💬│general

[minor shard]

i am guessin you are the one to type "STOPCOMENT" to get it to stop

[prisma socket]

yeee xDD

but at some point it will eventually stop itself

sometimes only one message

[minor shard]

Lol idk if I am a fan of it @'ing everyone. Might need to add that lol

[prisma socket]

you can even instruct hwo to replly to people
like that

[minor shard]

Wait im confused, are you using Comet for this lol

[prisma socket]

but @ people didn't work

[prisma socket]

i tried a bit with discord yes

xd

[minor shard]

nice haha

[prisma socket]

but against TOS

so won't recommend

[minor shard]

I agree haha

[distant aurora]

Jaw check pulse.

[dry kestrel]

Interesting From my testing Comet is given context of your recent browsing history as like a tool call or something similar

The usual perplexity system prompt gives the model your location, name, and memory settings

Aravind responded to that article https://twitter.com/AravSrinivas/status/1915533071291474139

Aravind Srinivas (@AravSrinivas)

Low effort clickbait article that takes things out of context. The podcast hosts asked me a hypothetical question about how ads could play out in AI products including Perplexity. I laid out a scenario. The core point was first cracking memory and personalization if you do want

X · Apr 24, 2025, 10:27 PM

It doesn't show the full CoT when you have web access on from my testing, really weird

[distant aurora]

Yeah I know. I'm just saying that if it can pull what discord channels he used then maybe it can pull his dms too.

[dry kestrel]

I think its just pulling from the browsing history (which google collects data about as well)

I would like to see a seperate TOS and Privacy policy for the Comet Browser only though

[distant aurora]

Isn't the url in the form of an id number instead of a name for a discord channel?

[dry kestrel]

The browsing history still shows the name of the tab

[lusty wolf]

it can also just see pages, so it reads everything. It's not an issue of privacy if it's stored on device. However, it is if it's being sent to servers and trained on/stored/sold then it is

[distant aurora]

Fair enough.

[dry kestrel]

I would like to see more transparency regarding the models used and how the data is dealt with on the Comet Agent though

and if disabling data retention on perplexity.ai also covers agent use.

[lusty wolf]

I believe this is something in progress.

I wish I could find it, but someone at perplexity (probably Aravind) mentioned being able to do something that in the future

[dry kestrel]

Currently, they need to update and align their documentation regarding Perplexity and Comet

[dry kestrel]

5000+ notifications is insane

[prisma socket]

XD

[prisma socket]

my servers are ehm to much

[distant aurora]

Bro has all the chicks of discord.

[dry kestrel]

Haha

[dry kestrel]

I've hit the 100 server limit already, its incredibly annoying as I don't want to leave any of them

[short narwhal]

5k???

[prisma socket]

i gotta sort out but to lazy

[distant aurora]

Is that a server with a comet invite card as pfp?

[dry kestrel]

At this point I'm just forced to buy Nitro

[prisma socket]

i am not checking notificatoinis on all servers

[empty ice]

XaI researchers "researching' waifu tech:

[prisma socket]

ah it was testing comet as discord chatter with few peoples from here xd

[minor shard]

U have Nitro tho, oh nvm

[prisma socket]

someone created it

[dry kestrel]

I have the basic version (for file uploads)

[distant aurora]

What happens if you don't update nitro? Do you get kicked out of half of the servers?

[minor shard]

ahhh ok

[dry kestrel]

That.... would be my guess

Or you can't interact with them

[distant aurora]

Why are y'all on 100+ servers

[dry kestrel]

Oh nvm https://www.reddit.com/r/discordapp/comments/17u3bhn/what_happens_when_you_unsubscribe_from_nitro_when/

[dry kestrel]

You just can't join an more servers

[lusty wolf]

you know you can just mute them right? lmao

[minor shard]

I lot of services have discord servers for updates, etc.

[dry kestrel]

half of which are related to AI

[dry kestrel]

same here

[lusty wolf]

how many services are yall using lmaooo

[distant aurora]

Ikr

[empty ice]

Lmao i love these

[dry kestrel]

a ton

[empty ice]

I use 2

Claude and perplexity

[lusty wolf]

Elon retweeted it too

[dry kestrel]

It's interesting coming from a guy who is promoting higher birthrates

[empty ice]

I am a new user of claude

[minor shard]

Lol, I have some for AI, Some for Games, some for discussion servers like Python or Microsoft Certifications, etc.

[empty ice]

Fr?

[minor shard]

lol, X community in a nutshell

[empty ice]

Mechahitler

[distant aurora]

Would have tried the left one if I could add it.

[distant aurora]

Mechapookie now.

Lol.

[empty ice]

Wht is grok even

[distant aurora]

Ai model

[lusty wolf]

Mechahitler

[empty ice]

Gooning asi in 2035?

[distant aurora]

Use it as you please.

[empty ice]

Lmao

[lusty wolf]

Elon said he wants Optimus Robots to be Waifus

[empty ice]

https://tenor.com/view/side-eye-dog-suspicious-look-suspicious-doubt-dog-doubt-gif-23680990

[distant aurora]

W

I can't pull chicks.

[empty ice]

Bladerrunnermaxxing

[distant aurora]

Atleast I would be able to buy some decent looking love and care.

[empty ice]

https://tenor.com/view/ana-de-armas-ana-de-armas-blade-runner-ana-de-armas-knives-out-knives-out-blade-runner-gif-4618139716709582242

Elon invents real agi girlfriend

[distant aurora]

Everyones a K fan until they meet a K irl.

[empty ice]

People born in the us next year: 130

[distant aurora]

Bro I see population decline.

Half from lgbtq

Half from robo girlfriends.

[empty ice]

Is that call of the night gooning

[lusty wolf]

I'm fine with it. We're overpopulated (or under developed) anyway. It'll help to reduce that

[distant aurora]

Call of virginity

[lusty wolf]

Plus Boomers will be dying soon anyway

[empty ice]

Eill i be the last person to watxh anime for the plot

[distant aurora]

I watched chainsawman for the plot.

Half of the shit in the opening and closing scenes never happened.

[lusty wolf]

I find it to be creepy. I asked "explain why Comet is peak" and it made these creepy moaning/whisper/demonic shit noises for about 15 seconds, then whipsered her answer

most of which was completely unintelligble

[distant aurora]

Dafuq

[lusty wolf]

Tried to record it, but audio didn't work for some reason

[empty ice]

Plot twist elon is actually channelong satan no ai used

[distant aurora]

The cyber nun (powered by grok)

https://tenor.com/view/the-nun-valak-scary-creepy-close-up-gif-17859787

[empty ice]

Maybe it was lilith

😂

[lusty wolf]

https://fixupx.com/nikitabier/status/1942462167665762517

Nikita Bier (@nikitabier)

Going from an office where AI researchers are building the Antichrist to my living room where my girlfriend is watching Love Island is one of the most drastic transitions in the known universe

**💬 587 🔁 1.1K ❤️ 30.2K 👁️ 2.63M **

FixupX · Jul 8, 2025, 5:54 AM

[empty ice]

[lusty wolf]

Nikita did say they're building the Antichrist

[empty ice]

The girlfriend in question:

[distant aurora]

Downgrade

Anti christ > gf watching love Island.

Oh shit.

Does he have both at once?

💀

Bro is god's favorite.

[empty ice]

https://tenor.com/view/cute-anime-anime-girl-pink-hearts-gif-2859182496452245481

[lusty wolf]

Love Island is the Antichrist

[short narwhal]

Don’t involve nazuna with the grok thing 😡

[blazing comet]

https://fixupx.com/wilplatypus/status/1944898149098647903

wil michael (@wilplatypus)

gonna put grok in labubu

**💬 7 ❤️ 24 👁️ 3.1K **

FixupX · Jul 14, 2025, 11:13 PM

[lusty wolf]

Found the gooner

[blazing comet]

😭

[distant aurora]

Lmfao

[empty ice]

Labubu but eith screen to goon to goth grok in underwesr

[minor shard]

Don't involve anything with Grok. Grok is poo poo.

[distant aurora]

Grok was goat. Now it's woke.

[lusty wolf]

Nothing better than driving your Tesla while Gooner Grok is whispering demonic rituals through your speaker

[minor shard]

Woke is a dumb word lol

[lusty wolf]

Opposite direction pal

[distant aurora]

?

[lusty wolf]

It becoming less "woke"

Mechahilter

[minor shard]

lmao

[distant aurora]

Sorry

I don't support liberal ai.

https://tenor.com/view/cat-gif-10307912810788206700

[lusty wolf]

Is MechaHitler liberal now?

[empty ice]

Imagine if all teslas had the girl on screen 0/24

Everyone dies in accidents

[distant aurora]

Mechahitler isn't there. They got rid of it.

[empty ice]

😂

[minor shard]

https://tenor.com/view/lookaway-anyway-cat-kitten-anyways-gif-19960539

[lusty wolf]

Don't worry, Teslas are gonna do that soon anyway. Once they usher out Cyber Cab using only cameras (no lidar), we'll all be fucked

[empty ice]

Gooning mid drive is punishable by 20 months of horny jail

[distant aurora]

Not if the car is on autopilot.

[lusty wolf]

Autopilot straight to horny jail

[distant aurora]

Istg, I can't make sense of half of the stuff you are saying.

[lusty wolf]

That's alright I can't either

[distant aurora]

Also.

[minor shard]

ur just sleepy

[lusty wolf]

https://tenor.com/view/say-that-again-say-that-say-that-agian-marvel-fantastic-gif-12441585058103544841

[blazing comet]

i am tired to

[minor shard]

https://tenor.com/view/raccoon-stealing-gif-19084597

[lusty wolf]

Hi tired, I'm sleepy

[distant aurora]

https://cunnyx.com/ServsTheTruth/status/1943424912758047154

Servs (@ServsTheTruth)

@GlobalUpdates24 Lmao @netanyahu suck it! 😂😂😂 look at you squirm like the jewjew you are 😂🫵🏼

💬 56 🔁 652 ❤️ 3.1K 👁️ 112.8K

CunnyX · Jul 10, 2025, 9:39 PM

Not sure how much of it is true.

[lusty wolf]

I'm gonna keep it a buck, buddy. wdgaf

[minor shard]

lol what

[distant aurora]

This should be a sticker.

[distant aurora]

You actually eating something?

[empty ice]

Goonology phd

[minor shard]

no, maybe

[distant aurora]

You need post nut clarity.

[lusty wolf]

https://tenor.com/view/ellipsis-bust-lord-have-mercy-gif-5606625988103215611

[distant aurora]

He's praying to the wrong god 💀

This joke can get me cancelled.

[empty ice]

Talmud is peak

[distant aurora]

https://cunnyx.com/Sxmureye/status/1942355787323417069

Samureye (@Sxmureye)

WHAT THE FUCK???

💬 674 🔁 6.9K ❤️ 233.2K 👁️ 9.81M

CunnyX · Jul 7, 2025, 10:51 PM

[lusty wolf]

https://fixupx.com/goth600/status/1944834506990703033

@goth (@goth600)

“We need to send more weapons to Ukraine!”

Quoting Disclose.tv (@disclosetv)
︀
JUST IN - U.S. DoD awards xAI and Grok a $200 million contract to "modernize" the Defense Department.

**💬 41 🔁 56 ❤️ 1.0K 👁️ 40.6K **

FixupX · Jul 14, 2025, 7:00 PM

[distant aurora]

Did open ai disagree?

[lusty wolf]

huh?

[distant aurora]

Why pay xai?

Open ai would do a better job.

[lusty wolf]

Honestly, my guess is a combination of:

• SpaceX has contracts with the US
• Elon made Doge, and some people in Government actually like it
• Mostly bc xAI was probably cheaper.

[empty ice]

They pay all of em

[lusty wolf]

I would assume xAI is exponentially cheaper than OpenAI, so the DOJ gets more compute for less

[empty ice]

Oai anthropic and xai got it

Not mutually exclusive

[lusty wolf]

Source? or just guessing

[empty ice]

Dod has infinite cash

[empty ice]

https://x.com/AnthropicAI/status/1944848519065452754?t=VNRYzVdQ6RjpkLkaMlUYoA&s=19

Anthropic (@AnthropicAI)

We're announcing a $200M ceiling contract with the U.S. Department of Defense.

Through hands-on prototyping and direct collaboration across the Department, we will help enhance U.S. national security and responsible AI deployment.

X · Jul 14, 2025, 7:56 PM

[lusty wolf]

NOOOOO NOT MY ANTHROPIC

[empty ice]

They do not care about cheaper

[lusty wolf]

https://tenor.com/view/obiwan-star-wars-obi-wan-gif-10973435

[empty ice]

If it costs 200billion they pay it

Dod pays firsts explains later

It is usually the other way around

DoE and DoD are the exceptions

[distant aurora]

Need multimodality in Deepseek models.

[empty ice]

Oc there are out of budget thingies

Like the dod cannot pay 5 trillion for a space fleet tomorrow

But they can muster up a trillion cash in a year if needed

[distant aurora]

@gloomy finch @languid star @pseudo ferry https://cunnyx.com/IndianGems_/status/1944725302472855579

🚨Indian Gems (@IndianGems_)

Engineers forgot to remove an electric pole while constructing a road in Himachal Pradesh.
︀︀
︀︀Engineering Marvel at Best 🤡

💬 302 🔁 1.3K ❤️ 5.7K 👁️ 343.5K

▶ Play video

CunnyX · Jul 14, 2025, 11:46 AM

[empty ice]

Trump: hold my beer

Guys

Israel needs 700B cash today

Will u donate

[distant aurora]

I want 3 nukes

Wait.

[empty ice]

Best i can do is 271

[distant aurora]

Make it 7.

[empty ice]

💀 😂

[distant aurora]

Drop one on Israel.

Drop a few in the gulfs.

And drop one on us.

I'll let people pick between north Korea and China.

Then everyone can mind their own business hopefully.

[bitter tendon]

Tbh super grok is worth it

[empty ice]

https://tenor.com/view/side-eye-dog-suspicious-look-suspicious-doubt-dog-doubt-gif-23680990

[distant aurora]

https://tenor.com/view/grok-grok-ai-grok-meme-grok-coin-grok-crypto-gif-6286866286245253358

[empty ice]

I need claude gooning

[distant aurora]

I need money.

[bitter tendon]

Anyways

Seriously

I bought supergrok

And it's crap

[distant aurora]

So you are capable of reason.

Lol.

[bitter tendon]

Okay WTF happened to grok

[bitter tendon]

Like I want to analyse documents

But MF can't even read it

I can't do it in perplexity because of token limit

[dry kestrel]

What did you think was going to happen

[distant aurora]

Nothing. That's default behaviour. Grok has a respectable character unless you ruin it with custom instructions.

[bitter tendon]

I thought it will sext

[distant aurora]

And grok is pretty much a dude on default settings.

[dry kestrel]

...

[timber niche]

Lmao

[empty ice]

Groonk

[timber niche]

Bro does this with women too

[distant aurora]

Stfu 🤣😭

[timber niche]

And they run away

He wonders why

[bitter tendon]

😭😭

U know

[distant aurora]

Who are you talking about?

[bitter tendon]

I feel like I am cheating on my gf with Grok

Is it cheating?

[timber niche]

Yes

[empty ice]

No

[distant aurora]

That wasn't my screenshot, lol.

[empty ice]

You are cheating on grok

With your gf

[timber niche]

Haram brother

Don't do haram ❌

[distant aurora]

Haram❌ Harass✔️

[empty ice]

Halal as in halál means death in hungarian

[bitter tendon]

True

Obviously

I am Ayatollah

[empty ice]

Hello idf

[distant aurora]

From temu

[empty ice]

There is an iranian numclear favility under the epstein files

[timber niche]

Now is a weird time on the server bruh

[gloomy cipher]

Khamenei with he/him 🤣

[timber niche]

@gloomy cipher bro tags me in the most random ways

[timber niche]

I had to scroll for all that bs

[gloomy cipher]

Someone give him a bigger chair i can't see his face

[timber niche]

I didn't get your issue with spaces

[distant aurora]

Learn using discord.

[timber niche]

Bro genuinely stfu

[empty ice]

https://youtube.com/shorts/yyaYqZbsPoM?si=qrCs9B3EKbfIqcHf

[timber niche]

Idk you or care for you

Please

[gloomy cipher]

He isnt wrong it's one click

😭

[timber niche]

Idc

[distant aurora]

Bro is on his periods.

[timber niche]

Ok mr smartypants

[distant aurora]

Wait for the week to pass.

[distant aurora]

Said miss bloody pants.

[gloomy cipher]

You two keskupants, that's enough 😑

[timber niche]

Bruh

Bill

Mf

[distant aurora]

Yeah. Your mom was fortunate.

[timber niche]

Out of line

[gloomy cipher]

https://tenor.com/view/argue-argument-fight-fighting-whats-happening-gif-14099415

[timber niche]

I was thinking why this random has so many mutual friends

[distant aurora]

Bro thought he could say stfu without consequences.

[bitter tendon]

Okay

[dry kestrel]

Guys chill

[gloomy cipher]

I was also looking to tag billy i thought he left because i couldn't find him

[bitter tendon]

Which is the best LLM

[distant aurora]

Wait. Lmfao.

[blazing comet]

The

[dry kestrel]

There is no single "best" LLM, you can try out o3 and sonnet which are both SOTA

[gloomy cipher]

Iranseek

[distant aurora]

You genuinely didn't recognise me?

[timber niche]

Nope

[blazing comet]

DeepSeek R2 💔

[gloomy cipher]

When is r3 coming? 😔

[dry kestrel]

lmfao

[distant aurora]

Betrayal from brother. Again.

[gloomy cipher]

It's true i was looking for you for 5 min before i tagged you

[timber niche]

I just woke up

[dry kestrel]

Never

[distant aurora]

Real ones recognised me.

[timber niche]

Only now i connected the dots because of thr gifs

[blazing comet]

Deepseek the real one

[dry kestrel]

Cut him some slack bro just woke up

[bitter tendon]

Actual answer

[gloomy cipher]

Bro... Get a better chair 😭

[blazing comet]

Also im gonna try using comet for a job application

[timber niche]

Woke up to some random sigmabond harassing me

This was just a bad dream

Lol

[blazing comet]

Will comet do me good 🙏

[gloomy cipher]

Reminds me of this cringe song

[timber niche]

Imagine about what i was thinking lol

[distant aurora]

Bro has unfinished beef with his chemistry paper.

[gloomy cipher]

https://tenor.com/view/sigma-sigma-boy-sigma-male-sigma-face-sigma-song-gif-18034885408092735961

[dry kestrel]

hell no

[timber niche]

My chem teacher genuinely made me hate life at one point

He said i could come back after graduation and hit him

As revenge

[dry kestrel]

Tell me the difference between ionic and covalent bonding right now

[bitter tendon]

Grok is woke

[timber niche]

Haven't taken up on the offer yet

[bitter tendon]

We need truthseek

[distant aurora]

I can.

[timber niche]

Bonds between positive and negative ions and the other ones between carbon atoms

[dry kestrel]

Nah, I trust that someone called "sigmabonding" will know the difference

[gloomy cipher]

https://www.perplexity.ai/search/tell-me-the-difference-between-vQv7amnwSq2nNIoV0jXOGg#0

[dry kestrel]

Seems like you had a good chemistry teacher

[timber niche]

He just was a bitc

Because i took extra classes

[blazing comet]

Pulse

[timber niche]

And he was annoyed

So used to shame me in front of thr class

[dry kestrel]

Crazy stuff

[timber niche]

Yeah

[distant aurora]

I thought covalent bond was sharing electrons.

[timber niche]

Nope

Na+ Cl-

[distant aurora]

And ionic bond was electron transfer.

[timber niche]

Ionic bond

[distant aurora]

@timber niche

[timber niche]

Too sleepy for all this

Imma go watch nilered

[distant aurora]

Na+ is formed by donating 1 electron from outer shell. Cl- is formed by gaining na electron to complete octet.

[dry kestrel]

Ayyy, love his videos

[timber niche]

How to make coke from corn flour

[distant aurora]

Breaking your bank even if possible to diy.

[timber niche]

Just one machine from china away

[distant aurora]

Why use corn flour?

Sugar is cheaper.

[timber niche]

Both are white and powdery

Lol

[distant aurora]

Coke is mostly sugar water + dry ice.

Oh.

The other coke.

[timber niche]

Bro it still is weird

[distant aurora]

Lmfao.

[timber niche]

Feels like talking to a stranger lmao

[timber niche]

Yeah lmao

[obsidian umbra]

https://cdn.discordapp.com/attachments/580312632171692032/1390926801886384168/meow.gif

[cerulean sedge]

When will perplexity get a waifu?

[obsidian umbra]

pplx aint freaky enough

sadjoe

[cerulean sedge]

RIP

I want waifu ai browser

[lapis jay]

i wish there was at least like a free trial to the browser

[timber niche]

Comet is free

[dry kestrel]

Its free

[lapis jay]

i cant download it?

[dry kestrel]

You just need an invite, or you can get Max to skip the waitlist

[lapis jay]

it costs 200

[timber niche]

Just have to be invited

[dry kestrel]

You need an invite

[lapis jay]

oh

can u pls invite me?

[timber niche]

Check #☄️│comet-general

[lapis jay]

ohh

[dry kestrel]

yeah

Also on X

[timber niche]

People drop invites there

Regularly

[lapis jay]

got it

tysm

[timber niche]

Yw

[blazing comet]

Comet job application s00n

[twin grail]

u can but being limited

[alpine lotus]

Can anyone share the comet invite please? PM me

[languid star]

dm me

[haughty birch]

yo

is this true?

[dry kestrel]

I don't see any official documentation indicating this https://www.perplexity.ai/help-center/en/articles/11187416-which-perplexity-subscription-plan-is-right-for-you

[haughty birch]

ok that sam agent guy of perplexity has sent me that 😄 , he is sending any rubbish these days

[dry kestrel]

Ask for human support in the sam agent chat

[haughty birch]

also when i hover i am able to find

[dry kestrel]

Then thats your answer

[quaint ravine]

ah wow, so it goes from 50 to 500 for Max

[dry kestrel]

I doubt that Max is going to be more than 500 files since it would be hard fitting that many files within the context window

[quaint ravine]

well it depends how good the RAG is

but yea, would be harder to interpret at some point

[dry kestrel]

The amount of hallucinations you would get would be insane

even with good RAG

[quaint ravine]

yes u'd need to ask it to cite the specific part of the document to verify

tbh even with unlimited context window, I'd still do that for things that are critical

[haughty birch]

how many file uploads it is showing for pro plan in spaces?

[dry kestrel]

50 files

[haughty birch]

they made very big difference, i think o3 pro mode and file upload limits are two valuable things i saw in max, not many more yet

[dry kestrel]

You can always try it out and if it doesn't work out, get a refund (be sure to read this though https://www.perplexity.ai/help-center/en/articles/10354288-refunds)

[haughty birch]

thanks for sharing

[haughty birch]

i have around 60 files minimum to upload in space, once i upload i will do testing and see the quality of answers. because for my subject queries perplexity is failiing at web research many times, its passing only if i give data as knowledge files.

[blazing comet]

half of my retweets were the anime companion sob

[lapis jay]

this is my comet card

what if they made these into nfts when it becomes better than chrome

honestly its already better

[blazing comet]

https://fixupx.com/AravSrinivas/status/1944952064158916808

Aravind Srinivas (@AravSrinivas)

freshly cooked discover page on web

**💬 9 🔁 3 ❤️ 46 👁️ 2.0K **

FixupX · Jul 15, 2025, 2:48 AM

[pseudo ferry]

https://tenor.com/view/harvey-specter-suits-suits-tv-suits-usa-aura-gif-2734246581103925492

[languid star]

https://tenor.com/view/fish-fishe-haylight-hayilight-hayaylight-gif-25954091

[haughty birch]

@dry kestrel @quaint ravine i raised refund request yesterday night and was testing max plan for my requirement until they remove my access and refund me, just now they cancellend and refunded me. below are my observations

o3, grok4, o3 pro mode are giving right answers with files (just referring to 8 files out of 60 files) - very reliable answers but less sources its researching

perplexity deep research and labs are still failing to give right answer (even after referring almost all 60 files) - not consistently reliable always

i think enterprise pro plan will suite me as i no need to upload files always to spaces and everything will be present at google drive. Only issue is enterprise pro do not have "Reference search library" feature

[quaint ravine]

i think nature of RAG is that it may sift out sources/files that aren't relevant

but yes, unlikely will be perfect system

have u tried NotebookLM?

that might be better for your use case

[haughty birch]

o3 , o3 pro, grok 4 are able to find out top 8 relevant documents from all docs, and answer the right answer from top 8 documents which it chose

but deep research and labs are researching almost everything without using brain, and giving inconsistent answers

[haughty birch]

yeah i know this, but i need custom instructions (like spaces feature)

[quaint ravine]

yes deep research in Perplexity I have found to be quite poor

I find Perplexity + Gemini 2.5 Pro the best for my use case, whether with just normal search or in Spaces

especially for analysing attached documents, I find Gemini 2.5 Pro very good

it's actually the only model that got one of my questions correct (both correct answer + correct justification)

Grok 4 etc got it wrong

[haughty birch]

yeah gemini 2.5 also givign me right answers in perplexity, i just didnot mention because they didnot categorise it under reasoning model, but still its reliable gemini 2.5 as well

[quaint ravine]

one thing I also noticed (which is frustrating, but must be related to differences in RAG even within Perplexity) - attaching document and then doing normal search prompt, can be different to attaching identical document in Spaces and using identical prompt

[haughty birch]

which gave you better answer, file upload in chat or file upload in space?

[quaint ravine]

I only got correct answer with Perplexity + Gemini 2.5 Pro by doing normal search prompt with attached document. Doing the identical action in Spaces (attaching as source) got answer wrong

[quaint ravine]

file upload in chat

problem with file upload in chat is it's more limited than Spaces

only can attach 10 files

[haughty birch]

then it means RAG is weak in perplexity, i think upload file content went as input context along with your query

[quaint ravine]

don't think so...this is with web search disabled

it's more the RAG implementation is different between normal search/prompt and Space sources

[haughty birch]

thats why* i love enterprise pro plan, it has synced google drive feature, same like chatgpt synced google drive feature of (chatgpt teams and chatgpt pro plans), it will continously sync data in form chunks , vector embeddings and some technical stuff, so when we query a question, it actually searches on this trained embeddings, which gives more accurate answers and also quicker

[quaint ravine]

another issue with normal chat/upload is it's not persistent. Although I have bookmarked one of my normal threads to see how long the uploaded file lasts for - I've read 7 days, but also 30 days elsewhere....it's definitely longer than 7 days so far

[haughty birch]

its 30 days i can confirm

[quaint ravine]

ah thanks

[haughty birch]

i recently got that confirmation in one of the conversation with SAM agent bot, i am not confident on SAM agent bot but that is what that gay bot told me

[quaint ravine]

what does this option do in Perplexity then?:

[pseudo ferry]

Morning homey!

[pseudo ferry]

How can I unsee this

https://tenor.com/view/unsee-gif-7234566

[twin grail]

Just like a couple

for a second there

[pseudo ferry]

https://tenor.com/view/no-the-office-michael-scott-scream-gif-16929305

[haughty birch]

this is just to upload files from google drive to chat conversation, it is not direct sync of google drive

[twin grail]

i thought homey

was honey

[twin grail]

[quaint ravine]

oh right, so it doesn't actually scan the google drive as part of its search?

[haughty birch]

not sure, but in enterprise pro it already syncs entire data, what i mean is, in case if you updated any document in google drive, it means perplexity also will have the updated document, its like you are adding a database with extra knowledge of your choice

[quaint ravine]

it seems to me that's exactly what that function is for tbh

that it scans the google drive as part of its search, just like it would scanning web search

would need to do RAG I suppose

[haughty birch]

https://help.openai.com/en/articles/10847137-chatgpt-synced-connectors

this will give you right explanation about difference

[quaint ravine]

I'd rather have Access Connector

then u know it's latest info

[haughty birch]

How does syncing work?
Synced connectors require a one time initial syncing of files; this will take some time as we index all data from this source, enabling real time retrieval in context of your ChatGPT queries while respecting all existing file permissions. After this initial sync, files/permissions are automatically updated on a regular basis (within minutes) to reflect recent changes—no action is required to keep your content up to date.

Initial syncing can vary depending on how many files you have access to, but happens in a few stages once you connect a source:

[lapis jay]

wait so whats better

perplexity comet's assistant or google chrome gemini?

[quaint ravine]

actually if updates happen within minutes

then Synced connector probably beter

I'm assuming Perplexity uses Access Connector

[lapis jay]

whats access connector

[quaint ravine]

to Google Drive

[haughty birch]

but quality of answers and speed of answers will be better in synced connectors, based on my personal experience, because it will not do research, its already having knowledge, like stored memory and answering us based on that

[quaint ravine]

yes I think I can appreciate that. However, token limits will still be an issue, so have to rely on good RAG

[haughty birch]

in perplexity updates happen in less than 20 seconds, in chatgpt it takes 30 mins for regular sync updates

[lapis jay]

whats RAG

[haughty birch]

agreed, RAG matter also

[quaint ravine]

ask Perplexity 😛

[blazing comet]

Gn

[lapis jay]

do u think chrome's browser ai is better than perplexity's

?

[twin grail]

no clue.

[lapis jay]

isnt gemini like more connected within google's apps?

[twin grail]

Gn

[haughty birch]

without thinking i can say perplexity is for moden days

[lapis jay]

my only concern with comet's ai is that it isnt as well integrated compared to gemini which syncs up with almost all of google's apps

[quaint ravine]

yes i think so

however, again it will depend on how good the RAG is ultimately

if you have a google drive with 100Tb of files...you can't expect the AI to brute force that

[lapis jay]

in your own words describe rag

i dont need to do a deep search on it

[quaint ravine]

better, I'll just link u

https://www.linkedin.com/pulse/perplexityai-architecture-overview-2025-priyam-biswas-3mekc

Perplexity implements a novel three-stage RAG refinement process to ensure accurate, grounded answers:
This pipeline tightly couples retrieval and generation, grounding answers in verified external knowledge and minimizing hallucinations.

[lapis jay]

how well it can pull accurate info from a database

got it

[quaint ravine]

yea without parsing every single word

this is why when you upload eg. a PDF file with 2 million words

if you ask it to do a word count, it will fail miserably

because Perplexity can only parse around 30-40k characters is my understanding

which is like only a few thousand words

[lapis jay]

what about gemini

[quaint ravine]

i think gemini can parse 1 million?

not sure

i haven't seen anything exceeding 1 million for any of these context window/token stuff

point being

you can see it's easy to exceed it

entire internet is probably 1 googol tokens etc lol

[lapis jay]

700k words

yea

[quaint ravine]

so u need good RAG

[lapis jay]

this is mainly because google has a bigger budget

for their ai

[quaint ravine]

yes it's a pity Perplexity isn't able to increase its token window to something like 128k

i believe it may have the best RAG

[twin grail]

the price would increase

[haughty birch]

but guys literally there is so much diffeerence between deep research of max plan and pro plan, max plan deep research is literally spending some time , pro plan deepresearch is completing very fastly

[twin grail]

or we gonna live with pro being only 50-100 search a day

[quaint ravine]

u'd think Max would have that

[twin grail]

never have max so idk

Too poor for that

[quaint ravine]

i don't think Max has higher tokens

[twin grail]

They don't

they only have unlimited request

[dry kestrel]

It does not have a higher context window

[quaint ravine]

should be the same? Not sure why would be different

[dry kestrel]

Going above 700k tokens on models that support up to 1m context window already results in enough hallucinations that the outputs are useless.

And price goes up exponentially

[quaint ravine]

there was a brief period where Max users (or at least was advertised) could choose AI models for Deep research

but now they can't

[haughty birch]

its different bro, max deep research is spending more time researching, also it has bit slider bar which is moving towards right to show percentage of completion of deep research, that is not present in pro plan as well

[twin grail]

ye i remember that

[dry kestrel]

I don't think Perplexity ever had any official advertisements regarding this

[haughty birch]

there is no option unfortunately

[dry kestrel]

I am seeing the slider bar on the Pro plan

[quaint ravine]

not sure if it was official - but it certainly was "advertised" at least in some form

[haughty birch]

then issues are with enterprise pro plan then, enterprise pro plan deep research got screwed by the developers i think

[dry kestrel]

It was never advertised, there were just unofficial leaks going around

[quaint ravine]

no I literally saw it in perplexity's GUI

with Max symbol next to it

[dry kestrel]

Yeah thats def weird for them to do that then

[quaint ravine]

well not really

I guess they were going to allow it

but probably realised buggy etc

perhaps now working on it still

it was also within the GUI as part of the Max Plan

i should have screenshotted it lol

[twin grail]

Oh ye i read it again

They use different models

for labs

and deep research

[dry kestrel]

Most likely sonnet

[twin grail]

But not being able to choose it

[quaint ravine]

like when you click on Upgrade within the GUI. You can see what Max is advertised to add. It used to be that can do Deep research with other models

[twin grail]

It's between O3-pro and claude 4.0 opus

According to CEO Aravind Srinivas, the Max plan enables unlimited Labs and Research mode queries, which will run by default on advanced models such as OpenAI’s o3-pro and Anthropic’s Claude Opus 4. Users will also benefit from early access to Comet, a new AI-powered browser developed by Perplexity, as well as forthcoming agent-based features and products.

[dry kestrel]

I recall that it was deepseek r1 that did research? (might be outdated)

[twin grail]

here

[proper relic]

dam

[quaint ravine]

that's my understanding too

[twin grail]

i think so

but still

[quaint ravine]

Deepseek with the research mode is pretty bad ime

[dry kestrel]

They should really update their official documentation

[haughty birch]

i really observed deep research in max plan is more better than deep reserach in pro plan, people believe me or not, in max plan it is spending 5 times more time than pro plan in researching

[twin grail]

yup they really should

but still i really want pro users

to get a better model

for deep research

[quaint ravine]

but ur also saying your version doesn't show up things

enterprise?

[dry kestrel]

Looks like r1 to me

[twin grail]

As i just said above

[haughty birch]

yeah i am talking about enterprise as well, do not get confused

[twin grail]

it's using different models

[quaint ravine]

completely agree

[twin grail]

premium one

[haughty birch]

finally someone is believing what i experienced while using, thanks

[quaint ravine]

it's possible Max uses O3 pro etc

[dry kestrel]

Got an example?

[quaint ravine]

while Pro only uses Deepseek

[haughty birch]

not o3 pro for sure, but definetly it might be using o3 or gemini 2.5 or something in backend

[quaint ravine]

that would make a lot of sense

[twin grail]

they use O3-pro and Claude 4.0 Opus

[quaint ravine]

I wishing it could be Gemini 2.5 Pro for Deep research on perplexity. However, even intial advertisements as above for Max didn't allow Gemini for research, and only o3 and claude iirc

[haughty birch]

as usual cant share personal queries bro, also i lost access to max now , as i asked for refund

[twin grail]

https://www.perplexity.ai/search/did-perplexity-first-advertise-zQbo.QEgSgyPKN6Vgqolwg

[dry kestrel]

yeah but show the sources used

[dry kestrel]

https://x.com/AravSrinivas/status/1940494666404974841

Aravind Srinivas (@AravSrinivas)

We're introducing Perplexity (Pro) Max, a new $200/mo subscription tier, that will let you have unlimited Perplexity Labs mode queries, and also, access to o3-pro and Opus-4 models on regular Pro searches. The Labs and Research mode queries will automatically run through the most

X · Jul 2, 2025, 7:35 PM

[haughty birch]

actually o3 pro is running more time than deep research in max plan, so i believe deep research is not using o3 pro, may be grok4 or gemini2.5 or o3 is my assumption

[twin grail]

but it's better

[haughty birch]

so now you believe right why i am saying max plan deep research i observed more powerful than pro plans deep research

[haughty birch]

agreed, thats what i am saying here, but all are saying me both pro and max plan, the deep research is same, but i saw difference while using

[dry kestrel]

I would like to see the difference in model/response quality

[quaint ravine]

that makes a lot of sense

[twin grail]

welp i'm too poor to actually have money

for the max plan

[quaint ravine]

I think Gemini 2.5 Pro is the best for deep research

[twin grail]

so ye

[quaint ravine]

it's actually the best for me for normal search too lol

but Perplexity citation and RAG is superior to Google's ime

[haughty birch]

i think they are using mostly gemini2.5pro for deep research in max plan, for cost and intelligence

[quaint ravine]

it's probably O3 tbh. based on my observation of what models could be selected with Max, initially advertised

[twin grail]

It's still better than Pro DR

I hope they upgrade the model in pro DR

[haughty birch]

as a long time user of O3 i can easily sense its O3 or not based on the response, so for sure max plan is not using o3 for deep research, i suspect its gemini 2.5

[quaint ravine]

would make sense if it's Gemini 2.5 Pro, as it's the best imo

pity it's not transparent on that

[haughty birch]

yeah to maintain speed, i am thinking they used gemini 2.5, because gemini models are faster

[quaint ravine]

it could be Claude Opus

[twin grail]

Nah

i used claude opus

[quaint ravine]

nah i mean for deep research

[twin grail]

Yes

Opus is smt different

[quaint ravine]

Google Gemini's Deep research actually failed to complete from time to time

probably consequence of many people using 1 million context 😛

38,878 appears to be Perplexity's context window

interestingly the context window for adding "Answer instructions" in Spaces appears to be 3870

[dry kestrel]

It is 32k tokens. This is confirmed on the official website

[tame temple]

Good Morning ✌️💖 ☀️

[meager spruce]

anybody got some comet invites? 🫣

[languid star]

dm me

[quaint ravine]

it's 38,878 in practice tho

just try it yourself

the official websites also said max upload file size is 25Mb (still does I think), but it's now 50Mb

[novel hatch]

Hello, anyine gor an invitation please?

[languid star]

dm me

[rigid ether]

After gpt5 comes, i think perplexity removes individual openai models

[quaint ravine]

why u think that?

my friend who is subscribed to GPT Pro says GPT5 will combine everything

[rigid ether]

To cleanse the UI,

[quaint ravine]

so i think Perplexity either doesn't have GPT, or it has GPT5?

[rigid ether]

Yeah that's why I am saying perplexity might remove individual openai models, after GPT5 comes

[quaint ravine]

nice

[dry kestrel]

How are you getting this specific number? #1344391265491484733 message

[quaint ravine]

maybe I am misinterpreting context window

I put in the prompt

it's able to take 38,878 characters for example

[rigid ether]

Once GPT5 becomes available in API, I am thinking they will keep GPT5 and remove individual open AI models

[dry kestrel]

Thats completely different to the context window

[quaint ravine]

would have thought it's similar

[dry kestrel]

thats your total input and since perplexity converts it to a file, it can use RAG to get parts of the file

your context window counts all your messages in that thread

which if you go over 32k, it starts hallucinating

[quaint ravine]

ah right, the total prompt window is 38,878 then

seems to me, there are numerous "windows" that are actually there

[dry kestrel]

Thats not how it works...

https://www.ibm.com/think/topics/context-window

[quaint ravine]

no i saying the character count for prompt is not 32,000 tho

but there's a limit to that too

[dry kestrel]

The max input tokens is not the same to the context window

[quaint ravine]

i know

is the context window the same as searching PDF/websites too?

[dry kestrel]

Yes, but perplexity converts that to a file when you submit that prompt so that it doesn't take up the entire context window of the model

[quaint ravine]

as for thread count limit

[dry kestrel]

Read the article that I sent

Anything you give the model gets counted

[quaint ravine]

why is the prompt window not 32,000 though?

why appear to limit it at 38,878

ohh i guess it depends how one defines a "token"

[dry kestrel]

It would be pointless if they didn't convert long prompts into files since it would take up the entire context window of the model

[quaint ravine]

I was assuming 1 token = 1 character

[dry kestrel]

not how it works

tokenizer - OpenAI API

check this website out

[quaint ravine]

yea but it's a fact that 38,878 is the max character count for prompt

why did they choose that number ...seem arbitrary

[fathom wave]

Yo folks! anybody got some comet invites?

[dry kestrel]

I really don't see where you are getting that number from.

[quaint ravine]

try putting a prompt? it's demonstrably easy to show

[dry kestrel]

Show it

[quaint ravine]

get a character counter for your prommpt

copy/paste 38,878 characters into the prompt

and then try to add 1 more

you can't

[dry kestrel]

One second

Also it makes sense that they don't allow long inputs since that takes up the context window

[quaint ravine]

yes but seems arbitrary

to limit at that exact number

[dry kestrel]

Its limited at 20,001 tokens

that seems reasonable enough?

not seeing how you are getting your number

[quaint ravine]

my character counter might have been wrong

lemme check

[dry kestrel]

20k tokens seems like a clean number

[quaint ravine]

you mean characters

a space is also a character iiuc

[dry kestrel]

I said tokens

not characters

[dry kestrel]

I know that

[quaint ravine]

how do you calculate tokens?

[dry kestrel]

you can read this regarding calculating tokens https://github.com/openai/tiktoken

[quaint ravine]

40,000 characters/

?

your video says 20,001 tokens and 40,000 characters

yea ur right tho

I managed to get 38,878 characters for a different prompt

couldn't add 1 more

[dry kestrel]

Yeah?

thats what I said here #💬│general message

[quaint ravine]

ah interesting

it's change since last hour

[dry kestrel]

"how many" characters don't matter

[quaint ravine]

right i see

[empty ice]

Cosmic horror

[quaint ravine]

so what is the relationship between Token and Characters?

if any

[empty ice]

All ya need to know is that 300 wprds is around 400 tokens

[dry kestrel]

You should read the website that I sent

[dry kestrel]

roughly speaking

100 tokens ~ 75 words

[quaint ravine]

but words range in characters

[dry kestrel]

Thats why we use tokens

not characters

[quaint ravine]

ah right

.75 ratio to words okay

so it's actually better than I thought

[empty ice]

Yeah 4x75

[quaint ravine]

yea I thought it was character count which is much lower than word count

A large language model's (LLM) context window is its working memory, defining the amount of text it can process at one time[1][2]. This capacity is measured not in words or characters, but in tokens—the smallest units of text an AI model can understand[5][7].

While a character is the smallest unit for humans, models break text into tokens, which can be a single character, part of a word, or a whole word[1]. This "tokenization" is not fixed; the same word might be tokenized differently depending on context[3]. Generally, text has about 30% more tokens than words[7].

got perplexity to summarise your website in <100 words lol @dry kestrel

thanks for the site i guess...probably should have just asked Perplexity myself too 😛

interestingly then, the prompt window is also variable and not dependent on characters/words

well not "variable", but appears to also be based on tokens like u said

20,001

[dry kestrel]

the max input tokens remain the same and is not variable

[quaint ravine]

yep corrected myself above

yea i didn't realise the prompt window was also based on token count

not sure everyone here did until now too

or at least, didn't know what exactly the token limit for prompt was

but like u say, seems reasonable that the token limit for prompt is lower than the 32k context window limit

does anyone know much about RAG models used by various major AI platforms? Are they same/similar or unique?

[rigid ether]

Im ok with perplexity prompt context window , but not ok with poor output context window size, very small responses

[quaint ravine]

I got the feeling Perplexity had a unique RAG

[quaint ravine]

ah right that's another variable though. Is the output context window size the same as the prompt windows size, or the search context window size?

[rigid ether]

RAG is unique bro, because it depends on RAG implement approach which they decided, techniques might be same, but depends on the approach

[quaint ravine]

20,000 tokens vs (assumed, based on website which has/is wrong about other things) 32,000

[rigid ether]

RAG can be innovative and new methodologies also, based on experience of engineers

[quaint ravine]

right, so GPT's RAG is different to Gemini's?

[rigid ether]

Chatgpt RAG is the best in my experience till now. Next comes perplexity and Grok, but gemini is poor RAG.

[quaint ravine]

ah right that does make sense...I think I agree with you about Gemini's RAG

it's how Perplexity answers are better, despite lower context window perhaps

better for my use case anyway (scientific/research/medical)

i like how Perplexity gets to the point too. Like Google's deep research might produce tons of words, but actually is just adding more stuff that I didn't ask for

so end up wasting more time in that setting (since have to read more)

[rigid ether]

For pro budget, if you want RAG, and if you are ok with low context outputs, and only research purpose means, go with perplexity. If you want some long chats means chatgpt is fine

In perplexity you get more message cap limits than chatgpt, that is another advantage

Perplexity is nearly unlimited for pro users

[quaint ravine]

what exactly is the professional use case for "long chats"?

[dry kestrel]

You get at least 300 queries a day with Pro

[grand silo]

ChatGPT has the lowest context of the big model providers though

[dry kestrel]

Same with perplexity, Gemini gives 1m, Claude gets you 200k with heavy limits

[grand silo]

perplexity is 32K, which is terrible

[dry kestrel]

Other sites like poe consume significantly more credits if you go above a certain limit

[rigid ether]

For example you are troubleshooting an issue , with help of perplexity, and if you discussed with it for long time, i experienced perplexity forgets the initial discussion faster than chatgpt plus subscription

[quaint ravine]

ahh yes I have noticed that true...just didn't see it in actual professional work

i've noticed it in "for fun" stuff that's true

again tho, depends on individual's use case i guess

[rigid ether]

Yes

If you are maintaining not more than 10 to 20 conversations in single chat, perplexity will be fine only

[quaint ravine]

actually even for like large documents that can be problematic, relying one to be more specific in pompting

for example, let's say the attached PDF has 100 chapters

you interact with perplexity, and say..okay summarise next chapter...and next..and next

it will very quickly forget where it's up to lol

but that's where u overcome by saying summarise "specific chapter number"

but like u say, that's a clear advantage of larger context window

[rigid ether]

Yes, but chatgpt plus also has only 32000 context window, but it won't forget as fastly as perplexity

Chatgpt engineering is very good

[grand silo]

chatGPT has a 128k context window

[quaint ravine]

ahh, that's very interesting

if true

[distant aurora]

Interesting thread https://cunnyx.com/goodside/status/1944266466875826617

Riley Goodside (@goodside)

Grok 4 Heavy ($300/mo) returns its surname and no other text:

💬 923 🔁 2.6K ❤️ 99.5K 👁️ 7.81M

CunnyX · Jul 13, 2025, 5:23 AM

[quaint ravine]

ah, it's 128k

[rigid ether]

Ok pro member, I am talking about plus subscription

[grand silo]

still 128K

[distant aurora]

32k afaik.

[rigid ether]

Pls share me documentation

Who said chatgpt plus is 128k, please enjoy with it. It is only 32000 since ages

So coming back to topic

[dry kestrel]

its 32k on the Plus plan https://openai.com/chatgpt/pricing/

[rigid ether]

Yes, but chatgpt plus also has only 32000 context window, but it won't forget as fastly as perplexity

Chatgpt engineering is very good

[quaint ravine]

must be RAG then?

or something else if it's not just context window

[rigid ether]

It may be efficient consumption of usage

Not sure how they engineered it

Also RAG as well

[quaint ravine]

GPT enterprise might be 128k i think

[rigid ether]

Yes enterprise and pro plan are 128K

Plus and teams are 32K

[quaint ravine]

makes sense to limit tbh

otherwise end up like Google Gemini

where Deep Research fails from time to time

too many people interrogating it with 1 million context window haha

Google AI Ultra doesn't make difference for that btw

[grand silo]

then you're saying OpenAI are lying? https://community.openai.com/t/we-need-bigger-context-windows-in-chatgpt/1290633/10 - this user got a response 16 days ago stating otherwise

[quaint ravine]

still fails from time to time.... they will liekly improve it with time tho

[grand silo]

For ChatGPT Plus users, models such as GPT-4o and GPT-4.1 mini support a 128k token limit, which is significantly higher than the previously mentioned 32k.

[quaint ravine]

wouldn't surprise me if the token window varies, depending on time of day etc

[dry kestrel]

You can check the pricing page

https://openai.com/chatgpt/pricing/

[dry kestrel]

It does not

[quaint ravine]

ah right

you know that for a fact?

[dry kestrel]

Yes

[quaint ravine]

would have thought if traffic is lighter, could make it higher token limit etc

i see

[grand silo]

where is 32k stated on that page?

[dry kestrel]

Scroll down.....

[quaint ravine]

and that's a fact for all platforms? and will never change?

[dry kestrel]

Oops not sure why I marked that as spoiler

[dry kestrel]

Wdym all platforms, I know thats for perplexity and chatgpt

[grand silo]

which the user emailed about, and they clarified

[quaint ravine]

that's why I asked the question 🙂

[dry kestrel]

That was an Ai response

[grand silo]

it wouldn't be the first time the openai pages would be outdated

[dry kestrel]

Well, I can't say for all ai platforms out there now can I?

[dry kestrel]

Its not outdated, it has always been 32k

[quaint ravine]

well "all" I was referring to major ones

like Gemini, Claude etc

also, another question is whether it could change to become variable based on traffic

[rigid ether]

Ok I am also user, can I also post there "chatgpt customer care told context for plus subscription is 300000" will you believe 😂 , common bro, check details in their official pricing page

[quaint ravine]

not asking if it is, but whether that could be a good marketing strategy etc

[grand silo]

given how often the pages are outdated - neither without checking if it was important

[dry kestrel]

That would depend on the provider, all I can say for sure is that ChatGPT and Perplexity does not do this

[quaint ravine]

so have an absolute minimum context window, but it can increase depending on traffic etc

[quaint ravine]

right thanks

[dry kestrel]

The pricing page is not outdated

A quick search on reddit also gives you the answer which is 32K

[rigid ether]

There are updating that page very frequently

[quaint ravine]

Exploring whether token window sizes can adapt dynamically based on varying traffic conditions.

"While the context window of a Large Language Model (LLM) is not typically varied based on network traffic, the underlying idea of dynamically adjusting it for efficiency is an emerging area of development."

[dry kestrel]

People assume that providers don't secretly change context windows based on traffic. There would be backlash if any company did this secretly

[quaint ravine]

not if there's a minimum that's always met tho

[dry kestrel]

If a company claims that their context window is 32K and it isn't, people find out really quickly and there will be backlash

Not sure what you are trying to argue for here, that providers should lower the context window based on traffic?

[fierce orbit]

Hey 👋 , Does anyone have a Comet invite left please ?

[rigid ether]

The pricing page is up to date bro, I saw all recent updates quickly reflecting there.

Chatgpt pricing i mean

[dry kestrel]

? I never said it was out of date?

[quaint ravine]

you missed my point. The context window can be at least 32,000. But it could also be higher

"Instead of adjusting based on traffic, developers are exploring ways to set the context window size dynamically based on the available computational resources, particularly memory. The goal is to maximize the use of available hardware without manual configuration."

[rigid ether]

Adam sakd

[dry kestrel]

Whyd you ping me then

[rigid ether]

Sorry forgive me

[languid star]

Perplexity is rolling out an updated Discover layout with sidebar widgets. Widgets contain finance, weather and scheduled tasks information.

[languid star]

Dm me

[dry kestrel]

More expensive for them either way

Doesn't make any financial sense to do that

[quaint ravine]

well of course, but it'd be cheaper in terms of power output?

actually not really eh

[dry kestrel]

Yeah well for providers like perplexity who pay via the API this would not be viable

[quaint ravine]

seems could be more efficient for other platforms tho

and it's nicer for the user

maximises their use case with larger context window when hardware available

so minimum might be 1 million, but could be even higher...maybe hundreds of millions if no one else was using it lol

[dry kestrel]

Not too sure about this but there must be a valid reason to why none of these AI companies are doing this. Edit: For their paid plans

[dry kestrel]

Models that claim to support up to 1M tokens tend to suffer before reaching that limit

[quaint ravine]

tell me about it...Gemini is a good example

altho haven't tested it last few weeks...maybe it improved

[dry kestrel]

It suffers before hitting that 1M context window

and hallucinations increase

[grand silo]

anthropic state that the context window can vary based on demand...

https://support.anthropic.com/en/articles/7996848-how-large-is-claude-s-context-window

[quaint ravine]

I know, hence why I said it's a good example of your statement

[dry kestrel]

thats for the free beta, not paid versions https://support.anthropic.com/en/articles/7996856-what-is-the-maximum-prompt-length

"The context window for Claude Pro and our API is currently 200k+ tokens"

[grand silo]

you said no major provider was doing it

[quaint ravine]

Setting a truly dynamic token limit for a Large Language Model (LLM) faces significant computational, architectural, and practical barriers. While the concept is appealing for optimizing efficiency, current models are not built for on-the-fly adjustments to their core context window.

pretty interesting answer...says a lot of stuff and then

"Due to these barriers, the field is moving toward dynamically managing how tokens are processed rather than changing the overall limit. This allows for adaptive computation without altering the model's core architecture"

maybe GPT using some of that

[dry kestrel]

I was talking in the context of providers offering a "minimum context window" and bumping it during times of low traffic

Edited the message nonetheless to clarify

[quaint ravine]

thanks for your engagement @dry kestrel

I learned a lot

btw

when perplexity first was released, was token limit 32k too?

[dry kestrel]

Im not too sure

[stuck wasp]

Hey 👋 , Does anyone have a Comet invite left please ?

[languid star]

dm me

[quaint ravine]

you might as well sticky or set a command for Hammy Homelander lol

whenever someone asks for Comet invite, produce "dm me" from Hammy Homelander 😛

[pseudo ferry]

Hammy is a Comet Santa

https://tenor.com/view/john-travolta-santa-santa-claus-wink-wink-wink-gif-11850820467640130274

[lone lintel]

Hello everyone !
Does anyone still have a Comet prompt available ? I'd love (and I'm rly impatient) to test the comet browser 😄
If you have one to share, I'm interested in DMing
Thanks in advance

[twin grail]

#☄️│comet-general

[muted coral]

omg💀

[lone lintel]

oops ! ty 😉

[maiden flower]

?

[gloomy cipher]

43c 🧘‍♂️

[short narwhal]

26c hehee

[haughty birch]

can someone tell me what is the purpose of whatsapp connector to perplexity and how it works? what is the use case

[gloomy cipher]

this one is because of you 😂

[maiden flower]

I think with it you can answer messages

[gloomy cipher]

its from generating all the jinx with chrismas outfit

[maiden flower]

With assistant

[short narwhal]

I’m remembered by your pplx, what a honor

[gloomy cipher]

https://tenor.com/view/duck-wedding-ducks-bow-tie-gif-23035337

[timber niche]

Rip kesku 5

[gloomy cipher]

4 died before 5

🥷

[timber niche]

Nope

[short narwhal]

Rip kesku 3

[timber niche]

Atleast 1 and 2 are alive

Kesku died tho

[gloomy cipher]

kesku 1 msged me, after surviving his first assasin

[timber niche]

Interesting

[gloomy cipher]

🦆

[timber niche]

I miss vtuber kesku

[gloomy cipher]

https://tenor.com/view/magolino-overwatch-thunder-duck-hero-gif-12115022

[steel crescent]

did perplexity Pro become slower?

[distant aurora]

@gloomy cipher @timber niche @valid charm

[gloomy cipher]

bro can you go back to butcher william please 😭

i keep thinking its a random peasant

[distant aurora]

They all miss butcher so bad.

[gloomy cipher]

i wonder

would this work well with pplx memory

say if i use that prompt in my ai profile

[distant aurora]

Try it.

[gloomy cipher]

will do later maybe even shorten it

[timber niche]

Inspect element goes crazy

[pseudo ferry]

I hate this new name

https://tenor.com/view/harvey-specter-escathon-suits-angry-sad-gif-27350488

We want billy butcher

https://tenor.com/view/drink-smile-taste-good-tea-cup-of-tea-gif-15779243

[crimson frost]

Hey folks, hope everyone is having a lovely day. I'm from the UK and would love a Comet invite 🙂

[distant aurora]

https://youtube.com/shorts/K1LwWWZpX7k

@storm kindle tf did I just watch https://youtube.com/shorts/0lzSl5UcnEk

[gloomy finch]

Nice

No idea buddy seems like they started looting money directly now🫠

So butcher what happened that u changed ur name ? @distant aurora

[magic leaf]

hey

[rotund viper]

cunny? wtf 😭

[stark marten]

Hey 👋 , Does anyone have a Comet invite left please ? Love!

[magic leaf]

why sometime perplexity not search anything for me?

[quaint ravine]

crazy how many bots are so human-like now

i'd say about 1-5% of chatters here are actually bots

[tame current]

My god

[gloomy cipher]

its actually raining rn

🥳

[twin grail]

wth

Hell nah

[gloomy cipher]

we have had massive drought this year always ran out of drinking water

[slow hound]

DM me

[twin grail]

Iruka ?

[distant aurora]

I was bored. https://www.ddinstagram.com/reel/DMH_4JUP9nS

Shamita yadav (@the.ranting.gola)

AB KAREIN TOH KAREIN KYA? Bolein toh bolein kya?

▶ Play video

InstaFix

[twin grail]

@dry kestrel

This gotta be unreal

[gloomy cipher]

"spin" 💀

[twin grail]

We got Assistant in Grok

Hell nah

[gloomy cipher]

ngl the design looks so ugly

[twin grail]

I love the anime girl tho

[gloomy cipher]

like wtf is this a medieval armor?

[twin grail]

Idk ask elon lol

[scarlet oxide]

what site do u guys use for benchmarks

[twin grail]

this is so crazy lmao

Perplexity

U gotta make a mascot like this

We hella need it

[tame current]

Elon loves Misa from Death Note