💬│general

[warm cave]

Could you explain what you mean about security?

[half venture]

Eh think about prompt injections

In longer contexts

They become more and more common

And attacks become more effective

Can cause the server to crash

Like with claude opus few days ago

Also scraping bots

[mossy fox]

So i guess 32k is the magical sweet spot

[warm cave]

But prompt injection dose not really affect the security of perplexity, they are api calls, they do not have agent ability for preform real world actions

Are you talking about flooding the server with api calls with large context

To crash it

[half venture]

Yes and no

Even with api calls

You can inject malicious code

That can cause problems

And yes flooding too

[warm cave]

Maybe I am behind on some of the latest techniques, but I am unaware of any methods that can be used against the api servers, is it internal interactions of the transformers architecture running on there compute? Or something else

[mossy fox]

I'm wondering aswell

[cinder comet]

With Enterprise Pro, our most robust offering yet, we add on even more functionality and features:

• Increased Data Privacy: Your data stays yours, period. We never train our LLMs on our enterprise customers’ data.

i think this should be the standard option for the pro version

[tame current]

Yeah. Everytime you add a message on, the entire message is reread, increasing the input tokens.

[cinder comet]

did they increase opus limit?

[warm cave]

Don't think so, still 50

[warm cave]

Use this to check: https://www.perplexity.ai/p/api/v1/user/settings

[cinder comet]

"opus_limit":41

[warm cave]

Yep, still 50 😦

[cinder comet]

NOOOOOOOOO

WHYYYYYYYY

😦

[mossy fox]

why aren't the labs capped then to 32k

[warm cave]

No idea, guess they forgot?

[mossy fox]

i smell something fishy

https://tenor.com/view/sigh-disappointed-wow-not-funny-womp-womp-gif-5209485905233272018

[warm cave]

It is very odd

One possibility is there code is a mess and they decided its too had to make the change now, but idk

Use to be no models over 32k

Or money 💵

But then why do they give free labs models with more than 32k idk

[mossy fox]

Imagine they are reading this and decide to kill off the context window lmao

in labs

D:

[warm cave]

Lol, yeah

[mossy fox]

"they started asking questions"

hahaha

[warm cave]

They know too much

[mossy fox]

cries in haiku

[inland bear]

Just to update the information about You.com (as some members expressed interest):

I had said that, when loading a PDF and asking questions in sequence, the answers from the 2nd one onwards no longer observed the document but rather data on the internet. Just now, I saw a comment on the You.com Discord saying that if you have the "Private Mode" option enabled, it erases all context after answering. I turned off the mode and it really continued to answer based on the PDF document normally.

An user just have posted these informations too:

Claude 3 Opus + You.com AI

[feral marlin]

output is limited around 2500 tokens i believe but i've tested with 150k or so input and it worked well other than sometimes producing errors

[inland bear]

I had already noticed that the answers there are shorter (I mentioned this in previous messages). Now, whether the models really use the full potential of the context window, I have no way of knowing.

[fluid mauve]

anyone watching the Rabbit R1 unboxing link?

[warm cave]

Is it right now?

[fluid mauve]

29 secs

1am GMT

[warm cave]

Thanks!

Came in clutch haha

Wish live streams had 2x speed

[fluid mauve]

yeah

this intro is killing me

[warm cave]

Yep, lol

Live unboxing, back stage before the cameras turn on, haha

[fluid mauve]

yeah, i had a chuckle at that

[warm cave]

If you got it, what batch are you?

[fluid mauve]

batch 6, june july time

[warm cave]

Same

Some perplexity back-end search

[fluid mauve]

looks like it wonder if hes using clause opus 😛

Claude Opus even

[warm cave]

Had is not use it the day before to make sure they had enough opus for the live demonstration 😆

[fluid mauve]

theres a lot of men in the crowd, not mant women, geeks are men

[warm cave]

Guilty 😅

It's kinda like a screen is useful, haha

[fluid mauve]

yeah. the keyboard is a nice feature for when you cant talk.

wonder if i can pair my mobile bluetooth keyboard

he didnt mention bluetooth

[warm cave]

Yeah

Oh lol

Yeah he did

[fluid mauve]

nice, bluetooth kb confirmwed

[warm cave]

Oh, lets see about translate

Most I have tried are too bad and are not good enough

[fluid mauve]

yeah, i use google translate to translate between languages, but never on the phone

[warm cave]

The handoff back and forth was always too confusing, maybe I was doing it wrong

[fluid mauve]

yeah. that was good translation

nice. surupticously recording meetings as no one knows what an r1 is

i suppose the twin tape reals are a giveaway

[warm cave]

It's my game boy 😅

The speed is impressive if it's the same speed we will get, humane pin takes ages

Just messed with an uber driver

[agile jay]

What event is this?

[warm cave]

Rabbit R1 live stream, watching rn

[agile jay]

Oh, so the useless product...

If your phone can already do the same thing, it's pretty worthless...

Since people aren't gonna carry two seperate devices around...

[warm cave]

Yeah, it's not going to beat out the phones, but I got one with my year of perplexity, so I'm going to have a little perplexity walkie-talkie toy

[agile jay]

Lol, the main problem with those kinds of devices is that they are voice activated.

Imagine being out in public and just saying out commands...

[warm cave]

Has a keyboard and screen like phone, but if we are being fr, just use a phone

[agile jay]

Yep, I can't imagine using voice commands in public to send a msg to someone else etc.

And I'm assume the R1 just uses API's

[warm cave]

Prob going to get a stand for it and have it and put it in my desk as a Alexa

[warm cave]

They say they don't, but idk

[agile jay]

... How expensive was the R1?

[warm cave]

Because they don't door dash barely works

$199

[agile jay]

And what hardware do you need to use even for weak models.

[warm cave]

100% clouds I expect

Wifi and cellular

[agile jay]

100%, at most the model that transcribes what you say might be local.

The rest is 99% cloud based.

[warm cave]

Yeah

[agile jay]

How much storage does the R1 even have?

[warm cave]

I never believed in the vision of Rabbot R1, and had no plans to buy it, but then as I went to buy a year of perplexity I saw that I could just buy an R1 and get 1 year. So now I get at worst a retro paperweight.

[warm cave]

Again, it's cloud, you have a web portal where it will store all of that

[agile jay]

Nice, free R1

Buy a few thousand more, and you can build a wall

I wonder how much the R1 costs to make.

If they are selling it for $200

And do you need a sub to use it?

[warm cave]

They say no subscriptions but you know how how that goes, haha

[agile jay]

Yep, if they need an API, the basically need a sub, unless they expect their users to not actually use it after the first week or so, lol.

[warm cave]

Yeah, I believe it when they say that they don't use API, if they did they would not have to ask it 3 times to order door dash, the subscription part is the Large Action Models, and the GPT 4 API

Btw, I mean no api outside of the LLM

[agile jay]

Wouldn't the R1 need enough storage and memory to run the model?

And a GPU

[warm cave]

The LAM is run on a virtual machine, and is kinda like open interpreter

[agile jay]

Must be a tiny model then...

[warm cave]

It's run on there servers afaik, thus they will add a subscription or make it so you can run it on your device. I don't believe for a second that they will run a virtual machine forever for you just because you bought the device, haha

[agile jay]

Hopefully they implemented OTA updates.

[warm cave]

That's what they promise

I'm pretty sure the search functionality is just perplexity, and the voice is by eleven labs

[agile jay]

As small models becomes more powerful, they might actually be able to run it locally.

[warm cave]

Yeah, like phi3

Or phi4 llama4 future models that are small

[feral marlin]

rabbit r1 seems pointless

using the internet but awful

[agile jay]

It pretty much is

Just using the AI hype to sell units.

@warm cave Did you try phi3 in openweb UI?

Since it's a small model?

[warm cave]

Yes, was far more coherent than other models like tiny llama and phi 2

[agile jay]

Are they releasing a few larger models too?

[warm cave]

Had decent logic, got the book problem right (tho it was prob trained on that one)

[warm cave]

It's called phi3 mini, so I expect that means they have larger versions

Maybe still in training

[agile jay]

Yep, wonder how their slightly larger model will compare to llama 3 8B

And I think they already have a large context too

128K

[warm cave]

Yeah, hopefully, it's good, one of the main things I noticed about phi3 is that it was the first 4b model that actually followed my instructions. Instead of mostly ignoring it

I wonder how many tokens it was trained on

[agile jay]

Lol, imagine the inference speeds with Groq

[warm cave]

Insanely fast

[agile jay]

Almost hald the size of llama 3 8B

which has around 800/ts

So maybe close to 1500/ts

[fluid mauve]

Seriously fast

[warm cave]

would be really great for things like predictive autocomplete, like code and normal writing

[agile jay]

Yep, especially with some finetuning

[warm cave]

Yeah, and easy to fix e tune with it being so small

[fluid mauve]

its a sushi travalator

[agile jay]

Yep, especially with a dynamically managed context.

Downside to open webui is that you need do clone the source

[fluid mauve]

right its bedtime methiunks. 02:14

[agile jay]

I guess that's what they get since it's mostly a TS project.

[agile jay]

Only 2:15

[fluid mauve]

i have a 5 yearold that will be up in about 3 - 4 hours

[agile jay]

Rip children. Just get an AI nanny, lol

[fluid mauve]

haha, my wife would kill me

[agile jay]

Night then

[warm cave]

Yeah, if you know a better option lmk, so far it's the best one I have found

[agile jay]

After I have mine mostly done, I'll integrate ollama into it for fun.

Does ollama rely on docker?

[feral marlin]

current music on rabbit is so bad

😭

[agile jay]

Will probably make my own integration on wasm, which should be a lot faster to respond.

[agile jay]

Rabbits can't sing...

[warm cave]

You don't need docker

[warm cave]

Bad speakers?

[feral marlin]

guy playing the music has no idea what he is doing

first ever beat

[warm cave]

If I remember correctly ollama is 100% or 99% Go

[feral marlin]

overlaying various sounds at different bpms that are all out of tune

[agile jay]

Oh, I am told that you need docker for ollama to work when doing a search.

[agile jay]

Nice

So is docker

[warm cave]

They had a docker image, but I never used it

[agile jay]

You're using the API right?

[warm cave]

For ollama?

[agile jay]

No, for the webgui

[warm cave]

Oh, yeah half and half, it has my local ollama models and my groq models

[agile jay]

Since I'm guessing the webgui only uses ollama/docker if you are running it locally.

[warm cave]

I am using docker for the webui, but I think you can use it without it

But have not tried

[agile jay]

Yep, I was talking more about the actual models, when deployed by ollama

[warm cave]

Yeah deploying models is done without docker

[warm cave]

So you are planning ollama integration for fun, or so my something like that?

[agile jay]

Yep, or my own version of ollama, to wrap around the models.

And to run them in wasm/wasi instead of a docker container.

That should also let me use the python libaries after compiling them to wasm.

[warm cave]

The joy of open source

[agile jay]

In my go code

[warm cave]

Nice

[agile jay]

But it will also mean that you could run the models locally, from a website.

No need to install something else.

[warm cave]

That's nice

[agile jay]

Yep, especially as the small models get more powerful

[warm cave]

Yeah, maybe one day you could use 2b local for the free tier

[agile jay]

Yep, or even other stuff like embeddings without needing to hit the API.

So completely private.

[warm cave]

I want to know what would happen if you were to pump a 1b model with 100 trillion tokens

[agile jay]

Pretty sure my pipeline will be python => cython => wasm

[agile jay]

That would be interesting

[warm cave]

I like that

[agile jay]

How long until it is overly saturated

[warm cave]

Idk, he said 70b had not show signs of saturation, but I'd about the 8b

[agile jay]

Yep, saturation could also be affected by the data quality too

[warm cave]

Yeah

[agile jay]

Well the other good thing about small models, is you can train them pretty quick to test stuff.

[warm cave]

Yeah, with phi3 it makes me want to look into fine tuning, fine-tuning larger models always seemed to expensive

And take too much time

[agile jay]

Yep, and you probably can't do CL in real time.

I'll probably try llama 70 with CL for a while, to see how much better it gets in my use cases.

[warm cave]

Yeah, and you have a good amount of local compute

Or you could do cloud

[agile jay]

Yep, I just wanna test it first to see if it's worth it.

So then I could use a lora for each paid user, or something, so it gets better the more they use it.

[warm cave]

That would be good. And when it's in a usable state you could setup RHLH (if I remember the acronym) with thumbs up and down, and we could help you build a training data set if that sounds helpful

[agile jay]

RLHF

[warm cave]

Haha, yeah that one

[agile jay]

Reinforcement Learning with Human Feedback

[warm cave]

Yeah, can never remember the acronym, but know what it is 😆

[agile jay]

I will probably use a small model to go through all your conversations and then rate the answer to each query instead.

Since most users are unlikely to do 👍 / 👎 all the time

[warm cave]

Yeah, the idea is if it's helpful me and anyone else who wants to help would help you build a dataset by using it a bit each day and rating each message

You could even have to prevent you from asking further questions without rating 😆

[agile jay]

Lol, I would rather have it automated

Don't think it's a hard task to label good answers from bad ones.

[warm cave]

Yeah, automated would be good

[agile jay]

I can pretty much predict it by what your next prompt was etc

Or if you keep on repeating it, becaus it didn't get it right.

And then use a better model to generate a better answer and add it for training.

[warm cave]

Yeah, and when there are errors in different parts like code execution

[agile jay]

Yep, I think automating most of the logs etc will make my life easier.

[warm cave]

Yeah, RLHF 😉 will maybe be even more accurate when there are no humans making the rating

[agile jay]

Yep, lol

[kind kettle]

So… basically every interaction with this app is sent to the US Gov, according to the iPhone Privacy Report.

[agile jay]

...

[warm cave]

We're any of your sources .gov recently? 😬

[kind kettle]

Nope

[agile jay]

Don't think that that is done on device

[warm cave]

I just turned on the auditing on my phone so I will see

[agile jay]

The web search is done API side, right?

Oh it is probably getting the preview of the website.

If it was a source somewhere

[kind kettle]

I literally just downloaded the app, created and account, and asked Perplexity to show me news of a sausage dog attack on the face of a woman in the UK that I saw on TV.

[warm cave]

I searched weather on perplexity

[agile jay]

Yep, I think it's source previews

[warm cave]

That's makes sense

[agile jay]

Did it work?

[warm cave]

I reset the audit data, that should I ask it now

Maybe site:apple.com

[agile jay]

How to make pie?

[warm cave]

Sure

The image of a pie is from Amazon and I see the Amazon right there

[agile jay]

yep

[warm cave]

Crisis adverted, haha

[agile jay]

Yep, I checked the site too and it's a news page where the government shares stuff.

So makes sense it appeared if you asked a question related to one of their topics.

[kind kettle]

It’s just really odd that my privacy report shows that site after searching a sausage dog attack lol

[warm cave]

We should delete our message and start some drama with it 😈 Reddit first

[agile jay]

Lol, maybe check that prompt and look at the list of sources

then click on it to see why it was chosen

[warm cave]

Your can reset you audit and then you can see if .gov is contacted again in the future

But there is also a counter next to it

[warm cave]

Just curious, what are your thoughts on the 01 Light from Open Intetpeter

[agile jay]

Nope

Too large, can't fit in a pocket comfortably, so it's doomed to fail

If you want people to carry something, it has to be super useful

[warm cave]

Yeah, but I guess as an overall product, you can make it into any form factor I'd you want

As an open source product

[agile jay]

So without the wall looking device?

Are you chatting with the model using voice?

If so, then it's doomed to fail too.

Have you seen people just trying to have a normal conversation with ChatGPT in voice mode?

From those experiences, I would say voice is not the right direction to go, for quickly iterating.

[warm cave]

I used the voice mode and the keyboard

But the API gets expensive

[agile jay]

Yep, IMO it would only be useful for UI people

Can you make the color more teal, etc

For actual devs, not that useful

[warm cave]

Yeah, honestly I though it showed a lot more promise than humane pin and rabbit R1. Mainly bc it's open source so you can build it into whatever you want, they are already bulding an app

And it can do Devin like stuff

[agile jay]

Yep, but imagine a room of dev people, and everyone is shouting out their commands, lol

[warm cave]

Like ask it to get open web ui running on my computer, it will search online and get it all setup using terminal commands

But I imagine the app will have a keyboard, haha

[agile jay]

Oh, is open interpreter other stuff and not just voice stuff?

[warm cave]

Yes

It runs on your computer

[agile jay]

If it's an actual code interpreter that is local, then it's probably better then.

[unique cloud]

I want to ask: Will the enterprise package have unlimited search instead of 600 compared to the regular pro package?
And how much is OPUS used per day for the enterprise?

[warm cave]

Yeah it is

[agile jay]

There should probably be a preview of changes before it makes them though...

[warm cave]

There is

[agile jay]

Especially if it can delete files, etc

[warm cave]

You have to use the -y flag to have it do stuff on its own

[agile jay]

Nope

As far as I heard from one of the employees

[warm cave]

He said that enterprise has the same limits as normal

[warm cave]

check it out: https://docs.openinterpreter.com/getting-started/introduction

[warm cave]

It can delete files, I was had it kill itself by uninstalling open interpreter, haha

[agile jay]

Lol

[warm cave]

but you have to confirm every action, unless you use the y- flag

[unique cloud]

if enterprise same limit as pro I dont need

[warm cave]

Yep, save your money

[warm cave]

If you are in the a place where you can get it, get Claude Pro, you will get full 200k and 45 messages every 5 hours, but it will use more usage if you use full context.

[unique cloud]

I cant register Claudepro in my country

[warm cave]

😦 Yeah, less options unfortunately

[obsidian pagoda]

Congratulations to the core team on the $1B valuation from the latest round of fundraising 🥳🙌

Perplexity is my primary interface for all LLM AI interactions and I love it. Please stay ethical 🙏

🌍🕊️

https://www.perplexity.ai/search/Perplexity-AI-becomes-SpOKfAvjRVmdpy.XYcXl_w

[warm cave]

https://discord.com/channels/1047197230748151888/1232523238886866955

[feral marlin]

Chatting with voice is just worse in most ways

If you’re good at typing you can type faster than the speed you speak for voice recognition stuff to understand it easily

And you can specify punctuation and spelling

[warm cave]

@livid mantle What happened to the Opus images? I was expecting to see Opus 50 images after the change

[livid mantle]

oh

lmao

quite a busy week and i just thought that this chat will always talk about opus limit caps

[sonic musk]

Currently opus is limited

[warm cave]

Yeah, we will never let Opus die

[warm cave]

Haha, its back 🫡

[livid mantle]

we really need Mr. Romanov's face as an emote in this server.

[warm cave]

Indeed:

@halcyon coral Hmmm... 🤔, care to explain? 🧐

https://tenor.com/view/new-york-islanders-alexander-romanov-islanders-isles-islanders-goal-gif-27621831 Ops, wrong alex

[west sage]

Opus is back? unlimited?

[warm cave]

Dont think so

[west sage]

I don't see the limit anymore

[warm cave]

use this to see how many uses you have left: https://www.perplexity.ai/p/api/v1/user/settings

they only hid the usage counter

[warm cave]

Opus is still limited to 50

[tame current]

sooo

that tweet from aravind....what was it about

nothing?

https://fxtwitter.com/AravSrinivas/status/1781902284844421624

[scenic ravine]

I still hate this opus limited to 50 requests

Like not worth my money

[buoyant kindle]

I'm curious to see for how long AI startups can keep making money off the product of companies making the LLM's.

[buoyant kindle]

Just curious - what do you need more than 50 for? Coding? Or do people use Opus to chat with their 10 year old anime girlfriend?

[placid coyote]

opus, for me, is better at programming. can more "intuit", I guess, what I actually want and I don't need to give it back 3 times like with GPT4Turbo which is also slower

yesterday it felt like opus is trolling me 😄 though to be fair, gpt4t didnt come even close to generating anything resembling a face

[warm cave]

Literally just releasing an enterprise plan

With encryption and privacy practices

[tame current]

sadde

[cinder comet]

I think for coding you will need way more than 50 msgs

So is it now fixed 50?

They said its temporarly?

[tame current]

tell that to antiviruses...

[stable radish]

I hope there's going to be a commercial gesture after all that "Claude 3 Opus message cap"

[harsh verge]

https://x.com/AravSrinivas/status/1782926084667011433

Aravind Srinivas (@AravSrinivas) on X

👀

Twitter · Apr 24, 2024, 12:14 AM

[tame current]

W

[quiet gorge]

Why can you execute code?

My code cannot be executed

[warm cave]

#😎│cool-projects message

[tame current]

will perplexity enterprise pro offer bigger context size ?

[tame current]

It should be. People feel rightly betrayed.

[tame current]

Well. They said it is temporary. So, if it isn't, people should walk away.

[upbeat tiger]

If that stupid joke is not over at the end of the month, they can go f themselves.
I will only use You or Cody (even if that one doesn't give sources and is only in vscode).

[unreal temple]

Sam Altman's Warning to Perplexity

"There are two strategies to build on AI right now. There's one strategy, which is to assume the model is not going to get better, and then you kind of like build all these little things on top of it. There's another strategy, which is built assuming that Open AI is going to stay on the same rate of trajectory and the models are going to keep getting better at the same pace. It would seem to me that 95% of the world should be betting on the latter category, but a lot of the startups have been built in the former category. When we just do our fundamental job - because we have a mission - we're going to steamroll you."

• https://www.youtube.com/watch?v=G8T1O81W96Y

[sweet jasper]

Sam altman where gpt sex

[glad pilot]

I use it to do coding, and it is quite common for me to rewrite the prompt and 50 is definitely not enough. Idk why you said that 10 year old girlfriend crap, if you are satisfied about the 50, how about just enjoying by yourself and stop talking trash here😅

[obsidian pagoda]

I don’t see this as a warning to Perplexity, as it’s acting as a portal/interface to the latest LLMs whilst providing the best online search experience via a specialised model of its own. I rarely use Google any more, because Perplexity is just way easier, faster and more accurate.

Sam is referring to apps that just use an LLM for some niche which the LLM will eventually be able to do natively.

[snow scroll]

hi just want to ask here again, because I didn't get the answer in the quick question channel: I have a trivial question, when I upload a file and ask something about this file with a prompt. How these inputted in a model, the prompt is appended after the file or before the file. Because it is quite different these two ways in term of performance if I uploaded a big file.

It can also determinate how I asked the question.

[stable radish]

This guy has been doing this for a week (even more) now... I would say, just ignore

50 is definitely not enough, if you have "serious" use of text generative AI

[austere kestrel]

neither.. the file is given to / embedded by another model (which has a context window of ~30k tokens), and then when you ask a question about the file, the LLM you are you are actually using will receive 'snippets' from the other model (i.e. it's a RAG system; the contents of the file are neither appended or prepended to your actual query, but handled by another model)

[austere kestrel]

^^ seems strange right.. But tbf it used to be a good / logical system (when GPT4-32k was the context window king).. Now imo it's kinda the embodiment of this lol " one strategy, which is to assume the model is not going to get better, and then you kind of like build all these little things on top of it."

[austere kestrel]

what a bizarre article to be highlighting after they just released an "Enterprise" product that is, presumably, targetted at enterprises...

and as the article itself highlights.. the practice seems more standard than unusual...

[harsh verge]

Yes, that's right. It's the timing of the article's release that seems odd here.

[austere kestrel]

maybe.. but like 50% of other stories written by this person seem to be the same kinda thing.. basically employees at tech companies giving her some tidbit of information about something that happened internally.. of varying significance / consequence..
just seems more like her beat than anything particularly suspicous imo

[harsh verge]

I had read this article in Business Insider: https://www.businessinsider.com/microsoft-blocking-perplexity-ai-employee-access-2024-4 , I hadn't paid much attention to the journalist in question. I agree with your analysis of the substance.

[austere kestrel]

yes same article. Again, it makes this point, which tbh, I think is a perfectly reasonable policy... I mean not least for a company like microsoft, given what some of their employees would be working on

And I believe users of perplexity's Enterpise product have their queries retained for 7 days...

[harsh verge]

Exactly, that seems to be normal in this sector.

[fervent needle]

I wonder what that duration is for deleted thread/queries by regular user with training opted out, their privacy policy only says data is deleted in 30days after deleteing account if I am not mistaken...

[austere kestrel]

this is what my understanding is: "if I opt out in Settings, none of my data is used for 'training'. But my 'queries' are still retained for 30 days, vs 7 days for enterprise users."

don't know if it's accurate / take with a grain of salt - just what i've pieced together fwiw

[fast meteor]

https://www.perplexity.ai/hub/faq/how-does-file-upload-work

[stuck bear]

For Ai model what’s difference between Default and Sonar?

[tame current]

Sonar is less censored, I think.

So, you can potentially write some naughty stuff.

But you shouldn't. You should of course go to church, and think of kittens.

[winged moon]

You really had to say kittens, didn't you?

[tame current]

I love cats.

[warm cave]

https://m.youtube.com/watch?v=W2pYTRdX5LA&pp=ygUJcmFiYml0IHIx

[warm cave]

Yes, sponsor block skipped 90% of the video, haha

It’s a very sponsored video

It’s funny how impressed they acted about the vision features, and how you can ask follow up questions. Like this is not new stuff, haha 😆

[winged moon]

How's the rabbit r1 vs the humane AI pin?

[strange rock]

https://x.com/raywongy/status/1783039023952335144

Some initial thoughts here

Ray Wong (@raywongy) on X

Because you guys loved the 20 minutes of me asking the Humane Ai Pin voice questions so much, here's 19 minutes (almost 20!), no cuts, of me asking the @rabbit_hmi R1 AI questions and using its computer vision to "look" at stuff

Some quick thoughts on the R1:

• The AI/LLM is…

Twitter · Apr 24, 2024, 7:43 AM

[warm cave]

This is what I compare the Rabbit R1 to, I don’t think that phones are going to be displaced by things like this, but with the retro look and AI integration makes this a fun Perplexity walkie-talkie toy.

$199 is the perfect price for an impulse buy, haha. Because of this it will receive a lot less criticism than the Hummane pin, $700 base + 25$ a month and $1000+ with add-ons, makes it so it has to be amazing and cant be a paperweight

[winged moon]

You'd just drop 200 bucks on a device you'll only occasionally use?

[warm cave]

No, I bought it for the 1 year of perplexity

Had no plans to buy the Rabbit R1, but then when i went to purchase 1 year of perplexity I saw the the promo and thought "wait, that is the same price, and i get a free device" but now that the promo is over I don’t see any reason you would buy it, unless you just had money you wanted to get rid off

[warm cave]

But generally speaking I would not drop $200 on a random device, but when i say $200 is the perfect price for a impulse buy, i mean that is a phycological way, where lots will do it without thinking (Impulsively), marketing tricks. Same thing with the playdate, i believe most people who bought it have it in a random place in there house collecting dust, but it sold very well from what i heard, if it was $400 I think it would be a different story. after the novelty wears off, i bet there will be a lot of dust bunnies 🤣

[austere kestrel]

not off to a great start...

[warm cave]

Looks like perplexity got an old article, but not as old: https://www.perplexity.ai/search/siteinversecom-Latest-article-RTM55N0YSuSx6ZaHZnKjpQ#0

Dose perplexity take time to index new article?

This one might not be its fault

Looks like his name is not indexed with the articles

[warm cave]

Yeah, but this is teh one i am looking for

i did put the wrond date for after: tho

[austere kestrel]

yes definitely. I think they index major news sites regularly (and presumably have some solution for weather / finance). but the rest of the web.. they're definitely indexing anywhere nearly as regularly as google (and who could blame them ig.. it's a massive undertaking ha)

[warm cave]

Yeah, looks like it

[austere kestrel]

it's there

[warm cave]

Oh it is there, good find, It is its fault, haha

[austere kestrel]

and while google is formidable.. here's duckduckgo ha

[warm cave]

Duck for the win 🏅

[kind kettle]

Is it worth it to correct Perplexity when it gives you a wrong answer? In other words, does Claude 3 learns from interactions? Example: I asked Perplexity what was the difference between “cognate” (English) and “cognado” (Spanish) and it said they both had the same meaning, which is not true. I corrected it by providing the corresponding meanings taken out from the dictionary.

Forget it, I just noticed that it doesn’t learn anything.

[vapid onyx]

If you click the three dots after the answer, you can mark it unhelpful. Usually the best approach in my opinion

[scarlet pagoda]

hey guy's i have a question

[half venture]

@iron basin llama 3 70b sonar when ?

[tame current]

Billion bucks is crazy

[upbeat tiger]

Great, now following the announcement, give us our f* Opus uses, thanks.

[weary flame]

is the 40$ of perplexity pro enterprise per user?

or for all the company

[thorny canopy]

dude I swear, raise the limit to 200 at freaking least 😭

[rain lodge]

Lol I like how the CEO of Amazon is now Andy Jazzy but nobody mentions him ever.

It's always still Jeff Besos

[tame current]

Robot man

[warm cave]

sounds like enough money for unlimited Opus 😉: #📢│announcements message

[upbeat tiger]

seriously, when will they wake up

[scarlet valve]

sounds like enough money to make searches work after uploading a single image

[loud terrace]

what limits?

[solemn orbit]

Is it just me or Opus is worse now compared to 2 weeks ago? It used to answer the clothes drying scenario correctly before but it can't now

[loud terrace]

there used to be a 600 pro search limit but i dont see that anymore...

[scarlet valve]

gonna need another hundred mill for that, bub

[austere kestrel]

you know what the most popular post in Announcements is..?

Unlimited opus lol

[scarlet valve]

Gotta get you to subscribe, y'know.

[cursive jacinth]

Hi sir, most of the announcement is money but less upgrades sir. Cant help to notice sir lol sir.

[austere kestrel]

don't forget the partnerships!

[cursive jacinth]

ah yes sir. I dont know anymore what is happening to perplexity sir.

oh well sir. Money makes the world go round sir.

I hope someone can match perplexity sir.

[fast meteor]

who controlled the board controls the company...

[cursive jacinth]

is you.com not good enough sir?

[scarlet valve]

Hey, @bug_reports! We have read your posts in earnest!

... Enjoy!

[rain lodge]

Opus is expensive cuz Anthropic charges big fees

Perplexity has gotta negotiate with em

[austere kestrel]

i had a one week coupon / trial thing, which I cancelled today after a week

[upbeat tiger]

They have more than enough.

[cursive jacinth]

how about poe, phind and you.com sir? should they also negotiate sir?

[scarlet valve]

Good way to stop churn is to ignore your customers, for sure.

[austere kestrel]

but if for no other reason than the fact i have enough / too many subscriptions already.. it needed to be better than poe and i would have swapped
and it's not up to perplexity's level when it comes to web retrieval

[rain lodge]

Well it might turn into an enterprise product , tech startup's get value from hype more than from profit

[rustic meadow]

The responses from you.com are actually better at the moment than perplexity ones

for me

[austere kestrel]

ha interersting - perhaps I should have given it more of a chance on that front

(i was admitedly kinda dismissive)

[rustic meadow]

but the UI is so buggy its not even funny...

[cursive jacinth]

the problem is sir, they are a bit unorganized sir, like their UI sir. well perplexity is good actually sir, but lack also of something sir. we aint live in a perfect world sir.

[scarlet valve]

Damn. Editing to make sure more sirs are in there.

[tame current]

Can it search multiple things at the time? It's the main perplexity feature for me that all others fail to provide

[cursive jacinth]

of course sir. that is my life sir. being disrespectful sir is not the way i live sir.

[rustic meadow]

for example?

[scarlet valve]

I'm a mam tho sir

[thorny canopy]

plex kinda crazy for doing the announcment, but not fixing what we want whatsoever... 🤔

[rustic meadow]

perplexity has big problems finding coding responses that work while you.com gets it. both with opus

[cursive jacinth]

mam is also sir for me sir.

[scarlet valve]

but my gender, sir???

[cursive jacinth]

Meanwhile sir, they dont know sir, we care more about improvements sir, not on their fundings sir. 😂

[tame current]

Ask it about several things that changed recently. Ie "who is president of argentine, who is prime minister of poland, what's the latest facebook ai model name" in one query

Perplexity will pass this test, most others won't

[scarlet valve]

Ah. F my gender it is, then

[solemn orbit]

This is true, all the codes written by Opus for the past week has been frustrating to work with

[austere kestrel]

lol yeah with all due respect goodawg, the female part is where it hits up against a pretty firm limit

[cursive jacinth]

Your gender is not my problem sir. saying sir all the time is my life sir.

[scarlet valve]

Many disrespect, sir!

This makes you a liar, sir

[cursive jacinth]

I object sir. we should respect people the way they see gender sir.

[lone citrus]

The GPT4T is half expensive, but now we have 600/day GPT4 meanwhile only 50/day OPUS

Not make sense

[tame current]

1/3, with poland being the only correct

What's the model?

[rustic meadow]

It says its Opus

[scarlet valve]

Not having seen vagina doesn't mean it doesn't exist, sir!

[warm cave]

In the very least use some of the money for a someone to make actually announcements to update the community and be transparent

[lone citrus]

Hire a part time worker isn't expensive for a "unicorn" company

It's not something Can do or Can not do, it's something Want do or Not want to do

[warm cave]

I’m sure Alex could even handle it on his toilet brakes

[rustic meadow]

Research mode gets it

[lone citrus]

and we do not get an announcements after about a week of limit OPUS

[austere kestrel]

yeah but that would've taken a while I'm guessing?

[tame current]

It's like pplx pro mode

[warm cave]

Elon musk dose it

[austere kestrel]

i saw that did alright; like kinda creating a mini research workflow - but it was slow and cumbersome to research what were perhaps par with perplexity results

[rustic meadow]

tbh i didnt watch the time but its def longer than pplx

[tame current]

Would be cool to see labs models in normal perplexity

GPT4 and Opus are way too slow for normal use

Yet no more providing that much better repsonses most of times

[elfin ferry]

i love sonnet

it does the job and fast

[warm cave]

which ones?

[tame current]

What...

[steel tiger]

define normal use?

i sometimes find myself writing out a follow up question in flight with chatgpt whilst its rendering the answer to the previous one

but i also find that it sometimes then covers the question im typing out too lol

[warm cave]

Btw, you can stop the response

[steel tiger]

i often wonder about that, does it result in muted context or does it just obfuscate from the user the follow up

[warm cave]

most of the time it will work well to cancel and then immediately ask the follow up

[steel tiger]

like if i asked it to count from 1 to 50 and i stopped it at 12, would it then continue from 13 or from 50

[agile jay]

How, very long...

[tame current]

To look up a single thing quickly. Pro mode makes it unbearably slow and without it it someitmes won't even acknowledge the news that are in the prompt suggestions

[warm cave]

If you ask it to continue

[warm cave]

lol

[lone citrus]

Even so, I still prefer Opus. Its more human-like tone than Sonnet makes me feel better.

[steel tiger]

theres definitely a big drive for context and i think that more context is never a bad thing, even if it takes longer because poor accuracy will be a major turn off

[tame current]

Llama3 is yet better. But no search

[agile jay]

Llama 3 70B comparison:

Here is a fictional character I came up with, along with a long list of traits:

**Name:** Lyra Asteria

**Physical Traits:**

* Age: 25
* Height: 5'8"
* Weight: 130 lbs
* Hair: Curly, platinum blonde, shoulder-length
* Eyes: Iridescent blue, almost purple, with flecks of silver
* Skin: Porcelain doll-like complexion, with a subtle shimmering effect in the right light
* Build: Slender, athletic, with toned legs and arms
* Tattoos: A constellation pattern on her left shoulder blade, with intricate silver ink
* Scars: A small, crescent-shaped scar above her right eyebrow, from a childhood accident

**Personality Traits:**

* Adventurous, always looking for the next thrill
* Introspective, with a deep love for philosophy and self-reflection
* Empathetic, with a natural talent for understanding and relating to others
* Quick-witted, with a sharp sense of humor and a love for sarcasm
* Independent, with a strong sense of self-reliance and personal autonomy
* A little reckless, with a tendency to take unnecessary risks and push boundaries
* Loyal, with a strong sense of loyalty to those she cares about

**Skills and Abilities:**

* Mastery over the ancient art of astral projection, allowing her to travel and manipulate energy across vast distances
* Mastery over the manipulation of ambient light, allowing her to create dazzling illusions and blindingly bright constructs
* Proficiency in a variety of martial arts, including karate, taekwondo, and capoeira
* Expertise in cryptography, coding, and computer hacking
* Proficiency in multiple languages, including Latin, Greek, and Mandarin Chinese
* ability to communicate with animals, particularly birds and small mammals
* Enhanced senses, including night vision, acute hearing, and a heightened sense of smell

**Equipment and Tools:**

* A worn, leather-bound journal filled with cryptic notes and sketches
* A custom-made, ornate dagger with an intricately carved silver hilt
* A small, polished stone pendant containing a fragment of meteorite
* A set of high-tech, state-of-the-art lockpicks and hacking tools
* A custom-built, high-performance motorcycle with advanced AI-assisted navigation
* A variety of ornate, gemstone-tipped throwing knives

**Personality Quirks:**

* Has a tendency to talk to herself when nervous or deep in thought
* Has a habit of collecting rare, exotic teas and brewing them with elaborate ritual
* Has a weakness for vintage, occult-themed fashion accessories
* Has a fascination with ancient mythology and folklore
* Has a deep-seated fear of enclosed spaces and being trapped

**Background:**

* Born to a family of esteemed astronomers and mystics, with a long history of celestial exploration and discovery
* Raised in a sprawling, isolated estate on the outskirts of a major city, surrounded by ancient artifacts and mysterious relics
* Has a mysterious, unsolved family tragedy in her past, involving the disappearance of her younger brother
* Has a network of secretive, highly skilled allies and informants scattered across the globe
* Has a personal vendetta against a shadowy organization known only as "The Umbra Collective"

hope you enjoy Lyra Asteria!

Not even close...

[harsh verge]

llama-3-70b-instruct on HuggingChat (https://huggingface.co/chat/) allows for web searches

[tame current]

Will fail the multi search test

Need something that will write on more than one thing at the time

[rain lodge]

Opus scales exponentially cuz of context on the front end of its like that in the API it makes sense

Well basically exponentially in the way they're charging for it

[eternal robin]

"advanced search engine"

[eternal robin]

this kept on printing "HELLO WORLD" and produced 2000+ tokens

[agile jay]

Lol

[eternal robin]

you want to see more wierd things?

[harsh verge]

he answered your multiple question: "who is president of argentina, who is prime minister of poland, what's the latest facebook ai model name".

[tame current]

ask it about the same thing but add "what's the latest phi model from microsoft"

the model already has that info in its data

nvm, it doesn't

seems pretty good then

[cursive jacinth]

Hi sir, who cares says perplexity team sir. joke sir. 😂

[cursive jacinth]

who cares about the Opus limit sir, when you have a big funds now sir and being a unicorn company sir.

[tame current]

I'd say it's bad when the prompts are simple and inherently don't need much

Well, that is if you need swift responses

[agile jay]

Use llama 70B then...

[tame current]

Sounds familiar

[agile jay]

https://console.groq.com/playground?model=llama3-70b-8192

You can try it on Groq

[tame current]

Thanks!

[harsh verge]

Command R plus on Cohere's playground answers your multiple questions as well (with your latest on phi too).

[rustic meadow]

What are the usage limits on those free ones? Like Huggingchat and cohere playground?

[agile jay]

There might not be. Likely just depends on the current load.

[tawdry scroll]

In the docs for the perplexity API it shows calling it via OpenAI. Is it also possible to call it directly through the pthon pyplexity library?

[heady quiver]

Cohere is based on what AI model??? Are they better than Mistral, Gemini or Gpt 4 or Claude Opus?

[rustic meadow]

Command R+

[heady quiver]

And whats that in benchmarking???? Is it good or…??

[harsh verge]

In "Overall" he is ninth

[steel tiger]

what is the link to that site?

oh where are my manners, please

[harsh verge]

LMSYS Chatbot Arena (https://chat.lmsys.org/?leaderboard&__cf_chl_tk=OPzDoa61uQx74cR96C3ZrLY5MaSvG5F2yWwIaO7vfnA-1713852824-0.0.1.1-1493)

[rustic meadow]

Im not gonna lie, i just tried cohere and huggingchat with the same questions as pplx and you.com and neither huggingchat nor cohere got it right

[agile jay]

Try llama 3 70B

It's a really impressive model

And looks like the 405B model will be around 95 MMLU

After it's finished training.

[rustic meadow]

I tried it with huggingchat, same question, wrong answer

[harsh verge]

Have you selected the right model?

[agile jay]

What prompt?

[rustic meadow]

I pasted the same prompt into Llama 3, CR+, GPT4T and Claude Opus 3. GPT and Claude got it right, Llama and CR didnt get it

It was about german law

Gemini 1.5 Pro got it wrong as well btw

[harsh verge]

Cohere playground requires 'connectors' for web search

[rustic meadow]

I did activate that

[faint bramble]

thanks for trying out our model 🙂

we just open sourced the interface you are using actually https://github.com/cohere-ai/cohere-toolkit/?tab=readme-ov-file

[cinder comet]

Is this new

Btw did perplexity increase the 50 limit cap

Pls say yes

[rustic meadow]

No news

[cinder comet]

Im f*cked

[thorny canopy]

damn

[rustic meadow]

do you need more of just Opus? and did you try you.com?

[cinder comet]

Before subscribing to perp i had many choices, but at that time( a week ago ) it was 600 usage vs like 100 on phind and similar on you.com, ive tried you.com, it still lacks some features

[rustic meadow]

you.com has no limit does it

[cinder comet]

Doubt that

It doesnt have a standard base model

Its you.com processing + base model

[rustic meadow]

i really start to lose track of all the limits

[cinder comet]

Apparently they are all lowering the cap

Opus is expensive, but 50 is definitly not enough at all

[rustic meadow]

i can say tho you.com solved coding problems perplexity didnt, although its UI is very buggy

[tiny plaza]

the main problem with you dot com in my experience is that you can't regenerate messages, just send them again. hitting the regen just duplicates your message with the previous response still in the context

idk if anthropic implemented their many-shot jailbreaking measures completely but if the prompts feel questionable the model breaks character and defaults to a really sanitized response

you can steer the model back but it's just easier to hit regenerate

so if you're a creative type you dot com is pretty annoying to use x.x

[rotund sedge]

I've noticed that after using the same thread for a while, perplexity will stop providing sources for answers, even when specifically prompted to do so. I've tried to use different language models, but had no success in generating sources in the same thread, having to create a new one/use a preexisting one. Is this a known issue?

[warm cave]

What’s the deal with chatGPT learning about you, has anyone tried it, is it helpful or a gimmick?

[regal sun]

Did I miss you guys talking about this https://huggingface.co/collections/apple/openelm-pretrained-models-6619ac6ca12a10bd0d0df89e

[valid yoke]

Why doesn’t the iOS app respect my default browser?

Very annoying.

I don’t want a built-in Safari viewer. I want to open webpages in my browser.

[civic folio]

gm

[sonic musk]

Thank You

[leaden fractal]

Noob question. But how is perplexity.ai able to search live?

[sweet jasper]

RAG + llm in a nutshell

[harsh verge]

Thank you for making it available to us! I'll take a look 😉

[leaden fractal]

Thank you

[odd mauve]

Does the regular perplexity still support Opus?

[uneven relic]

@blissful raptor

#📢│announcements message

Don't know about the product Aravind, but that ad is kickass! 🤘👊

[odd mauve]

2nd question: Is there a difference between regular Pro and Enterprise Pro in terms of the quality of search?

[wispy basin]

I love yall

[tame current]

I highly doubt it. But you can pay double the money for privacy! Top deal.

[halcyon coral]

Yes

[halcyon coral]

No, please check the screenshot for Enterprise Pro features

[leaden fractal]

what is in simplest terms of understanding LLM with examples? preferably with the context of NLP? I'm understanding that LLM revolutionized NLP LLMs but can you say all LLMs are cable of performing various natural language processing (NLP) tasks?

[tame current]

How many pro uses does the enterprise version get, for Opus specifically?

[halcyon coral]

Currently the same limit as the normal pro.

[tame current]

Oh. Fair enough.

[tame current]

Probably a hallucination. But, I just got this response.

[sonic verge]

hello there, can I log into my pro account both on my mac and iphone?

[placid coyote]

multiple devices shouldn't be an issue - I am logged in on desktop, android (browser and app). limits are understandably shared

[sonic verge]

yeah idk I couldn't use pro on my phone although I was logging the same gmail account, and now whenever I'm logging in on my phone a noti pops ups saying "There was a problem signing you in. Please try again later."

[vapid onyx]

You should be able to!

[placid coyote]

how is opus going? I am a bit afraid to try using it again, because last few (2?) days it never showed on web (pro slider nor in model selection) how many "daily credits" remain. only after I exhausted them and got kicked to sonnet, it showed 0 in model selection 😞

[stoic wave]

Most likely, i'd rather prefer to use one device at a time.

[solemn cedar]

On the pro version is GPT-4 limited or unlimited ? Also what's version is it ? The last one ?
Any plan to have Llama 3 70b with Groq lighting speed on Perplexity ?

[placid coyote]

GPT-4 Turbo is technically limited, but I think 600 per day? They used to show how many daily uses remain, but that was taken away some time ago. I somewhere read almost all request (80% few days ago?) are done via the new one

[sweet jasper]

Its limited to 50

[solemn cedar]

thanks

[placid coyote]

seeing how hard time gpt4t had with "drawing" in pillow (python), I thought it would be useless for basic visualization, but this isn't half bad. just two prompts, but the first one was fairly detailed (and, well, I used collection prompt, so that might be considered cheating). I can imagine with good pre-prompt, it could be possible to create a decent "chart collection"

[rigid kestrel]

why this gpt4 is so trash? is different than poe one?

[tame current]

Context size, probably.

[mighty fox]

GPT- 3.5 limited ?

[solid whale]

I thought they already had an option to opt out of having data used for training.

[lone citrus]

Intresting

I would use chatgpt plus if they do give more than 100 messages per 4 hours

[solemn cedar]

I don't believe them, I stopped believe what they say months ago

[lone citrus]

May i ask what made you feel that?

[solemn cedar]

They lied many times, for example they tweet gpt-4 laziness was a mass hallucination, then admitted the contrary a few weeks after ... Also they were never clear about plus caps, some people's got more than others. I don't even remember why I don't believe them, cold company, doesn't give a *** about their customers. I haven't subscribed or use their services for 6 months now

[honest basin]

yeah openai is the last company you want to be leading the industry

[south kindle]

anyone else here playing around with Udio?

[signal hamlet]

Hey, @sonic verge! Could you update to the latest version or reinstall the app and try again? Please let me know if you could sign in after that.

[vivid heart]

is there a feature request discord channel?

[halcyon coral]

#💭│ideas

[vivid heart]

ahh that's an interesting name, I thought it was just ideas of how people could use the service 😂

thanks

[warm cave]

Yeah, after you mentioned it, best i have tried

[boreal idol]

Are there web ui pages that support conversations based on pplx-api？like ChatGPT-Next-Web

[odd mauve]

Hi, if I use the AI Data Retention toggle in the regular Pro version (NOT Enterprise), will I opt out of my data being used to train only Perplexity's models, or will that opt out of training Anthropic's models too?

[half venture]

Is opus still 50

[pine valve]

yep

[agile jay]

You can check here:
https://www.perplexity.ai/p/api/v1/user/settings

[half venture]

Wait gpt 4 turbo also has a limit ?

571 ?

Or 580

[crystal elbow]

I can't see the Opus limit anymore, where is it now ?

[agile jay]

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

You can see it there, in the json

[half venture]

Raise it to 70 atleast perplexity come on

[agile jay]

That's just the normal limit I believe

[half venture]

Atleast offer as much as the official site

And mind you

It doesn't reset daily in website

It resets 8 hours

Come on perplexity

[agile jay]

Yep

[half venture]

Is there any app

Where I can try out

Stable diffusion 3

[crystal elbow]

Why are these guys activly making it harder and harder to find that, why?

[agile jay]

Perplexity has SD3 XL

I believe

[half venture]

Sd XL

Not sd 3

[crystal elbow]

just put it in the damn PRO thing again like before, jesus christ

[agile jay]

Isn't SDXL supposed to be better?

[half venture]

No

[agile jay]

Most likely because of the investment thing

They wanna hide any bad publicity

#📢│announcements

[half venture]

Yeah

[thorny canopy]

well supposedly if using your 600 on opus wasnt abuse, then why cant they raise the limit? Especially with the surge of investments.

[agile jay]

Money, probably

Don't know why it's been a week and they couldn't fix someone as simple as usage abuse. It's not hard to filter out accounts that have unusual usage patterns.

[tepid portal]

Perhaps they are putting money into the development?

[half venture]

No it's simple they are diverting resources to enterprise

That's it

They will probably restore opus a bit

As they get more gpus

[agile jay]

$40/seat for basically no additional usage etc is pretty much a rip-off

[thorny canopy]

bruh i just want 200 a day at least

[agile jay]

They don't host opus

[half venture]

Billion dollars companies don't care

[agile jay]

Most of the market is not billion dollar companies.

[half venture]

Most of the market doesn't use perplexity yet

Only high end do

Adoption takes time

[agile jay]

Have you seen their contracts with SoftBank and SKT?

[half venture]

Nole

[agile jay]

They are practically adding 100M users just through that

[half venture]

Bo

Nope

[half venture]

We will see

[agile jay]

Which is 10x their current monthly users.

But that should make it more mainstream in those countries, which will mean more businesses will want to use it.

[fluid mauve]

What idf the unusual usage patern was coming from a paying client? I had this all the time in Cyber when I worked for a trading companty. We came down on the side of it was btter not to cut off a paying client as opposed to 'shields up' and dropping everyone. I believe before I put some stuff in place we cut off a high value client. Lets just say Management were not happy. Mind you they also would not be happy if the platform got taken down in a similar fashion.

[tame current]

The unusual usage was about automating it and using in other products, and probably from trials and the $0 coupon codes

[harsh verge]

Stable Diffusion 3 is just available with their api (stability ai) But the Stable Cascade model built on the Würstchen architecture (https://stability.ai/news/introducing-stable-cascade) released on February 12 is available for trial on Huggingface https://huggingface.co/spaces/multimodalart/stable-cascade

[agile jay]

Yep, isn't the reason why abuse is wrong because you are losing money from having those clients?

So why does it matter if the bot user is paying if you are losing money by having them?

[fluid mauve]

we weren't though. they flew really close to the edge of the Ts&Cs, but ultimatly traded. We still amde money from the mtcro-finance and holding fees. They were just so close to be ing cut off due to terms violations

[agile jay]

That's a different situation then, since I'm pretty sure Perplexity would be losing a lot from a user using all 600 msgs every day.

[fluid mauve]

So they were abusing the free trial week and thats why that got canned?

[agile jay]

Yep

[agile jay]

Some were paying too

[fluid mauve]

paying and sharing accounts

[agile jay]

But $20/month does not cover 600x30= 18,000/month

[fluid mauve]

definitly dosent. i like the enterprise $40 a month but I am not sure I am ready for that yet. i was never anywhere near 100 Opus uses per day

[agile jay]

At an low estimate of 1000 tokens of output for each request, that would be 18M output, which would be 18x75=$1350

They likely have an enterprise deal with anthropomorphic, but it still isn't cheap.

[fluid mauve]

Definitly costly. I run ollama and a few models here when i am trying stuff out. No costs to worry about, well apart from electricity.

[tame current]

Hey mates. Why can I no longer use claude 3 opus with perplexity Pro? Is that limited?

[fluid mauve]

Limited to 50 a day, There is a pinned message about it

[tame current]

Jesus christ, that is terrible. Why i cant see it bevor i buy premium 😄 In europe we have transparency obligations 😄

[fluid mauve]

im in europe. platform was abused. This is 'Tempory measures' Feel free to sue.

[tame current]

-.- abused rly ? i hate some humans ... Hmm you know more avout that abuse ?

Nah broh to sue is to hard 😄

[fluid mauve]

just been told people were abusing the free week trial.

[limpid mason]

Where did you see that ?

[tiny plaza]

i just hope it's not a smokescreen to permanently kneecap opus usage TwT

[tame current]

Ah i understand, some guys abused then and takemoney back.. like this

[tame current]

Yes the same i think also 😄 i hope^^

[fluid mauve]

the took Opus uses down to 30 per day, and a few days later moved it back up to 50 per day. Bear in mind these reset after 24 hours from time of use. So if you use a opus at 9am on 1st of month, that will be refreshed at 9 am 2nd of month

[warm cave]

Dalle 3 is better at text than I thought

[agile jay]

It had a recent update which improved it a lot.

[austere kestrel]

you're saying SoftBank (an investment fund) and SKT (a south korean telco) have 100m employees?

[agile jay]

No, it's their telecom section

You get a phone contract and you get perplexity free with it

Which I think that's how it's being rolled out.

Maybe with broadband too

[austere kestrel]

south korea has like 40-50 million people (only a chunk of whom will be signing up to this telco in the future)

100m sounds like a stretch ha

[agile jay]

Yep, probably closer to 75m

But still a lot more than their current userbase

[tame current]

Which AI model other than Opus, (limited to fck 50)
Can analyze the books well and find information from them?

[austere kestrel]

bro that's still larger than the entire population of south korea

[agile jay]

GPT4 Turbo

[agile jay]

Yep, but not Japan

[austere kestrel]

anyway, hopefully perplexity does east asian languages well

[limpid mason]

Is it normal that there is 300 pro search per day I stead of 600 ?

[tame current]

Nah 😦 he tell me bullsh:D I ask him and he say. no specific inforamtion, but there are ^^

[agile jay]

Yep, hopefully they do a few finetunes of llama 3

[agile jay]

Free users have 300

[limpid mason]

No…

[tame current]

normal user have 5 Pro search per day ^^ pro 300 😄

[agile jay]

I don't know, I've only been a free user for 10 minutes

[tame current]

Why does Gpt 4 turbo tell me that it has no access to the document, I have switched off the Pro search 😄 before that it could read the document. I then tell him “but you can see it” why ?

[tame current]

for free 5 per day and pro 300 trust

the normaly search i dont know^^

[bitter hazel]

hasn’t the voice feature been available for awhile? i used it yesterday and it talked to me

[austere kestrel]

right..where they apparently some partnerships with local companies.. which is not quite the same as 70 million customers.. but anyway.. good for perplexity - can agree on that ha

[agile jay]

Yep, but can they handle the load? Lol

I guess now they can negotiate for cheaper prices easier too.

[raven sierra]

Can anyone explain what the new announcement is about

And how can I access it

( I have pro )

[austere kestrel]

그 사람들이 한국어를 다룰 수 있나요? would also be my question ha

[agile jay]

Doesn't Korea have the extra outside traffic costs?

Probably not the country I would have picked, unless I could host the servers in Korea.

[halcyon coral]

Sure, the computing resources are scaled dynamically.

[halcyon coral]

You access it with the latest iOS version. Please note that the feature is currently being rolled out.

[raven sierra]

And once it rolls out how would I access it - what exactly is it

I notice the UI

[agile jay]

Yes but I assume that Openai and Claude have a maximum request limit for their API?

[ripe void]

@signal hamlet the vocal modality update is so funny, the dude started screaming while explaining the divergence theorem lol

W update

[raven sierra]

Is it an improvement on the current voice system ?

Like a refresh

[halcyon coral]

You'll find a button in the search bar.

[muted pine]

Do you have the iOS update available ?

[raven sierra]

Not yet

[muted pine]

Oh ok

[halcyon coral]

Large companies get a lot of capacity

[agile jay]

Yep, but a lot isn't much when you also have a lot of users.
If the average number of requests is 10/day, that would still be 3B/month

With the user base doubling or tripping soon, that would increase even more to around 10B/month

[halcyon coral]

Then you make an agreement with OAI and get separate access

[agile jay]

Also how's the abuse thing going?

[halcyon coral]

It is still being worked on

[upbeat tiger]

Temporarily

[raven sierra]

I’ve updated and I don’t quite see the new button

[halcyon coral]

#💬│general message

[raven sierra]

Would it not roll out along side the update

[odd mauve]

Gemini used to be supported on Perplexity. It's not anymore.

[raven sierra]

I just updated it it was a new update from 5 mins ago… anyhow I’ll wait a bit

I see the new voices at least

OO I hear the response now but not the UI

new voice is nice

[warm cave]

Been gone for a while now

Guess they plan to never bring it back

[vital thicket]

Hey Are you from the rabbit server? I updated but I don’t find any new feature. I’m guessing it’s just a text to speech mode. Which I already had I am Perplexity pro.

[raven sierra]

It is I think some new voices - accessible via the settings

And UI which I don’t have yet

But just ask it a question with the mic as you would regularly

The new UI would be sick tho - I just don’t see it

Just the regular audio response screen

It’s visible now 🙂

[vital thicket]

Ah that would be cool. I don’t think I have new voices. There’s male and female so I don’t know if there was female already but now I changed it to it lol

[raven sierra]

There was one of each I think and now there are two

New UI came in for me

[vital thicket]

Would be nice that we could upload files directly from the app. I wish they would add that feature.

[south kindle]

#1233142008143548417 message

[junior adder]

guys, is there any chance to know the temperature used on opus and sonnet? Using writing purposes.

[warm cave]

lol, I got to see what this is

[junior adder]

It's kinda lacking if compare to poe

[safe wolf]

Hi! Sorry, what is this?!

[thorny canopy]

thinking its low, apparently I heard theyre considering adding an option to change it. Tho idk.

[south kindle]

check the linked thread. entirely creative writing use, not meant to be productive.

[junior adder]

it would be perfect if they did, changing temperature is crucial if you want to generate some novel

[thorny canopy]

couldnt agree more

[twilit halo]

perplexity for watch OS needs to happen

also ios widget for the voice feature would be great

[quasi gale]

I am using perplexity pro. Does anyone know: In the setting page why the Pro Support button does not work?

I emailed the support team, did not get any response!

[warm cave]

Im making a feature request, including cost saving measures, any suggestions, one will be being able to change the model from the main page so you dont start the conversation just to click re-write and use twice as many tokens

[wraith crow]

it's probably a browser extension blocking the pop-up, please try again in a private tab or disable some of your browser extensions and try again.

[quasi gale]

No, it's not. I tried with different account same browser with all the extensions. Works well those accounts.

[warm cave]

Hey, can you guys add to this if you have anything i missed: https://discord.com/channels/1047197230748151888/1233156627277021357

[vital thicket]

So far, you can use it with API someone did it in this discord. I am using that Shortcuts it works great

[tepid portal]

Where is the version of that for Android? I don't run iOS on my S23 Ultra.

[agile jay]

Nowhere

Android users are second class citizens

[tepid portal]

So a pro user is a paying customer regardless of the device. What do you mean?

Okay, paid apps over the store are a problem for Android because of piracy, but that's not the case with Perplexity Pro.

So...?

[agile jay]

Lol, it takes months for IOS features to get added to android

On android you probably get a better experience using the website.

[tepid portal]

Why does the app even exist?

[agile jay]

So they can say they have one...

And because Apple doesn't support PWA's as normal apps.

Then they could just focus only on web, and have all the users have access to the features.

[tame current]

How come. You can just install it from safari, works for Jellyfin perfectly

(Sort of) works offline even, but obviously not access to the content

But not many people know that. Easier to install an app and also get some traction from app store

[agile jay]

One of the main reasons we helped the EU and US sue apple is because of PWA support that they removed.

[tepid portal]

So Apple is a special kid in the class (as always, lol), so they decide on that one unsupported platform, rather than for everyone?

WHAT?

Android Market Share Worldwide According to the latest data, Android dominates the global smartphone market with a 70.69% share, while iPhone (iOS) has a 28.58% market share.

[agile jay]

Yep, except apple takes 30% tax, and you have to go through audits for all updates.

[tame current]

Can't test for iOS 17, but works on 16. I've heard that they readded PWAs in some update

[agile jay]

When with PWA's you can support mobile, desktop and web, with one codebase

[agile jay]

Yep, now we need to force them to open up the API's to PWA's

Notifications, accessing storage and other stuff is still annoying or not doable

[agile jay]

In the US the market share is around 70% IOS

[tepid portal]

More than 70% of mobile devices + desktop supports PWA...

[agile jay]

95% for people under 25

[warm cave]

I know USA has a lot of IOS users, but I thought Android was still over %40 or so. Seems like there is only a slight advantage having an iOS app over android be kept up to date

[agile jay]

Or just one codebase...

[tame current]

Telegram PWA has notifications, no advanced settings. It just asks you for the permission

[warm cave]

I guess age demographic also matters

A lot of young iOS users

[agile jay]

Yep, but I imagine the notification control they have is pretty minimal.

Compared to native apps.

[tepid portal]

Browsers supports notifications pretty well.

[tame current]

It's as bad as normal browser notifications

[agile jay]

Yep, but pretty well is pretty different to exactly like native apps.

[tame current]

Telegram is the only app that supports notifications properly, rest doesn't even try. Like slack or discord. They only work when running

[agile jay]

Yep, because it's such a bad experience

What apple always does is make some of it possible, but then make the API as hard to use as possible.

[tepid portal]

Eh? I've got Discords notifications even when I clear all my recent apps.

[agile jay]

Which is against anti trust

[tame current]

I mean, PWA notifications

Native apps of course have it done right

[tepid portal]

Oh, you mean PWA, okay

[agile jay]

Yep, it's pretty stupid you need multiple codebases to make a copy of your app for multiple devices.

Just support PWA's as much as possible

[tame current]

You can use React Native, like Facebook and many other apps do. Wonder why doesn't pplx do that

Or flutter, like google

[tepid portal]

Yeah exactly, right... Apple, do you heard that?

[agile jay]

Those are still native apps

the point of PWA's is they work on all platforms

With react native, flutter, xamarin etc you need to compile to each platform.

So it's only supported if the dev decides it's woth it.

[cinder comet]

https://twitter.com/btibor91/status/1783603187993252338

Tibor Blaho (@btibor91) on X

Yes, Sonic - SNC (SearchGPT) agent seems to be already in eval phase with:

• image search
• widgets (weather, calculator, sports, finance and timezone difference)
• option to ask follow-up prompts

Using GPT-4 Lite (Scallion; POR), GPT-4 or GPT3.5 (Sahara-V) as model and…

Twitter · Apr 25, 2024, 9:05 PM

[tepid portal]

And then end up like Discord. Discord runs best in the browser because it's fully hardware accelerated, not like the old Electron build that Discord shipped and used.

[cinder comet]

seems like perp has a competition here

openai is working on their own search engine

[agile jay]

Yep, how it's integrated will matter quite a lot though.

[tepid portal]

As soon as they use Bing, it will be pure garbage.

I tried Copilot from Microsoft, but it redirected me to a URL that doesn't even exist.

[tame current]

It even hallucinated gpt4.5 turbo release

Redirecting to the half year old article that was deleted

[stoic wave]

😄

troll ai

[agile jay]

That's why you need some type of source validation

[tepid portal]

So, let's wait to receive new features on Android, because Apple is the unique kid in the class that won't behave normally and can't support PWA naturally.

[tepid portal]

Yeah and in case of Bing, you can be 100% sure it will provide wrong URL.😁

[tame current]

Don't forget gemini [url invalid]

Yet better

[half venture]

Is voice to voice available in website or Android

[tame current]

Seems it's only on ios

[half venture]

What kind of bs is that

[agile jay]

Nope

[half venture]

Literally discrimination

[agile jay]

Yep, they only care about IOS

[half venture]

Not only Android

But windows user too

[agile jay]

Should be web first

So all users get access

[half venture]

Pretty sure chatgpt doesn't have voice in there

Website either

Can we clone voices

Or just oreset

Preset

[honest basin]

sounds like Aravind is putting the 62 million dollars into a new mansion and a couple supercars instead of hiring talent

[half venture]

True

[agile jay]

Or a nuclear bunker for the AI apocalypse...

[agile jay]

@raven sierra what does the voice thing look like in IOS? @atomic crane Was wondering

[feral hazel]

anyone has tried out both perplexity and you.com as paid subscribers? Do you mind sharing your experience about your preference? thanks. I would like to know how perplexity compared against its closest competitor

[feral marlin]

with you.com sometimes you need to send it your code or whatever source you're trying to base stuff off of and then ask your prompt in the second question but other than that its nice

cant do image uploads with most models on it im pretty sure but the context is convenient and student benefits are nice

[raven sierra]

Just like the announcement

[thorny canopy]

its been a week now since the limit, yeah?

[arctic spindle]

Since opus limit?

[thorny canopy]

yeup

[arctic spindle]

I believe that to be the case.

[thorny canopy]

big rip on this 'soon' business

[arctic spindle]

Well, I am only using gpt 4 turbo at the moment, waiting for perplexity to completely add the new version of gpt 4 turbo.

[agile jay]

It's around 70% done, last week

[arctic spindle]

I'm aware of that. Do you happen to know the timeline for full implementation?

[agile jay]

You would have to ask Alex or Denis

I guess Aravind would know too

[arctic spindle]

I'm uncertain about pinging them.

[thorny canopy]

alex usually responds with pings, but I havent down it myself

[agile jay]

@signal hamlet @neat elk How's the rollout for the latest GPT4Turbo model going? Last time we were told it was around 70% new update, and 30% old update?

[feral hazel]

indeed. half price for students. This is something which perplexity loses to you.com.

[feral hazel]

I think that's ok for coding questions. For questions on stackoverflow, one has to show some code first before getting answers. That's the norm.

[feral marlin]

true but i just meant because it turns it into a text file then says review the prompt, usually that messes it up if you ask the question in your first prompt, so you have to send one, let it analyze it, then send another as the actual question

[feral hazel]

the latest, most powerful gpt4 turbo is not yet available on perplexity?

[arctic spindle]

The availability is currently at 70%, and they are working to achieve 100% soon.

[feral hazel]

sorry. I don't understand what is meant by 70% availability. To users, it's either there or not there

[arctic spindle]

Approximately 70% of the messages are directed to the latest version, while the remaining 30% are routed to the older GPT-4 Turbo.

[feral hazel]

oh ... as a user, that's bad experience. When I ask a question, I expect the question to be directed to the LLM that I specified. I don't encounter this problem with other AI chatbots that I use. Of course, If other chatbots do it secretly, I won't know but if I know, I will stop dealing with them for the dishonesty.
Good that perplexity is transparent about this limitation at the moment.

[signal hamlet]

Currently, the ratio is still the same, we are working on it.

[ancient flare]

Yo. hows perplexity rate compared to phind and you.com? No sneaky truncation or context length limiting?

[agile jay]

32K context

[ancient flare]

ah. seems to be the standard for API resellers

that includes claude API?

[agile jay]

Yep, but currently rate limited because of abuse to 50/day

[ancient flare]

transparency is good. theres a reseller that claims unlimited but you betcha theres truncation and all kinds of nuisances

[agile jay]

Or they use RAG

Long context is gonna get cheap though, because of models like llama 3

[ancient flare]

i wish. it goes off task less than 8k in

[agile jay]

They've already gotten them to 160K context, and are still increasing it

[ancient flare]

only coherent context offering is claude, google's 1 million is false, at least for code

[agile jay]

It's 160K coherant context

[ancient flare]

bring it on. along with a nice tune. and im keen on meta's bigly bot

[agile jay]

Yep, you can see it stays perfectly green until around 160K

24 hours ago it was at 64K

So they are making steady progress.

They are also making performance improvements too, so it will be pretty nice once a bit of time has gone past

[ancient flare]

hmm i wonder if that test is good enough for coherence, if google uses the same one

[agile jay]

Yep, it's just needle in the haystack test

[ancient flare]

its only retrieving a value, but what if you want it to reference a line of code 100+K tokens in which gpt4-128k and gemin 1.5 pro fail at providing many values or keeping a codebase in scope

they probably all use a standardized test except claude

[agile jay]

Yep, but I think needing to reference a line of code 100+K lines in probably means it's being implemented incorrectly. Unless you are actually loading the whole codebase.

https://x.com/mattshumer_/status/1783531642344067159

Matt Shumer (@mattshumer_) on X

It's been a week since LLaMA 3 dropped.

In that time, we've:

• extended context from 8K -> 128K
• trained multiple ridiculously performant fine-tunes
• got inference working at 800+ tokens/second

If Meta keeps releasing OSS models, closed providers won't be able to compete.

Twitter · Apr 25, 2024, 4:20 PM

[ancient flare]

i use them to develop apps, implementation is only doable with opus, since the others forget the code its given plus the original context code you've provided. it becomes evident once it gives you examples not within the codebase scope. if its staying in context, it will give examples within the codebase. gpt4-128k is actually worse than gemini 1.5 pro at this

[agile jay]

That's more a client problem, since my own client implementation even outperforms Opus on a lot of tasks.

The main issues come from doing 0 shot and using only a very basic prompt.

Doesn't give the model much to learn from.

[ancient flare]

i use very developed prompts, the performance is showing once you get deep into the context window

[agile jay]

Oh, my context is dynamically managed, so it doesn't grow too much.

[ancient flare]

thats how i work around it. just reset the chat and send it the current code

[agile jay]

Yep, but that's more a problem of the interface.

A chat dialogue is not a good solution for writing code

[ancient flare]

wont be that way for long, but its early days yet

[agile jay]

Yep, I'm coming up with my own way to do it.

[ancient flare]

Matt is good, but he tests LLMs with the python snake game lol

[agile jay]

Yep, that is too basic

[ancient flare]

they're all trained on that now. he needs novel testing methods

[agile jay]

Yep, but it's also that there are probably many demo's of making a snake game online. You need a test which shouldn't be in the data, and that you need a good code structure to complete.

Also llama finetunes on different languages will be nice

[shut plaza]

Yo

[arctic spindle]

Do you have an estimate for the total implementation time?

[meager sparrow]

I am very sick and tired of people in my school yelling and constantly saying my address all over the place. I ask for theirs and they say “it’s confidential. It is very annoying. Is they a way I could get their address so they could shut up?

[ripe void]

Stop trying to larp