💬│general

[fossil sentinel]

https://labs.perplexity.ai/

Tfw pplx models are available to try out on perplexity labs

[robust quartz]

Okay. That doesn't tell me anything about pplx-8x7b though

[fossil sentinel]

What are you trying to know then

[robust quartz]

Anything about it

Besides the using it in labs

[valid yoke]

You wrote “very specific question” lol wtf?

[robust quartz]

There's absolutely zero documentation on it
I just wanna know something about it
That's not coming from the model chat
Cause it's gonna hallucinate

Id say its specific because nobody but the people from perplexity can answer

[fossil sentinel]

No idea but probably related to the mixtral-8x7b-instruct model

[robust quartz]

It's probably the same concept 8 different models and only 2 get activated per needed request but it would be nice to confirm that
Cause it would be faster than pplx-70b

[runic garden]

pplx-8x7b is most likely a fine tune of the mixtrial model, similar to 7b and 70b, being a fine tune of mistrial-7b and llama2-70b

[tame current]

Do you think Perplexity can read PDF better than ChatGPT?

[austere kestrel]

Interesting article.
Personally love the sound of this

[tame current]

Both good options, I wonder what perplexity would do with the upcoming LLaMA-3 model

Well, I say upcoming, it’ll still be released sometime this year, at best

It’ll still be huge though if it turns out to be a GPT4-class model

[serene sand]

#💬│general message

[inland wolf]

Will the standard GPT-3.5 model be replaced by the lama-2? I just didn't get it a little bit

[austere kestrel]

I don't know. The article makes it sound like they are very determined to not be using commercial models as the underlying LLMs for their product (don't want their search engine to just be a wrapper of an OpenAI model). He said they are highly focussed on using open source models (mentioned that the day lama-2 was released, they started fine tuning it). And obviously they're interested in Mistral too. So not sure

[south prawn]

Where did the "edit or adjust sources" icon go? We cannot find it anymore. { } View Sources
https://blog.perplexity.ai/faq/what-are-threads

[tribal wigeon]

Hi, did something change in how to operate the pplx online model through api? Some specific settings it want? Because it has been behaving different for me. Lenghty nonsense replies or it elaborates on some no_search stuff and bing

Also when deleting all my settings to use it at default, it performed inconsistent

[fierce sand]

Hello everyONE, I'm not sure if this is something for you but I thought I just share it here. I hope this is acording the rules 😉 ... otherwise let me know or just remove it.

https://www.linkedin.com/posts/bryancassady_ai-summit-ai-powered-creativity-activity-7162444182446419968-aJze?utm_source=share&utm_medium=member_desktop&lipi=urn%3Ali%3Apage%3Ad_flagship3_search_srp_all%3B0osXCoG5TGGCwMiYJ2nd%2Bg%3D%3D

[pine dagger]

@signal hamlet Does Perplexity plan on open sourcing their models:
PPLX-7b
PPLX-70b
and, PPLX-8X7b

[shrewd crown]

which model is best to use with perplexity?

[inland wolf]

Which neural network will be better? pplx-70b-online or pplx-8x7b-online?

[hybrid bison]

Anybody experiencing strange results via api today or is it just me? pplx-70b-online

[woven orbit]

Yes since yesterday I have been having issues with both 7B-online and 70B-online. It start responding find and then starts producing gibberish.

[hybrid bison]

OpenAI is having issues too

interesting

[pulsar rampart]

GPT4 is not working with images?

[halcyon coral]

Hey @pulsar rampart!
GPT-4 is working with images, you can simply upload a file with the + symbol and ask something about it https://www.perplexity.ai/search/Transcribe-pD43J2NtT12SPjKLA5nbdw?s=c

[hybrid bison]

Getting a lot of timeouts from pplx70 online API

[teal peak]

I would recommend trying out GPT-4 and Claude to see which one you prefer.

GPT-4 is smarter, Claude is more conversational.

[tame current]

Perplexity Push sounds very exciting!

[worn monolith]

Yeah it does! I'm excited for that.

[tame current]

Me too, it’ll be nice to have a regular AI curation of News

[eternal quartz]

I'm a little confused, is it only for Slack? I don't use slack :(

[bronze matrix]

Yeah, looks amazing! When is it launching.

[grave saffron]

You do now

Haha

[tame current]

Team discussions ?

[boreal pasture]

oops pplx might be dead https://www.theinformation.com/articles/openai-develops-web-search-product-in-challenge-to-google

[eternal quartz]

That will only have their own models. Perplexity has way more options. Unless they pull something unexpected I don't see myself switching back to OpenAI.

Honestly I'm happy to see more competition for Google. Maybe they'll make their search results actually good again...

[hallow linden]

It has to do multiple things to be better than perplexity. 1. Be as fast 2. cite sources 3. be accurate as much if not more than perplexity.

[tame current]

Perplexity will always be more versatile just because of the amount of models to pick from

But it won't surprise me if Perplexity becomes the first of a series of new web search browsers powered by AI

[boreal pasture]

yea that was a troll

i think these 2 are going to have different use cases

i bet that pplx will look way different in 2 years

[glad crypt]

does anyone know how perplexity builds off of the gpt base model to provide embeddings of online sources?

is their implementation for this open source? i am curious to how they are doing it

[valid yoke]

you know that GPT is closed source right

there’s nothing open about OpenAI

[brisk aspen]

so it isnt really gpt 4 ?

[glad crypt]

yea of course, but I am assuming the perplexity takes the output from gpt then manipulates it to get what we see. wondering how they are doing that.

[valid yoke]

Don’t assume, ask it!

[wise edge]

#📘│faq
is there 32k token context length model in labs or api ?
mixtral-8x7b-instruct 4k, pplx-8x7b-online 32k ?
I don't know the maximum context length for these models, and don't know how to choose when using labs and api.
mistral-medium is also work nice, i am confused
some body can help me ?

and i think we need a python sdk for online models

[oak flax]

Hey Ai fam , glad to be here 🙌

I am so stoked, yesterday Pre-Ordered Rabbit R1 and got $200 credit for one year subscription of Perplexity 💜 thx for great VD present 🎁

Free R1 and dope tool for one year, insane 🙌🔥

[eternal quartz]

Did you guys disable the &copilot=true URL param? I was using it as a custom search shortcut and it stopped working.

[cursive jacinth]

Hi, Gemini Pro out from perplexity?

[queen palm]

Deleted all my personal information as it seemed that pplx just added it to my search results with no added value or sense.

[austere kestrel]

I think it can be useful if you want the responses to be in a certain 'style' ("I'm still learning English, so keep the language super simple" – or something like that I imagine works pretty well)
But if you tell it that your day job is scuba dive instructing or whatever, it will just add some random sentence at the end of each response ("As a scuba diving instructor, the long term price of wheat will be unlikely to directly impact your work" or something like that - it actually can be pretty funny sometimes)

Or if you use something like, "I am a professional investigative journalist. Make sure you use every advanced information foraging technique you know to surface obscure and hard-to-find data from the web to help with my investigations."
It won't actually use more sophisticated techniques and return substantively 'better' information - tends just to tack on the "As an investigative reporter, yadayad

[fossil sentinel]

Man one thing about gemini that bugs me is that it doesn't render latex

[austere kestrel]

If you tell it that it can, maybe it can?
(this is coming from somone who didn't know what 'render latex' meant until a few moments ago - so apologies if this is way off base lol)

[fossil sentinel]

WTF

Ngl idk why I never bothered to ask it

I just expected it to do it on its own 💀

[austere kestrel]

It says it can't see uploaded images sometimes as well 🤷‍♂️

Obviously annoying to have to "tell it" what it can for it to do it, but I think these kind of wrinkles will be ironed out pretty quickly

More important that a capability exists than whether it has been implemented perfectly on release (though that would ofc be preferable ha)

[fossil sentinel]

Agreed

Thanks for pointing it out for me

[tranquil orchid]

Hey @fossil sentinel!

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.

[fossil sentinel]

Oh

[fossil sentinel]

You know what

It's not that it can't render latex

The issue is I was using the Gemini app

Which can't do it

It can only do it on the site

🤦‍♂️

[austere kestrel]

@fossil sentinel aha I see! Well hopefully soon it works across platforms. Was thinking about how annoying it would be, for people working with equations - like yourself by the sounds of things - to not have this rendering. Seems like it would really limit the readability/usefulness of the outputs (even if the content actually is useful/accurate)

[tame current]

there is a gemini app ?

[cursive jacinth]

Hi sir, saw someone saying about gemini app. I couldn't find any. Seems they could have beta access sir.

[tame current]

I found it, looks like it's only available in US for now

[solar trellis]

Gemini Pro is out from perplexity?

[cursive jacinth]

Yes sir, due to low utilization and performance, said by sir @halcyon coral

[fluid mauve]

I really should read the docs. I'm a silly muppet. This is exactly what I was looking for, so I didn't have to hardcode model names into my testing app. @halcyon coral Is this the best place to get accurate and up-to-date model names? Or, where do you suggest we pull them from?

[halcyon coral]

https://docs.perplxity.ai is mostly up to date or you can just ask in #🧪│api-general

[fluid mauve]

Thanks. I wanted to programmatically pull the list so I can play with which ever model I have. I will go ask im #🧪│api-general

[tranquil orchid]

Hey @fluid mauve!

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.

[somber helm]

Just for the first few random tries: Perplexity (I'm on the paid plan) is not a coding master. Far away from GPT-4 quality.

[halcyon coral]

Did you set the model to GPT-4, did you try Copilot and also Writing focus?

[somber helm]

I could. But why does Perpelxity not automatically provide the best possible model for a respective use case, on their default settings? Why am I forced to manually play around with model settings?

[halcyon coral]

So that users have freedom of choice.

[somber helm]

I guess you missed my point. I think the choice of a model can and should remain possible, but Perpelxity's base model should cover any use case by automatically applying the most convenient/most suitable model on a specific case/query. That would be most convenient for a user and a real added value in the marketplace.

[hybrid bison]

Which is the right chat to talk to people using the API

[teal tulip]

To discuss the API?

#🧪│api-general #1161804761247526912 #1161803596870979664

Depending on what you’d like to discuss

[hybrid bison]

Bugs that might come up, tips and tricks on better api usage, use cases, etc

Just general api discussion

[glad crypt]

haha, is there anyone who is able to shed some light on this question?

[halcyon coral]

Such internal company information is not shared here, thank you for your understanding!

[devout hemlock]

anyone has "error - refunded" when tring to generate an image - Ive tried for 15 or so time - all the same

[austere kestrel]

https://blog.perplexity.ai/blog/introducing-pplx-api

[heady quiver]

What have u done to this tool????? It doesn’t recognise youtube links anymore; I tried both Claude and GPT 4 and can’t make a summary of a YouTube video.

[upbeat dawn]

been reported these problem but no update still

[sleek vortex]

id assume theyre using fine tuned llama/mistral models to summarise websites

possibly with a embeddings model as a cache layer in between

then feeding the summarised sources to gpt when you make a query

they also have some like broken related model on labs

probably some internal model that finds related web sources to a query from their cache or from the web

also with "copilot" mode the model that asks you questions has to be different to the model doing the actual output

if you switch between gpt/claude/perplexity/gemini but ask the same prompt often the copilot questions is the same

plus copilot often asks you what you asked it, so it's probably some lower-level model thats worse at tasks

[tame current]

https://blog.google/technology/ai/google-gemini-next-generation-model-february-2024/amp/

Google is already working on a 1.5 version for Gemini

It’s going to have a whopping one million tokens

The way they’re wording this blogpost is that 1.5 is going to be available sooner rather than later

[fossil sentinel]

I mean yeah they gotta work fast considering how underwhelming 1.0 was compared to the hype

Still good but not as good as expectations

[tame current]

Mhm

The context window is very impressive

I think OAI would have a hard time reaching 1M Tokens in the short term

[fossil sentinel]

Oh wait

Just took a quick look lol, so Gemini 1.5 is for the Pro model?

[tame current]

Yeah

Interesting choice considering

But Google has been known to have really poor marketing lol

[fossil sentinel]

Hmm

The video upload thing seems cool

[wraith crow]

thanks for sharing your agario gameplay! but this server isnt meant for that.

[quartz jungle]

is perplexity push still in beta - cant find in slack apps

[halcyon coral]

Soon, you'll be able to subscribe to topics & pipe updates directly into Slack channels for lively team discussions. Stay tuned!
https://discord.com/channels/1047197230748151888/1047204950763122820

[cursive jacinth]

Hi sir, is it only for slack apps? I thought in perplexity sir?

[halcyon coral]

It's only planned for Slack

[jolly otter]

Can't wait to get perplexity as default search on safari. Anything one person can do to push Apple to offer that?

[echo blaze]

Main reason I deleted stopped using Safari. Went to Chrome for the Perplexity extensions. Not really looking back

[wraith crow]

for safari you can use the extension xSearch to set perplexity as your default search engine (this works for ios and macos). or you could use the the "Quick Website Search" on macos but this doesn't work as a default search, you will have to type "perplexity.ai" first.

[cursive jacinth]

ah ok sir. Thought gonna be something new for perplexity user. thanks

[tranquil orchid]

Hey @cursive jacinth!

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.

[fossil sentinel]

How come we can upload files to google's ai studio but their actual site is so bad with it

[jolly otter]

I would do xsearch but I am a cheapo and it's $1.99 haha. I'll just set my homepage to perplexity for now

[obsidian acorn]

Has anyone seen sora open ai video model

[cursive jacinth]

do we have gpt 4 turbo now sir

[glad crypt]

ah seems like the 4th point is what im looking for. hoping to generate output that is like LLM with citations

[cyan creek]

I personally don't seem to

[glad crypt]

could you explain why you make this guess? usually the embeddings are static within the model after training, so you wouldn't need a separate layer for cache right?

[glad crypt]

or are you guessing that the embeddings change due to context during inference?

[stiff orbit]

hey, im looking for alternatives of poe and i stumbled upon perplexity. can anyone tell me how much is the context window for gpt-4? i couldnt find this info online.

#📘│faq

[fossil sentinel]

I think 32k model

[stiff orbit]

does it accept images as input as well?

[halcyon coral]

Please have a look at https://blog.perplexity.ai/technical-faq/what-is-a-token-and-how-many-tokens-can-perplexity-read-at-once
https://blog.perplexity.ai/faq/what-is-perplexity-pro

[stiff orbit]

thank you. as a final clarification i have 3 questions.

1. so i get 300 messages/day in total, it doesnt matter if i use gpt4 with vision or claude?
2. is it possible to have multiple threads/chats with different models? for my studies i prefer to use claude for text research and gpt for explaining/searching things. it says we can use "rewrite" command but can only change in settings 🤔
3. is it possible to only use claude 2 instead of 2.1? as we all know how trash 2.1 is

[halcyon coral]

1. 600/24h, yes
2. There is no default model for collections, but there are a few great suggestions 👀
3. No, you can only use Claude 2.1

[stiff orbit]

damn, i see. claude 2.1 is honestly very bad and that switching model thing might be an issue. thanks though. i definitely wanna try it once and see if it can work for me.

[tranquil orchid]

Hey @stiff orbit!

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.

[errant moss]

Will Gemini 1.5 be available via Perplexity?

[signal dirge]

based on the fact that they removed Beta version today (because of the release of v 1.5) it will be for sure.
@tranquil orchid do you have any plans to decrease subscription amount?

[errant moss]

I don't even know what I would do with a million tokens.

@tranquil orchid any idea if a 'family plan' or 'team plan' is on the roadmap?

[fossil sentinel]

Upload a movie like they showed 😂

[inland adder]

@tranquil orchid Is there an RSS feed for the discover page: https://www.perplexity.ai/discover?

[placid spoke]

did perplexity for discord stop working?

[valid yoke]

How did it work?

[placid spoke]

like a bot you tag it with @ and it answers questions

it stopped working for me

[valid yoke]

Guess so

[placid spoke]

do we know why?

[valid yoke]

I didn’t even know there was one until today

Wish they had a telegram bot

[signal hamlet]

Hey, @placid spoke, #📢│announcements message

[sleek vortex]

by embedding model i mean like i think they have a RAG somewhere in the chain which is how they’re able to make it so fast

Asin storing the like summaries of large amounts of webpages then when you search they could use a small model to form search queries then search it into their cached summaries of websites

well somerhing like that

Because there’s no way pplx-70b-online is actually hitting websites in real-time while being so fast

or any of the pplx models on labs

copilot might work slightly differently i feel like they actually get that to make the summaries in real time

which is why it’s more up to date but also much slower

[meager sparrow]

Please elaborate on what you are saying

[marsh temple]

Hey question, apologies if this is incorrect location. Where can i find an update on my rabbit r1 pre order?

[austere kestrel]

Yes I think you're right, certainly for pplx-70b-online, it's not literally conducting a google search in real-time and pulling snippets of results, it's referring to a cached index from perplexity's crawler bot (which is fine, to be clear - don't understand how their online models would be possible at such speed either otherwise)

Something I hadn't noticed before in one of FAQs is the reference to 'partners' (some kinda serp / search results API maybe?) as part of the underlying process for perplexity. Not sure if this applies to the API/pplx-online models (and doesn't particularly bother me in any case), but just an observation

[austere kestrel]

A

[lament spoke]

Finally a definitive answer.

[somber pelican]

+1 to the family plan! I have a family of 5 of which most are infrequent users. I'd love to get them pro features but think $100/month is too steep. At the same time, I think it's worth more than keeping them on the unpaid plan.

[halcyon coral]

Please have a look at your order email or contact Rabbit support 🙂

[halcyon coral]

Currently no RSS feed is offered, but thanks for your interest.

[quiet gorge]

？

[cursive jacinth]

I dont know also sir. Maybe glitch sir

[quiet gorge]

It seems to be another pilot test, only a select few have the privilege to experience it.

[fluid mauve]

Can I ask perplexity to reference 5 site: and pull the contents into one coherent document?

[tame current]

isn't that what perplexity does by default

[fluid mauve]

I don't know I'm on a train and don't have access to my pc to test

Just tried it on the perplexity android app trying to get a news report from a combined bbs, al jazera and cnn sites: but perplexity couldn't do the request as asked for

[wraith crow]

you can if you mentioned the websites with copilot on, here is an example https://www.perplexity.ai/search/6319eb7c-15e9-4111-adbb-ccf80b2e0b20

[gray prawn]

https://fixupx.com/kantrowitz/status/1758167517514658284?s=46&t=ma1yL55mNVPFHkyaDymViw

Sheeeeeeeesh

Google really must’ve liked that guy

[fluid mauve]

AHH thanks. So I preface it with websites: for generic names. Or does it also work with full website URLs?

[tranquil orchid]

Hey @fluid mauve!

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.

[wraith crow]

should work with the url too

[fluid mauve]

Excellent. That might have saved me 2 hours between 3 and 5am this morning 🙂

[placid spoke]

Can't find related announcement, please save me time and tell me if it's pay walled to get the discord bot back?

[halcyon coral]

hey, quick update, we are discontinuing our discord bots. we are going to bring image gen to the main product soon, stay tuned.

[normal crest]

Yes sir 3 bags full sir 😎

[cursive jacinth]

Hi Sir, I dont know what you are talking about sir.

[austere kestrel]

Was your screenshot from a pplx mobile app (whereas @quiet gorge 's seems to be from the web)? Perhaps that's why they're different?
(in any case, seems like there are some changes afoot!)

[cursive jacinth]

Hi sir, that was from Mobile and i opened kiwi Browser for it sir. I dont think it makes any difference sir.

[austere kestrel]

ahhh gotcha. I see I see, thanks! (I'm seeing the same as the other person; no turbo)

[tranquil orchid]

Hey @austere kestrel!

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.

[austere kestrel]

Is there any performance difference with turbo selected as the model?

[cursive jacinth]

But I still dont know sir. Why it says turbo sir.

[austere kestrel]

Understood (that suggests the answer to my question is no ha). Thanks!

[tranquil orchid]

Hey @austere kestrel!

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.

[cursive jacinth]

I have not tested yet fully sir, i am at work sir

[austere kestrel]

Roger that - I'll leave you to it (gracias again)

[austere kestrel]

yikes, well if true that makes me think Google probably won't be giving any kind of early or preferential access to Gemini 1.5 endpoints 😬 😅

[humble swan]

I assume they might give access to it in the next month, maybe end of two months

I think it would be quite a good move, because they offered 2months of free gemini access on release and if they would just at the end of the free trial give access to gemini 1.5 that would be a powerful advertising method

[jagged pike]

Theres any way to load multiple pdfs at the same time and make prompts to use that info? (i got Pro)

ignore me! i just realiced i can upload many files just droping them 🙂

[jovial mantle]

Yes, the problem however is that you need a certain degree of editing afterwards because subsequent replies will always feature heavily the sourced text

Say, if you're tring to write a story and inputed a pdf as source

[hot surge]

what is a good/user-friendly synonym for the word "temperature"? I want to enable users to set the temperature of LLM.

[tame current]

randomness ?

most sites just call it temperature tbh

[hot surge]

the app is not for users with technical background, it might looks confusing to them.

[tame current]

you can add tooltips to explain it

[hot surge]

maybe stickiness?

I want to use it in a paraphraser tool

[tame current]

how about "creativity" ?

[hot surge]

creativity sounds more appropriate!

[gray prawn]

Google is also likely trying to protect its Ai team because I’m sure everyone is trying to poach everyone right now

Must be good to be an AI/ML employee at F500 companies rn lol

“Hey boss, ___ is offering me $20—“

Boss:

bruh my message was deleted lmao

“quadruple this man’s salary immediately!”

[ornate verge]

Will perplexity offer image generation like DALL-e and Midjourney?

[halcyon coral]

#⚡│ask-community message

[tame current]

I don't think midjourney has an APi

but perplexity already supports Dall-E afaik

[jovial mantle]

yes it does, DALLE3

[glad crypt]

if i use the pplx-70b-online model through the pplx api, will i get the sources citation as well?

or will i just see the text output?

[signal hamlet]

Currently, you'll see the output, the team is working on providing sources, as well.

[barren hawk]

Regarding asking for sources or adding outbound links to paragraphs generated through perplexity (using Claude), is it normal for it to add all dead links? Or is this a Claude specific issue?

[signal hamlet]

Claude might sometimes hallucinate some links if you ask to output them within the answer. The links in the Sources section backing up the answers should be fine, you won't get a dead link there.

[barren hawk]

Ok makes sense. Seems I'm finding Claudes reputation for being the biggest hallucinator is valid lol

[glad crypt]

okay thanks! is there a timeline for this? and is this what point number 4 is referring to?

[tranquil orchid]

Hey @glad crypt!

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.

[glad crypt]

also, i dont quite understand what a dedicated pricing structure means. isnt there already a pricing structure found here? https://docs.perplexity.ai/docs/pricing

[halcyon coral]

The blog post is from the time when https://docs.perplexity.ai/docs/pricing didn't exist yet.

[serene sand]

deep dive ➤#💬│general message

second paragraph, check first and third link. PRO users can choose between DALL-E and SDXL, 50 images/day

[glad crypt]

im also wondering what use case is codellama that isnt finetuned for instruct

wouldnt it render such a model quite useless since its hard to interface with it?

[glad crypt]

whats the difference between the pplx-70b and pplx-8x7b? is there somewhere that have measured their performance?

[serene sand]

#⚡│ask-community message

[meager sparrow]

What’s happening?

[echo mica]

Hey everyone

[solid lily]

#💬 I recently was charged $200 and I want a refund because I was never notified of this incoming charge.

[echo mica]

ouch

[halcyon coral]

Perplexity uses a normal subscription system, which means that as long as you do not cancel, the amount for the subscription will be debited automatically. If you have any further questions, please contact [e-mail].

[winter atlas]

are there any parts of the docs that are commonly overlooked that help with understanding things beyond search?

[arctic spindle]

Perplexity uses GPT-4 Turbo?

[jovial mantle]

I believe so

[snow pond]

Senior rust developer is here!!! Cosmos-sdk ecosystem, cosmwasm smart contract, solana native program, etc. Please ping me anytime.

[raven socket]

Crazy how google is scared bruh, we just have to admit it that perplexity with gpt5 (when it releases) will break the entire internet.

honestly, gemini 1.5 seems to be impressive but I don't expect much

by the news they quadrupled one employee's salary just seem a relatively out-of-scaredness approach to me

[valid yoke]

I wish the perplexity app had a share sheet thing on iOS

[raven socket]

investing in tokens just change nothing, making the predictions better and faster is the solution

[valid yoke]

Rather I mean: I wish perplexity appeared in the system share sheet

[tame current]

how to add perplexity to my discord server ?

@helpers

@tranquil orchid hi

[valid yoke]

They turned the bot off

[fossil sentinel]

Anyone try out NotebookLM? Curious if anyone finds it useful for them

[cursive jacinth]

Hi sir, GPT 5 sir AFAIK sir, is a long waiting sir, relax sir.

[sweet jasper]

No

Gpt4 turbo hasn't been rolled out yet in perplexity

Only gpt4 32k

[austere kestrel]

I'm not from teal (and don;t know what that is), but this may or may not be relevant / helpful #🦙│feedback-playground message

[glad crypt]

are there other online models that rival perplexity?

[gusty blaze]

YO new maps results in search????

[steady linden]

Hey guys

What channel shall I use to send something to the developers of Perplexity?

Specifically, a discrepancy between the subscription and FAQ pages

[halcyon coral]

Please describe the problem in more detail.

[steady linden]

So in here, the File Upload in the subscription page on the left shows that you can upload Documents, Images and more file types. While when I press on the button "Learn more", in the FAQ page, I can see that it only supports text documents.

[halcyon coral]

Under the item "How does File Upload work?" there is another item with "How does Image Upload work?" ☺️
But everything is listed on this page: https://blog.perplexity.ai/faq/what-is-perplexity-pro

[steady linden]

I see, thank you!

[tranquil orchid]

Hey @steady linden!

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

[tame current]

Hey everyone do anyone have problem on Claude 2.1 when you upload documents In and when you ask or want a summarize Claude complain about copyright Issues and then refuse to generate content?
But I uploaded my own document on there.

Something It answer Is "Unfortunately I should not reproduce potentially copyrighted excerpts or materials without permission. As an AI assistant without clear licensing information, I must refrain from providing verbatim quotes to remain in compliance. However, I'm happy to have a thoughtful discussion about any public domain works or general concepts that were mentioned. Please feel free to rephrase the question without extended quotes, or to discuss your own original ideas. I want to be helpful while also respecting content ownership."

Only have them a few times tho just curious.

[jovial mantle]

Question, how can I see other people's articles and stuff? Right now I can see only this series of searches pushed by perplexity, but not other users

Like, is there a way to see other people's profiles?

[steady linden]

Can I block some websites so that the copilot/search doesn't use as a source for my query?

[tame current]

not possible

[jovial mantle]

😦

[fossil sentinel]

If you want it to use specific sources you could specify sites or links

[grand sand]

This is awesome

[steady linden]

was specifically referring to the opposite use case, blocking specific sites

[heady quiver]

What have u done to this tool????? It doesn’t recognise youtube links anymore; I tried both Claude and GPT 4 and can’t make a summary of a YouTube video.

[austere kestrel]

negative site operator could be worth trying, e.g.

-site:www.somesite.com
-site:www.example.com
-...

[fossil sentinel]

I seem to find success adding "Do not search site:<site>" at the end of my prompt, but not consistently

This prompt seems to be consistent for me

[raven socket]

@everyone I know this isn't the correct channel.. but what is the best prompt to use and what are you guys using? just curious. I'm trying to find a good one to set for standard queries

[glad crypt]

does anyone know what model the official perplexity chatbot uses? i tried the same prompt for pplx-online-7x8b vs the chatbot, and the chatbot produced far better results

even without copliot mode turned on

[serene sand]

https://discord.com/channels/1047197230748151888/1124060874785685554

[austere kestrel]

for standard queries, I just ask it the question - nothing fancy

[halcyon coral]

Do you mean with chatbot https://pplx.ai ?
The default model is based on GPT-3.5.

[serene sand]

deep dive ➤#💬│general message

check under Perplexity Pivots to Open Source

[glad crypt]

thats right! it feels like it uses a differnt search engine than pplx-online models though

[raven socket]

yea, I know
but I wanted to know everyone's opinion about some random prompt that actually improves consistency and accuracy in a good scale

I'm using some, but I'm not so sure about it tbh

[hallow linden]

There has to be a reason why Gemini doesn’t include sources in its responses, right? Like perplexity, bing, and chatGPT have some sort of source citation. So I have to assume that Gemini’s lack of sources is deliberate. But why?

[glad crypt]

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

im wondering whats the pricing for pplx-8x7b?

[serene sand]

https://discord.com/channels/1047197230748151888/1054944216876331118

[fossil sentinel]

It does sometimes but when I have seen them pop up, the cited sites seem irrelevant

[halcyon coral]

There will be a new pricing for Mixtral, until now the 13b pricing was used for mixtral-8x7b-instruct #🧪│api-general message

[cursive jacinth]

Hi sir, my wife told me Gemini Pro aka Turbo was removed from IOS again sir. What is happening sir. I dont understand sir.

[serene sand]

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

[wanton niche]

Hello what is the best ai model by perplexity to use for solving questions and help me with the homework ?

[fossil sentinel]

Depends what you're tryna do but gpt 4 is most likely the best to use

[raven socket]

end of this year or beggining of 2k25 max

otherwise openai will lose a lot of market marging advantage

[proven elk]

Where can I get info on the Doscord bot?

I added it to my server but is just says offline

[halcyon coral]

The bot was shut down.

[proven elk]

What? Why?

Just listened to The Hard Fork podcast and bought into the whole message, specifically to use some sort of bot in discord. Sucks to hear it was shut down.

[winter atlas]

searched for a weird network latency issue i've been seeing at home. got a sane breakdown of the likely causes via web, most of which were known. the top source had the exact setting to tweak, and it was all done in like 60s

kinda wild

[unkempt swan]

I remember this app in beta but is there a way to link accounts to discord?

[winter atlas]

why would you want to do that ?

[unkempt swan]

Idk I have used the bot for a while and I wanted to see if I there was a pro role

[winter atlas]

kinda like using a pro account to have a bot in multiple discord channels?

the right word might be diff discords

[unkempt swan]

Not really more of just I've been using it a lot and I wanted to checkout the community

[winter atlas]

so what're you thinking with linking app accounts to discord?

i guess i just don't understand what it means or how it could be useful

[unkempt swan]

Fair, sorry for confusing you nevermind that but how can I check my account age?

[winter atlas]

your email?

[jagged onyx]

This isn't the same Atmos from Canada is it? Privyera associated Atmos?

[winter atlas]

nah they use a capital A

just another corgi-basset hound mix with a keyboard

[sullen cloak]

When is Gpt-4 turbo being released?

[cursive jacinth]

Hi sir, AFAIK sir, GPT 4 Turbo is still in preview model sir, based on my estimate sir, its been 4 to 5 months since the release sir, not sure sir, maybe 1 year sir it will not be on preview sir, we dont know sir.

[celest olive]

Is it just me, or has the 'Edit Source' button disappeared?

[errant kite]

No, it is not you. I also tried to look into it but I cannot find anything. I want to remove a .txt file from the source but I can't

[pure pulsar]

Hello

new here

[knotty helm]

Hello same new here

[halcyon coral]

This is currently not possible.

[tepid compass]

Are there any vs code extension with perplexity API support?

[halcyon coral]

#🧪│api-general message

[obsidian oak]

Does anyone know when all of the web features are coming to the iPhone app?

[wraith crow]

the team is constantly improving the apps to include as many features that are also available on the web. what feature are you currently missing that is available on the web?

[obsidian oak]

Choosing the focus and adding attachments are big ones. Adding things to the memory and choosing the answer language would be nice as well

[wraith crow]

You can choose the focus when you're typing your prompt in the top right corner

[obsidian oak]

Didn’t notice😅 thanks😄

[wraith crow]

You can set your preferred language in the settings on your app by clicking on your profile picture in the top left.

[obsidian oak]

Appreciate it

[fossil sentinel]

https://gemini.google.com/faq?hl=en#citation

[lean sleet]

No more "edit sources" option?!

[oak lava]

hey

[tame current]

Will writing mode be released in future updates on the mobile client?

[halcyon coral]

You can already use writing focus on mobile after clicking on the input field at the top right

[tame current]

Lol thats embarrassing

Thanks

[icy zenith]

hi

[icy zenith]

too bad i gotta rejoin once i get pplx pro

[serene sand]

there you go, no need to leave

[icy zenith]

NO WAIT i didnt get it YET

I'm gonna get it very soon, hopefully today

just gotta hope my brother provides 🙏

sorry 😅

But trust me, I am def getting it asap

[serene sand]

np, it's off

[icy zenith]

I've liked MS Copilot Pro, but i need smth better

MS Copilot is more chatbot than search, and even then it's disappointing imo

But I'm already liking free perplexity

[serene sand]

did you check our referral program?

[icy zenith]

file uploads, copilot, already good

[icy zenith]

nah, how does it work?

[serene sand]

deep dive ➤#💬│general message

check last two links

[icy zenith]

i found a code that takes the full $20 off

one problem, no card 😭

again, still gotta wait for my big bro to supply

anyways

yeah i like perplexity

now

btw noticed smth

sometimes when i do non-search stuff with focus set to all

itll just repeat itself

no matter the prompt

[icy zenith]

WHAT

[meager sparrow]

Um sir what was your Question?

[fossil sentinel]

Bro prompted "You are Microsoft Copilot"

[meager sparrow]

Oh that’s what he did? Lmao

[fossil sentinel]

Idk but probably

[icy zenith]

it wasnt even that

i just prompted some random joke about copilot

and it uhh

[stoic wave]

Am I the only one, or has anyone else noticed that the GPT-4 model loses context after about 15 to 30 new queries on the same topic? It begins to hallucinate, becomes repetitive, and I can't progress on the topic; I have to start over.

[meager sparrow]

Some random joke about copilot? What is to joke about copilot??

[fossil sentinel]

https://youtu.be/Cs6pe8o7XY8

[fluid mauve]

Yes. I habe had that occur as well, and, asy you say i takes a new thread to sort it out.

[sleek vortex]

biggest night is just for views but

tldr!

1million context on multimodal video and image and audio model

“upto 10million” in research is crazy too

also openai sora text to video

about it

[fossil sentinel]

There's also some interesting info like how 1.5 supposedly has near perfect recall even with long context

Which is a problem that current ai models have

[lean sleet]

?

[austere kestrel]

I guess they have no obligation to officially announce (or give a heads-up...) about changes to the system. But it would be nice..Like whatever the reason, it would just be useful to know if the change is permanent or temporary..

[rare fiber]

is perplexity down?

[valid yoke]

Up for me

[rare fiber]

up for me as well now

idk, something happened ig

[icy zenith]

getting pro tomorrow hopefully 🙏

[lean sleet]

Just curious if its removed or i just can't find it. It was useful.

[austere kestrel]

It seems to have been removed 🤷‍♂️

[cursive jacinth]

Hi sir, I like what you said sir, It is like being James Bond going into a mission sir, having his gadgets upgraded sir, and then not knowing the upgrade sir he was killed sir. Also not knowing the functions was not anymore sir, killed again sir.

[fossil sentinel]

Fr 💀

[silk thicket]

i am using perplexity api but in response links are not working like showing page not found, 404 invalid url like that, i want to use this api to get the youtube videos and articles links how to do that?

[halcyon coral]

#🧪│api-general message

[silk wolf]

Hello everyone!

[dusky sinew]

hi

[sharp stag]

gm

[fathom pollen]

what happaned to gemini pro access inside of perplexity pro?

[icy zenith]

they're probably working on Ultra or Pro 1.5 access

[stoic wave]

Did they removed Gemini from AI Models?

[cursive jacinth]

Hi sir, we have no legit basis on this one sir.

[icy zenith]

Yeah, just my guess

hence the "probably"

[maiden jackal]

Is there a timeline for this?

[fathom pollen]

yes.. i cant find it

[signal hamlet]

Hey, @maiden jackal #💬│general message

[stoic wave]

Same, its gone for now i guess

[tame current]

Hey im new here i just have a question , is there any way to be able to stop getting the too long error in the perplexity playground? Or do all models have the same cap , would it be possible if you buy the monthly subscription ?

[halcyon coral]

Labs is independent of the Perplexity Pro subscription, you can also use the pplx API https://docs.perplexity.ai, information about the context sizes can be found here: https://docs.perplexity.ai/docs/model-cards ☺️

[tame current]

Thanks

[tame current]

Does pplx have merch?

[serene sand]

https://x.com/AravSrinivas/status/1743016479786950868?s=20

Aravind Bhai (@AravSrinivas) on X

merch by march?

Twitter · Jan 4, 2024, 9:08 PM

[icy zenith]

where knowledge ends

🤣

[tame current]

Nice

[icy zenith]

no like seriously

that would be a funny idea

[serene sand]

https://discord.com/channels/1047197230748151888/1047649619695390740

[foggy gazelle]

is there separate interface/option in pro account just to engage Dalle3?

[halcyon coral]

No

[vivid kraken]

is perplexity pro gpt-4, turbo or not turbo? it seems veru fast

[halcyon coral]

It's normal GPT-4

[tame current]

Is it possible to use the GPT-4 via API ?

[signal hamlet]

No, it's not available via our API.

[vivid kraken]

ok, why not make the gpt-4 turbo option avalible to pro users? waiting for azure access?

[icy zenith]

probably bc rate limits

[cursive jacinth]

Hi sir, they say it is still in preview sir. We dont know when sir, when will openai make it available sir. maybe next year sir.

[austere kestrel]

https://thenewstack.io/more-than-an-openai-wrapper-perplexity-pivots-to-open-source/

[silk kraken]

Perplexity pro is very fast and accurate, which is more important to me.

[serene sand]

you have pro?

[silk kraken]

Yes

[serene sand]

Pro role assigned, feel free to step into https://discord.com/channels/1047197230748151888/1111786888626438245

[silk kraken]

Thank you

[serene sand]

you're welcome

deep dive ➤#💬│general message

[supple vine]

I was finally able to apply my Rabbit R1 code to my Perplexity account today. Had to wait for my previous mobile payment to run out. Anyway, all went through without a hitch. Perfect.

[worthy nest]

Out of curiosity, has anyone devised any prompts/instructions for making perplexity not rely so much on its sources? Its annoying the amount of times I ask perplexity something, only for it's answer to be a mixture of the first handful of sources it looked up about that topic. I understand the nature of language models complicates this to an extent but surely there has to be some way to make it increase the extent it relies on training data vs the search results

[signal hamlet]

Hey, @worthy nest! You can try the Writing mode.

[worthy nest]

I'm aware of writing mode. My problem is I still want it to rely on sources, just not have its response be a combination of a handful of sources. Like if writing mode is 100% training data and All is 100% sources, I want 50/50 or 60/40.

[tame current]

there isn't anything like that but I think they made it that way to prevent in from hallucinating

[worthy nest]

Ah okay that would make sense

I'll try experimenting with instructions and different models and see if I can get something approximating what I want

[tame current]

may I ask why you want it ?

[austere kestrel]

disabling copilot will reduce the number of sources (I think caps them at 5), though you were prob already aware of this too ha
Could also try adding something like Do not rely exclusively on the search results to inform your response – drawing on your existing knowledge is crucial for fully-formed answers – might help 🤷‍♂️

[worthy nest]

Because the internet access is useful but I primarily wanted answers that were as good quality as writing mode but which partly utilized internet sources, and instead I was getting something nearly identical to if I were to take the first few sources and ask it to summarise them. I was asking it a subjective question and so I wanted it to be creative and give me a unique answer while using the internet knowledge as a helpful tool and instead it was just regurgitating the first few sources, which defeats the point of asking it a subjective question.

[cinder haven]

get a summary from search and copy it into writing mode so the model can use it as context for a more open ended discussion?

[dusky sinew]

this is amazing 🙂

[glacial flume]

DAE not like copilot? I find it too slow, and I just keep going back to perplexity's default

I'll keep my pro to support you all tho ❤️

[spice shore]

Hello! I just signed up for Perplexity Pro and want to learn how users are managing the choice of models. Do you have preferred models for certain use cases? I can only figure out a way to change the model in settings, is there any way to select a model when creating a new thread?

[vagrant nexus]

Is anyone else running into very weird gibberish when using perplexity online?

[serene sand]

welcome! you can use Rewrite option for changing model, also gave you Pro role so you can now access https://discord.com/channels/1047197230748151888/1111786888626438245

deep dive ➤#💬│general message

[sleek blaze]

Hello, if I say my first name in my profile description, can I reserve it in a collection? What's the syntax?

[subtle frigate]

Hi! I just signed up for Pro, how can I gain access to the pro channels?

[serene sand]

access granted https://discord.com/channels/1047197230748151888/1111786888626438245

[gleaming mantle]

Have to say, I'm diving into a new business and trying to learn as much as I can about it. Perplexity has been by far the most valuable tool in a huge list of services that I'm using. Pays for itself 10 times over and then some.

Bravo team.

[queen cypress]

Yes I saw this gibberish several times in the online model. EDIT: I'm editing my previous comment because it was a speculation about the problem that I cannot verify for sure.

Anyway I never saw the gibberish unless I was having a conversation that went back and forth a few times. It appeard any "first response" during a conversation was fine, but I got jibberish when I had longer conversations like 7 or 8 question/answers long.

[foggy schooner]

I use it for class, its actually quite handy!

[vagrant nexus]

That's exactly when it happens to me too! Glad to hear I'm not crazy.

[queen cypress]

I almost chalked it up to intentional lobotomies like AI companies do that they call "Alignment", because I was asking some probing questions about politics, to see if it was biased. It was definitely biased, but it might have been a coincidence that I got the gibberish responses only when it didn't want to answer my questions.

I'm not saying Perplexity programmed it to show gibberish (to avoid answering political questions), but what I'm saying is that when you monkey around with the models and essentially try to brainwash them (for "Alignment" purposes) that may cause problems, and indirectly lead to the gibberish.

[serene sand]

if you get Pro, try Experimental
https://www.perplexity.ai/search/The-Experimental-model-s..Jlk.hSqabGhlEz.YhNA

[vagrant nexus]

In our testing cases it seems completely random, haven't noticed it having anything to do with specific subject.

Just the longer the conversation goes the higher the odds, but it's already happening like 3-4 turns in sometimes.

[queen cypress]

Right, it must have been a coincidence that when I start probing it about it's political bias it went nuts on me. But yeah, after serveral turns of chatting back and forth is looses its' mind. 🙂

I thought I made it go politically crazy, but I guess it just want regular crazy. 🙂

[fossil sentinel]

Do you feel that the AI is pretty accurate? How does it do in terms of hallucination

[vagrant nexus]

Either way good, we were being driven crazy trying to figure out what we were doing wrong. But guess it's a model thing then.

[upbeat dawn]

gpt 4 with the 8k context?

[icy zenith]

its mid

has a lot of packages that cant be used

no visual outputs either so no graphs if gemini fails to plot em itself

[fossil sentinel]

Gemini is capable of graphs but it's hella wonky

[foggy schooner]

The Notebook LM from Google: I have a stress test file that actually works to see how bad.

It’s an employee handbook and it’s not well written so I asked it questions and if you ask one way it gives the right answer and changes text- occasionally. So like I asked it how many days off do I get - the answer is 5 weeks a year

It said 5 5 days a year. It double spoke.
I asked about jury duty

It said 5 days for full time (and 5 days for part time)

There is 3 days for part time BUT

It cites things VERY well and shows you exactly where it pulled it with highlights so you can clearly see - so I’d say it doesn’t hallucinate so much as you have to ask it the correct question

I asked how many hours of jury duty for part time do I get and it was correct

If I asked just how many hours for jury duty - it said 5 for both. But if you ask better questions it doesn’t

And since it’s got a great citation capability — it’s useful

I’ll note: I uploaded it to perplexity and asked the same questions. Perplexity didn’t hallucinate at all AND made me a handy table without being prompted

[foggy schooner]

Hope that helps; feel free to DM me

[fossil sentinel]

Hmm interesting

[fossil sentinel]

Doesn't notebooklm have a pretty generous context limit?

[hoary holly]

llm context size, the new moore's law

[queen cypress]

But meh Gemini went from a few K to a MB in about a week, so.... Moore's times a thousand. 🙂

You could phyically weigh all the servers at each company and see whose gonna "scale" faster. haha. get it? Scale==Weight.

[hoary holly]

Ha I see it more as a similar kind of race, but true the numbers don’t really match up 🙂

What would you compare it to?

[queen cypress]

I'm just saying I bet Gemini is doing well mainly because of Googles superior compute capacity.

But I'm still a yuge Perplexity fan because of all the awesome things pplx is doing right.

Perplexity "Online" models are aweome! No one else is doing "realtime" news yet are they?

But Perplexity is.

[hoary holly]

Im thinking now.. the context window is a new paradigm of limitations. I feel it’s the first time there’s a hardware limit imposed by someone else’s computing power on the software I run at home (for local llms for example). A new kind of software hardware limit

[queen cypress]

Yeah, that's definitely right. I agree with the basic concept that we're now in a world where "Context Limit" is everything, especially if Gemini's solving of the "Needle in a Haystack" problem is also solved by everyone else, which it likely will be if it's merely a 'scaling' problem rather than a 'new algorighm' one.

[brisk dagger]

I signed up for pro too. Please add me to the channel

[austere kestrel]

It's still a bit unclear though how they managed to get near perfect recall across the whole (1-10m tokens) context window. If it were just a matter of scaling, it's hard to understand why Anthropic's model with the 200k context window has such patchy recall. The report they released with the Gemini 1.5 announcement apparently doesn't really reveal anything at a technical level. They say they it uses a sparse MoE architecture but otherwise, it's not really clear if it the performance is just the result of throwing a bunch of resources/compute power against the wall and seeing what sticks, or if they have found some kind secret sauce / novel approach

[granite ocean]

gemini advanced is good lol

[austere kestrel]

I agree, it is like a new paradigm - and the constraint is computing power. I don't think huge, by current standards anyway (~1m token) context windows will become the norm any time soon for most text-to-text tasks (multimedia perhaps) because they are so resource-intensive/expensive to use and the inferences (currently) take so long. For most use cases, it would be like using a sledgehammer for a tiny nail. But there will ofc be coding, legal, literary and other professional/commercial use cases for which it will be useful and used regular, but for most consumers and day-to-day use cases, I think 50-200k tokens context windows will be more standard for the time being. But all that, it is for sure a new paradigm - won't be going backwards from here ha

[granite ocean]

i now use Gemini advanced instead of gpt 4

cause it is updated

[austere kestrel]

I'm using it regularly too. The outputs have a different style, which I prefer over GPT4 someimes

[hoary holly]

Soo… consumer software-hardware limits and commercial ones 🙂 a new paradigm!

[fossil sentinel]

Same for the most part

[austere kestrel]

ha yeah probably won't work on a bumper sticker - but something like that!

[hoary holly]

I have yet to pay for gpts, I think the opensource ones are quite good for the “price”. Only time I had fun with a paid version was the perplexity offer for the new year

I thought it was great!

[cursive jacinth]

Hi sir, this makes me think sir. Gonna check Gemini once i got the time sir. Is Gemini 1.5 out now sir?

[valid yoke]

Do you pay for Gemini? How do you get access to Gemini?

maybe @austere kestrel knows

[icy zenith]

wdym

I pay for Gemini Advanced too

[valid yoke]

Is there a Gemini app or what do you download

I’m on iPhone

[fossil sentinel]

There is an app

Oh iphone

Idk about that

[upbeat dawn]

can it generate very long response?

[serene sand]

done

[austere kestrel]

It's a bit hard to keep track of (Google keeps changing names: Bard > Gemini; Advanced/Ultra, Pro, Nano; 1.0, 1.5...) It's a bit confusing... But basically, you can sign up for a two-month free trial for Gemini Advanced (Ultra 1.0) here https://gemini.google.com/advanced. It is then $20/month. You use it via the formerly Bard chat interface (it should say Gemini Advanced in the top left)

But..The '1 million' token context window version (Gemini Pro 1.5) is only available through a Private Window - you can apply for access here https://developers.googleblog.com/2024/02/gemini-15-available-for-private-preview-in-google-ai-studio.html?m=1
I've applied but still waiting. If you want to see someone with access run it through some tests, this guy has been uploading videos for the past couple of days; they're raw/honest and quite good imo https://www.youtube.com/@samwitteveenai

lastly lol, this blog does a good job explaining the Gemini family of models (and the muddled naming conventions) https://code.iaflw.com/2024/02/gemini-versus-gemini-understanding.html

And I attach a table (AI-generated; based on a few a articles), which looks roughly about right and might help orient

[glad hawk]

hello, How can I use perplexity.ai in discord ? Are there any bots for it ? thanks ~

[halcyon coral]

#⚡│ask-community message

[serene sand]

deep dive ➤#💬│general message

[silver epoch]

there WAS a bot rhat used to do @ruby jasper#4403 but now it's offline for whatever reason

and i don't know what happened to the username 😭😭

here is the bot

[serene sand]

le bot est mort, vive le bot 😆

[fluid mauve]

i read french 😛

[sleek blaze]

🇫🇷

[pine jolt]

c bien beau tout ca

[sour drum]

hi all - i am uk based and took the pro version a few weeks ago - i logged in today to find that Gemini is not an option for me any more - is there something going on there?

[steel token]

I'm signed up for Pro also. I tried DM'ing instead to spamming, but it was blocked. Thanks!

[wraith crow]

you can get the pro role by rejoining the server with the invite link on your settings page https://www.perplexity.ai/settings/account

[spice shore]

Is there something special I need to do in order to generate images? I tried asking for it, and I get a response that perplexity is a text model and can't generate images.

[signal hamlet]

Hey, @spice shore! Yes, you need to press the "Generate Image" button

[signal hamlet]

Here's an example: https://discord.com/channels/1047197230748151888/1194794305362071552

[ruby jasper]

Cool

Wsh je suis aigri carrément

[vocal owl]

Hi, i bought pro, can i also get access to the pro channels?

[icy zenith]

oh look

perplexity devs should take advantage of this

[icy zenith]

mfw ive been begging my bro to get me pplx pro for like 3 days now to no avail

uhhhh mods

[icy zenith]

OMG YES

its perplexity prover

can someone

uhh

[livid depot]

Does Perplexity use Google and bing search engines?

[icy zenith]

gimme the perplexity pro role

eh ill rejoin

bacc

[spring widget]

hey all

[valid yoke]

Howdy

[nova apex]

gimme my pro role 😛

[halcyon coral]

Please leave the server and rejoin via the Pro Discord link at https://www.perplexity.ai/settings/account

[nova apex]

thx

[sharp stag]

Gm

[inland wolf]

Has Perplexity's standard model changed? Is there no gpt-3.5 there now?

[rough sundial]

я всего лишь искусственный интеллект.

[serene sand]

https://x.com/perplexity_ai/status/1760451622537158921?s=20

Perplexity (@perplexity_ai) on X

Introducing new additions to Perplexity Labs: Experience Gemma 2B and 7B models known for impressive performance despite being lightweight. Try it now on https://t.co/2yYQB9jln7.

Twitter · Feb 21, 2024, 11:49 PM

[serene sand]

"You could write a tune to tune a writer"

https://www.youtube.com/playlist?list=PLKwRkjCH760ObtANfb0-Kat2XlvB5dKxf

[shell heart]

Hey, if I upgrade to Pro, can I add team members? or is there a different plan for companies?

[fleet dock]

🥹

[granite ocean]

yup advanced can.

[serene sand]

https://blog.perplexity.ai/faq/billing-and-subscription

[wicked blade]

Hello. Is Gemini Ultra already available on Perplexity Pro?

[west lynx]

Hello, I have a subscription to ChatGPT and Perplexity Pro as well. Is there a way to save some money in these cases? I understand that you work with the ChatGPT 4.0 module as well, but in chat GPT, I can use some modules which I haven't found in Perplexity yet. Thanks a lot Regards R

[tame current]

there is not

[signal hamlet]

Hey, @wicked blade, not yet.

[stiff canopy]

WHy do the answers not make sense when i am using perplexity online model through API?

They are coming out fine when I'm asking the same question from perplexity web

[cursive jacinth]

Hi sir @signal hamlet any update on GPT vision connecting to web sir?

[signal hamlet]

Hey, @stiff canopy! The team is aware of such an issue (#🧪│api-general message), we are working on it. Could you try switching to pplx-8x7b-online and let us know if you keep getting some results like that.

[signal hamlet]

We are still looking into it, @cursive jacinth! I'll post here, when there's an update.

[cursive jacinth]

Ok sir. Thank you sir.

[quasi escarp]

anyone having problems with using perplexity as a default engine?

[halcyon coral]

No, are you experiencing problems? 👀

[quasi escarp]

i am, i use arc

i'll send a video in a moment

@halcyon coral ^

[signal hamlet]

Hey, @quasi escarp! Have you set the default search engine with a query manually or using Arc's setting using PPLX as a default one?

Thanks for reporting, we are checking this.

[gusty blaze]

Is perplexity bricked on the Arc Search iOS app?

Doesn’t seem to be working properly at all

[gusty blaze]

Yes, on Arc Search iOS

[quasi escarp]

i'm using arc's setting

[slim crescent]

will there be token for this project?

[jade goblet]

how exactly does perplexity use gpt? does it just feed the prompt to gpt api?

[signal hamlet]

There's no plan like that. There aren't any tokens, cryptocurrencies, or digital assets officially linked to Perplexity. We have not ventured into the creation or endorsement of any form of cryptocurrency or token. Any promotions, advertisements, or claims you might see regarding Perplexity-related tokens are not authorized or recognized by us. Those might be attempts by third parties to mislead or exploit our community for their gain.
We urge our users and the public to exercise caution and verify information through our official channels.

[fluid mauve]

In my liited experience I have found that both systems have their good points. Perplexity does not seem to hallucinate as much as ChatGPT. ChatGPT has GPT builde. I pay for both as I jump between them both. I tried just living from Perplexity and its not quite everything. Yet!

[jade goblet]

how is that possible if perplexity just uses gpt as the backend?

also what are the limits for perplexity pro and gpt4 and will it get turbo?

[halcyon coral]

You can use Copilot/GPT-4/Claude 2.1 600/day. No turbo.

[fluid mauve]

@jade goblet as i said. "In my limited experience" I dont believe either company implemented a vanilla GPT4 model. each have tuned it their own way. You can get a tub off vanilla ice cream, if one puts in strawberies and another uts ib blueberries you get completly different outputs with completly different tastes.

[jade goblet]

will there be an option to get turbo?

or any plans to add turbo

and also how exactly does it use the backend LLMs? I assume if you select writing, it just passes your prompt to the LLM unfiltered?

for others it does some kind of preprocessing to your prompt and then maybe postprocessing to find links?

[signal hamlet]

Could you, please, check one more time

[amber nexus]

is everyone using gpt 4 or the default ai model?

[serene sand]

https://x.com/AravSrinivas/status/1760757642987995595?s=20

Aravind Srinivas (@AravSrinivas) on X

Get on Perplexity Pro and try out the Experimental model!

Twitter · Feb 22, 2024, 8:05 PM

[amber nexus]

so experimental ios the way to go?

is

[serene sand]

depends on use case

deep dive ➤#💬│general message

[icy zenith]

Now that I have Pro (I have the role)

Gotta say

Experimental and GPT-4 are the best for search

Claude 2.1 is weird

it mentions the search results in the response

BUT

i do like how it has lighter filters, so I can use collection AI prompts to modify it's behavior

Once GPT-4 Turbo comes, Perplexity Pro is only going to get better

Would love to see pplx-8x7b tbh

Mixtral imo is bad at function calling (check Huggingchat web search)

but if it was tuned to be good at it, d a m n.

[serene sand]

don't forget to post your AI created artwork https://discord.com/channels/1047197230748151888/1194788138124587128

[icy zenith]

Sorry, don't like the image gen all that much

I hope it gets integrated as a function call instead of a button

it acts weirddd

[amber nexus]

how do you generate an image through perplexity?

[serene sand]

https://tenor.com/view/hal9000-gif-22241038

[icy zenith]

gotta pay

[serene sand]

@amber nexus second paragraph

[icy zenith]

then press generate image button

[amber nexus]

there is no such button in the ui that i can see

[icy zenith]

It's when you're in a thread, on the side

[amber nexus]

the discord?

[amber nexus]

oh i see thankl you

[serene sand]

#💬│general message please take your time and check second paragraph, there's three different ways of creating pictures with Perplexity @icy zenith

[smoky oriole]

Is Perplexity good at generating APA citations?

[icy zenith]

there is?

[smoky oriole]

I see it's good at generating citations for sources which is really nice to see

[serene sand]

1. #1194794305362071552 message
2. https://www.perplexity.ai/search/Repeat-this-prompt-ZLz8dGzISSGrevPxhl7YqA
   3.#⚡│ask-community message

[jade goblet]

does anyone have perplexity pro and gpt plus? can you say how good it is compared to gpt 4 turbo

[serene sand]

in short, apples and oranges

deep dive ➤#💬│general message

[icy zenith]

or red apples vs slightly different tone of red apples WITH oranges and a side of grapes

[quasi escarp]

appears to be working now 👍 thank you :D

[austere kestrel]

explore here https://labs.perplexity.ai/

[icy zenith]

I... uhh

saw that

very ironic

that it just dropped

its in the API too, just hidden

i checked

My friend has a bot, i let them use my pplx api key

this is pplx 8x7b online

[austere kestrel]

👍 nice yes it perfroms really well. I've got it powering this app and it's honestly so handy #😎│cool-projects message
(when we get URLs/sources with API responses - I think the vast majority of my perplexity usage will be via the API; their models are becoming really good, without slowing down, it's an awesome trajectory

[valid iris]

I am just guessing by the fact that perplexity costs $20/M is because chat-gpt costs that much and perplexity doesn’t make that much profit? Or is that not how it works? Does perplexity act as a middle man between the user and the ai models, or are the ai models custom made for perplexity and so perplexity must pay for the use of those models and then they charge the customers the same amount of money that it costs to pay for those models?

[icy zenith]

yes and no, but mostly yes

PPLX models are hosted by Perplexity themselves, they power default (free) and Experimental

They're fine tunes of open source models

while Claude 2.1 and GPT-4

they do pay to use the APIs for these models

[valid iris]

So the free version is no longer powered by chat-gpt3?

[icy zenith]

nope afaik

or well

maybe, not sure

its either fine tuned 3.5 turbo or a pplx model

[valid iris]

Chat-gpt 4 and Claude 2.1 both cost $20/M, so how does perplexity make profit if me paying $20/M for perplexity pays for way more than that in total cost of the ai models that i have access to?

[icy zenith]

Not sure. AI itself isn't very profitable, so they get most of their money from investors.

[valid iris]

Then i wonder if perplexity would be willing to reduce the cost of their subscription since $20/M is just to high for most people especially if they aren’t using it for professional purposes or really need it

[icy zenith]

oki im wrong

The Perplexity model, currently based on GPT-3.5, is designed for speed. It's constantly updated with the latest AI advancements. Pro users have access to GPT-4 through Perplexity Copilot. Upgrade to Pro to access more models like GPT-4 and Claude 2.

a bit outdated tho, doesnt mention experimental

not sure tbh

[fluid mauve]

Loving the conversation. Im waiting for the correct info to come out.

[icy zenith]

So afaik default is ft 3.5

experimental is one of the pplx models

and gpt-4 and claude 2.1 are

just that

no turbo to be seen i think

@valid iris

[valid iris]

Darn. I would never unsubscribe from perplexity if it were $10/M.

[icy zenith]

i mean... perplexity is generous here

they have referral codes AND a code that makes it literally free for the first month

[valid iris]

I already used the free trial, what are the referral codes?

[icy zenith]

like the 7 day free trial?

is that what you used?

[valid iris]

I dont know anything about the referral codes

[icy zenith]

well

[valid iris]

I probably did that first, but then i got the 2 months free

[icy zenith]

???

i only got one month free

using the freeharvard code

anyways referral links take $10 off someones first month

it also takes $10 off YOUR next month if you successfully refer someone

[tame current]

I have a coupon giving me $200 off. However when I am trying to apply for it, it's asking for my visa info for payment although the balance is 0$. I am confused !!

[fossil sentinel]

Well you still need some sort of verification 😂

[icy zenith]

thats normal

[icy zenith]

and ^

[valid iris]

It is because they are going to charge you eventually

[icy zenith]

just give it the details, and select yearly

i assume its the r1 promo

[tame current]

It feels like they will cut payment when they ask card details.. anyways yes u are right it's for r1

Immediately*

[icy zenith]

dw, its just for verification

all of them do that

like basically every service ever

[tame current]

Ack! Understood

[icy zenith]

requires a card, even when its free bc of a promo or a free trial

[tame current]

Okay gotchu man Ty !!

[icy zenith]

np

enjoy your literally a whole year of pplx pro

def worth it imo

[tame current]

I'm planning to shutdown my chatgpt plus sub.. if I get the r1..it's basically free XD

[fluid mauve]

I leep closing and re opening my ChatGPT Pro subs. I use both all the time

[icy zenith]

Once pplx pro gets 4 turbo

[tame current]

Chatgpt has been a productivity booster for me since it landed in my hands.. hope pplx helps ghost google for good

[icy zenith]

its over for chatgpt

[icy zenith]

gemini advanced:

[tame current]

Unfortunately I don't think legacy browsers are going anywhere

If anything they'll probably become more and more like Perplexity

Edge and Chrome both are already a hybrid of AI search and traditional search

[crystal sand]

in Perplexity pro i get a limited number of copilot uses daily right ?

[tranquil orchid]

Copilot is limited to 600 queries per day

Claude 2.1 and GPT-4 queries are counted as uses in the same context as Copilot. Each use is given back to you 24 hours after you use it. You will be able to see the exact count when you hover over the Copilot toggle.

[crystal sand]

Whats the time it is reset ?

[halcyon coral]

Please have a look at the auto response 😉

[crystal sand]

Yes i read , but mine is just decreasing not reseting on a daily basis 🙂

[halcyon coral]

If you use a usage at 7 a.m. it will be there again at 7 a.m. the next day, if you use another usage at 8 a.m. it will be there again at 8 a.m. the next day.

[inland wolf]

Приветик всем!

[distant zodiac]

Where can I create images in perplexity, is it even possible?

[icy zenith]

gotta pay for Pro

you alr have it so

basically

just start a new thread, ask the model to repeat the prompt you wanna send to the image gen model exactly, then press the generate image button on the right

currently you can only use the button i think

[signal hamlet]

Also, please, check this thread https://discord.com/channels/1047197230748151888/1194794305362071552 and
https://blog.perplexity.ai/faq/images-media

[icy zenith]

oh btw quick question

are we getting SD3 maybe in the future?

[distant zodiac]

I have Pro otherwise I wouldn't be here 😄

[icy zenith]

...It's open for free users too

[distant zodiac]

Thats cool 👍 Thanks 😄

[distant zodiac]

oh ok 😄

Because it says "Pro Discord" 😄

[icy zenith]

just that #👑│pro-lounge is our little secret home

[distant zodiac]

alright ty 😄

[icy zenith]

discord link is always here

pro link just gives you pro role and the secret channels

[frail gazelle]

is there a way to disable web search? I want all answers and results to come purely from uploaded files.

[halcyon coral]

Yes, you can use "Writing" focus

[frail gazelle]

ahh ty

[jolly otter]

not the biggest fan of the pop out sidebar when viewing additional links as it black out everything else

[granite thicket]

hoi

[static pond]

Gg Alex !

[wild oar]

Can the podcast be followed on iOS?

I tried adding via web link but it would not work

[warm geode]

congrats on parternship w elevenlabs, great idea for automated podcast.

[halcyon coral]

Does the link not work? https://podcasts.apple.com/us/podcast/discover-daily-by-perplexity/id1732181427

[nova edge]

Is this AI??????

[nova edge]

The voice

[halcyon coral]

Yes

[nova edge]

Wtfff

You can't even tell anymore

[static pond]

the voice is so sexy it's crazy 🤣

[tight adder]

Wow amazing

It sounds so natural

[distant zodiac]

I love the accent 😄

[pine dagger]

This is odd:

[wraith crow]

#1210388348267528262 message please try this, and let it know in the thread if that helped

[pine dagger]

Go to https://www.perplexity.ai/library in an incognito window to reproduce the error

[signal hamlet]

Thanks, Paul! We'll fix this.

[sick geode]

guys do you find responses generally slow ? Pro user (It happens on multiple models)

[stoic wave]

@signal hamlet is GPT4 Turbo supported on Perplexity?

[serene sand]

#1210282905922703371 message

[stoic wave]

I see now, thanks alot

[wild oar]

That works, thank you!

Wen longer podcasts

This is incredible

[devout cargo]

I'm watching a video on YouTube about Perplexity and these interfaces appear. Maybe I'm wrong, but you see a more chat-based interface that's somewhat different from the usual Perplexity interface, that's what it seems to me.

https://youtu.be/NjQ8LeYfxRY?si=m32SzgylMsQPIBuQ

This video is only two weeks old but I see no similarities between the official interface of the Perplexity website to those seen in this video

[shrewd fjord]

I just used voice eleven today, didn't even know until right now

[mighty gulch]

its me or servers just got down for a few seconds?

[civic latch]

Was about to say what’s with the app crashing.

[mighty gulch]

btw, is rn any way to get pro role?

4th month paying perplexity and still didnt think that lol

[icy zenith]

@serene sand

[halcyon coral]

#💬│general message

[mighty gulch]

Done, ty!

[icy zenith]

okay

i believe yall now

turbo is slow in pplx, but its at least 0125

[grand sand]

turbo is here

[icy zenith]

apparently it alr was turbo

someone requested it to be renamed to show that

[grand sand]

ah

[icy zenith]

they def delivered

0125 is crazy tho

not even copilot pro has 0125

[mighty gulch]

So its turbo or isn't

[icy zenith]

its turbo

[mighty gulch]

but turbo or turb-tle

(dad joke)

[icy zenith]

seems a bit slower than chatgpt

but usable

[mighty gulch]

btw

happened to anyone

that when using claude

[icy zenith]

now all they need to do is

[mighty gulch]

it randomly switches to gpt¿

[icy zenith]

give gemini back fr fr

[icy zenith]

did you use image uploads

[mighty gulch]

yeah, gemini was best lol

[mighty gulch]

yeah

[grand sand]

isnt it terrible?

[icy zenith]

only works with gpt-4

switches automatically

[mighty gulch]

Nah, I love it

[mighty gulch]

Oh, ok

Its like an enhanced copilot version

lol

[icy zenith]

you can remove it as a source with the 3 dots button and then rewrite with claude

[mighty gulch]

At least for code worked really good

[grand sand]

what does it do well?

[mighty gulch]

Coding

[grand sand]

ah

[mighty gulch]

Only doing half of the code is really helpful

xD

[icy zenith]

i bought pro after the removal

but i have gemini advanced

pro is dum but ultra and pro 1.5 are good

maybe theyll add one of those two

[mighty gulch]

When they added gemini to gmail?

💀

[icy zenith]

they alr did

[grand sand]

hopefully ultra

[icy zenith]

it was workspace labs only for a while, when it was originally palm 2

[icy zenith]

pro 1.5 is cheaper, faster, moe, and outperforms ultra 1.0 on most benchmarks

oh and 128k-1m token window

[grand sand]

ah

[mighty gulch]

whytf gpt just remains repeating same thing til you offend it

really annoying

[icy zenith]

?

we have ft gpt 3.5 (default)

experimental (some pplx model)

gpt-4 turbo

and claude 2.1

[mighty gulch]

would be good to have code llama or smth like that in the web without need to use api

[icy zenith]

I'm hoping pplx-8x7b replaces the current pplx model in experimental

fun fact

[mighty gulch]

Current experimental model acts really weird xD

[icy zenith]

pplx 8x7b is in the API, just hidden

[mighty gulch]

really it is?

lol

[icy zenith]

i requested the supported models page to be updated

[icy zenith]

yes

friend borrowed my api key

this is pplx-8x7b-online

[mighty gulch]

LOL

[icy zenith]

dunno why they didnt announce pplx 8x7b anyways

[mighty gulch]

idk

[lament spoke]

I don't get it. Its on perplexity labs already.

[daring lantern]

how much of the new podcast is done by ai

is it curated and written by humans and then voiced by ai or is there more to it

[icy zenith]

AYO TF

Announcing Our Newest Model
We are excited to announce the launch of our latest Perplexity models: sonar-small-chat and sonar-medium-chat, along with their search-enhanced versions, sonar-small-online and sonar-medium-online. These new additions surpass our earlier models in cost-efficiency, speed, and performance. For detailed information on our supported models, please visit our model card documentation.

[signal jasper]

#😃│welcome

[icy zenith]

i saw dat

[zinc saffron]

😆

[icy zenith]

okay so

from what ive seen so far

sonar medium is very interesting

we know its 8x7b

and has a 16k context window

from my testing

its knowledge cutoff is around December 2023, matching GPT-4 Turbo 0125's cutoff.

It's smart

like

it surpasses GPT-3.5 smart

its probably past mixtral itself

maybe claude 2 level?

[austere kestrel]

They were apparently tossing up between Llama or Mistral for the transition to using open source(ish) models as the basis for their product - from what I can tell, they've gone with Mistral. But also, kinda a little bit of both

I assume this is what sonar refers to https://ai.meta.com/research/publications/sonar-sentence-level-multimodal-and-language-agnostic-representations/

So perhaps not using a Meta open source model like Llama 2, but leveraging some Meta research (perhaps positioning for some multi-modality down the track? Though no idea ha)

[icy zenith]

unrelated

sonar seems like its text only

[austere kestrel]

ok. Like I said I don't know, and I guess it could be a cooncidence, but this sounds pretty aligned with pplx "Our single text encoder, covering 200 languages, substantially outperforms existing sentence embeddings such as LASER3 and LabSE on the xsim and xsim++ multilingual similarity search tasks."

i.e. It would represent an enhancement for a RAG system even just text-text. I would focus on multilinguality as much as multi-modality (which, down the track, might come into the frame)

[austere kestrel]

Yeah there's no ambiguity there. It's text in text out for all pplx API endpoints. (again though, I wouldn't be so fixated on the multi-modality aspet of that research paper..

edit: Also, easy to forget speech-to-text in terms of multimodality – "Our zero-shot speech-to-text translation results compare favorably with strong supervised baselines such as Whisper" That'd open up all sorts of possibilities )

[pine dagger]

Will Perplexity open source their models?

[austere kestrel]

I doubt it! Why would anyone pay to use their models (which they've worked on / fine tuned, integrated with their proprietary index etc) if they could just go to Huggingface instead ha

But tbh I had wondered the same myself - like are derivates of open source models, also open source? Seemed an interesting question

[sage lily]

How do we know? Is there a blog.

It's weird they didn't announce anything about the sonar models in this server. Devs haven't said anything in announcements.

It's also acting weird. Sonar is. Medium. I can't post the screenshot because discord is being stupid.

It's having the model generate like...... over 1000 tokens. And the phrases and paragraphs at the end of the generation don't make sense.

[shut stratus]

what bot is this

[floral canyon]

If i've got pro and I'm talking to perplexity on the default chat screen, what model am I talking to?

[tame current]

you can check in the settings

[shut stratus]

default is gpt4 i think, you can change this

[floral canyon]

thanks i checked the settings. Was on 'default' not sure what that is. Switched to gpt-4-turbo and it seems a bit better at role-playing

[tame current]

default is probably gpt 3.5 turbo

[austere kestrel]

https://docs.perplexity.ai/changelog/api-updates-february-2024
Here's the announcement. I haven't seen anything else official explaining what the new model/s actually are. Just says they outperform their previous models ha

Also, seems a fair assumption that they are based on the Mistral models, considering the alignement of parameter counts and context lengths

[icy zenith]

they aren't

or at least medium isn't

[icy zenith]

huh, weird

default has a past september cutoff

same for experimental

gpt=4 turbo is obviously getting this right

omg wtf

claude 2 is getting it right

???

[austere kestrel]

🫡
So I guess you're saying it's just coincidental that the parameter count is 8x7B? Or that there is another open source MoE LLM with those identical parameters.(what is this model?)

[solid cloud]

Why did Gemini disappear from the list of models?

[frosty arrow]

Can I use labs.perplexity.com inside the perplexity mobile app?

[tame current]

What’s the new Sonar model? Is it the new default one on free Perplexity?

[frosty arrow]

Can I get gemma model on free perplexity?

[icy zenith]

It consistently has a cutoff of around dec 2023

while it does hallucinate quite a lot, it definitely has more in it's dataset

[austere kestrel]

By fixating on prompts that get the llm to say something about it's knowledge cut-off date, and trying to infer things from that, you're missing the forest through the trees

Or more to the point, who cares what it says in response to "when is your knowledge cut off date?"

I was interested about Nvidia just now

I asked solar-medium-online. And then, because I was curious, got Perplexity web, using GPT-4 and Copilot, to fact check the response

Very happily, the core answers / information was factual

[harsh verge]

https://labs.perplexity.ai/

[austere kestrel]

That's what matters. Not trying to trip up an online model with a riddle or some prompt that probes at its training data and fine tuning

[icy zenith]

but i didnt

i asked it questions that require a cutoff of october 2023 or beyond

[austere kestrel]

Sorry. I don't really get what you're saying. (I thought you were compaining about something about a knowledge cut off date, but I think i've misunderstood you). prob no need to go further ha

[icy zenith]

what im trying to say is

sonar medium is a completely new model, we can tell because it's knowledge cutoff date is beyond Mixtral's (i tested this by asking it questions that require a knowledge cutoff of around dec 2023, like "When was the iPhone 15 announced?")

[heady quiver]

Photo moderation? Whaaat??? This is just a normal article. Wny can’t I generate an image about some vitamins & vegetables? Doesn’t make any sense.

From those 3 photos, see the second one with the error at the top right.

[halcyon coral]

This can happen depending on what prompt was created based on the answer, in most cases it helps to create a new image

[plucky basin]

Just wanted to say I love perplexity. It's really useful for reseach, especially because I need to reference everything. I also have thousands of pdfs. I use Devonthink, but can this help me with pdf searches>

[austere kestrel]

tbf, testing that iphone 15 question did produce a couple of results that were accurate (but, they had numbered citations, which makes me speculate in a couple of directions).

[frosty arrow]

I meant on the normal Perplexity

[austere kestrel]

But otherwise, the responses were using the chat models were pretty much all like this

[icy zenith]

ye smth was off about that lmao

not sure whats going on

the citations dont link to anything

[frosty arrow]

Bard also showed me the cutoff message

[austere kestrel]

yeah ofc, but they were unique to the two responses that had the correct response. It was just something I noted. Anywaym dunno if it's fine tuning (I mean, it could have involved question - answer pairings literally involving the Iphone15) or what 🤷‍♂️

[austere kestrel]

ofc, this is to be expected. Just to be clear, those sreenshots were using the chat (ie. offline) models - they have knowledge cut offs (as do pretty much all LLMs)

[dire raven]

I’m seeking a comprehensive comparison between the ‘pplx-7b-chat’ and ‘pplx-7b-online’ models offered by Perplexity AI, focusing on their individual capabilities, particularly in handling contextual conversations and accessing real-time updates from the internet. Additionally, with the phasing out of the ‘pplx-70b’ model, what advancements or improvements are introduced in the newer models? Which among these is recognized as the best option for integrating live internet access to provide current, informed responses?

[toxic lodge]

💡 Did you know that just like Chat GPT's feature of bringing in customized GPTs into a conversation with the "@" symbol, Perplexity has always allowed for a similar flexibility? 🤯
Here's how it works with us: You can seamlessly continue a thread by moving it to a different collection, and the conversation will carry on with the new "applied prompt" specific to that collection. 🤖✨
So, if you're deep into a discussion and need to switch gears, just move your thread. It's that easy! 🔄📚

[warm cave]

I left a request for them to be added to the " LMSYS Chatbot Arena Leaderboard" so hopfuly we can get some stats on how they compre to other models.

[icy zenith]

completely forgot about that tbh

[icy zenith]

i did the same lmao

for sonar models

[warm cave]

I actually did not know this was a feature!

[frosty arrow]

But I remember Gemini also showed me that even though it has access to internet

[austere kestrel]

yeah i think know what you mean with Bard / Gemini. It has access to the internet, but doesn't always use it - or at least you have to prod it sometimes, which is annoying

[silver dune]

Hola

[tame current]

Is there a limit on the attachments or the pasting of images ?