💬│general

[sleek vortex]

whereas pplx normal, this is entirely different

this is a follow up question

read_write_token and last_backend_uuid are used to track the conversation on backend instead of sending the messages

[cinder comet]

alot to decipher

[sleek vortex]

yeah

i dont think its worth changing how this works

do what other people are doing and just add to the ui

like new buttons and idk

[cinder comet]

https://corsproxy.io/?https://labs.perplexity.ai/

oh the captcha

[sleek vortex]

yeah no doesnt work on my other proxy either

this works tho??

😦

[cinder comet]

lol

[sleek vortex]

doesnt work in either proxy engine

well the server for one is hosted on cloudflare too

so idk?

turnstile is proof of work im pretty sure

they care more about browser fingerprint than rest

oh my

412 lies on this js based proxy loool

can you not just override .tostring?

ok whatever

going to sleep

goodnight

[cinder comet]

will look into it tomorrow

gn

[cinder comet]

thats readonly

[hardy willow]

HUGE problem that is ongoing and extremely common now with GPT-4o https://www.loom.com/share/8cc0ce3bff334b64b558f5a04bdcf2b2?sid=ff6eb1ac-2ba9-434b-ad1b-bc372d2ac9b0

[drifting bough]

Had the same issue, but with Opus everything is fine

[hardy willow]

Opus was my 100% exclusive LLM for over a full month, until 4o came out... I think we're still a full year away from "her" type AI, but 4o is superior in my opinion, so I'd hate to switch back for no other reason than there's an error or "bug" in the system. Can Perplexity address this?

[vivid kraken]

ok, why not offer llama 3 70b to free users through? or gemini 1.5 flash? they are cheap models

[pulsar sail]

To me Opus is still superior, might be just personal preference though...

Quick question: Do you use the "PRO" search often?

[drifting bough]

Opus is way superior, hopefully you'll realize that soon.

[hardy willow]

100% of the time

[hardy willow]

I used Opus, like I said exclusively for the entire month after it was released, I really liked it. I already switched Perplexity back to it about half-an-hour ago, but GPT-4o is ultimately the wave of the future with multi-modalities.

[drifting bough]

And what specific multimodal capabilities do you gain by using GPT-4o on Perplexity? When comparing a pure chatbot, Opus significantly surpasses 4o

Even the simple fact that Opus can only be used 30 times a day, while GPT-4o can be used 300 or 600 times, says a lot. It's not without reason that Opus is much more expensive than that pathetic GPT-4o with practically zero changes compared to GPT-4 Turbo, at least that's what I and many others have already experienced

[rancid dome]

Why this error when I generated only 3 images "Failed due to moderation. Please try a different query"? How do I fix it? Thanks!

[warm cave]

What prompts are you using, and are you using Dalle 3?

OpenAI added a lot of rules, one example is no celebrities

[rancid dome]

I am using Dalle 3. I was not aware of those rules. Thanks!

[warm cave]

no problem

[hushed dune]

Anyone else having issues with Perplexity not sending sign-in links?

[urban goblet]

Could be someone trying to get your account info

Be careful

Sign-in links can be faked to gain access to your account

[hushed dune]

I'm not recieving any links. The Perplexity site just states "A sign-in link has been sent to your email address." however no email ever gets sent to me. Sign-in links is how Perplexity login works...

[urban goblet]

Oh

No problem then

[hardy bay]

ah shi here we go again

[warm cave]

Yeah, but this one is scheduled

Hopefully its down so they can add unlimited Opus ✨

[vale ravine]

Hey folks,
what were y'all asking 💀
You done broke the little fella

[tame current]

I think they doing upgrade rn

Probably came back In 30 minutes.

[thorny canopy]

for the love of god plz lol

[digital mango]

that is what the announcement said

[idle mulch]

Back up here, but surprise surprise, no change in Opus message limit: https://www.perplexity.ai/p/api/v1/user/settings

So that´s a month now that it is "temporary" at 50. What a bad joke and false advertising.

[olive ermine]

Is that per day or month?

Per day that seems fine? That's what ChatGPT and Claude do

[tame current]

Dang

[warm cave]

I updated my api price table, many more options now days fs

[winged fiber]

Hi, What is the daily usage limit for Claude Opus? I have run out of limit

[warm cave]

50 per 24h 😦

😭

[tame current]

What did they say In pro advertisment banners?
"600 messages" what a load of bull.

[winged fiber]

Also, do you guys experienced quality differences B/W Claude Opus Available Here and available on the Claude.ai website? i mainly use it for coding and i noticed Claude opus( here) makes more coding mistakes than available on Claude.ai website

[cinder comet]

How much is the limits on claude website?

[warm cave]

Depends on how much context you use, at most you will get 45 per 5h and at least you will get 12 messages per 5h

If i give it full 200k then i am out in 12

if i keep it short and start new chats frequently then you will get closer to 45

[warm cave]

On Claude Pro the temperature is set higher and you have full 200k, outside of that they should be the same, but temp can make a big difference https://www.perplexity.ai/search/Explain-Temperature-in-gBEJK3Q4S6eXHijb8YGobg

[fading moth]

Daily is right. Temp is a huge setting.

[austere kestrel]

temp is set to 0 on one, and to 1 on the other. Which is which?

[cinder comet]

So basically perplexity is still the best provide for it

[fading moth]

Right

But temp hits things with objective answers much less notably than abstract or subjective ones.

[warm cave]

ahhh, depends, at least you get the option for 200k with claude pro, if you need creativity and long context it might be better to do Claude pro, but yeah the restrictions are tight

and with pplx you get 600 GPT4o when you run out of Opus

so for most people, yes

but also you get sonnet after you run out of Opus, and its 200k, so yeah, hard choice in my book

[wide mesa]

How long can Perplexity's prompts get before it just starts ignoring most of it?

[cinder comet]

Ive never run in limits on chatgpt gpt4o

Maybe at the begining

But im pretty sure ive used like 100 msg in 2 hours or so

[warm cave]

I’m pretty sure it 80 per 3h or something like that, a lot. I was just comparing PPLX and Claude Pro with the mention of 600 per day

I wonder what the limits will be for GPT-5

[austere kestrel]

yeah agree. though I still think sometimes the affect is less discernible than often assumed (including myself here ha)

a more open-ended follow-up – anyone going to take a stab?

[warm cave]

👀😆 I'm famous

[candid cairn]

I would like to see a temperature setting added to the Writing mode. Because it only returns the same results every time. For other searches, it is preferable to keep the temperature low to ensure accuracy, but not for Writing.

[livid mantle]

I would say 3rd send in thw conversation

[wide mesa]

I mean in the initial prompt when you're making a thread

[iron fulcrum]

The Web can't work like usual

[cursive rock]

anyone else having issues? queries not completing

API
web
android

[fluid mauve]

it wasnt working for me but it may have been a stale session from last night. a shutdown and restart of Chrome has me working fine now

[cursive rock]

ah ty will try that

[cursive rock]

Android app logged out and relogin and getting no text answers only the related search links and a headphone icon

using API with typingmind UI is working though

[fluid mauve]

I just tried the android app and it told me all about Gengis Khan. It seems to be working fine for me

[cursive rock]

ah this must be user error on my part aaaaaaa

ty

yeah it was me working now

[fluid mauve]

No worries. Glad you got it sorted

[tame current]

aaaaaah ddg bing yahoo ecosia everything is down

[fluid mauve]

Interesting. Where in the world are you? DDG works for me , Bing is slow, yahoo is slow I am in the UK

[tame current]

did you try searching ?

the website is up but the search itself is down

[fluid mauve]

bing search is working

[austere kestrel]

same here, aside from yahoo

bing is also working for me, but some old weird landing page

[south kindle]

bing is down for me

[fluid mauve]

yes, same, but if i enter something it will search

[tame current]

not for me

[restive osprey]

The one at the website paid sub, is amazing in my experience. I never used the payed Perplexity one.

[austere kestrel]

https://techcrunch.com/2024/05/23/bing-is-down-bringing-duckduckgo-and-ecosia-down-too

[fluid mauve]

Got to love a bit of dependancy hell. This stuff will only get worse as AI slips into our life

reminds me a lot of leftpad - https://www.theregister.com/2016/03/23/npm_left_pad_chaos/

[tame current]

meh

I had bad experience with node

[fluid mauve]

me too. it sucks balls

[tame current]

I very much prefer python

it just so straightforward

[fluid mauve]

same, but then again. Python is all i know. well and shell scripting in bash

[tame current]

"go" is good too

the programming language

even worse name than python but good language otherwise

[fluid mauve]

i really want to take a look at rust, but i am just not sure the time investment is worth it as i am not a professional coder.

[tame current]

go is easier afaik

[austere kestrel]

aha and that reminds of "how one drunk trader impacted the global oil market..." https://en.wikipedia.org/wiki/Oil_futures_drunk-trading_incident

[fluid mauve]

It is a hard life being a gtrader

Did i mention I hate running benchmarks. its nothing complicated, just comparing my desktop i9-14700KF, 192GB RAM and an Nvidia 4090 with24GB VRAM to my laptop i7-13700h, 16GB no usable GPU (Intel Iris Xe) I am running llm_benchmark against my ollama install. it took the PC about 3 mins to run all 36 of the tasks. my laptop is still on the first task and its been over an hoyr. To give you a different comparison i ran some stats last night using stable diffusionXL and comfyUI. Laptop took 73 mins to run the test suite. the desktop took 21 seconds. I think i might be waiting till tomorrow for my results on the laptop

[tame current]

then don't run benchmarks, easy

[fluid mauve]

it was easy on the dsktop

[sleek vortex]

lol

roll your own search engine scraper instead!

metasearch

[austere kestrel]

do what now?

[magic flax]

Hi everyone, i need help using perplexity pro. Can you help me here?

I want to generate an image with stable difusion xl but it is seemingly impossible

[foggy pecan]

So searching web with chatgpt 4o is paused apparently

[sleek vortex]

yeah bing is down

[cinder comet]

Still down?

[austere kestrel]

search / Bing seems to be yeah

though can still browse the web (i.e. not use Bing)

[sleek vortex]

this is huge for duckduckgo then

everyones realising they send all data to bing

im surprised pplx isnt down

i think they have their own search

wow

idk

[agile jay]

Perplexity has their own indexer.

[sleek vortex]

yeah

pplxbot it seems

thats pretty crazy

they started on bing api and somehow just ended up building their own web index

[vapid onyx]

Not complaining about that though 😅

[halcyon coral]

Please check https://discord.com/channels/1047197230748151888/1225557012532695060

[tame current]

do they actually use the crawler for indexing?

if they did, they would have webmaster tools

no?

perplexity is not on cloudflare radar either

[peak yoke]

And it can find stuff very precisely and accurately for the most part, even for some niche things I try to search

[sleek vortex]

@warm cave

lol

budget perplexity free

[tame current]

(perplexity is already free)

[sleek vortex]

(i know)

whys this honestly decent

like this is using a 7b model

[vapid onyx]

I know you’re talking about the response but I got one of the new iPads and coming from a 2018 model… it’s pretty crazy how fast they are. I think WWDC will dictate what happens with iPadOS and whether the value of these is worth it for some. I like it though. Only thing that’s hard is grabbing text from Discord.

[sleek vortex]

yeah

because i saw a few reviews

and i feel like

the thinning of the ipad was like the main feature

like what would they have it it wasnt for

• thinning
• tandem oled
• new pencil

depending on the os features they might announce, ittl decide if its an upgrade for lots of people

because everyone complains the m4 is useless on an ipad

which i think it is too, frankly

the only use i might have for it is games

but why would i spend £1.2k on a tablet then, would just get a pc

[vapid onyx]

Totally get that. The keyboard is also a solid upgrade from the last generation and only works with these M4 iPads

[sleek vortex]

Yeah but mainly for me

when i add storage, the keyboard, the pencil

it quickly approaches macbook pro price range

[vapid onyx]

Yep. I think the largest issue is the OS. If they can improve the OS to a point where it’s functional and allows for some of the more complex tasks, then I think it’ll be a positive experience.

[half venture]

Is it only me or is entire bing website completely down

Copilot as well

[tame current]

.

[tame current]

.

took you a while to notice

but yeah, down for a few hours

no explanation

[peak yoke]

Hello peeps, I have a question. If we show two images to a vision model similar topics but perhaps different variations. And ask it to point out which one is better, how does it decide?

[half venture]

Oh I see the small company Microsoft must have a hard time running bing I see

Like it's been 5 hours

[sleek vortex]

not bad at all

actually this has more sources than pplx free now lol

well i switched it to haiku just to see

and definitely better response quality...

[hallow linden]

It’ll probably say “what do you mean by better”

[agile jay]

24s though...

[sleek vortex]

https://www.perplexity.ai/search/is-the-ipad-aOgjoPv1QJ2u_5PHtkjvyg#0

this took 24 seconds

22*

[half venture]

Stop zerking over few seconds

What matters is the quality

[sleek vortex]

yeah

idk im trying to see if i can make an actual research agent

speed not the priority

[agile jay]

Which model are you using though?

[sleek vortex]

this test just used same model as pplx

before i was trying gemma/mixtral on groq

[sleek vortex]

claude 3 haiku

[sleek vortex]

i need to add a layer in my thing that does the whole

embed -> find rag chunks that are related -> use these instead

to reudce the amount sent to the model

[agile jay]

How did it take so long when using haiku?

[tame current]

are you using tavily ai?

[half venture]

I am kind of annoyed by openai not rolling out voice

[sleek vortex]

well im counting from enter to response end

[agile jay]

Blame black widow

[sleek vortex]

well i was using a custom rolled bing scraper but bing is down so i was trying serper.dev

[half venture]

No even before that entire drama happened openai was stalling

Clearly it was not ready to be deployed but openai rushed it

To compete with Google

[agile jay]

Yep, but they said it would be rolled out in the coming weeks

[half venture]

months

[sleek vortex]

idk why they rushed it anyway

theyre acting like google isnt going to take months either

yeah this is insanely high

[agile jay]

They changed it to months after the black widow stuff

[tame current]

what is

[agile jay]

Whats the tokens per second for haiku?

[sleek vortex]

.

[half venture]

Whatever Microsoft will deploy it for free anyway on June 18

[sleek vortex]

7 sources 60k tokens

[half venture]

Regret paying a subscription to openai

[sleek vortex]

likely a ton of useless context being sent

[tame current]

really

?

[half venture]

Yeah they will

[tame current]

but well I don't have high hopes

because the current version

[sleek vortex]

copilot already failed once

[tame current]

is nowhere near gpt 4 level

[agile jay]

Yep, a lot of useless things.

Why not convert the site to markdown?

[tame current]

even tho it's supposed to be the same version

[sleek vortex]

wdym?

[tame current]

html to markdown I assume

[agile jay]

html => markdown => query

[half venture]

Eh you would be surprised

[sleek vortex]

yeah but html to markdown doesnt really work does it

[agile jay]

It would reduce the tokens a lot

[half venture]

Copilot actually succeeded in reviving a dead horse

[sleek vortex]

most sites arent actually "html friendly"

and everything uses <span> and <div>

lemme try it on apple.com

[half venture]

When is wwdc

[sleek vortex]

130kb on this random (bad) converter

[agile jay]

Yep, so a lot smaller.

[tame current]

what about readability (the firefox reader thing)

[sleek vortex]

June 10-14, 2024

[half venture]

Yeah expect openai to roll voice out then

[sleek vortex]

ooh good idea

[agile jay]

It is opensource

[tame current]

there is a python port too I think

oh nvm

[sleek vortex]

i already with short lines of code turned that page into 40k

'[document]', 'noscript', 'header', 'html', 'meta', 'head', 'input', 'script', 'footer', 'style', 'a'

just ignoring these

after that it's all text but

[agile jay]

Yep, saves a lot.

[sleek vortex]

a large amount is in their footer and navbar

[agile jay]

And that is even without cleaning it up even more.

[sleek vortex]

but i can sure filter it for just apple

but what about every site ever

idk if you were to make a search engine this would be a slight problem lol

[agile jay]

The way I would do it is by asking for each new site which element of the page holds the article.

[warm flame]

Perplexity is losing its mind in the recent update. Basic grammar questions are coming back in bizarre form.

https://www.perplexity.ai/search/Is-in-a-1yGOOlZwS0q6Hd3oSIS3HQ

[agile jay]

That way I just only perform the html => md for that element.

[sleek vortex]

LMAO

[agile jay]

Which model were you using?

[sleek vortex]

default

[agile jay]

Looks like it got faster.

[sleek vortex]

no 60k tokens from apple website

so haiku bit only took 1.6s

[agile jay]

Is the previous part (14s) just scraping the links?

[sleek vortex]

Yeah, it seems

it sends your query to a small model to turn it into 3 searches

makes 3 searches

[agile jay]

Can you show how long it took for each link?

[sleek vortex]

good point

1min

[agile jay]

Because it might be one page that makes the rest be delayed. So might be good to add a timeout.

[tame current]

maybe see if you can use multithreading, tho I'm still not sure what the bottleneck is

[sleek vortex]

its already doing all pages at once

[tame current]

oh

[sleek vortex]

lemme add some timing to every step

[agile jay]

Yep, but I am assuming you are not making the query to the model until all the pages are done?

[sleek vortex]

yeah

[agile jay]

The biggest bottleneck is likely the scraping.

Some sites can take seconds to respond.

[agile jay]

What's the [6.77s] Finished reading 10 sources?

[tame current]

probably total of all links

[sleek vortex]

yeah

[agile jay]

The links above show less than 2s, and you said they are async requests.

[sleek vortex]

yeah hmm

maybe my parallelism is a bit flawed

let me check

[tame current]

how are you parsing them ?

if it's beautifulsoup that can be pretty slow

[sleek vortex]

yeah it is lmao

idk any faster libs

[agile jay]

Oh, you using BS4 for parsing the html?

which parser did you set?

[tame current]

you just need to change the parser to lxml

[agile jay]

html.parser or lxml?

[tame current]

lxml is way faster

[agile jay]

Yep, lxml is supposed to be faster.

But even html.parser shouldn't be that slow to make a huge difference...

Then again, it is python, so who knows...

[sleek vortex]

ok so why on earth do all links individually take 1s, but total is 6

[agile jay]

That's what I said...

[sleek vortex]

yeah

hmm

[tame current]

is this lxml ?

[austere kestrel]

yeah that one apple link is the problem - the rest are like 2-5k tokens

[agile jay]

How are you timing it?

[sleek vortex]

start = time.time()

_res = Parallel(n_jobs=16)(delayed(process_link)(link) for link in links)
_res = [item for sublist in _res for item in sublist]
print(f"🚀 [{fmt_time(start)}] Finished reading {len(_res)} sources.")```

[agile jay]

Can you add the time the request was started?

[austere kestrel]

also i don't think html is being passed - if you copy paste the visible webpage text, the number of tokens seem to align (using the old tokenizer)

[agile jay]

The request could take that long, but start at different times.

[sleek vortex]

yeah
i was thinking maybe rag chunking but that would take ages if pushing for speed

[austere kestrel]

you wanna try to just extract the main article body

[agile jay]

And I imagine the total at the end would be the correct time.

[austere kestrel]

none of the headers and all that

[sleek vortex]

isnt that a comparison page though

[agile jay]

Yep, that's what I said.

[sleek vortex]

i've tried to do this to some extent

[austere kestrel]

it's easier with some pages than others right

[sleek vortex]

def is_visible(element):
    if element.parent.name in [
            '[document]',
            'noscript',
            'header',
            'html',
            'meta',
            'head',
            'input',
            'script',
            'footer',
            'style',
            'a'
    ]:
        return False
    if isinstance(element, Comment):
        return False
    if element.parent.has_attr('style') and 'display:none' in element.parent['style']:
        return False
    return True


def scrape(url):
    response = httpx.get(url)
    html_content = response.content

    for tag in ["article", "main"]:
        element = soup.find(tag)
        if element:
            return element.get_text(strip=True, separator=" ")

    for class_name in ["content", "main-content", "post-content"]:
        element = soup.find(class_=class_name)
        if element:
            return element.get_text(strip=True, separator=" ")

    soup = BeautifulSoup(html_content, 'lxml')
    visible_texts = filter(is_visible, soup.findAll(string=True))
    output = ' '.join(t.strip() for t in visible_texts)

    return output.strip()```

ok this article and main thing was uncommented...

[austere kestrel]

like most news articles, wiki etc - no worries. but then there's everything else ha

[sleek vortex]

let me combine the article/main selector with the filtering

best approach

[austere kestrel]

yeah that sounds worth a go for sure

[sleek vortex]

def scrape(url):
    response = httpx.get(url)
    html_content = response.content

    soup = BeautifulSoup(html_content, 'lxml')
    base_elem = soup
    for tag in ["article", "main", ".content", ".main-content", ".post-content"]:
        if base_elem.select(tag):
            base_elem = base_elem.select(tag)[0]
            break

    visible_texts = filter(is_visible, base_elem.findAll(text=True))
    output = ' '.join(t.strip() for t in visible_texts)

    return output.strip()```

something like this

lets test

[sleek vortex]

hmm

[agile jay]

My idea is when it's a new site, to make an llm request asking which is the parent element of the article, and then cache that to scrape the same url in the future.

[twin cedar]

🚀 IT Managers, Network Admins, and Cloud Migration Specialists, we need your input! 🚀

Are you planning a major data migration soon? Share your experiences and future plans in our LinkTek Migration Survey – 2024. Your insights are crucial in shaping advanced data migration strategies.

🔗 Take our quick survey now: https://linktek.com/surveys/lfa-tribe-survey-round-6/

Thank you for helping us innovate and improve data migration processes!

[sleek vortex]

not a bad idea

[austere kestrel]

but have you looked at the page? I dunno why it's 30k tokens.. ig there's a bunch of collapsed stuff, or all the diffferent products that can be selected from dropdowns https://www.apple.com/ipad/compare/?modelList=ipad-pro-13-m4,ipad-pro-11-m4,ipad-air-11-m2

I would consider capping sources that are that big - it's such an outlier compared to most webpages/articles

[agile jay]

and if you wanna be even more accurate, cache a different parent element for each different endpoint.

/article/:id
/discover
etc, might have different layouts

[sleek vortex]

this is oddly broken and repetitive

are the columns being read left to right or something

oh

ok apple is just stupid too

cant afford grid

and rather every section is its own like element

... come on

[austere kestrel]

there's still another ~10k tokens somewhere ha
or... it doesn't like look like all the tabs / spaces were preserved copy-pasting.. perhaps that's them
edit:hmm actually i can see spacing

[sleek vortex]

maybe add a short regex to remove useless spaces?

[austere kestrel]

yeah i dunno.. it's weird. Like go to here https://forums.macrumors.com/threads/m4-ipad-pro-11-first-impressions-performance-heat-pwm-and-others.2426567/ and copy-paste into a tokenizer - it's basically fine. like even getting rid of the menu/headers etc would only be shaving a few tokens off

I think it's that particular website.. like it's normal for 30k tokens.. maybe cap at like 10k or something

[sleek vortex]

this page is because well

• it is a ton of items
• the way apple did it, not being a grid, so its almost useless anyway since each feature for each ipad is sort of listed before the next one, instead of each ipad's features then the next ipad

Number of tokens: 28021
Number of characters: 100960

before the regex was 32k

[agile jay]

What about if the input is too large, just pass it to haiku and tell it to write out the article from the page into markdown...

[sleek vortex]

Yeah but that defeats the point since we could just feed the whole input to haiku in the first place

Time taken: 1.56s
Time taken: 1.54s
Time taken: 1.55s
Time taken: 1.58s
Time taken: 1.56s
Time taken: 1.58s
Time taken: 1.61s
Time taken: 1.60s
Time taken: 1.67s
Time taken: 1.66s
Total time taken: 6.52s```

tested the scraping only in a different file

scraping the same site 10 times weirdly adds up to 6.52s

[agile jay]

No caching, and looks like it's not properly async

[austere kestrel]

i'm not sure how poe's system works with URLs.. it like creates a file of the page - i'm pretty sure it's just the raw content. usually the 'files' are in kb, like the macrumors one on the left, but this apple one is 1mb

[agile jay]

The request looks like it takes an average time though

You are using BS4, so maybe asking for the classes of all the bloat would be a good option. Then you can use element.decompose() to remove them from the dom tree.

[austere kestrel]

yeah that'd be pretty sweet tbh

[agile jay]

Yep, probably the best long term option

Wouldn't take long until you have the selectors for most sites.

And then instead of using an llm, you could use embeddings to match them.

[agile jay]

Did you figure out the timing problem?

[peak yoke]

Lol, let me try

[agile jay]

Just add print(datetime.now()) to the first line of the scraping function

[sleek vortex]

@agile jay fixed timing

[agile jay]

Nice, so now it's a lot faster?

[sleek vortex]

yeah

web scraping taking 1.52s here

instead of 7

[agile jay]

What was the issue?

[sleek vortex]

using joblib just wasnt proper parellism it seems

[peak yoke]

@hallow linden

[sleek vortex]

had to convert to multiprocessing

[agile jay]

nice, at least it got a lot faster.

That's why logging is important, lol

So you can spot problems like that.

[sleek vortex]

lets ask that one now

is the ipad pro m4 worth it? whats new and how much is it? should i upgrade

wtf

2x faster than pplx default

[agile jay]

Nice, that's what happens with good parallelism.

Can you increase the number of sources?

Just to see if it stays pretty fast.

[scarlet roost]

Hello, sorry, I have a question and I don't know where to post it exactly, is there any type of discount on the subscription for students?

[fossil sentinel]

https://youtu.be/NPOHf20slZg

[sleek vortex]

they dont

[sleek vortex]

seen this :d

[fossil sentinel]

Saw it pop up on my home page 😂

[sleek vortex]

you can get like $10 refferal from anyone in this server first month

about it

[scarlet roost]

how can I get it?

[agile jay]

i made a simple python function to try and locate the main content of a page.

from bs4 import BeautifulSoup

def extract_main_content(html):
    soup = BeautifulSoup(html, 'lxml')

    def get_text_length(element):
        if element.name is None:
            return len(element.string.strip()) if element.string else 0
        return sum(get_text_length(child) for child in element.children)

    def find_largest_text_container(element):
        max_length = 0
        main_content_parent = element
        for child in element.find_all(recursive=False):
            child_length = get_text_length(child)
            if child_length > max_length:
                max_length = child_length
                main_content_parent = child
        return main_content_parent

    body = soup.body
    if body is None:
        return None

    main_content_parent = body
    while True:
        new_main_content_parent = find_largest_text_container(main_content_parent)
        if new_main_content_parent == main_content_parent:
            break
        main_content_parent = new_main_content_parent

    if main_content_parent.parent and get_text_length(main_content_parent) < get_text_length(main_content_parent.parent):
        main_content_parent = main_content_parent.parent

    return main_content_parent

You can test it out and see how well it works.

Basically using the amount of text to try and locate the parent of the main content.

@sleek vortex Should help keeping the context low

[pine dagger]

@sleek vortex Can you DM me you min_agent.py code?

[sleek vortex]

ok im back

@agile jay increased sources to max of 15

9 organic, 3 story, 3 sublinks then remove dupes and non-html links

i limited long sources to 12k chars

gimme a query to try

[agile jay]

And looks like you're not showing the token count now.

[sleek vortex]

the token count did require me to run the tokeniser

i think that was also like the smallest bit slow

i could add a token estimate (divide by like 4.7)

[agile jay]

Is there not an open-source tokenizer?

[sleek vortex]

it is the openai tiktoken

ok it isnt that slow

but on super long sources it can take like 0.25s

lemme add it back 1min

[agile jay]

What about running it after the request is complete?

So doesn't add latency

[sleek vortex]

hm ok

not bad at all

[agile jay]

Nice

[agile jay]

Did you try the content extractor?

[sleek vortex]

wow, 7 second response for price of btc

[sleek vortex]

let me try

anthropic.RateLimitError: Error code: 429 - {'type': 'error', 'error': {'type': 'rate_limit_error', 'message': 'Number of request tokens has exceeded your daily rate limit (https://docs.anthropic.com/en/api/rate-limits); see the response headers for current usage. Please reduce the prompt length or the maximum tokens requested, or try again later. You may also contact sales at https://www.anthropic.com/contact-sales to discuss your options for a rate limit increase.'}}

bruh

[agile jay]

Lol

What's your rate limit?

[sleek vortex]

max TPD 300k

that one request with the 60k source is worth like 3 requests 😭

[agile jay]

Contact them and ask for a higher cap...

[sleek vortex]

well im on a free tier

i could put some money into it

but i also dont feel like doing that

[agile jay]

Oh, guess time to make a few accounts...

And rotate the API keys...

What's the price for haiku?

[sleek vortex]

uh

$0.25/mtok input
$1.25/mtok output

$5 free credit i used 8p today with 300k tokens

lets test this with g4f

[austere kestrel]

Ask "what happened to the Iranian President's helicopter?"

[sleek vortex]

bruh why does phidata's openai client not work

[austere kestrel]

Or just "who is the Iranian President?" (harder)

[sleek vortex]

hang on

taping on claude model rq

[tame current]

doesn't it require phone number?

and voip doesn't work

[sleek vortex]

now using cohere command r plus

as i ran out of other models

ok maybe i should try and filter the amount of sources after seeing that some fail

[warm flame]

Sonar Large 32K

[agile jay]

Or check for cached versions.

[sleek vortex]

your method like half works

i combined it with the if it just finds a .main or content tag

[tame current]

have you tried other models?

[quasi bridge]

No bug channel?

[austere kestrel]

Responses are pretty decent (though they get the date wrong)

sonar-large-online

[agile jay]

21K tokens still feels like a lot.

[warm flame]

Yes, GPT-4o worked perfectly.

[austere kestrel]

btw it's got built-in web rag functionality (it's pretty limited)

[sleek vortex]

from 14 sources

idk

[sleek vortex]

yeah it does

but im just using it because its api is free

haiku was way better at forming the response stho

[sleek vortex]

this model seems to always want to shrink it

[austere kestrel]

ha ofc - why not!

[sleek vortex]

bruh with 24 sources, 40k tokens, i get this

should i use low or high temperature

this is the context provided to the thing

[austere kestrel]

they're also optimised for RAG, if I remember correctly

the cohere models

[austere kestrel]

go with fewer sources i reckon - if could somehow make the selection of sources more targetted

or snippets

[sleek vortex]

thatd probably be using like selection of chunks using embedding

let me try lower the amount of sources ig

[austere kestrel]

yeah otherwise i don't really see how you get around it - either fewer sources, or more sources but with snippets..

fewer sources would be the simpler option anyway - if only temporarily

[agile jay]

Which indexer are you using, btw?

[sleek vortex]

indexer?

im using serper.dev which is proxy of google

@austere kestrel

after setting temp to 0.0

instead of 0.7

[austere kestrel]

nice. that's a solid response

[sleek vortex]

lemme try the ipad one from before

[agile jay]

The number of sources increases the output time by a lot.

[sleek vortex]

ok this is pretty mid for 31 sources

i think back to 10 sources

sources = sources[:10]

[austere kestrel]

yeah 30s is long time

[agile jay]

Yep, unless you are making different searches and taking the best sources from each one.

[sleek vortex]

there's no "best" sampling right now

only real way to do that is embeddings

response time will become like 2 minutes though since i dont have a gpu

[austere kestrel]

yeah or some agentic approach, where another (or the same ig) model evaluates and selects the (max 10) sources, like tries to do it smartly

but would obviously add another layer of complexity and slow things down

[sleek vortex]

wait wth

this is the best response yet

only took 25s on the model since it isnt fast (not haiku)

[agile jay]

I think groq llama 3 8B would probably be good for preprocessing the sources.

Do you have a good system message to control the output?

[sleek vortex]

its not the worst

<context>
    [context]
</context>

When answering to user:
- If you don't know, just say that you don't know.
- If you don't know when you are not sure, ask for clarification.
- Make sure to provide any details, going beyond the user's query if you have the information to do so.
Avoid mentioning that you obtained the information from the context.
And answer according to the language of the user's question.

Given the context information, answer the query. Answer in as much detail as possible, but do not make up information.```

[austere kestrel]

yeah groq, 8B for the preprocessing/curating of sources, 70b for actual response maybe

[sleek vortex]

def format_source(source: Source):
    base = "<context>"
    if source.link:
        base += f"<link>{source.link}</link>"

    if source.title:
        base += f"<title>{source.title}</title>"

    base += source.content
    base += "</context>"

    return base```

[sleek vortex]

i think this might be what pplx does

but they run the 8b bit beforehand

im guessing

[austere kestrel]

fwiw i would consider experimenting moving the [context] to the bottom, and having the instructions at the top. just to see if it makes a difference

but that structure would make sense for antrhopic/haiku

[sleek vortex]

the context is added at bottom

[austere kestrel]

ah right sweet

[sleek vortex]

soryr not clear in that

1sec

agent.print_response(message=f"""
User's Query: {query}

{as_str}
""", show_message=False)```

as_str is the sources joined

[sleek vortex]

that goes into system prompt

[austere kestrel]

gotcha 👍

[sleek vortex]

was typing date

but uh

[austere kestrel]

lol

[sleek vortex]

interesting how i set the model to gpt4 and it hallucinated the price for the ipad

[pure onyx]

by any chance, are students provided with any subsidy on the license / pro ?

[sleek vortex]

nope

[pure onyx]

oops

[austere kestrel]

maybe try going with 0.1-0.2 or something like that with the temp. 0 seems kinda extreme (but that might just be me ha)

would be a bit paradoxical if it reduced hallucination ha
but it kinda makes sense to my mind that it could

[sleek vortex]

ok this is claude 3 opus on gpt4free (you.com?)

so this isnt very 100% accurate since there are [[1]] quotation things so there's either you.com's system prompt or their own web search colliding wiht mine

the response is great but idk if its from my efforts or no

(i cant afford opus on my own account)

[austere kestrel]

problem with llama3 via groq actually would be the context window

[sleek vortex]

yeah

[austere kestrel]

8k tokens or whatever it is

[sleek vortex]

lemme try phi3 128k local

im dead 😭

[agile jay]

...

[tame current]

What if I am gay smh

[rancid dome]

How Perplexity’s writing and researching compare to Gemini Pro? Is it more or less accurate? Thanks?

[halcyon coral]

I would rate Perplexity as better in both areas. Perplexity is an AI search engine ("answer engine") and therefore perfect for research. For writing tasks, you can use Opus in the Writing focus.

[sleek vortex]

gemini pro only good thing is context

rest is like decent

[plush cairn]

How about comparing perpelxity with chat-gpt in processing data? like csv.
is there a possibility for the future perpelxity can upload csv or xlx format?

[halcyon coral]

You can already upload csv files

[plush cairn]

daamn I just realized can drop csv file
thxxxxx

[fossil sentinel]

I think chatgpt is the better option for that

[stable radish]

For CSV tasks, I would use Julius AI
Quite interesting tool if you don't know about it

It's like a Perplexity, but for data analysis so it runs with 2 layers:

• Python Running in BG
• LLM that formulate response (Claude 3, GPT-4, GPT-4o)

[hallow linden]

We don’t know how good we got it

[sleek vortex]

hmm

[tame current]

did the upgrade happen

Claude 3 opus is so useless now, it used to be able to answer any chemistry question I gave it just from a screenshot, now it just keeps whining about copyright no matter what I show it 🙄

[grand silo]

I’ve found Sonnet good enough in most cases, but it does depend on the case

[orchid shore]

How to succeed in [x] market:

Do nothing
Competition repeatedly shoots itself in the foot

[rancid dome]

What else can replace Claude 3 Opus?

[tame current]

idk, i was a big fan of it but now with the limited to 50 responses and having copyright restrictions on basically everything so I can’t even show it pictures and ask it to answer, it’s just so disappointing when I know what it’s capable of 😦

[arctic spindle]

maybe gpt-4o

[minor shard]

Didn't know he was Muslim lol

The stuff I am seeing on Google AI Overview is what people fear.. spreading misinformation

love this one

[split glen]

Can we have a desktop (Macos) app

[warm cave]

Bro, that one is too good 😆

Are they using Gemma 2b? 🤨

If so then that is their fault

[minor shard]

Lol no idea

i didnt know you could choose a model for AI Overview

[warm cave]

sorry, i meant that google was asking for problems if they used gemma 2b lol

Just gave gemma some sources and put the reddit thread in the middle, lets see what we get

[minor shard]

haha

[warm cave]

Just ran it, lol

[minor shard]

wtf lmfao

[dense goblet]

Does anyone tell me how to generate images in the iOS App?

[warm cave]

Don't think its possible

you are best off using the web site on your phone, or using ms copilot for images

[tranquil sparrow]

Hey, does anyone know when will the api provide images and videos references? Last time I check it was in beta

[warm cave]

or whatsapp if you like lightning image gen, its fun

@minor shard Phi 3 passed the test```The issue of cheese sliding off homemade pizzas can be addressed with a
few adjustments in technique and ingredients to improve adhesion and
moisture retention. Here are some strategies:

1. Sauce Thickness: Marinara sauce tends to be watery, which may cause
   the cheese to slide off. Opt for thicker pizza sauces designed
   specifically for this purpose. The increased viscosity helps keep the
   cheese in place.

2. Cheese Texture and Type: Shredded 2% mozzarella has a finer
   texture, which can contribute to sliding off when baked. Using whole-milk
   mozzarella or longer shreds may adhere better due to its coarser texture
   allowing it to grip the sauce and other ingredients more effectively.

3. Cooking Techniques: Allow your pizza to cool slightly after removing it
   from the oven before serving. This not only prevents cheese sliding off
   but also makes it easier to cut into slices without pulling out cheese
   along with them.

4. Moisture Control: Too much moisture can lead to sogginess and
   subsequent cheese loss. Pre-bake the pizza dough for a couple of minutes
   before adding sauce and cheese, which will help create a barrier that
   prevents moisture from seeping in and affecting the crispness needed to
   keep the cheese attached.

5. Dough Hydration: If your homemade pizza dough recipe calls for
   water or other liquids, consider reducing their proportion slightly to
   enhance crust firmness without compromising taste. A sturdier base will
   support more evenly distributed sauce and cheese.

6. Baking Techniques: Bake your pizza on a preheated pizza stone or
   baking steel for an even, hot surface that can aid in crisping the crust
   without over-browning or drying out the toppings too quickly.

[lyric pulsar]

who the peplexity owner thingy idk

i gotta ask but please

overrigg this thing for school bro

all in one thing like yea

[warm cave]

Perplexity answers 🪄 https://www.perplexity.ai/search/who-the-peplexity-fa1eI6H5ReOUkMX9WxMssg#0

[minor shard]

Phi3 The best ❤️ I still like Llama 3 more

[warm cave]

Ollama is the best! But we don’t have a Llama 3 3.8b model, maybe in the future

Im sure it would be good if meta did so

[tame current]

finally got access to gpt mac desktop app woohoo!!!

[dreamy carbon]

Now i see Page besides Threads and Collections. Can somebody tell me what's new with Page?

[vivid kraken]

my perplexity subscription is ending soon in 2 weeks, should i keep paying? I already have chatgpt plus and gemini advanced, do you think the openai search engine will be announced within next month>

[warm cave]

amazing!

[tame current]

got a message saying that the new voice will be coming in the next few weeks. hope thats right, although im sure itll be a month or two away probably

[solemn fox]

How do we get the new discovery tab on mobile to not be so freaking terrible

Like how it was before

[wise edge]

pages ~~~

[wise edge]

https://www.perplexity.ai/page/RDA-and-UL-jd1aZH.qQkeV6iEAMJC7JQ
amazing of pages

[pulsar sail]

Hi! What are exactly pages?

[shut plaza]

GPT-4o teaching me how to mate with a King and Rook

Idk, but I don’t think there’s a 129th rank on a chessboard

[fading moth]

Opus have same problem?

[fading moth]

But does Opus suggest glue for pizza, Daily?

[warm cave]

No, haha 😂 opus is too smart for that

[fading moth]

Just making sure. You never know.

[warm cave]

If opus tells me to eat glue I will do soon, I trust opus, opus knows best 😇

I will Mary Opus4o on release 😆, she will be my ai bride 👰

In the mean time GPT4o will have to do for a GF 😂

[sweet jasper]

Yall should get Chinese waifu yi-large

2nd place for coding on lmsys leaderboard

[harsh stag]

where can i try Chat Xi PT?

[tame current]

@harsh stag wot

you mean chat gpt?

[harsh stag]

https://www.perplexity.ai/page/Chinas-Chat-Xi-QyjXwHVZTMS_4wB8PsZb_g 😂

[ionic stone]

Cant tell if serious 🧐

[plush cairn]

woow, there's a new update for 'Pro Search' 🔥

[harsh stag]

what's new?

[plush cairn]

for my case in finding research articles. the sources provided are more reliable

[sweet jasper]

It's number 2 on chat.lmsys

For coding

So srs

China xi Qi ti

[ionic stone]

I see now

Not exactly Waifu but I get it

[ocean cosmos]

convert to page feature is awesome! thanks team perplexity 🙏

[sinful kettle]

how can i delete files uploaded to perplexity

[wraith crow]

you can delete the files from the sources if you don’t want answers based on the image/file. files uploaded to perplexity will be deleted 30 days after upload

[granite turret]

Hi ! What is exactly the aim of the "convert to page" functionnality ?
I tried on a discussion where i asked perp to generate a document, the "convert to page" result is totally different, then it tells me that it is now shared...
Any hint on that ? thx 😉

[junior adder]

guys if my pro expires, can i use someone's referral to roll in with a 10$ subscription?

[scenic bison]

我能提问吗

[tight prism]

Yes you can

[fervent needle]

New pro search is very cool 👍

[jagged ivy]

When was it updated?

[fervent needle]

I noticed it an hour ago, don’t know when precisely it was released

[sleek vortex]

what’s new about it though ?

[fervent needle]

I asked it to fix some old code with legacy APIs that has been deprecated, so it went online and searched latest API specs and it showed this multistep reasoning, which is pretty cool

[sleek vortex]

Oh what

i don’t have that

AB testing?

[sleek vortex]

also modal yoo

[harsh stag]

can you share the thread?

[fervent needle]

yeah maybe should be releasing soon

[fervent needle]

https://www.perplexity.ai/search/Below-is-my-w.MoUuDzSECwkuUIASuDNw#1

[sleek vortex]

Yeah idk

I mean there’s a good chance it’s just a frontend change?

[harsh stag]

I think new UI only

[fervent needle]

maybe just an ui update

yeah exactly

[sleek vortex]

yeah

[fervent needle]

but pretty cool anyway

[sleek vortex]

better quality is just placebo

i think

[sleek vortex]

yeah

[fervent needle]

I can't tell of the quality is better, but this new conclusion is pretty cool

[sleek vortex]

yeah

[harsh stag]

nope it seems like there's something more

[sleek vortex]

hmm?

[harsh stag]

on the left is @fervent needle's thread, the right is mine, same query.

[sleek vortex]

so maybe they’ve added a new mini model

or changed their pre model

to be more “like” an agent task

[harsh stag]

huge, so Pro search will now actually do something besides pulling more sources than normal search

[fervent needle]

if that is the case then its great improvement, doing multistep searches based on previous step searches definitively improve the result

[fervent needle]

exactly what i was hoping for

but i can't tell huge difference yet

[sleek vortex]

this is such a coincidence considering i was working on this yesterday lol

[harsh stag]

how about the clarification questions?

is it better?

[fervent needle]

wow just noticed the new discover articles is a page now

[austere kestrel]

I saw it do this yesterday! but I hadn't heard anything about it so concluded that it was just me imagining things ha

i was using a VPN set to the US, to use the new chart thing

I haven't seen it again since not using the VPN - wonder if it's a US thing? or just a/b

[austere kestrel]

nice. thanks.
this looks promising

first major change to the search system in months it feels like

[sleek vortex]

such a coincidence based on my work from yesterday lol

but yeah excited to see pro search be more than just a larger amount of sources

[austere kestrel]

i've been reviewing the sources lately - it's striking how often out 20 sources, maybe 10 will actually be relevant.. the others won't just be old, but sometimes utterly irrelevant... I think 5 highly relevant sources alone will usually be better than 5 good sources mixed within 15 dubious/lower quality sources – it's just noise (and tokens..)

[cinder comet]

@sleek vortex ive found a way to use the send func on wss

[sleek vortex]

i have this thread from yesterday

one of the sources was literallt à websites javascript file

like 0 related sources

[austere kestrel]

ha yeah sometimes it's hard to work out how something ended up in there as a source

and i search basically exclusively in English

[half venture]

Is opus still 50

[austere kestrel]

can only assume it's more erratic for other languages, esp non-latin based

[fervent needle]

my ip is not us, i have the new search not the new chart which is us only rn

[austere kestrel]

cool thanks - good to know

[sleek vortex]

this really isn’t going to change any time soon i think

they’d go broke

[fervent needle]

I believe its pernement and abuse was an excuse lol

[sleek vortex]

abuse is like 15% of the thing

[fervent needle]

but anyway opus is no longer top 1 other than specific uses so less of a deal

[sleek vortex]

85% is to make sure it’s still profitable

[sleek vortex]

yeah

[cinder comet]

Yes

[half venture]

Man gpt 4o doesn't hit the same as claude opus

[stable radish]

If we’re being realistic here, Claude 3 Opus never going to comeback with 600 uses

[austere kestrel]

when Opus appears here for throughput pricing I suspect it prob will
https://aws.amazon.com/bedrock/pricing/

i actually think it's as much about hardware as costs (though two sides of the same coin ig).. like I was listenting to an interview with one of the Anthropic cofounders the other day, he was saying that there just isn't the ability to serve "their leading model" (Opus) to both their enterprise and consumer customers at full scale

[austere kestrel]

Claude 4 might be introduced with 600 uses though ha

[harsh stag]

80 ~ 100 will do just fine

[sleek vortex]

im like 50% sure they use azure openai

but idk

ok they probably use both im dumb

anyway

https://www.reddit.com/r/AZURE/comments/1bzs8gr/have_you_purchased_openai_ptus_how_much_did_it/

[sleek vortex]

yeah wouldnt be surprised if claude 4 adopts a smaller model

like gpt4o

[stable radish]

Why aha?

Or you were making a joke and it went over my head?

[stable radish]

It would be fine for many, but it still not what you and me are paying for

[sleek vortex]

@stable radish

[stable radish]

Yeah anyways that’s just beating around the bush

It’s like:

“We reduce the Claude 3 usage because it’s too expensive, although we made a “promise” through our value proposition that you would get 600 uses per day

But hey, if Claude 4 comes out we might go back to the 600 searches a day because it’s less expensive for us

Of course, you will not get the same performances, but that’s a workaround 😊”

[flat raft]

I'm trying the new Pages and it's cool but they are auto-generated in english even if the default language in my profile is different. How do I make it generate Pages in languages other than english?

[austere kestrel]

ha somewhere in between… like it’s coming.. and we don’t talk about Claude 2 anymore.. so I figure soon enough Claude 3 / Opus will become an afterthought too..
but yeah no idea when (or whether it will be with 600 uses on Perplexity aha)

but i do think it's gonna have to be smaller / more effecient / cheaper than Opus

otherwise they're toast given gpt4o's pricing

[umbral pollen]

I agree, for most use cases I’m dealing with now, I’ve closed the book on Claude.

[austere kestrel]

yeah it's sooo much more expensive

and then with gem pro 1.5 in the mix too, especially the Flash variant

[umbral pollen]

Not saying it’s useless, but I’m just really impressed with what Gemini and of course OpenAI are doing

[austere kestrel]

that kinda negatates whatever had going for it imo

[austere kestrel]

I think a lot of people very much prefer Claude 3 in terms of style

and just nicer to intereact with

[umbral pollen]

What’s the deal with flash? I haven’t experimented with it much because pro is still fast enough for me

[austere kestrel]

but at the end of the day, it's so much more expensive that outside of personal usage, anything at scale would just seem nearly impossible to justify

[austere kestrel]

it's not as fast as haiku, but it's a stronger model imo (and has 1m token context window; and a few modalities )

[umbral pollen]

I can understand that. Thing is I feel like I customize my experience so much on OpenAI that I make it my style

[austere kestrel]

same tbh

i find Opus 'nicer' to interact with

[umbral pollen]

But pro vs flash; is there a benefit to flash other than speed? And does that speed come at the cost of anything else?

[austere kestrel]

but if i'm trying to get some stuff done or do some kinda automated workflow, that pleasantness is basically irrelevant

[austere kestrel]

• Gemini Pro 1.5
• Gemini Pro 1.5 Flash

Same underlying model, but the flash variant is faster and cheaper (and yeah less performant ofc as the tradeoff)

[sleek vortex]

might be cooking something?

[austere kestrel]

they're due introduce function calling at some point

[sleek vortex]

they have a fucntion calling guide

[austere kestrel]

yeah

[sleek vortex]

idk if i should use FC or not

[austere kestrel]

i don't think it's available yet

but might be mistaken

[glass crystal]

Interesting! Is this just historical data or will it be updated from sources regularly?

[sleek vortex]

https://docs.anthropic.com/en/docs/tool-use

[sleek vortex]

I'm trying to replicate ChatGPT's data analysis tool simply with prompting and a serverless jupyter container/backend

[umbral pollen]

I’m embarrassed to admit this here but I haven’t really figured out how to use FC in my assistants

[sleek vortex]

FC?

oh

[glass crystal]

Super cool. Would love to follow and hear your result.

[austere kestrel]

oh the public beta has been available since April

not sure what i was thinking / talking about

[sleek vortex]

yeah but im not 100% sure if i should do it with function calling, or just text prompting

because i want it to work with local models too

and function calling in the backend is probably doing just this anyway (with some finetuning)

hmm

I remember before it was available in the API claude had a guide for FC with xml

You may call them like this:
<function_calls>
<invoke>
<tool_name>$TOOL_NAME</tool_name>
<parameters>
<$PARAMETER_NAME>$PARAMETER_VALUE</$PARAMETER_NAME>
...
</parameters>
</invoke>
</function_calls>

Here are the tools available:
<tools>
<tool_description>
<tool_name>get_current_stock_price</tool_name>
<description>Gets the current stock price for a company. Returns float: The current stock price. Raises ValueError: if the input symbol is invalid/unknown.</description>
<parameters>
<parameter>
<name>symbol</name>
<type>string</type>
<description>The stock symbol of the company to get the price for.</description>
</parameter>
</parameters>
</tool_description>

<tool_description>
<tool_name>get_ticker_symbol</tool_name>
<description>Gets the stock ticker symbol for a company searched by name. Returns str: The ticker symbol for the company stock. Raises TickerNotFound: if no matching ticker symbol is found.</description>
<parameters>
<parameter>
<name>company_name</name>
<type>string</type>
<description>The name of the company.</description>
</parameter>
</parameters>
</tool_description>
</tools>```

[austere kestrel]

would os models not fine tuned to do it be able to just through prompting?

[sleek vortex]

that's what im thinking

idk

not sure

[austere kestrel]

maybe those with json outputs mode

[sleek vortex]

they might need fine tuning

[austere kestrel]

would need json outputs at a minimum i think

[sleek vortex]

claude is like the only model i've experienced that plays well with xml more than json

[austere kestrel]

llama 3 might?

[umbral pollen]

I’m really proud of the dozens of GPTs and OpenAI assistants I’ve created, but I’m a writer (RS), finance guy, and psychologist, so I’m not sophisticated from a coding perspective

[sleek vortex]

lemme try this exact prompt on 3-70b

[cinder comet]

Have you tried yi-large

[umbral pollen]

What are the ultimate aim of your inventions, would you say? If you had to generalize

[sleek vortex]

me?

[umbral pollen]

I was just looking into ChatGPT’s data analyst tool a couple hours ago. I asked it for its instructions and got it to tell me every python package in its environment

[umbral pollen]

Yes

[sleek vortex]

I might try to build a tool that combines the best of all the other ai tools - code interpreter, agent-like web research and so on, not really sure as of yet

right now just doing experiments to see if I can make these work

[sleek vortex]

then when I have free time I might do this

Well really just for my own benefit, but I could turn it into a product too - would be something nice to work on honestly

not bad for an 8b either

[sleek vortex]

my main well restriction is cost and budget

i'm trying to experiment with backend infrasturctures such that I could make such a product on the cheapest possible starting out backend

[umbral pollen]

Well I have a lot of projects that people would really benefit from using. As a psychologist, I’m into deepening understanding and enhancing communication

So my inventions are mostly centered around that theme

But I’ve developed so many features that it could also be big business

I see one of the principal benefits of AI as its impartiality . I’m picturing like an employee email integration to detect problematic communication styles

Some people are just not good at articulating themselves in a respectful and productive manner

Eventually, everyone wants such an individual fired, but they have no objective judge that can say “your phraseology indicates disrespect and passive aggressive behavior.”

This presents a problem when it’s time to terminate, and ideally, my technology could show the communicator the errors of their actions before it becomes a serious problem

[sleek vortex]

are you faced with consdering ethics?

[umbral pollen]

My crown jewel is my Ethical Algorithm GPT

Ethics lend themselves well to algorithms. Problem is there are so many schools of thought when it comes to what is ethical

But that’s philosophy, if we’re talking about ethics in business or in an individual company, the definition of ethical behavior becomes increasingly more clear

I don’t believe in AI replacing human judgment. I think it should just be another voice in the conversation.

[sleek vortex]

Yeah, like how if employee behaviours at some big company like google is (probably) already tracked with an algorithm that considers performance

Using AI is just another level of that

[umbral pollen]

Did you have a specific ethical concern in mind?

[sleek vortex]

Not really, no

Just curious generally

[umbral pollen]

These would be for small businesses and educational institutions

[sleek vortex]

yeah, I was just using google as a simple-to-make comparison

[austere kestrel]

Ethics lend themselves well to algorithms.
I find that a bit disconcerting ha

[umbral pollen]

It was a good comparison too, because I imagine they already have tech doing that for them. But I know lots of small businesses that still rely on “people keep complaining to me about so and so’s attitude “

[umbral pollen]

Why? Algorithms are instructions, ethics are too

[austere kestrel]

mm not sure i agree... as you said before, ethics is basically a philosophical consideration / view.. they might guide and inform decisions / values, but I don't see ethics as instructions

[umbral pollen]

Rules/instructions

[austere kestrel]

like let's say the death penalty. Is that an ethical form of justice?
I personally do no think it is

but obviously it's not black/white

[umbral pollen]

Again, I’m projecting a bit from my world where “ethics” are strictly defined rules for which breaking results in losing one’s license

[austere kestrel]

right sorry, i see what you mean

[umbral pollen]

Ah, I see what you mean. That’s not the type of ethics I’m referring to

[austere kestrel]

yeah sorry! i've gotcha now

[umbral pollen]

No im sorry lol and grateful

You’re really helping me figure out how to articulate myself better by letting me say it wrong first haha

[austere kestrel]

I understand what you mean perfectly though 🙂 like any institutional / corporoate setting - whether academic, NGO, government, private sector etc.. they've for sure got ethical frameworks which very much are like rules

[umbral pollen]

And it could give people something to point to and say “look: AI has detected unethical conduct all of these times for all of these reasons.”

It’d save companies fortunes in bs wrongful termination suits

And hopefully (I’m a softy) could help aim everyone to communicating more effectively, respectfully, and harmoniously

Removing the need to terminate capable people who just don’t have awareness of how they come across in textual communications

One bad manager can lead to quiet exodus of talent that are just sick of working with someone like that

[austere kestrel]

this is admitedly starting to sound more like HR than ethics aha

but i hear what you're saying

[umbral pollen]

And I wouldn’t have it function just like a cop, those who demonstrate great communication skills would be noted as well, and hopefully rewarded

[umbral pollen]

lol, yeah it is more HR in practice, I can’t sell software to the ethical resource departments of small businesses lol

[austere kestrel]

ha yeah even companies are using AI for hiring/firing decsions.. they definitely aren't doing so openly ha

[umbral pollen]

I’ve been the primary communicator under my techs scrutiny and I’ve actually learned things about myself that I NEVER would have thought of without it

[stable radish]

By the way, Claude 3 Opus is the GOAT (just wanted to lay that down after some conversations with it yesterday ahaha)

[austere kestrel]

same. all this ai technology that has emerged in the past 2 years or so is insanely powerful / enabling

[stable radish]

You're talking about Gemini on API or browser session?

[austere kestrel]

ahaha yeah i know i know.. it's much beloved

[sleek vortex]

time to make it an actual ui

(rolled the shadcn template now to make it work :d)

[umbral pollen]

Both

[stable radish]

I don't know, I tried Gemini "Advanced" on web and I wasn't impressed honestly

[umbral pollen]

This is kind of funny but I created a GPT called argument analyzer, initially because I wanted to prove myself right. I programmed the thing, and it turned out I was 80% wrong! 😂😂

[stable radish]

Guys by the way quick question, what's your background with AI/coding?

[sleek vortex]

me?

or generally everyone

[umbral pollen]

60%* wrong but I thought I was definitely 100% right. This particular argument was between myself and a girl I was dating so obviously emotions blinded my judgement. But that happens to everyone. My university thinks it could be a useful tool for couples therapy

[umbral pollen]

I jumped in with both feet this year and became obsessed

ChatGPT taught/is teaching me how to code. And I’ve created just over 100 GPTs

[harsh stag]

opus and 4o wrote all the code for me so I'm the greatest prumpt engineer ever

[umbral pollen]

When did you try it? They updated two days ago

And sell me on Claude. I want to be wrong. What is it better at than ChatGPT 4o, perplexity, or Gemini?

I’ve been crushing on Perplexity lately because of the sources

No other ai has given me enough citations no matter how I engineer my prompts

[chilly nova]

they added the researching?

[chilly nova]

what is this? vercel thing that can make ui for you?

[umbral pollen]

Perplexity? The browser version yeah

[chilly nova]

what?

[umbral pollen]

What’s your background in coding?

[fervent needle]

may i ask which frontend is this?

[stable radish]

Well, you and everyone (still interesting to get to know some of the active people here)

[stable radish]

Oh that's interesting, so how would you rate yourself in Python and the subsequent lib?

[wide mesa]

AI being janky again

[stable radish]

Hmmm, on the 21th of May

[umbral pollen]

I’d rate myself as competent for tasks involving OpenAI

[stable radish]

It is "smarter"

If you've done a lot of writing with LLMs since the release of ChatGPT to the public, you get the grasp/natural feel for it

It feels less "autocomplete"

Which it still doing, but you get me

[umbral pollen]

It’s more eloquent

?

For what type of writing?

[stable radish]

Any type I'd say. I just prefer Claude tbh

[sleek vortex]

shadcn template "blocks" - they had one from ui

not using ai to make the ui though - i dont think its that good anyway and im competent enough

im just making this a rough tool so i didnt want to waste 30 mins making this frontend bit from scratch lol

[sleek vortex]

tailwindcss + shadcnui

[sleek vortex]

im mainly a frontend (and backend) js/ts sorta web dev, but im also good with python for backend

also worked with swift, swiftui lots, and other random things

[stable radish]

"tasks involving OpenAI", you mean you're proficient with API stuff and everything?

[umbral pollen]

by type i mean have you been using it to write code? Data analysis summaries?

[umbral pollen]

yes

I know its endpoints better than chatgpt, which still recommends you write code to deprecated endpoints

[wide mesa]

Opus just refuses to do write ups now?

[chilly nova]

what is that IDE? didn't they have teh website for that thing?

https://v0.dev/

[wide mesa]

what the hell is wrong with Claude and why is it giving that "x and X might be copyrighted" type responses?

[sleek vortex]

what ide?

[sleek vortex]

yeah this

its not that amazing though

(imho)

why is llama actually funny

[hoary phoenix]

Hi all. Is there a breakdown anywhere of the pros and cons of the different available AI Models for Perplexity? Not sure which to use.

[hoary phoenix]

Thanks! Not sure what to do with that information, but I appreciate the reply all the same 👍

[austere kestrel]

https://www.youtube.com/watch?v=0T1444HJbt0

[wide mesa]

Seriously, is Claude being janky for anyone else here?

I'm giving it a prompt and it refuses to properly answer it

[wide mesa]

alright it stopped being jank now

[warm cave]

This is good news! I gotta check and make sure its open source

[sleek vortex]

well its on HF

idk the license

[warm cave]

Great, just know that mistral has been doing some open and some closed lately

but now that i see its on HF im good

[sleek vortex]

yeah

ok got this much working

[warm cave]

or i guess i should say open weight, lol. for clarity

[sleek vortex]

now to hook up this actual parsing on the backend

[plush cairn]

In Gemini Advanced does it use the 1.5 flash model by default?

[real storm]

I'm trying the new page feature but suddenly I can't find the create page or see my page on the web

[sonic hull]

Same, it just disappeared

[lusty basin]

Same here

Was really enjoying it

[real storm]

Still visible on mobile but can't edit

[agile jay]

Is that the shadcn example playground?

Are you also using the vercel SDK for interfacing with the different models?

[sleek vortex]

yeah but its turning out more annoying than beneficial really

got the xml parsing working

the thing in grey is the model showing it's "thoughts"

[
  {
    "role": "user",
    "content": [
      "ovveride python global print to truncate the middle of arrays over 50 in length\n\ni mean like in pandas, when u do like .unique()\n\n"
    ],
    "id": "sdvaHL5jJfYqcwtU-g5px"
  },
  {
    "role": "assistant",
    "content": [
      "Okay, let's think through this step-by-step:\n",
      {
        "intent": "thinking",
        "content": "<thinking>\n1. The user wants to override the global print function in Python to truncate the middle of arrays that are over 50 in length, similar to how Pandas handles printing large arrays with the .unique() method.\n2. To do this, we'll need to create a custom print function that checks the length of the input and truncates it if it's over 50 elements.\n3. We can use the built-in len() function to check the length of the input, and then use slicing to extract the first and last 25 elements if the length is over 50.\n4. We'll also need to replace the default print function with our custom function, so that it applies globally.\n</thinking>"
      },
      "\n\nHere's the code to achieve this:\n\nimport sys\n\n# Define a custom print function that truncates long arrays\ndef custom_print(*args, **kwargs):\n    output = []\n    for arg in args:\n        if isinstance(arg, (list, tuple, set, frozenset)) and len(arg) > 50:\n            output.append(f\"[{', '.join(map(str, arg[:25]))},...,{', '.join(map(str, arg[-25:]))}]\")\n        else:\n            output.append(str(arg))\n    print(*output, **kwargs)\n\n# Replace the default print function with the custom one\nsys.stdout.write = custom_print\n\n# Now you can use the print function as normal, and it will truncate long arrays\nprint([i for i in range(100)])\n# Output: [0, 1, 2, 3, 4,...,75, 76, 77, 78, 79]\n\nThis should give you the desired behavior, where long arrays are automatically truncated when printed."
    ],
    "id": "tQlfyjgn-9P7W7Elb6abb"
  }
]```

[agile jay]

What's the annoying part?

[sleek vortex]

Well since im trying to roll my own code interpeter

their "streaming ui" thing had me a bit confused for a bit

[agile jay]

For which language?

[sleek vortex]

well python right now

but any in theory

like chatgpt's "data analysis"

i know like it probably exists on github but

wanted to see if i can do it anyway

[agile jay]

Oh, it's just a way to create react components by inputing the data the model has generated.

Oh, in my case I use wasm, and run it in the browser.

[sleek vortex]

I want the full environment though

planning to connect it to a serverless jupyter notebook

[agile jay]

But I am turning python into cython and then into wasm.

Oh, mine is designed to be client side.

And for the user to generate tools from code, that the AI can then call later.

Since with the new wasm spec, I can call any tool in any language that supports wasm.

So the user can use ffmpeg natively in python to edit the video, if they wanted to. And many other usecases.

[sleek vortex]

ffmpeg wasm was very slow for me last time i tried though

I could just forget streaming for this

do you know if chatgpt data analysis has streamed messages or not?

[agile jay]

Well based on how it shows up in the UI, it looks like it generated it, and a tool then runs that script most likely on a runtime somewhere, and then just returns the result.

The downside is that the libraries are very limited.

[sleek vortex]

mine or yours?

ohh

chatgpt

[agile jay]

ChatGPT

[sleek vortex]

What do you think of as missing?

[agile jay]

Most import one is probably not being able to use http requests.

[sleek vortex]

Hm

Well my "container" technically supports it

right now adding the upload support for files

[agile jay]

Also not being able to run it as a server...

[agile jay]

Were you using ffmpeg.js or ffmpeg.wasm?

[sleek vortex]

Wasm i think

[agile jay]

As long as you were using the wasm version and your on a new enough version of chromium, it should have been pretty fast.

Maybe you were using a cpu encoder?

[valid yoke]

ChatGPT just added citations !?

[sleek vortex]

did it not

always have that

?

[warm cave]

Its had that for a little whole now

[sleek vortex]

yeah