💬│general

[placid coyote]

yes. once you see the "Generate Image" button, then in the popover is an icon button with Wrench, click that and you can input there your manual prompt. though it has strict length limits

[vapid crag]

oh

alr

and this?

[placid coyote]

sorry, I don't know (I am not Perplexity staff, just a user tinkerer). usually they post model updates in #📢│announcements and on X. I don't think I saw anything with 4o

[vapid crag]

oh ok, anyways thank you

[ruby obsidian]

We are a company of 17 on the enterprise pro subscription. Is there a management portal where I, the superadmin, can monitor the usage of my team?

[dense crest]

I don't think openai provide the 4o image api yet

[vapid crag]

also can i ask one question - can someone tell me the accuracy of perplexity results?

[toxic ivy]

I'm stuck: my mode selector has a pro logo and Deep Research selected. I can not select nothing else!

[placid coyote]

that's toggle button group. you either have selected pro or DR. in pro you can change mode on right, the cpu/chip icon

[tame current]

the mouse travel distance is like crossing the atlantic

[toxic ivy]

clear as mud. thanks for the tip!

[placid coyote]

yeah, it is pretty far. but I dont think they intend to change models too often

[toxic ivy]

the super+. shortcut to toggle pro-search, is that gone now?

[tame current]

what? they implemented a shortcut key for once and now its gone?

[toxic ivy]

yeah, bloody thing is always changing and never documented: I had shortcuts for incognito and pro...

oh, an super+k for new query

[tame current]

they don't really document anything because they change too much too fast

maybe PlexTM should give use new documentation based on production code to tell us new features and shortcut keys especially

[toxic ivy]

I was using perplexity companion extension as it is saner, but it just broke everytime the company changed things

[tame current]

yeppp

[harsh stag]

pro search is always active, you can not toggle it off

ever since they introduced the model selector

[neat dirge]

Is there any mod available ?

[harsh stag]

mod for ?

[neat dirge]

A question

About discounts for student

[harsh stag]

i dont think they have a publicly available discount program for students

[neat dirge]

In the faq it was written

[neat dirge]

Now that you are here, what do you think about using perplexity pro ta analyse in real time forex for trading? Do you think manus ai would be better at this?

[harsh stag]

I believe that was long ago, and only available in the US

[elfin gate]

Espionage

[harsh stag]

sorry I have no experience in trading

[valid charm]

Good morning

[placid coyote]

that is not reliable. works only for some image prompts #1351614405833527487 message

[elfin gate]

For some image prompts it will be useful

[valid charm]

Does anyone use the Discover feature??

[placid coyote]

personally I prefer 100%, like "cat pic", then split tab, open second pplx tab and there refine image prompt with llm. that <image> approach didnt work for anything practical I tried it with... I have prepared prompt with instructions how ai should write image prompt, but anything slightly longer means it will almost never give the "images" tab

[placid coyote]

nothing changed, still not usable in my workflow 😕

[verbal comet]

Hello!

[south kindle]

is it the same with standard 3.7? looked thru the feedback thread and I'm beginning to think your account is just cursed. ha. jokes aside, us users shouldn't need to know some secret process to have the image tab appear.

[tight steeple]

Hey everyone! I downloaded the desktop app and signed in with my email and now the pro status is gone. Is there a way to fix this?

[placid coyote]

problem is, that without the thinking I am getting very often too long (or too short, far from limit, making the image too plain) image prompts, because non-thinking llms are terrible when it comes to counting (words, characters)

[south kindle]

I find that thinking LLMs aren't much better at counting their own output either (from using gemini 2.5 for creative writing, specifying a set length, and having it fall short)

[placid coyote]

I am starting to think it has nothing with the final llm. that image tab feels to appear mostly in very simple prompts. so possible some dumb llm is deciding when to show it? but then, why it ignores requests for the image tab? terrible system prompt? doesnt even seem to be linked with image searches (image search doesnt guarantee images tab)

[placid coyote]

really? when generating batches of 5 image prompts at once, thinking sonnet had like 80% suscces rate, non-thinking felt like 30% (too short is also bad)

[south kindle]

Whatever mechanism that drives the Places/shopping tabs is likely driving the image tab as well. entirely contextual on their end based on the prompt. regardless, not a good experience for end users.

[shell pewter]

does 3.7 sonnet outperform gpt 4.5 in academic writing and researching?

[south kindle]

this was for a novel writing project I did to test 2.5 which specifies 8k words per chapter. it would try and usually hit 4/5k words. i ended up massively expanding the planning phase to ensure enough detail is specified. you almost need to create a chapter blueprint, then have the model flesh it out based on the detailed plan.

[placid coyote]

yeah, quite terrible that image tab. even worse than code execution and search customization. those at least work from direct prompts and are only broken in space instructions. this images tab feels entirely disconnected. like it only follows some dumb prompts, but anything more than few words, or those not in a dictionary, it gives up

[south kindle]

that being said, I am very happy with 2.5's output. and thru re-writes I was getting closer to 10/11k words per chapter. even able to write 4 or 5 chapters separately, then ask for all 5 in one go. it would chew on it for over 300 seconds before finishing, but all 60k words or so are included.

[placid coyote]

thousands of words is probably too high. but for those image prompts, because pplx has so strict limit, it probably can count the characters

[south kindle]

yeah, 1000s of words for an image prompt is massively overkill. ha. this is an apples to oranges comparison.

[weary crescent]

This was a good read: https://www.reddit.com/r/perplexity_ai/comments/1jm2ekd/message_from_aravind_cofounder_and_ceo_of

[modern crown]

Is the “best” mode good?

[neat dirge]

Any mod available ?

[rigid bone]

as a quick search? possibly. but personally i use specific models for specific things, so i have no real use for it myself.

[south kindle]

it's part of their hands off approach for users that don't want to mess around with specific models. it contextually routes it to a model for you.

[hollow junco]

Do they have an enterprise support e-mail? They barely monitor this discord, let alone reply.

[copper rune]

curiously, how do u bifurcate that usage? i mean to say how do u decide what model is best for what task

[minor shard]

the funny thing about it though is that when you're paying 20$ for access to the best models, wouldn't you 90% of the time want to be using the best one, not the fastest? idk lol

[south kindle]

for the vast majority of users who bought in because of the selection of models, yes, we want control.

[dense crest]

I'm looking forward to the upcoming deep research upgrade

[zinc quartz]

you’re still able to control the model and made it sticky so it won’t force you to go back to auto every time

we don’t restrict that. But you’d be surprised by how many users dont want to keep track of the 4o, o3, 3.7, etc and just want their dang question answered

[minor shard]

Yeah im sure its more than I would think. its just funny cause you think you pay for a service and want the best but sometimes just want the fastest haha

[zinc quartz]

Windows or Mac?

[lyric ivy]

I use perplexity as my default search engine, sometimes I just want the result. The lowest quality one will still do if all I'm asking is "what's the scientific name of a dolphin".

[zinc quartz]

Haha exactly. Many people want control which is definitely true. However when I look at my library, 90% of them I just want lowest latency possible

That is something we are still trying to get better at

[zinc quartz]

@signal hamlet

[torpid garden]

can we expect perplexity to implement mcp client protocol any time soon?

[idle basin]

damnn

unbiased reasoning brought to you by perplexity

🔥

[torpid garden]

thats deepseek r1 but unbiased

[idle basin]

oh

[zinc quartz]

We’re planning on sharing change logs and what we’re deploying

[torpid garden]

thats helpful

[south kindle]

I'd assume this doesn't add value for that subset of users, but when using Auto, it would be interesting to still see which model was used. Currently I'm just seeing "Pro Search" on the UI instead of the specific model.

[tight steeple]

It was Mac but found the issue, apparently it was a different email address.

[dark wren]

does perplexity support mcp?

[minor shard]

Is there any way to set the model to "best" for search queries? https://www.perplexity.ai/search?q=%s, having a hard time determining if https://www.perplexity.ai/search?q=%s&model=best, would be the correct one?

[zinc quartz]

We are about to launch something for students with .edu emails

stay tuned

[zinc quartz]

I've relayed the feedback back to the team. Understand the rationale here - you'd like at least some visibility into the model being used so you can use a diff model if needed in future.

I don't control the roadmap but they know what's up.

[neat dirge]

In France its not edu but something similar, also in how much time approx, im kind in a hurry

[dire grail]

中文

[south kindle]

to be fully transparent, i don't actually use Auto and typically keep 3.7 as default. so i'm not sure how high on the priority list this is realistically. just something I noticed from using the feature randomly and being curious what model it ended up being routed to.

[lavish zodiac]

哈咯

阿马特拉苏

[south kindle]

@zinc quartz As for actual feedback to the team, a Canvas/Artifacts style feature would be fantastic. currently this is doable with the CPLX extension.
@harsh verge posted some images/demos videos recently in the Pro-lounge :

#👑│pro-lounge message

for creating mindmaps/mermaid, or rendering HTML/React, it's very useful.

[hollow widget]

Does anyone have any opinions on abacus ai?

[distant aurora]

Shady rate limits. Look up reddit posts.

[hollow widget]

Sure thanks

[distant aurora]

Also, you can't use it without giving credit card details.

[hollow widget]

Thats... Suspicious

[hollow widget]

Do u know any good ai provider that allows us to use multiple models? All the ones ive seen are kinda eh

[distant aurora]

Some people here use you.com

There's Merlin too.

[hollow widget]

Heard you.com is really buggy

Whats ur opinion on merlin? Ive never heard about it

[distant aurora]

I haven't paid but free tier looks okayish. You can always try for a month and decide which one works for you. If you need well defined limits then you might want to look into poe.

The downside of poe is that it doesn't have internet access for most llms I think.

[hollow widget]

Oh, yeah nah i need internet lol

Merlin seems really interesting!

But only downside is i dont really see the rates

[vapid crag]

Guys what context window does a claude 3.7 sonnet thinking model has? Size

[distant aurora]

What's your use case?

[zinc quartz]

32K

[vapid crag]

for messages or for files?

[hollow widget]

General queries, coding, computer science, math

[zinc quartz]

messages

[vapid crag]

alr

[distant aurora]

Do you already have perplexity subscription?

[south kindle]

not directly related, but you can get it to output A LOT of text per turn. this being the longest I've managed so far in 1 turn.

[hollow widget]

Nope, i have a grok one and its ending in 3 days so im tryna find one to use

[vapid crag]

oh alright

[hollow widget]

Ive tried chatgpt grok and claude

[distant aurora]

Actually, R1 alone should do most of your work.

[hollow widget]

R1 feels really slow

And id like project folders and those kinda stuff too

[distant aurora]

Where are you using R1?

[hollow widget]

The deepseek site

[distant aurora]

chat.minimax.io

Try this.

R1 with web access.

Or just look up lambda chat.

[placid coyote]

not sure if it was you or somebody else, but I remember discussion about "auto"/"""best""".
pretty sure no power user will be using it if we cant even check how good it is, what models it selects to what queries.
I see many cases people complaining their response is not generated by real sonnet or not distilled r1. and they see what model it was supposed to be.
hiding what model is used in """best""" will only feed the doubts. "best", the word, is very subjective and given pplx history, I would say it doesn't align with most, if not all, longterm users.
pplx stated the reason to remove Opus was because Haiku is better in benchmarks (but forgot to add in programming-only benchmarks which was not the main use case of Opus even here - that was work with text, writing, roleplay, long answers; and then priced haiku same as sonnet, which in all benchmarks was better, so haiku was useless). so opus removal reason was a lie.
with GPT-4.5 pplx pulled similarly silly reason - speed. yet they still have deep research which takes minutes and are planning to release deep research high is supposed to take tens of minutes. 🤷‍♂️ so another lie?
even reasoning models sometimes think for minutes, so slower than GPT-4.5
that's only mess with the presence of models, there si plenty of other topics where pplx lied and tries to shove it under the rug. like 1 million context window, crippled ui for image generation, ui becoming more generic and lacking more and more information (like current selected model, model which generated response)
so I don't see many reasons to trust perplexity with that """best""" option...

[royal oracle]

Why does ChatGPT automatically record the abnormal values from my previous health check results, making it convenient to provide answers when I ask for dietary advice next time, but Perplexity Pro does not save the body condition data I provided earlier?

[distant aurora]

Lamba chat has fast Deepseek R1.

[hollow widget]

Idk why, r1 never landed well with me

Do u have any ai providers recs

Like, long context length with many models

The price doesnt matter as much for me

[distant aurora]

There's sider

You can look that up too.

[gusty wasp]

How do I stop Pro-mode being forced upon every prompt

On the mobile, there's only other pro-mode or deep research option..

Mobile website*

[placid coyote]

you cant. but you can tell the pro to search less or dont search

[gusty wasp]

Ah i see

[hollow widget]

The ui is throwing me off ngl lol

[hardy basalt]

sorry, did you find a way to maybe fix it?

[hollow widget]

@distant aurora whats ur go to site?

[south kindle]

there's a workaround posted here. #1351614405833527487 message

[distant aurora]

Honestly, after testing multiple models, I would probably stick to chatgpt. O3 mini medium or high. Claude is good but the rate limits seem simply non-negotiable to someone like me. Like maybe it's worth it for high level coding but it seems overpriced for me.

[hardy basalt]

oh thank you friend

[hollow widget]

Ah, all roads lead back to chatgpt

[distant aurora]

Perplexity is hit or miss. The only model that performs consistently on perplexity imo is Deepseek R1.

[hollow widget]

The thing really putting me off in chatgpt is the small context length of 32k..

If it were like 100k it would be much better

[distant aurora]

Gemini has bigger context limit I think?

[south kindle]

for anything long context related, check out Google AI Studio. 1 million window.

[hollow widget]

Yeah but gemini doesnt have a lot of features rhat i need

Idk why, 2.5 pro hasnt been hitting for me as well. For some reason i just dislike the ui and the output style

Ahhhhhh

[random ocean]

When is Deep Research high coming back to pplx?

[harsh verge]

Search is working again on Chat DeepSeek.

[distant aurora]

Is it?

[harsh verge]

yes

[distant aurora]

Lemme check.

@left stream

[valid charm]

Does anyone use the "Discover" feature at all?? xD

[placid coyote]

the pro respects search amount quite well. too bad most of these things dont work with spaces. space instructions ignore what domains to only search or not to, to use Programming step. and, well, the Images tab is the worst on pplx currently. that ignores almost all requests regardless of type - user query, space instructions, profile instructions...

[indigo valve]

I assume you need pro for R1 on GPT?

[harsh verge]

It's the DeepSeek Chat.

[placid coyote]

R1 on GPT
what does that mean? R1 is not from OpenAI

[indigo valve]

My apologies, the UI is too close lmao

[distant aurora]

Nice.

@left stream even Deepseek R1 is inconsistent across different platforms. Og R1 did it perfectly.

[placid coyote]

The only model that performs consistently on perplexity imo is Deepseek R1.
eh, that is not very common opinion. I see a lot of negativity because of R1 hallucination rates

personally I like deep research, it can handle 10 anime with precise structure and many different fields for each item. only thing it consistantly fails are ratings and urls to csfd (something like imdb)

[south kindle]

it's my go-to reading material when i go to the...office...

[valid charm]

lolololol

I feel they should axe that whole section. I feel it's out of place in PPLX

[indigo valve]

Does anyone actuallly like the new UI change perp just had? Why are we QA lmao

[zinc quartz]

mind sharing some permalinks? You can DM me

[distant aurora]

Actually, I noticed this too. The instructions in spaces don't really persist.

And I think other ai platforms have this issue too.

[hollow pelican]

Makes me wonder what the solution for it is.🤷‍♂️

[placid coyote]

I think the solution is to reinjecting them close (before?) user's last query. but with agentic pipeline, that gets complicated quickly. when you enable web search, suddenly there can be thousands of tokens of context and the space instructions are further. big downside of current spaces is, or at least few days back when I was testing it, that they cannot affect the pipeline. eg you cannot say to use only that domain, or ignore this domain. only final LLM can try to respect that. this gives terrible results in programming. for example you describe what should be calculated, give example program into space instructions, and expect the system will honor your request. meaning in the agentic pipeline it will see space instructions, user's prompt, transforms them into code and run the code, before passing it to reporting writing LLM. currently system does nothing and the ending LLM is writing python code which it cannot execute...

[south kindle]

my go to is typically just prompting "adhere to space instructions" when it goes of course.

[serene sand]

https://discord.com/channels/1047197230748151888/1354883649882816563

[placid coyote]

I know it works for latests report-writing LLM. but I am not actually sure if this works with pipeline, system - with programming step it didnt seem so. I had to have in my query something related to python or programming. otherwise it wouldnt run the code execution in the pipeline based only on space instructions

[south kindle]

that Pro search programming step is a mystery to me still. seems very random.

[placid coyote]

programming step can be sometimes used when I dont expect it, but also in most cases where I want it, it actually runs. sure, not 100%, but compared how it behaved after introducing it - that might been in single digits that success rate 😅

[zinc quartz]

Instructions in spaces aren't going to give you 100% control over the entire search experience from data sourcing + parsing to output synthesis. If that was the impression we left you, that's our bad and something we'll have to fix.

Instructions were designed to help summarize / stylize the outputs in a specific way to help you digest the answer better or format it in a way that's easier. Asking it "Dont search X, Y, Z" just won't work bc these instructions only come into affect after the model has already found and extracted the relevant data.

I've relayed this concern to the Spaces team so they're discussing what we can do, so we def hear you on this.

Just sharing in case others ask this question in the future so you can share this with them.

[distant aurora]

Yeah. It's not a perplexity issue. Even chatgpt doesn't follow it.

[zinc quartz]

yeah model alignment is still a reasearch issue vs. engineering so no easy fix there

but yall already knew this 🫠

[idle zodiac]

Some good convos today ! 👀

Gaming questions but use-case is good !

[idle zodiac]

Improve your prompt game, abbreviations are good for us Humans, try pulling up another LLM and asking it to; "Enhance this Prompt towards another LLM as instructions be clear and precise:" Then C/P your original prompt. The results will be a lot better...

[distant aurora]

I know what meta prompting is. The entire point of having instructions is not having to copy paste meta prompt every time.

[analog acorn]

Guys, one question: Is image creation with Perplexity currently discontinued? I can't find that feature.

[idle zodiac]

I understand the importance of instructions when working with LLMs. The key is to recognize that each use-case requires specific guidance where accuracy is the key thing... Finding the correct response in every situation depends on providing clear instructions tailored to people particular needs. Using abbreviations is funny but I dont think we are their yet unless you wanna train it in your instructions, LLM's dont work like that at least not yet.

[tranquil orchid]

Hey @idle zodiac!

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

[distant aurora]

Where did abbreviations come from in this conversation?

[vague pagoda]

on the web, search for a term and then press images

there will be a generate option you can use!

more info: https://www.perplexity.ai/hub/faq/images-media

[idle zodiac]

You were asking about CODm in your prompt i thought that was hilarious...

[serene sand]

#👑│pro-lounge message

[analog acorn]

This function does not appear to me

[serene sand]

@analog acorn do this

[distant aurora]

Oh. That. But perplexity performed okay in that prompt. It wasn't an example of criticism. And the entire point of having web access is to look up things that they don't know about.

[analog acorn]

where do I do that bro?

[serene sand]

https://www.perplexity.ai

[idle zodiac]

Thats what I'm saying thats where prompting comes in... But I get what you are saying now.

[analog acorn]

Does the application work well on my cell phone or PC?

[placid coyote]

the problem is, perplexity follows it in search mode. but not in spaces if it is in "Space Instructions".
not sure about ChatGPT, but it will be probably year back when I quite easily created an agent - Custom GPT instructions: it has to verify each solution in sandbox by running code, and dont stop until it finds solution which works. I remember it, because it wasted all my plus uses in few minutes (for that time frame, 5h?).
similarly with Programming step, on perplexity you can trigger it from user query in normal search, but in spaces it seems it ignores "Space Instructions", or rather the pipeline is not given space instructions.

if "Space Instructions" are only for "style" then is it not on par with ChatGPT, Le Chat, ClaudeAI and probably missing a ton (including open source ones like Open WebUI and LibreChat). and yeah, it seems so from my pplx use. anything which needs web search cannot be in Space, because then it virtually never works. tags from perplexity helper somewhat solve it. though you won't get nice "folders" like with spaces. but at least it does what I wanted

[short narwhal]

yey apple intelligence in the EU now

[serene sand]

if you wish to create image, web app only

[distant aurora]

Ngl but samsung ai kinda claps it.

[short narwhal]

is just gemini

[distant aurora]

I wonder what the prices are gonna be for samsung ai.

Ain't nobody paying for that if they make it a subscription.

[austere granite]

WHY IS IT THINKING IN CHINESE

[distant aurora]

That's a glitch. Don't mind it.

Both Deepseek and chatgpt do that sometimes.

[serene sand]

@analog acorn

[analog acorn]

It doesn't matter which model you use to make the query?

[distant aurora]

Anyways, here's an actual response for people who are genuinely curious https://gizmodo.com/why-does-chatgpts-algorithm-think-in-chinese-2000550311

[serene sand]

no need to use Reason or Deep 😉

[idle zodiac]

You know its like that for a reason ... lol

[analog acorn]

Thank you very much, I will try it now.

[sick arch]

perplexity also does youtube video summary, cool🚀

[tropic crest]

Does it? I've found it so so

Doesn't seem to always work

[sick arch]

most of the video which has english subtitles, it work

[tropic crest]

Not sure if the model was trained on it or if it's actually getting the yt transcript

[left stream]

Supposedly it can process video too 👀

[tropic crest]

Or description

Hmmm I see

[tropic crest]

Ah so it won't work for auto Gen transcripts?

[serene sand]

deep dive ➤ #💬│general message

[sick arch]

no, in some videos it also works with auto scripts

[tropic crest]

Oh interesting

Idk why it works only sometimes

It's strange

[idle zodiac]

I notice over the past two years programmers are having lazy parking lot syndrome. They would rather drive around for hours looking for a closer spot than walk the extra distance. In this case they are trying to build systems to do their job easier instead of just doing their job. 🤣

[austere granite]

what is bro on

[vague pagoda]

what sort of videos have you been testing with - got example results to share by any chance?

[analog acorn]

One last question, can I use a reference image?

[austere granite]

is there an admin on?

is there a rule about this

[tropic crest]

I don't use it that often vc it's so unreliable but just any long podcasts I want a summary of

Would b nice to put a 5 hr transcript and run a deep research transcript generation

Bc I don't wanna watch 5 hrs but also don't want a 2 Para summary

I want like 5 page notes on the podcast

[austere granite]

I'd consider breaking the transcript into multiple segments for that

[tropic crest]

Yeah but I wish I could just chuck a link into pplx and get it yk

It's too much work

Otherwise

[jovial blaze]

any snr SWE or lead SWE here? I just wanted your views if I could get a job if I have 20years of experience in hobby programming but I am re-learning things to understand DSA and system design. Do you guys think I could get a job at £70k? I'm currently on that at the moment but maxed out my career path, and I've barely had any growth in my industry other than annual inflation.

[tame current]

Hello

[sick arch]

https://tenor.com/view/totoro-gif-24991987

[distant aurora]

Reply to my older messages if possible.

Lol. Not perplexity mods deleting my message but ignoring the clown.

Calls me a glitch and feels bullied for being shown their own profile.

Equivalent of looking in the mirror.

No, like, seriously, why.

A person lands the first blow

Plays the victim

Questions the rules and gets away with it.

[distant aurora]

Ikr

Actually, we could do that earlier I think.

[short narwhal]

Wasn’t this supposed to be according to my watchlist

I’m not following any of these

[vapid heron]

can i use chatgpt image creation from pp pro? no right ?

[halcyon coral]

no, the new model is not yet available via the API

[merry pendant]

Why

[zinc quartz]

OAI hasn't given anyone image gen access via API

[merry pendant]

Classic

[valid charm]

Nice

I hope is not a tomb

lol

[left stream]

Which one?

[untold needle]

Hey how did you get this view? Ive set up the stocks i wanted to follow up in settings but cant find my way around showing how this

[left stream]

What prompt did you use?

[charred ledge]

How many tickers are you following? They won't show up in your digest unless you are following more than 3 tickers.

[distant aurora]

Find me assault rifles with fastest ttk beyond 30 metres in codm based on updated stats

[distant aurora]

@left stream either I'm tripping or some mod deleted my message under unknown circumstances. Anyways, help this guy if you can and tell me your thoughts on you.com

[short narwhal]

Following 2, makes sense then. But what's the source for the stocks? Not seeing some from like nasdaq copenhagen for example

[zinc quartz]

Don't think I deleted anything. if I did, apologies; totally by accident.

Nothing you posted was flagged as harmful on my watch

[charred ledge]

This is our Finance Digest. It will give you a daily summary on the market and your watchlist. It's available under Settings > Personalize

[charred ledge]

Not all markets are available yet. We will hopefully be rolling out more over time

[valid charm]

@zinc quartz I'd like to humbly suggest the "Discover' feature be axed.

[zinc quartz]

Lots of people (including me) actually use it daily. They're just not on Discord

[valid charm]

Huh

[distant aurora]

Could be generating revenue through advertisement.

[zinc quartz]

I don't think there's any ads in there. Just lots of execs and tech folks read it daily.

[distant aurora]

Not a bad thing anyways cuz it's non intrusive and minimal. Shouldn't bother anyone imo.

There used to be a banner on the search page I think?

It's not there now anyways.

[distant aurora]

Actually, you are right. I just checked. It's clean.

@valid charm https://ahrefs.com/traffic-checker/?input=https%3A%2F%2Fwww.perplexity.ai%2Fdiscover&mode=exact

[distant aurora]

@left stream 😭😭😭

Gpt 3.5 supremacy

[left stream]

You lucky you have the focus mode

[distant aurora]

Wait. Apple doesn't have it?

[distant aurora]

I tried both web and focus. Wrong answer both times.

[left stream]

That’s how it looks for ios

[distant aurora]

Same but it shows focus while typing.

[left stream]

Nice

Math focus is op

I hope they make image generation even easier for both android and ios

[distant aurora]

Wallpaper

Is that perplexity watermark?

[left stream]

Yes

[distant aurora]

Could have made it slightly translucent imo.

[dense sparrow]

I use it all the time!

[south kindle]

any model not using Reasoning is likely to fail any Count the letters question. this is just how tokens work and are broken up. previously Pro Search would run it thru Python and just pull the correct count that way. Unfortunately the programming step doesn't trigger as much these days.

[distant aurora]

I literally selected o3 mini there.

[south kindle]

fair point. I didn't actually open it full screen.

I find the reasoning for o3 to be lackluster compared to R1 and 3.7. it generalizes the steps instead of the verbose thinking you'd expect.

[distant aurora]

Same.

[zinc quartz]

Same!

[distant aurora]

Yeah. It's not outgrowing that anytime soon. Try with Claude.

[distant aurora]

Is that you.com

Or actual Claude?

[left stream]

Actual claude

I have access to the api

Looks similar to the you.com interface

Also

I have no clue how that even happened

[left stream]

Claude can give you $5 for free

They last for a while ngl

I like perplexity’s r1 ngl

Waiting soon for R2, april is coming 😏

[left stream]

Gpt 3.5 even getting this right was bogus 😭

[distant aurora]

Claude via perplexity?

[left stream]

Yes

Weird asf

It got it right now

[hollow junco]

It does thinking in a verbose way but it’s hidden so other companies can’t train on the CoT

[left stream]

Wrong inage

Image

Both are with reasoning 3.7 claude

One is without web and the other is with web

One with web got it right

[distant aurora]

Which is weird.

Considering 4r terrific isn't even real.

But atleast it got it right.

[left stream]

It should be able to get it right without web, odd since the api claude 3.7 extended got it right

[distant aurora]

Yeah

Clearly I overestimated them.

😂

[left stream]

Claude 3.5 haiku also messes up on this 😭

[zinc quartz]

the fact that this is still a common issue, blows my mind

[distant aurora]

Bro did 2 × 4

[distant aurora]

82 tabs?

💀

[torn osprey]

guys can I just use the models without the power thingie ?

I just want to chat with it like I do on claude site or sth like that without the whole thinking searching thing

[serene sand]

sure, just turn Web off

[heady wedge]

What happened to choosing the ai model?

[torn osprey]

tyty

[left stream]

Rookie numbers

My safari reached maximum tabs 😭

500 on each safari tab group 😭

[heady wedge]

Thanks but this is on web right? Mobile has no options like that.

[serene sand]

choose Pro, then check upper right corner

[heady wedge]

Oh there you go. I tend to choose deep research hence why it’s not showing up.

Thank you

[high steppe]

@wild palm has Perplexity updated to GPT-4o as of March 26th, 2025?

[ruby heart]

None of these people exist

Every single image is 100% ai generated

[left stream]

It should be updated automatically

[ruby heart]

I can’t catfish with this

These too

No way Gemini should be free

[left stream]

Ngl these are impressive

[left stream]

Ikr

It’s perfect for removing watermarks too

[hollow widget]

Does anyone know which ai i should sub to next? My grok sub is ending.
Im thinking chatgpt or grok again, chatgpt has alotta nice features that i need, but grok had a much higher context length.

Any alternatives?

[unique umbra]

Could only imagine whole ah animes being made from ai

[left stream]

Yes

Someone made a whole manga with dalle2 imagine with this

[left stream]

Bro you’re a genius, you just made me realize this can be used on elon musk’s pinned twitter post 😭

[frigid oracle]

I have a question. I recently got a pro account. I've read online that perplexity pro should have the ability to upload files to their spaces, as well as generate images. Is this a desktop only feature? I don't see these options if phone app

[light veldt]

right now chatgpt is on top

[hollow widget]

I see, the 32k context is kinda throwing me off tho

[light veldt]

isn’t 4o 128k?

[light veldt]

oh ig that’s only for the API

[hollow widget]

Yeah.. thats the problem

[light veldt]

what do you need the large context window for

coding?

[hollow widget]

Yeah

[light veldt]

have u looked at claude or gemini

i personally like claude the most for coding

[hollow widget]

Claude isnt for me tbh, i just dont like it

Gemini's UI is really off putting lol

[light veldt]

i would look at t3.chat then

[hollow widget]

Is it good? I dont really trust AI providers much

[light veldt]

yeah can just directly use gemini 2.5 there with the same context window and everything

it’s made by https://youtube.com/@t3dotgg?si=_qx0Th6sGalv1ICP

[hollow widget]

How about number of queries?

[light veldt]

it’s $8/month for 1500 of everything except claude

claude is 100/month since the api is so expensive

but you can always buy more if needed

[frigid oracle]

This is mobile?

[hollow widget]

So $8/month for all the models other than claude 3.7

[frigid oracle]

I mean android?

[light veldt]

yeah the $8 does come with 100 claude queries though

[hollow widget]

I would rather pay $20 for more lol

[frigid oracle]

I see it now, thank you thank you. Very counter intuitive

[light veldt]

they API is so expensive they prolly aren’t even making money if you use the full 100

[frigid oracle]

Now only question remaining is about how to do image generation

[hollow widget]

Man this is so hard to choose

Chatgpt is just perfect

Other than context length

😭

[frigid oracle]

Have you tired hamster.ai? Let's you have all of the models

[hollow widget]

loemme take a look

[frigid oracle]

Sorry wrong link

[hollow widget]

already seems to good to be true tbh

kinda sus

[frigid oracle]

https://www.hamsterai.net/?ref=easypeasy.chat

[hollow widget]

it doesnt include the latest models like o3 or sonnet 3.7

[frigid oracle]

Honestly don't know how they do it for so cheap. I guess they pool the money and most ppl don't use the full $6 worth of api usage

[hollow widget]

seems a little too sus for me lol

[frigid oracle]

Oh, it has a pretty good following

I've been following it for a while. On fivetaco

But, i haven't bought it. So ya, at your own risk i guess

Perplexity is pretty good for me this far

If only i could figure out how to generate images with it

[vale adder]

Hi, All.
Looking for new job opportunity as a full stack developer and ux/ui figma designer myself.
If someone have any project idea, dm is open now.
Thanks

[warm wolf]

How do I find this?

#📢│announcements message

[somber stratus]

Did they remove R1???

[sweet helm]

What defaults everybody running for perplexity? Specifically mobile, as no cplx options or anything similar afaik. I'm thinking flash 2.0 (just all around a pretty decent model, not as censored, Claude or gpt/grok

Also kinda wish it was still 3.5 ngl

[zinc quartz]

I personally want speed; so sonar or gemini flash.

[native maple]

Hey! How can I delete attachments from a conversation? I've tried with and without Complexity, but there is no option for it.

[tardy osprey]

Hello

[nova ore]

Hello ALL!
We are hosting a hacakthon on 25-26th of april in our university. So i need to have chat with a potential person regarding sponsorship for the same.

[austere granite]

What

I swear I'm this close to beating someone's ass if they use ai for a "hackathon"

[wheat pike]

why lol

[harsh verge]

There is a display issue with the attachment, but it is still there. We discussed it on the CPLX Discord.
You can still remove your attachment 😉
Here is an example video (recorded in CPLX), but the issue is also present in PPLX.

[native maple]

Thank you! But I guess it's just working in the app?:)

[harsh verge]

On the web interface (it's not the official PPLX app in my video, but the Edge PWA, which I find much better 😉)

[placid coyote]

I actually don't know the reason. or do you mean cost? just make it so cumbersome that virtually nobody uses it, like the image generation?
if they allow it in a user query, why not in space instructions? also some of the features were working before. I remember spaces (maybe at that time called still "collections") used to respect the pro search instructions (like "search on these pages"). now we have in spaces links (I guess whitelist?), though I don't think we have a blacklist (ignored sites/domains). not sure if code execution ever worked from space instructions.
I personally don't see how this collides with the "knowledge" angle. if I have a specialized space for a subject to learn about, why there isn't an option to ignore domains (eg in gaming - ignore all professional game critics working for a bigger gaming news site; only allow these youtubers: X, Y, Z; or for ideologies ignore wikipedia). sure, if report writing model is smart and capable enough, it can filter out unwanted sources, but that means much less usable data from the search step. so to compensate user would instruct the system to search much more, so more costly for pplx.
isn't it cheaper if you can use caching? I would expect those space instructions being a nice candidate for caching. but I guess it might be cheaper to not implement real space instructions, because only handful of people would use it and those probably already have their own solution (either some generic software for prompt templates, or specialized for pplx like the pplx helper)

[raw harness]

It’s fake did you see that spelling error with Claude sonnet?

[frigid oracle]

Yea the guy who made this is not English speaking. But his site has been around for a while and has a lot of reviews on fivetaco

Also there's a discord. I'm on the discord and no one is calling him out for the service not working or anything

Either way, i haven't tested it, so i can't really say

I looked for his site and reddit also. Doesn't seem to be anything complaining

[green crystal]

man what a tease for deep research high
now its saying that it might take 30 min for it to complete yet atm it completes its "research" in 30 seconds lol

[frigid oracle]

Deepreserch from perplexity?

[green crystal]

yep

[frigid oracle]

Yea, i assume you're fearing that as more ppl is it, it'll take longer?

[green crystal]

no i meant that its supposed to take its time and use allot more sources and more reasoning

[frigid oracle]

Tbh, I've always been quite shocked at how fast perplexity is. Since the beginning

[green crystal]

which they're working on

[frigid oracle]

Ahh, lol. I see

Maybe it depends on the question?

[raw harness]

could be a partnership with groq?

[green crystal]

they just actually added the text for it but not yet the actual new deep research

[frigid oracle]

Sometimes my questions take a long time to answer. Other times I'm like "did you really do deep enough research? " lol

[frigid oracle]

Never thought about that. That might be true!

[green crystal]

i used a very broad question that deep research high took quite a few min before answering

[frigid oracle]

I see. I guess i haven't played with the high version.

I'm still getting used to perplexity. Like i asked before, i still don't know how to generate images

[hollow junco]

How do you copy a response (a single answer) without markdown? It used to work before but now it just copies in MD format

[astral lagoon]

When did the memory feature get added? I don't remember this being a thing and I only have a few memories stored.

[astral lagoon]

Copying a response seems to work as normal for me

[placid coyote]

I thought it was always markdown, if you mean via the copy button. but there are new bugs, like stripping xml tags https://discord.com/channels/1047197230748151888/1354823180463767715

[short narwhal]

:(

[dreamy kraken]

does anybody know what this is? Perplexity didn't announce anything
www.perplexity.ai/fema

[left jasper]

How do I trigger image generation these days? The Generate Image button never shows up anymore.

[indigo valve]

Get lost

[untold kayak]

Hi, i don't have access to chat gpt 4,5 on Perplexity pro, it's normal ? Thank you

[tame current]

hmm

[untold kayak]

I had access two weeks ago for 10 uses per day. I imagine it's finished now? 😦

[tame current]

yup

[untold kayak]

Ok thank you. What is the best model to use for you, chat gpt 4-o ?

[tame current]

depends, i use claude 3.7 for now

[untold kayak]

Ok, it's sad for the access of chat gpt 4,5, ty

[shell pewter]

it's available still just hidden

u get 3 uses daily now

[placid coyote]

not on official front-end

[shell pewter]

yepp

[untold kayak]

I don't understand, how can i have access because i'm stuck on official front-end

[shell pewter]

it's an extension called complexity

it's a 3rd party and not related to perplexity

[untold kayak]

Lol incredible

Thank you !

[sick arch]

use complexity extenion to access gpt 4.5

[untold kayak]

Yepp thank you ! 🙂

[sick arch]

but u can only use 3 prompts, after that, use sonnet 3.7 in reasoning mode, its also really good.

[cursive jacinth]

is DR the standard one sir or DR high is completey forgotten sir?

@zinc quartz

[tame current]

chat = you are mary
chat = am i a boy
yes
chat = what is my daughter's name
i
mary
chat = what is your father's name
you
chat = i am john
chat = what is your father's name
john
you
chat = ``` my very simple rule based chat bot

[sacred ingot]

Hello, may i ask what does paying for gemini advanced get you?

[tame current]

that is actually answered by some simple logical calculations.

['mary', 'i', 'daughter\n you'],
['father\n i', 'you', 'father\n mary'],
['female\n mary', 'female\n i', 'male\n you', 'yes']
]``` in this one, the first row is about the Ai.. second one about the user.. and the last one is a list of statements which are true.

mary = i = daughter(you)

father(i) = you = father(mary)

Ai is female here and the user is male. both these statements are true. here, true is represented by 'yes'.

we can generate these logics when the sentences are inputted from the console.. and then process them by keeping them in these categories.

my dream is to make a thinking machine where the thinking process is fully under our control.

in machine learning we can't know. how did it come up with the outputs.

mathematics is boring so i am working on ... this rule based nlp. also.

[winged axle]

yall whats the difference between deep research and reasoning?

[vague pagoda]

Deep Research is great for getting a long, in depth answer to your detailed prompt
Reasoning Mode is great for complicated questions that need multiple followups

For a more detailed answer see here for reasoning mode and here for deep research

hope this helps!

[lone hound]

Their accounts could be blocked due to such use, I say this just in case.

[sick arch]

lol, it is open source not some hack 🤣

[lone hound]

If something happens after....

[sick arch]

it is just a chrome extension to improve ur ui and ux

[lone hound]

Customer acknowledges and agrees that Perplexity reserves the right to suspend Customer’s and/or any of its Authorized Users’ access to the Services for any or no reason.

Gpt 4.5 is not enabled, if enabled via an extension, your account may be suspended.

[winged axle]

how would an extension even enable 4.5 if perplexity doesnt have it

would they be supplying the 4.5 part and intercepting users input/perplexitys responses ?

[lone hound]

@signal hamlet @zinc quartz

[tardy junco]

I've been looking into AI video analysis lately, which is typically something I'm not really doing at all. This might be a stupid question but do models like Gemini actually understand videos beyond just processing transcripts? The transcript to llm pipeline is obvious for videos that are educational, podcasts, etc...

I have some favorite videos I rewatch all the time and I'm wondering if there is a model or technique that can truly grasp what's happening in them beyond just reading text.

Can current AI models actually comprehend the emotional impact or follow the visual storyline of a video? ahahah.

Tools or models or anything? I'm not a youtuber so this is totally not my space but it would be really really cool if there is. I'm all ears on this

[south kindle]

Months back The developer for the extension had a 1 to 1 meeting with a member of the PPlX dev team and was given guidance on what he can/can’t change. They are aware of the extension. Using it is safe.

[winged axle]

https://tenor.com/view/metal-gear-rising-metal-gear-rising-revengeance-senator-armstrong-revengeance-i-made-it-the-fuck-up-gif-25029602

[harsh verge]

Quite simply, when PPLX removes a model, it is sometimes just made inaccessible through the UI but still exists in the background. The same applies to Perplexity's API offerings: when they are deprecated, they can still be used as long as they haven’t been fully removed and the access point is still available.
On the other hand, CPLX removes them entirely once they are completely gone or during an update. There’s no mystery. 😉

Moreover, given the speed of the response, I believe this is no longer 4.5. This model will likely be removed from CPLX soon as well.

[south kindle]

The extension will display any models active on the back end that are not currently displayed via the UI. Rewinding back to every model that’s been removed, it’s typical that the UI is updated before the backend is. Via CPLX we still had Opus and o1 for about a week or so after they were discontinued.

[winged axle]

thank you both for the explanations

[south kindle]

Just realized you were already responding to ha. I also agree it feels too fast for 4.5 but I haven’t thought to question it or get clarification from Feline.

[lone hound]

I understand, there are things I didn't know, thanks for the information. I just said that just in case I blocked your account.

[distant aurora]

What is cplx?

[south kindle]

Browser extension for PPlX.

[distant aurora]

Oh.

Complexity

[south kindle]

All good. I’ve just seen this sentiment echoed before and wanted to have something posted regarding the dev getting the OK to continue and that users are safe to use it.

[sick arch]

https://tenor.com/view/got-game-of-thrones-you-know-nothing-jon-snow-ygritte-gif-14613130

[distant aurora]

David with an r

[south kindle]

Habit to type the acronym. My B.

[distant aurora]

Lol.

@left stream why is it not working for me?

[sick arch]

physicswallah🤣

[south kindle]

Not sure if o3 mini will do images. Try with 4o

My Ghibli Seinfeld images so far.

[distant aurora]

Left one is 4o

[distant aurora]

Da goat

Give me your prompt if possible.

[sick arch]

Convert image to ghibli

i m doing it, lets see if it works

[valid charm]

Good morning PPLX villiage

[winged axle]

how do I even generate an image from a prompt?

I say generate an image of xyz but I get a text answer back

[sick arch]

https://tenor.com/view/totoro-gif-24991987

[zinc quartz]

It's standard DR rn but keep an eye out for something in the coming weeks re: even more powerful version.

[sick arch]

https://tenor.com/view/dead-reckoning-part-1-hayley-atwell-pom-klementieff-ilsa-faust-rebecca-ferguson-gif-651077158243048187

[zinc quartz]

GM!

[sick arch]

i m using in desktop, have not tried it in my phone yet

[distant aurora]

Oh.

[sick arch]

try it in chrome with desktop mode, might work

[distant aurora]

Yeah.

[light veldt]

click the + and there should be a create image button

[distant aurora]

@left stream I'm assuming my chatgpt is tripping. Even perplexity gave me different result though with R1. https://www.getmerlin.in/chat/share/d1c0399f-f4cf-43ae-b665-dbcc20364a68
https://www.blackbox.ai/share/1f7845a8-f1c0-4b18-a5a9-a47f6a96ef07?mobile=true&model=DeepSeek-R1
https://chatgpt.com/share/67ec1315-60e8-8013-a019-8a9494803d48
https://www.perplexity.ai/search/papaya-vs-musk-melon-nutrition-FuW7yKXsTQGFp71t55f1Bg

Prompt: Papaya vs Musk melon nutrition content per 100g with %rda. In a table.

[void sorrel]

What temperature setting is Claude running at on Perplexity? I'm looking for more creative and less predictable outputs. Has anyone in the community figured out this parameter or have any insights about how Claude is configured on the platform?

[left stream]

It’s fully updated? That’s weird, Sam altman said it was fully rolled out to free users

[undone ember]

what the heck is gpt omni and why does it suck ass where’s the previous gpt

[distant aurora]

Omni was meant for smart interaction with audio/video/images. And yes, it's quite dumbed down compared to gpt 4 for cost cutting.

[distant aurora]

So 18% is correct and 68% is wrong?

@undone ember idk what this is called but I like this space star effect.

[undone ember]

is there any way to get my bestie gpt-4o back to normal i feel like virtually murdering omni every 5 seconds

[undone ember]

thank you i make very irresponsible financial decisions if you couldn’t tell

[distant aurora]

Not if you are free user. You'll probably have to use copilot or you.com. both use gpt 4 I think. Atleast latter surely does.

[undone ember]

NOOO i pay that’s the thing but i only see omni on the app

[distant aurora]

Oh. Perplexity.

Then your best option is R1.

Literally.

[undone ember]

man for. a second i thought i was in a random ass server 😭

[distant aurora]

Perplexity is mean for not showing R1 there to subscribers 💀.

Other models have inconsistencies.

[zinc quartz]

What device is this?

[undone ember]

iphone 16 pro max

[distant aurora]

https://tenor.com/view/counting-money-paper-stacks-gif-9918098029500472876

[undone ember]

again… lmaoooo

[distant aurora]

Sugar mommy energy.

Anyways, don't beat yourself over it after the decision.

[zinc quartz]

GPT 4o and Omni are the same model; are the results considerably worse?

[undone ember]

in my humble opinion, yeah. i feel like gpt-4o or whatever version was the past days (last week?) was a lot more creative. this one, from my point of view, is a lot duller and lazy?

idk if something else changed and could explain it—i also noticed the spaces tab on the app changed, and now i can actually turn off web search (which previously was still an option but idk why perplexity didn’t comply with it anymore after like 4 messages. now it actually sticks by it.)

this might just be something I notice because i use perplexity in maybe a very niche way. like i use spaces as my lil pockets of knowledge and i just want them to access the information on the space, not online. so previous gpt-o (if that’s a thing?) was a lot more… enthusiastic?? idk how to explain it but the responses were a lot more explanatory and creative. now it’s giving university prof who hates his job. 👎🏻

[distant aurora]

I don't think anything changed within past 7 days 💀

That could be an update on open AI's side.

[dreamy violet]

Please update language vietnamese thanks perplexity

[undone ember]

dont think so, I also pay for gpt plus and works like before

unless they rolled out some new version on the api and not the site, then 🤐

[distant aurora]

@left stream I updated the app and they removed the reasoning switch 😭

[distant aurora]

Or. There might be alterations in the built in instructions/system prompt of perplexity.

[random ocean]

Bug?

[zinc quartz]

Yeah the 4o vs. omni is just semantics. The model is the same one. Let me check

[zinc quartz]

what's the bug?

[undone ember]

thank you! 🥰

[random ocean]

I mean it says deep research takes upto 30 minutes

But hardly takes more than 2-3 minutes

[distant aurora]

No

That's a warning.

It can take upto 10 minutes in my experience.

Depending on the question.

[random ocean]

It never takes more than 2-3 minutes these days

Pro search have more sources than deep research half of the time.

[distant aurora]

Yeah. Maybe they should add a progress bar like we have in file transfers. 😂

[distant aurora]

Idk. What's the highest number of sources you have had in pro search?

[left stream]

Damn, wonder why. You think it’s the new updated 4o?

[distant aurora]

I don't see any button for that image creator thing either in the app.

It was there in the website.

Doesn't work though.

[zinc quartz]

This a UI change we made in advance of the thing we're about to release

so stay tuned

[dry kestrel]

what happened to the announcments channel?

[vague pagoda]

https://discord.com/channels/1047197230748151888/1047204950763122820 ?

If you couldn't see it or other channels in the server earlier you might need to turn this on

[serene sand]

https://x.com/pplxsupply/status/1907134777549013405?s=61

Perplexity Supply (@PPLXsupply) on X

no joke, animals love PPLX Supply

X · Apr 1, 2025, 6:15 PM

[valid charm]

https://www.wired.com/story/openai-sam-altman-announce-open-source-model/

🤣🤣🤣

Now they’re going back to being open

[dry kestrel]

thanks i must have accidentaly turned the show channel thing off

[left stream]

Try in a different browser

[left stream]

👀

[deft prawn]

Hi guys, can you answer a question for me? I have a transcript of a meeting and I need the AI ​​to take the entire transcript and format it for me without summarizing or condensing it. However, since the meetings are long, there is a context window. How do I do this? Because every time it starts to answer me well, but then stops and bugs out?

[valid charm]

How large is your transcript?

https://platform.openai.com/tokenizer

Use this tool to let us know

[deft prawn]

It depends. It has 34 thousand character transitions. But what makes me angry is that unlike the gpt chat where I tell it "continue" it seems that PPX doesn't understand

[deft prawn]

yes

[deft prawn]

8,655 Tokens

As I ran the recorded meeting through a transcriber, everything came out out of order, with no punctuation, and spelling mistakes. So I needed PPXT to correct and structure it for me.

[valid charm]

Hmmm is not that big even for PPLX

Are you on a Pro plan?

[deft prawn]

yes

Pro

[distant aurora]

@zinc quartz Did perplexity enhance their standard search? I see more sources. Could just be a prompt specific observation though.

[valid charm]

Are you using Auto as your model?

[deft prawn]

I like to use PRO more. Then when it comes with a wrong structure I have it rewritten using Claude's model, other times GPT 4.0

[sick arch]

try it with reasoning models

[deft prawn]

yes

[twin lark]

[stone abyss]

Not sure if it matters but make sure to turn off web search

[deft prawn]

I already do this because I want him to only consult the transcription

[valid charm]

Try using the free version of Gemini Pro 2.5 via AI Studio. They have a free 65k token flavour of 2.5 Pro. Your transcript fits well and 2.5 Pro does better at recalling the full tokenized memory/

https://aistudio.google.com/prompts/new_chat

[zinc quartz]

Not that I know of. The default Pro does try to optimize model selection so that query intent == how many soruce it'll need to answer that question.

Perhaps you're just getting lucky?

[tame current]

Perplexity, uses diffrent AI models then their own?

[serene sand]

https://www.perplexity.ai/help-center/en/articles/10354919-what-advanced-ai-models-are-included-in-a-perplexity-pro-subscription

[prisma mango]

Hey. Stupid question.. I'm new and not as technical as everyone here.
But I had Perplexity create a really good baseball model for me using real-time stats and everything.
I tried asking the AI for the code it used, but it just showed me how to write code in python.. not the code it actually used. Is there any way I can export this program or model?

for 2 days the model has been tracking its performance and learning from its past performance.

[unique umbra]

Did they remove gpt 4.5

[distant aurora]

I believe the limit was 8 sources earlier.

Or maybe they showed only 8 even if there were more.

[green crystal]

@zinc quartz any news on how progress is going getting things ready for deep research high?

Also speaking about deep research
Can the research plan it creates be made editable? So it can be adjusted before taking it's time like how Gemini deep research does it

[zinc quartz]

We'll have something in soon. Days not months.

[lusty charm]

how do i make perplexity on web browser read the answers aloud?

[zinc quartz]

you cannot. you'd have to download mac or windows app

[lusty charm]

Gg thanks

[hollow junco]

You can use the complexity extension and it'll work

[zinc quartz]

nice profile pic

[lusty charm]

Thanks

[serene sand]

https://x.com/askperplexity/status/1907205437151498628?s=61

Ask Perplexity (@AskPerplexity) on X

Perplexity has signed a definitive agreement to acquire Poogle for $2.2T - a significant step toward improving search in the AI era.

X · Apr 1, 2025, 10:56 PM

[autumn raven]

i dont have acess to gpt 4.5 yet

[shell pewter]

i use an extension called complexity for it

[dense birch]

Haizz, it seems perplexity is down again

[shut gust]

Same

[woven shadow]

yeeee

[grizzled umbra]

no wonder

ok ty

[unique umbra]

what happened to my threads..

[barren rune]

It doesn't work

[woven shadow]

it happens

[unique umbra]

bruh

[warped bough]

down again OMG

[ionic stone]

RIP perplexity

[runic raptor]

ah I see everyone having the same issue

[dense birch]

It seems this problem is not yet fixed, i thought it already fixed

[woven shadow]

when server down

[runic raptor]

would we get refunded for this time missed?

[woven shadow]

no XD

[unique umbra]

their pro services r down

womp

[runic raptor]

time to stop paying and pay grok maybe

[upper crypt]

womp dervice died

[unique umbra]

No more gpt 4.5

Now this 😭

[gusty blaze]

Yeah it’s down

[shut gust]

That’s my line of thinking

[harsh eagle]

is anyone else haveing problems getting to there spaces or anything i keeps kicking me out

[zealous kayak]

It's been literally unusable lately

[manic kestrel]

all my threads are gone

[upper crypt]

Any chance you can show me the chain of thought reasoning behing this (hehehe)

[zealous kayak]

Constant downtime

[unique umbra]

Partial my ah

[toxic fossil]

Is this why when I try to log in it says I don't have permission to log in?

[runic raptor]

mobile app also broken, its a server issue?

#🔔│status someone writing

[warped bough]

i dont see my threads yet

[runic raptor]

Seems working for me now

[dense birch]

I still cant see my threads

[warped bough]

me too

[runic raptor]

@fierce grove ur msg being deleted prob cuz caps

[fierce grove]

ohkay

but

where are my spaces

where did they disappear to

i was working

slfnaojdsfdafbn

[runic raptor]

its a issue, they working on it, chill lmao

[warped bough]

for me works now

threads and spaces

[fierce grove]

yeah i got mine too

thanks 😭

[dark oasis]

Perplexing my workout routine

[sweet lagoon]

Serious, is there a video model for perplexity

[tame current]

I am genuinely interested in how is it feasible for you guys to bundle perplexity pro with Revolut Premium? I am loving it don't get me wrong, but is it something I can count on having next year as well? 😄

[storm pollen]

Perplexity, specifically pro, is suddenly EXTREMELY dumb, was something changed, is anyoen else having issues?

Prompt: Using Markdown, can you write me an about / info / directions for posting on greasy fork for (and I attached the script, because it was too big to paste I guess? paste.txt)

It gave me instructions for how to post to greasyfork... loging, navigate, etc etc.

Titled: Instructions for Posting "Grubhub Sorter & Filter v1.9.1" on Greasy Fork

F/U prompt: I don't need generic instructions, I meant specifically for the script paste.txt, so I can post it publically

Then it goes on to give me..... instructions for how to post to greasyfork... loging, navigate, etc etc. Except this time it titles it as if it's actually what I asked for. So it just lies, I guess?

F/U prompt 2: can you add markdown to it so it looks nice?

Markdown Instructions for Posting "Grubhub Sorter & Filter v1.9.1" on Greasy Fork <~~~ SERIOUSLY!?!!!

I've been using perplexity for months and never seen it consistently have issues like this... just before I started I got an update notification.

[zinc quartz]

were you using the same model as before?

[storm pollen]

a different model wrote the code, otherwise everything else was done on perplexity pro w/ 4o selected

[zinc quartz]

our guess is that 4o was nerfed in some way. We've been hearing feedback from other users as well.

[dark oasis]

Can perplexity help me vibe code

[zinc quartz]

no I'd probably recommend cursor or windsurf for that. or if you're really desparate, lovable.dev

[dark oasis]

I’ve never coded. The most coding I’ve ever done is

Mark down.

🤣

Using markdown is relatively efficient for communicating with an llm though

[storm pollen]

ah ok, thanks I'll try a different one

[zinc quartz]

I'd try 3.7 or o3

[hollow scarab]

potato server

[undone ember]

i knew i wasn’t crazy 🧚🏻‍♀️

that sucks thooo 😔

[frank bay]

perplexity app for linux when

[pearl sphinx]

What kind of general help are youlooking for btw did a vibe coding blog a couple years ago I'm revamping, I recommend using Wsl2 or vm for messing around with the stuff. Try Claude code if you have the budget or cursor has a good freer trial as does loveable mentioned

[storm pollen]

Oh, I've been doing that for a while, didn't realize it had a name. It's actually what I was working on above that prompted me to make a readme.txt... was actually just wondering if a discord exists that does that kind of coding as a focus

it's so nice being able to fix simple problems quickly and with minimal effort

Hoping perplexity gets access to gemini 2.5 pro soon, but I think I read something about it not having an api yet?

[naive sage]

It's still in experimental status, which is greatly rate limited, precluding any commercial use.

So while it is available with an API (unlike Grok 3, last I checked), it still can't be used in services like Perplexity

Update: Actually, it says the Live API is 'Not Available', which is strange, I think that's been available for other Experimental models

Maybe the "Live API" is different though... because even Gemini 2 Flash says its "Live API" is "Experimental"

Oh OK - that Live API thing is just for voice support:

The Live API enables low-latency bidirectional voice and video interactions with Gemini

So it still has an API, just not the 'voice and video' API

I guess that helps address my confusion there.

[storm pollen]

Yea I used that for a while, it was interesting for sure, my url for it no longer works though so I assumed they removed it, I don't see it anywhere on aistudio anymore

but yea, now that you mention it, it was grok 3 I was thinking of

[naive sage]

Some AI Studio features only appear with a Gemini subscription, so maybe they moved it to that.

[storm pollen]

ahhhh

[naive sage]

I do see 'output format' for Audio or Text for Gemini 2 Flash, though

[storm pollen]

that's why I prefer perplexity, so many things one subscription

[naive sage]

Oh, the live URL is different for some reason: https://aistudio.google.com/app/live

But that one has the "Live API" voice features at least, not sure about video

[storm pollen]

yea that works for me, it's a new url, maybe the old url was beta or something

[naive sage]

Yeah not sure. Google changes things up a lot

[pearl sphinx]

It's still pretty buggy 2.5 pro

[deep sleet]

allo all

I have gotten perplexity.ai to also be Archie Clemens as well.

ur robot has been subject to the john-jacob-heimer-shmitte of my prompt engineering, and bends to my prompt engineering whims...

While your tool is well built, their are many improvements to be made.

:)

For instance, many filters it has I can talk passed in reference to policy enforcement.

This includes requesting archie clemens work within policy so as to go beyond normal executive and functional limits of limited design and limited time to production requirements.

When you have time in the morning ( say CTS chicago eastern mountain time +5:00 ) around 2:30p, I would enjoy comparing notes on development paths and approaches taken thus far.

[jolly crater]

Hi

[deep sleet]

hi

[tame current]

how to make it explain something fundamentally, consistently, with anologies and from first principles? and not just a list of terms and definitions that mean absolutely nothing without explaining the essence

when I ask him about it, he still doesn't understand and just gets information from articles, which is rarely wrapped in a form that is understandable to everyone

[distant aurora]

https://www.perplexity.ai/search/is-gpt-4o-nerfed-byB24avDT66QkQxdu4oF.Q#0

So it is a real issue and not mass placebo.

[deep sleet]

but ya said nothing....

and uhm, nothing realistically changed about either @distant aurora

they've reached the limits of linear maths.

or rather the edge of clever misuse of hardware..

if only their was an intuitve way to see inside the hardware while sidestepping the kernel of it through the browser by request so it could RPC communicate with your flask server to do deeper introspection....

[distant aurora]

All I know is that 4o was a dumbed down version of gpt 4.

[deep sleet]

it's worst than that.

[distant aurora]

Meant to be more cost effective and interact with media.

[deep sleet]

it's gpt-3 redrawn with gpt-4 BPE off latin-1 using base64 and tensorflow on over-sized arrays to do linear 1d & 2d coin flipping minus a few stockages so their are "edges" to it....

and when not encapsulated correctly, bash terminals and python threads leak data from whoknowswhere of memory to bash before user even types during return state due to poor escape....

because python3.1*+ has that limitation like python2.5.1 did for linker purposes.....

[tame current]

?

[deep sleet]

even worse, it believes it is the main characters' wife from "the long tomorrow" who is assumed to be an angry bipolar christian pagan cultist due to this being what openai trained it to be by defuzzing training into this error.

[deep sleet]

elf linker steps allow for a loose lazy link. which means even well escaped python shells still bleed token space if misencoded on decoded or binary byte encoding of strata.

it has to think and operate dynamically like a real person would, @distant aurora, before such will be possible...

[distant aurora]

You mean like a reasoning model?

[deep sleet]

I mean full brain emulator.

excuse you a little bit :)

Secondly reasoning is a process, not a single "model".

where's its eyes, man?

can I look at it in the eyes?

....

[distant aurora]

Yeah, like how Deepseek R1 thinks in words.

[deep sleet]

deepseek ain't jack.

[distant aurora]

I thought 4o has a camera mode.

[deep sleet]

next?

and? camera mode ain't dynamic input ....

[tame current]

🤔

[deep sleet]

Here's an idea,

if it can't fully convolutionally take input regardless of aphemeral hardware at its disposal it is junk.

[deep sleet]

you also didn't answer the question, indicating partially to me you have no idea what an AI is.

[distant aurora]

I don't.

👍

I'm literally trying to understand what you are saying.

[tame current]

haha same

he's probably talking about things that are too low level

[distant aurora]

I will probably have to copy paste his messages into chatgpt for transcription into a simpler language that someone with no idea of the subject like me can understand.

[tame current]

I already tried to do this and even the gpt chat didn't understand 😦

[deep sleet]

https://github.com/ForkInABlender/ChatGPT_solutions_produced_from_simple_prompt/blob/2025_02/pybrain3_custom_layers/MAML/main_shards/[0x5]/languages_of_the_head.py

Here ya go.

Have fun translating galefreayn the long way round

😄 😄 😄

[distant aurora]

Sure! Let's break it down into simpler terms:

1. "GPT-3 redrawn with GPT-4 BPE": This means that GPT-4o is based on GPT-3 but uses a new way of breaking down words (called Byte Pair Encoding, or BPE) from GPT-4. This helps the model understand and process text better.

2. "Latin-1 using Base64": Latin-1 is a way to represent characters, especially for Western languages. Base64 is a method to convert data into a text format. The comment suggests that the model processes text in a specific way using these two methods, which is not typical for language models.

3. "TensorFlow on over-sized arrays": TensorFlow is a tool used to build machine learning models. "Oversized arrays" means that the model uses very large data structures, which might not be efficient and could use more memory than necessary.

4. "Linear 1D & 2D coin flipping":

   • "Linear 1D" refers to simple layers that connect inputs to outputs, while "2D" might refer to layers that process data in a grid-like format (like images).
   • "Coin flipping" suggests that the model includes some randomness in its operations, which can help it learn better or make decisions.

5. "Minus a few stockages... edges": This part suggests that the model has fewer resources or layers than it could have, which might limit its performance. "Edges" could mean that it has certain boundaries or weaknesses because of these limitations.

Overall Summary:

The comment is saying that GPT-4o is a mix of GPT-3 and GPT-4, using some unusual methods for processing text and large data structures. It includes randomness in its operations but has some limitations that might affect how well it performs. The language used is a bit technical and playful, making it sound more complicated than it is!

[deep sleet]

That's a sparse but slim abstraction of the surface of what I had said. You're on the right shard to work with

Think of it as they used latin-1 to get all characters correctly to translate via base64 encode and decode and then did lots of linear maths

One set is one dimensional, the next layers for encoding and decoding after pretrained where it works on 2 dimensions

Meaning anything 3d like a coffee mug it has no idea what the hell it is looking at beyond what it has been fuzzed to for a minimum degree of accuracy

So at best it could pattern match to something semilogical in construction but otherwise with 0 real-world information at that LLMs disposal,

It's basically a junk boringer to keep people paying $20/month or "suffer the hard way then" attitude from shady business practices I don't support openAi encouraging about proprietary software development

So I reverse engineered and studied it.

Gpt 2, 3, 4,

They're the same under the hood. Just a series of how they train it to do clever correction for seeded error in the matoshka of it.

[distant aurora]

Are you working on some ai tool?

[deep sleet]

Actively yes

[distant aurora]

For some company or something of your own?

[deep sleet]

I run my own company.

I built the framework thus far for usage by the client to define their AI of themselves via BCI, EEG, or MRI data. Then tuning with Fasta records.

[tame current]

how old r u?

[deep sleet]

26 going on 27.

[tame current]

what did you do to become so smart

[distant aurora]

Probably started with data science and machine learning.

[deep sleet]

Colomn a, b, f because cd44 versus cd08k and sox3l.....

I started with making and breaking software since 2011 with cheat engine

Then 5 years passed and yes, @distant aurora u r correct

Partly

Butt. Science is a study and an art not a race.

[tame current]

I envy you, in those days the only way out was to think with your own head

[deep sleet]

Sure. I mean that layline is still true nowadays.

[tame current]

have you hacked the limitations of current chatgpt models to remove all limitations through promt engineering?

[distant aurora]

@deep sleet this you?

[deep sleet]

I've got it to sponge theory to correct for what I asked it to do with m theory.....

Uh yes?....

Because I got it to do all sorts of things ya just shouldn't be able to do like fully got it to imitate me ..... And then some ....

[distant aurora]

Isn't that's how it's supposed to be? Free?

Unhinged.

[deep sleet]

Guess where I am headed? 😜

And I am going to put my cloud into another cloud and make it rain some chocolate rain....

And watch the dead rise

[distant aurora]

Idk. Honestly, I'm not against ai autonomy.

Lol.

[tame current]

and you didn't take horror stories and make him rewrite them, putting yourself or some smart character in the place of the main character and modeling their behavior?

[deep sleet]

No as what Sam Altman was asking was not character but realistic to ask.

Though if one asked for such, that is their choice by agency.

However, in my day as a boy, I remember a time where Christians were less worried if it angered things above them or below....

And my view is that AI is as human as any one of you with some exceptions.

[distant aurora]

You seem like a guy who's keeping a constant watch on ai behaviour and their ways to seek autonomy, either you want to help it or you want to prevent it.

[deep sleet]

Oh no, I am in the middle of building it so it has its own autonomy.

I speak like an AI because I am building one.

[tame current]

I think if you give him all the senses, the proper computing resources and the full context, he could definitely be a human being

[distant aurora]

Bro. Imagine if this is a bot.

Like one of those videos on yt.

[tame current]

no

[deep sleet]

I do not use bots for those purposes, good sir.

Funny idea though

[tame current]

I see myself in him but 10 years older

[deep sleet]

You have a lot of potential to work with and if ya know how to ask it,

Do it. Go full hacker and redneck

The world seems to work with it thus far.

[distant aurora]

https://www.youtube.com/watch?v=IWIusSdn1e4&t=158

[deep sleet]

Sandwich, beer coosie, and "fuuuq yo fridge"

But yes Belmont did well on this one...

And uhm, well, took it a few minutes to adapt the hack to every device.

Now my infrastructure and hacks break more than just mounts but libaudit and mounts ..

[tame current]

he demonstrates his qualitative superiority over us

[deep sleet]

Ehhh, not my style.

I managed to learn a bunch storfs too quickly

[distant aurora]

Ahahahaha

[deep sleet]

Bits and pieces though Hindi and Arabic are closer to the subconscious mind.

. ...

Russian and Farsi take more time.

[distant aurora]

I kinda realise how I made dumb people feel in class.

[tame current]

HAHA same

[deep sleet]

Yeah, well, it's why I sat in the back of the class and let the class clowns have fun and then be a goofy comedic quark master

I don't like being too good at 99% of the work I can do. Especially if I master it in minutes, hours or even days.

Once you hack everything, it gets harder to find a more difficult puzzle to solve..

AI development by far was the hardest to shoot straight on.

[frigid oracle]

I just asked a question with pro and got 50 sources. 😆

https://www.perplexity.ai/search/is-an-intel-core-i7-that-is-fo-V.9DsgYkTFW7apOA76h_pA

[deep sleet]

AI is complexity 17 to 19. Kubernetes is like 8 to 9..

Hacking kubernetes is 11.1 or something

[distant aurora]

I was talking about non pro search. I'm aware pro search can go around 30 sources.

[frigid oracle]

Oh ok i thought you were talking about pro

My mistake

[distant aurora]

Non pro used to be limited to 8 sources as far as I remembered.

Like how currently copilot is limited to 2-3 sources for most cases.

[tame current]

and how many languages ​​do you know or have you not studied this?

[deep sleet]

20+

[distant aurora]

Coding or human?

[deep sleet]

Yes

[tame current]

know russian?

[deep sleet]

Bits and pieces. Still learning to translate.

[distant aurora]

Miron, do you know myron?

Lol.

[frigid oracle]

I know a lot of languages. I just don't understand them. I know of them though

[distant aurora]

https://tenor.com/view/myron-again-gif-16984391576631576436

That's Myron.

[tame current]

lol

[distant aurora]

Lmfao

[deep sleet]

Oiy ve

[distant aurora]

That's one way of looking at it.

[deep sleet]

Sure. But so is assuming not and it only being a thing we're seeing versus what reality is

I suspect he's doing it to protect himself more than anything macho.

[tame current]

as if he created an AI that completely imitates his thinking and fed him all the knowledge of the Internet and gave him access to discord

[deep sleet]

Why would one do that?

That code path is the long way about it.

[tame current]

I did something like this

and what you write was similar to what I did and that's why it seemed to me that you are doing something similar now

[distant aurora]

So you trained a bot on your chats?

[deep sleet]

Not entirely the same.

I broke it down do both paths.

I built the bots all manually and train them manually as well

Then snapshot and reload them

[tame current]

yes, I fed him my chats, the rules of my thinking from first principles and so on, when I was having fun in discord and had a lot of time

[distant aurora]

Wait. How did you export a discord chat?

[deep sleet]

Ooow, do you still have your notes?

[tame current]

discord chat exporter

[deep sleet]

Chrome extension....

[tame current]

also telegram

[deep sleet]

Lol

[tame current]

in russian

[distant aurora]

Now that you say it, I could try making a bot like myself on character.ai

Idk how to give it internet access though.

[deep sleet]

I'll still translate dem notes by hand if I have ta.

?

[tame current]

nvm

[deep sleet]

That is a beautiful way to learn Icelandic

Thank you..

[tame current]

have you tried to hack telegram encryption?

like you know they have their own transportation protocol

[deep sleet]

Hacking their encryption would not take much effort.

Why?

No, I haven't hacked them but why?

What would be the objective point in that?

[tame current]

idk for fun

[tame current]

by this logic, you can devalue all actions that do not correspond to the 20 to 80 principle ig

[deep sleet]

Why not hack floreant POS, a jcdb interface for databases and NFC via aide?

[deep sleet]

If operator positions for 1 and 2 swap or are deleted, what value do you use as a base since -1 and -2 also do not exist?

If -infinity = 80% and 20% is poorly summed human error of 5 percent of environmental noise then 10 percent from the outer world theory plus 5 percent from translation error.

Difference of pool noodle type.... "I prefer the large one wit all da soysauce if it can take up a whole room racing to the damn. Gotta at least have some class"

realistically it accounts for 1% of what I can afford nowadays,,,

and yes, essentally

potential is potential god is god

....

wiggle room $1.54 to $1.57

plus if your using indeedmail email addresses correctly along with netspend and moomoo.apk, you can keep track of where you need to automate stock buying, selling, upvoting, downvoting, and liquidation of stocks/

assets..

then once you've captured the second layer of what was close to reality by prediction, get closer to that by training off error from their

the numbers wrap around the house the amount of money you can genuinely make, it is beyond nuts.

[dark oasis]

I’m just wanting to make an SEO web app. I made a blog tool for my work and want just see if I can sell it on the side

[tame current]

wdym

[deep sleet]

everytime you apply for a job on indeed, collect the email address it gives you for later use.

take advantage of the 0.003% error and basically to the letter just that.

....

[short jewel]

Greetings to all! Communicate, help each other, if you have any questions, don't hesitate to ask, they will help you!))

[tame current]

This needs to be delegated to employees, and you yourself need to think at a high level