💬│general

[agile jay]

Yep, I've mostly given up on closed source models.

Even if they come out good, they get worse over time...

[lone citrus]

It's Opus, what made you think it's not real Opus may I ask?

[limpid warren]

I also subscribed to Claude pro. In terms of experience, Opus's perplexity has decreased a bit, and occasionally there are hallucinations or answers that are opposite to the original version.

[agile jay]

They use different system messages, and probably temps too.

System messages can affect the output by a lot.

[floral lotus]

test it against the opus on chat.lmsys.org

[limpid warren]

Thank you, I will compare it.

[ember sparrow]

I have a billing issue. Who / where do I reach out to get it fixed?

[agile jay]

[e-mail]

I think

[ember sparrow]

Ah okay. Tried that and didn’t hear back so came to this discord. I’ll send another email.

[agile jay]

You can also try #1140622008086970420

[steel tiger]

anybody noticing GPT-4 is really slow lol?

http://i.popz.top/u/FbgNhVduSoOI.gif

[agile jay]

Hasn't it always been slow?

[tame current]

A lot more services will probably switch to that and autogpt will increasingly become more pointless

[steel tiger]

nah it was pretty quick for me most of the time

[agile jay]

I guess they all feel slow, since I'm used to groq speeds.

[steel tiger]

on the backend doesn't gpt have a timeout window too - so its kind of weird that perplexity can keep pulling so slowly

[tame current]

Gpt is rll a 1 year old model bc it took so long for anyone to access it

[agile jay]

They already had access to it one year earlier.

In comparison the llama models are released a week or so after training is finished.

Doesn't make sense to release a better model if you are already ahead, lol

Especially if all the other companies are using your top public models to generate training data.

[tame current]

With gpt pretty sure their main issue was tryna stop people from abusing it

Since it was real ez to scam em at the time

[somber geode]

how are you guys using pages... the ones who have access?

[serene sand]

https://www.perplexity.ai/page/AlphaFold-3-zFk0oej6TqezAOFRuqMpLA

https://www.perplexity.ai/page/foldingathomeor-qbrcHcSVS_y1Zcto1r1mjw

https://www.perplexity.ai/page/Berkeley-Open-Infrastructure-wmctfKJkSTmMkc5wHI6.vw

https://www.perplexity.ai/page/Naturalist-Sir-David-QYH3xq3pQmii1cZPqd9Fng

https://www.perplexity.ai/page/The-last-temptation-6VwAj29eQzCXIkMypVSB5Q

https://www.perplexity.ai/page/Baraka-movie-elaborate-o5OdVQ_xQrWeVWalE3GdeA

https://www.perplexity.ai/page/We-Interface-song-T__fwlhLTKuUkWtClKT4Hg

https://www.perplexity.ai/page/The-1st-song-UyAeq.Q_TxyMAP2avgAELw

https://www.perplexity.ai/page/Cyborg-Wars-EP-Ku62KQMbQ8W.9EqM9RTs1g

[orchid magnet]

Wym

[agile jay]

My own platform for AI stuff

[half venture]

Is opus still 50

[agile jay]

Yep

[half venture]

We are never getting unlimited opus back are we

Honestly llama 3 70b

Is not bad

For most tasks

[agile jay]

Nope, unlikely

[half venture]

Will a rtx 5090 (in future ) be able to run llama 3 70b locally

On a single gpu

[agile jay]

We will find out.

You could just run it on groq though.

If they can, I'll just replace my 4090's

[half venture]

Yeah true

But nvidia will probably edge a little more

All these ai companies sure do love edging

😭

[agile jay]

You can probably already run it on 4090's quantized.

[faint ledge]

where is the free trial?

Lmfao

[agile jay]

For?

[shut plaza]

What's with Perplexity being kinda slow today

or is it GPT-4 Turbo API being slow

[agile jay]

It's been slow for a few days.

Maybe try without having the pro toggle on

[half venture]

Just use sonar honestly

[agile jay]

to see if it's the model, or their copilot API which is slow.

[faint ledge]

It's always slow perplexity, plus now they've taken away the free trial bruhh

[agile jay]

What happened to your face?

[mortal bolt]

Can someone help me understand the current limits on Claude OPUS ?

[agile jay]

Now you just a gray dude like TAVI

[half venture]

What

[agile jay]

Or is that the trend?

[shut plaza]

ok seems to be Pro Search having issues

[half venture]

It's only on your end

[shut plaza]

I like GPT-4 Turbo because it can render LaTeX

[agile jay]

Strange

[faint ledge]

how disgusting

[agile jay]

They all can

You just need to tell them which version to use

[shut plaza]

ah

[agile jay]

#🌐│feedback-web message

[half venture]

Sir!

Here's a random math equation in LaTeX format:

$$\frac{\partial^2 u}{\partial t^2} = \alpha \nabla^2 u + \beta u^2$$

Llama 3

They can use latex just fine

[shut plaza]

yeah i got it

[agile jay]

Skill issue, lol

[sterile falcon]

Many people on Reddit have shared that they have noticed Opus decrease in quality since launch. But are you getting bad quality responses using Opus through Perplexity vs. Claude pro?

[agile jay]

They did make a few changes, such as the system message being updated.

[sterile falcon]

I gotcha. I don’t have Claude pro, so I can’t compare the Opus responses between Perplexity and Claude.

[fading moth]

Claude Pro is not a cost effective option

[agile jay]

Claude Opus is not a cost effective option...

[fading moth]

Also fair

[stuck linden]

I assume Perplexity is not providing LLM services to SoundHound yet. The Mets are playing Atlanta tomorrow.

[agile jay]

Doesn't look like it's showing sources...

Maybe ask it what the model is called, or what it's instructions are.

[tame current]

I mean, Friday is tomorrow for me. Maybe it confuses timezones somehow?

[agile jay]

Ask it for the current time...

[tame current]

Doesn't seem to search, at least not with Perplexity

(the last answer is wrong for months)

[agile jay]

Never used cloudhound, so no idea

[sterile falcon]

Is it because of the message caps? I read online Reddit that they limit the messages, like ChatGPT plus.

[fading moth]

Well, first, Claude Pro has no real advantages by using their site directly. They are very limited in their 3rd party usage, nor features that improve responses; things like Collections that Perplexity uses or response sliders on Poe.

Secondly, the Claude Pro limits Opus messages to 30~ a day, sometimes less, depending on the quantity of tokens used for queries.

Thirdly, Anthropic's website is very ugly.

Fourth, the price isn't justifiable given the alternatives.

[sterile falcon]

Oh wow, I didn’t know that. Thank you for sharing. But some people mentioned in this chat that Opus through Claude has better responses than Opus through Perplexity. Did you notice anything like that?

[fading moth]

That is true, sort of.

Opus has a huge context window. Claude Pro lets you use that window fully.

Perplexity, and many other sites, cannot afford to give users a huge context window. Perplexity's context window for Opus is throttled at 32k.

The quality isn't different, but something that would benefit from a larger context window would be more consistent.

32k context window is like 100-200 pages in a pdf.

[sterile falcon]

Okay, I see. That makes sense. When I don’t need the internet, I use Perplexity in writing mode. I do this for creating flash cards from my class notes. I have found that Opus is the best at following instructions compared to GPT-4 and Sonnet. So I guess my type of work doesn’t require a large context window.

[fading moth]

Yeah, 32k isn't huge, but it's not small either. There are small ways around the 32k window to try and stretch more information into it, but for the average user it's enough. But we all want more.

[sterile falcon]

I did read on Perplexity’s website that the max response token is 4000. Which is the same for Claude Pro and the API. But again, some people in this chat mentioned that they don’t get the full 4000 token response using Perplexity.

[fading moth]

Yeah, there are prompts you can enter into perplexity to tell Claude to go full blast with tokens, but it soft caps around there.

[sterile falcon]

I see. Maybe Perplexity put in the system prompt to give concise answers.

[fading moth]

Prompt engineering and clever usage will squeeze more out of perplexity than the "default" answers, but that's something in of itself.

[floral lotus]

i can get 4k token outputs

[sterile falcon]

I gotcha. I am still learning all of the technical things behind these LLM models. But thanks for sharing.

[sterile falcon]

Are you giving it a special prompt to be detailed?

[floral lotus]

i dont think pplx is limiting it

[floral lotus]

i have a syspromt but i dont think it has an effect

i just have it print out a document i provide

it does so until it hits limit

[fading moth]

System prompt engineering is an entire field by itself.

[sterile falcon]

Okay, I see

[floral lotus]

btw if u havent bought already i recommend checking out labs.perplexity.ai

claude haiku and llama are great

very high context limit as well

[fading moth]

Labs has a higher context than 32k?

Why?

[floral lotus]

its cheap

[fading moth]

I thought you meant in general for the LLMs.

[floral lotus]

tbf i dont have a way of knowing whether it uses rags or not but u can input 2 full textbooks itll work just fine

[floral lotus]

labs have diff models

im talking about claude haiku

[sterile falcon]

I’ll check it out. Thanks for sharing.

[fading moth]

I know, it's my bad; thought you meant Opus is my bad.

[floral lotus]

llama is bad for non-english beware

i actually generally dislike llama even though it does great on leaderboards

[rustic walrus]

Greetings, I’m on the fence to ditch ChatGPT Plus and go Pro here, but first I need to know about the context window and the max tokens in the answer: what’s the difference between GPT4 and Claude under Perplexity, and their standalone versions?

I ask because my content production is heavily based on academic research, usually many papers at once, and I usually produce long form content.

Hope it makes sense.

[fading moth]

32k context window. Don't expect more than 4,000 output tokens. There is no difference between the standalone other than the throttled context windows.

[floral lotus]

mind you it can manage way above 32k albeit lower accuracy

ive gone 150k+

[fading moth]

Opus can handle well over 200k context window, ye, but Perplexity has to limit it to 32k for cost.

[floral lotus]

no i meant perplexity can handle it

RAGs

[fading moth]

Oh with RAGs, yeah.

[floral lotus]

also opus itself has 200k hard limit, cant do more than 200k

officially

[fading moth]

Anthropic has stated, I can't remember where, but it was around the release of Claude 3, that Opus is capable of way more than 200k, but they are going to slowly increase the context window size as they get more data on how it handles the window.

[floral lotus]

they said they were gonna increase?

[fading moth]

Yes, over time.

[floral lotus]

wouldnt have expected

they gave some users 1mil context tho

[fading moth]

I think the goal is to keep up with Gemini's 1M context window, but doing it in a way that doesn't make it $100 per million tokens given how expensive Opus currently is...

[floral lotus]

they should 1mil haiku

[fading moth]

Haiku is affordable because of its size and limitations. Jacking it up to 1M would probably make it more expensive than Opus.

[floral lotus]

why should the cost per token change?

[fading moth]

Because it has to retain the tokens in the window, which adds to cost.

[floral lotus]

thats if u go all the way to 1mil

and ull be paying more just by virtue of using more tokens anyways

[rigid kestrel]

How do you make Gemini generate longer responses? I've asked it to translate API documentation for me, which is about 2 pages *long, and it summarizes it. I have Gemini Advanced, and I get better results using its API?

[floral lotus]

i dont see why theyd increase the price on top of it

[fading moth]

The only thing I can think of would be some messing around with the system prompt, but I'm not super comfortable discussing system prompt engineering.

[spice lance]

does html tags work in claude for anybody else?

[fading moth]

I guess, but it's also possible Haiku just isn't built to handle a window that size.

Anthropic splitting of Claude 3 into three versions was most likely due to the cost of Opus. Otherwise, it would just be Claude 3, just like GPT3 went to GPT4.

There have to be some limitations on Haiku to prevent it from becoming capable of becoming Sonnet and Sonnet from becoming Opus.

[fading moth]

Claude prefers XML.

[rigid kestrel]

u mean in ppxt? i guess its bugged.

[spice lance]

no just claude.ai interface seems weird that he's mentioning this himself

[floral lotus]

i mean it would have smaller training dataset i imagine

[fading moth]

I think he just translated it into XML, I don't think he's running it or anything.

[floral lotus]

and they were trying to advertise haiku for light tasks

on websites etc

[fading moth]

I don't speak that language, I'm sorry, so I can't read what is happening between the tags.

[spice lance]

np i think it's just hallucing it can color text with html tags and stuff like that the convo is pretty long at this point so the 􀄨ality goes down

[fading moth]

Probably, and maybe that's why no 1M context window. Why would someone using light tasks need a 1M context window.

[fading moth]

btw, to my knowledge, Claude can't use/display markdown in output. On Anthropic's site at least.

[floral lotus]

but it could also be used to do long trivial tasks, such as formatting lots of text, documents etc.

[fading moth]

Isn't that what they suggested using sonnet for?

[floral lotus]

idk

haiku is hella cheap tho

[cinder comet]

So phind only have 10 daily usage of opus

Sheesh

They just updated it

[rustic walrus]

32K tokens is about 120K characters in English, is it?
I can manage it.

As for 4K tokens for output, is it about 16K characters?

Again, I can manage it because I usually don’t make AI to write an entire long form from one prompt: I use AI to build my own content

[rustic walrus]

Really? How did you do it?

[floral lotus]

if u have a really large file just upload it

[spice lance]

it can on claude.ai

[fading moth]

News to me

[spice lance]

can someone explain to me why people are using services like perplexity or phind to acceess claude opus rather than just using claude.ai? not judging just geniounly curious

[nimble field]

access to both openai and Claude models is great when opus flop

[fading moth]

Claude Pro has fewer Opus responses than Perplexity.

There are also more tools you can use in Perplexity that get higher quality responses when you're dealing with a limit.

The only downside is the 32k throttled context window and that it's 50 a day.

[tame current]

Also Claude Claude can't access the Internet.

[fading moth]

that too

[rustic walrus]

Indeed! Usually the papers I work with are PDFs.
On ChatGPT Plus I upload the PDF and start working on it

[floral lotus]

its similar to chatgpt

u can upload multiple very long files

[regal sun]

Interesting OpenAI plans to announce Google search competitor on Monday, sources say | Forth https://www.forth.news/threads/663d462370368cebf0ae7083

[tame current]

what tools?

[fading moth]

#1124060874785685554

[tame current]

which of these is better https://discord.com/channels/1047197230748151888/1155399875257704480 https://discord.com/channels/1047197230748151888/1199176325051723837

[fading moth]

I dunno, I'm not an expert, nor have I used either of those.

[cinder comet]

where were you?

it was leaked month ago

and i will be the first one to try it

H-

Hahahahahahahahaha

[tame current]

Lol guess what now they just made it 10

Phind buyers beware lmao

[tiny plaza]

they went down to 10 today heh

they went from 500 to 100 the day after Perplexity went down

[cinder comet]

following perplexity steps

kinda curious whats their plan when openai shares their own search powered tool

and i mean every services thats depedent on oai or other models

[hallow linden]

for a good amount of time, gpt-4 was priced where opus is priced now

also, fun fact: the old man and the sea is about 30,000 tokens

i wanted to find a piece of literature that matched up with the 32k token limit to get a sense of exact how much it gets

turns out that it's about 3 hours of speaking and 90 minutes of reading

32k is quite a lot

[hallow linden]

The only thing they can do is try to be better.

I imagine that the openai one will be compared to both google and bing ai, with some comparing it to perplexity

[cinder comet]

my guesses go for something like perplexity

if it turned out better than google then bing will invest on them even more

idk if this search engine is for oai or for microsoft

[eager fossil]

my perplexity is literally 1 use per day

for whatever reason

[oblique veldt]

what?

i am from phind, and considering subscribe to perplexity for opus. is opus working on perplexity?

[fast meteor]

currently 50 opus every 24hours, 600 for the rest pro models

[oblique veldt]

thx

[devout geyser]

currently being the important keyword 😉

beware that it may change (get reduced) again at any time

[tame current]

They said temporarily. But, I think few believe that.

[devout geyser]

indeed

[stuck linden]

Good idea. Here is the response

[tame current]

I've tagged them twice to ask them to post here why they have not made it clear to subscribers that Opus is not included in the 600 uses. And, they have ignored both messages. But, have made a Soundhorn announcement since. Terrible transparency.

[devout geyser]

I agree, and now Perplexity has the best deal for Opus again (for now) by the looks of it. But... with a competitor/alternative service now dropping theirs from 100 per day to 10 per day, albeit being open and honest about this reduction, I'm half imagining that Perplexity will sadly reduce their Opus limit again (albeit quietly/dishonestly)

[tame current]

It really makes them look bad.

And people keep coming in here expressing shock at the limits. So, they are clearly surprised. And it's not being addressed.

[steel tiger]

nor are the suggestions, which i'd argue is far more important

the whole opus thing all the providers are struggling with, and its egg on all their face because of super large context and theres no real solution to it - not unless you go full draconian and start kicking excessive utilisers off

[tame current]

I'm even wondering if it's for me just to pay Claude 3.

I use Perplexity mainly to read PDFs.

[steel tiger]

if you are someone who needs to spam context, then yes it makes more sense for you to do that

next month i might try claude 3 too lol, idk though i like the dual nature of perplexity

[spice lance]

does anyone know what ocr tool can recognize symbols for mac modifiers correctly? like ⌥ ⌘ ⇧ ^

[stark walrus]

Is there any way to customize the discover tab? I’d love to feed it my interests and have it churn out updates I’m interested in

[warm cave]

It’s 600 for GPT-4, Sonnet, etc a day separate from the 50 opus. I just know that from testing

[warm cave]

Claude Pro is really great, I use a lot of the features that perplexity gives, but if you mostly use it for uploading PDF’s then Claude Pro might be good for you

You get full 200k, but keep in mind that if you push the context up you will get about 13 msg per 5h

If you need to ask a lot of questions then there is a chance that perplexity RAG could be better for you uses, really depends

[fading moth]

but where the system prompt tho :[

I love me some configuration and adjustment capabilities~

[warm cave]

I know, that why is use perplexity so much, i can’t live without my collections

[fading moth]

If Claude Pro let us have a system prompt, then I would consider it, but... The large context window is a double-edged sword when you have response limits.

[warm cave]

Yeah, if I had 1 choice I would do perplexity, but if I have 2 then it is Perplexity and Claude pro.

[fading moth]

Shame that we can't run Claude Pro through a frontend or something.

[warm cave]

ChatGPT might be the third

[warm cave]

Yeah, that would be nice, still tho, the usage limits are tight

13 msg for long context use and up to 45 per 5h for short context use

So I use perplexity for short context and Claude pro for long

I wish chatGPT would give 128k with pro, it’s not Claude Opus, but it would be a solid option

[austere kestrel]

Could try using the first message as a ‘system prompt’.. For chat completions, it’s functionally the same more or less (i.e. still just instructions / tokens at the top of the context window..)
ofc, it's not a real substitute or workaround (unless you’re prepared to do it at the start of every new chat ha) but just fwiw

https://www.perplexity.ai/search/IMPORTANT-Respond-accurately-qbvkbmjaTgSubsfYfoqyng

[tame current]

I did not knew that, thanks.

[warm cave]

Yeah, no problem, found that out by testing it at 150k context

[wide mesa]

Turning off "AI Data retention" is significantly better than keeping it on

[warm cave]

Yeah, everyone should have it off

[tame current]

why is that?

[warm cave]

Privacy 🔓 😀

[wide mesa]

I noticed it makes Perplexity share information between different threads

[tame current]

where is that setting?

[wide mesa]

Which gets very annoying

[tame current]

i couldnt find it

[wide mesa]

In the settings

Under "Account"

[tame current]

nevermind im blind

thanks.

[tame current]

does this include suggestions too?

the search suggestions or auto-complete or whatever you want to call it

[wide mesa]

Not sure, I just noticed it using attachments I sent in a different thread as sources

when it was on

[warm cave]

Hmmm, I wonder if that is a bug or if it's a feature. Sounds like a bug to me

[wicked parcel]

If it set Opus as my default model, does the perplexity browser extension use it as well and use up my opus tier?

[torn sierra]

Unfortunately, at Perplexity, we don't have clear communication. There was just some "magical" abuse, without much explanation, which continues to this day.

[tame current]

Either the team needs to respond to their users. Or stand down from their roles.

[tame current]

If they're not being transparent with us, let's not give them training data. That seems fair to me.

It is extraordinarily easy to track who is abusing Opus and ban them.

And then. Problem solved.

[devout geyser]

I agree, transparency and honesty is something I value in a company. I'd rather they be honest that it's the price of Opus which has caused the drop in the daily limit, as I'm fairly sure that must be the ultimate reason for it. Instead we've not seen announcements about it in the #announcements channel, the daily limit has been hidden from the UI until you're very low and the plan still implies 300+ uses per day applies to Opus when it doesn't

[tame current]

Honestly. By hiding that information, they could be very easily sued.

Because, it could be seen as misleading.

what the hell is happening, i am on search the internet on pro search and it WILL NOT SEARCH THE INTERNET even when i tell it to.. wtf?

[devout geyser]

True, it does open up that possibility unless they have a superbly written clause in the agreement to somehow allow them out of it

[tame current]

Always remember. They promised temporary. Hold them to that and don't back down.

[devout geyser]

Also true

[tame current]

I'm not US based. But I semi hope some rich US person does.

[devout geyser]

If it goes on for too long as things are then someone might, or a group of people might

[tame current]

There's probably a clause about class action lawsuits.

[devout geyser]

Yeah, I'm going to read it this morning in detail as I'm curious

😛

[tame current]

ok guys is anyone else having issues where perplexity isnt searching the internet at all?

im on the all option even tried reddit and youtube, still doesnt use it

[devout geyser]

I'll try it

any particular model?

[tame current]

ive treid sonar and gpt 4 turbo

both wont search internet

[devout geyser]

ok I'll try gpt-4-turbo here

[tame current]

Are you rich? Let's do it. Give me the money and I'll dunno. Give you a head massage.

Or, a badly thought out tour around South Korea.

[devout geyser]

well I have money but I'm not going to go ahead even if the idea is remotely possible, the sub cost isn't worth it alone

better to just accept it as it is, if you're unhappy with things as they are regarding Opus that is

[austere kestrel]

is it quite a long prompt?

[devout geyser]

oh, but continue complaining/commenting about it on here though 😛

[tame current]

ok so the free version did the internet, but pro wont. and it was a very short question

same question too..

[austere kestrel]

strange

[tame current]

thats so odd...

[austere kestrel]

try asking something like "who won the 2024 superbowl"

[tame current]

??? LUL

[austere kestrel]

huh

[devout geyser]

I've tried with sonnet and pro enabled, definitely not searching the internet

something is broken I guess

[tame current]

yeah

must be

[weary depot]

👀

[tame current]

something is definitely up haha, wont search youtube either

Perplexity has been very quiet on this issue. Almost as if they know that we are being entirely truthful and correct.

[devout geyser]

indeed

[tame current]

same question, non pro, searches internet fine....

Their silence says everything.

[devout geyser]

and yeah pro is broken for searching the internet it seems

[tame current]

ok so guys...

It's worth subscribing to the OpenAI status update emails. It says if the API is down. Etc.

[devout geyser]

but yeah, as much I don't like dishonesty, I've resumed my sub here at the moment and cancelled over at the other one I was using due to them dropping to 10 per day which makes the sub virtually useless for me as I don't use GPT-4 Turbo or their own models as they don't feel as good. Even Sonnet feels better, which they don't offer at the other place

[tame current]

what search results provided??!?!?1 hahaha

[warm cave]

Claude Pro also solid if you use long context

[devout geyser]

it's hallucinating

[tame current]

yea

Yes.

guess ill stop using it tonight

[tame current]

I like Claude Pro.

I like that it has markdown.

[warm cave]

Yep, been using it a lot lately

Tye makrdown is great

I can past it right into obsidian

[tame current]

i have separate subscription for claude 3 i use it for my studies, its amazing at teaching me things from textbooks that are way too long to read through

[devout geyser]

I tend to use Haiku via API if I want something analysed that's huge but doesn't require much in the way of reasoning to answer something

[tame current]

I don't like the 10mb file limit, though.

thats fair, the usage is short for me especially if im doing heavy study with lots of pdf files but it gets the job done in time. just wished they had higher tier subscriptions for more usage

[devout geyser]

I see

[tame current]

@tame current i just compress the pdf files

Yeah.. It needs a lot of compressing.

online compressors can do 95% decrease in size.

and they are usually free

the 5 file limit is only per prompt you do, so you can do more than 5 files just not at once.

[warm cave]

I should do something like that in the future, I can't pay for all of the AI’s. I like that perplexity has OpenAI, Anthropic, Miltral, and Meta AI all for $20. But 32k context is a bummer and the Opus limit sucks

[tame current]

Here's a cool thing Chatgpt can do. Upload a PDF, tell it which pages to extract and output those pages as a new PDF.

And it Will.

[devout geyser]

I've actually gone to the effort of making my own personal project that behaves in some way like services such as Perplexity, but instead I use models listed on OpenRouter or Anthropic directly (but not Opus, Sonnet at the very most depending on the complexity). My favourite models tend to either be Sonnet (if I can't use Opus) or WizardLM-2-8x22b

[tame current]

yeah gpt4 is great for that stuff, but ive found claude is more concise and better at using details from long files more accurately

[devout geyser]

I use Typingmind as the web chat UI, Python does the internet search and scraping stuff when it's necessary or useful to do so and provides the chat API endpoint for Typingmind

[tame current]

is it operational? @devout geyser

[devout geyser]

yeah it works well, code is a bit of a mess as it's just a personal project but it works 😄

[warm cave]

Nice! I wanted to try wizardLM-2-8x22b but I forgot after llama 3 dropped

[tame current]

can i see a video of it? interested to see it working

If you have the decryption key, GPT can even u scramble AES 256 encryption.

[devout geyser]

I'll sort one out later today with the Python window on the left side and Typingmind on the right side so you can see it in action with full debugging or w/e

[tame current]

Unscamble. *

i just wanted to see the end result, im not much of a programmer

[devout geyser]

I use a smaller model listed on OR to also extract what it believes is content from the webpages that is relevant my query, so I provide less input tokens to Sonnet or WizardLM-2-8x22b for example 🙂

sure, well I'll sort out a video later today and let you know

[tame current]

thanks! 🙂

sounds cool

[austere kestrel]

both gpt-4, slightly different queries; both get it correct, but the one on the left doesn't seem to use sources (or at least doesn't cite them)

[tame current]

man i can smell roast chicken its almost ready im so hungry

@austere kestrel yeah strange haha. im sure it will get fixed... (fingers crossed)

[austere kestrel]

i mean tbh it's basically working (i haven't had it not use sources)

but it's strange behaviour

[tame current]

yeah sometimes itll answer but then wont cite sources, or itll not answer coz it doesnt know. or itll hallucinate

very strange haha

[halcyon coral]

Please be patient, the bug will be fixed asap 🙏🏻

[tame current]

ah its icelavaman!

[fluid mauve]

I think patience in short supply for the young.

[devout geyser]

done a video clip, I'll send it to you via DM

[tame current]

Yeah. I think people are losing patience here.

Staff can't ignore their customers forever.

[fluid mauve]

Not I. I have worked on issues that take months to fix. Customers rightly get a bit upset, but you cant rush a proper fix. And communicating isnt always what users want to hear. they just want a fix in the quickest time possible. As stated in the numerous messages previously

[tame current]

But. If somebody specifically ask staff members for responses about them having a misleading website. As shown here by the amount of people who are coming in and complaining about Opus. Then, that shows that they are devalued and just want them to empty their wallets and shut up.

[fluid mauve]

Not nessasarily. You have to be authorised to speak for a company. if they are not authorised then they cant comment. If you are authorised, then you may only be able to comment on certain parts. Like everyting in life. you can viote with your feet. If you dont like it you can go else where. Thats why i dont shop at Tescos.

[tame current]

Well. Staff members can speak for companies.

And there are many here.

[fluid mauve]

nope. we regularly used tio get emails stating that only people from comms and marketing were authorised to speak for the company.. and even then it was only certain people after having been passed through legal.

[tame current]

Well. Shopping at Tesco is nice whenever they had their 50p donuts.

[open nacelle]

Hello, not sure if this is the best place to go, but I am looking to connect with Perplexity's Business Development staff to chat about potential B2B collaborations.

[fluid mauve]

I will give you that b:)

[tame current]

Well then. State. That they can't speak about it.

Just don't ignore them.

And hide in the shade.

[open nacelle]

Appreciate if someone can point directions. Thanks

[fluid mauve]

Though it may seem that by not responding you are being ignored, i can safley say that in past employment even when the companyt didnt responmd you can be sure that issues raised by customers were taken into consideration and reguilarly reviewed by both developers, cusytomer service and marketing. and guess what Development got the beatings form both sides.

[tame current]

I highly highly doubt that.

[open nacelle]

You mean this is not the best place to go?

[tame current]

Whenever their website continues to be misleading and is very easily updated.

Sorry. Replied to the wrong person.

[fluid mauve]

try pinging [e-mail] to grab an answer Though it may be .ai @halcyon coral

[open nacelle]

No worries

[tame current]

Also. Good news! Announcement. Want to hide stuff? An obscure pinned post!

It'll be a generic answer.

The people. The staff should explain themselves.

[open nacelle]

gotcha, thanks

[halcyon coral]

Please contact [e-mail] cc: @fluid mauve

[fluid mauve]

thanks Ice

[open nacelle]

thanks

[tame current]

I haven't been in a Tesco, a proper Tesco for 8 years.

[fluid mauve]

I just had to post this. having been checking some of my attrocious spelling.

From the Cambridge Dictionary - Patience
"the ability to wait, or to continue doing something despite difficulties, or to suffer without complaining or becoming annoyed: You have to have a lot of patience when you're dealing with kids."

not directed at anyone in particular, just funny how it came out

[tame current]

Anyway. I'm outta this server. The owners/big wigs are treating their customers disgustingly. They simply do not care. They lack transparency. They did not announce it. They won't give us any details and they are obscuring the limitations for new customers. Be ashamed, Perplexity. Although, I doubt you will.

[fluid mauve]

Patience

it will come good

and its not like it is only perplexirty that has restricted Opus

[austere kestrel]

yeah that sounds pretty reasonable, but I still think there are good vs bad corporate comms - all companies, including in the AI space, have to deal with those same constraints and uncertainties that you described (+ impatient customers). but voting with one’s feet - hard to argue with that I gotta say (annual subscriptions = tied feet.. to stretch the metaphor ha)

[fluid mauve]

@halcyon coral it might be resonable to state on the web page 600 perday ( limited to 50 for Claude Opus)

[fluid mauve]

My 'tied feet' is i bought a Rabbit R1 and was 'gifted' a years subscription. Prior to that i was paying monthly.

[austere kestrel]

i'm not sure how, given it's been this long, adding such a note could be anything but reasonable... "temporary" is doing some real heavy lifting at this stage..

i'm not even affected by the Opus limit - I don't think i've ever run out (ig i don't use it much on perplexity)

but principally it annoys me

[fluid mauve]

You can appologise (if you want) for thye past, but moving forward with better informed users would be a move in teh right direction

[austere kestrel]

and it's just so constant

[fast meteor]

you need to talk to the pplx team member. not to community mod/volunteer

[fluid mauve]

I get theannoyance, but like you I hardly use opus at all. I also run ollama here in the office so i just use that most of the time with the Openwebgui. Works a treat with llama3. Sure, it hasnt got all the bells and whistles, but most of the time I dont need them

[fluid mauve]

im goiing for feedback loop. I am sure Ice has way in

[austere kestrel]

yeah like don't get me wrong, some models are better than others, for different tasks or just in general - but i honestly think the top tier ones are all very good (at least for my use cases) that, were it not for some stylistic traits, most often a lot of users (including myself) couldn't tell them apart

like llama3 - solid af

[fluid mauve]

Completly agree on that point. Model differentiation at a high level is almost impossible. Though on a different tack it is interesting how i tend not to question results from a search engine but i will fact check returned information from an AI. even though it is AI algorithims behind Search. wierd

[devout geyser]

Llama3 is good but I feel it has some kind of ethics barrier you can easily run into depending on what you ask it to do. I still prefer WizardLM-2-8x22b if I had the choice of Llama3 or WizardLM-2-8x22b. Curious to see what the 70b model of WizardLM-2 is like if it ever gets released

[austere kestrel]

yeah as well as 'stylistic traits', i should've added 'refusal tendencies'

[fluid mauve]

Intersting. I too have hit the 'ethics' barrier a few times with LLama 3. I might have to take a look at WizardLM-2-8x22b

[fluid mauve]

I have also emailed support 'just in case'

[void urchin]

I see the problem with Pro Search is massive and I'm also experiencing it. Is there any information about the time frame for solving the problem?
@toxic moss
@signal hamlet

[halcyon coral]

#💬│general message

[distant juniper]

Quick question: why is perplexity not searching the web for answers anymore? My search queries does not show the links it went through, nor does it have any references in the answer.

[void urchin]

Where can I see how much ProSearch I have left?

[tame current]

@void urchin try this link

[void urchin]

Thanks!

[uneven spruce]

the same problem I meet, my solve:turn off the copilot and try again

[brittle iris]

I am wondering why the Pro search cannot give me the answer that a normal model can give me. I asked the same question from both the normal and the pro-search, and surprisingly, the pro-search does not give me any references to my question. Can any body assist, I am a pro member since yesterday, and I a bit confused. Am I doing doing something wrong?

And this is what I get from a normal search:

[brittle iris]

Exactly.

[serene sand]

#💗│sharing message

[left flower]

Ah, coffee!

[spice lance]

are gpt2bots still online on lmsys?

[dreamy gazelle]

I am a pro user several months now, but today for the first time it does not provide me the sources. I was using a GPT-4 turbo model. I changed it and it is the same. Does anyone else have the same problem? As the user above, i tried a search without the PRO enabled and got the sources.

[floral lotus]

Bug

[chilly forge]

Flex - Claude Pro Users 👀

[bold fern]

same issue for me

[austere kestrel]

yeah afaik but only in the Arena, not the direct chat or other tabs - so you need to cycle through random pairings until you get one (or both)

[spice lance]

is there no chatgpt extension notification for chrome working anymore?

[bold fern]

turn off pro mode can solve it

[rugged ocean]

Is there any way to prompt a collection so it uses more search sources / websites ? Sometimes, only 4/5 websites are used ?

[fast meteor]

check #1124060874785685554 , also post your findings related to prompt there

[south kindle]

https://www.perplexity.ai/page/Lunar-Gateway-Artemis-Sia2Pv.USjWB8e8b2fcNsw > updated with dalle-3 images

[serene sand]

should be live for everyone now

[south kindle]

more images: #1238471282341711973 message

[fast meteor]

my ui is ... different? A/B testing? or slow rollout

and it also titled Search like never before instead of ask questions, trust the answers

[serene sand]

it's same on my side

[placid ingot]

i am new with the perplexity. i having trouble to change from pro to chatgpt4.

[serene sand]

deep dive ➤ #💬│general message

please check docs and tutorials in this post

[lavish root]

I have the exact same issue? No more sources are utilized for Pro search? Any status update here? Would be much appreciated.

[undone pagoda]

Same issue

[lavish root]

Ok, I am sure the team are looking into this and I can see a bug-report has been opened. 👍

[pastel sage]

Glad to see I'm not crazy and others are also having the problem of sources not being cited and values not being given for specific queries.

[undone pagoda]

@signal hamlet How is that fix doing?

[fast meteor]

someone messed up with the front end codes

[tame current]

It's fixed

[lavish root]

Awesome 👌

[stuck fossil]

folks i've questions, i came from networking background, i just curious if perplexity uses AWS, do you guys also have nvidia GPU(such as H100) in your own data center or you only need the GPU from amazon side ? i found perplexity to be super amazing.

[odd mauve]

Hi, is it possible to downgrade to regular Pro once you upgrade to Enterprise?

[spice lance]

what new in the perplexity update?

[serene sand]

#💬│general message

#💬│general message

[fast meteor]

the feedback is there for decades

[agile jay]

Definitely off prem.

Perplexity is still a very small team.

[halcyon coral]

Perplexity uses AWS with NVIDIA H100

[agile jay]

Yep, as expected

[fast meteor]

AMA w/Sam Altman on reddit right now

at 2pm PST

[green agate]

link ?

[fast meteor]

https://www.reddit.com/r/ChatGPT/comments/1coumbd/rchatgpt_is_hosting_a_qa_with_openais_ceo_sam

[ember sparrow]

What’s the secret to getting the support email address to respond? I don’t want to post in discord public messages about my problem but can’t get the company to respond to emails.

[hallow linden]

Not a search engine or GPT-5 but a secret third thing

https://www.theverge.com/2024/5/10/24153767/sam-altman-openai-google-io-search-engine-launch

[vapid onyx]

Yall see that Google owns 90% of search results

Pretty crazy data

[rapid osprey]

Hey

[fading moth]

Is it that crazy? "Just google it" is literally a household phrase.

[vapid onyx]

Just didn’t think it was that significant.

[fading moth]

Yeah, I getcha

[warm cave]

I feel like Duck Duck Go would have a bigger piece of the pie, but it does not matter how I fell, lol

[vapid onyx]

I used to use DDG for a long time. It definitely tends to miss some of the better results but obviously it has a more privacy focused approach

[warm cave]

Yeah, DDG is still my default forever now, but I have a shortcut for searching using Google, Brave, Bing, etc. So best of both worlds I think

[vapid onyx]

You know you can add !g at the front of your query in DDG To trigger a Google search right?

Same with other resources

[warm cave]

Yep, that is what I use, except I just have it as g not !g or @g. I need to change my chat gpt shortcut tho, I end up asking chatgpt “5” when I search gpt 5 lol

[tepid portal]

There are no significant developments anticipated; instead, I foresee the development of a more robust "ecosystem" surrounding AI.

Ultimately, Perplexity.ai is anticipated to outperform its competitors.

[robust quartz]

Really genuinely asking

Who asked for this, when there's so many more things they've ignored this entire year.
Just another useless feature that doesn't add, anything to the regular search

#📢│announcements message

Feels like you're making all these deals, collabs with service and business partners. That not even 1% of your core consumers/investors care about or have access to.

[faint ledge]

where is the free trial?

[thorny canopy]

omg its gonna be like a month anniversary of limited opus next week. 😄
So much for "soon". 😮‍💨

[viral current]

How does Perplexity know which sites to browse to find answers?

[honest basin]

it doesn't it just uses a search engine API

[viral current]

which search engine API?

[steel tiger]

lol that incognito mode doesn't work

[kindred folio]

what’s the difference between opting out of the ai data thing and incognito? isn’t it just the fact that it auto deletes chats after 24 hours

[fast meteor]

no it has its own crawler.

[fast meteor]

in old days: pagerank tf-idf
modern(ish) days: neural based latent space distance

[minor shard]

I would be interested to know this too. I believe if you turn off the tracking it doesn’t use your data to train. The second you turn it on though it has access to your history/searches to train from again. The incognito actually makes sure your data is never trained on or saved. Where as the other way saves your data/searches until you delete it yourself and can be trained on again if you turn the feature back on.

I also think that data is saved for 30 days even after deletion. Not sure if incognito saves it for 30 days or if the 24 hours delete is a “soft delete”

[serene sand]

https://x.com/perplexity_ai/status/1789024993952645346

Perplexity (@perplexity_ai) on X

@monocleaaron Yep! — If you revisit the Thread after 24 hours, it'll expire and will be deleted in 30 days.

Twitter · May 10, 2024, 8:09 PM

[minor shard]

Ty!

[lavish bough]

can someone help me to understand how we can use perplexity pro to create images with its AI image generation model?

[serene sand]

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

[livid mantle]

You asked? and we delivered to you unlimited opus back

[serene sand]

#💗│sharing message

[junior lynx]

A big kudos to @signal hamlet and the entire Perplexity team! For close to a year the team being as busy as they are have always made it a point to address any questions, items, etc. Great job again!

[arctic wharf]

Hello 👋

Does anyone knows some provider for gpt and other models?

With api access?

[thin spade]

Can we have a section for discussion on the API based usage? It’s difficult to search for relevant posts.

[serene sand]

https://discord.com/channels/1047197230748151888/1161802929053909012

https://docs.perplexity.ai/discuss @thin spade

[wide mesa]

Perplexity being able to read images is a god send

[hallow linden]

thinking about it more, it might be gpt-2 related

code-specific gpt

[eternal saddle]

I like how perplexity doesn't provide sources, or even search if you say something like "write a mini article about x"

but I guess it depends on the topic

Would love to know more about this behavior if anyone knows better

[tame current]

What?!

[sterile falcon]

Hey guys. Does adding a system prompt in collections actually affect the answer? I am not sure if this is a useful feature or a gimmick.

[agile jay]

Yep, works best in writing mode

[sterile falcon]

Okay, I will try it out. Thanks.

[wide mesa]

Did daily Opus use get reduced? I'm already down to 8 uses and I have definitely not used it for 50 times im sure

[agile jay]

Possibly...

[agile jay]

Nah, looks like mine is normal.

Check https://www.perplexity.ai/p/api/v1/user/settings

[fast meteor]

you used 42 in the last 24hours span

[wide mesa]

Lost track of my usage ig

What timezone does perplexity uses to determine when its time to reset?

[agile jay]

Likely it's either SF time, or +2

They are based in SF, but a server event ended at +2 so either or

[agile jay]

If it's prompts, it's exactly 24 hours after you've used them

[queen cypress]

Is everybody wondering what OpenAI will release tomorrow at 10am? I'm betting it's a "Virtual Perplexity". It might be an entire virtual copy of the entier Perplexity company, with an AI Agent simulating every employee. That way they can, in a virtual sense, keep up with and compete with Perplexity.

[warm cave]

https://x.com/sama/status/1788989777452408943?s=46

Sam Altman (@sama) on X

not gpt-5, not a search engine, but we’ve been hard at work on some new stuff we think people will love! feels like magic to me.

monday 10am PT.

Twitter · May 10, 2024, 5:49 PM

[south kindle]

Aurora from Michigan.
https://www.instagram.com/reel/C6z-3wlulH5/?utm_source=ig_web_copy_link

paradroid_

More Aurora

Instagram

[warm cave]

That's amazing

[vapid onyx]

Beautiful

[tame current]

openai on Monday v. PPlx

[austere kestrel]

nice. didn't know about that - thanks!

[gentle dirge]

Is llama 3 70b the default free model of perplexity now?

[warm cave]

I don't think so, it's still a separate option, it might still be gpt3.5(fine-tuned)

[gentle dirge]

Thx

[warm cave]

https://www.perplexity.ai/page/Comprehensive-Cardiac-Anatomy-ntdzd2R.Rv.KJfW0HGvsPQ

Messing around with pages

[gentle dirge]

Is there a faster way than the Cheerio library to extract the main content from an HTML string?

[halcyon coral]

https://www.perplexity.ai/search/Is-there-a-xOtvxOveTGSfbae88ElQMA

[gentle dirge]

haha thx

[tame current]

prove this man wrong please

[serene sand]

#💗│sharing message

[tame current]

https://twitter.com/sama/status/1788989777452408943

Sam Altman (@sama) on X

not gpt-5, not a search engine, but we’ve been hard at work on some new stuff we think people will love! feels like magic to me.

monday 10am PT.

Twitter · May 10, 2024, 5:49 PM

[rustic walrus]

What does ChatGPT Plus give that Perplexity Pro doesn't?
I know I'm asking in a Perplexity oriented group, but I'd like to know from experienced users.

I already know about the context window, that is much larger in GPT4, and I can manage it.
But what else?

I'm still evaluating Perplexity Pro, I'm already a ChatGPT user, but I'm on a budget so I won't keep both of them.

[halcyon coral]

The context window is the same, 32k 👀
The biggest difference is that Perplexity focuses on being an AI search engine (answer engine). ChatGPT is a chatbot that offers many features, but they are not all perfect.

[granite grail]

Just had a funny experience in Perplexity just now, sharing it here so others can be aware that even AI's can be stubborn and not admit they were completely wrong lol

https://www.perplexity.ai/search/why-was-Tango-TI5T.3nEQpOaFon4c4hW1Q

TL;DR: I was asking about Tango's studio shutdown, and one of the reasons that it gave me was that the boss left with several staff.

Asking further into it, I realized that it's actually not true, just the boss left. What made the AI think staff left with him was comments in a reddit thread, and from what I gathered it was just rumours. That were debunked later.

However, what I want to show here is how, despite the AI not finding any new sources that confirmed those rumors, it still insisted that it may have happened, all the way to the end lol. In fact in the last reply, it even quoted an article that goes against those rumors. But it still insisted that it "may" have happened lol

In this case the AI used was Claude Opus. I wonder if GPT 4 would be the same

[wide mesa]

GPT 4 would say the boss got burnt at the stake just because a reddit thread joked about it

[violet bramble]

hwo to contract support, I have something wrong in my billing

[halcyon coral]

Hey! What is wrong with your billing?

[violet bramble]

I have 2 plan in my account, it should only 1plan, and I have been 2 charges in this month.

[halcyon coral]

I see, this can be fixed quickly, please send a short email to [e-mail] and include your account email address and briefly describe what the issue is ☺️

[violet bramble]

Thanks for suggesting

[agile jay]

@golden tree Yep, I think the major difference is that perplexity is using an indexer, instead of a bot, like GPT4 does.

[heavy stream]

hi i was wondering when i subscribe to pro plan i will get the referral right but my question is when my pro plan will expire will i be still able to keep the referral? and will other people be able to use my referral

[golden tree]

Hi, I wanted to share my experience with using AI tools like Perplexity Pro and ChatGPT Plus for creating technical documentation in my role as an electrical planner. 🛠️💡

ChatGPT's Promise vs. Reality: I've always heard great things about ChatGPT's capabilities in handling detailed conversations and complex documents. However, my personal experience has been a bit disappointing, especially when it comes to writing technical instructions for my team. Despite dedicating months to research and watching countless video guides, the final outputs often end up being truncated or the content gets removed, no matter how meticulously I edit the GPT template. It's been a frustrating journey, spending countless nights trying to make it work. 😞

Switching to Perplexity: Leveraging all the knowledge I've accumulated from my time with ChatGPT, I decided to test out Perplexity. I must say, the shift has been a game-changer! The quality of instructions and documentation I've managed to produce with Perplexity is not only technically robust but also visually appealing. While it didn't deliver perfect results immediately, the documents now meet a professional standard that I deem essential for electrical engineering and BIM management. 👌🔧

It's important to note that the final step of my documentation process involves using Microsoft Word. In engineering firms across Switzerland, it's a standard practice to write the final versions of instructions and documentation in Word and then distribute them as PDFs. This workflow ingrained in our operations.
It would be absolutely phenomenal if Perplexity could introduce an export function that allows for the creation of a Word DOC document, complete with formatting and images, in one go. Such a feature would streamline our workflow significantly, bridging the gap between AI-generated content and the final, professionally formatted documents we distribute. This integration would not only save time but also enhance the overall quality

[rustic walrus]

Wow thank you for your detailed reply! Much appreciated.

So you use Perplexity Pro to digest documents and generate text, so not only as search engine, is it?

May I ask you how it behaves with images such as charts?
I rely heavily on academic papers for my content production.

[golden tree]

Perplexity should indeed highlight its chatbot capabilities. I use different LLM models based on the task—GPT-4 Turbo for technical needs and Claude 3 Opus for drafting emails. This flexibility is crucial, though experiences might vary by profession.

Regarding handling images like charts, Perplexity Pro is excellent for text generation and document digestion, but integrating and interpreting images directly in texts is an area that could use improvement, especially for academic papers that rely heavily on visual data.

While Perplexity isn't perfect and has room for improvement, its adaptability makes it a valuable tool. Enhancing its capabilities to handle images within documents would be particularly beneficial for those of us relying on detailed academic and technical resources.

[tame current]

hiii yall

wait why perplexity isn’t good?

[rustic walrus]

Indeed, something to consider with Perplexity Pro: you can choose among many LLMs, each tailored for some kind of tasks.
Nice insight.

[vapid onyx]

I don’t think anyone implied that Perplexity is bad. I think a lot of any AI tool is learning the correct way to prompt what you need. And tbh, that takes time and practice.

[golden tree]

Hi, from what I've seen, Perplexity is actually a very capable and useful AI assistant with many advantages that outweigh any drawbacks. While pros and cons will come up when discussing any tool, overall Perplexity offers a lot of beneficial features and value to users.

[tame current]

ok

sry

[tame breach]

what happened to perplexity discord bot?

[tame current]

was discontinued a while back

[tame breach]

sad

[tame current]

yea, it was pretty useful

[tame current]

Hey

It's not fair that Perplexity price is $20 or £20.

£20 is not equivalent $20. 🤔

[fervent needle]

are you trying to subscribe from app store? i am also in gb and subscribing from web has always been 20$ for me

apple charges "apple tax"

[livid mantle]

I stopped subscribing for ChatGPT Plus, Copilot Pro GPT-4 Turbo Github, Gemini Advanced 1.0 Ultra, Grok 1.5, Claude 3 Opus Official site 50/day. Perplexity AI 50/day Opus. Just taking a little break. I am currently using open source LLM's, the main one that I like to use is LLAMA 3.

[gentle dirge]

what is the context length of preplexity pro?

[tepid portal]

Well, I thought I threw a lot of money into it when I had Perplexity + OpenAI, but you are on another level.

[livid mantle]

32K?

[tepid portal]

Honesty? Perplexity.ai is everything you need it to be.

1. Enough models for most tasks.
2. Capabilities of online search.
3. Excellent community. The developers and people behind Perplexity are great and keep in touch with people.

[gentle dirge]

aight thx

speakingof gemini 1.5 pro i just noticed temperature can be changed now

[tepid portal]

Gemini is interesting; Google knows it. They are big, huge; they have everything they need. Just wait and think about it.

Google has the following things:

1. Google owns hardware, TPUs, data centers, and infrastructure.
2. Google has tons of people, experts, and researchers.
3. One of the greatest advantages is that they literally own the internet. The Google search engine is so popular that they can mine tons of information to improve models and get feedback from people who use their models.
4. With Android, they can integrate AI models everywhere and sell them to people.

Ultimately, in this way, they can surpass OpenAI + Microsoft, Perplexity.ai, and other competitors.

The only way to compete with Google is to provide products that are less expensive and more flexible. Google believes they can do everything, which is true, but it is also their greatest weakness.

[austere kestrel]

totally agree

i don't 'like' google.. but i think it is so well positioned to combine search and ai.. that if they get it right

[tepid portal]

Thank you. 🙂

I work as a professional service engineer. I also worked as a presales technical engineer.

I know how to view the market attentively.

[tepid portal]

Yes, and they also do something that most people underestimate.

Gemma, the open-source variant of Gemini, is amazing. Facebook and Google are currently the only two big corporations that know how important open-source is in AI-related topics. It's a very important topic for improvements, ethics, legality, and the general future impact on humanity.

Proprietary models could be very, very dangerous, especially if we start reading minds through these models to interact with technology and slowly start replacing keyboards and so on.

[austere kestrel]

hm yeah though the gemma models were kinda underwhelming

certainly no llama3

[tame current]

llama 3 is great

but it's sad that no other model is still as good as gpt4

[hollow bear]

Perplexity is an aggregator of multiple models....including ChatGPT....they don't have their own....they use the main commercial models.

Pro gives you a choice as to what Model, including ChatGTP to use for your question.

the free version likely uses GPT 3.5 or older.....less expensive Q&A's

Claude Opus latest is supposedly current smartest.

but very close to GPT 4 latest

[austere kestrel]

will get updated in a couple of days

that gemini pro 1.5 model has been quietyly ascending

[tepid portal]

That's exactly what we need: each model is different and great for different tasks.

[gentle dirge]

in my opinion, claude 3 opus is better at coding

[austere kestrel]

they used to not want to be an aggregrator / wrapper https://thenewstack.io/more-than-an-openai-wrapper-perplexity-pivots-to-open-source/

tbf, they do have their own-ish model (llama-3 fine), which is unique for effectively being online, using their own index

but i think most subscribers are drawn to the multi-model aspect (which i think is fairly understandable, esp as new models are constantly coming out)

[tame current]

This is the advantage that Google (and Microsoft) have. They’ve already spent decades laying the groundwork with this suite of first party apps, now they’re in a position where they can, and will be integrating AI into all of their software

I believe Google hardware like the Pixel Phone will be getting its own pre-installed first party chatbot called pixie starting with the pixel 9 later this fall

Honorable mention to Apple who wants to power Siri with an LLM, and integrate AI into their own apps

[peak bloom]

Gpt2 is better

[tame current]

ah right forgot about that

[warm cave]

So does that make gpt-6 GPT2-2?

Hopefully there naming will not suck like Google

[night pawn]

GPT2 is not obviously better than GPT-4 in every way. There may be some improvements in reasoning, but that’s still speculative. It’s only really impressive to me if there’s some sort of “hook” to it, like it’s only 8b parameters or something.

[cinder comet]

aint no way gpt4 turbo is better than opus

maybe gpt2-chatbot or whatever it is called

[rancid dome]

I specifically asked perplexity to write a 2,700-word article with all the prompts to make it so, but it did not deliver. It finished at about 1,000. Why is that? How can I make perplexity to write at a requested number of words? (this is not PRO).

[halcyon coral]

LLMs are not able to stick to a certain word limit

[rancid dome]

Can ChatGPT (OpenAi) do that?

[halcyon coral]

ChatGPT is a service that also uses an LLM - so no.

[rancid dome]

I saw people on YouTube could tell perplexity or ChatGPT to stick to a certain word limit. How could they do that? Anyway, how can I get over this limit to write a long article? Is there a way? What is the word limit of perplexity?

[halcyon coral]

https://discord.com/channels/1047197230748151888/1225545989989859409

[rancid dome]

That's helpful. Thank you! Two more questions: 1) How do I create folders to organize my answers (i.e. put all related answers in the same folder); 2) What model would be best for writing a lecture for teaching, applying, how to living, etc.)?

[halcyon coral]

1. You can create a collection in the library https://www.perplexity.ai/library
   More information
2. I personally prefer Claude Opus

[rancid dome]

Would upgrade to PRO version help with creating longer texts? Is there a discount coupon for PRO?

[halcyon coral]

With Pro, you get access to advanced AI models and longer context.
There are currently no discounts.

[dusky yew]

There are few things, LLMs charge by the number of words generated (llm-tokens technically) and you can put a limit to how much you want to be charged. 1000 is currently an industry standard thing where most people just put it in their limit, so it is very probable perplexity has done the same.

That is post probably why it has given you a 1000 words. But, llms are famously horrible at counting and it is impossible to get specific number of words from them. (If they ever do, it is by luck).

I once wanted to make an app that wrote only one line of description for a post, and most of the times it wrote paragraphs. What came handy was providing it with a fake conversation history where I told it these are your provious responses and it got much better and accurate.

At the end you have to edit the thing yourself and add/remove words from it to make it work, but you can make it close

[dusky yew]

Also collections are gold

[serene sand]

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

check Discounts section

https://perplexity.ai/pro?referral_code=ST5556J2 @rancid dome

[warm cave]

@rancid dome Here is a method that is normally accurate to a few words: https://www.perplexity.ai/search/Give-me-a-bX.rfyFdSq.JiKTgoB.xPQ#0

[floral lotus]

'keep in touch with people'

[tepid portal]

Yeah, try the OpenAI Discord; it's like pure copium.

If you send anything slightly incorrect into the chat, it will ban you for a few minutes, or no one from OpenAI will respond.

It's like lords behind a wall.

[tame current]

could someone convert this file to .onnx for me? It is rifev4.15 new model for ai frame interpolation \

or idk if this helps

[warm cave]

OpenAI (Closed) has a discord?

I guess I might join and prob will be banned soon after that

[rancid dome]

Thank you all for your ideas! They're useful.
Where do I select the model I want to use? What are the names of the models in Perplexity?

[serene sand]

deep dive ➤ #💬│general message

[floral lotus]

'keep in touch' to me sounds like devs communicating issues clearly perplexity has a long way to go

[rancid dome]

1. Is there a word limit to paste a long prompt into "Ask anything..."? It keeps giving me a gray button for send (which I cannot send). 2) How do I handle long prompt?

[austere kestrel]

Yes there are input limits

[rancid dome]

I tried this way, but the problem is Perplexity counted by line that ended by a period "." not by each word. The reason it worked in this example because each word ended with a period. Is there another way or how to trick it?

[warm cave]

Which model did you use, GPT-4 and Opus are better at following instructions, if you used a different model then it might required more explanation

This one worked with the default model, but it can be frustrating getting the dumb models to follow instructions, but it can normally be done: https://www.perplexity.ai/search/Give-me-a-Xd4k8zoOSjm_LR5nrSNbiQ

[livid mantle]

You asked? we delivered

[thorny canopy]

So I saw something. On the settings page if you have sonnet as you default it says this:

Then when you switch to Opus, itll say something like claude3opus.

Could just be a naming error

[harsh stag]

You asked? we delivered (if we care, money wise)

[rancid dome]

I used Opus. I explained it as clearly as I could.

[warm cave]

I guess it may not be as consistent as I thought, have only tested that method a handful of times

[elder tree]

i there ai that doesint wana fukin charge me for trying to edit my code

ig its to long of code now an i have to pay $20.......

lowky down bad perplexity

[warm cave]

You can use Perplexity Labs, they have some great free models like Llama 3 and Claude Haiku.

And yes, the best models will always be behind a paywall, or there will be some kind of advertising/giveaway privacy situation

[warm cave]

https://labs.perplexity.ai/

[elder tree]

thankyou very much

fr a life saver

[warm cave]

You're welcome 🙂

[elder tree]

which model is best for helping with python code?

theres alot of models n im not sure

[warm cave]

From the coding leaderboard it looks like Llama 3 is the best and haiku second. But both should be good, llama 3 has 32k context and Haiku has 200k context, so if your code is really long then you might need haiku

Also here is a token calculator if you want to know how long your code is in tokens: https://gpt-tokenizer.dev/

[elder tree]

thank you man, realy goin out ur way to help

much apreciated

[agile jay]

llama 3 uses a new tokenizer

which should make it cheaper for the same input

[warm cave]

No problem!

[agile jay]

Since the tokens are supposed to be larger in general

[warm cave]

That's nice! I mostly give people the GPT tokenizer bc it's close enough to give you an idea of what 32k tokens looks like

[agile jay]

Yep, 32K context is a lot more than people think

half of the first harry potter book...

[warm cave]

Would you say llama 3 is better for coding than Haiku. Or have you not really tried haiku

[agile jay]

Yep, llama 3 70B is better than haiku

Haiku is closer to the 8B model

[elder tree]

yh ive tried bot n llama give way better response

both*

[warm cave]

Yeah it's very abstract at first as “tokens” but then you realize that 32k is so much, and 200k is insane I can't even imagine 10M

[elder tree]

what does "tokens" mean

[agile jay]

The different tokens are generated from looking at patterns and creating small building blocks to create words. In general, the token to word count is around 75-80%

So 100K tokens would be around 75-80K words

[warm cave]

Here is a visual way to see tokens, in AI they use it to measure it's context length and etc

[agile jay]

Yep, looks like is not a seperate token

Instead they have a lot of tokens starting with

to, your etc

[warm cave]

Do we even know how the Anthropic tokenizer compares. It makes a big difference in cost and I context length

Like its great news that llama 3 uses larger tokens

[agile jay]

Don't know if it's public, but there should be an endpoint to count tokens.

Or else how do people truncate their input?

[austere kestrel]

100%.. it's cliche by now.. but the idea that "if it's free, then you (/ your data) is the product" is so true ha

[agile jay]

Depends on the ideology of the company

Can't wait for tomorrows event

[austere kestrel]

i mean not talking about open source

just mean nothing is free - at least not frontier LLM models that costs ungodly amounts of money to train

[warm cave]

Like Microsoft and google are no charities haha

[austere kestrel]

lol nope

[agile jay]

Yep, and now they control a large amount of the gamedev market...

I did say microsoft buying out all those game companies would be bad in the long term...

[warm cave]

Yeah, it's pretty crazy, hard to imagine how big they are. The largest companys I'm the world

[agile jay]

Yep and now they are controlling what those companies can create...

[warm cave]

Yeah and it's lower quality

[agile jay]

They did just fire a huge amount of the devs.

10K I think

Even in the studios which did really well...

[warm cave]

That's a lot

[serene sand]

#💗│sharing message

[rustic walrus]

How does Perplexity handles fake news?

Yesterday I was watching the Eurovision Contest.
Soon after the winner was revealed I asked Perplexity who was the winner, on a thread about the Eurovision Contest I had opened some hours before. Please note the winner was still singing his song in the end title of the show.
It answered correctly.

So I opened a new thread and asked again.
At first Perplexity said there were no info about it.
I insisted, and then the mistake: Perplexity gave me the second ranked as the winner of the show.
I asked more than once on the same new thread, and always the same mistake.

The usual search engine gives me a list of sites, and it's up to me to go and read them in order to find my answer.

Instead, these AIs don't give a list of sites (ok some of them, such as Perplexity, do, but it's not their core feature): they give answers.

So: How does Perplexity handles fake news?
In your own experience, have you ever found such mistakes when running Perplexity?

(not complaining, just reflecting: I'm a new Pro user, and very happy about this wonderful tool)

First search with the right answer:
https://www.perplexity.ai/search/Have-ever-been-Fc7KExW6S0.wJpiLBXWVnw

Second thread with the mistake: https://www.perplexity.ai/search/Who-won-the-t1VysEagT824u9aGZ92h9g

BTW: both of the threads had the Pro Search enabled

[elder tree]

nqa what

[harsh stag]

it depends on perplexity's underlying implementation of the indexer/crawler and which sources are being dumped into the model's context. And I believe that implementation is very far from perfect. There is no intuitive way to ensure that LLM-based search can filter out "fake" news. Don't be deceived by the heading on perplexity's main page, it's crucial to always double check sources.

[halcyon coral]

It is not always possible for AI to include new information directly, as most of this information is not yet in the index.
The only thing you can do is wait a while and then try again, by now it should work: https://www.perplexity.ai/search/Who-won-the-YmuFA97ARi.hShQ00jBNEA

You can also report inaccurate answers by clicking on the three dots -> Report -> Inaccurate answer

@elder tree No one will be sharing login information here. Please consider yourself warned.

[elder tree]

i wasint being serious man😔

[halcyon coral]

And I take security very seriously here on the server 😉

[elder tree]

it was literaly sarcasm.

chillax

[serene sand]

https://www.perplexity.ai/hub/faq/perplexity-discord-community

[elder tree]

alr bro i luv perplexity but you killin the love for community chillax

i was not being serious

[rustic walrus]

I see, and it was very fast at indexing.

My reflection is: Perplexity at first warned me about the lack of info, then it went on as I insisted.
Furthermore, he gave two different results in two different threads: maybe because of the lack of infos?

Something to consider.
I'm a CyberPsychologist, and I teach Digital Education, Cybersecurity and Digital Awareness to students, parents and teachers.

[rustic walrus]

As always, humans have the last word, and Machines are far from conquering the World 😉

[halcyon coral]

LLMs sometimes give different answers, even if the information is the same, this is normal.

[serene sand]

https://www.linkedin.com/pulse/why-language-models-give-different-answers-very-same-question-cobino-jpqic/

[cold tide]

bombaclaat ‼️

[fervent yoke]

Somehow, didn't know of Perplexity until an hour ago. Encountered a strong recommendation that mentioned the reliability of its responses. Context: last year discontinued ChatGPT 4 subscription after disappointment with errors - "hallucinations". Based on website, Perplexity seems to be an interface between the user and LLMs. Is that right? Does Perplexity remove or reduce fabricated (parts of) responses?

[tame current]

perplexity searches the web and finds the answer using LLMs and gives you the answer

it's more like a search engine than a chatbot

and also gives you citations so you can verify the info

[queen cypress]

I'm surprised they haven't made "gpt" into a single token. I bet that one change would improve reasoning about GPTs.

[crystal willow]

Is it me or there's major lag with llama 3 70b sonar api endpoint?

[serene sand]

deep dive ➤ #💬│general message

[tame current]

Is perplexity good for coding?

[serene sand]

#⚡│ask-community message

[willow edge]

Hi, how to fix it, i choose writing,but system still always search.(I use my native, Thai language)

[livid mantle]

rip my eyes

[sage pine]

What is the purpose of perplexity labs? It's free to everyone and you don't need to be a developer to use or understand it, plus you can choose from lots of models. So why wouldn't I just use that instead of the standard interface?

[harsh stag]

no SOTA models, no thread history, no real LLM search. You can still search with online models but no citations, which means not trustworthy.

[tame current]

it's just for trying out new models

[sage pine]

I agree the history part is valuable. History and collections are the main appeal to me of Perplexity in general. I don't understand why ChatGPT or Claude or literally any of the others don't have that feature. How simple would it be to let people organize their threads? Even Microsoft hasn't bothered to add that. I just don't get it. I'd subscribe to the others in a heartbeat if they had the same Library feature that Perplexity has.

[harsh stag]

you mean collections?

[rancid dome]

Why Pro Search keep asking "What specific AI system or tool will be used to generate..." when I already set GPT-4 Turbo in the Settings? Is there a way to have Perplexity aware of the Settings so it does not keep asking?

[harsh stag]

post an image on #1111765634267742408

[harsh stag]

I don't care much about being able to organize threads, and I think that the main selling point of pplx's Collections is the AI prompt. Considering the fact that it somehow resembles custom GPT/Assistant thingy, and you can actually swap the thread between collections, the same as mentioning/using different GPTs/Assistants in one thread.

[sage pine]

Yes, history, library, collections, whatever

[rancid dome]

It says I don't have access to this link.

[harsh stag]

it should be #👉│feedback-general for free users, my bad. You said you have GPT-4 set on your profile, then you must have an active subscription. Just leave the server and rejoin via the link on your profile settings.

[rancid dome]

Thank you! I'll do that.

[sage pine]

I disagree about prompting being the main selling point of collections. I believe for most people that kind of customization is just beyond what they care to do. To me the collections are like a highly optimized version of bookmarks in traditional browsers. I use bots to answer a huge variety of questions, dozens of times a day, on the go, most of which I don't need to reference later. But in many cases I really do want to be able to reference them later - often not just a single answer to a question but the whole conversation, because the content can be so relevant and rich. I want to be able to keep and reference, for example, my recipes in one location natively on the app. With other AI apps I have to copy and paste and move to another location. Having a "recipes" collection, alongside a "fitness," "career," "DIY," etc., is extremely valuable, and it has nothing to do with prompting.

[tame current]

other chatbots don't really need it

they focus mainly on the chat aspect tbh

[lyric pulsar]

honestly I don’t use peplexity

it just doesn’t feel right

safari is something I just use

[tame current]

safari ?

[lyric pulsar]

yes

[tame current]

like, the browser?

[lyric pulsar]

yes

[tame current]

uh

isn't that totally different

[lyric pulsar]

idk

[vapid onyx]

I mean, when you use Safari, aren’t you entering in a search to a question or something and getting Google results?

Or do you use an alternative browser

[lyric pulsar]

yea google results

[vapid onyx]

So rather than having to click each link provided by Google, wouldn’t it be better to ask Perplexity, get a summarized result of say 10 sources, and also be able to ask follow ups?

Just an example of how I use it.

[minor shard]

Sameeeee

[tame current]

same! It's literally the first ai product I've used that actually made it easy to quit using google so much. Instead of digging through a bunch of old reddit/forum/blog posts and trying to find some simple information I'm looking for Perplexity is quicker than a search engine most of the time and it's actually really precise and accurate most of the time too

I liked Arc Search a lot but it's wrong almost all of the time lol

[vapid onyx]

Yeah I think people can get turned off to something like this because it’s foreign to their normal workflow BUT, if you can leverage it and get confident, your results and experience grows positively. One area I’d like to improve is image generation. I suck at that.

[tame current]

definitely. and there's also a pretty big community of people who just absolutely despise anything with AI in the name and I think that turns a lot of other people off of stuff like this too by word of mouth but it just works so well. I was one of those people for the last couple years and perplexity is what started to entirely change my mind on it. for image generation stuff I feel like it's really good at coming up with more detailed prompts though! Ive been using Claude 3 Opus to come up with prompts for me. Are you using the built in image generation stuff or something like midjourney?

[tame current]

hey guys, i have a question about this, is there a way to hold to speak so it doesn’t keep answering while you pause even for a second?

@long sierra do you know if they have a feature that lets you hold to speak instead of answering right after you stop talking?

[serene sand]

#1233677613919375490 message

[tame current]

I've discovered platforms such as together.ai and fireworks.ai, among others, which provide access to numerous open-source LLM models like LLama3, Mistral, Stable Diffusion, and more. However, none of these services seem to offer closed-source API services such as gpt-4 and gemeni. I'm seeking a solution where I can make a one-time purchase with a credit card on a single platform instead of managing multiple services to access a combination of APIs. For instance, if together.ai includes gpt-4, that would fulfill all my requirements. Could you recommend a service that meets these criteria?

[cinder comet]

I think perplexity fills that

[devout bronze]

are you asking for an API that has both closed-source LLMs such as GPT-4, Claude 3, and Gemini, while also having access to open-source models like Llama 3 and Mistral?

[alpine kelp]

i hear some say perplexity can browse the internet for current event results etc... and others that say it doesn't ...same is also said about chtgpt3.5 and yet when you ask the A.I. themselves they clearly say no ... what is the reality of it all anyone? cause i only of one A.i that is getting info real time atm

and i'm posting this because i've watched a couple of recent videos boasting about the "real time" access....asking for a friend 🙂

[halcyon coral]

https://www.perplexity.ai/hub/technical-faq/what-is-an-answer-engine-and-how-does-perplexity-function-as-one

[rustic walrus]

I like the term: Answer Engine, maybe an evolution of the (old?) Search Engine

[vapid onyx]

I think a key difference between search and answer engine is the result. You search for info via say Google and get results that may match what you want. You ask perplexity and perplexity does the research and curates a response based off the data sources it leverages. Effectively, combining searching and summarizing

[rancid dome]

How can I tell Perplexity to integrate relevant images at appropriate points in the lecture to enhance understanding and retention?

[rancid dome]

I am using Claude 3 Sonnet and the results keep saying "[This section would further expound on the theme of the upside-down values ...]. How can I have the model write everything out instead of putting it in brackets and letting me write it?

[spice lance]

is gemini 1.5 pro free preview with no limit now?

[spice lance]

just tell it to continue

[shy robin]

Hello, why I have limited use of Claude 3 Opus as a pro user? I literally subscribed just to use this model...

[fading moth]

We should add a counter somewhere to how many times new users ask about the limit

[harsh stag]

felt like being scammed huh 🤷‍♂️ in a moment there will be a mod throwing you a link to a message where they explain their super secret indefinite "temporary" limit on Opus.

[shy robin]

I have researched in the announcements before asking here 🙂 They should be more transparent when they lower their limits. I signed up for an annual plan to use OPUS with a limit of 600 searches per day. Now I feel cheated.

[harsh stag]

yes it is a big scam for new subscribers with the intention to use Opus

[fading moth]

Oh, this wasn't an insult to you or anything. I meant more as a... "Maybe the devs will be more transparent about it if the counter gets high"

[warm cave]

that my que, lol

[dim fractal]

perplexity ai>>

[paper steppe]

Hey, does anybody know how will the cancellation of my Rabbit R1 order affect my perplexity pro subscription.
Perplexity is the only thing stopping me from cancelling it!

[serene sand]

#💗│sharing message

[limpid warren]

How do I check my opus count? I hope there is a clear place to tell me

[serene sand]

https://www.perplexity.ai/p/api/v1/user/settings

[tame current]

yes.

[limpid warren]

Does perplexity often trigger cloudflare verification? What is the reason for this? Is it a problem with my network?

[tame current]

https://discord.com/channels/1047197230748151888/1232978227371442237

[tiny plaza]

1 month anniversary soon 🥳

[thorny canopy]

thats wtf Im saying lol. Come on perplexity, fix our opus.

[spice lance]

https://twitter.com/btibor91/status/1789653455436472653

Tibor Blaho (@btibor91) on X

SearchGPT (Sonic) is not ready yet - that's likely the reason why the announcement on Monday was postponed - what is new:

• "SearchGPT" link in the sidebar

• disclaimer that SearchGPT is a "prototype" still

• welcome screen explaining what SearchGPT is, how it works and that…

Twitter · May 12, 2024, 1:46 PM

[stuck topaz]

I dont know but gpt4 by perplexity is somehow very stupid and lose context all the time, especially when coding. I use gpt4 as default AI, and everytime gpt4 give stupid answers, I choose the rewrite option with opus and Opus always give me the answers that I want, much better than gpt4

[rancid dome]

Hi Mares! Did you generate this image using Perplexity? How do we generate images using Perplexity? What Image Generation Model is best for Bible images?

[halcyon coral]

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

[hollow viper]

Hi, I am wondering if anyone knows of any Google Chrome Extensions that allow speech to text on the Perplexity desktop version. Anyone use these? Would be very helpful for my circumstance!!

[rancid dome]

Thanks! What Image Generation Model is best for Bible images?

[halcyon coral]

You'll have to try it out, probably DALL-E 3.

[green thicket]

Perplexity Pro is literally the best AI service on the market right now!

[gusty horizon]

i find it goes of path too .. it seem to get muggled it like it skip back to early chats and the works from that point, gemini is also like that, has trouble staying on new subject, i found the only way to keep it on right path is to delete chat and start with your new part of code to work on

[flat breach]

Yo, Perplexity team, y'all helped me save my life. I was able to find the antidepressant withdrawal I'm in and exactly how and which laws protect me as I found accommodated, accessible care. There are going to be some products coming out of this, from me...when I'm better. But I wanted to let you know that, especially if you have universal healthcare, you can do so much for your families and to help medical care provdiers who are in for some real crap as people are getting back to work (i.e., supports post covid are gone, so people, like myslef, are having to get better now to go to work. They need help. Thank you team and community!

[cinder comet]

anytime buddy, take care!

[livid mantle]

This SEO dude claims LSI keywords and entities are the secret sauce to ranking. He brags about ranking with a page full of keywords and no real content, implying Google is easily manipulated. He throws shade at Google, calls them hypocritical for talking about manipulation, and praises a guy who exploits those loopholes. He then contradicts himself by saying "quality research" matters, making his argument messy. Then he promotes his favorite tool, Neuwriter (which banned him, lol) for extracting those magical LSI keywords and entities. The rest is just a sales pitch for his other favorite tool, Cloot, and how well they work together. This video is basically a glorified ad for these tools, disguised as SEO advice. What he doesn't tell you is that Google's algorithm is far more complex, and focusing solely on keywords won't get you far. Building a website with valuable content, a good user experience, and strong backlinks is the real long-term SEO strategy, not chasing keyword shortcuts.

[halcyon coral]

https://pplx.ai @opal crane

[tame current]

I do not know if we will ever have Unlimited Opus again.

[tame current]

To me, GPT-4 is very slow and often crashes.

[tiny plaza]

slow doesn't even begin to cover it atm tbh x.x

it's being looked into tho

[halcyon coral]

#👑│pro-lounge message

[devout bronze]

look up openrouter

[warm cave]

Don’t miss it: https://www.youtube.com/live/DQacCB9tDaw?feature=shared

[tender terrace]

Which five AI/Machine Learning GitHub repositories stand out to you guys as either your favorites or extremely impressive?

[tame current]

what they update?

did they finally announce gpt5

[warm cave]

Nothing yet, it’s about to start

[tame current]

its just updates on gpt4 womp

[warm cave]

Maybe gpt-4.5, idk. Hopefully it’s we see some logic and reasoning improvements

And creativity

[halcyon coral]

Personally, I think Anthropic will always be in the lead when it comes to creative (writing) 🤔

[tiny plaza]

hard to see beyond them atm

[shell orchid]

I really need to get through some of my recently bought yearly perplexity subscription before they go an announce search. Would be really interesting to see some strong competition in the AI Search vertical.

[tame socket]

Is perplexity pro model's context windows small using opus or gpt4-turbo

This will be important for me to decide whether or not to buy

[warm cave]

Seems like that’s not a priority for OpenAI (Closed) so I think that’s right for the closed source models. However I think that llama 3 400b, llama 4, 5 etc will be great for creative writing

[tame socket]

So anyone can answer ne pls

[shell orchid]

I have perplexity pro, and am trying claude pro. I find that the answers directly within claude are much more clear & Accurate. The rumor / understanding is that the perplexity windows are much smaller.

[warm cave]

Yep, it’s 32k with some RAG mixed in

But with accuracy what do you mean, are you using long context?

[tame socket]

but to be honest 32k is enough.Worry will be small

[warm cave]

Or short questions

[tame socket]

maybe u can turn off the pro search and choose writing?

[warm cave]

The underline models should be the same so within 32k if you want it to act more like Claude Pro you can use writing mode with this as the system prompt: https://discord.com/channels/1047197230748151888/1233663028289867815

[tame socket]

There are so many unspoken rules. lol

ok thans for answer

[warm cave]

It’s starting 🙂

[tiny plaza]

thoughts and prayers

[warm cave]

Time to meet out new ruler 📏

May agi be benevolent

[tiny plaza]

and consumer friendly

[minor shard]

lol little buggy still it sounds like

[warm cave]

Sounded better now, but yeah was pretty choppy

[untold edge]

those ai voice cracks are wild

[nocturne leaf]

Did I buy a Pro subscription for chatgpt 4 in vain? Lol

[fossil sentinel]

What's the spring update about

[halcyon coral]

Free GPT-4 for every1 as well as: Data analysis, image input, GPTs and improved voice chat.
All powered by the new GPT-4o model.
2x faster, 50% costs of GPT-4 Turbo via API

[fossil sentinel]

Free gpt 4 on their website?

[halcyon coral]

Yes

[fossil sentinel]

Wait that's kinda goated

Not even kinda

[warm cave]

Perplexity when GPT4o?

Hopefully it will be cheap enough to be a free model

[halcyon coral]

Probably very soon, the team usually integrates new versions of existing models quickly.

[gentle dirge]

50% costs of GPT-4 Turbo is still expensive ngl lol

but damn its crazy that gpt-4o will be free to try

[acoustic wraith]

https://openai.com/index/hello-gpt-4o/

For the spec of the model

Some stuff that they didn't say on the youtube live is it is also rendering 3d and integrating logo on photo mockups

[tame current]

What's the context size? It looks like they already killed the Gemini's main advantage, audio input

They can never keep up

[tepid pilot]

They alr had audio, it was just a separate model

[halcyon coral]

128k

[acoustic wraith]

They seem to have new tokenizer compression so context size will be way improved for non-english users

[tame current]

It wasn't the same, it was just transcription, and only to english. This one can talk about the input

[tepid pilot]

Yea

[tame current]

Hope it won't be lazy like the last GPT4 Turbo was

[untold edge]

🚀 New AI Model Update: GPT-4o! 🚀

1. Multimodal Capabilities: GPT-4o can reason across audio, vision, and text in real time. It accepts any combination of text, audio, and image inputs and can generate outputs in any of these modalities[1].

2. Improved Response Time: The model can respond to audio inputs in as little as 232 milliseconds, with an average response time of 320 milliseconds, which is similar to human response times in conversation[1].

3. Enhanced Language Support: GPT-4o has shown significant improvements in handling text in non-English languages while maintaining strong performance in English and coding tasks[1].

4. Cost Efficiency and Speed: The model is 2x faster and 50% cheaper in the API compared to GPT-4 Turbo, with 5x higher rate limits[1].

5. Advanced Vision and Audio Understanding: GPT-4o has enhanced capabilities in vision and audio understanding, surpassing previous models[1].

6. Unified Model Architecture: Unlike previous versions that used separate models for different tasks, GPT-4o is trained end-to-end across all modalities, meaning a single neural network processes all inputs and outputs[1].

7. Safety and Limitations: New safety systems have been implemented to provide guardrails on voice outputs, and the model has undergone extensive external red teaming to identify and mitigate risks[1].

8. Language Tokenization Improvements: The new tokenizer used by GPT-4o compresses text more efficiently across different language families, resulting in fewer tokens needed for processing text[1].

9. Model Availability: GPT-4o is available in the free tier and to Plus users with up to 5x higher message limits. It is also accessible in the API as a text and vision model, with plans to launch support for audio and video capabilities to a small group of trusted partners[1].

[tepid pilot]

It’ll prolly start good and then get progressively worse

It always happens

They get a bunch of money from ppl

And then make is as bad as possible to increase profits

Save on computer I assume

[warm cave]

It’s so fast

[mossy fox]

Is it lazy?

Try asking it to create some code

[warm cave]

Idk, have only asked it to write me an essay about frogs, I really hope they fixed the lazy code issue

[untold edge]

coding is slow @mossy fox , same as before tbh

everyone & their grandma is probably on the website rn tho

[mossy fox]

ye

I mean i need it for reasoning about code more than i want it to be a code generator for me

[untold edge]

@mossy fox this took around 10secs https://chat.openai.com/share/7d54f2f8-bd70-4665-9a84-492dd9b63921

[dusk shadow]

So, are we going to have gpt4o in Perplexity too?

[halcyon coral]

#💬│general message

[dusk shadow]

Sorry, I just came here

[mossy fox]

o

[tame current]

It's already in ChatGPT? I don't see it. Is it Pro-only for now?

[mossy fox]

i'm looking at the demos here https://openai.com/index/hello-gpt-4o/

the mobile app is insane

(if they are not lying lol )

[dusk shadow]

What about the Vision feature?

[nocturne leaf]

Does anyone have a problem with Perplexity responding very slowly? Time to switch on the chatgpt?

[halcyon coral]

In principle, there is no reason not to use improved models if they cost less and are better.

[tame current]

Nvm

[dusk shadow]

yeah, I get that. But how would you implement this "Vision" feature in Perplexity?

[halcyon coral]

Just like the current feature

[untold edge]

update to assistants api as well https://platform.openai.com/docs/assistants/whats-new

[dusk shadow]

Which current feature?

[halcyon coral]

Image upload

[devout geyser]

well GPT-4o looks impressive so far, very fast to respond. The only possible downside is that it appears it might be limited to 2048 output tokens, not 4096

[dusk shadow]

Ohh, but in real-time it would work through a web-socket solution instead of a REST-API, right?

[warm cave]

This is the only bench that they reported it did worse in: DROP https://www.perplexity.ai/search/What-is-the-E8wtVa7OTBGy1AJlDH7YKw

[pastel pecan]

DO you have a section where I can find resrouces regarding your Overall API implimentations on a tool that im building?

[acoustic wraith]

I am trying it in Playground on enterprise version, and so far I still prefer Claude Opus vanilla expression, gpt4o still sounds like chatgpt😕

[mossy fox]

Main takeaway im getting from this is the amazing phone app

[fading moth]

I don't know if GPT will overcome Opus ever, honestly. Even if they dropped GPT5, it would have to be a big jump.

By the time they drop GPT5, Opus will have gotten updates.

[dusk shadow]

apparently it's gonna have around 1.5T Parameters

[tepid portal]

It looks like they copied Perplexity.ai search, basicially it lowered gaps between ChatGPT and Perplexity.ai

[fading moth]

Yeah, but it's not about the size. It's how you use it.

[dusk shadow]

multiple factors, but yeah

[untold edge]

out of curiosity, what other factors?

[dusk shadow]

Training data for example

[tepid portal]

But honestly? I triying my prompts again in ChatGPT, which reacted badly or gives me bad results, but Calude 3 Opus can do it. Now? ChatGPT works perfectly.

[mossy fox]

4o or standard chatgpt?

[tepid portal]

Even regular ChatGPT, I'm shocked xD

[tame current]

Maybe they reverted to normal GPT4, not the Turbo brainrot

[tepid portal]

idk, let see how GPT-4o will perform.

[tame current]

For now, it seems less lazy, but maybe it will be as bad as before when I'm going away from generic stuff

[tepid portal]

Yeah

But GUI of OpenAI, damm it so good 😄

[stuck topaz]

hopefully gpt-4o better than opus. So that I will cancel perplexity and switch to openai once again. perplexity free default AI is good enough. Right now the gpt4 from perplexity lose context everytime and opus is limited 😃

[warm cave]

When Claude Opuso

And Sonneto

Haikuo

[nocturne leaf]

Opinions
So
Is perplexity pro worth it now?

[untold edge]

@nocturne leaf its been worth it

[warm cave]

Depends on why you are here

[untold edge]

https://x.com/inafried/status/1790083063374033046

Ina Fried (@inafried) on X

A couple tidbits I've confirmed as well. 1) The mysterious GPT2-chatbot that showed up on benchmark sites was GPT-4o. 2) OpenAI did desktop version first for Mac because "we're just prioritizing where our users are." @miramurati told me. https://t.co/SydKkXyF4P

Twitter · May 13, 2024, 6:13 PM

[warm cave]

I think it’s live, I have GPT4o on chatGPT rn

[kindred folio]

same

is gpt4 o gonna be added to perplexity mobile and web

[warm cave]

I bet so, idk why they wouldn’t

50% cheaper

[chilly nova]

so is this gpt4-o on perplexity or not?

and why should i use it instead of claude 3 opus

[lone citrus]

Omg, Gpt4-o

[untold edge]

go watch the stream

[chilly nova]

which

[untold edge]

@chilly nova https://www.youtube.com/watch?v=DQacCB9tDaw #💬│general message

[warm cave]

I think it’s still the same story, GPT is the logic reasoning king with top end STEM reasoning and Opus is great at reasoning but is geared to creative work and nature writing. But I have yet to really try GPT4o so I guess we will have to see how everything lined up

[chilly nova]

that's bad, there are better videos on their website

oh right, there's actually a good use case i have for this. i often get those bs iq tests. this would be perfect to test if it got better at solving matrix puzzles

[untold edge]

is this bait?

[chilly nova]

https://openai.com/index/hello-gpt-4o/ watch videos here

the guy walking around london maybe most interesting to see how fast it reacts. but different use case

still, this can definitely kill all tourist guide jobs

[torn sierra]

Same here 💪🤗

[cursive jacinth]

hi sir, if i subscribe to Chatgpt today sir, is this GPT-4o or what is it called sir, can it be used now sir?

[tiny plaza]

apparently

sorry i meant, apparently sir

[cursive jacinth]

nice reply sir. How is the performance sir?

[untold edge]

its not bad sir

[cursive jacinth]

you mean it has a significant improvement sir? in what way sir?

i am tempted to subscribe sir

[idle mulch]

Just for info for everyone who is annoyed about the 50 / day Claude 3 Opus limit that Perplexity silently introduced and never lifted again (although promised), YesChat offers 100 Claude 3 Opus / day (3000 per month) along with 200k context length (capped to 32k at Perplexity) for $20 a month including tons of other AI bots (Midjourney, Suno, etc.).

https://www.yeschat.ai/pricing

[tiny plaza]

i don't know yet sir but sir i wonder if it is related to my name sir can any sir in the chat please confirm sir?

[tame current]

For now, it's MUCH better than gpt4 turbo sir. Not lazy anymore. Will do stuff without asking it million times or being stupid