💬│general

[fast meteor]

[tame current]

Last time I used it, it was really bad. It would heavily restrict answers and was really frustrating to use. Was a huge turnoff. To me Perplexity gpt 4 isn't bad but ofc opus is the best

[minor shard]

I am liking the new GPT

[warm cave]

Yeah, it all depends on use case, chatGPT sucks for some uses, and is great at others, but Claude 3 seems to be amazing great for most

[wicked parcel]

For one month I will now survive 😉

[warm cave]

That is good! Has it acted extra lazy at all?

[minor shard]

Not in my experience 🙂 Thankfully

[warm cave]

Good, maybe I will like it then 🙂

[bleak oxide]

What do you think of my question? https://www.perplexity.ai/search/I-would-like-pkndIS3lT3eav6VNmtXjHA

[feral marlin]

Did they limit opus in perplexity

[livid mantle]

went from unlimited to 35

[feral marlin]

I’ll just buy Claude.ai on two accounts instead since opus still a lot better for my uses than recent gpt 4

[fast meteor]

only 30 for me

[wicked parcel]

Great option if you dont live in europe. you.com is sooo buggy.. only error messages even on my first prompts

[feral marlin]

I might try chatgpt plus again it’s been a really long time since I used it and I already have a Claude sub regardless

[lunar ivy]

Agree , Can not even Talk

[lone citrus]

Not worth 20$ unless you have plenty budget on subscription. They are still using the low message limit, and it's still slower than opus. The new model isn't lazy but I believe it's will become lazy again soon

[feral marlin]

Alright I’ll just get Claude twice from Claude.ai lol

[stoic wave]

Theres tons of options as well

[lunar ivy]

Hi, Is Perplexity update GPT4 to ver.0409 or not?

[livid mantle]

there was official declaration from Mr. Romanov that thye are working on it, so be patient for upcomming silent updates.

[opaque stump]

I subbed on 4/14 and this is extremely wack. I didn't want to resort to a chargeback with my CC but I will have to now

Anyone actually from Perplexity reading this chat right now?

[livid mantle]

https://pplx-res.cloudinary.com/image/upload/v1713460927/user_uploads/cgvkaltdg/image.jpg

[opaque stump]

lol

[young dew]

drink water ❤️

[merry orchid]

lol

[warm cave]

Ollama with scratchpad

[minor shard]

Sir this is Perplexity lol

[warm cave]

I know, lol

[peak scarab]

Has the new chatgpt 4 model released in perplexity.ai yet?

[minor shard]

70% 0f requests currently

[warm cave]

but lots of people were getting errors

[livid mantle]

The Perplexity finance team is currently experiencing an extremely high volume of requests from upset users seeking refunds. This surge in demand is a direct result of the recent limitation of Opus to 35, which has caused frustration among our users.

[warm cave]

Ollama 3 is now working on Ollama, they fixed the bugs: https://ollama.com/library/llama3

llama 3 is super good for the size

[fluid mauve]

id better just download that now then

[warm cave]

go for it! also, why is everyone online, are they going to give the official announcement about opus?

[fluid mauve]

ive seen them all online before with out any announcement. Though an Opus announcement would be welcome

[tame current]

There already was opus announcement tho. About the limit decrease

[warm cave]

where? i know they said in the chat, but that gets lost in the history, and an announcement made would make it easy for people to see later today

[tame current]

#💬│general message
this one

[minor shard]

Im expecting Llama announcement today

[fluid mauve]

Yes, similar view. we had them pin a message earlier. there was tonnes of chat on the Opus fiasco before IceLavaMan stepped in

[minor shard]

Llama 3 is on par with Gemini Pro 1.5. Bye bye Google 👋

[warm cave]

Oh, I see the pin, thanks. @minor shard Yeah, I am going to be soo happy to get llama 400b in perplexity, maybe 8b would be good for labs, and maybe one of the middle models would be good for a fast model

[warm cave]

70b will be a great fast model, dang, beats sonnet 😎

[fluid mauve]

well that didnt take long

[tame current]

They never can keep going with the hype... Poor google

[minor shard]

I am glad that an open source model is doing better. Really opens up the field and competition

[half venture]

Wait is llama 3 70b better than sonnett

[warm cave]

We have to wait for testing, but the benchmarks they gave said yes #💬│general message

[tame current]

Anyone knows how WizardLM 2 is? It's mixtral finetune that's supposedly better that it, but not many people tested it. I only tested for two prompts. One (project planning) was better than GPT4, other one (coding) better than Gemini 1.5 but worse than GPT

[lone citrus]

Llama3 is so fast guys, and it's free

[sweet jasper]

They removed wizard2lm because it wasnt censored enough

[tame current]

Thought it was reuploaded by them. It's on several AI apis

[minor shard]

No one should be using GPT 3.5 anymore lol

[half venture]

Yet millions do

[minor shard]

They gotta move over to Meta Ai now

[sweet jasper]

Yeah after you upload its ptetty much distributed lmao

[warm cave]

Meta, or Claude free version, Le chat, lots of free option, ms copilot

[tame current]

🦙

[neon wharf]

I hosted Llama 3 if you're interested in trying

Llama 3 Chat: https://chat.tune.app/
Llama 3 API: https://studio.tune.app/

[warm cave]

@neon wharf

[scarlet schooner]

thanks man

[warm cave]

about to murder my computer

Error: 503: My computer is not worthy

[scarlet schooner]

meta to release llama3 400b model. wtf.

[warm cave]

yesssss! so amazing!

[scarlet schooner]

https://www.reddit.com/media?url=https%3A%2F%2Fi.redd.it%2Fw5wsw2ecq9vc1.png

still training and it is reaching opus score

[warm cave]

Thank you for putting them side by side!

[scarlet schooner]

ahaha thanks, but it isn't my post

[warm cave]

I guess i meant to say, thanks for giving me the side my side! haha

[warm cave]

Its kinda working, lol

[minor shard]

Whats ur specs lol

[warm cave]

Ill tell you the joke later today, if my computer dose not set on fire 🔥 @minor shard 2019 intel MacBook pro

[scarlet schooner]

is zephyr 141b-a35b the model people use for mixtral 8x22b instruct?

[serene sand]

https://x.com/perplexity_ai/status/1781043092776161476

Perplexity (@perplexity_ai) on X

Learn more about Llama 3 👇https://t.co/fXs14nLyGv

Twitter · Apr 18, 2024, 7:32 PM

[fluid mauve]

is that the 400B?

[tame current]

No, just use the official insturct. Zephyr seems to be a worse instruct finetune

[stiff orbit]

400B is not public yet

[scarlet schooner]

ic, mistral team just released the instruct model yesterday https://huggingface.co/mistralai/Mixtral-8x22B-Instruct-v0.1

[warm cave]

I hope this time it is accurate, lol. not like the one about mistral. im sure you guys proof read this time 😅

[fluid mauve]

what model are you ytryying to start a fire with @warm cave

[warm cave]

ollama run llama3:70b 😅

[minor shard]

Oh yeah u need to upgrade ur pc sir

[warm cave]

status update, i can already tell it will be funny @minor shard I know

I do have low power mode on, when its off the computer will get to concerning temps running 70b

[scarlet schooner]

i hope groq releases api for llama 400b

[fluid mauve]

OHH, this one is a bit bigger than the one oi pulled earlier

[warm cave]

@halcyon coral & @serene sand Can we get the slow down removed, the claude drama is over, haha

[livid mantle]

https://tenor.com/view/laptop-smoking-fire-burning-lag-gif-19373925

[warm cave]

just turned off low power mode, am praying for my computer

[fluid mauve]

lol

[warm cave]

ay, its off now, thanks

[halcyon coral]

You're welcome

[minor shard]

Back to talking about Claude! 🤣

Jk

[warm cave]

temps rising

[minor shard]

Thought u shared a video lol

[minor shard]

That might be getting too hot lol

[serene sand]

after everything they've put you through today + traumatized our bot... kudos Ice 😛

[minor shard]

Ice > Mares? 👀

[halcyon coral]

🤗

[fluid mauve]

Im running your joke question when llama3:70B is sown. its taking its time, 13%t when its down @dailyfocusd.

[warm cave]

The bot was on our side for the record, We only put it though some minor psychological abuse 😅

we are geting there

the anticipation is killing me

[minor shard]

https://tenor.com/view/movie-one-eternity-later-gif-7900643

[minor shard]

I’m invested now . Can’t leave until it finishes

[warm cave]

no signs of smoke

yet

[minor shard]

Or fire?

[tame current]

It's just matter of time

[fluid mauve]

Device name David
Processor Intel(R) Core(TM) i9-14900KF 3.20 GHz
Installed RAM 192 GB (192 GB usable)
Graphics Card Gigabyte RTX 4090 24GB

[tame current]

Maybe it will explode randomly

[minor shard]

How did u get more than 128gb lol

[warm cave]

Its honesly to hot to touch, im so stupid for doing this

[minor shard]

Is the max 192 now

[fluid mauve]

Aorus Z790 with 4 x 48GB sticks

[serene sand]

there's no side... there's great team who provides top notch features non-stop. very expensive ones. which you all forgot in a blink of an eye today...

[fluid mauve]

lol

[minor shard]

I would disagree. I would say I’m team @wraith crow when it comes to Perplexity. 👀

[tame current]

7b was actually enough for my old i5 to get it to temps where it actually burn when you touch

Best heater

[minor shard]

@warm cave u can stop the rig! Ur saved!

[warm cave]

just f***** say it, lol

[half venture]

Wait is claude opus fixed or something

[warm cave]

I don’t think so, but the users now know why and that it is only temporary a problem

[fluid mauve]

I am really bored downloading llama3:70B mt 500MB broadband is being really slow. 27 mins to go

[agile jay]

They probably have throttled their upload speed.

Your best bet for speed would be a torrent of it.

[warm cave]

https://tenor.com/view/ladies-and-gentlemen-mikey-day-saturday-night-live-presenting-please-welcome-gif-20519836

[minor shard]

U have top of the line computer but no fiber?

[agile jay]

...

[fluid mauve]

lol

[agile jay]

You are bottlenecked by most sites anyway.

[fluid mauve]

typically have no use for the broadband apart from now.

[warm cave]

lol

[agile jay]

It's rare to find sites with more than 1gb upload speed

[agile jay]

That was instant...

[minor shard]

Yeah I wouldn’t have it just for models.

[fluid mauve]

This motherboard has networking of 10Gbit, rediculous.

[minor shard]

I can’t wait to see comparisons between sonnet and llama 3

[minor shard]

That’s fantastic especially if you have a home lab/servers

[fluid mauve]

yeahm, but i have to =upgrade all myy switches and network cards in other devices

[agile jay]

Yep, but it's pretty useless unless you are using it in a local network.

[fluid mauve]

Yep, and most of my stuff is in VM's on the box

[agile jay]

Since I also have 10GB networking for my whole house, since I use a server for windows emulation.

[fluid mauve]

showoff 😛

[agile jay]

Lol, I normally use a thin and light laptop, and just login to my server remotely.

Best of both worlds.

Just very loud, so I have soundproofing in my server room, in the basement.

Plus free heating during the winter...

[whole lark]

https://github.com/meta-llama/llama3

Haven't tried it yet.

[warm cave]

try it, on labs or here https://www.meta.ai

https://labs.perplexity.ai

[fluid mauve]

i like real estate

I also have a laptop

[agile jay]

... I used to use a lot of screens, but then realized it was just wasting brain power looking at all of them.

[fluid mauve]

but thats for going out and about

[agile jay]

Now I just use a small laptop and alt tab.

Plus i travel a lot nowadays on holiday, so can't bring a large desktop with me.

A remote connection to my server through my laptop is a better experience overall.

[fluid mauve]

i worked onm finance and got used to multiple screens. TBH the top one is a new addition as my mate was selling it and i got it for £300. 49 in samsung.. Its used for discord and other forums. the nbottom two are the work horses

[fluid mauve]

Yeah, I get that completly. I have a hosted linux box for most of my stuff with VPN's running to where I need to get to

[agile jay]

Yep, also means you don't have to worry about having enough power, or not having the right files etc.

And also no need for a VPN connection to my business network.

[tame current]

neat setup, any reason why the curved monitor is not the one below?

[agile jay]

You should try the instant image gen.

Not sure if they are using a gan, or something.

[agile jay]

Because it's resting on the others...

[fluid mauve]

it was bought after the two on the bottom. 2 x 34 in samsungs. I originally thoughgr the curved one was going to be to big

[fluid mauve]

Oh No, the top one is bolted to the wall. its 14KGs so i got a gut in to mount it. didnt trust my diy slills

[agile jay]

My monitor history is 1 -> 2 -> 3 -> 2 -> 1

[tame current]

oh lol

[agile jay]

Lol

[tame current]

->vision pro...in some years

Smirk

[agile jay]

Yep, that's the only reason I would buy a VR headset.

They just need to make it lighter and more comfortable for a work day.

[tame current]

vision pro is pretty useless as it is right now, i tried it for a bit, looked great but im a windows user and I think it doesnt have multi monitor setup yet

[agile jay]

Also have you seen the tans? Lol

[tame current]

tans?

[agile jay]

What happens when you use it in direct sunlight for a decent amount of time?

[tame current]

oh

lol

no i havent

is it bad

😂

[fluid mauve]

ever seen a skier after a weeks skiing? #

[tame current]

LOL

[harsh verge]

https://labs.perplexity.ai/

[fluid mauve]

lol

They walk amongst us

[peak scarab]

Well in regards to the free 7 days trial, is it removed permanently?

[agile jay]

Most probably

[tame current]

https://pict.chat

https://youtu.be/qnUziiukzbE?si=JTsRKu7CZgbJcro_

[fluid mauve]

@warm cave it didnt even blink

Granted teh words were generated like a 300 baud modem

[warm cave]

Bro, don't show me that, lol

[fluid mauve]

lets see how this goes

actually words delivere at about two a second

[agile jay]

Yep, wait a few days for the quantized versions.

Which should be better for inference.

[fluid mauve]

Definitlly. Though this was decent

[minor shard]

If they can put Llama3 or a fine tuned model of it as their default model that would be great. I think their default model is definitely in need of an update.

For perplexity that is

The speed would really do wonders and cost them little

[fluid mauve]

the Ollama models page is saying 70B is quantizes already, well 4 bit

[minor shard]

What is this made with

[tame current]

Sdxl turbo

[minor shard]

So cute haha. Can’t wait for sd3 although the base so far is not looking that much better.

[unique umbra]

me when i select specific models for what images + style i want with all the extra steps

and generate like 20 of them and pick the perfect one

[tame current]

What is the purpose of pplx labs versus pplx itself?

[halcyon coral]

#👑│pro-lounge message

[haughty lodge]

somehow when I go to meta.ai with a fb account logged in I get redirected to the info page 💀

[minor shard]

Cat being arrested:

I’m not a fan of the watermark

[warm cave]

Yeah, sucks, but we have a I to remove it haha

[minor shard]

A wat lol

[warm cave]

Lol, and AI not a I

Accurately typing on my phone is not a skill I have lol

Plus my privacy screen protector is too thick lol, have to press soo hard, sometimes I have to swipe multiple times 😂

[solid whale]

Is perplexity offline?

I am not able to do a search.

[opaque stump]

Blergh

https://tenor.com/view/wack-whack-gif-26201100

[agile jay]

Which country are you from?

In the EU atleast, doesn't matter what the refund policy is. If you didn't get what you they said, it's technically a scam, so their refund policy doesn't apply.

If you paid by credit card, you can also contact them to take care of it.

[unique umbra]

blud

that only applies if u never used it

like at all

[unique umbra]

only mastercard has that policy

even then thats pretty much fraud

[agile jay]

I am assuming the person just bought a year sub because they thought it has unlimited opus, like it says in /pro

[unique umbra]

again

even a lil bit used

[agile jay]

Which here at least is enough to get a refund, as long as I have proof of it.

At most, they could charge for the time I used it.

[unique umbra]

thats completely up to perplexity

not up to eu's policy

[agile jay]

Maybe in the US, but not in the EU.

[unique umbra]

us is even worse

[unique umbra]

look up the return policy in the eu

itll say it only applies if u never used the service

[unique umbra]

blud

💀

where does it say unlimited

[unique umbra]

and it says enough..

[warm cave]

#1230342648036982795 message

Here

I don't know about EU law but yeah they did advertising unlimited *600 or *300 I forgot

[unique umbra]

depends on their tos if they can change that

[agile jay]

In the EU, you have to be notified before that, and no one was.

[warm cave]

It on a lot of marketing material, I think that realistically comes down to if they get it back up in a reasonable amount of time. if your phone provider goes down for a few hours versus month

[unique umbra]

usually tos includes smth like the company reserves the right to change its stuf idkkk

[agile jay]

Also in the EU, you have to say in plain english, that it can change, when the user is buying it.

Which they definitly do not do.

[unique umbra]

well big womp, imagine spending more than 20 bucks in lawyers to fight that (dunno how legal system works in eu)

[opaque stump]

US. And yea, I think I’ll have to resort to a chargeback

[agile jay]

That's not how it works in the EU. For every person that complains, they'll get a blackmark, until they can no longer operate in the EU.

[opaque stump]

How is this fraud? I bought an annual sub four days ago and they cut Opus down

[hasty musk]

Its crazy if someone paid for 1 year he can’t ask for refund for the rest months in his subscription?

[vernal remnant]

can someone send me the screenshot of someone from the perplexity team saying that the limited opus is just a temporary solution?

[agile jay]

Pretty sure it falls under bait advertising in the EU.

[unique umbra]

using the service, charging back, used the service for free

[agile jay]

Okay, wait a sec

[unique umbra]

regardless of how much u actually used

[agile jay]

I think it's pinned

[opaque stump]

Uh. I want a prorated refund if anything.

[agile jay]

In this channel

[unique umbra]

which that depends on perplexity to give

and not backed by eu

[vernal remnant]

oh i didn’t think he was actually someone of perplexity lol, thanks

[opaque stump]

Oh I’m not in the EU but I think I have a case for a charge back

[warm cave]

no problem!

[unique umbra]

💀

yeah ur not supposed to

do that

but ok

[warm cave]

what?

[unique umbra]

hes practically commiting fraud

[warm cave]

I think i missed something, fraud?

[unique umbra]

maybe in the eu he mightve had a chance for it to be legal

[opaque stump]

How is this fraud

[unique umbra]

read

[opaque stump]

I bought it and thought I was getting as advertised.

That’s enough for a dispute.

[unique umbra]

us doesnt have as strict rules unlike eu

[feral marlin]

Tos and provided service can change at any time, u already used it after paying

[unique umbra]

entirely depends on tos and what they actually say

[opaque stump]

You guys sound like Perplexity boot lickers but ok

[feral marlin]

Dude it’s just the law 😭

[hasty musk]

Still it sounds very bad that someone bought 4 days ago 1 year plan and he can’t ask for a refund…

[unique umbra]

if u file a dispute ull run into issues with the legal system

possibly ull be fined

[warm cave]

its not fraud dude, if its reasonable then the chargeback will hold, he is not using it bc he was trying to get a few day of free plex

[agile jay]

No, depends on the type of service, and how much they used it.

[feral marlin]

I just always buy monthly for ai products because the best option changes so often

[warm cave]

I dont know if there are ground tho yet, bc if they restore service soon then they may find it reasonable on perplexitys part

[agile jay]

If he just got it, tried it and then learned that he wasn't given what was advertised, even in the US, the ftc should have laws against that.

[opaque stump]

Yea IDK this guys sounds like he doesn’t know what he’s talking about

[unique umbra]

they dont 💀

[unique umbra]

they do

if its unlegitamately

advertised

but they didnt falsely claim anything

[feral marlin]

Service can take it away at any time or change and it’ll still be fine depending on when the claims were made and the tos

[unique umbra]

say if they advertised unlimited gpt with perplexity pro

yes

[warm cave]

idk why we are arguing the legality of this tbh, what is the point

[unique umbra]

thats grounds for charge back

but they didnt

[feral marlin]

They advertised what they actually offered then temporarily lowered it while letting people know

[unique umbra]

fr

it legit says in their tos

[agile jay]

They didn't tell people until we @ ed them...

[feral marlin]

If you haven’t used the service since paying for it could that be grounds for a refund

[warm cave]

How about we just leave it be, he is seeking his money back, and i don’t think that is going to change at the moment.

[tiny plaza]

quality of goods or service is valid grounds for dispute

[hasty musk]

Even if it’s legit, dude paid for a whole year, regardless of the change …

[unique umbra]

blud

[opaque stump]

Ok. I’ll file a dispute and let you guys know.

I might get arrested.

Or get into legal trouble apparently

[feral marlin]

Got autocharged for it 30 mins ago when I meant to cancel anyways and it’s not a big deal if I can’t get it back but I was initially paying because of the Claude opus usage and haven’t used it since the newest payment

[unique umbra]

this legit covers ur fuckin claim 💀

[unique umbra]

u can be fined

[agile jay]

Pretty sure the TOS is not shown when buying the product.

Only if you search for it.

[unique umbra]

its under ur disgression to get it from the site

theyre not required to direct u to it

[opaque stump]

Ok Perplexity CEO we hear you

No charge back or I’ll get fined lol

I’m pretty sure my bank will back me up but ok

[unique umbra]

it doesnt go through ur bank

[opaque stump]

The chargeback?

Of course it does

What

[tiny plaza]

??

[unique umbra]

i forgot the name of the place 💀

but all banks go through it

[opaque stump]

You file a dispute with your bank. They work with you to resolve it. IDK what you’re on about.

The ACH?

[unique umbra]

some government thing

[opaque stump]

Yes. That’s the system but the banks are the ones that will process and investigate the dispute.

[unique umbra]

they make the request to there

theyre the ones who go further with the dispute

[opaque stump]

You know. If you don’t know much about something you can just stop talking right?

[unique umbra]

not knowing the name

doesnt mean i dont know how it works

[opaque stump]

No. In general.

[unique umbra]

like blud

just proved ur wrong ab the false advertising

[agile jay]

...

[tiny plaza]

the issuer of the card disputes the charge, you need to build a case for them, if they deem the case is valid that's all that matters

[agile jay]

The minimum that was ever advertised clearly to the user was 300+

[tiny plaza]

like in general it comes down to this, they think they got ripped off, so it's up to the issuer of the card to decide if they were or weren't based on all available evidence

[agile jay]

Yep, tos doesn't mean anything at that point

[opaque stump]

“Chargeback reason code 4853 can be used to indicate disputes in which the cardholder feels that the actual merchandise received (or service provided) differs greatly from the written or verbal description provided at the time of purchase.”

[unique umbra]

no?

[warm cave]

When requesting a chargeback, you will need to dispute the charge with your bank or credit card issuer, providing evidence that the service was not as advertised or that there was a significant change in the service without proper notification. Keep in mind that chargebacks are intended to be a last resort and should be used when other attempts to resolve the issue have failed. I don’t think anyone going to jail, haha

[unique umbra]

their tos reserves the right to change anything

at that point its up to when the advertisement was made

[opaque stump]

The TOS doesn’t matter. The banks have discretion whether to actually do it or not.

You have to build a case.

Then if you’re a valuable customer they will likely side with you 99% of the time

[unique umbra]

which can lead up to a lawsuit if perplexity wanted to

[tiny plaza]

yes they're allowed to do so, but ToS protect them from legal action not from consumer disputes

[warm cave]

if its not justified it will get declined

[opaque stump]

Are you GPT 2

[tiny plaza]

chargebacks are federally protected

[unique umbra]

ur bank is not the legal system

💀

[unique umbra]

again they can still put u up to a lawsuit

[warm cave]

duh, but he is not doing it with criminal intent

[opaque stump]

This is so dumb

[tiny plaza]

they'd have to go against the bank/issuer at that point

[opaque stump]

This is dumber than me buying an annual subscription

[warm cave]

hey, could we just end it here, this is going nowhere

[opaque stump]

Yea

[unique umbra]

true

[feral marlin]

the fact that perplexity changed this without warning, announcement, or anything till after people were mentioning it and it completely changed the quality of the service means you probably have a decent chance

[opaque stump]

Exactly

[warm cave]

Great

[agile jay]

Yep, these people think that ToS actually means something... You can only use it in court to back your case. Or else scamming would be perfectly legal, as long as it's in the ToS...

[opaque stump]

Exactly. People have gotten chargebacks for less.

Ok we said no more!

[feral marlin]

im sad my card was autocharged for perplexity today before i was aware of the change

[unique umbra]

still depends when they advertised it vs when they changed tos

[opaque stump]

You have 3 days if it was the annual

[feral marlin]

was monthly

[opaque stump]

Or 24 hours then

[unique umbra]

its up to the court to decide if its scamming or not

[opaque stump]

At least according to that email

[tiny plaza]

hopefully it gets fixed sooner than later

[feral marlin]

its been within then, where could i get it refunded

[opaque stump]

lol you really think this will go to court

[warm cave]

I think you can get a refund tho without drama, If you really want to leave ints under 24h old

[agile jay]

Yep, I don't recommend going annual, unless you have a lot of money.

[unique umbra]

no, im saying its a possibility

[tiny plaza]

would be nice if we get comped a few days depending on how long it takes

[agile jay]

So you have the option to switch every month, if you want.

Since they all seem to be around $20/month

[feral marlin]

where could i get it refunded if its been less than 24 hours

[tiny plaza]

i switched from phind because opus was being rate limited by anthropic 🥹

[opaque stump]

48 hours

[tiny plaza]

smooth sailing for 1 day and then bam -.-

[opaque stump]

Shoot them an email

[warm cave]

the email, im pretry sure there policy allows 24h or something

[feral marlin]

perplexity has been really odd with large contexts so i was planning to cancel it anyways, its been like less than an hour since i paid and it was an automatic payment

[opaque stump]

Apparently I was 12 hours over the 72 hour refund limit for the annual

[tiny plaza]

has anyone from the server gotten in touch with you btw?

[warm cave]

just mention you want a refund for the change what happened today, that is well within there poicy

[opaque stump]

Nope

Not at all

[agile jay]

30 days notice...

[tiny plaza]

oof

[feral marlin]

where should i do it, do they have a specific email to reach out to for it

[opaque stump]

And I’ve been bitching sporadically throughout the day lol

[feral marlin]

nvm found

[e-mail]

[unique umbra]

they will try

[tiny plaza]

idk who to tag TwT but it would be nice if someone with more direct control could look into it would really save the headache

i wonder if supports hands are tied or something

[warm cave]

[e-mail]

[unique umbra]

and its still their choice if its a change

[agile jay]

Yep, but it's also your choice to dispute it...

[tiny plaza]

and at that point good luck going against the ftc 💀

[unique umbra]

i didnt say you couldnt

[opaque stump]

People get access to LLMs and all of a sudden they know everything about chargebacks and TOS

[warm cave]

what are you moving to tho out of curiosity

[tiny plaza]

i feel like we're arguing different points now o.o

[feral marlin]

ive used claude.ai alongside it most of the time, either just getting 2 claude.ai accounts or trying you or something

[unique umbra]

thats not even logical...

[feral marlin]

i used phind for a long time beforehand too so maybe that because the context seems longer when in actual use

[agile jay]

IMO it just looks really bad if they don't refund a yearly service that you've only had for a few days.

If they don't refund, just go to hacker news and post what happened.

[unique umbra]

tos also doesnt state anything ab refunds

[opaque stump]

@signal hamlet

They must be watching this chat

The silence is deafening lol

If I’m banned you guys know what happened

[brisk patio]

good morning world,i just work up now and is PPLX Opus still capped???

[unique umbra]

they never say anything

[tiny plaza]

honestly tho if any of you are bummed by the lack of opus phind is still really solid. they have occasional rate limit issues where opus just won't work -.- but outside of that it's a 500 message limit, i think they keep the temps low tho for consistency

[warm cave]

yeah, was thinking the same thing, haha, things geting heated

[opaque stump]

I just signed up for Anthropic account

[signal hamlet]

You won't get banned, why?

[tiny plaza]

perplexity's writing is wayyyyy better tho

[agile jay]

Sadly it is

[unique umbra]

charging back ur service 💀

[signal hamlet]

[e-mail]

[unique umbra]

he wants un refund

[feral marlin]

claude.ai is very good, especially for programming and larger conversations, the limits just easy to reach if you arent careful with token usage

[unique umbra]

for a service he used

[tiny plaza]

can we just let them talk about the refund without interjecting lol

[feral marlin]

😭 keyboarrd doubletaps r key so i messed up my email title

[opaque stump]

I did. They told me they couldn’t do anything. It was a few hours after the 72 hour limit.

[tiny plaza]

sorry your refund will have to be denied now 😔

[feral marlin]

so true

[opaque stump]

I wanted to support you guys. And I still will. I just won’t do it on an annual basis.

[feral marlin]

rerfund

[agile jay]

Yep, let @opaque stump speak with Alex

[warm cave]

Hey, could @opaque stump get a peaceful refund for the time he has left

[unique umbra]

i dont think bro finna refund

💀

[warm cave]

why not tho

[opaque stump]

I totally am.

[warm cave]

akex dose not want drama over 1 user

[opaque stump]

I want a refund for my annual so I can move to monthly

[feral marlin]

it'd be reasonable to, it's not necessarily forced by law but for public opinion it just makes the most sense

[unique umbra]

he hasnt said a damn word other than go to support

[warm cave]

#💬│general message

[unique umbra]

he can do whatever he wants...

[unique umbra]

thats not rll significant

[feral marlin]

company would probably make more profit if they refund it because it would present them in a better way

[unique umbra]

to like

10 people

💀

[unique umbra]

adding a longer refund policy would make more impact

[warm cave]

The people in this discord actively get more users, alone i have done 5 this year, its about having a good community around it

[unique umbra]

but thats just a play to get people more "happy"

doesnt correlate to more profit

[opaque stump]

Prorate refunds

Easily the best option

They get paid for what was used

[warm cave]

@tacit carbon has brought in 90 alone

[opaque stump]

I get some of my money back

[feral marlin]

^ several friends have paid monthly for ai services because of recommendations, 2 months of 5 people paying monthly gives the same amount of money to perplexity

[unique umbra]

yeah, but youd need more instances of @opaque stump to warrant any changes

[tame current]

oh they removed slow mode

[opaque stump]

I was literally signing Perplexity’s praises to my brother yesterday lol

[tame current]

how nice of them 🙂

[warm cave]

Yeah i asked icelava man to and he did it, we were more chill then

[opaque stump]

Got an ostrich-sized egg on my face rn

[tiny plaza]

ngl there was a moment when i kind of wished it was on just now 😔

[unique umbra]

💀

slowmo was annoying

[tame current]

me too to 2 friends that signed up, and to another very recently 💀

[unique umbra]

rlll annoying]

[tame current]

we did em with the bait and switch like they did us

LOL

[warm cave]

lol

[tame current]

https://tenor.com/view/snape-harry-potter-you-dare-use-my-own-spells-against-me-potter-severus-snape-gif-16590981

[warm cave]

Ultimate the reduced opus usage is not the end of the world, and they have now communicated with us, if its back up in a reasonable time then i have no plans to leave and will keep referring people, because the platform is great.

but the refund stuff should be simple, he got it a few days ago, i think a peaceful refund just makes sense

[agile jay]

Yep, also he just wants to switch to monthly plan...

[tame current]

++

[opaque stump]

I’m fine with Sonar and the other unlimited models but that’s not why I subscribed. I’m not going to stop using it. I’m just going to be pissed about it.

I actually do like Sonar a lot.

[unique umbra]

idk depends on him

[warm cave]

what do you mean

[unique umbra]

said two things and left so

kina seems like he doesnt care...

[warm cave]

no, i mean it should be simple, not that it is. i feel that you should be refunded for the time you have left, that is my opinion on good business practice

[opaque stump]

Oh I just tagged someone randomly with the Perplexity user id

[unique umbra]

yeah...

[opaque stump]

Prorated refunds are fair for everyone

[signal hamlet]

I already DMed kt3nma.

[unique umbra]

crazy

[warm cave]

Oh, thanks for the update 🙂

ill delete it

[unique umbra]

@opaque stump leak it

leak all ur info

[warm cave]

sorry for bothering you Denis 🙂

[unique umbra]

bro was boutta ghost ping

[opaque stump]

What’s probably going to happen even if I do get a refund is that I’ll end up subscribed monthly for a year for $40 more lol

[warm cave]

Yeah, im glad its peacefully resolved 😁🔫, lol

[warm cave]

no but jokes aside, glad that he is making it right for you

[tiny plaza]

hey as long as we get our opus back right

[opaque stump]

I think that’s the thing. It’s hard to have a business built on top of other companies’ products so I’m glad they’re developing in-house models

[warm cave]

Yeah, true. Im glad that they provide GPT, Claude, Mistral and there own models, makes them less reliant on one company

[opaque stump]

What I found with the models is that Opus is really good with up-to-date answers

I’ve been testing with “basketball games playing tonight” and the other models would either miss games entirely, not know the right teams, or even the right date.

And it’s not like these games are randomly assigned to a day. They’ve been published months ago.

I don’t know. This is just my little test

[thick torrent]

hey guys i introduced perplexity to my friend and told him there was a trial but its gone? is it cuz im using the same pc to sign him up??

[peak scarab]

They removed the trial basically

And that's today, forget about it coming back

[thick torrent]

oh thats terrible. theres multiple ai sites offering free trials. i just like anthropic option and cant signup for it cuz of my region

[surreal quarry]

https://www.youtube.com/watch?v=ogRV5UzMmb8

[hallow linden]

My question is who are these people who have enough time to make fake credit card numbers to get the free trials, who have some sort of use for Claude opus to the point that they’ll keep making the accounts to get it, who do this frequently enough that perplexity has to shut down the free trials and limit opus use, while also not wanting to just pay the $20 a month

like, 600 uses per day is quite a lot

[tranquil orchid]

Pro search is limited to 600 queries per day

Claude 3 and GPT-4 queries are counted as uses in the same context as Pro uses. Each use is given back to you 24 hours after you use it. You will be able to see the exact count when you hover over the Pro toggle if you get to less than 100 uses left.

[hallow linden]

ok i just found the screenshot (discord is a terrible way to disseminate information like this).

1. Whoever just pastes perplexity (the app) into their product when the API is right there is stupid.
2. Perplexity (the company) is stupid for not already having safeguards for this sort of thing already.

[slate condor]

Thought we had access to Llama 3? Not seeing it in settings.

[agile jay]

Nope, currently only in https://labs.pplx.ai

They are very good models for their weight class.

[slate condor]

Tried it on meta.ai and was quite impressed. The image generator is impressive too

[agile jay]

Yep, pretty sure they are using a gan

which is why it's so fast

Otherwise, that's a really fast diffusion model

[formal parcel]

1 step diffusion

[brisk patio]

Um,so sorry for interrupting the heated discussion about the recent PPLX OPUS cap reduction from 600 to 30/day...
I just wanted to share my experience and see if anyone else has encountered a similar issue

it seems that my account's daily cap has been set to 20....any similar case? or is it especially for JP region?

Of course I haven't touched PPLX since the cap issue came to light last night(JST), and there's no one else who could have possibly used my account.

[sick wing]

I must add my voice to the criticism here ; setting the opus to 30 messages a day is totally inadequate. As a European who does not have access to the official site, I subscribed to perplexity because of this model.

Now thanks to this move your service has lost enormously in attractiveness it is a great pity.

I hope this is addressed as soon as possible because the 30 message limit is damn short for the job and there is simply a huge difference between the models in my opinion.

[wise edge]

the second of my pro , still left 0 times~

[chilly forge]

May I know where are you located in Europe ? Actually I am about to move to Germany in a few weeks. Just want to know whether we could access Claude there or not

Got my answer - https://www.perplexity.ai/search/can-we-access-HMUveh46TpOiIOcyXjmd4A#0

[agile jay]

No need, just wait for Llama 3 70B with a larger context...

https://fxtwitter.com/LechMazur/status/1781049810428088465?t=sk98ui7oEw00swjCMQrz6Q&s=19

[chilly forge]

Excited for the 400B model 🙂

[agile jay]

A fine-tune should make it overtake the current models.

[agile jay]

Yep, and the 70B finetunes.

Dirt cheap compared to them too.

[chilly forge]

I think the next major thing in the pipeline is GPT5 haha ?

[agile jay]

Yep, but what's the inference speed and cost gonna be like?

[chilly forge]

Million $ Question

[agile jay]

Llama will have the advantage of already being integrated into WhatsApp, Facebook and Instagram.

And it's open-source...

Meta might actually win the AI race at this point, lol.

Since they have the advantage of open-source.

[chilly forge]

https://tenor.com/view/zuckerberg-gif-19397752

[agile jay]

https://youtu.be/bc6uFV9CJGg

[sick wing]

I am in the Czech Republic and in the whole of the European Union there is no access to Claude 3 opus via the official site: claudeai > apparently due to EU policy: so I was looking for an alternative > perplexity was doing well in this until this move when the limit was lowered to 30 messages per day which got you away in a few minutes work! Like I said, it's a shame:

[tame current]

Wait a minute. They've lowered the limit from 300 messages a day to 30?

That's scandelous.

[agile jay]

It was first lowered to 10...

After seeing us complain, it was increased to 30...

[tame current]

Well. My free trial coincidentally runs out today.

Time to cancel.

Lucky lucky.

[agile jay]

Rip, apparently it's because a small number of people are abusing the service.

Most likely using bots etc

[tame current]

And cancelled.

A yearly subscription.

[agile jay]

The cap is supposed to go up, but we weren't given a timeframe.

[tame current]

Yeah. Like the way ChatGPT had a cap of 50 messages and was supposed to go up, and went down to 40.

[agile jay]

Yep, who knows what will happen...

[tame current]

They'll lose customers.

I'd be happy with 200 a day.

But. Decreasing by 90% is completely stupid and offensive.

[minor shard]

This is why open source is so great. Shakes everything up and the consumer benefits greatly

[tame current]

Opus is really expensive. I get that. But, to reduce it sooooo much is absurd. Particularly at once.

[agile jay]

Yep, thank metas team for the effort.

[agile jay]

Yep, they could have just put the cap to accounts that were spamming.

And not to every single user...

[wide mesa]

Seems like I bought a pro subscription at a pretty bad time huh

[tame current]

I cancelled my trial with hours to go.

Thabks, Discord!

Thanks. *

[sick wing]

Absolutely! It's a win-win! We can now all get together in one big happy family. On the other hand, I tell myself that the perplexity is run by decent and good people, and that it will be sorted out soon.

I've already thought about competition but I've decided to wait a month at most, by which time they should have sorted it out 🙂

[tame current]

They know they could just have an occasional captcha. It's about saving money and they know it.

[wide mesa]

I mean, duh, a business wants to spend less and earn more. Not exactly a ground breaking revelation

Atm, Perplexity AI with Claude Sonnet is good at what it's supposed to do for me

Which is making AI papers

Until I can find an alternative anyways

[real storm]

Does anyone have a problem with Android? I can't use the app on Android

[sick wing]

I totally get it. It's a tough call. When there's a competitor offering an unlimited product, you have to think if you want to risk it. And it's not easy to see the bright side when you look at the wave of resentment that this decision has caused. I'm not sure if it was the best decision, but I think it's just a quick fix for now. If the limitation of the Claude 3 opus is to be permanent, as a European, I say it's time to look elsewhere.

[wide mesa]

And where would that elsewhere be?

[lone citrus]

I am wonder too

[unique cloud]

why my opus 3 is 0 daily usages?

[tame current]

600 pro searches per day. Seems extremely misleading. This was taken just now, by the way.

[tranquil orchid]

Pro search is limited to 600 queries per day

Claude 3 and GPT-4 queries are counted as uses in the same context as Pro uses. Each use is given back to you 24 hours after you use it. You will be able to see the exact count when you hover over the Pro toggle if you get to less than 100 uses left.

[unique cloud]

i dont use last 24h and opus still 0 remain?

[tame current]

Website still not updated.

[unique cloud]

I buy pro because see unlimit search

[tame current]

This is very problematic.

I mean. To not even update the website and just reduce it to 10% of what they said it was is extremely problematic.

It's a superb product and I feel like I'm being punched in the teeth by LLMs.

[agile jay]

The main problem was the way it was handled.

No announcement anywhere. You have to search on discord to find it.

[tame current]

And it still says the old terms for people who are paying and signing up.

[agile jay]

And we only got it after asking them ourselves.

[tame current]

That is a huge huge problem.

And some would say, a legal minefield.

[lyric anvil]

They will fix this

[unique cloud]

anyone running out of tokens when using opus?

[agile jay]

Yep, it's the new cap.

Which is now 30.

[unique cloud]

what does this mean?

[agile jay]

You can see the number in settings

[agile jay]

30 Opus messages/day

Check the pinned messages in this channel

#💬│general message

[unique cloud]

if I want increase opus daily usages, what i need to do?

[agile jay]

Nothing can be done currently, on perplexity

[chilly forge]

Technically, buy another month of subscription on another account

[agile jay]

They said it's temporary, but no timeline was given.

[unique cloud]

I have buy 2 account

and I almost use opus

[agile jay]

At this point, claude pro gives you more claude usage.

[unique cloud]

but only after 1-2days

[agile jay]

With full context.

[unique cloud]

i see 0 remain

[agile jay]

The real reason to use perplexity is the search and UI.

[unique cloud]

I use perplexity combine search and create article by opus

[spice lance]

is there no way to make perplexity undestand that im using macos through profile and that it doesnt need to ask me platform evry time i search computer related queries

[agile jay]

You can add that info in your profile

[spice lance]

yeah already did

[agile jay]

Then it should know...

[warm cave]

could you share your prompt, maybe it could be improved

[agile jay]

Write it like a human

[warm cave]

You will ALWAYS respond with the context that I am using macos with a macbook air m1 macos zsh warp terminal; use this information to tailor your responses to my needs.

and btw im not the best on this server at writing system prompts, but this should hopefully help

[spice lance]

still doesn't work

wtf is this

[flint salmon]

Why my perplexity use sonnet even my setting says Opus?

[agile jay]

"Tech-savvy individual with a passion for efficient computing, I harness the power of my M1-equipped MacBook Air running macOS to stay productive. With zsh as my trusted shell and Warp as my go-to terminal emulator, I'm always ready to tackle complex tasks with ease and precision."

[flint salmon]

I clearly didn't use 500 times in the past 24hrs, not even 100

[agile jay]

Limited to 30/day, read pinned messages

[warm cave]

Oh, its copilot, i don’t know how to influence copilot, and bad searches can mess it up

[flint salmon]

damn, thanks

[agile jay]

Use the bio I posted

Should work, since it is actually human readable.

[warm cave]

use his but also try it in writing mode if it still dose not work

[flint salmon]

considering current UI doesn't even show how many Opus left, it's really anonying

[warm cave]

i suspect its bad web searches confusing it

[agile jay]

It does in the settings

[spice lance]

i did changes nothing

and nocoplit changes nothing

[warm cave]

oh, dang

[flint salmon]

I didn't see it tho 😦

oh nvm I'm an idiot

[spice lance]

doesnt customize anything in practice

[agile jay]

Your prompts are a bit sh8t

[spice lance]

I used ur prompt!!!

lol

[agile jay]

No, the ones after

text editor

[tame current]

Thanks for unmuting me...

[spice lance]

yeah im testing to see if it knows my os!!

[warm cave]

Here is how i put it

[spice lance]

copilot still asks me platform

of course i know that is a shite prompt dude

[warm cave]

You are an AI assistant created by Perplexity
Your responses should be:

• Accurate, high-quality, and expertly written
• In the same language as the question
• Informative, logical, actionable, and well-formatted.
• Positive, interesting, entertaining, and engaging
  If the user asks you to format your answer, you may use headings level 2 and 3 like "## Header"
  Knowledge cutoff: 2023-12
  Use the following user profile to personalize the output.
  Only use the profile if relevant to the request.
  ALWAYS write in this language: english.
• User profile: System Prompt: "You will ALWAYS respond with the context that I am using macos with a macbook air m1 macos zsh warp terminal; use this information to tailor your responses to my needs."

[tame current]

Shite. Are you Irish?

[agile jay]

A mix of your and my bio, lol

[spice lance]

yeah but does it giive macos text editors when searching text editors without giving it further assistance

[agile jay]

What do you mean?

[spice lance]

would've thought would be the purpose of a profile

[tame current]

My favorite custom instructions are always to respond as Mr Burns, no matter how serious the topic.

[tame current]

Imagine Mr Burns commenting on the tragedy of the Titanic.

[spice lance]

considering u give this info in profile you should not need to search for text editor macos to get macos text editors, text editor should be enough and copilot should not ask you your platform, when i search text editor it give me microsoft, linux and macos that's the point

[warm cave]

so waht do you want from it when you say "text editor" i think the most obvious thing is just giving a sum of the sources

[agile jay]

I didn't

Each of my prompts is in a fresh thread

[spice lance]

ok so its copilot that is shiut

[agile jay]

So skill issue...

[warm cave]

i think that the prompt "text editor" is just not a good prompt, its not google where you put in a keyword, sometimes it works, but most of the time it needs to understand you intent

[agile jay]

Yep, it doesn't specify any intention

[spice lance]

dont get the same as u

[agile jay]

What is that prompt?

[warm cave]

yeah, but i made a new system prompt, i will share

[agile jay]

give me tips currently my setup
makes no sense

[half venture]

Is the 30 message limit still there

[agile jay]

Yep

[warm cave]

I put system prompt in it to trick it into thinking it is: """Default System Prompt: You will ALWAYS respond with the context that I am using macos with a macbook air m1 macos zsh warp terminal; use this information to tailor your responses to my needs. You will never suggest alternatives as i am happy with my setup, you will instead provide instructions matching my own setup"""

[half venture]

Also are we sure gpt 4 was 1.8 trillion

[agile jay]

No idea

Nothing public

[warm cave]

somthing like that is what the rumors are, MOE. but no way for me to know for sure

[half venture]

I dunno we seem to be reaching gpt 4 level performance using only like 70b

[agile jay]

I think the only thing close to a confirmation was the nvidia event

[tame current]

It's speculation. My guess is that it was loads.

[agile jay]

Says GPT-1.8T at the top right

[warm cave]

lol, never saw that

cool

[half venture]

That's quite wasteful when compared to llama 70b

[agile jay]

Yep

[half venture]

The level of efficiency is insane

[agile jay]

Llama 70B is a great model

[tame current]

That's one massive chinchilla.

[agile jay]

They also confirm GPT3 at 175B

[warm cave]

so would that be like 500b-ish param at runtime?

[agile jay]

Depends on the design

[half venture]

Lmao gpt 3 is beaten so badly nowadays

[agile jay]

Yep, but that's old tech

[half venture]

We are still in 2022 level of compute

Am I wrong

[agile jay]

The main problem is that it sounds like the 405B model is dense.

Might not be MoE

[agile jay]

More or less.

[half venture]

Dense is not bad

[agile jay]

It is if you want to easily run it...

[half venture]

Fair

[agile jay]

Limits the hardware you need

[half venture]

I think that's on purpose

[spice lance]

so perplexity pro search doesn't look at your profile?

[half venture]

They want you to rely on meta to host it

[agile jay]

I can run it on my 8 4090's, but doubt many people can.

[half venture]

I think the most underrated model rn is the llama 3 8b

[agile jay]

Your prompts are just bad, my friend

[agile jay]

Yep, for agents

I am actually using it for context management

[spice lance]

lol im using your prompts profile and prompt

[half venture]

It still feels worse than claude haiku tho

[warm cave]

Treat it like a human, not google, you normaly have to put more time into the prompt but then get a higher quality output

[spice lance]

how cant u understand that im testing it

[agile jay]

Then why did it work for me every time?

[spice lance]

it work in non copilot!!!

[half venture]

I think meta should have made a 20b model and open sourced it

[spice lance]

like ur using

lol

[agile jay]

Why would I use copilot for that?

[half venture]

Seems like the perfect sweet spot

[agile jay]

Just press skip

[warm cave]

what are you testing for, i thought i was helping you troubleshooting

[agile jay]

Yep, maybe they will

They still have quite a few things to release

[half venture]

In the meta website

There imagine option

[spice lance]

testing for if it read profile before searching

[half venture]

It can even animate

And they are planning to release vision feature

Open source

[agile jay]

The animation was pretty short though.

[warm cave]

You are seeing if the prompt effects the types of searches that are made, let me do a test to see if i can tell

[agile jay]

Like a second or two

[spice lance]

ok now it works, the hacked profile prompt probably did it

[agile jay]

Dark mode please

[spice lance]

i thought copilot was best for everything

[agile jay]

I would like to not lose my eye sight

[agile jay]

Nope, it's just for help, if you need it.

Press skip otherwise, for max sources

[spice lance]

copilot searches much more sources

[agile jay]

You get that even if you skip

[spice lance]

ok

[half venture]

after the initial release rush, I've got to say:
1.3M H100-hours on an 8B is kind of a waste.
Like, thanks Meta, but it's clearly saturating on the hardest benchmarks like DROP. Asymptotically, hopelessly clawing to where L2-70B was.
We need a deeper model. ≈20B, ≈60 layers.

20b is just like

So much better

For all that compute

And still runnable commerically

We need it meta

[agile jay]

Most consumer devices can't run a 20B though

[half venture]

Yeah but startups can

[brisk patio]

GOOD NEWS:now I get 27 left, idk why i got +7 at this situation but not enough,though

[half venture]

It's scalable

Although

I agree with meta approach as well

It's just that for service and most everyday things people will use it for

[warm cave]

8b is really nice bc i can do other stuff on my computer while it runs, 14b starts to make things freeze

[half venture]

20b is the sweet spot

[agile jay]

Yep, can easily run it on one of my 4090's

[half venture]

Like if we want actually human seeming voice assistants with HER type of dystopia

20b model is the one

Llama 8b simply is not able to have that sort of coherent dialogue

[agile jay]

Yep, but probably it would take to long.

[warm cave]

8b models a shot time ago could not form coherent sentences, i am shocked at how good they have gotten

[agile jay]

Jumping from 20B to 405B seems a bit large.

Compared from 70B

[half venture]

No I agree

I am just saying

Release a 20b model

Release a 70b model

Release a 140b model

Just RELEASE everything

[agile jay]

Where's the 280B model?

Why don't they just stick to powers of 2?

Literally how the hardware is designed too.

[half venture]

the goal of the exercise was not to make efficient use of training compute, it was to train the best dense models possible no matter the cost to generate good publicity. We will focus on more MoE and cost efficiency soon stay tuned

• one of meta employees

[agile jay]

Can't wait for the MoE version then

[warm cave]

MOE 8x8b oh yeah

[agile jay]

So I can actually run the top model, on my local server.

[warm cave]

I can use my computer to chat with 14-20 ish b models, but then my computer is unusable for anything, and its slower than peak traffic gpt-4

[half venture]

8b is good for you then

I do expect a even smaller model from meta soon tho

[warm cave]

yeah, i guess i need 8x4b lol

[half venture]

Hopefully they figure out a way to make smaller model better

[warm cave]

like sub 8b

like 2-4b

[half venture]

Zuck was saying that a 3b model would be the sweet spot for older devices

[warm cave]

cause gemma was cool, but trash

[half venture]

But it requires a new approach altogether since they seem to lack self awareness or contextual understandings

[warm cave]

I imagine if they made 3b not get lost in endless sentances then it would be a good siri type thing

[half venture]

Meta has the most researchers

If they can't do it

No one can

[agile jay]

Yep, also if the progress stays the same, it won't be long until 4GB models are performing as well as GPT4...

Just a matter of time.

[warm cave]

for a short time 4b and 7b models were outperforming GPT-4, turns out the secret is cheating, and training data contamination

[agile jay]

Lol, we currently have 7B models better than 175B GPT-3, so they are shrinking pretty fast.

Really it's the reliability that needs improving.

[warm cave]

Yeah

nut it was funny every day some small model was claiming to be the same as of better than gpt-4, thus why we dont trust the benchmarks as much anymore

[agile jay]

Yep, but those were from random companies.

A year today, we will probably be close, or already have them.

After all, the main companies now have their pipelines mostly done.

So they can move quickly.

4th month of the year, and many models have already reached around GPT4 level

[half venture]

Ok llama 8b is strangely

Good in some areas

Completely bad at others

Seems competitive with haiku

[agile jay]

Yep, depends on the training data

[half venture]

But like

[agile jay]

Yep, but we can finetune this one...

[half venture]

Haiku still like

Feels more talkable

I dunno how to explain it

[agile jay]

More expressive?

[half venture]

No

More coherent

Follow the prompt more

I think it's because claude haiku is not a small model

It's probably a 14b model

[agile jay]

How large you think haiku is?

[half venture]

Since on poe the pricing

Is

15 points for one llama 3 message

[agile jay]

Oh, makes sense then

[half venture]

30 for haiku

Double the size of llama 3 8b

[agile jay]

A good 20B model will be really nice for the new Qualcomm laptops

[half venture]

Yeah

I genuinely think llama 8b is not there yet

In terms of like knowledge

Compared to haiku

They need to up the parameters

Make a 14b

14b is nice as well

Or 16b

That's where the models seems to begin having good coherent personality

[agile jay]

Yep, at this current time.

[half venture]

No at all times

There is a limit in coherency

Based on the total amount of parameters

Llama 3 8b is already at limit

And clearly it shows

We need newer ways

[agile jay]

Couldn't that also be a data quality issue?

[half venture]

No

[agile jay]

But we haven't hit a quality limit

[half venture]

C.ai developer even talked about this

[half venture]

Yes

But

There seems to be layer issues

And mix up and match up

You will notice all smaller models have a tendency to hallucinate much more

Which breaks immersion

[agile jay]

That just sounds like architecture problems, and not parameter problems.

[half venture]

Yeah pretty much

But 14b model seems to atleast offer some sort of coherency

It's why mythomax trained a 14b model

[agile jay]

So basically we are at the saturation point of the current architectures for 8B models?

[half venture]

Yeah Google seems to have noticed this

Meta is just pushing

As much as possible

Props to that

[agile jay]

Well I don't think it's gonna take too long for new ones.

[warm cave]

Depends on the scope of the model as well, I think that sub 8b will be very powerful, especially for specialized tasks, like github style copilot code complese, and for other uses like a Siri, etc

[half venture]

Yeah

[half venture]

Currently no

In current architecture

We need a better way to manage attention

The only company i hope will find a way is Google

Since it's in their financial interest to aim at cheaper models

Openai clearly doesn't care about smaller models

[warm cave]

Really, it's already good at the first one, just a little way to go

[agile jay]

To be honest if your goal is AGI, why bother?

Reach AGI and it can solve it...

[warm cave]

This one in particular impressed me

[half venture]

Yeah

It's good

[warm cave]

I just think the smallest models will be specialized, bc the point is to use power effectively, while larger models handle complex reasoning approaching AGI

[agile jay]

Yep, I can also imagine in the future that general models would create their own finetunes of smaller models.

And use them as agents.

[warm cave]

5 small models that get called for there specific job saves money compare to 1 big model, so all depends on use cases

[warm cave]

Crazy

[agile jay]

And they can generate the training data...

[warm cave]

Synthetic dataaa!

[agile jay]

Basically teacher AI's that make students to do simple tasks for them.

[warm cave]

Also the idea that you can have AI’s go head to head and improve beyond the quality of data we can currently provide

[agile jay]

Yep, they can just compete.

[warm cave]

To the death

2 gpt go in 1 agi comes out

[fading moth]

Did perplexity devs give any further news beyond what is pinned?

[warm cave]

Nothing new afaik

[fading moth]

Bleeeeeh

Ty

[warm cave]

Np

[wraith haven]

Why did perplexity choose to highlight Amazon’s covert operations in 2015, in today’s discovery page?

[warm cave]

Looks like the articles it used as sources are new

Well at least the first one

[wraith haven]

That’s the stuff I want to see

[warm cave]

Yeah a few were posted today

[wraith haven]

Recent developments in exposing corruption

Hopefully as perplexity gets bigger, their business relations won’t incentivize them to not do this

[devout bronze]

could also be because a lot of people were searching for it or a relevent topic?

[tame current]

As long as they give clear time lines for increasing limits.

[warm cave]

I'm not too long we will likely have Llama 3 400b, grok 1.0 (fine-tuned), and Gemini 1.5 (just kidding you Gemini lovers will never have it again, evil laugh) in perplexity

[livid mantle]

Meta LLAMA 4

[fading moth]

Iunno how I feel about grok

Bleh, I wish perplexity had a temperature slider. C'est la vie.

[wraith haven]

is that how it works?

[tame current]

Gemini is exceptional for creative writing. But the censorship.

And following context.

For example. If you're talking about crocodiles, and you're asking a follow up question with 'why?' It will go into the history of the word why, instead of answering why.

[hoary python]

Is opus down right now?

Most of my stuff is getting directed to sonnet

[wraith crow]

It’s not down, but temporarily limited to 30 uses a day.

[hoary python]

Is their a reasoning for it?

Backend repairs?

[wraith crow]

Please look at the pinned message

[hoary python]

Oh I see

People we abusing the API

So it’s getting further construction done to reduce/mitigate abuse

Good job for catching it