💬│general

[tame current]

^

[snow briar]

Mhm ok. Yea I tried it and it was okay but gpt4 was more thorough with answers

[warm cave]

GPT-4 still king, but if we have full 200k context then it’s a great addition

[sweet jasper]

Lol no

Claude 3 blows gpt4 out of the water

Its crazy even in coding

[snow briar]

What about using sonnet tho?

[warm cave]

Benchmarks, or personal use testing?

[sweet jasper]

Personal testing

But need more tjmd

Time, so far I'm very impressed

[snow briar]

The bit about less hallucinations than gpt is great

[warm cave]

Then it may be the best tool for you, it sucked at some calculus and reasoning questions I gave it, and gpt 4 had no problem

I’m glad to see competition

[sweet jasper]

Maybe so, I might just switch back if it starts flunking too many questions

[warm cave]

I’m planing on running a more in depth test on a questions set I put together, it’s hard to really compare models that have such broad use cases

[paper patrol]

Anyone using Claude 3 through Perplexity Pro while being from a country not on the list of supported countries by Anthropic?

[snow briar]

Try asking it to summarize a page and comlare the results

Compare*

[tame current]

it isn't lol, claude 3 opus is genuinely mind-blowingly better

particularly in programming tasks

[halcyon coral]

Old benchmarks, also without Turbo

[tame current]

i know

but it is still better

even with turbo i reckon its at minimum on-par

[snow briar]

I’m kinda interested in how sonnet compares with gpt4 because I don’t think having 5 uses of opus is really practical.

[warm cave]

I would take benchmarks with a grain of salt, from my experience GPT-4 is much smarter, but I also may have different use cases, and different models have pros and cons, plus I have used gpt-4 many times a day for a long time so I might know how to have gpt work for me, as I spend more time with both Claude 3 and gpt-4 I may change my mind, but I feel more impressed by gpt, but if you like Claude 3, it’s probably the best tool for you

[tame current]

then what're your use cases

i definitely wouldn't say 'much smarter'

[sleek tapir]

I have my default browser search set to perplexity, so I'd likely waste all five searching a website name

[tame current]

the only area where gpt-4 is noticably better is raw maths and some logic questions

[sleek tapir]

[tame current]

creative & code = claude

[brittle dagger]

i personally use a keyword shortcut in the firefox address bar since i don't always need to use perplexity for search

[south kindle]

i think we've posted this image like 12 times since it was released. can we just get a bot command to summon it. lolol.

[warm cave]

Again, I am saying from my uses I feel that way, I am making an objective statement, and you are fine with disagreeing, I feel it’s smarter when I interact with it, I have only done limited testing, but calc 2 and hand made logic problems make be believe gpt is still the king

[sleek tapir]

That's smart

[snow briar]

So if you had to pick which model to use between gpt4 and Claude 3 sonnet for summarizing text in a concise simple way, which one would you choose and why?

[hallow niche]

Not seeing Claude 3 in the Perplexity iPhone app (on Pro)

[halcyon coral]

Claude 2.1 = Claude 3 😉

[warm cave]

If it really long text where 200k context is needed prob going to use Claude 3, but if it’s a few pages, gpt-4. Creative writing probably Claude 3. Math and logic GPT-4

Benchmarks have a big problem with data contamination, application to real life scenarios, and more. If you want to know the best one for your use case, use them side by side for a while

[tame current]

^ side by side comparisons are best done on chat.lmsys.org : )

[warm cave]

Yes

Oops, I said objective instead of subjective

[snow briar]

Yea im gonna test out both models more and compare. I use it usually to summarize lectures and articles into concise points with topics, along with analogies and examples. So far gpt4 was good with that but I haven’t tried it much with sonnet yet. I’m hoping sonnet requires less prompts to extract info I need

Sometimes gpt4 hallucinates and generates some info that slips pass by me then I gotta re-prompt

So if the other hallucinates less, that’s a plus

[tame current]

sonnet is very good in terms of recall from context

so at 200k context successful retrievals happen >90% of the time

although i think there are still some hallucination problems

[warm cave]

Yeah, they talked about less hallucinations, if this is accurate than it’s a great benefit!

[warm cave]

I don’t often required more than 32k context, but the needle in a haystack looked good, and I could see this being useful for many applications

When I judge an LLM I think I put most of my weight on short term logical reasoning, word problems, math, technical analysis, explanations. But for context and other features I see it as a nice bonus. But I know context is becoming an increasingly important part of what makes a good model.

[stoic wave]

logical reasoning is the issue all the models have

[warm cave]

Yeah it is, and why we see so much talk about mode 1 vs mode 2 thinking, how it differs from how we think and if we think AI will reach a plateau.

[snow briar]

If it can explain/summarize a block of text I need without missing any key info and in a way that a 5 year old can understand, I’m cool with that.

Very curious to try opus but I really dislike only having 5 uses lol

[fossil sentinel]

Gemini 1.5 will be better for long context imo

[fossil sentinel]

Not as good as gemini 1.5

[distant juniper]

Sorry to ask but other than marketing, whats the point of having 5 queries on Opus?

[warm cave]

I think they said that Claude 3 was limiting usage and they have to wait on them

[distant juniper]

Ah, makes sense.

Another thing I really never understand is the exact context size for the models and the exact model in use. For example, is GPT 4 using 0125? I'm also pretty sure they limit the context to like 4k or something like that.

I feel like this could be more transparent.

I also wonder if the model used to search the web is their own model which then summarises things for the chosen model to then actually process and give an answer to rather than the chosen model being the one who goes through the web search.

[harsh verge]

Yes, and that's no problem. Claude Opus and sonnet will work in Perplexity Pro!

[sweet jasper]

This is the most important part, we don't want models to misconstrue bits of code that were previously correct into something else

For gpt4 the recall drops significantly after the 64k token range

Before that it has 99% recall I think

[heady quiver]

I reported this issue as well

[fluid hawk]

the problem with Opus is the outrageous pricing

[queen flame]

Claude is better than GPT-4 at coding.

[livid aurora]

no support for images

[queen flame]

Devs out of a job by the end of the decade

[fluid hawk]

I tested it on LMSYS chat, it wasn't at all worth $15 dollars per million input tokens and $75 per million output

[obsidian acorn]

question, how do you use the 5 Opus cause whenever I just use it it uses sonnet

[queen flame]

Also why am I getting more than 5 opus requests, I am on my 10th

[livid aurora]

You already reached the limit

[obsidian acorn]

oh really I don't remember using it at all

[heady quiver]

What’s this in price and out price??? With examples please

[livid aurora]

yeah strange I heard other people talking about it

[queen flame]

It might mean 5 opus requests per day and they rollover

[livid aurora]

Guys did you know image support is not there yet with claude 3

[fluid hawk]

any proof besides the sketchy benchmarks? humaneval is very unreliable, like MMLU and benchmarks in general just aren't a good way to compare, they should do blind testing with 1000 people, and see who prefers claude 3 opus over GPT-4

[fossil sentinel]

Anthropic had to have used some different prompting methodology to get their benchmarks

[queen flame]

been working with gpt-4 coding since it came out. the difference is palpable

it doesn't truncate code nearly as much

it doesnt try to get you to do work

[fossil sentinel]

I've tested some subjects across math, chemistry, physics, and stats but gpt 4 usually outperforms Claude in my experience

[queen flame]

in the end the difference could be due to the fact that openai is trying to deal with scaling by getting gpt4 to output shorter responses

[fluid hawk]

is it better at debugging tho? in my testing, it couldn't debug code nearly as well as GPT-4

[queen flame]

better for python, can't speak for real programming languages

[fluid hawk]

read the last sentence

[fluid hawk]

I tested it in python, and it couldn't debug.

[fossil sentinel]

[queen flame]

doing it rn and it's working. 🤷

[fluid hawk]

I just chose problems on edabit's Python "expert" challenges

[fluid hawk]

on lmsys or claude.ai?

on lmsys, where I tested it, it was garbage compared to GPT-4

[queen flame]

api key 👍

btw anthropic ppl if ur reading this I am just pumping and dumping your bot. can't wait to move on to the next big thing after clod, I mean claude

[odd mauve]

I would prefer to use Claude 2.1. It's better than Claude Sonnet. So in essence, this change is for the worse. Unless you do only 5 queries/day. Please restore the Claude 2.1 option!

[fluid hawk]

weird how claude 3's cutoff is august 2023 but it doesn't know about the open source Llama models, neither version 1 nor 2.

I think they are lying about the cutoff to some extent

[peak bloom]

We’re sorry but the Perplexity Phone (2a) offer cannot be redeemed as you cannot purchase Nothing products via nothing.tech in your country. 😕

[lament imp]

Yeah, seems like it is US/UK/India

[odd lava]

I would love to bought that 2(a) if there's an autograph by Axwell or Ingrosso or Steve Angello

[ivory bluff]

Hard agree from my experience today

[marble trail]

Hi admin, I have been testing the Collections feature, I imagine it has been created with the intention of doing something similar to the GPTs. However, it is not possible to add documents for it to extract information from them. Is the incorporation of this function planned?
I would also like to know if it is planned in Perplexity's roadmap to open it up to the development of APIs specialized in various topics, just as it happens with the GPTs. Thank you in advance for your response. Regards."

[lament imp]

I'm not an admin but I am curious what do you mean it is not possible to add documents and extract information from them?

[marble trail]

Yes, in general search, but not inside Collections feature

[lament imp]

You should be able to do it with each individual conversation/question, but I don't think there's a general option where you can upload the file to Collection so that it can reference it with each conversation in the Collection.

[lament imp]

Also, when you refer to "open it up to the development of APIs", do you mean specifically opening up collections to APIs, or Perplexity being open to APIs in general?

[marble trail]

I mean Perplexity in general. Something similar to GPT´s of OPEN AI

[lament imp]

Perplexity does have apis, where you can prompt the service with an inquiry and a model.

[livid aurora]

It's really code at coding

good

lol

[supple ermine]

Does anyone know how to get a higher token limit using the API? I got pro but it is still 4k

[lament imp]

The more so becomes available if you have perplexity pro, but it does come with a few limitations monetarily.

[halcyon coral]

You can't get a higher token limit for the API.

[supple ermine]

😦

https://blog.perplexity.ai/technical-faq/what-is-a-token-and-how-many-tokens-can-perplexity-read-at-once this page should be updated to say that explicitly

[halcyon coral]

Perplexity.ai and pplx-api are independent

[lusty mantle]

so is perplexity experimental ai your own ai?

is it less offended by everything and more flexible in answering questions? cause i got it to make a violennt zombie gore scene which was really cool so i figured that its less restrictive

[agile jay]

I believe that default/free tier is a finetuned GPT3.5Turbo model, while the experimental one, is a 70B finetuned llama model.

[halcyon coral]

Perplexity and the pplx-api belong together but are two different products, one is the WebUI and the other is a set of LLM's including custom online and chat LLM's for developers via API.

[tame current]

The Nothing brand is pretty cool, but it’ll be awhile before I pickup their phone

iPhones and Galaxies are simply unbeatable

[agile jay]

There is the same offer for that rabbit device.

[lusty mantle]

im so down for them releasing or putting a non restrictive model for users who like writing horror books and stuff i think elons ai allows that

i think im not sure havent tried his yet

[halcyon coral]

And yes, the own models have a lower morale and don't refuse to answer often.

[tame current]

I’m even more skeptical of the rabbit device

Part of what makes phones so dominant is versatility

You can browse the internet, listen to music, snap photos, and play video games

All on one device

[lusty mantle]

i just bought this yesterday im honeslty impressed im able to use all these ai models for 20$ a month its insane and i just found out you can just turn google off and use it like normal

why is this not more popular

[tame current]

Marketing

The product is great, but perplexity just isn’t very well known in mainstream circles

[lusty mantle]

kinda good tho maybe allows us to have faster response times do to lower volume?

[round plover]

hmm 2a phone how much it cost??

[tame current]

349 dollars

[lusty mantle]

what is a 2a phone?

[round plover]

ok, so only $149 more than the Rabbit device

hm

[tame current]

It’s a phone from the startup nothing

[lusty mantle]

i looked at it but what makes it special?

[tame current]

Nothing tries to make your phone less intrusive, basically

The phone is only black and white, both with the frame and with the software

[round plover]

woah

[tame current]

Their android skin makes all your apps black and white, and very pixel-esque

[lusty mantle]

kinda cool i guess i thought it would have like ai built in or something

[tame current]

I think they have a lot of potential as a company, their CEO comes with a lot of experience from OnePlus

[round plover]

interesting

[tame current]

But for now, their products aren’t as good as a company like Samsung or Apple

[lusty mantle]

yea i think samsung just released there new phone with ai

[tame current]

Yes, though their chatbot isn’t very good

It was moreso using AI to optimize existing phone features instead of creating a home grown chatbot

[lusty mantle]

what is wolfram alpha?

and yea havent tried it yet

[tame current]

It’s an AI math model

[lusty mantle]

is it good at math?

[tame current]

Yeah I would say so

[agile jay]

No, it's not an AI model.

It's an API

From wolfram alpha

https://www.wolframalpha.com/

[graceful osprey]

I love the fact that I can access different models and have them look at search results to supercharge their responses. But sometimes I just need to jump in and have a quick exchange with the AI to get ideas or help me shape ideas. However, I find having to switch to writing mode and disabling copilot a bit cumbersome. Is there a way to quickly go into “chat only mode” as it were? Good old ChatGPT style. How do you guys do it?

[round plover]

anyone bought rabbit recently / can verify still giving code? How can i tell if still <100k sold?
still need to think about it but might buy it if its still going on

[minor shard]

They are still giving out codes during Batch 6 (current batch)

[paper patrol]

As a brand new Pro subscriber, I tend to agree, it could be less intricate to get to image generation. Perhaps it`s just me being a newb, perhaps some kind soul could point me to a streamlined method to get to image gen without wasting all my daily Claude 3 Opus prompts unknowingly...

[rough haven]

Usually I say: Summarize the highlights of our conversation. Then I copy paste that into a new dialogue with write mode.

I continuously start new dialogues and reopen old ones etc.

As if you're working in a team with people and have to bring them up to date again before you can make use of them

[round plover]

ok, i bought a rabbit

lets see when i get the code

:)))

im part of the future now!!!!!

:DDDDDD

[round plover]

ok, got the code

anyone know if i can redeem the code and make it add it to the free trial?
for some reason i can't redeem, possibly due to free trial being acgtive?

[warm cave]

Its good at basic math, but for harder tasks it really needs to have code interpreter integrated into perplexity

[signal hamlet]

Right, please send us a message to [e-mail] with all the relevant details, we'll help to apply it.

[warm cave]

Hey, do you guys know how long it takes for Claude 3 to be added to the "LMSYS Chatbot Arena Leaderboard" I know its already available to chat with, but how long dose it take to get ranked and placed?

[round plover]

Thanku

[prime ravine]

What’s a rabbit?

[lusty mantle]

has anyone used claude opus? i heard its good

[round plover]

Rabbit M1 device with built in ai and it does stuff for u

[warm cave]

https://www.rabbit.tech/

Perplexity has a deal with them where you get 1 year of perplexity if you buy one of the first 100,000 units

[manic anvil]

Pretty expensive for such a small plastic thing with a circuit board inside

[gaunt crater]

I want one tho 😅
And despite my recent whinging over the loss of Poe 2.1 and not being crazy about Poe Sonata Perplexity is such a solid AI search and even more solid vision of how it will improve as the models continue to improve that ~-$240 for it is... I wouldn't say just subtract it from the price but take it into consideration if you were gonna do Perplexity usage anyway.

[warm cave]

It could be cool or not, but i was about to buy perplexity for a year anyways, so I thought, hey free device

[gaunt crater]

What if you open it up and there's no circuit board, just a paper coupon for Perplexity

[warm cave]

In the very least it will be a fun little toy

I almost bought it before i saw the promo

[manic anvil]

Save money and buy better internet things, this thing isn't worth it. Buy prime for a year or Netflix whatever, you get more out of it.

[gaunt crater]

I think the issue with Rabbit is that it's an exciting idea (multimodal AI for controlling other devices) that is only more than a novelty if it's really good. And that's a lot of faith on a preorder.

[warm cave]

I didn’t buy it for the little device, i bought it for the year of perplexity, I use perplexity 100s of time a day, i think I will get more use out of it than netflix

[gaunt crater]

But will it support Netflix-style Sora video by the end of 2024

(probably not)

(probably)

[warm cave]

I want sora in apple vision pros so i can go full ready player one and enter the matrix

[manic anvil]

Ki isn't there yet, you can tell it's light years before when a video was created with Sora

[gaunt crater]

For all of their ClosedAI flaws, OpenAI is such a leader in the field

[warm cave]

by a lot

[manic anvil]

I find Gemini better, the free version of ChatGPT is unfortunately very crumbly

[gaunt crater]

Even Sora's failures are fascinating, like the liquid bubbling up from the table on the spilled glass, etc.
Is it even true that it trained on a bunch of Unreal Engine interactions or is that just made up nonsense until OAI does a talk or a paper? I don't remember reading anything like that on the article that went up on the Sora site.

[warm cave]

Yeah, free gpt-3.5 is very old now, gpt-4 is the cutting edge, and hopfuly not too long gpt-5

[manic anvil]

With Google you get better human answers, with ChatGPT it's always the same approach and you realize that it's superficial.

[warm cave]

system prompt effect the style a lot, GPT-4 is very steerable from my experience

[manic anvil]

And costs money

[warm cave]

GPT-4 has a focus on concise and accurate answers, not everyones favorite style

Claude 3 can be great for longer answers, gemini i think i will only use for the 1m to 10m tokens if i ever get to use it

All of the models have good use cases, but OpenAI has been the golden standard for a while, and only recently has competition started to give them pressure, I wonder if there will be a big gap when GPT 5 is here, we will see

[warm cave]

Yeah, I mean the best free model is its own conversation

there are lots of pros and cons with free models i thinks its best to try a few and see which fits you use case the best

[craggy prairie]

Is the Claude 3 Opus limit real? I’ve done like 15 messages back and forth and the UI still says I’m using Opus???

Is the UI not updating properly or there is no real limit?

[warm cave]

I would guess it the UI not updating, I have heard from others that the name only changes when you start a new chat after already using up all of your messages

[craggy prairie]

Wow if this is sonnet then I’m very impressed. I was impressed thinking it was Opus this whole time!

It’s nowhere near as lazy as GPT-4 for coding…no more annoying back and forth. It’s almost like the original gpt4

[warm cave]

Thats good, I have only tested it a little, and i am really glad have it as an option!

[austere kestrel]

from their Discord

[warm cave]

Thank you! I look forward to seeing where it ranks

[austere kestrel]

updated leaderboard btw - two main GPT4-t variants tied at the top; Mistral Large making a solid debut,

[warm cave]

Nice, I don’t know about you, but I look to this leaderboard more than all of the benchmarks coming from the company’s.

[austere kestrel]

totally

[craggy prairie]

it still shows opus for new chats for me… weird

[warm cave]

Maybe you have a good glitch then 🙂

[sand storm]

Yes I used claude opus

Its fun

[sand storm]

Yeah I tested it but I got bored.

I have tested opus, gemini advance, gpt4

pi by inflection ai

Here is my rating:

Image Generation:
Midjourney: 10/10

Music:
Suno: 6/10

Empathy:
Pi: 8/10

Conversation:
Pi: 7.5/10

Storytelling:
Gemini 4: 8/10
ChatGPT 4: 5/10
Claud Opus: 5.9/10

Eastern Philosophy:
Pi: 5/10
ChatGPT4: Better than Pi

Coding:
ChatGPT 4: 7/10
Gemini Advanced: I haven't tried it yet
Claud Opus: Haven't tried

I haven't really tried testing it on math.

[round plover]

anyone know when Rabbit M1 ships?

[hallow linden]

claude opus is only 2x the price of gpt-4 turbo. I feel like 5 per day is pretty low

[ivory bluff]

I just want my 2.1 back. Opus and Sonnet's writing is repetitive garbage

[gaunt crater]

At least we'd have a middle ground of Claude's better writing quality on 2.1 than Sonnet. Sonnet has good prose but comparatively poor logic, and I don't really have enough requests with Opus to compare.

[ivory bluff]

I've tried with Opus using poe, and honestly I think 2.1's logic is better than Opus's too, in the writing arena at least

I still really think 2.1 has its place.

[violet venture]

Any of you have experience with validating content, if it's written by AI or not?

[minor shard]

Do you mean R1?

[agile jay]

It's not possible unless there is a clear error.

The main way I do it is by looking at stuff a person has written in the past, and seeing if the vocab and sentence structure is similar or not.

But you can't paste it in a machine and get a Yes this is AI, and No this isn't

[violet venture]

I didn't think it would be, with latest generations of LLM. Still, sites out there claim to be able to. I have this 2,000 word document I've knitted together with AI help (Perplexity, varying models), which https://gptzero.me/ is 88% confident is human written, 12% mixed, and 0% AI - but why should I trust that?

Oops, sorry, didn't intend to spam with that link :>

[agile jay]

Yep, openai tried to make one when ChatGPT blew up, but they realised it was basically impossible.

Any site that says they can, is just straight up lying.

[violet venture]

We are way beyond the Turing test...?

[agile jay]

Yep, way past it.

GPT 4 writes better than 90% of people already.

[violet venture]

Thank you for confirming my suspicions - although, you could be a lying AI 😉

[agile jay]

Take all the stuff you've written before and run it through the checker, lol.

If you are in uni, take stuff your lecturers have published and see what results they get...

You will quickly see how bad they are.

[violet venture]

It's the first semester term paper, they don't have anything to compare with 😛

Besides, I am confident with my vocabulary, both in English and Danish.

Besides, it's not cheating if I actually understand and check references (AI still sucks at that, paraphrasing citations), is it?

[agile jay]

Yep, and with the sources from perplexity in academic mode, it should look pretty legit....

[violet venture]

I suffer from doing everything at the very last minute. Had a week to write 7-10 pages, it's due in ~6 hours. Still need one more page to reach the minimum, did everything since... yesterday.

I'm happy with my Perplexity Pro 😄

[agile jay]

What's the paper on?

[violet venture]

Psychology, Trait Theory, perspectivation, etc.

[agile jay]

Ah, should be pretty easy then.

Not a topic which needs very accurate calculations.

[violet venture]

Well, it's my first paper on this level, first five days I went in circles trying to find an approach. Easy is relative, but...

No, that's true.

Yet... I can't just make up stuff 😄

[agile jay]

Lol, if it's psychology, you normally have to reference the case studies.

[violet venture]

Yeah, there is no case study in this one. It's a re-exam (mental breakdown first attempt), they changed it from case study to theoretical account. Is that the correct term in English?

Not sure if that's what you meant, there are experiments in the curriculum texts of course.

[sullen cloak]

How come Gemini was removed?

[violet venture]

I did prefer Bard to GPT-3.5, don't ask me why.

[agile jay]

No idea, didn't use perplex when they still had them.

[violet venture]

Same, but I just arrived a few days ago.

[languid nebula]

They said it wasn't being used that often as was generating too many bugs.

[round plover]

yeah

[violet venture]

I tried to be accurate, it still effed up. Not a totally invalid answer, but some hallucination is going on.

[sand storm]

lol

[sand storm]

I find it's writing not the intriguing

[lusty mantle]

claude on there site lags alot when generating code

or just typing in general

[plush fjord]

can perplexity models do functions call?.

[hallow linden]

Technically just the one. But that’s all I’d trust an LLM to do at this point

[plush fjord]

I tried replacing GPT4 for Perplexity and it does not call the function

[hallow linden]

The function is the web search

I was making a play on words

Or a joke or something

But to actually answer your question, no, it doesn’t do functions

[warm cave]

As far as i know LLM's don’t call functions, your code calls a function, its weird since OpenAI called it function calling, but really its simply GPT-4 returning valid json, then your code parses it and then uses the info to call a function, so really almost any llm can "function call" the question is what LLM is consistent at returning the valid format that the code will recognize and then call the function, but maybe someone else understands it better than me and can explain it or correct me

[humble swan]

#📢│announcements this seems like a excellent deal

[sweet jasper]

@languid nebula are you using the api, you have to have application logic to check if the message follows a certain format

[tall plover]

hi

[livid mantle]

whats the experminetal wha

whats the difference

one is slower and the other is faster

[serene sand]

https://www.perplexity.ai/search/The-Experimental-model-s..Jlk.hSqabGhlEz.YhNA

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

[livid mantle]

thankyou sir

[serene sand]

#💗│sharing message ☕

[magic loom]

People

How do I add an anonymous thread to a collection

I have been using a thread since before I had an account

And now when I login to other devices, the thread doesn’t show up

Nor am I able to add it to a collection

@serene sand is there a support section in the server

[grand sand]

Funny how the CEO of perplexity is the complete opposite of elon musk

He's open and honest

[wraith crow]

You can ask your questions here in #💬│general or #⚡│ask-community.

[wraith crow]

That is currently not possible. If you kinda want to save that thread you could fork it with your account by asking a question while logged in on that thread from the anonymous user.

[knotty frigate]

Has anyone had any success with eleminating code snippets from responses from GPT-4 unless being explicitly asked? I'm trying to fine tune AI prompt but no matter what I write it doesn't stop giving code examples.

[livid aurora]

Gpt 4 is lazy and dumb

[fluid mauve]

Regarding the nothing 2a and perplexity pro subscription, I have already pre-purchased the Rabbit R1 which comes with a Perplexity Pro annual subscription. If i purchase the Nothing 2a phone can the $200 credit be used on for my year two perplexity subscription? Basically stack em up?

[fluid mauve]

Perhaps this is a question for the Nothing marketing team whos budget I am sure this will come out off.

[fluid mauve]

Just used Google Gemini Pro, ChatGPT Pro, and Perplexity Pro to answer a question from the same prompt. From the results of my tests. Perplexity Utilizing Mistral Large , or Google Gemini gave the best results from my consistent prompt. Interesting excercise but one that i dont fancy carrying out again. too much reading.

[cursive jacinth]

hi sir

[fluid mauve]

Hey @cursive jacinth Ask your question. I know you have one

[cursive jacinth]

just mic testing sir.

[agile jay]

Indian?

[agile jay]

Lol, you could pass the results to a model to evaluate instead.

[rich wren]

Hi, I do have only Cloude 2.1 available in my app on iPhone. I have no updates available in AppStore. Is Cloude 3 available only on desktop?

[agile jay]

Yep, on web. The apps haven't been updated yet.

Not sure what technology they are using for their apps. If they are using react native, the update should be pretty quick. If they are using flutter or something similar, it will probably take a week or so. Since apple is really annoying with their vetting process.

[foggy whale]

What was your question?

[fluid mauve]

It was a long and detailed question with specific information required. I will post the full prompt when I get back.

[fluid mauve]

I could. Or I could get the wife to evaluate as she is the Subject Matter Expert

[hearty quartz]

CLAUDE 3 OPUS is clearly superior to CLAUDE 3 SONNET... Can you put CLAUDE 3 OPUS back officially, 5 uses of CLAUDE 3 OPUS is too little

[agile jay]

I assume anthropic has limited the opus API which is why it's currently limited to 5/day, unless you know the glitch.

[austere kestrel]

bring back Claude 3 Haiku

[agile jay]

I used Claude opus > 35 times yesterday.

It's pretty good at programming but still needs steering.

[solar trellis]

which better with gpt 4 turbo?

[agile jay]

Yep, Turbo is too lazy IMO

Leaves a lot of code out with comments etc.

[sweet jasper]

Uh

Dawn you printing money

[agile jay]

Lol, it was pretty easy to figure out. I'll test it again today, and see if it's still working.

Too busy with a birthday today, so don't know if I'll have enough time.

[fluid mauve]

Here's my prompt - Any constructive criticism appreciated.

Question
Write a report based on the Chartered Institute of Public Finance and Accountancy (CIPFA) teachings, about carrying out the monthly outturn process.
The report should be written as a student studying the subject but who has 5 years of experiance in the area.
The audience is a CIPFA examinier
The report should adddress the deficiencies shown in the deficiancies section
Each deficancie is prefixed with a ##
Each deficancy has an index denoted by capital letter, a numeral, a period, and a numeral. Examples of these deficiancy indexes aere S1.1, B4.5,S2.3

Highlight in the answer where each of these index numbers are referenced.

Deficancies
BUSINESS INSIGHT - Influence the impact of business decisions on relevant and affected communities based on an appreciation of different organisations and the environments in which they operate.
S2.5 - Uses financial management expertise to generate clear strategic alternatives, based on data and evidence gathered from multiple sources.
S2.6 - Seeks out examples of best practice to inform business improvement processes.

COMMUNICATION - Communicate in a clear, articulate and appropriate manner. Adapt communications to suit different situations, individuals or teams.
S3.5 - Build networks to support ideas that strengthen financial knowledge and analysis across the organisation.
S3.6 - Deploy a range of communication skills to change perceptions or perspectives on strategic financial issues.

[rough haven]

When will the 5/day limit of Opus be removed?

[rough haven]

That means you were using Claude Sonnet... After 5 uses it automatically changes. The label says opus, but it's sonnet apparently

[rough haven]

glitch??

[agile jay]

Yep, you can check the network requests to see what models it's hitting.

[austere kestrel]

fwiw I'd say that's asking a lot of the llm, in terms of the requirements for a single output. The way you have structured it actually shows it quite clearly - there are a lot of parts to it. To the extent possible (and just fwiw), if you could consolidate / simplify it, or execute it as flow in segments, I think that may yield better results (assuming that is that you want the output to be detailed / granular etc)

[fluid mauve]

@austere kestrel Thanks. I was under the impression that you gave it enough to get it going. The report output from both Mistral and Google were both detailed and good. I could have promnpted further but I feelthat for my tests it was good enough

[austere kestrel]

good to hear – if it works, it works!

I just find the more tasks that are piled into a prompt, the quality of the output almost degrades proportionally

[jagged pike]

Best model to create dungeons and dragons characters portraits?

[fluid mauve]

hmmm, from what I have seen and looked at more is sometime less. especially around prompting. Have you looked at github.com/danielmiessler/fabric. If you look at the prompts he uses in the patterns directory of say extract_wisdom, which is used to analyse a transcript from a video, his prompt is as follows

IDENTITY and PURPOSE
You extract surprising, insightful, and interesting information from text content.

Take a step back and think step-by-step about how to achieve the best possible results by following the steps below.

STEPS
Extract a summary of the content in 25 words or less, including who created it and the content being discussed into a section called SUMMARY.

Extract 20 to 50 of the most surprising, insightful, and/or interesting ideas from the input in a section called IDEAS:. If there are less than 50 then collect all of them. Make sure you extract at least 20.

Extract 15 to 30 of the most surprising, insightful, and/or interesting quotes from the input into a section called QUOTES:. Use the exact quote text from the input.

Extract 15 to 30 of the most surprising, insightful, and/or interesting valid facts about the greater world that were mentioned in the content into a section called FACTS:.

Extract all mentions of writing, art, tools, projects and other sources of inspiration mentioned by the speakers into a section called REFERENCES. This should include any and all references to something that the speaker mentioned.
Extract the 15 to 30 of the most surprising, insightful, and/or interesting recommendations that can be collected from the content into a section called RECOMMENDATIONS.

OUTPUT INSTRUCTIONS
Only output Markdown.

Extract at least 10 items for the other output sections.
Do not give warnings or notes; only output the requested sections.
You use bulleted lists for output, not numbered lists.
Do not repeat ideas, quotes, facts, or resources.
Do not start items with the same opening words.

INPUT
input:

[livid mantle]

I would say dall e 3

[austere kestrel]

Interesting. This is long, but it has a pattern, and it is basically about retrieval (e.g. extract xx, extract xx); I think that's different than like (just for example) "do a SWOT analysis, then explore the Strengths and Weaknesses in depth; then propose solutions to xxx. etc etc"

[fluid mauve]

good spot

[austere kestrel]

Like extraction is more about retrieval than creation; if you're asking for analysis - the more focussed the beeter

[fluid mauve]

nice. I can work with that

[austere kestrel]

just my 2c ofc - I don't really have any clue ha
(but hopefully was constructive :))

[fluid mauve]

definitly Constructive

[cursive jacinth]

hi sir, why i am experiencing failed message sir

[halcyon coral]

With Claude?

[cursive jacinth]

hi sir, on discord sir. maybe just a bug sir. i am just mic testing sir.

[azure crown]

what's with the sirs @cursive jacinth

[granite grail]

idk sir but I kind of like that sir

[cursive jacinth]

I was raised well by my parents sir. Since birth sir, they told me to always say sir to every sentence sir. I am 33 now sir. Can you blame me to undo it sir?

[azure crown]

sounds like an llm

[fluid mauve]

Or a reespectful Indian

Im assuming. But then again if you make an assumption it makes an As out of you, and umption

[rugged ether]

I want to confirm again when I turn off the AI Data Usage option, none of the attachments or messages that I input and upload in each mode will be used to train the model, right?

[austere kestrel]

just someone respectful ?

i mean who cares

rather that than the opposite lol

but that said for the females here, I'm not sure it works gooddawg - you might need to reassess that bit lol

[fluid mauve]

pfft, it might work in the west, not necessarily in the east

[austere kestrel]

i mean i find emojis generally annoying

but I use them myself, and try not to judge people who use them excsessively

[fluid mauve]

Yes those are, but respect is tolerable, if not a little annoying, but we cant over ride anothers cultural leanings

[austere kestrel]

I do hear ya

It's a global platform

people from all over the world here I assume

[fluid mauve]

Yeah, im geting that vibe,

[proud patio]

Loving the new Claude opus. Huge leap in coding competency

[cursive jacinth]

Hi sir, sound racist to me sir, i am half indian btw sir

[fluid mauve]

Thats not what was intended. Racism has no part in a conversation. we all just need to chill

[halcyon coral]

👀

[sand flower]

new claude answered all my coding questions wrong

[warm cave]

Someone correct my if I’m wrong but when you use the slider perplexity will not use anything you input for product improvement/advertise, etc. however they do use API for some of there AI models, Like GPT-4, I know OpenAI has taken the stance that they will not use data from API usage, but I don’t know about Claude, Mistral, Gemini (When it comes back), etc

So I would check the privacy policy of whatever AI model API you are using along with any API tools, such as wolfram alpha. It’s also possible that perplexity might have enterprise API access and that might have different privacy policies

[tame current]

ok

[glass ibex]

The nothing .tech /perplexity link in #📢│announcements is giving me a 404 error.

[rose fern]

Perplexity down right now? Cant get into the website or the android app

[rustic fern]

Does anybody know a good text to speech extension for claude on chrome one that sounds natural not robotic

[granite grail]

don't know any extensions, but try out microsoft edge if you're using windows or mac. It has built in text to speech natural voices

[pseudo linden]

Can't purchase the Nothing phone in Canada 😦

[vapid onyx]

No issues here.

[rose fern]

Yep, sorted itself shortly after my message 🙂

[tacit meteor]

Hello! Can I limit a search database scope? I want to get informations only from here: https://www.ableton.com/en/live-manual/12/welcome-to-live/

Are there any methods for this?

[brittle dagger]

I have a trial. Do you want me to fill in a specific prompt and compare?

gotcha. writing mode or use perplexity search as well?

Here is the same question, ran through multiple models https://www.perplexity.ai/collections/GPT4-vs-35-e_xK7SZcRDqXJ2iCpZ80Dw

[knotty galleon]

Canada is literally a 3rd world country.

[fierce mantle]

If you click “Pro” (formerly Copilot) does it work with whatever model you have chosen, or does it revert to the Perplexity default model?

[wide stratus]

Guys, why does Perplexity (Android) sometimes not post ANY sources at all? (This list I mean) 😳🤨

[deep mesa]

Hi! I am new on this discord. I preordered the R1 few weeks back and was wondering how to get that free perplexity pro for a year?

[sweet jasper]

Third world countries usually get access first

At least for google models

Because gdpr and studf

[warm cave]

“Pro” still uses whatever model your have set is setting, it just helps in search and clarifying questions

[rich widget]

There’s so many questions I ask that don’t get a straight answer from perplexity. The app usually makes the call on ethics and safety. Is this happening for anyone else? I need more than just a “basic” answer I already know

[agile jay]

Yep, I think the copilot/pro toggle just makes the model ask you questions before responding, to refine your prompt.

[warm cave]

It’s Opus for the first 5 messages, and then switches to sonnet, you will get your 5 back 24 h after you used it

[agile jay]

Seems a lot of people don't realise how LLM's work...

Yep, unless you know the hidden method to get infinite Opus...

[warm cave]

I forgot sometimes it’s not common knowledge

[agile jay]

Claude 3 seems a lot better at speaking than GPT4T

Less robotic, I would say.

[warm cave]

Nice having Claude 3, good addition to perplexity

[sweet jasper]

More extensive testing, do you find sonnet better at coding than opus?

[agile jay]

Yep, i think the current limit is because of anthropic.

[agile jay]

Opus is the GPT4 equivilent, so shouldn't be.

[sweet jasper]

I have a mixed feeling bag that sometimes sonnet outperforms opus

[agile jay]

You need to try the exact same prompts to both of them. Since it could just be bad prompts.

[sweet jasper]

I'm testing it at 0.7 temp , will decrease temp

[agile jay]

...

For code, you want low temp

Temp controls the amount of entropy.

Good for creative writing, bad for code.

[warm cave]

I notice that sometimes Sonnet is more verbose than Opus, not sure if it’s just a few examples that stood out.

[fluid mauve]

there is a link in the mail that rabbit sent you click on that and follow thw directions

[deep mesa]

Awesome thanks

[sweet jasper]

I'm aware, sonnet outputs are way more verbose

Opus tends to output just what you ask

No filler code so to speak

[agile jay]

Could be on purpose.

Since output costs a lot more for Opus

[fluid mauve]

if you get to checkout and it dosent say $0.00 to pay, then dont continue or you will get debited and have to reclain. if its more than zero email pplx support. they fixed min out in a few hours. i already had a monthly pro account.

[agile jay]

So would make sense to make it's outputs more concise.

[sweet jasper]

That is a good theory

[warm cave]

Thy makes sense

[agile jay]

Claude 3 should actually be better than GPT4, it's basically GPT4 created with more modern techniques.

Which is why it's so good at needle in the haystack and other similar problems that GPT4 isn't.

Then again GPT4 is almost 2 years older, so...

[warm cave]

So with sonnet being more verbose it would probably be a better option for creative writing and opus would be the programming/logic king

[agile jay]

I am guessing the sources are cached.

[agile jay]

Yep, you also have to remember that it is likely a system message telling it to be shorter or more concise. You could tell it to not be concise, by adding your own system messages in a collection.

Yep

There are two steps when you input your query.

[steady bolt]

Which AI language model is best for programming codes?

[agile jay]

They probably do a search in their indexer and get a list of search results. And then they pass that to the model you chose with your prompt, and a system message telling it what it's job is, which in this case is to respond to your question by using the sources.

The first part of getting the sources is likely cached.

[agile jay]

Claude Opus is pretty good from what I've seen.

Before that I used Mistral Large, since it would give me more verbose responses.

[sweet jasper]

Claude sonnet / opus and gpt4 turbo

[agile jay]

GPT4Turbo seems to have gotten lazier.

[sweet jasper]

Yeah with claude you don't have to wrestle the model

It's a nice change of air

[agile jay]

I would ask it to write the code, and literally all of it was just comments...

[sweet jasper]

At work we have github copilot but honestly it gets so many things wrong

[livid mantle]

yes i stick with claude 3 opus

https://tenor.com/view/cat-dont-care-didnt-ask-didnt-ask-i-didnt-ask-gif-25429803

[warm cave]

Here is what i found:

[agile jay]

Github copilot is good for code you've already written in your code base. It's bad when it has to come up with stuff from scratch.

[livid mantle]

opus became sentient

[sweet jasper]

I agree, code autocomplete is very good

But chat is kinda meh

Like 8k context

[warm cave]

lol, all the times the researchs have said that there ai became alive is funny

[agile jay]

There is a gradient of conciousness.

The question is where is current AI at.

[warm cave]

Yes, indeed, its the early stage where we think its smarter than it is, but it will not be long before its not

[agile jay]

Yep, especially when each model is 100X the compute of the last...

It is really quickly overtaking us.

Did you see the example where they did the needle in the haystack test, and claude pointed it out, and assumed it was being tested?

[warm cave]

I know it performs very well and is very competent at so many tasks, but it lacks a certain level of understanding, Where the AI can explain complex topic but cant recognize when it is being tricked into doing something it previously refused, maybe this is a training data blindspot, maybe we need a less focused data set, I dont think i can quantify it, but its amazing until its not

Some people say this is the limits of mode 1 thinking, where we have mode 1 and mode 2 thinking

[agile jay]

I don't think so. I just think the models are just too small.

The problem is just solved by scaling.

[warm cave]

They are really small, and that is a big part, but we are there feedback loop, they can not engage with a though over a year, i do think the biggest factor will be scale, get it up to 100T parameters and it will be a whole new beast, but i think there will be a few changes the token feedack loop, i guess we will see

[livid mantle]

why perplexity removed gemini

waiting for gemini ultra 1.5

[agile jay]

They said something about bugs etc

[livid mantle]

okay

[agile jay]

I think CL more or less solves that.

The problem with current models is that they are static after training.

You are basically talking to a clone, and each prompt in your conversation has likely been give to a different clone too.

And the model doesn't remember your conversation at all.

CL (Continuous Learning) more or less solves that by having the model learn while talking to you (and everyone else).

We probably still need a way to distribute and update all the models though.

However that is currently too compute intensive.

[small hemlock]

Is there a way to interact with perplexity using just my voice and getting a live audio answer. I do know I can activate audio using the mic icon but would like a live audio response. Thx

[agile jay]

Nope, as far as I know, they haven't got a voice to read the response to you, like 11 labs or the openai speech models.

Not sure if their model is called whisper, or if it's a different one. I think whisper it speech to text.

[warm cave]

CT dose seem like like a important part

[small hemlock]

Thanks for the info

[agile jay]

Np

[tame current]

Hello everyone

[grand sand]

https://www.perplexity.ai/search/If-i-had-YPptn3R5SWSUtFWL7hG2vw?s=m

Perplexity created a diagram

Didn't know it could do that

[gusty wagon]

I really like "plex it" that @blissful raptor has been tweeting about. Can we get plex.ai ??

I hate typing perplexity every time because it's long and complex, and when I tell friends about it, they say the name is confusing. plex.ai would be perfect. Its a better name.

[round plover]

Anyone know best model to use that also doesn't censor like ChatGPT?

GPT4 and Claude 3 censor pretty heavily

Default = GPT 3.5 right? modified somehow though?

also what is the experimental one

[brittle dagger]

Experimental is basicaly a fork of either LLaMa or Mistral 7B. And it's supposed to be less restrictive.

[grand sand]

Try the experimental model

[round plover]

thx guys

[gaunt crater]

Per https://blog.perplexity.ai/blog/introducing-pplx-online-llms: our PPLX models build on top of mistral-7b and llama2-70b base models.
Somehow it's both.

I'm guessing they have one for each size.

Is the one on the site the 7B? Given how quickly it streams I guess it's still possible it's 70B, but unless they just have it running on extreme hardware it feels too fast not to be the 7B.

[livid mantle]

my claude 3 opus ran out for the day for perplexity, i went and subscribed to calude 3 opus on anthropology.

[sleek vortex]

i feel like default is 7b sonar or something

and the “experimental” is probably mixtral based

because it’s still fast too

[sleek vortex]

well done

now in 2 weeks there’s gonna be some new model which beats it

🙏

[livid mantle]

maybe i will go back to gemini advanced ultra 1.5 when it comes out

[sleek vortex]

1.5 pro is promising idk

i just use api on all of them

don’t feel like paying $20 each

[sleek vortex]

well other than perplexity

[final sedge]

Hey any technologists here

[honest basin]

it's pretty ass don't get ur hopes up

[thick dock]

Hi

[thick dock]

I think it's worth it and I tell the creators that their chatgpts are pretty good

I'm buying it and I'm surprised that the pro has so many options

[sleek vortex]

context and video input is promising

actual model is yeah worse

[thick dock]

Yes

[honest basin]

it's probably not even real context it's probably just a summarization model

[thick dock]

But it really has potential as chatgpt

Everything has its time

[thick dock]

I tried Claude and even videos he recommended he gave me what I wanted, he can interpret and do almost anything.

Search IDS music many things I found it I make it easier

[vapid onyx]

Define technologists

[haughty lodge]

I think the reason 1.5 is so trash right now is because google forced the temp to be 2. At least 1.0 pro seems significantly smarter in logical reasoning when I set it's temp to 0 compared to the default 0.9

[final sedge]

if you have solved practical problems with existing tech(not just computer tech) or created a new technology

[vapid onyx]

Ah gotcha, like network architecture related items or more deeper?

[final sedge]

like people working on building perplexity

[final sedge]

.

[vapid onyx]

Ah not me. Sorry. Just a volunteer

[ripe void]

@signal hamlet arigato gozaimasu (^.^)

[smoky niche]

discord怎么使用

[heady quiver]

English Lol

[livid mantle]

sorry sir

[tame current]

y only 5 queries for claude?

[fossil sentinel]

I've been using so far

Been very impressed

The only sad part is it's gonna launch at 128k Tokens or sumn...

[hoary python]

it's expensive

pplx bugging fr:

asking it a code problem and it gives me the weather and treatement options

that's crazy

[foggy whale]

That's super interesting. I havent been using models in that way, but certainly that seems like the kind of framework I could follow if you're going to be populating a field of information. Do you have a background in accounting to be able to qualitatively evaluate the results?

[serene sand]

#💗│sharing message ☕

[native robin]

for the opus , is it 5 chats or 5 to and fro questions only?

[halcyon coral]

Five questions

[native robin]

then i guess gotta wait for them increasing the limit , 5 sounds too low

[fossil sentinel]

Definitely not trash when uploading a subject related textbook

Otherwise... results may vary

[small hemlock]

Sadly the cutoff of Claude is Aug 2023 so it can't access current info which is a real bummer

And perplexity's so called real time access leaves a lot to be desired. It misses a lot of stuff out. Lots of room for improvement

[outer relic]

you .com api has url citatiton ı think anyone experiance it

[hearty quartz]

Why is Claude 3 Opus limited to 5 times?
Is there an unlimited usage plan?

[halcyon coral]

#💬│general message

[grand sand]

Currently it's really expensive so they had to limit it

[wispy gulch]

I am ashamed. I can't get the photo generation to work. In the settings I found a choice of Playground v2.5, DALL-E 3, Stable Diffusion XL but I do not know how to start them.

[wraith crow]

Please have a look here #1194794305362071552 message

[wispy gulch]

Unfortunately I have read the thread but the buttons as on the screenshots do not show me. I have purchased Ai mainly for the feature of correcting my images with the elements I need to apply and also to test Cloude 3 Opus which also cannot be enabled ? Can you indicate directly how to use the generation and how to use cloude 3 opus ? Thank you in advance 🙂

[queen quiver]

Guys, tell me, what is Perplexity Discover? How it works?

[round plover]

click on interesting news, read, ask followup questions
I'm guessing someone writes the titles and AI researches/writes the story
So far I've been using it and enjoying it

[livid mantle]

theres a button below that says skip

[digital hare]

opening perplexity app and then searching for stuff takes higher load time then it does while doing the same using its widget, that's super fast

[livid aurora]

Bro Claude 3 got faster and it's fast af now 🙂

[lusty mantle]

Quick question what is copilot exactly?

[halcyon coral]

Please check https://blog.perplexity.ai/faq/what-is-copilot

[digital hare]

fast and easy 😁😁☝🏽

[grand hearth]

nice phone! 😄

[digital hare]

its nothing phone 1 😁

[serene sand]

deep dive ➤#💬│general message

please check tutorials in this thread

[fossil sentinel]

wait I wonder if gemini 1.5 would code more reliably if a relevant book/reference book for whatever language was uploaded to it

[modern crown]

I think someone else asked here but it was a good question but when using Cluade or GPT-4 do you get the full context if you upload files like Claudes 200k context?

[fossil sentinel]

Kinda relevant

As for full context might have to test it out yourself

[modern crown]

Yeah, I will have to test that

[lean fossil]

Hi, Im planning to swtich to perplexity pro, Is it better than chatgpt 3.5 and 4? Im a software dev, will be using it to boost productivity.

[fossil sentinel]

Gpt 4 is one of the options you can use on pro but doesn't have code interpreter if that's something you find important

[lean fossil]

I see. What's your experience wth perplexity pro? is it better or more or less the same as chatgpt?

[fossil sentinel]

Might be better to ask some of the people here who are on multiple paid plans for the different models 😂

[agile jay]

https://inflection.ai/inflection-2-5

Looks like it's GPT4 level release week...

[lean fossil]

yes, I tested the free one in detail, it took 7 -8 tries to get what I want vs 2 or 3 with chatgpt. Just wanted to confirm.

[agile jay]

Rip

[fossil sentinel]

🤔 world class huh

Bold statement to make

[spice shore]

What differences are there between using ChatGPT or Claude 3 directly through their provider, versus using the writing mode in Perplexity? Is it mainly the prompt engineering that happens before the user prompt? I have seen people suggest that using ChatGPT through OpenAI provides a better multi-turn conversational experience than using Writing mode with the model set to GPT.

[agile jay]

Pretty much. Could also be faster too.

Btw pretty sure pi2.5 is free too...

[zealous totem]

I mean it works fast

Don’t know any other reason you would use it for if you have other LLM’s avaliable like ChatGPT or Claude

[agile jay]

It has just under their performance for free...

[warm cave]

Its fast and the voices are good, except the first 2 sound like they are almost from elmo

[agile jay]

Lol

It's fast because the model is a lot more efficient.

2.5x more efficient

[warm cave]

Its like its trying to be an assistant, but also a friend, haha. I wonder how it will do at coding

[fossil sentinel]

I remember trying it out while back

[zealous totem]

That’s really good

[zealous totem]

Wait can you try it out now or is that release for later? (This)

[agile jay]

Already released I believe.

[zealous totem]

Dunno if the one talking to right now is the 2.5 one or earlier

[zealous totem]

Ah i see, that’s impressive speed then

[zealous totem]

Is this free
?

[agile jay]

Yep

pi.ai/talk

I believe

[fossil sentinel]

I feel like "competitive with" indicates it's not actually as good 😂

[zealous totem]

Oh that’s really good

[zealous totem]

Gemini….😭

[fossil sentinel]

Even Claude 3 say they "outperform" on benchmarks but that's just thanks to their prompting smh

[warm cave]

Interesting, they are taking a different approach, they have built in dating advice

[fossil sentinel]

💀

[zealous totem]

I really like their approach, where a personal ai instead of an assistant for work and stuff

[fossil sentinel]

Wait til the dating becomes digital partners...

[zealous totem]

Already has happened…

😭

[fossil sentinel]

[warm cave]

Will be insane, then sora + apple vision pro, we are doomed, haha

[warm cave]

I know its already happened but It will one day be normalized 💀

[fossil sentinel]

😭 this is what the peak of humanity brings us to

[zealous totem]

Kinda crazy that Sora was actually built a year ago (I think?), kinda scared what they have in store for now

[zealous totem]

Definitely

[warm cave]

Hmm, i like how its hard to see the percentage, like 95 ish

[fossil sentinel]

Have they even fixed the MMLU yet or is it still messed up

[zealous totem]

Actually I think AR is kind of cool but that’s just my opinion, the apple vision pro is useless for normal people at this point of time

[agile jay]

Yep, they have it in their technical report.

[warm cave]

Interesting

[agile jay]

Yep, and the efficiency of the model is why it's so fast.

[fossil sentinel]

Lol I tried out a simple statistics question and failed

For now I'm just gonna think of pi.ai as a conversation ai still

[sleek tapir]

Is there a way to use perplexity to make a news feed of a particular topic

Like the discovery feed, but for local news. Or video games, or movies, or whatever

[agile jay]

Yep, if you look at the graph, that's where the biggest gap is compared to GPT4.

#💬│general message

[warm cave]

It got this calc problem wrong and wont be my GF, haha

[fossil sentinel]

💀

Bring back OG Sydney bing

[warm cave]

but for most people i think this AI would be plenty, it looks to be solid but not for power users

[fossil sentinel]

I think Claude sonnet would provide more value though? Also free

[warm cave]

Yeah, I just like to see new AI options, but yeah I think claude would be better, but it would be more than enough for things like this

It seems like a simple AI, not bad, pretty good, the only thing about it that is really different is its fast, but i think sonnet is too, and its super friendly, which I guess could be a selling point, haha. just not really worth it compared to all of the great options, Claude Sonnet, Mistral Large (Not Free but pretty cheap), Gemini, GPT-3.5

Oh, the voice it good tho, so if you want an AI to explain something out loud, this is good, i think you have to pay on chatGPT to get voice, and google might have voice, but i think its was just standard tts, sounds robotic

[agile jay]

Voice is free on the ChatGPT app.

[warm cave]

Thats good. I never use the app, but I guess gpt-3.5 with voice prob is solid for most

[humble swan]

I am disappointed with gemini tbh

[warm cave]

looks like it did better then mistral, a bit surprised

[humble swan]

I tried it and I am fascinated of how my experience with it is much worse than with perplexity pro

[warm cave]

Yeah, I hope 1.5 makes it worth it

[humble swan]

I Think they rushed it

I think they felt like they had to release it asap or something

because it's quite unfinished

they even had to include free cloud storage in order to compensate, or at least this is what it feels to me

I think the whole vision they have with gemini is making it a whole ecosystem, that's what I assume they want to make, what would be better than the competitors

but as of now they do not have the ecosystem, 5 extensions only

[humble swan]

This is Claude right?

[hoary python]

You'd have to use the API and write some logic to query on a interval

[fossil sentinel]

I just treat it like Google search

[humble swan]

The results were disappointing especially simple math explanations

Also no profiles, libraries, collections

The only nice feature is drafts

Also for me personally it took up to 4 seconds for simple prompts such as "explain asymptotes of functions"

But maybe my use case is wrong, maybe it's aimed at making stuff rather than explaining math concepts, like creating essays, texts and so on

[warm cave]

No, pi

[mild citrus]

Hello Perplexity folks, which channel is appropriate for discussing sonar and the online API?

[humble swan]

the font looks really cool

[warm cave]

prob pplx api genral

[halcyon coral]

Hi, please use https://discord.com/channels/1047197230748151888/1161802929053909012

[warm cave]

Yeah, looks nice

[mild citrus]

thanks much

[humble swan]

also that %, love it

[zealous totem]

is it worth upgrading to nothing phone 2a

[subtle fjord]

I reserved my R2!!

[livid mantle]

[agile jay]

How many messages have you been sending? Lol

Or are you on the free tier?

[livid mantle]

opus

i used up 100 since yesterday

[fossil sentinel]

Did gemini 1.5 always have a pop up message for file uploads saying only text was converted and no images?

Or am I tripping

[raven tree]

is everyone just using claude 3 now? how is gpt 4 turbo vs claude 3 sonnet

[fossil sentinel]

From my personal tests opus is not as accurate as gpt 4

[sweet jasper]

Claude 3 sonnet sometimes better than opus

[raven tree]

opus was more verbose/helpful? for coding questions, but yea hard to say with only 5 uses a day 😄

im testing sonnet over 4 turbo as a daily driver, jury is still out

[sweet jasper]

Depends on the prompt really, sonnet was able to figure out a book by ballard, while opus mentioned said book was by stephen king

Chat.lmsys has a nice side by side view when testing prompts

Wish perplexity had it to compare model answers

[fossil sentinel]

Yeah that's how I was testing em out

Seems like I need to try out sonnet more

I just assumed opus would be better in all cases

[azure crown]

just subbed to pro

[raven tree]

opus results are more "human" and definitely parses your prompt better

[vapid onyx]

welcome

[digital hare]

got my perplexity premium 😁😁☝🏽☝🏽 enjoying it

what does pro discord mean lmfao

[wraith crow]

Pro support will give you access to live chat with the support of perplexity when they're online. And Pro Discord gives you access to the pro channels when you rejoin the server using that invite.

[digital hare]

i didnt got it ????

how to get pro channels

pls give

[wraith crow]

#👑│pro-lounge

[digital hare]

noway i have this lmao i didnt saw

thanks

[azure crown]

cool

[final sedge]

are everybody here only users of products?

[worn monolith]

I have other roles in my life...like eater of products

[livid mantle]

i have 100 messages available for claude opus 3 but idk what to do now

[scenic grail]

Anyone get trolled by perplexity before? I keep asking it to elaborate on somethinf specific but it keeps answering it in vague ways to avoid answering me completely. And no its not at all nsfw or anything bad

[livid mantle]

why perplexity can only do 3 messages for claude 3 opus

[scenic grail]

It keeps saying the same thing but in different ways. It refuses to get any more specific about the location

"Located in the realm where the events take place" lmfao ok

[livid mantle]

jerr bezos is here

[royal niche]

You might have to change the default AI to something else sometimes. Some of the models have a tendency to do that. The small Gemini one had a really bad problem repeating itself or this type of behavior.

You can fix it by switching the default to GPT, asking the question, then switching it back. Or you can be more specific with your prompting. Giving it more context tends to stabilize the answers. This is true with all LLMs.

[scenic grail]

I didnt post the others but i asked that question in three different ways and got the same response

[fossil sentinel]

I was curious about tencent's stake in discord so I googled... this was the first article

Uses chatgpt

[livid mantle]

Gemini Advanced Ultra 1.5 vs ChatGPT-5 vs Claude 3 Opus vs Copilot Pro (Github GPT-5 Turbo) vs Grok-1.5

Who will win?

[feral hazel]

what's the difference between using GPT4 on OpenAI website versus GPT4 on perplexity AI? Is there a noticeable difference between the answers?

[fossil sentinel]

Gpt 5

[cyan creek]

Whatever is the best model on Perplexity.

[livid mantle]

we need grok in perplexity!

[cyan creek]

I've heard so little about Grok, no benchmarks come to mind, I've only heard of one person using it recently.

[livid mantle]

according to hallucinated statistics, it performs better than chatgpt 3.5

[shadow saffron]

It's only super useful for twitter real time search rn, and for that you gotta be a blue checky

[cyan creek]

Generally 3.5 hallucinated a ton for me.

[livid mantle]

wtf is this groq that has been trending very recently

[shadow saffron]

It's INSANE how that was only a year ago

[livid mantle]

its so overwhelming to think of the fact that GPT-4 update is now lazy mode

sorry bill gates but i disputed my subscription

sorry sam almatn if ur in this discord too

[cyan creek]

I find myself not necessarily needing the best models, I don't code, and any medical advice is rejected anyways.

[gaunt crater]

It's incredibly unremarkable. Has the feelings of a mid AI that is very X/Twitter centric, doesn't keep threads, doesn't seem bad but not worth using either. But it's also kind of inoffensive, I guess. Just not a contender.

That was me trying it only briefly maybe 6+ months ago, so it may have gotten an upgrade since, or I just didn't feel inspired to give it more time.

[tame current]

Beating 3.5 doesn’t cut it anymore tbh

Even Open Source AI beat 3.5 pretty frequently now

Having Grok would be pointless when we can use almost every major AI model to date

[cyan creek]

for me I'd like a model that was less censored, follow instructions better, and was snappy tier of fast.

[minor shard]

I do not trust any information coming from Grok

[stark python]

So Perplexity give us up to 5 times for Opus, How do we use it? For now all responses are sonnet.

[fossil sentinel]

Based

[oblique pier]

Do I need to press the Pro button when I want to use Claude 3 Opus?

[cyan creek]

no that's for what was called "copilot" it just understands your question and doesn't change claude 3 models. you get 5 opus responses total, then change to sonet.

[oblique pier]

Is there any way to test the authenticity of Claude 3 Opus?

[halcyon coral]

In general, you cannot check the models 100% for authenticity with any provider, but for the reasons mentioned in this thread, you would inevitably realize it at some point: #1198438847877480629
But don't worry, Perplexity always uses the real models.

[dense field]

i am a developer, but use the products to discover some of the code i write

[heady quiver]

Can we manually switch to Claude 2.1 please? For Certain tasks that was better than the Opus/Sonet

[oblique pier]

Why I use Claude 3 but the Q&A is Perplexity?

[sweet jasper]

System prompt

[halcyon coral]

I previously mentioned the thread 🙂

[buoyant shard]

claude 3 available in perplexity now?

[knotty frigate]

When we give an AI prompt in a thread, does it modify the system prompt or just add some extra text as pre-prompt?

[sweet jasper]

My guess is it appends to the system prompt

[granite grail]

extra, pretty sure

[halcyon coral]

The prompt is System Prompt (from Perplexity) + AI Profile

[halcyon coral]

Claude 3 is available for all users

[serene sand]

#💗│sharing message ☕

[distant juniper]

Anyone know what model and what the actual context window is for perplexity? On my tests they don’t have the 100k context window the gpt 4 0125 model has.

[shy grove]

please help!

I have a serious bug disturbing me

whenever I try to use the rewrite function and use probably mistral large or gpt 4, it doesnt work. instead it brings out this prompt telling me to subscribe to perplexity pro even when i use a Perplexity pro account already.

[halcyon coral]

Hi, I assume you just recently subscribed to Pro?

[shy grove]

no

[halcyon coral]

The problem should be easily solved by deleting all browser data, cookies and so on. Then log in to Perplexity again with your account and everything should work.

[shy grove]

i am a regular

[shy grove]

let me try this thanks

[shy grove]

no this doesn't solve the problem. i have perplexity on my IOS and on my android. i even signed in on another browser to try and it is still give the same bug.

[halcyon coral]

Then please contact [e-mail]

[shy grove]

alright then

[dense bear]

I'm a little bit disappointed that we get just 5 claude 3 opus requests per day

[halcyon coral]

#💬│general message

[shy grove]

@halcyon coral Hey!

no reply from support, no response from the ticket submitted and no response from the messages i sent

[shy grove]

this issue still persist

[halcyon coral]

Please be patient, the response time is several business days, if you write several messages it just takes longer, thank you!

[placid skiff]

Guys I am wondering perplixity is 20 a month with 600 messages, chat gpt is the same but with 40, is there a catch?

Is anyone here?

[halcyon coral]

Hey 👋

[placid skiff]

Hi👋

[halcyon coral]

There is no catch, you can see all the advantages of Pro here: https://www.perplexity.ai/hub/faq/what-is-perplexity-pro
Now Claude 3 instead of Claude 2.1 and a total of three different image generation models.

[placid skiff]

And also, Just wondering why does chatgpt have a higher user base if perplexity offers much better service and messages with the model?

[placid skiff]

Thanks!

[halcyon coral]

Perplexity is not so well known yet, most people know ChatGPT because the well-known models like GPT 4 were developed by OpenAI 😉

[placid skiff]

So was GPT4 in perplexity released recently? And also which would you suggest for a student. Appreciate the answers.

[fossil sentinel]

Gpt 4 has always been part of perplexity as far as I remember lol

And way better for students imo

[placid skiff]

So perplexity is better for students than chatgpt plus , is that what you re saying.

[fossil sentinel]

Imo yes

[halcyon coral]

GPT-4 has been available in Perplexity for some time, and it was recently upgraded to GPT-4 Turbo.
I would recommend Perplexity to any student, as a student you are dependent on up-to-date factually correct data. ChatGPT is usually unable to provide this.

[placid skiff]

Very True. Thanks guys im mostly upgrading to Pro instead of Plus just needed some research.

[fossil sentinel]

Goated research tool

[placid skiff]

Also smart? Like answers hard questions or writes essays like chatgpt with the Pro

Thanks guys, Hyped to access the power Perplexity can bring, Hope it flies in sucess.

[shy grove]

is there a way one can increase the daily limit of 600?

[halcyon coral]

If you decide to try Perplexity you can get the first month for only $10 instead of $20 with this link 😉
https://perplexity.ai/pro?referral_code=ST5556J2

[placid skiff]

600 is already a lot

[halcyon coral]

600 uses should be more than enough

[shy grove]

not if you are conducting massive research papers

hehehe

[placid skiff]

What, How thats amazing.

[placid skiff]

It it legit

[halcyon coral]

sure

[placid skiff]

is it legit?

[fossil sentinel]

😂 damn needing over 600 is crazy

[placid skiff]

fr

[placid skiff]

Do u have pro?

[fossil sentinel]

I did but not anymore

[placid skiff]

oh

[placid skiff]

didnt like?

[fossil sentinel]

No I just don't have a need at the moment for pro

[placid skiff]

Alright, if you wanted to would you rather get Gemini Advanced, Chatgpt plus or teams or perplexity?

[fossil sentinel]

Perplexity for sure

Just wish it had code interpreter

[placid skiff]

oh

doesnt gpt 4 in it have it?

[shy grove]

hehehe

[warm cave]

Me too

[serene sand]

deep dive ➤#💬│general message

[dense bear]

I think 200$ per year overall is more then fair in terms of what you get. Perplexity has a value, for searching, even tho it’s not the best search engine. But if you need answers that are backed by sources, you need perplexity.

[vapid onyx]

I think the source gathering that perplexity does is really great and helps reinforce your result entirely

[late oracle]

Yes.

[sweet jasper]

At what time do opus queries reset

Is it dynamic

Aka just like the 600 used

[cyan creek]

if you sent your first response at 1:00PM, it would reset at 1:00PM the next day.

[dense bear]

I have access to opus, but no in Claude.ai, it depends on the task, it’s a bit better then gpt-4T

[round plover]

perplexity so much better than chatgpt, i hate openai so much

[devout bronze]

without OpenAI, Perplexity wouldn’t be able to access GPT4

and the development of LLMs as a whole would be much slower

therefore we wouldn’t have even seen most of the incredible AI advancements

[round plover]

OpenAI still needs to be competitive and ethical IMO.

(they aren't, which is why I don't support them)

[autumn birch]

Hey guys, I am actually curious how does Perplexity works? I mean it uses Bing APIs to search the web and then build a RAG???

[round plover]

I'm not sure if it's Bing. But I can say that their Pro Search also looks at Reddit and more...

[autumn birch]

It is Bing trust me. I saw the query responses on both platforms and it’s definitely Bing. Nothing bad even ChatGPT uses it.

But I am curious does they build a RAG for answering questions or not?

Because building a RAG will get better responses but really really slowly

[round plover]

Idk what a RAG is

[autumn birch]

Retrieval Augmented Generation, it’s a whole different beast but I am still learning about it.

[round plover]

They don't seem to say they do...

[sleek vortex]

well im sure theyre using some form of RAG with their search results already

it cant in time:

• search 25 sources
• feed all into ai
• output

all in like 10s

searching is prob done in parallel and cached

most of the 25 sources are prob from a cached database, using RAG to find relevant ones

idk

but then yeah its using kinda rag on the pages?

bc its giving it snippets of each page

[sleek tapir]

caching the internet is probably the first step

[violet venture]

https://www.youtube.com/watch?v=3NN5L-f0cDo
What do you think?

[violet venture]

Define "ethical" in the context of AI, except not trying to make AI of course.

[agile jay]

A likely scenario, except that by then gen AI quality is likely to be higher than the actual internet...

[round plover]

I dunno. It's more of a general thing, I guess. How they collect so much data, and the new plan that doesn't collect data requires you to get at least 1 other person to subscribe. And censoring.

[violet venture]

Yah, that's what I thought, although I can't quite grasp what "higher than internet" would look like.

[agile jay]

The Internet has high and low quality data. The difference is like a random YouTube video vs a Netflix video.

[violet venture]

Censoring is now unethical? Not the threat of Skynet? Collecting data is happening... everywhere. We're beyond getting used to that.

[agile jay]

What happens when all YouTube videos look like AAA movies.

With CGI better than current movies etc.

[round plover]

Censoring always has been unethical

[agile jay]

There is also ethical censorship.

[round plover]

Idk what ur point is, I'd rather have less data collected than more

[violet venture]

Ok, so there's a threat to creative content, but we've been through that before (video killed the radio store). In regards of quality, that's highly subjective.

[agile jay]

The problem is when you are only censoring opinions you disagree with.

Since you end up in an echo chamber, and don't challenge your ideas.

[violet venture]

Okay, don't expect AI builders to set the standard. You'd have to ask NSA to do the first steps :>

[round plover]

Wym?

[agile jay]

Yep, the difference is likely to be that at one point, the human is only the client.

And not the producer.

[violet venture]

I see this as pandering to politicians, we all know how to trick the AI into giving an answer anyway, right?

[violet venture]

I agree, but what's different than it is now? Supply and demand?

[agile jay]

As the agents get more intelligent, jailbreaks become more and more difficult.

[round plover]

I'm making my vote by financially supporting Perplexity and spreading the word

[agile jay]

It's why I say the safest AI is the most intelligent.

Since it can actually follow your rules.

[violet venture]

Perplexity subscribes to GPT, hence you indirectly support OpenAI.

[round plover]

I'm paying $20/month for perplexity. CHATGPT is only getting a portion of that...

[violet venture]

True. I agree. There is reason for concern, I'm just too old for doomsday scenario speculation.

[round plover]

Also I'm not supporting OpenAI with my data now

[violet venture]

You're right, you're only fractionally a hypocrite 😉

[round plover]

How?

[agile jay]

Yep, imo the main danger is from making a tool instead of an agent.

[obsidian acorn]

Perplexity probably spends a crap ton to openai

[round plover]

My problem was with overpaying and data collection. I'm fully avoiding that now?

[violet venture]

If you stand by your values, you'd refrain from touching anything OpenAI/ChatGPT is involved in.

[round plover]

Why?

[obsidian acorn]

I tried using the gpt4 api and just me alone in less than a month it was 15 dollars

[violet venture]

I don't care, actually... never mind me.

[round plover]

Cause you're wrong that's why

[violet venture]

pff ^^

[round plover]

Or better said, you're fractional correct

[agile jay]

Well in the future if you want a good experience, there needs to be some sort of data collection.

Or you don't want the AI to know the past context of your conversations.

[violet venture]

Why not live feed? Why does there have to be a cache?

[obsidian acorn]

That's how all these language models work

[agile jay]

Because a cache is far cheaper, hardware and response time wise.

[violet venture]

Counter-intuitive, the web is huge.

But so far you've earned my trust, I believe you.

[agile jay]

Not really, if you look at what actually happens when you make a request.

It's mostly key value pair look ups.

url => DNS => html webpage

Not very compute expensive.

A search engine just adds embeddings and SEO as an extra step, for finding relevant urls.

[violet venture]

Yeah, I think I see your point, but I actually don't know how search engines work. I used to work with SEO (15 years ago), everything has changed since.

[agile jay]

The problem with current LLM's is that all compute is the same.

[violet venture]

I mean, I understand the fundamentals of web crawling (spiders), just not in detail how it's processed.

[agile jay]

You want to do 1+1 or differential equations, it uses the same compute...

[violet venture]

Hmm, is that a problem for LLM or AGI, you mean?

[round plover]

You listened to the Yann podcast with Lex? :)

[agile jay]

LLM's

[violet venture]

Quantum computing will change that game, right?

[agile jay]

No, that's not how quantum computers work.

What they will be useful for is for the AGI to make predictions and test them in an emulated environment.

[violet venture]

Probability curves?

[agile jay]

So it can expand its knowledge past the training data.

[violet venture]

Yeah, that's what I meant, AGI-wise.

Thanks for the short explanations, sir.

[agile jay]

A few years away from AGI, so it looks like fun times ahead.

[violet venture]

Yes, but the preliminary steps matter...

[agile jay]

Of course that depends on the definition you use, lol.

[violet venture]

This

[agile jay]

Yep, the current problem models have is that they don't get to test their output.

[violet venture]

Needs user feedback to verify, you mean?

[agile jay]

So they probably create a lot of unneeded structures early in training, which makes them less efficient.

[violet venture]

Well, babies don't walk perfectly when they begin exploring legwork 😄

[agile jay]

Lol, yep. But what if you can make them learn faster?

[violet venture]

Hmm, speed is relative.

[agile jay]

Yes and no. In isolation that's true. But in an environment you are in competition.

[violet venture]

I mean, you'd have to use quantum computing to "cheat" the learning curve, right?

[agile jay]

Yep, but it's less cheating, and more letting the model play.

Like children do.

[violet venture]

Yeah, hence the citation marks.

[agile jay]

Basically what we do, which is exploration.

[violet venture]

Also, as I understand, quantum computing can test several models in parallel instead of serial.

[agile jay]

That largely depends on the architecture.

[violet venture]

I mean, that's the goal, also the challenge in actually achieving quantum computing.

I see... well, I don't, but that's where I lack knowledge and will leave it at that ^^

[agile jay]

Yep, well having low decoherence.

[violet venture]

Decoherence... lovely word.

[agile jay]

Yep, states are very volatile at those scales

[violet venture]

Cognitive dissonance is a human thing.

I mean, if we want it to think like us? 😄

I guess we don't...

I wonder if there's a future in AI psychology.

[feral hazel]

Can a user select when to use Claude 3 opus, since there's only 5 times it can be used per day?

[devout geyser]

they can't

[clear finch]

Hello, is it normal to get very different results from the API & website? Using the API it hallucinates a lot, vs using the website it returns results that would be expected. I've tried online small & medium models

[agile jay]

Which website? Perplexity?
If so, were you using search? Since search makes it hallucinate less.

[clear finch]

Yes using search. How can I get it to use search through the online models in the API?

I've tried doing Search the internet for the answer using "${website}" as the primary source and it still comes back with hallucinations half the time..

[agile jay]

What model were you making a request to?

[clear finch]

I tried using sonar-small-online and sonar-medium-online.

[agile jay]

Were you trying it on: https://labs.perplexity.ai/

[clear finch]

Nope through the API requests. Thanks for trying to help me btw, much appreciated 🙂

https://docs.perplexity.ai/reference/post_chat_completions

[tranquil orchid]

Hey @clear finch!

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

[agile jay]

Might just be because smaller models tend to be more prone to hallucinations.

[dapper spade]

i was logging into perplexity tonight and stumbled onto a new tab under Library labeled "Article" which took me here https://perplexity.ai/article/new.

This only happened once, but I found it interesting o.o

[agile jay]

https://www.perplexity.ai/hub/blog/introducing-pplx-online-llms

This has info on how the online models work.

[sweet jasper]

@agile jay do you think we can achieve agi with classical computers?

[agile jay]

Yep

We are already not far off, just need to scale

At most, you use new chips, like photonics, to make quicker and more efficient hardware.

[coarse venture]

claude 3 opis is only like 5 for perplexity pro?

to my knowledge perplexity pro is the only service that offers internet search with claude 3 opis right now though right?

[agile jay]

Wheres GPT4Turbo?

[agile jay]

you.com

[coarse venture]

ohh nice, i'll check it out thx dude

[agile jay]

I believe they also don't have a low rate limit too.

For Claude.

[coarse venture]

what one gives better searches though?

[agile jay]

Try the free version of both, and see.

[coarse venture]

Perplexity i have always been a bit disappointed on how limited it's searches are, and have been using phind pro because i have been getting better answers