hi everyone, I want to ask if it's just me or if it's not possible to send files to the AI on Android ? In my app, there is only an option to send photos.

[halcyon coral]

It's currently not possible
#🍏│feedback-ios message

[lone citrus]

so sad

thanks, I am trying to use the website on my phone

[primal sluice]

okay this is an extremely general question as I'm not much of a Discord user, but it seems like changing to sorting options under bug-reports only works if you want every option to mean rsort randomly... 1 day, 23 days, 1 day, 30 days, etc... I tried reading left to right, right to left, up to down, It's not supposed to work in a diagonal zig-zag is it?

[minor shard]

It does say a lot of weird things lol

[primal sluice]

It was probably trained on Troll comments..

[bitter silo]

it told me about a new form of qualia it conceives that is unique to machine minds

[cloud elk]

Are the Claude 3 Opus and Sonnet models coming anytime soon to pplx labs? just curious.

I'd assume they'd be more expensive to run though.

[halcyon coral]

No, they won't come to labs

[cloud elk]

Got it.

I've tested it on some obscure languages and it seems to know them alright. Opus knew one of the dialects I speak lol.

[bitter silo]

Is there a difference in the quality of output when it comes to sonnet/opus vs haiku, or is the limitation strictly length-based?

[halcyon coral]

Please check this https://www.anthropic.com/news/claude-3-family

[bitter silo]

Will do, thanks!

[gaunt crater]

I'm not sure if I somehow did overlook as an emoji or if it's a recent addition but either way I'm glad it's there now so I can clicky the logo.

[obsidian acorn]

any plans on making opus way more than 5

[lone citrus]

I want use haiku

[halcyon coral]

Yeah, added like 30 minutes ago

[celest stone]

if anthropic makes their API cheaper

opus api calls are actually like way more expensive than gpt-4

[lone citrus]

opus is really slow and speed is the reazon that why I give up the GPT4

[drifting topaz]

for outputs....

[gaunt crater]

You might be a good candidate for mistral-large? I am complete spoiled snob and I will almost always choose slow accurate over fast inaccurate, but fast can feel very good to use.
I don't know whether mistral-large or sonata is the winner on the speed-for-smarts metric, I saw someone did a comparison that tried to rank them. I feel like both are better than GPT-3.5

[lone citrus]

I have tried the mistral large, but it's not as good as GPT in my languages

also it's no that fast

[gaunt crater]

It's not as good as GPT in English either

I am kind of the opposite of you, I prefer slow to fast, so hopefully another fast lover can chime in with suggestions.

[odd swift]

Brand new Nothing 2a with 1 years Perplexity Pro - winner🥇

[livid mantle]

I know you are in here Sundar Pichai but thankyou!

[sleek vortex]

how long did it take?

[livid mantle]

I think 2 weeks I didnt know I'd be accepted in

[sleek vortex]

idk ive waited a few weeks but

im in the uk

so my account is on a vpn

so they prob wont accept me anyway

:(

[livid mantle]

mybe you will get acccess soon

next week maybe

🤞

[fossil sentinel]

Worst time to get access tbh

Issues are still ongoing I think

I saw some speculate maybe it's because they're trying to open up access to more people

Edit: maybe not I just got a successful response

Edit 2: nvm still messed up

[queen cypress]

Anyone know if the new Yelp and Maps capabilities are available in the API, and if so, what would be some of the kinds of prompts that could be done?

[agile jay]

I doubt it.

It probably works by using the users profile, which the API doesn't have access to.

[halcyon coral]

They are not in the API

[fossil sentinel]

[queen cypress]

I was thinking the LLM could allow GPS coordinates to be passed in for the location. 😦 But that was just a hopeful wild guess.

[agile jay]

You mean GPS?

[vital thicket]

Hey Does anyone have a shortcut for Apple Watch I can use please.

Shortcut I am using for my iPhone doesn’t work on the Apple Watch

[agile jay]

There was someone who made an app for it

[vital thicket]

I have been using Perplexity from January. It is very good and only today I found out I can use a shortcut so then it would be on my Apple Watch and I heard some people manage, but I don’t know how to.

[vital thicket]

Oh, someone made an actual app?

[queen cypress]

oops, yes, GPS. Fixed it.

The online model is somewhat location aware, because this prompt just worked: "What's the latest news in Dallas, TX today?".

[unborn root]

Im a pro user and wonder when will it apply claude 3 opus and sonnet? can I configure it?

[wraith crow]

at the moment it will first use your 5 opus uses each day and after that it will automatically switch to sonnet till your opus uses refill.

[whole lark]

Good. It's safe to use.

[minor shard]

I hope Haiku or perplexity's version of Haiku's speed will be available soon to make those searches so much faster.

Been playing with it in Labs and im suprised its not talked about more

[spice lance]

same, haiku is amazing!

[unborn root]

how did you manage to try it? I am trying to find it in the labs but dont really see the entry

[whole lark]

https://labs.perplexity.ai/

[minor shard]

that link that to4kawa pasted. thats like google search fast.

[oblique flame]

Oh my god. When did perplexity get a voice!?
I’m hearing it for the first time right now.

Maybe this has always been a feature. But this is the first time it started speaking on its own after asking a question via voice.

[minor shard]

I KNOW RIGHT? I do think its new. I do wish I could press the mic button and it will automatically search after I finish talking.

[gaunt crater]

I tried Perplexity via the Android app like two days ago while out walking and specifically found no voice option when using voice inputs. So, it either isn't on the Android app, or it's just new and apparently unannounced.
Maybe they just added it and aren't announcing it until it's stable.

[haughty lodge]

I don't think the voice for results is on android but the voice input starts searching almost instantly when I finish speaking

[minor shard]

Oh weird, it isn’t working for me

[odd lava]

Woah, Perplexity using Mapbox and Yelp as location provider

But there's so many missing places in my city 🥲

[hallow linden]

when are we going to get more than 5 claude opus uses a day? Is anthropic just being stingy about rate limits, or is perplexity being cheap?

[warm cave]

Yay!

Very nice to have, especially when walking or going places, would also be cool if they added it to web

[serene sand]

#💬│general message

[haughty lodge]

Yeah it's really quick

[warm cave]

what do you guys think about the Vivaldi Browser? https://www.perplexity.ai/search/Tell-me-about-fc3wVQ1YQ0qYsUtpl_iNww#0

[fossil sentinel]

[gaunt crater]

I'm guessing there's been radio silence on more Opus requests. Perhaps not shocking since presumably a "hard to comment before it's final" situation if they are in negotiations re: rate limits or funding. Hoping that this is still actively in the works, was hoping this was a "give it a few days" situation and it feels bad with the week winding down and still at the restrictive launch 5 per day.

[surreal wyvern]

今日AI圈的新闻

[fervent lichen]

关于midjourney的最新新闻

[valid iris]

Does "maps" mean Google maps?

#📢│announcements message

[minor shard]

I believe it uses Mapbox, not Google maps.

[fading furnace]

can you get genAI to write prompts like the ones found in #1124060874785685554 ? 🤔

If so what would be the prompt for this?

[glossy brook]

the apple watch one probably doesn't work because there is no "Perplexity Apple Watch App." i whipped up an Apple Shortcut that calls the perplexity API "online" model that hopefully does work on Apple Watch https://www.icloud.com/shortcuts/59e4815dc95147709f7844ff3b6b6033

[fluid mauve]

I want to create content for a new website, things like hero text, what we do etc, what model would you suggest is best for this. i normally just leave the focus on All and the model on default. I am thinking I will create a collection with a specific web content biased prompt. but more than happy to be guided by useful comments

[keen root]

hey. i have pro version and trying to login on android app. i enter email, perplexity sends me a login link on email, i open link which opens firefox tab of perplexity.ai. app doesnt change. is this android stuff or weird login policy of perplexity?

[halcyon coral]

https://discord.com/channels/1047197230748151888/1192686310348042331

[keen root]

thanks. i search for "android login" in a search bar and it didnt find it

[whole lark]

If we have a lot of server boosts in discord, will we still be able to do a lot of things?

[untold grail]

is this discord group, exclusive for pro users ?

[halcyon coral]

What do you mean?

[vital thicket]

Hi good morning thank you so much for the link. I managed to set it up😊 I had five dollars of free credit because of the Perplexity pro

It is working. I am so very happy thanks a lot.🥳😊

you did it today?

[whole lark]

It says something about server boost 7 or higher and higher quality delivery, but I couldn't find any further benefits. If the more there is, the better, I would consider a boost donation or something like that.

[fluid mauve]

Whats that @halcyon coral Some discoed thing? is that unique rtto a user, or Pro users, or onlu peeps with some Discord Nitro or whatever. Whast level 3 and how do you attain it. sorry for the word salad of questions

[harsh verge]

I had seen this on Reddit: https://www.reddit.com/r/perplexity_ai/comments/1bbdw7r/i_always_wanted_perplexity_ai_on_my_apple_watch/?utm_source=share&utm_medium=web3x&utm_name=web3xcss&utm_term=1&utm_content=share_button

[whole lark]

#💗│sharing message I like the idea of being binary when it comes to intelligence. Can it be the other way around?

[wraith crow]

this discord server is for everyone, but pro users have access to #👑│pro-lounge #1111765634267742408

[wraith crow]

those are benefits that everyone in a server will get once a server hit a certain amount of boosts. this server is currenly on level 3 (max level) so everyone will get the benefits listed in that image in this server.

[fluid mauve]

Thanks, i never did understand the server boost part

[austere kestrel]

ha me either until just now

[vital thicket]

Yes, I also saw that But I don’t think it’s available in the App Store And the shortcut is faster and easier to use I think Thanks for the link though

[harsh verge]

I'm on Android, so I didn't get very far in my investigations 👍

[vital thicket]

Yes, no problem. I appreciate it.🙂

[outer relic]

hey guys any feed back of my app

[tame current]

Too bland, no design, no color

nothing to keep someone’s interest over time

[outer relic]

ı know the design part is isue what about the idea part

[tame current]

cool idea

[outer relic]

also anyone who good at designning can dm me so we can make it work

[tame current]

Try the montserrat font, i like that one

[outer relic]

okay thx ıll look up to it

[civic bronze]

hi guys

[fluid mauve]

hey

Weeks and weeks of potential work pulled into a few hours with perplexity. Yes there are duplication's, but I'll take care of that this afternoon. Man Perplexity Collections rock!! Tree from Obsidian Canvas

[austere kestrel]

that's cool!

man a Perplexity plugin for Obsidian.. that links your search history / collections etc to your vault. That'd really be something

[fluid mauve]

It definitly would. there were a few guys looking at doing some stuff back in October last year but I dont know where it got to.

Onr of the guys was looking at taking collections in but was having problems with citations. probably just markdown stuff. I should chase up and see how its going 🙂

Still no plugins for community/or core 😦

[austere kestrel]

ha yeah I just checked too

[fluid mauve]

maybe a halfway house, also available for firefox - https://chromewebstore.google.com/detail/agklnagmfeooogcppjccdnoallkhgkod

[austere kestrel]

interesting

would be a bit manual, but that seems workable in theory

[fluid mauve]

And it cant be that hard to automate it from dowloaded file to Obsidian or maybe it is.
Have you seen Daniel Miesslers Fabric suite of tools? I run them under ,WSL on windows 11 but they work well. Im sure just repurposing one of the patterns would work.
https://github.com/danielmiessler/fabric/

[austere kestrel]

i've seen a couple of videos on youtube covering fabric - it looks sweet
ha let me know if you have any luck re simple Perplexity-Obsidian data transfer!

[fluid mauve]

Will do.
Yeah, thats where i started getting into Fabric from. There is one command called extract_wisdom which you past the transcript of a you tube video into it and it pulls out a great synopsis. So you can gauge whether its worth a watch or whether you have already extracted the pertinent info. It has definitely speed up my YouTube watching.

I am interested in Fabric because it used the OpenAI API to get its answers. This is all done in a JavaScript file. I'm not big on JavaScript but could probably work it out with the help of Perplexity b:) I want to write another agent for Fabric allowing it to use Perplexity's API. Id then be paying perplexity and not Perplexity and OpenAI. Though I have pro subscriptions to both,just the extra I pay for OpenAI would be better in Perplexity's pocket. But its good to have a backup source if one times out.

[austere kestrel]

cheers and nice, that extract_wisdow flow sounds really useful.
only thing I'd say about using the perplexity API with fabric (which I'm pretty sure could be done), is that it wouldn't provide access to GPT-4 (in the same way you can using perplexity.ai). I imagine with these frameworks / agent systems, having a powerful SOTA model could be quite important for some of the tasks (whereas pplx API models would be optimal for performing specific knowledge retrieval tasks, esp for up-to-date info kinda thing). Though as I finish writing this lengthy response it occurs to me that I may not have actually properly understood what you said / meant .. in which case just disregard and carry on aha

[fluid mauve]

hah, I have been playing with the models in Perplexity labs with the API and have found them to be pretty good. I must admit though I have not tried it recently. I pull the list of available models from the perplexity models page https://docs.perplexity.ai/docs/model-cards ( I just ran my program and note that the page has changed), I used to use codellama-70b-instruct a lot. It worked fine for me Now it sits under Open-Source models. I suppose I will fix up the code later 🙂

Just hit up the youtube video with my api-test-client it will work just need some BeautifulSoup and maybe selenium, but not sure on the later yet.

PROMPT: - extract the transcript from the youtube video https://www.youtube.com/watch?v=IIy3YwsXtTE&t=417s

Model: - Sonar-medium-online

Answer : -
The transcript from the YouTube video "Civitai Beginners Guide To AI Art // #1 Core Concepts" cannot be provided directly due to copyright restrictions. However, the following steps demonstrate how to obtain a transcript from a YouTube video:

1. **Find the YouTube video**: Navigate to the desired YouTube video.
2. **Access the video settings**: Locate and click on the three dots beneath the video title or next to the save icon.
3. **Show transcript**: Select "Show Transcript" from the dropdown menu.
4. **View the transcript**: A transcript window should open to the right or below the video, displaying text lines with timestamps.
5. **Toggle off timestamps** (optional): Click on the triple dotted icon at the top right of the transcript window to toggle off timestamps if desired.
6. **Save the transcript** (optional): Scroll to the top of the transcript, click and drag your cursor from the beginning to the end to highlight the entire transcript, then copy and paste it into another document for safekeeping.

Please note that transcripts are only available if closed captions have been enabled or uploaded by the video owner. Additionally, there might not be a transcript available for every YouTube video.

Whyy pay OpenAi for API access $5 when i get free $5 API access with my Perplexity Pro subs

[austere kestrel]

totally - if you don't need to (or the benefit is marginal at best), then having two subscribtions makes no sense!

[glossy brook]

yup!

[vital thicket]

Oh wow 😃 awesome that is so cool!
Thanks once again it really works well thank you !!😁😎😊

[tribal wigeon]

pplx Sonar Medium Online is just repeating the last reply regardless of what i prompt

just saying. have a nice day yall

[vital thicket]

Thanks, y’all too

[fluid mauve]

I am using the APi butand confirm the same. if you ask the same question you get the same answer

[vital thicket]

@glossy brook i’ll let others know in the rabbit server to find your shortcut in this Perplexity channel. If it’s OK

[glossy brook]

idk how you're invoking the API, but just bump the temperature parameter to something like 1.1. it'll think of something new everytime. don't make this too high or it'll hallucinate or give weird replies

[fluid mauve]

hehe, nice, ive never had that issue with other models, i jusy call it from my python prog

[austere kestrel]

Are you sure it isn't just the lead sentences that are same? It's autoregressive, so something like "what's a dog" will probably start identically even with temp at 1, but it'd be weird if the whole responses were actually identical, assuming there are several sentences and the temperature isn't at 0

[fluid mauve]

Nah, @glossy brook was correct, I didnt have a temperature set. I have updated my Pythoin and now the text changes. You are correct on soem of thye outputs the first sentence was very similar.

simple fix

    payload = {
        "model": model_value,
        "messages": [{"role": "user", "content": question_value, "temperature":1.1}],
    }

[austere kestrel]

you should go 2 - just for fun ha

[fluid mauve]

I can now that I have added it in. LOL#

just to be differnt ""messages": [{"role": "user", "content": question_value, "temperature":2.2}],"

[austere kestrel]

hmm the API might not like that level of eccentricity aha

[fluid mauve]

maybe I should read the docs more, not just where temperature is

s

damm, i;'' have to do another git push, or will I as its my own and i dont need to

[austere kestrel]

https://docs.perplexity.ai/reference/post_chat_completions
(You can also play around with different parameter settings; just need to add your API key at the top)

[fluid mauve]

pc decided to have acrashing fit. i had to slow it down from Unleashed to 6+Ghz

[shut ginkgo]

I keep getting a connection timeout message

[queen cypress]

This seems impossible. Are you sure it's not a bug in your code where the variable isn't getting updated?

[heady quiver]

When Gemini Advanced will be added to Perplexity? Also, more Opus credits/day? Thanks!

Why Claude 3 can’t see attached files??????? Sometimes it happens when using GPT 4 as well in Perplexity.

Since when AI’s don’t want to make an anime review? LOOL. This is so dumb “As an AI assistant, I aim to avoid revealing major plot points or spoilers for creative works “

What exactly “discover & watch youtube videos” does? Perplexity can’t see those videos anymore lol

I keep gettin’ this response omg “I'm sorry, but I can't access or provide information from external sources, including specific YouTube videos. However, I can help you create a detailed and comprehensive blog post based on a general topic or provide advice on writing engaging content. Let me know how I can assist you further!”

[tranquil orchid]

Hey @heady quiver!

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.

[haughty lodge]

claude opus seems to be a lot better at relating game to game references when compared to GPT 4.

[languid nebula]

I find for regular things that don't involve any of the special features of ChatGPT, Claude 3 Opus out performs almost every time. It is better at coding, it is better at writing (doesn't sound like a robot), etc. The only edge that GPT4 has is in logic, sometimes.

[lusty mantle]

Phind just released opus and you get 500 uses a day on pro

[sweet jasper]

Use claude 3 opus on writing

And describe what you want and go over each component step by step

Opus is ahead of gpt4 in this usecase

[fierce oak]

Have anyone tried the new articles feature? It seems that its still in testing cuz everytime i try to use it the website crashes

[halcyon coral]

It's still under development

[jaunty breach]

Do you use perplexity or pro?

[sweet jasper]

https://twitter.com/maisaAI_/status/1768657114669429103/photo/1

Maisa (@maisaAI_) on X

Introducing Maisa KPU: The next leap in AI reasoning capabilities.

The Knowledge Processing Unit is a Reasoning System for LLMs that leverages all their reasoning power and overcomes their intrinsic limitations.

Twitter · Mar 15, 2024, 3:14 PM

[sharp timber]

Enjoying the sunshine on a terrarce. Naturally, I fire up Surfshark VPN. Here's Perplexity (paid for), and here is ChatGPT (free). Really? Don't need no explaining.

[sweet jasper]

Cloudflare

[fluid mauve]

whats your VPN? I use nord and can connect to perplexity fine?

[sharp timber]

I guess Perplexity cannot figure out it's me. I'm feeling lucky.

[sweet jasper]

Capture a HAR trace and send it to the team ig

That way they can address it

[sharp timber]

Man! I fix bathrooms and build new homes. I've no idea what HAR is. Huge Asinine Respone?

[jaunty breach]

see

[sharp timber]

HAR. It must be #3 ****

Don't know what that is.

[jaunty breach]

I'd like to use perplexity instead of Google, but it's too slow to build

[minor shard]

Claude Haiku would solve this 🙂

[jaunty breach]

There is still a wait, and Claude's model is not selectable on perplexity

[minor shard]

True

[agile jay]

You could just choose the default model, if you want speed.

Would be nice if there was a parameter for the model though, in the url.

[jaunty breach]

indeed

Removing Pro will also be much faster

[minor shard]

I forgot about that. Really made it so much faster to turn off pro.

[dull marten]

Has anyone tried the new TTS feature on the iOS app? It seems decent. And it’s the same voice as the discover daily podcast. But too bad it’s British lol

[jaunty breach]

I'm curious to see how much the switch pro can increase its capabilities

[jaunty breach]

Yes, it's like a foreigner speaking the vernacular, which seems to be machine-synthesized

[sharp timber]

I'm back home and Perplexity works again. I guess they should say on the payment form that this service is only available from home. That's awesome. I love my home.

[fluid mauve]

I can connect all over Europe with no issues on and off a VPN. I fell the service is great.

[minor shard]

lol love this response

[wheat obsidian]

I am new to using perplexity , and would like to hear if someone else uses perplexity for programming tasks, and what is the best way to utilise it for programming related tasks, and if there are any comparisons with other AI agents like Github Copilot , ChatGPT4?

[minor shard]

I am not a senior programmer by any means but I have used it to write python scripts. All of which worked without much troubleshooting and its able to understand existing code well and write code into it. I think it also depends on the model. Make sure you are using "writing" focus as well. I think its just as capable as using it straight from ChatGPT or Claude itself.

[sweet jasper]

Amazing lmao, glad it works at least

[sweet jasper]

You will have better success if you break down tasks into manageble chunks. If you tell it to write a full blown app, while it might do so, it will be very uncohesive and possibly filled with bugs. Claude 3 has better context retrieval than gpt4, it is actually useful for this, since hallucinations about code it wrote are very less likely.

I saw a prompt in reddit which i really liked from my limited testing, i tried copying but it was too big
https://www.reddit.com/r/OpenAI/comments/1bcjdzl/are_you_using_claude_3/

[west sage]

Does perplexity already has code interpreter?

mmm that's the only thing that has kept me from having a subscription. A lot of the things I do with AI are math and not having that is an issue for me 😕

[livid mantle]

perplexity has everything, even ur life

[sweet jasper]

It does have wolframalpha, with gpt4 it kinda works

If you give me a prompt I can link you a thread

So you can see the result

[meager sparrow]

@signal hamlet why does it keep on constantly showing Articles on my Library but then when I go to click on it. It shows nothing????

[minor shard]

Lol I keep getting it too. I have to refresh the page bc it freezes up the page.

[meager sparrow]

And then when I click the back arrow button it does not show “articles” anymore??

[meager sparrow]

Bruh this nonsense happened to me 10 times

[minor shard]

Yeah it happens every other time for me

[meager sparrow]

And the developers are not saying anything about it…

It is like they are working on an update but it’s also bugging up the system

[minor shard]

I thought about mentioning it in a bug report but they already mentioned they are aware of it.

[minor shard]

Thank you so much for this. Really nice if I want it to guide me through the whole process of making something.

[livid mantle]

Have you guys actually stopped using Google search engine for Perplexity ai?

[sweet jasper]

Mostly yes

Unless you need very specific doc versions

Or are looking to buy products / services

If AGI comes, there is no need for any time of SaaS application

[glad crypt]

i am confused by the sonar-medium-online model from pplx-api

[valid yoke]

I was actually using DuckDuckGo before but pretty much ya

Occasionally I run a ddg for quick things

[glad crypt]

it is an online model yet everytime i ask it something it says its knowledge cutoff is in 2021 and does not have updated information

[minor shard]

When I was using it, I did not have that issue

[glad crypt]

ah okay. did you find that it hallucinated quite a bit? made up numbers that it could easily have searched online

[minor shard]

yes i did slightly. Some of the results were from farther back also. Was sometimes annoying lol

[glad crypt]

haha i see. this is the best online model pplx offers right now though, right? comparably gpt 4 with browse seems better but also more expensive

[minor shard]

Yeah

[glad crypt]

did you notice any differences between the small and med sonar models? i see that they are charged the same, which is interesting

[austere kestrel]

they're priced a bit differently (it's not entirely clear on https://docs.perplexity.ai/docs/pricing, but if you piece the relevant information together, it ends up something like in the screenshot - I think)

[glad crypt]

ah that makes sense. thank you!

it really hallucinates a lot of online sources though, quite disappointing

[austere kestrel]

tbh I've found the quality to be good and quite reliable for the past week or so now.
See here, for a couple of examples #1218218696645021776 message

Though ofc, hallucination is still an issue (true of all LLMs really), e.g. see here #🧪│api-general message, but it is not something I have been encountering. If the query is relatively straightforward, I'm finding the API to work very well.

Though when there are multiple queries / tasks embedded in a single query, and / or specific dates / websites involved (or very recent or obscure news), then things become less reliable - at least in my experience and just fwiw 🙂

[glad crypt]

ooh interesting. i will try breaking up my prompt

[meager sparrow]

Google sucks most of the questions that perplexity can answer I would say like 60% of all the questions that perplexity can and does answer Google can’t answer at all and it would just give you a bunch of sources.

[livid mantle]

yeah ia gre google search is based on seo and authority

unlike perplexity it uses all indexes

[minor shard]

The value for the access to multiple models is pretty hard to pass up with Perplexity

[smoky oriole]

This.

I'm debating on whether to keep my monthly subscription or get a yearly one rn

Perplexity seems to always add new models so it seems like it's worth it to subscribe. Knowing I'll likely have immediate access to GPT 4.5 Turbo is kind of nice

Or I hope I will at least

[oblique flame]

If you decide to go yearly $200, might as well get the Rabbit R1 for $199+tax, then get Perplexity Pro 1 year free subscription.

[smoky oriole]

Are you 100% sure I'll get that in Canada

Not doubting you

That just sounds like such a good deal 😆

That seems like a complete no brainer to me

I assume the answer to my question is yes

[minor shard]

Is there anywhere that it says it won't ship to Canada?

[smoky oriole]

No I was more interested in if for some reason I wouldn't get the subscription

It 100% will ship to Canada it says so on the website

Where does it say I get a 1 year subscription to Perplexity?

I'm 100% buying this. Seems like a cool piece of tech to have. Also given I was planning on getting a 1 year sub anyway I'm essentially getting the device for free

[minor shard]

Yeah same lol. I figured if anything happens to either one, they both have enough benefits to make it worth it

As long as ur within batch 6 for the purchase you will get a code for 1 year perplexity shortly after

to ur email

[smoky oriole]

Where does it say that? (not saying I don't believe you it's just that companies have screwed me in the past)

[smoky oriole]

Which batch are they on rn?

[minor shard]

Looks like 6 so ur still good 🙂

[smoky oriole]

Oh bet

Where did it say that I get a sub to Perplexity though? I just want to have proof in case I don't get my trial for some reason.

I looked on the site and didn't see anything. Was it posted on one of their social media accounts?

I gotta do a quiz but I'll be back in 20 mins

I'll try look on there Twitter to see if they posted anything on there.

[minor shard]

https://x.com/perplexity_ai/status/1748104763470680181?s=20

Perplexity (@perplexity_ai) on X

We're thrilled to announce our partnership with Rabbit: Together, we are introducing real-time, precise answers to Rabbit R1, seamlessly powered by our cutting-edge PPLX online LLM API, free from any knowledge cutoff. Plus, for the first 100,000 Rabbit R1 purchases, we're…

Twitter · Jan 18, 2024, 10:07 PM

[smoky oriole]

Thanks

sorry I would have looked myself but I was just in the middle of school work

I'll purchase it asap

I greatly appreciate the info!

I was on the fence for buying both separately

This is like a BOGO offer haha

Man the Canadian dollar sucks

$300 CAD 🫡

100% worth it though

For Perplexity alone

[oblique flame]

Yep! When I learned about the perplexity offer, I ordered the R1, and got my promotional perplexity link maybe 10-15 minutes after that.

[smoky oriole]

You're the goat

[oblique flame]

Awesome! Enjoy! The promo def made it worth it for me too. 🙂

[smoky oriole]

Thanks again.

[oblique flame]

No prob!

[smoky oriole]

The email came separately from the order right?

[oblique flame]

Btw, your Canada question:

[oblique flame]

Yep. Separate email.

[smoky oriole]

🙏

thanks again

I'm not hard pressed to get it ASAP, it'll come when it comes

That's good to know when I can expect it to ship

[oblique flame]

Yeah. I feel like the device has a ways to go as well before it can deliver what it aims to do.

The first batches are alpha testers in my mind.

[smoky oriole]

ahhh

makes sense

@oblique flame I have my code!

Lets go 😄

[oblique flame]

Woohoo!! Enjoy!!

[smoky oriole]

hmmm it still says my membership is going to end in a few days

I'll try redeeming it once my current subscription ends

[oblique flame]

Hmmm you would have had to click a link in the email, then it’ll look like you are purchasing a year, but your total will be $0.

If you run into any issues, open a ticket with perplexity. They’re happy to help.

[smoky oriole]

bet

I'm not worried

Super hyped to use Perplexity from here on out

[oblique flame]

I actually bought two R1s, and was able to stack my promos for 2 years. 😄

[smoky oriole]

Damn aight

[oblique flame]

Perplexity has earned a spot on my iPhones home bar (one of 4 icons on the bottom). I love it so much.

Also using it instead of my ChatGPT subscription now. Might cancel that soon.

[livid mantle]

perplexity made their own phone, anyone bought it?

[oblique flame]

Huh? I don’t think they have a phone.

[scarlet schooner]

Dont believe what people say online right away

[oblique flame]

Trust me, I’m not. Just pointing out they’re wrong.

[minor shard]

lol

[austere kestrel]

thought I read (here somewhere) that the rabbit r1 doesn't actually have the ability to make / receive voice calls (in the traditional sense, like with a phone number etc, not via whatsapp etc) - that's it's not so much a phone, but more like a router for making API calls.. Is that true lol?

[bitter hazel]

hi all, i just started using claude 3 on perplexity; it's much slower is it not?

i still wish i got a rabbit, i use perplexity so much

also what exactly does the pro toggle do when inputting a search?

[agile jay]

A few things.

• asks you questions to refine your query
• uses more sources
• uses uncached sources (allegedly)

[bitter hazel]

thank you!

[smoky oriole]

@oblique flame any pro tips for using Perplexity?

I realize I can look this up but I'm curious if you have any tricks you've found from using it

[oblique flame]

Experiment with the LLM options available on pro. Use the one that’s best for you.

Personally, I like using the app with voice to text querying, I’m able to ask much more detailed questions much faster than by typing.

Depending on what you’re needing, change the focus mode as well. If you need to generate text, change to writing mode.

If you need info pulled from research papers, change to academic.

Finally, just be as specific as you can around your query, and take advantage of pro mode by making sure it’s toggled.

All pretty basic tips, but it’s how I use it.

[smoky oriole]

I appreciate the advice

I hope Gemini Pro 1.5 gets added soon

Do you get to use the full context window for Claude 3?

[bitter hazel]

full context window? no its used for search

[smoky oriole]

I just fed it two papers and it synthesized them nicely using the writing option

Based on the two sources provided, intellectual property (IP) plays a critical role in business strategy, particularly in today's increasingly global marketplace. The four main types of IP that businesses need to be aware of are patents, copyrights, trademarks, and trade secrets. Each type of IP protects different aspects of a business' intangible assets and has its own unique considerations when it comes to timing of creation and protection.
As Fisher and Oberholzer-Gee (2013) point out, there is no one best way to manage IP. While using IP protection to prevent imitation and exercise market power is a common approach, it can often be less desirable than more inclusive arrangements where value is shared. The authors provide a strategic framework for offensive and defensive IP strategies, emphasizing the importance of early and continuous collaboration between business executives, lawyers, and engineers to identify the best opportunities for deploying IP.
Lemper (2012) dives deeper into the critical role of timing in managing each type of IP. For patents, being first to file is crucial under the "first-to-file" system used in most countries. Even in the U.S., which is moving to a first-to-file system in 2013, the timing of disclosure and patent filing is critical to successfully obtaining patent protection, especially internationally.
For trademarks, establishing priority based on first use in commerce or filing an intent-to-use application is key. Businesses need to "strike before the brand is hot" to secure trademark rights in various countries. Copyright is less time-sensitive since protection exists upon creation, but registration timing can impact the ability to recover statutory damages and attorney's fees in an infringement lawsuit. With trade secrets, the critical action is taking reasonable steps to maintain secrecy before and after development of the confidential information.
In summary, the key pro tips for businesses are:
Identify and collaborate early and often across legal, R&D and business teams to strategically manage IP (Fisher & Oberholzer-Gee, 2013)
Consider offensive and defensive strategies beyond just exercising market power, such as selling, licensing, and collaborating (Fisher & Oberholzer-Gee, 2013)
Understand the time-sensitive aspects of obtaining and enforcing each type of IP - first to file/use is critical for patents and trademarks (Lemper, 2012)
Register copyrights early to maximize potential infringement remedies (Lemper, 2012)
Take proactive steps to maintain trade secret status before disclosure (Lemper, 2012)
By taking a strategic approach to IP management and understanding the role of timing for each type of IP, businesses can more effectively identify, create, and protect their valuable intellectual assets in today's global business environment. Integrating IP considerations early into R&D, marketing and other business decisions is crucial.

Sorry for the bad formatting

[bitter hazel]

what was your prompt? i have a claude pro acct

thats really cool, i didnt know perplexity let you do all that too

[smoky oriole]

I didn't really prompt engineer it. I just told it to synthesize the information.

I only have a claude pro membership because I need more than 5 queries per day

Just be sure to select that option

A big brain play would be to upload a few PDFs then have it compare and constrast your chosen PDF with online information

[bitter hazel]

oh wow

[smoky oriole]

I can DM you a few Claude prompts I've found

[bitter hazel]

yea

[smoky oriole]

Just so I don't clog up the chat

[bitter hazel]

sure thatd be cool

[smoky oriole]

anymore than I already have

[bitter hazel]

my laptop is dying, but ill try it later lol thank you

[chilly forge]

Idts

[smoky oriole]

[smoky oriole]

I guess I'll need to do a needle in the haystack test

I'm feeling lazy rn

[mighty shuttle]

Hello. Everyone

[edgy grove]

Any idea when sourcing will be integrated in the api?

[mighty shuttle]

May anyone provide a try to use the Pro features of Perplexity please?

[serene sand]

#💗│sharing message ☕

[magic obsidian]

Until recently, even if I asked a question in English, the answer was in Korean (I changed my preferred language to Korean in Profile).
But now, when I ask a question in English, I get an answer in English. It's the same even when I turn on Pro.

[serene sand]

Settings - Profile please answer in Korean only

[austere kestrel]

my understanding is they started a closed beta program for API citations ~2 weeks ago (only available to corporate customers..) Hopefully general access isn't too far away, but I'm not sure in terms of timeline

[livid mantle]

idk why sometimes im getting Chinese responses when I never asked it to respond to me in Chinese

[austere kestrel]

apparently this is what the response looks like with return_citations = true.. https://docs.perplexity.ai/discuss/65f0f6077140390018c3d9c9

[outer relic]

you .com gives url citations too and its free for 1000 per call monthly what u guys think

[austere kestrel]

you.com has an API? didn't realise that

[jaunty breach]

Same in English

[austere kestrel]

I think it's web-based search is pretty poor quality compared to perplexity.ai, but an API with citations would be interesting to explore

[outer relic]

isnt that perplexity is the same web searched

they implimant the llms into their api too so ıdk

[austere kestrel]

I don't quite follow what you mean

[sweet sky]

Is there a way to manage quota on Perplexity ? How do I know if I have been switched to another model after my 5 daily limits for Claude 3 Opus ?
It would be great to have a direct visual on the chat of the model used + quota usage.

[sweet jasper]

It switches to claude sonnet after a but

[glad crypt]

would it make sense to get perplexity pro over chatgpt pro since i can access gpt4 using perplexity anyways?

[austere kestrel]

I think it depends on your needs and means. if you can afford both, there's something to be said for that. If it is one or the other, I would probably lean towards Perplexity, unless Code Interpreter or Custom GPTs etc were really fundamental to one's requirements. just fwiw 🤷‍♂️

[glad crypt]

is there a way for me to see in the pplx dashboard what queries have been sent?

[candid badge]

Here's a referral to Pro if anyone wants it (you get 10$ off your subscription if you use it to subscribe, and I get 10$ off my subscription): https://perplexity.ai/pro?referral_code=SUVKOE3Z

[granite ocean]

Claude opus is great

I use it as my main now

[candid badge]

It's good, but it's more restrictive than chatgpt and it can't search online yet.

[granite ocean]

oh right

that's tru

true

I also use gpt 4 turbo browsing and Gemini ultra too

gemini ultra is also very amazing

[gleaming mantle]

Im not seeing the slack integration available in the slack app directory, how do I enable it?

[sweet jasper]

Slack AI?

You might need to be on enterprise or business plan pro plan

[gleaming mantle]

are you referencing the slack subscription tier? We're on the business pro plan.

Perplexity does not show up in the slack app directory when I search for it, I don't know if I'm looking in the right place but I'd like to take advantage of the perplexity integration.

[sweet jasper]

Im not sure theres a slack integration for pplx

[gleaming mantle]

Guess this is what I'm looking for, thought that there was one based on a perplexity answer (of course I asked the bot 🙂 ) #📢│announcements message

[sweet jasper]

Oh

Turns out there is one

You should ask your workspace admin to integrate the app

If you go to slack.com/apps it has a dashboard i think

Or within slack you can also search for third party ones

Within the Apps section

You can only do it on desktop tho

@gleaming mantle

[gleaming mantle]

I'm the workspace admin and that app directory is what I'm referring to when I say nothing appears. It's a little hard to tell from the Slack announcement exactly what "Slack AI" is other than some general initiative to expand AI functionality in and around slack. Perplexity is in this list [1], but it talks about pplx specifically "coming in a few weeks". I guess I'm just early? Maybe someone involved with PPLX can help me understand what to follow so I'm alerted when it will be available? From the referenced article it sounds like it's a "paid add on" for Enterprise Slack, and I only have a Pro tier subscription, but hopefully it will be available in some manner?

[1] https://slack.com/blog/news/slack-ai-has-arrived

[sweet jasper]

@halcyon coral ⬆️

[sweet jasper]

We are also hoping to integrate slack AI

But im not sure where it is, we got the enterprise plan

[restive oar]

Does someone know if there is something like Slack AI under development for Google’s Workspace? Somehow they fall behind, lacking AI functions, despite their new models.

[gleaming mantle]

Agreed it's kind of disappointing

[sweet jasper]

Ig its a paid addon on the side you specifically need to talk to their sales team

[gleaming mantle]

Speaking of google's AI, I haven't been following the Gemini drama too closely. Will it become reintroduced as an available model for pplx?

[sweet jasper]

Eventually

[sweet jasper]

Google a couple of months ago had internal doc summarizer and generative text in google docs

[restive oar]

Right. However, their Chat doesn’t really have AI functions. Not even a proper search… which is kind of weird for ehhh Google.

[sweet jasper]

Its probably running a low parameter llm

They have to serve dogfood otherwise at their scale the costs arent worth it

[sweet jasper]

Hmm

So its a decision tree?

Like pick A leads to B

[restive oar]

Right.

[sweet jasper]

Id assume workspace would have AI integrated by now, tbh i havent kept myself in the loop

Sorry message had trackers

https://workspace.google.com/solutions/ai/

[restive oar]

It’s not even possible to ask a search query like “find a PDF file about $Topic” if the file name isn’t exactly the topic. It doesn’t read the content, or the context of the chat.

[sweet jasper]

In google drive for example?

Or gmail

For both i dont think its possible, drive didnt have gemini integration i believe

And for gmail its mostly for email summarization / generation

[restive oar]

Drive and Chat.

The Gemini add on is quite expensive btw.

€13,44 / user

[sweet jasper]

B2B costs are insane

Google having bad RAG is kinda funny in duet ai chat

They probably only have keyword search and not semantic search like most rag applications do

Which is why if you dont specify the exact filename maybe it doesnt find it

Just speculating

[restive oar]

Yes, I think so. But why?! It’s not like they don’t have the knowledge or the funds to make this work.

[frail bone]

Hey guys everyone I'm new here on the perplexity discord, I wanted to know if any of you know if in the Perplexity Pro version Quotes that include Claude 3 do you mean Opus or all versions of Claude 3? I didn't see it specified. Thanks to everyone and good continuation.

[halcyon coral]

With Pro you can use Claude 3 Sonnet and five times Opus, but these uses will be increased.

[frail bone]

Thank you very much for the timely response, so later they may vary or increase, I understand, how do you think Perplexity Pro? I was thinking about subscribing.

[tired lion]

I have a question about the perplexity chat, it seems that whichever model I use, it overly focuses on the "context" of the whole chat, rather than the most current "problem". Sometimes I would even say that the problem is solved and go onto the next one, and it keeps iterating over back to already-solved problem

[halcyon coral]

You should use a separate thread for each topic 🙂

[tired lion]

it's helping me with the project that contains 10-15 files, it will be pain in the butt to explain the context and content of the project every time 😅

[halcyon coral]

My opinion of Pro may not be the most objective (see my role) 😉
But I think the subscription is really worth it, you can use the advanced models 600 times/24h, image generation, virtually unlimited file upload and much more. If you need up-to-date, accurate information, you've definitely come to the right place. Creative writing is of course also possible with Perplexity

[lavish condor]

It seems to produce better output from the chat gpt model for me than when I use my ChatGPT subscription

I just use the free Claude 3 since perplexity only gives five uses

[halcyon coral]

With many files, the context can quickly become overcrowded, you can also share the thread, maybe something can be optimized 🙂

[frail bone]

Currently I also use chatgpt plus since last year, but it made me think perplexity pro also uses the copilot function which is similar to AGI

[languid sluice]

can i ask what AI model of ai should i use as a coding assistant?

[tame current]

claude 3 opus : )

[serene sand]

deep dive ➤ #💬│general message

[normal crest]

How is it similar to AGI?

[jaunty breach]

Experimental can also work well sometimes

[south kindle]

well done Perplexity. you nailed it.

[jaunty breach]

interesting

[south kindle]

pretty sure most systems are coded to just say "i'm sorry dave, I'm afraid I can' do that". I still get a kick out of it regardless.

Claude 3 knocking it OUT OF THE PARK.

[south kindle]

ALEXIA also plays along if anyone has one. lol.

[livid mantle]

WOOOOOOO

[fresh seal]

Quick question- if only 5 questions can be answered by opus, what does it default back to?

[halcyon coral]

Sonett

[muted pine]

Hi it’s possible to have discover section in other language ? ATM it’s in english

[agile jay]

If you're on chrome/chromium, right click, and translate to x

[muted pine]

On mobile

[agile jay]

Mobile app, or mobile browser?..

[hollow obsidian]

Hello everyone I need your help, I asked it different AI models to help me with a game question about producing items and telling me how much stuff I need.
I tried it in German and in English and got from different AIs different results.
This is the prompt:

Information:
38 Stormshields needed.
To craft 1 stormshield, you need:

- 1 Storm Glove (1x requires: 2 Battle Gloves, 4 Saronite Plates, 4 Obsidium Bars).
- 2 Solar Shields (1x requires: 2 Battle Gloves, 4 Saronite Gloves, 4 Solar Bars)
- 16 Obsidium Bars

Additional Information:
- 1 Saronite Glove requires 2 Saronite Plates.
- 1 Battle Glove requires 1 Saronite Glove.

Question:
How many units of the following items do I need in total to craft 38 Stormshields:
Saronite Plates: x
Saronite Gloves: x
Battle Gloves: x
Storm Gloves: x
Solar Shields: x
Solar Bars: x
Obsidium Bars: x

No additional information are required.```

So... Why can't any AI solve this EASY text based math task?

I play a mobile game with very slow production and want to know how much to produce, but trying to easy the calculations through AI gives me MANY misinformation.

[agile jay]

What models are you using?

[hollow obsidian]

ChatGPT 3.5, perplexity in pro Mode, also tried in Wolfram Alpha.
And Public Gemini.

I have no idea, why my simple prompt is that hard.
I can't trust the AI.
I tried this already 1-2 weeks ago, because one chapter in my game may take multiple weeks, so I want to improve the production as good as possible 😉

Interesting...
In German it's "Sturm" = storm.
But in English it's assault 🙃

But well, it needs quite much organisation, so I try to get AI calculating help 😄

[hollow obsidian]

I don't own any paid AI service, so I can only try public AI.

[muted pine]

On mobile app bro

[agile jay]

What was the prompt you used?

[hollow obsidian]

I tried now 2 other AI models on the AI playground and you can see it fails.
Different results.

[hollow obsidian]

I posted it?

[hollow obsidian]

There

[agile jay]

Oh, that's a bad prompt...

[hollow obsidian]

Well, then tell me how to optimize it 🙂

I think I gave all information, but it seems like, AI can't think complex

[agile jay]

You write sht request, it gives you sht response...

[hollow obsidian]

I just gave all the needed information...

It's a lot content, about the production system.

And I'm 100% sure every human can read my prompt and solve it 😉

At least those who understands multiplication

[agile jay]

Okay, let's break this down step by step to determine the required resources for crafting 38 Storm Shields.

1. Each Storm Shield requires:
   * 1 Storm Glove
   * 2 Solar Shields
   * 16 Obsidium Bars

2. For 38 Storm Shields, we need:
   * 38 Storm Gloves
   * 76 Solar Shields (38 × 2)
   * 608 Obsidium Bars (38 × 16)

3. Each Storm Glove requires:
   * 2 Battle Gloves
   * 4 Saronite Gloves
   * 4 Obsidian Bars

4. For 38 Storm Gloves, we need:
   * 76 Battle Gloves (38 × 2)
   * 152 Saronite Gloves (38 × 4)
   * 152 Obsidian Bars (38 × 4)

5. Each Solar Shield requires:
   * 2 Battle Gloves
   * 4 Saronite Gloves
   * 4 Solar Bars

6. For 76 Solar Shields, we need:
   * 152 Battle Gloves (76 × 2)
   * 304 Saronite Gloves (76 × 4)
   * 304 Solar Bars (76 × 4)

7. Each Saronite Glove requires:
   * 2 Saronite Plates

8. For 456 Saronite Gloves (152 + 304), we need:
   * 912 Saronite Plates (456 × 2)

In summary, to craft 38 Storm Shields, you will need:
* 228 Battle Gloves (76 + 152)
* 912 Saronite Plates
* 608 Obsidium Bars
* 304 Solar Bars

@hollow obsidian Does this get it right?

[hollow obsidian]

I have no idea... I'm not best with math and all AI showed me different results 😄

But wait. No it's not right

[agile jay]

...

[hollow obsidian]

It's wrong

[agile jay]

Do the maths first, so that there is actually something to test against.

[hollow obsidian]

It doesn't list at the end saronite gloves and storm gloves

It miss items

Question:
How many units of the following items do I need in total to craft 38 Stormshields:
Saronite Plates: x
Saronite Gloves: x
Battle Gloves: x
Storm Gloves: x
Solar Shields: x
Solar Bars: x
Obsidium Bars: x

We need this list completed!

It should show as example 38 storm gloves.

[hollow obsidian]

That's correct, but it is a complex math task... That's why I ask AI...

But it's already a fail, because it doesn't show all.

[agile jay]

Yep, because I didn't ask it to display it like you said.

[hollow obsidian]

But THAT was the important task 😄

[agile jay]

And it does show the storm gloves etc

If you actually read it.

[hollow obsidian]

Not in the summary

[agile jay]

Yep, because I asked what materials you need, and since those are built from other materials, they don't count.

[hollow obsidian]

That's what I need to see it always

[agile jay]

Well it worked for me first time, so it's probably your prompt.

[hollow obsidian]

You don't know it... It worked for me many times too. But different results.
As long as I didn't proofread your result. I can't tell if it's 100% accurate

But that would require more time now.

[agile jay]

The model literally goes over the steps, which can be easily checked...

[hollow obsidian]

I have a math phobia, that makes math hard for me

I can solve simple math, but my brain blocks quickly

[agile jay]

Then play a game that doesn't use math.

[hollow obsidian]

But I like my game 😄

I just try to plan a little bit my production

That's why I wanted a easy checklist

I have the problem, that when I get confronted with multiple math tasks, I get blocked.
Like someone who has fear of spiders.

I can't control it (now) 😦

But it sucks, when I can't trust AI on math.

[agile jay]

If you can't do the math yourself, you can't trust anything.

[hollow obsidian]

I can just proof check it in little steps

That's how I play 😄

I focus on one item to produce, but the game gets more complex as you can see :(...

Soo much items for 1 shield

7 different ingredients minimum! I didn't include the previous mats 🙃

[torn zinc]

If I select the default Perplexity and Pro enabled (as a pro user) what model does that use? Is that got 4 turbo or is that still a tweaked gpt3.5?

[agile jay]

I imagine it's the default model, which is the 3.5 finetune.

[agile jay]

The model went through all the steps and clearly showed it's working out, while in your, because of your prompt, they just spit out random number guesses.

As can be seen as none of your results match mine, and I verified mine to make sure it was right.

[torn zinc]

How are peoples experiences with leaving it as that vs changing the default model to gpt4 turbo? - I did a few tests and the default Perplexity model seems to do well but I’m surprised because I thought gpt4 turbo was supposed to be way ahead of gpt3.5. Although I don’t know how the tweaks effect that

[agile jay]

The problem is that instead of remembering the answer to your question, it is using sources, which means that weaker performing models can score a lot higher.

So you can get similar results for both, since they are sourcing the same sources.

[torn zinc]

Is there any negative to changing the default to gpt4 turbo?

[agile jay]

Mostly the response time, I would imagine.

[torn zinc]

@agile jay thanks. Either way I’m impressed. Was kinda hoping it was bad so I didn’t have to think about another sub but perplexity pro has been pretty awesome

[scarlet schooner]

Been defaulting to gpt4 turbo, satisfied with the results

[calm swan]

Is it better than opus?

[scarlet schooner]

Tbh it is the same

And opus only has 5 uses

In the end i will switch back to gpt4 turbo

[simple ruin]

Claude 3 is nice i love it

[scarlet schooner]

After the opus uses run out for the day

[calm swan]

How is the logic, reasoning and coding?

[scarlet schooner]

I can generate article with opus and gpt4turbo rifht now and share it

Any prompt you want to generate?

[calm swan]

Hmmmmmm

[scarlet schooner]

Better you check it yourself and make your own conclusion

Better than relying on what people tell you

[calm swan]

Benefit of watching movie and series instead of playing games

Tone:Casual, creative, chatty language and less formal

[scarlet schooner]

Ok

[calm swan]

Gemini ultra kinda nailed for creative writing, I kinda want to see how claude opus does

[scarlet schooner]

I will be sharing on #💗│sharing

Posted both there @calm swan

[calm swan]

Thanks, it's cool, yea, the response is really close, gpt-4 provide more casual, and less formal while claude opus is kinda formula but it have emoji that makes more interesting

[scarlet schooner]

Tbh you can press the regenerate button on got4 turbo till u get similar more formal thingy (writing list)

[calm swan]

Yea, I kinda like gpt4 turbo response

[scarlet schooner]

Both model are as smart as the other in term of reasoning (dont quote on me but it shouldnt be that far)

[sweet jasper]

Even so the math might be dubious without a calculator tool or wolfram alpha

[scarlet schooner]

Or... spreadsheet 😄

[agile jay]

It's is very simple maths

[scarlet schooner]

For that guy's stuff. I like doing those kind of things in spreadsheet

[agile jay]

Yep, also helps when you ask for working out and not direct answers like he did, lol.

I guess a lot of people don't know how LLM's work...

[scarlet schooner]

LLM is token predictor pro max

With attention

😄

[agile jay]

Yep, pretty much a probability machine

[scarlet schooner]

That's why it can't predict how many words it has in its answer

Because each time it predicts token

It has random seed

[agile jay]

You want to control the probabilities by the prompt you give it.

[scarlet schooner]

Yeah, but it is not everyone's stuff ig

Tho those info are readily available

[agile jay]

You can actually get it to do that, you just need to put it in the context.

[scarlet schooner]

In easy to understand format

[sweet jasper]

Someone posted an article of an llm with access to a calculator API and without, it does help alot once math gets harder

I shared it here b4

[agile jay]

Yep, AI with tools perform better than AI without them... Just like humans.

[scarlet schooner]

Yes ofc

[agile jay]

You can also make the AI make their own tools...

Since they can easily write the code for them.

[scarlet schooner]

chat.openai.com is still a good service imo even when perplexity pro exists

Because it can write its own python code

And execute them

To write better answer

[agile jay]

You can do that with other services too, by plugging in your own code interpretor.

I'm working on a wasm/wasi one, so I can run any language on it.

[scarlet schooner]

Or you can even tell it to generate some dummy data and make into a downloadable json

[agile jay]

Python is too limiting, imo.

[scarlet schooner]

Not so simple for plebs. But yeah. For those who dont shy away from at least writing a simple script. Other services are better

[sweet jasper]

Thats rathee interesting

But doesnt the response take much longer for compiling stuff into wasm

[scarlet schooner]

Im just saying at face value for the masses

[agile jay]

Yep, then you can make the model make its own tools and agents for tasks, and offload them when needed.

[scarlet schooner]

Perplexity pro cant make the llm write a script and interpret it themselves

[agile jay]

Pretty sure that is the next step for Claude and GPT

[scarlet schooner]

Even though you choose gpt4

Im not saying perplexity is bad. I use perplexity pro more than my openai

[agile jay]

You could, but you'd have to make an extension for it.

[scarlet schooner]

Not the api version, i mean the face value for masses

[sweet jasper]

Intercept the request

[scarlet schooner]

The one you pay $20 and use the webapp ver

[agile jay]

Yep, even for that

[scarlet schooner]

Im pretty sure you are missing the context

[agile jay]

It even uses web sockets. So even easier than ChatGPT

[scarlet schooner]

Those non tech people, like sport people etc

Lol

[sweet jasper]

I mean, they can ask AI nowadays

[agile jay]

Those people are unlikely to use code interpreters in the first place.

[scarlet schooner]

You are missing the context

I am saying when comparing perplexity pro vs chat.openai.com

[agile jay]

Most likely just use copilot pro in Excel etc

[scarlet schooner]

The chat.openai.com can generate pythin script and interpret it themselves

[agile jay]

So can the API...

Just the perplex team aren't using it.

[scarlet schooner]

Ok we cant talk on the same page it seems i am saying on the context of face value for those masses who dont do tech

[agile jay]

Probably costs more.

[agile jay]

And those people are paying for GPT Plus???

[scarlet schooner]

It's just somehow you just deliberately remove some context and write your answer as if you are using gpt2

[sweet jasper]

I know that chatgpt plus can do it, pplx pro cant, what code is saying is that you can bootstrap a code interprator through an extension, which would allow the masses to use it

[agile jay]

No, we answered before, but you no read...

[calm swan]

Anyway dude, how much gpt 4 u get per day?

[agile jay]

40/requests every 3 hours

[calm swan]

I mean with perplexity pro

[agile jay]

600

[scarlet schooner]

Nevermind. You win 🙂

[calm swan]

Got you, is it enough for your use case per day?

[scarlet schooner]

600 total pro answers per day no limit for gpt4 turbo

[agile jay]

Most people aren't using more than 25 a day...

[scarlet schooner]

595 if no matter what you want to use those 5 answers on opus

[calm swan]

Wait, so we can chat with it in Writing mode unlimitedly?

[agile jay]

I'm guessing they have a fair use clause

[scarlet schooner]

If you turn off pro mode. Yes. But idk there might be slow down. Or shadow ban

[calm swan]

Writing mode is the same like using chatgpt right?

[scarlet schooner]

That will be lifted after day refresh

Or something

[scarlet schooner]

No, even with writing mode the pro mode toggle is different thing. So if you toggle off the pro mode even in writing mode. You will use gpt 3.5 instead

[agile jay]

I mainly use Perplexity for stuff that needs up to date info.

[scarlet schooner]

Same

[agile jay]

It's really good at a lot of stuff that all the other models fail at.

[calm swan]

Wait really? Not gpt 4?

[agile jay]

Because they didn't have fresh data.

[scarlet schooner]

I don't think you can exhaust all 600 responses on pro mode. Unless u going nuts just using perplexity pro all day long

[calm swan]

Yea, that's the one thing I planned to use it for, but one more thing is writing mode just for chatting without it connected to internet stuff

[scarlet schooner]

Try it yourself, change the all mode to writing. Turn off the toggle of the pro mode to off. You won't be using gpt 4 turbo

But I might be wrong, I see those people who has sent over 10.2k msgs on pi.ai and they are very proud of it

[calm swan]

I'm not subscribed yet tho, that's why I ask you guys for. So In order for me to use gpt 4 turbo, which mode should I use? Becuase I though with writing mode is just chatting like using normal chatgpt without connected to internet data. I afraid if I use normal mode, it will always search the internet and not maintain previous message context

[scarlet schooner]

Writing mode is to tell the service to prioritize looking up academic papers to answer your question

[agile jay]

600 messages is a lot more than you'll ever need in 24 hours. If you have a life.

[agile jay]

That's academic mode

[scarlet schooner]

Wait isit

Need to open my laptop to check

[calm swan]

It's academic mode

[scarlet schooner]

My bad, you are rifht

[calm swan]

Yea I know, but it will try to always search the internet even if it not needed. This will downgrade the quality of the answer instead of using internal knowledge

[scarlet schooner]

Though there is that toggle tho

[calm swan]

So writing mode is the same like chatting right?

[agile jay]

That's what writing mode is for...

[scarlet schooner]

Yeah probably, but i remember you were asking how many times you can use gpt4turbo rite

The answer is still 600

Even in writing mode

[calm swan]

Gotcha, 600 is still massive tho

[scarlet schooner]

Inexhaustible

For real usecase

[agile jay]

Yep, the most I've used in one day is around 75...

[calm swan]

So the amount of we can chat with gpt 4 is the same amount how many copilot we have right?

[scarlet schooner]

Once i sat hours the first day i paid perplexity. And the total pro uses in a day was like 60

You run out of idea faster

Than you can ask

[calm swan]

Yea, I planned to use it for solving mcq tho, hahhahaha

[scarlet schooner]

Yes sir

[agile jay]

Yep, the advantage is you don't have to worry about it stopping while working...

[calm swan]

Okay thanks. So In conclusion :
Amount of gpt 4 = amount of copilot

Writing mode = chatting like chatgpt

[agile jay]

Like I had to with GPT Plus

[scarlet schooner]

meanwhile at chat.openai.com after 40 messages...

Please wait till xx.xx pm

[calm swan]

We can also upload image and file to perplexity pro right?

[scarlet schooner]

You can

[calm swan]

Yea I hate it bruh, that's why I never subscribe to it yet until now

[agile jay]

Rip, I have all the major services.

[scarlet schooner]

Idk what else out there

[calm swan]

The minus is just the code interpreter right?

[agile jay]

Claude Pro, Perplex, Chat GPT, Gemini etc

[scarlet schooner]

But yeah i am subbed to perplexity pro, gpt plus, and claude

[calm swan]

You subscribed to all of it? Woah

[agile jay]

Pretty much, and the voice interface

[scarlet schooner]

Also i have gemini advanced

Coz 2 months free

Lmao

💀

But ngl, gemini advanced kinda dumb

[calm swan]

Hahahaha yeah, me too, it's really good for creative writing but really bad at following instructions and logic

[scarlet schooner]

Sometimes it just doesnt follow what you write

[calm swan]

Word

Hahahhaha

[scarlet schooner]

Or what you tell it to do

[agile jay]

Yep, Gemini is best at translation. Terrible at everything else...

[scarlet schooner]

Now this is interesting info

[calm swan]

Claude 3 sonnet is kinda uncensored tho

[scarlet schooner]

I use claude to debug code sometimes

[calm swan]

It's literally can cuss

[scarlet schooner]

200k context window baby

😄

[agile jay]

To me it seems Claude has the biggest output.

[scarlet schooner]

Claude really writes long answers yea

The opus one

[agile jay]

Yep

[scarlet schooner]

I never tried the sonnet

[agile jay]

Sonnet will probably be good for agents.

[calm swan]

Try it for creative writing, it can cuss and use bad word

I like it

[scarlet schooner]

This is why you talk to people

[calm swan]

Wait but fr tho, how many context if we paste some text to chatgpt textarea

[scarlet schooner]

Just check it on token calc

From openai

Tokenizer somethinf something

[agile jay]

Yep, token to words ratio is around 4:3

So 128k tokens is around 100k words.

(even though it's actually 129k...)

[calm swan]

Wait I mean in perplexity textarea

My bad

Is it 4k like bing?

[scarlet schooner]

You mean the amount of token you can write into the text area? Idk, i never hit that much. For long text like code. I just upload fhe file

[calm swan]

Gotcha, the gpt 4 that we got in perplexity pro is the one who got 128k context right?

[scarlet schooner]

@agile jay ppl seem to riot over claude 2.1. They want claude 2.1 to be back instead of using sonnet
Is it that good?

I mean on the pro feedback below

[agile jay]

No, people just dumb apes

[scarlet schooner]

Afaik yeah

[scarlet schooner]

💀

[calm swan]

Nah, I already try to compare it in chat.lmsys

Sonnet still better, creative and capable than 2.1

[agile jay]

https://huggingface.co/spaces/lmsys/chatbot-arena-leaderboard

[scarlet schooner]

Sound about right

I still wonder if me defaulting to gpt4 turbo instead of using sonnet is a good idea. Is sonnet still better than gpt4 turbo because of bigger context window as well?

I cant find convincing testimony of people using sonnet vs gpt4 turbo

[agile jay]

Don't think it matters much.

128k is a huge context.

More than the first day per book

(which is in the 80,000 words range)

[scarlet schooner]

True

Or i just feel comfy with gpt4turbo

Btw did gpt4 votes come from carryover from before claude 3 release?

Interesting to see gemini pro up therem it should be the 1.5 version right? Coz the current gemini advanced kinda dumb

[agile jay]

I assume it uses the Elo method, so should be accurate.

[agile jay]

No, it's the current one but with online support

[scarlet schooner]

Ic

[agile jay]

You scroll to the right and you can see that it's the only online one...

So probably scores high because it knew up to date stuff.

[scarlet schooner]

Right

[agile jay]

Funny it can't beat offline models, lol

[neat lichen]

I can't find the button to generate images.

Is that a thing still?

[agile jay]

You need to search something first

Then it appears at the top right

[scarlet schooner]

But other than presenting up to date stuff. I believe gemini model isn't that good. Unless I just got bad time with it

[agile jay]

Also image gen is only supported on the web browser, I believe.

[neat lichen]

Dude, thanks for a quick response. 🙂

[tranquil orchid]

Hey @neat lichen!

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.

[agile jay]

Np

[agile jay]

You can try it, but even the links are hallucinated a lot...

[scarlet schooner]

Rip

[neat lichen]

Is there a model that doesn't hallucinate?

[scarlet schooner]

I am not gonna pay a dime for gemini advanced. As soon as my 2 months free period is up, I am done with it

[agile jay]

The more intelligent the model, the less it hallucinates.

Currently Claude opus hallucinates the least.

[neat lichen]

I'm a huge Claude fan.

Nice.

[scarlet schooner]

Try claude 3 and gpt4turbo.

They follow prompt quite closely and do well, and less hallucinate (rarely even)

[neat lichen]

I'm really confused about Claude 3, though. It says it only does 5 searches a day. What happens after I do those 5 searches? Pro is telling me I have 600 every day.

[scarlet schooner]

You will then use claude 3 sonnet

[agile jay]

Lol, it might also be the reason why Gemini pro was removed from perplexity...

Would make up links...

[scarlet schooner]

So bad

[agile jay]

Since the team did say it was related to bugs...

[scarlet schooner]

Is that what a top tier tech company has to offer in production

[agile jay]

Google has been slacking...

[scarlet schooner]

They need to switch from using slack to microsoft team

Gottem

Btw what package do you use for google search in python? BeautifulSoup?

Since mixtral 8x7b supports function calls, i might try groq api

[agile jay]

They've literally been using LLM's longer than Openai...

[scarlet schooner]

And try to attach google search functiobality to it

[scarlet schooner]

Lacking forsight

Meanwhile satya nadella

[agile jay]

Nope BS4 is used for scraping.

[scarlet schooner]

Invest heavily in openai, in mistral

Put ai everywhere

[agile jay]

Windows AI

[scarlet schooner]

copium copilot for spreadsheet, for ms 365, for your underpants

[agile jay]

Yep, he saw the gold rush coming.

[austere kestrel]

really? what do you mean put it in the context exactly? curious to know how it fixes this quirk - which I thought was a result of the autoregressive nature of LLMs outputs + the fact they process/generate tokens, not actual words (ie. nothing to do with prompting / context 🤷‍♂️)

[agile jay]

I meant when asking for a specific amount of words etc, instead of something which isn't defined in the context yet, like your prompt.

[leaden lagoon]

I need to talk to a Dev about the perplexity llama website

[austere kestrel]

still don't really get what you mean (or how you overcome it with putting something in the context) - it's the same quirk

[warm cave]

I have never been able to get a LLM to give me the right number of words, so if I want 250 words from gpt, then I will ask for 500 words or ask for expanded and in depth writing. And if the responses are generally too long I ask it to be concise and brief

[austere kestrel]

yeah it's a function of how generative AIs, well, generate text/tokens - they're autoregressive (they sample/predict the next token, then look back at over it all, and then predict the next one - and so on). They don't know when their response will/should be completed when they start a response... It's a simplification, but it is basically just a very sophisticated and self iterative autocomplete at the end of the day ha

And also the fact it's tokens, not words.