#💬│general | Perplexity | Page 57

sleek vortex May 28, 2024, 10:15 PM

#

before it got like none of them right

#

this is a good test query to test lots of parts of the system

#

ive changed my searxng instance to have way better and updated source rankings

#

now i need to see why its reporting the wrong info for 1.5 pro

#

🔗 [SearchAgent] [0.00s] Picked 6 links total
🚀 [SearchAgent] [1.98s] Finished pulling 3 sources - 8874 max chars/source
    * 🔗 https://ai.google.dev/pricing - 3699 chars
    * 🔗 https://www.cnet.com/tech/services-and-software/google-gemini-pricing-1-5-pro-and-1-5-flash-compared/ - 2526 chars
    * 🔗 https://beebom.com/how-use-gemini-1-5-flash/ - 7446 chars```

#

📎 message.txt

#

hmm

#

this has the info, just not organised well?

devout geyser May 28, 2024, 10:51 PM

#

you have got me curious, so I'm trying a similar query with my own project just to see how it handles it. I don't expect it to work as I never tested this aspect 😛

#

well mine missed some models unfortunately (such as GPT-4o). it did a lot of search queries though, what's in the table is somewhat correct, a few anomalies from what I can see

I'll try your exact query instead and see what that results in

#

if you're wondering what my current result is with my personal project 😂, as you can see, it's got a few problems in the result

sleek vortex May 28, 2024, 11:08 PM

#

devout geyser if you're wondering what my current result is with my personal project 😂, as yo...

can you try gemini 1.5 pro price per mtok

#

im getting really annoyed at this single query

devout geyser May 28, 2024, 11:08 PM

#

just that alone?

sleek vortex May 28, 2024, 11:08 PM

#

since my one can never seem to get it perplexity-like correct

#

yeah just that

#

* gemini 1.5 pro price in million tokens
* gemini 1.5 pro price per mtok```

devout geyser May 28, 2024, 11:09 PM

#

trying both variants now

sleek vortex May 28, 2024, 11:09 PM

#

when i did mtok

#

it thought mtok was a crypto

#

....

#

and when i did the first one half the time it doesnt take in the information from the main site

devout geyser May 28, 2024, 11:09 PM

#

🤣

sleek vortex May 28, 2024, 11:09 PM

#

ai.google.dev/pricing

#

half the time it gives me vertex ai pricing

#

this is so infuriating

devout geyser May 28, 2024, 11:10 PM

#

should know what mtok says in a moment

#

for mtok it said:

The pricing for Gemini 1.5 Pro is as follows:

- For input prompts up to 128K tokens, the cost is $3.50 per 1 million tokens.
- For input prompts longer than 128K tokens, the cost is $7.00 per 1 million tokens.
- For output, the cost is $10.50 per 1 million tokens for prompts up to 128K tokens and $21.00 per 1 million tokens for prompts longer than 128K tokens.

sleek vortex May 28, 2024, 11:11 PM

#

asked it for the comparison and i got this

#

come on omfg

#

why has it given so many variants

devout geyser May 28, 2024, 11:11 PM

#

lol

sleek vortex May 28, 2024, 11:12 PM

#

trying once more...

devout geyser May 28, 2024, 11:12 PM

#

trying the million one now

sleek vortex May 28, 2024, 11:13 PM

#

#

so close yet SO FAR

#

why is it only this stupid gemini one thats only right like 10% of the time

devout geyser May 28, 2024, 11:13 PM

#

well million said this:

The pricing for Gemini 1.5 Pro in terms of million tokens is as follows:

- **Input Tokens:**
  - $3.50 per million tokens for prompts up to 128K tokens.
  - $7.00 per million tokens for prompts longer than 128K tokens.

- **Output Tokens:**
  - $10.50 per million tokens for prompts up to 128K tokens.
  - $21.00 per million tokens for prompts longer than 128K tokens.

This means the cost will vary depending on the length and complexity of the input and output text.

sleek vortex May 28, 2024, 11:13 PM

#

yeah spot on

#

but again im budget limited on context

#

which might be the main issue

#

see the issue is it also doesnt know when it's wrong

#

so my feedback loop idea from before didnt really work...

devout geyser May 28, 2024, 11:15 PM

#

yeah probably, I'm using Gemini 1.5 Flash to extract relevant content from the webpages scraped, to reduce wasted input context on GPT-4o. This however does slow things down a bit, although I get an accurate answer at the end of the day so it's worth waiting about half a minute 🙂

sleek vortex May 28, 2024, 11:15 PM

#

#

right now using llama3-8b-8192 for the per-intent source handling

devout geyser May 28, 2024, 11:15 PM

#

I see

sleek vortex May 28, 2024, 11:15 PM

#

which works like

#

80% of the time

#

but some queries like this

#

it decides to not

devout geyser May 28, 2024, 11:16 PM

#

yeah I had a similar problem with some models, where it would either make something up or mix up details on the webpage

sleek vortex May 28, 2024, 11:16 PM

#

its mainly because it keeps thinking this vertex page is the main source

#

    * 🔗 https://ai.google.dev/pricing - 3699 chars
    * 🔗 https://cloud.google.com/vertex-ai/generative-ai/pricing - 11353 chars
    * 🔗 https://blog.google/technology/ai/google-gemini-next-generation-model-february-2024/ - 11572 chars```

devout geyser May 28, 2024, 11:16 PM

#

I see

sleek vortex May 28, 2024, 11:16 PM

#


Here is a detailed summary of the pricing information:

**Gemini Models Pricing**

* Model          | Price (input) | Price (output) | Notes
* Gemini 1.5 Flash | $0.0001315 / image | $0.000125 / 1k characters | 
* Gemini 1.5 Pro    | $0.001315 / image | $0.00125 / 1k characters |
* Gemini 1.0 Pro   | $0.0025 / image | $0.000375 / 1k characters | 

**Context Caching for Gemini**

* Model          | Price (input) | Price (output) | 
* Gemini 1.5 Pro | $0.0006575 / image | $0.000625 / 1k characters | 

**Imagen Pricing**

* Model          | Price (input) | Price (output) | 
* Imagen         | $0.020 per image | $0.003 per image | 

**Multimodal Embeddings Pricing**

* Model          | Price (input) | Price (output) | 
* Multimodal Embeddings | $0.0002 / 1k characters | No charge for output | 

**PaLM 2 for Text Pricing**

* Model          | Price (input) | Price (output) | 
* PaLM 2 for Text  | $0.00025 per 1,000 characters | $0.0005 per 1,000 characters | 

**Partner Models Pricing**

* Model          | Pricing | 
* Claude 3 Opus   | Input: $15 / million tokens Output: $75 / million tokens |
* Claude 3 Sonnet | Input: $3 / million tokens Output: $15 / million tokens |
* Claude 3 Haiku | Input: $0.25 / million tokens Output: $1.25 / million tokens |

**Gemini 1.5 Pricing**

* Model          | Price (input) | Price (output) | 
* Gemini 1.5 Pro | varies depending on context window size (starts at $0.0006575 / image) | varies depending on context window size |

It's important to note that the prices listed are subject to change and may vary depending on the specific use case and context window size. Additionally, there may be additional costs associated with using these models, such as latency and computational requirements.

It's also important to note that the prices listed are in USD and that prices may vary depending on the user's location.

#

so thats what this "worker" kinda outputted

#

which might be right but its interms of images??? and idfk what

devout geyser May 28, 2024, 11:17 PM

#

😂

sleek vortex May 28, 2024, 11:17 PM

#

well just not what im looking for

devout geyser May 28, 2024, 11:17 PM

#

yeah understandable

sleek vortex May 28, 2024, 11:17 PM

#

now theres only so many options i have here

#

i could use haiku instead of 8b

#

but that would increase the price per query quite a bit

#

lemme run that calculation

#

right now average

(0.59/1000000)*1500 + (0.05/1000000)*350 # groq/llama-3-70b router 
+ ((0.05/1000000)*6656 + (0.05/1000000)*768))*4 # ~4x groq/llama-3-8b search summarisers
+ (0.25/1000000)*1750 + (1.25/1000000)*400 # claude-3-haiku final

#

and with haiku

(0.59/1000000)*1500 + (0.05/1000000)*350 # groq/llama-3-70b router 
+ ((0.25/1000000)*6656 + (1.25/1000000)*768))*4 # ~4x claude-3-haiku search summarisers
+ (0.25/1000000)*1750 + (1.25/1000000)*400 # claude-3-haiku final

#

$0.012336 per query

#

which is more than double

#

😦

devout geyser May 28, 2024, 11:19 PM

#

ah so I see. I'd like to drop gemini for something else primarily because of the moderation sometimes falsely flagging, but I've yet to find a viable alternative. I'll have to try tweaking the prompt I use for extracting the relevant content with another model and see if I can get it to eventually give me similar results on some specific tests

#

ideally the model needs to be fast

sleek vortex May 28, 2024, 11:20 PM

#

i mean i could use gemini flash but again im trying to minmax costs

#

so that i could maybe turn this into a real service lol

devout geyser May 28, 2024, 11:20 PM

#

yeah of course, that's understandable

sleek vortex May 28, 2024, 11:20 PM

#

well the question here might be

#

whats actually at fault

#

whats wrong in searchworker?

websearch rankings
the scraping output
the 8b model

devout geyser May 28, 2024, 11:21 PM

#

a mystery to solve

sleek vortex May 28, 2024, 11:21 PM

#

    * 🔗 https://ai.google.dev/pricing - 3699 chars
    * 🔗 https://cloud.google.com/vertex-ai/generative-ai/pricing - 11353 chars
    * 🔗 https://blog.google/technology/ai/google-gemini-next-generation-model-february-2024/ - 11572 chars
⏰ [SearchAgent] [2.60s] Model response time (llama8b)```

#

so its shipped the context of the main source which is ai.google.dev/pricing

#

but the thing is my extraction for that page also isnt the best due to google's rubbish layout

#

📎 message.txt

devout geyser May 28, 2024, 11:22 PM

#

yeah that won't help for sure

sleek vortex May 28, 2024, 11:22 PM

#

can you show me what sources yours used

#

to answer it

devout geyser May 28, 2024, 11:22 PM

#

sure let me see

sleek vortex May 28, 2024, 11:22 PM

#

what im trying to do is like a needle in my head

#

few accurate sources per element of the query

#

what you and pplx are doing is more like search and verification by the masses?

devout geyser May 28, 2024, 11:23 PM

#

sleek vortex can you show me what sources yours used

and the ai.google.dev is the /pricing page

#

at least on the first one

sleek vortex May 28, 2024, 11:24 PM

#

hmm

#

do you have a debugger to see how it was extracted?

#

curious to see the differences in our impl

devout geyser May 28, 2024, 11:24 PM

#

yeah I do have such output, let me see what it did for that specific page

sleek vortex May 28, 2024, 11:26 PM

#

hmm

#

idk there really isnt any other good models

#

deepseek v2???

devout geyser May 28, 2024, 11:27 PM

#

<webpage url="https://ai.google.dev/pricing" search_query="gemini 1.5 pro price in tokens" publication_date="">
<result>
The Gemini 1.5 Pro model is priced at $3.50 per million tokens for input prompts up to 128K tokens and $7.00 per million tokens for input prompts longer than 128K tokens. For output, the price is $10.50 per million tokens for prompts up to 128K tokens and $21.00 per million tokens for prompts longer than 128K tokens.
</result>
</webpage>

<webpage url="https://ai.google.dev/pricing" search_query="gemini 1.5 pro price in millions" publication_date="">
<result>
Gemini 1.5 Pro is currently in preview and is free of charge. Starting May 30, 2024, it will be priced on a pay-as-you-go basis. The pricing for Gemini 1.5 Pro is as follows:

* **Input:** $3.50 / 1 million tokens (for prompts up to 128K tokens) and $7.00 / 1 million tokens (for prompts longer than 128K tokens).
* **Output:** $10.50 / 1 million tokens (for prompts up to 128K tokens) and $21.00 / 1 million tokens (for prompts longer than 128K tokens).
* **Context Caching:** $1.75 / 1 million tokens (for prompts up to 128K tokens) and $3.50 / 1 million tokens (for prompts longer than 128K tokens).
* **Storage:** $4.50 / 1 million tokens per hour.

The pricing is in USD.
</result>
</webpage>

There are a few more but I think two of them are adequate examples. What it extracts as 'relevant content' depends solely on the search query used basically.

sleek vortex May 28, 2024, 11:28 PM

#

are you using scraping like me or a headless web browser?

#

like do you know what text content came extracted out of that page

devout geyser May 28, 2024, 11:28 PM

#

requests in Python, so not a headless browser

sleek vortex May 28, 2024, 11:28 PM

#

thats what im curious about

sleek vortex May 28, 2024, 11:28 PM

#

sleek vortex like do you know what text content came extracted out of that page

?

devout geyser May 28, 2024, 11:28 PM

#

sleek vortex ?

to answer that question, I don't explicitly output that in the console but I can certainly add that output

sleek vortex May 28, 2024, 11:28 PM

#

sleek vortex May 28, 2024, 11:28 PM

#

devout geyser to answer that question, I don't explicitly output that in the console but I can...

yeah if you can 🙏

#

just would like to see :d

devout geyser May 28, 2024, 11:29 PM

#

sure, one moment

#

I think I'll save the output to file instead, as it will be messy to find it in the console

sleek vortex May 28, 2024, 11:30 PM

#

avg ~$0.00642752 per query with deepseek-v2-32k (in theory)
vs
avg ~$0.0034784 per query with llama3-8b-8192 (in theory)

#

interesting

cinder comet May 28, 2024, 11:30 PM

#

sleek vortex

How abput yi-large

#

Is it available yet through api?

#

It seems slow on lmsys

sleek vortex May 28, 2024, 11:31 PM

#

platform.01.ai seems to exist

#

on a waitlist...

devout geyser May 28, 2024, 11:38 PM

#

sleek vortex yeah if you can 🙏

Here you go

📎 message.txt

sleek vortex May 28, 2024, 11:38 PM

#

ok well in the end nobody is beating the price of groq

tame current May 28, 2024, 11:39 PM

#

devout geyser Here you go

whar

sleek vortex May 28, 2024, 11:39 PM

#

Hmm, your parsing is very similisr to mine

devout geyser May 28, 2024, 11:39 PM

#

That's what I send to Gemini 1.5 Flash to extract relevant content from that

sleek vortex May 28, 2024, 11:39 PM

#

so it must be down to a case of the model just being confused due to the extra sources and not being able to pickup on the first source

#

man

#

idek how to fix this without changing model

#

changing model isn’t really an option i’ll be entirely honest

devout geyser May 28, 2024, 11:40 PM

#

possibly a combination of a good prompt and right temperature/top_p, but if the model just isn't good enough to do it then rip 😦

sleek vortex May 28, 2024, 11:40 PM

#

WAIT you’re right i haven’t set temperature

#

WAAAIT

devout geyser May 28, 2024, 11:40 PM

#

so far the only model I've had actual success with is flash, haiku worked but it wasn't as detailed with the docker compose release notes

sleek vortex May 28, 2024, 11:40 PM

#

you’re onto something

#

did i forget temperature setting

#

am checking../

devout geyser May 28, 2024, 11:42 PM

#

I found temperature alone didn't work for me, I had to also adjust top_p to stop Flash randomly getting the docker compose release notes "test" wrong every so often

sleek vortex May 28, 2024, 11:42 PM

#

I’ve never really played with top_p

#

yeah… i don’t think i’ve set a temperature

#

what’s the default

#

1.0???

devout geyser May 28, 2024, 11:43 PM

#

nope, mine is 0.3 actually

sleek vortex May 28, 2024, 11:43 PM

#

but what’s the default temperature for groq api

devout geyser May 28, 2024, 11:44 PM

#

not sure on that one

agile jay May 28, 2024, 11:45 PM

#

sleek vortex ok well in the end nobody is beating the price of groq

Yep, or the speed.

devout geyser May 28, 2024, 11:45 PM

#

I tried asking about that in my personal project, it didn't find the answer

sleek vortex May 28, 2024, 11:45 PM

#

what’s your top_p set to?

sleek vortex May 28, 2024, 11:45 PM

#

devout geyser I tried asking about that in my personal project, it didn't find the answer

yeah the answer doesn’t exist i think

devout geyser May 28, 2024, 11:45 PM

#

0.9

sleek vortex May 28, 2024, 11:45 PM

#

nowhere on the groq docs

agile jay May 28, 2024, 11:46 PM

#

sleek vortex nowhere on the groq docs

What is?

sleek vortex May 28, 2024, 11:46 PM

#

sleek vortex but what’s the default temperature for groq api

.

agile jay May 28, 2024, 11:46 PM

#

1 I would imagine

#

devout geyser May 28, 2024, 11:47 PM

#

yeah I'd probably say it's 1 by default, most APIs usually go with 1 by default

sleek vortex May 28, 2024, 11:47 PM

#

agile jay

i know what this means due to that 3blue1brown video

#

lol

#

have you seen it?

#

the ones on transformers

agile jay May 28, 2024, 11:48 PM

#

Yep, good for removing unlikely tokens.

sleek vortex May 28, 2024, 11:48 PM

#

oh my god

#

#

it worked???

agile jay May 28, 2024, 11:48 PM

#

Looks like the default is 1, so all probabilities are considered.

agile jay May 28, 2024, 11:48 PM

#

sleek vortex

Nice, what change you make?

sleek vortex May 28, 2024, 11:48 PM

#

just the temperature

sleek vortex May 28, 2024, 11:49 PM

#

sleek vortex WAIT you’re right i haven’t set temperature

i realised all of them were at default (1) in my code

agile jay May 28, 2024, 11:49 PM

#

Lol, to make it more deterministic?

sleek vortex May 28, 2024, 11:49 PM

#

oh my god youre telling me i had such good results already with 1 temp

#

?????

devout geyser May 28, 2024, 11:49 PM

#

setting the right parameters helps

sleek vortex May 28, 2024, 11:49 PM

#

silliest oversight from me but

agile jay May 28, 2024, 11:49 PM

#

Lowering it should reduce hallucinations.

sleek vortex May 28, 2024, 11:49 PM

#

yeah

#

worked well

agile jay May 28, 2024, 11:50 PM

#

And lowering the Top P should help too.

sleek vortex May 28, 2024, 11:50 PM

#

yeah

#

llama70b = Groq(model="llama3-70b-8192", temperature=0.6, top_p=0.9)
llama8b = Groq(model="llama3-8b-8192", temperature=0.2, top_p=0.9)

claude = Claude(model="claude-3-haiku-20240307",
                api_key="",
                temperature=0.3, top_p=0.9
                )```

#

ive gone with this configuration

agile jay May 28, 2024, 11:50 PM

#

It's probably set like that since the default is best for general tasks.

sleek vortex May 28, 2024, 11:50 PM

#

70b is the initial model, 8b is the source summarisers, and claude is the final response

devout geyser May 28, 2024, 11:50 PM

#

llama70b is for what? generating the search queries?

sleek vortex May 28, 2024, 11:50 PM

#

(api key redacted)

agile jay May 28, 2024, 11:50 PM

#

Such as writing etc, where randomness is seen as a feature at times.

sleek vortex May 28, 2024, 11:51 PM

#

devout geyser llama70b is for what? generating the search queries?

i've adopted an intent/destructing model (if youve been up-to-date)

agile jay May 28, 2024, 11:51 PM

#

devout geyser llama70b is for what? generating the search queries?

I am guessing it's for writing the intent and steps.

devout geyser May 28, 2024, 11:51 PM

#

ah ok I see

#

sorry, nearly 1am here so I'm half sleepy 😛

agile jay May 28, 2024, 11:51 PM

#

sleek vortex i've adopted an intent/destructing model (if youve been up-to-date)

Yep, is it the style I recommended?

sleek vortex May 28, 2024, 11:51 PM

#

yeah

#

[
  {
    "intent": "Get the prices of the AI models",
    "steps": [
      {
        "tool": "search",
        "query": "Claude 3 models pricing"
      },
      {
        "tool": "search",
        "query": "GPT-4 Turbo pricing"
      },
      {
        "tool": "search",
        "query": "GPT4O pricing"
      },
      {
        "tool": "search",
        "query": "Gemini 1.5 Pro+Flash pricing"
      }
    ]
  },
  {
    "intent": "Convert prices to million tokens",
    "steps": [
      {
        "tool": "calculator",
        "query": "convert prices to million tokens"
      }
    ]
  },
  {
    "intent": "Create a price comparison table",
    "steps": [
      {
        "tool": "write_answer",
        "query": "create table with prices in million tokens"
      }
    ]
  }
]```

agile jay May 28, 2024, 11:51 PM

#

Intent is what I use, since it's the models guess of what you wanted.

sleek vortex May 28, 2024, 11:51 PM

#

maybe i ought to lower temp more

#

right now it ignored calculator since i havent impl'd it

agile jay May 28, 2024, 11:52 PM

#

Yep

#

0.1 is probably good.

#

You want it to use sources, and not make stuff up.

sleek vortex May 28, 2024, 11:52 PM

#

yeah but then im thinking

devout geyser May 28, 2024, 11:52 PM

#

yeah play around with the temperature and see what results you get basically

sleek vortex May 28, 2024, 11:52 PM

#

will it just go really bad on the queires like the japan one

agile jay May 28, 2024, 11:52 PM

#

sleek vortex will it just go really bad on the queires like the japan one

No, it shouldn't

sleek vortex May 28, 2024, 11:52 PM

#

so when i did it before and got a crazy good response, it was on temp 1

#

so idk

#

it did like 7 searches on different aspects

#

it was really good lol

devout geyser May 28, 2024, 11:53 PM

#

for curiosity I'll try that japan query with mine 😛

sleek vortex May 28, 2024, 11:53 PM

#

devout geyser yeah play around with the temperature and see what results you get basically

yeah ok

sleek vortex May 28, 2024, 11:53 PM

#

devout geyser for curiosity I'll try that japan query with mine 😛

here it is

#

plan me a trip to japan somewhere in the month of may 2024 from london

devout geyser May 28, 2024, 11:53 PM

#

thanks

sleek vortex May 28, 2024, 11:53 PM

#

raycast clipboard is godly

#

agile jay May 28, 2024, 11:53 PM

#

Also, you can change the temperature for the different steps, if needed.

sleek vortex May 28, 2024, 11:53 PM

#

yeah

#

i did

#

oh like

#

search vs calc

#

yeah fair enough

agile jay May 28, 2024, 11:54 PM

#

Yep

sleek vortex May 28, 2024, 11:54 PM

#

for now ive hidden the nonexistent tools that we mocked up though

#

#

calulcator doesnt actually work rn, but the thing is i have it in the shot prompts so to avoid confusion/keep consistency its there

agile jay May 28, 2024, 11:54 PM

#

sleek vortex

Yep, my infrastructure is an intent based model, and I also think intent just sounds cool too.

agile jay May 28, 2024, 11:55 PM

#

sleek vortex calulcator doesnt actually work rn, but the thing is i have it in the shot promp...

Could plug it into wolfram alpha?

sleek vortex May 28, 2024, 11:55 PM

#

intent makes me think of like

#

google assistant

sleek vortex May 28, 2024, 11:55 PM

#

agile jay Could plug it into wolfram alpha?

yeah

agile jay May 28, 2024, 11:55 PM

#

For the calc

sleek vortex May 28, 2024, 11:55 PM

#

searxng has wolfram

agile jay May 28, 2024, 11:55 PM

#

Nice

sleek vortex May 28, 2024, 11:55 PM

#

i can use that

#

well ive hosted my own searxng

agile jay May 28, 2024, 11:56 PM

#

searxng is pretty nice.

sleek vortex May 28, 2024, 11:56 PM

#

since the public ones have mostly outdated/cached responses

#

and arent as reliable

agile jay May 28, 2024, 11:56 PM

#

Yep, also better for latency long term.

#

I am hooking up tools using rpc, so they are pretty seamless.

#

And not coupled.

devout geyser May 28, 2024, 11:57 PM

#

ok I have no idea how much of this is correct but this is what the japan query said for me on my personal project. Mine isn't as comprehensive in that it won't search for plane ticket availability or such

📎 message.txt

sleek vortex May 28, 2024, 11:58 PM

#

ooh but it provided budget

#

whats your end model?

#

gpt4o?

devout geyser May 28, 2024, 11:58 PM

#

gpt-4o yes

sleek vortex May 28, 2024, 11:58 PM

#

fair enough then

#

haiku aint telling me about a simcard any time soon lmao

devout geyser May 28, 2024, 11:59 PM

#

😂

sleek vortex May 28, 2024, 11:59 PM

#

gemini 1.5 pro price in million tokens

#

back to this query

#

sometimes it works

#

sometimes it doesnt

devout geyser May 28, 2024, 11:59 PM

#

for the sake of curiosity, I could try a different end model if you want

agile jay May 28, 2024, 11:59 PM

#

devout geyser for the sake of curiosity, I could try a different end model if you want

He's using Haiku atm

sleek vortex May 28, 2024, 11:59 PM

#

im using haiku

#

mainly since i do so much testing id go broke in gpt4 credit

agile jay May 29, 2024, 12:00 AM

#

Yep, and the speed too...

devout geyser May 29, 2024, 12:00 AM

#

yeah understandable, I'll try Haiku with the same query on my setup

sleek vortex May 29, 2024, 12:00 AM

#

🔎 [0.00s] Searched for "Gemini 1.5 Pro price"@SearXNG - got 5 links, 0 snippets
🔗 [SearchAgent] [0.00s] Picked 4 links total
🚀 [SearchAgent] [7.63s] Finished pulling 3 sources
    * 🔗 https://ai.google.dev/pricing - 3699 chars
    * 🔗 https://cloud.google.com/vertex-ai/generative-ai/pricing - 11353 chars
    * 🔗 https://artificialanalysis.ai/models/gemini-1-5-pro - 11572 chars
⏰ [SearchAgent] [2.41s] Model response time (llama8b)
Based on the provided context, the Gemini 1.5 Pro model is a multimodal model that can be used for various tasks such as text generation, image generation, and multimodal fusion. The pricing for the Gemini 1.5 Pro model is as follows:

* Input token price: $0.001315 per image, $0.001315 per second, $0.00125 per 1k characters
* Output token price: $0.00375 per 1k characters
* Context caching: 0.0006575 per image, 0.0006575 per second, 0.000625 per 1k characters
* Context cache storage: 0.0011835 per image per hour, 0.0011835 per second per hour, 0.001125 per 1k characters per hour

The pricing for the Gemini 1.5 Pro model is based on the number of input and output tokens, as well as the context caching and storage. The prices are listed in US Dollars (USD) and are subject to change.

It's worth noting that the pricing for the Gemini 1.5 Pro model is different from the pricing for other models, such as the Gemini 1.5 Flash model, which has a different pricing structure. ```

#

really...

#

come on

#

😐

#

back to square one 😭

devout geyser May 29, 2024, 12:01 AM

#

ah crud, I can't actually do that easily as I forgot I'm using OpenAI directly, not using OR for the final model. Well I'll try it tomorrow when I'm more awake 😛

sleek vortex May 29, 2024, 12:01 AM

#

fair enough

#

lol yeah 1am so

#

well idk

#

even pplx pro does the crypto thing

agile jay May 29, 2024, 12:04 AM

#

...

#

Confused I guess

sleek vortex May 29, 2024, 12:05 AM

#

yeah but mine takes it too seriously

#

how do i stop it from thinking million tokens is a crypto

#

ok for the sake of testing

#

let me try sonnet's output

devout geyser May 29, 2024, 12:06 AM

#

it wasn't as difficult as I thought it would be to change it to haiku, here's what Haiku said, which isn't as detailed:

Here is a suggested 10-day Japan itinerary for a trip in May 2024 from London:

Day 1: Arrive in Tokyo
- Check into your hotel
- Explore the Asakusa district, including the Sensoji Temple and Nakamise shopping street
- Visit the Imperial Palace East Gardens

Day 2: Tokyo
- Visit the Meiji Shrine and Yoyogi Park
- Explore the Shibuya Crossing and Harajuku district
- Attend the Sanja Matsuri festival in Asakusa (mid-May)

Day 3: Tokyo to Kyoto
- Take the Shinkansen bullet train to Kyoto (approx. 2.5 hours)
- Visit the Kinkakuji (Golden Pavilion) and Arashiyama Bamboo Grove
- Explore the Gion district and watch a traditional geisha performance

Day 4: Kyoto
- Visit the Kiyomizudera Temple and Fushimi Inari Shrine
- Explore the Nijo Castle and Nishiki Market

Day 5: Kyoto to Nara
- Take a day trip to Nara
- See the friendly deer in Nara Park
- Visit the Todaiji Temple and Kasuga Taisha Shrine

Day 6: Kyoto to Hiroshima
- Take the Shinkansen to Hiroshima (approx. 2 hours)
- Visit the Hiroshima Peace Memorial Park and Museum
- See the iconic Itsukushima Shrine on Miyajima Island

Day 7: Hiroshima to Osaka
- Travel to Osaka (approx. 1.5 hours)
- Explore the Dotonbori district and try the local cuisine
- Visit the Osaka Castle

Day 8: Osaka
- Take a day trip to Himeji Castle
- Explore the Kobe Harborland and Kitano Ijinkan district

Day 9: Osaka to Hakone
- Travel to Hakone (approx. 2 hours)
- Ride the Hakone Ropeway and enjoy the views of Mount Fuji
- Relax in an onsen (hot spring)

Day 10: Hakone to Tokyo, depart
- Return to Tokyo (approx. 1.5 hours)
- Explore any remaining sights in Tokyo
- Depart for London

This itinerary allows you to experience the highlights of Tokyo, Kyoto, Nara, Hiroshima, Osaka, and Hakone, with a focus on cultural attractions, festivals, and natural scenery. Let me know if you would like me to modify or expand on this suggested Japan trip plan for May 2024.

agile jay May 29, 2024, 12:08 AM

#

So just a simpler and less accurate response.

devout geyser May 29, 2024, 12:08 AM

#

yeah

sleek vortex May 29, 2024, 12:08 AM

#

yeah

agile jay May 29, 2024, 12:08 AM

#

It's probably still affordable, as long as there is more filtering before passing into 4o.

sleek vortex May 29, 2024, 12:09 AM

#

wha

#

no like he replaced the 4o at the end with haiku (like i have)

agile jay May 29, 2024, 12:10 AM

#

sleek vortex no like he replaced the 4o at the end with haiku (like i have)

I know

sleek vortex May 29, 2024, 12:10 AM

#

oh

#

agile jay May 29, 2024, 12:11 AM

#

Some finetuning of llama 70B could also do the trick.

sleek vortex May 29, 2024, 12:11 AM

#

ok im moving on from this one gemini query

agile jay May 29, 2024, 12:11 AM

#

Can't wait for groq to add it.

sleek vortex May 29, 2024, 12:11 AM

#

i cant get it to work

#

but whatever 😐

agile jay May 29, 2024, 12:11 AM

#

Yep, asking for the cost of all the different models has always been hard.

sleek vortex May 29, 2024, 12:11 AM

#

no

devout geyser May 29, 2024, 12:11 AM

#

although slower to reply, here's what wizardlm-2-8x22b also said for anyone curious 🙂

📎 message.txt

sleek vortex May 29, 2024, 12:11 AM

#

this is different

sleek vortex May 29, 2024, 12:12 AM

#

agile jay Yep, asking for the cost of all the different models has always been hard.

i got that query down

#

this is gemini 1.5 pro price in million tokens
problem 1: keeps quoting https://cloud.google.com/vertex-ai/generative-ai/pricing (this is completely unrelated and has prices in per character)
problem 2: thinks million tokens is some crypto

agile jay May 29, 2024, 12:14 AM

#

I think making a knowledge graph would probably be useful.

#

And then injecting the knowledge if it's relevent to the query.

sleek vortex May 29, 2024, 12:14 AM

#


For prompts up to 128K tokens:
- $0.35 per 1 million input tokens
- $1.05 per 1 million output tokens

For prompts longer than 128K tokens:
- $0.70 per 1 million input tokens  
- $2.10 per 1 million output tokens```

agile jay May 29, 2024, 12:14 AM

#

sleek vortex ```According to the information provided, the price for Gemini 1.5 Pro in Millio...

?

sleek vortex May 29, 2024, 12:14 AM

#

well its like 50% right...

#

it didnt do the crypto thing

#

but thats the flash prices, not pro

agile jay May 29, 2024, 12:15 AM

#

Maybe ask it to repeat what it found before giving the answer?

#

But as a summary, not the whole number of sources.

sleek vortex May 29, 2024, 12:16 AM

#

🔎 [0.91s] Searched for "Gemini 1.5 Pro price in Million Tokens"@SearXNG - got 22 links, 0 snippets
🔗 [SearchAgent] [0.01s] Picked 6 links total
🚀 [SearchAgent] [7.17s] Finished pulling 4 sources
    * 🔗 https://artificialanalysis.ai/models/gemini-1-5-pro - 11600 chars
    * 🔗 https://indianexpress.com/article/explained/explained-sci-tech/google-gemini-pro-1-5-1-million-tokens-9166398/ - 11600 chars
    * 🔗 https://ai.google.dev/pricing - 3699 chars
    * 🔗 https://www.cnet.com/tech/services-and-software/googles-gemini-1-5-pro-will-have-2-million-tokens-heres-what-that-means/ - 2797 chars
⏰ [SearchAgent] [2.83s] Model response time (llama8b)
"""Based on the provided context, I understand that you are looking for the price of Gemini 1.5 Pro in Million Tokens. According to the text, the pricing for Gemini 1.5 Pro is as follows:

* For prompts up to 128K tokens: $0.35 / 1 million tokens (input) and $1.05 / 1 million tokens (output)
* For prompts longer than 128K tokens: $0.70 / 1 million tokens (input) and $2.10 / 1 million tokens (output)

Please note that these prices are subject to change and may vary depending on the specific use case and requirements."""```

#

time to look into these sources

#

well first, both these sources have some parsing issue...

#

agile jay May 29, 2024, 12:17 AM

#

Yep, must be really confusing for the model.

#

What about suing 70B and asking it to remove any errors from the scraping? Maybe it's already knowledgeable to do it.

sleek vortex May 29, 2024, 12:18 AM

#

it would slow the query too much tbf

agile jay May 29, 2024, 12:18 AM

#

Just as a test.

sleek vortex May 29, 2024, 12:18 AM

#

if this was to become a site where i can get active like rlhf feedback from users then maybe i could see when these sites make certain issues

agile jay May 29, 2024, 12:19 AM

#

Since if 8B can do it too, then it should be pretty fast with groq.

sleek vortex May 29, 2024, 12:26 AM

#


$10.50 per 1M Tokens for prompts up to 128K tokens
$21.00 per 1M Tokens for prompts longer than 128K tokens

The key details are:

- Gemini 1.5 Pro has a pay-as-you-go pricing model
- For prompts up to 128K tokens, the price is $0.35 per 1M Tokens
- For prompts longer than 128K tokens, the price is $0.70 per 1M Tokens
- There are also additional charges for output prompts
- Billing for Gemini 1.5 Pro starts on May 30, 2024

Please note that these prices are subject to change and may vary based on the specific use case and requirements.
⏱️ 12.199028968811035 seconds```

#

oh my god FINALLY it works

#

ok well again its only 75% correct

#

indigo plank May 29, 2024, 12:38 AM

#

yo sneakyfishy

agile jay May 29, 2024, 12:38 AM

#

Yep, I'm just working on an easier way to clean up the input data.

sleek vortex May 29, 2024, 12:42 AM

#

indigo plank yo sneakyfishy

yo

sleek vortex May 29, 2024, 12:42 AM

#

agile jay Yep, I'm just working on an easier way to clean up the input data.

you mentioned ages ago

#

markdown

#

im trying that rn

indigo plank May 29, 2024, 12:43 AM

#

i dont want to sound like dumb or anything but once you use perplexity what like website do you sue to bypass like the ai detection

sleek vortex May 29, 2024, 12:43 AM

#

perplexity since it cites the web usually doesnt sound like ai much anyway

#

if you want you can click copy, remove the sources at the end and ask ai to rephrase it, or if youre super paranoid something like quillbot

indigo plank May 29, 2024, 12:44 AM

#

its for presentation

sleek vortex May 29, 2024, 12:44 AM

#

but if its for important research, or maybe an assignment youre turning in, id reccomend you rephrase anything manually really

indigo plank May 29, 2024, 12:44 AM

#

and i have to include all like the sources and everything

sleek vortex May 29, 2024, 12:44 AM

#

indigo plank its for presentation

you could ask it to give the main points for each slide/etc

#

you can copy it over and change a few words

#

then add sources as seems fit

agile jay May 29, 2024, 12:45 AM

#

Or you can just use perplexity to quickly find sources you can use.

sleek vortex May 29, 2024, 12:45 AM

#

yeah

patent rapids May 29, 2024, 12:47 AM

#

Can't you summarize YouTube videos?

sleek vortex May 29, 2024, 12:48 AM

#

gemini is decent at that

#

you can but idk if it works well

agile jay May 29, 2024, 12:54 AM

#

Yep, also thinking how to handle graphs and svg's.
I'm guessing removing the actual svg content and only leaving the class and stroke/fill will be enough for those.

#

For graphs, probably just leaving them as plain html would work.

#

Think I'm gonna make a quick interface to measure how many tokens are saved from each method, compared to the average end result of the query output.

sleek vortex May 29, 2024, 12:58 AM

#

i just stripped images and svg entirely

#


$10.50 per 1M Tokens for prompts up to 128K tokens
$21.00 per 1M Tokens for prompts longer than 128K tokens

The pricing details are:

- Input: $7.00 per 1M Tokens (for prompts up to 128K tokens) and $21.00 per 1M Tokens (for prompts longer than 128K tokens)
- Output: $10.50 per 1M Tokens (for prompts up to 128K tokens) and $21.00 per 1M Tokens (for prompts longer than 128K tokens)

Please note that these prices are subject to change and may vary depending on the context window and other factors.```

#

i will take that

#

[
  {
    "intent": "Plan a trip to Japan in May 2024 from London",
    "steps": [
      {
        "tool": "search",
        "query": "flights from London to Japan in May 2024"
      },
      {
        "tool": "search",
        "query": "best places to visit in Japan in May"
      },
      {
        "tool": "search",
        "query": "Japan weather in May"
      },
      {
        "tool": "search",
        "query": "Japan travel guide"
      }
    ]
  },
  {
    "intent": "Create an itinerary",
    "steps": [
      {
        "tool": "search",
        "query": "7-day Japan itinerary"
      },
      {
        "tool": "search",
        "query": "things to do in Tokyo in May"
      },
      {
        "tool": "search",
        "query": "things to do in Kyoto in May"
      }
    ]
  },
  {
    "intent": "Book accommodations and flights",
    "steps": [
      {
        "tool": "search",
        "query": "book flights from London to Japan in May 2024"
      },
      {
        "tool": "search",
        "query": "book hotel in Tokyo"
      },
      {
        "tool": "search",
        "query": "book hotel in Kyoto"
      }
    ]
  },
  {
    "intent": "Return the final answer",
    "steps": [
      {
        "tool": "write_answer"
      }
    ]
  }
]```

#

interestingly does a huge levelled response for japan query

#

like before

#

even on low temp

agile jay May 29, 2024, 1:02 AM

#

Yep, because lowering the temp just reduces the randomness

#

Which is good when using sources.

sleek vortex May 29, 2024, 1:03 AM

#

[
  {
    "intent": "Plan a trip to Japan in May 2024 from London",
    "steps": [
      {
        "tool": "search",
        "query": "flights from London to Japan in May 2024"
      },
      {
        "tool": "search",
        "query": "best places to visit in Japan in May"
      },
      {
        "tool": "search",
        "query": "Japan weather in May"
      },
      {
        "tool": "search",
        "query": "Japan itinerary for 7-10 days"
      }
    ]
  },
  {
    "intent": "Get accommodation options",
    "steps": [
      {
        "tool": "search",
        "query": "hotels in Tokyo"
      },
      {
        "tool": "search",
        "query": "best areas to stay in Japan"
      }
    ]
  },
  {
    "intent": "Plan transportation and activities",
    "steps": [
      {
        "tool": "search",
        "query": "Japan train tickets"
      },
      {
        "tool": "search",
        "query": "things to do in Tokyo in May"
      }
    ]
  },
  {
    "intent": "Return the final trip plan",
    "steps": [
      {
        "tool": "write_answer"
      }
    ]
  }
]```

#

ran it again

#

even better???

#

groq.BadRequestError: Error code: 400 - {'error': {'message': 'Please reduce the length of the messages or completion.', 'type': 'invalid_request_error', 'param': 'messages', 'code': 'context_length_exceeded'}}

#

uhh

#

    * 🔗 https://www.selectiveasia.com/japan-holidays/weather/may - 366 chars
    * 🔗 https://top.his-usa.com/destination-japan/blog/a_guide_to_japan_-_may_and_june.html - 9822 chars
    * 🔗 https://www.japan-guide.com/e/e2273.html - 10000 chars
    * 🔗 https://www.holiday-weather.com/tokyo/averages/may/ - 7460 chars```

agile jay May 29, 2024, 1:04 AM

#

Rip, too long

sleek vortex May 29, 2024, 1:06 AM

#


Flights:
- Depart London on May 1, 2024 and return on May 15, 2024.
- Based on the search results, the cheapest flights from London to Tokyo during this time period are around £402 roundtrip. The flights will take approximately 13 hours and 49 minutes each way.
- The most popular airline for this route is Iberia.

Accommodations:
- For your 2-week trip, I would recommend staying in a mix of hotels and traditional ryokans (Japanese inns) to experience both modern and cultural aspects of Japan.
- In Tokyo, consider staying at the Hotel Gracery Shinjuku, which has an 8.3 rating and rates starting around £128 per night.
- In Kyoto, you could stay at a ryokan like Yoshida-sanso, which offers a more authentic Japanese experience.
- In Hakone, the Park Hotel Tokyo is a luxury option with stunning views of Mount Fuji, starting around £165 per night.

Itinerary Suggestions:
- Spend 4-5 days in Tokyo to see the top sights like the Imperial Palace, Sensoji Temple, and explore the diverse neighborhoods.
- Take a day trip to Kamakura to visit the famous Daibutsu (Great Buddha) statue and historic temples.
- Spend 3-4 days in Kyoto to see the Kinkakuji, Kiyomizudera, and Arashiyama bamboo forest.
- Visit Hakone for 2-3 days to ride the Hakone Ropeway, see Lake Ashi, and try to catch a glimpse of Mount Fuji.
- Consider a day trip to Nara to see the friendly deer and historic temples.
- Spend the remaining days exploring other areas of interest, such as Hiroshima, Miyajima, or Kanazawa.

Let me know if you need any other details or have additional requests for your Japan trip planning!
⏱️ 42.565316915512085 seconds```

#

the hotel pricing is decently accurate

#

the flight not so much?

#

well its literally may 29

#

how is it showing info for a flight in the past

agile jay May 29, 2024, 1:08 AM

#

Looks like they can save a lot. Now I wanna find out which filters have the most gain, and for what sites. So I can create embeddings to choose which filters to apply.

sleek vortex May 29, 2024, 1:08 AM

#

agile jay May 29, 2024, 1:08 AM

#

sleek vortex May 29, 2024, 1:08 AM

#

whys your cleaned html so high

#

my cleaned text was like only 1.2x markdown size

#

markdown just added like the bolding and title separation

agile jay May 29, 2024, 1:09 AM

#

Your cleaned html is the text content right?

sleek vortex May 29, 2024, 1:09 AM

#

yeah

#

well i developed it decently well

agile jay May 29, 2024, 1:09 AM

#

Mine is actual html

sleek vortex May 29, 2024, 1:09 AM

#

oh lol

agile jay May 29, 2024, 1:09 AM

#

And tailwind css adds a lot to the amount of characters lol.

#

The initial cleanup is stuff like removing tags like script, and element with classes which include nav, sidebar, header, footer etc.

#

But I'm gonna make more advanced ones to remove stuff like all the tailwindcss classes.

#

But the aim is to remove any extra clutter which isn't part of the main content.

#

So stuff like the apple webpage would be doable after cleanup.

#

There's the cleaned html is getting smaller.

agile jay May 29, 2024, 1:31 AM

#

sleek vortex

After the cleanup, and then passed back to llama 3 70B with the initial prompt, I get this as the source:

### Gemini 1.5 Pro: Quality, Performance & Price Analysis

**Quality:** Gemini 1.5 Pro is of higher quality compared to average, with a MMLU score of 0.819 and a Quality Index across evaluations of 88.

**Price:** Gemini 1.5 Pro is more expensive compared to average with a price of **$10.50 per 1M Tokens** (blended 3:1). Gemini 1.5 Pro Input token price: $7.00, Output token price: $21.00 per 1M Tokens.

**Speed:** Gemini 1.5 Pro is slower compared to average, with a throughput of 56.2 tokens per second.

**Latency:** Gemini 1.5 Pro has a higher latency compared to average, taking 0.95 s to receive the first token (TTFT).

**Context Window:** Gemini 1.5 Pro has a larger context window than average, with a context window of 1.0M tokens.

I think making the input super small like this is probably the way to go for price.

wise edge May 29, 2024, 2:58 AM

#

harsh stag May 29, 2024, 3:38 AM

#

wise edge

you can't

warm cave May 29, 2024, 4:08 AM

#

harsh stag you can't

True

mighty gale May 29, 2024, 5:19 AM

#

wise edge

Try disabling the Pro button.

wide mesa May 29, 2024, 5:59 AM

#

Really only writing mode doesn't have "search"

plush cairn May 29, 2024, 7:41 AM

#

I got an issue, in the last two days while using the Perplexity answer repeating the previous answer with different prompt

austere kestrel May 29, 2024, 8:09 AM

#

agile jay Yep, because lowering the temp just reduces the randomness

yeah I think this is part of the issue.. and for all the models involved - like if one is told to find links relevant to X, and another is tasked with producing a response about X, the smaller models are prone to just making it up if the information (about X) isn't actually there (or if it's there but is embedded within a bunch of unstructured and non-relevant information)

#

lowering temp would presumably help

#

but it's also kinda just a limitation with the smaller models imo.. they just can't parse lots of information and stay focussed on multiple requirements effectively.. they get lost in a way that models like GPT-4 and Opus don't (or at least are less likely to)

#

btw came across this yesterday. haven't tried it out, but looks interesting (/potentially relevant @sleek vortex ) https://www.firecrawl.dev/

Firecrawl

Turn any website into LLM-ready data.

stable radish May 29, 2024, 11:18 AM

#

austere kestrel btw came across this yesterday. haven't tried it out, but looks interesting (/po...

Firecrawl looks very interesting! Thanks for sharing here

devout geyser May 29, 2024, 12:25 PM

#

I recall seeing that website, not cheap

sleek vortex May 29, 2024, 1:29 PM

#

devout geyser I recall seeing that website, not cheap

yeah, looks expensive

#

i mean i think its open source

#

might take some of these
https://github.com/mendableai/firecrawl/blob/main/apps/api/src/scraper/WebScraper/utils/custom/website_params.ts

#

well ive done similiar things

#

just isnt organised lol

devout geyser May 29, 2024, 1:31 PM

#

yeah looks like you can self host it for free

sleek vortex May 29, 2024, 1:35 PM

#

there was this very fast scraper

#

https://github.com/projectdiscovery/katana

GitHub

GitHub - projectdiscovery/katana: A next-generation crawling and sp...

A next-generation crawling and spidering framework. - projectdiscovery/katana

#

it doesnt do any of the parsing

#

this is just a crawler

#

but its pretty fast

#

written in go

devout geyser May 29, 2024, 1:37 PM

#

I see

sleek vortex May 29, 2024, 1:40 PM

#

so what if we kept a function warm on like aws lambda

#

with a chromium browser backed by s3 scraping cache

#

then maybe when one query hits a site, we can go in the background and scrape the whole site into s3 cache

#

idk

#

but is there really much benefit from that

fleet sandal May 29, 2024, 2:08 PM

#

I subscribed to Perplexity a day ago, but I'm having a problem that I haven't had with others. It responds quite slowly and, worse, it repeats itself. For example, if I have an error in my code, it gives me a solution. I then ask something else, and it repeats the same thing I asked in previous questions. Is there any solution? Am I using the chat incorrectly? I don't know. These things never happen to me with the official GPT page or even with Phind.

south kindle May 29, 2024, 2:14 PM

#

fleet sandal I subscribed to Perplexity a day ago, but I'm having a problem that I haven't ha...

I've noticed this with GPT-4o recently. Try to have it re-write using OPUS.

proper sage May 29, 2024, 2:23 PM

#

I subscribed to a pro plan today, and it seemed like we had a limit of 600 requests per day. Where can I see the remaining credits? It's not displayed anywhere in my account?

sleek vortex May 29, 2024, 2:28 PM

#

proper sage I subscribed to a pro plan today, and it seemed like we had a limit of 600 reque...

hover on the pro button

#

if youre enterprise yeah it just doesnt say

#

if you really want to check then open this link https://www.perplexity.ai/_next/data/hoDUGcZA5-ruYK-5Lc8a9/en-US/settings/org.json and look for this section

proper sage May 29, 2024, 2:32 PM

#

Over Pro button i just see CTRL .

sleek vortex May 29, 2024, 2:32 PM

#

sleek vortex if you really want to check then open this link <https://www.perplexity.ai/_next...

yeah then youll have to do this if you want to really check

#

they hide it for enterprise pro, idk why

fleet sandal May 29, 2024, 2:33 PM

#

south kindle I've noticed this with GPT-4o recently. Try to have it re-write using OPUS.

yes the bad thing is that I ask 5 things and they make me wait a whole day to use it again 😦 I have a 20 dollar subscription.

sleek vortex May 29, 2024, 2:33 PM

#

well i dont think ive ever hit 600 anyway

proper sage May 29, 2024, 2:33 PM

#

ok it's was just for check but i'm not enterprise

sleek vortex May 29, 2024, 2:33 PM

#

oh

#

then they hide for all

#

not sure

proper sage May 29, 2024, 2:33 PM

#

yes just for curiosity 😄

sleek vortex May 29, 2024, 2:33 PM

#

they used to show it when it was 300

#

¯_(ツ)_/¯

proper sage May 29, 2024, 2:33 PM

#

BEcause i played so much today haha

sleek vortex May 29, 2024, 2:33 PM

#

fleet sandal yes the bad thing is that I ask 5 things and they make me wait a whole day to us...

pplx has bad followup context management i think

#

idk

livid mantle May 29, 2024, 2:54 PM

#

devout geyser May 29, 2024, 3:00 PM

#

yeah the counter is now hidden until you get 'low'

#

in Opus's case it's a bit misleading as you don't get the same amount, you get nearer to 1/10th compared to other models such as Sonnet or GPT-4o

agile jay May 29, 2024, 3:02 PM

#

sleek vortex written in go

My scrapers are already written in Go, so likely won't see a large difference.

sleek vortex May 29, 2024, 3:11 PM

#

devout geyser in Opus's case it's a bit misleading as you don't get the same amount, you get n...

8.33%

agile jay May 29, 2024, 3:12 PM

#

Looks like the advantage of katana is being able to easily switch between http and headless requests.

hollow sorrel May 29, 2024, 3:24 PM

#

I just don't understand something:

I upload a PDF into Perplexity,
I ask a question or two about the contents of the pdf,
I create a Collection and in the "AI Prompt (optional)" section I indicate that I don't want Perplexity to use any outside sources when it responds to questions,
I come back the next day, open this collection and ask a question and it consults outside sources! I ask it in the prompt NOT to do this and it still does it!

What am I doing wrong? I appreciate any help someone can provide. 😬

sleek vortex May 29, 2024, 3:40 PM

#

turn off pro mode

#

if you have pplx pro then select a model in settings like gpt4o or opus

hollow sorrel May 29, 2024, 3:43 PM

#

I do have the Pro slider set to off....

#

And I am using GPT-4o: https://app.screencast.com/8BgvwUhLJLOYT

TechSmith Screencast

mikebritt

2024-05-29_11-44-22

World's leading screen capture + recorder from Snagit + Screencast by Techsmith. Capture, edit and share professional-quality content seamlessly.

austere kestrel May 29, 2024, 3:47 PM

#

it doesn't have any kind of persistent memory. each thread starts fresh. your previous messages and uploads don't carry over

sleek vortex May 29, 2024, 3:49 PM

#

upload carry over in follow up

#

i’ve had good success with that

austere kestrel May 29, 2024, 3:50 PM

#

yeah but i think it's for a collection. as in you need to reupload the files each time. The prompt for the Collection is the only part that is "stored"

#

though i may have misunderstood

hollow sorrel May 29, 2024, 3:50 PM

#

austere kestrel it doesn't have any kind of persistent memory. each thread starts fresh. your pr...

I didn't know that. I really thought it had this capability. Surprising.

hollow sorrel May 29, 2024, 3:51 PM

#

sleek vortex upload carry over in follow up

can you explain what you mean by "upload carry over in follow up"?

austere kestrel May 29, 2024, 3:51 PM

#

hollow sorrel I didn't know that. I really thought it had this capability. Surprising.

be wary of what these large language models will tell you!

#

they're very convincing, always obliging - but not always accurate ha

hollow sorrel May 29, 2024, 3:52 PM

#

The problem is that it told me that certain concepts were contained in the uploaded document, but when I searched for those terms in Adobe Acrobat it - correctly - told me that those terms were NOT mentioned in the pdf!

austere kestrel May 29, 2024, 3:52 PM

#

#⚡│ask-community message it all stems from the 'hallucination' proble with the technology

hollow sorrel May 29, 2024, 3:53 PM

#

I guess I should just use Adobe Acrobat's AI Chat feature.... Or is there another AI tool I should use when I only want to search withing a pdf?

#

I could use: https://pdf.ai/ but I just don't want to subscribe to yet another AI tool.

PDF.ai | Chat with your PDF documents

We built the ultimate ChatPDF app that allows you to chat with any PDF: ask questions, get summaries, find anything you need!

austere kestrel May 29, 2024, 3:54 PM

#

hollow sorrel I guess I should just use Adobe Acrobat's AI Chat feature.... Or is there anothe...

you might want to try using the "Writing" mode on Perplexity; it disables the web/online part of perplexity, which will help keep the llm focused on the doucment

gentle dirge May 29, 2024, 3:54 PM

#

yo guys whats the best way to scrape web for giving llm latest context? im doing brave search api + cheerio currently

austere kestrel May 29, 2024, 3:55 PM

#

but fwiw for what you were describing earlier - wanting to upload a repository of files - you might want to look into chatgpt's Custom gpts

hollow sorrel May 29, 2024, 3:55 PM

#

austere kestrel you might want to try using the "Writing" mode on Perplexity; it disables the we...

How do you turn on "writing mode"?

austere kestrel May 29, 2024, 3:55 PM

#

Press the 'focus' tab then select it from there

hollow sorrel May 29, 2024, 3:56 PM

#

austere kestrel Press the 'focus' tab then select it from there

Excellent idea. Thank you!

austere kestrel May 29, 2024, 4:01 PM

#

google's Notebook LM is also good for working with documents

tidal sail May 29, 2024, 4:09 PM

#

https://news.itsfoss.com/openai-google-search/

It's FOSS News

OpenAI Plans to Challenge Google With its AI Search Engine

Another ChatGPT-powered wave incoming with the new search engine?

tidal sail May 29, 2024, 4:10 PM

#

tidal sail https://news.itsfoss.com/openai-google-search/

Perplexity is going to compete with open ai lol

fading moth May 29, 2024, 5:41 PM

#

Well, they've got more funding, but there is such a thing as wearing too many hats

#

I'd rather have 4.5/5 fix 4/4o's weak points than stretch its capabilities to other things that might suffer from those weak points

mystic basin May 29, 2024, 5:45 PM

#

Hello!
Can anyone tell me what is the limit of Claude 3 Opus and when does attempts renew

fading moth May 29, 2024, 5:45 PM

#

Given that Google's main game is data and being synonymous with Internet searching, and how they have been struggling with Gemini that has 10x the context windows of GPT while having Google's 80% of all user data; plus, experience as the world's biggest search engine...

#

It shows that OpenAI would need to do a lot of work in areas that currently have better solutions, even if that search functionality was added.

#

It would be better for OpenAI to be innovative, improve their weak links, and make their foundation unshakable before trying to branch out into areas GPT wasn't built to handle optimally without fixing the weak spots.

#

Gemini was far less known compared to GPT to the general public and GPT still hallucinates without pulling fringe/troll search data from sites like reddit.

It would be a large blow to their credibility; just as Google's AI currently is losing credibility by suggesting people eat rocks or jump off bridges.

fading moth May 29, 2024, 5:54 PM

#

mystic basin Hello! Can anyone tell me what is the limit of Claude 3 Opus and when does attem...

50 and I can't remember, sorry.

sleek vortex May 29, 2024, 6:02 PM

#

fading moth Gemini was far less known compared to GPT to the general public and GPT still ha...

I have no clue how google made it such a joke honestly

#

I was able to make a better search summary than google with like a few days of work

#

@agile jay what do you think i should make the backend in

agile jay May 29, 2024, 6:04 PM

#

I would have probably made it with Go.

sleek vortex May 29, 2024, 6:05 PM

#

true but i know 0 go

agile jay May 29, 2024, 6:06 PM

#

sleek vortex true but i know 0 go

Luckily it's probably one of the easiest languages to learn...

sleek vortex May 29, 2024, 6:06 PM

#

true

agile jay May 29, 2024, 6:07 PM

#

And the simple binaries as output is nice.

warm cave May 29, 2024, 6:12 PM

#

agile jay And the simple binaries as output is nice.

Yeah I find that super nice

agile jay May 29, 2024, 6:13 PM

#

And the build times too.

sleek vortex May 29, 2024, 6:17 PM

#

hmm, okay

#

maybe i will learn go

near imp May 29, 2024, 6:29 PM

#

I just switched back to Android from iOS, and perplexity app seems to not support the same 'voice conversation' features on Android as on iOs? is that correct?

agile jay May 29, 2024, 6:29 PM

#

near imp I just switched back to Android from iOS, and perplexity app seems to not suppor...

Yep, android users are second class citizens on perplexity...

near imp May 29, 2024, 6:30 PM

#

.... wow.....

#

that is #sadge

agile jay May 29, 2024, 6:31 PM

#

Yep, that what happens when the devs are from SF...

near imp May 29, 2024, 6:31 PM

#

specially since Apple and iPhone are literally the worst platform for Ai and will see signifcant dropoff in the next few years, due to their lack thereof, which they won't be able to make up for anytime soon.

#

sad.

agile jay May 29, 2024, 6:31 PM

#

Yep, in the US it's like 70% apple users, and since SF is likely richer, their percentage is likely even higher.

fading moth May 29, 2024, 6:32 PM

#

Yeh. The GPT assistant is macOS only too.

near imp May 29, 2024, 6:32 PM

#

yeah, but that's OpenAI, but indeed, I was suprised considering they literally have microsoft as their top investor.

#

no idea why anyone wants to be inside the digital prison that is AppleVerse.

#

I tried an iPhone for 2 years, and yes it has some nice features, but it was a prison.

agile jay May 29, 2024, 6:33 PM

#

They like spending money to feel relevant.

fading moth May 29, 2024, 6:33 PM

#

It's also weird since Apple is so restrictive with what they approve of and disapprove of in hardware/software.

#

So much more red tape.

agile jay May 29, 2024, 6:33 PM

#

Yep, and no side loading.

#

But in the EU there is supposed to be sideloading this year.

fading moth May 29, 2024, 6:34 PM

#

EU is doing a lot to help everyone fight Apple's "isolationism" tech

#

In the US we wouldn't be able to hold any ground against Apple

agile jay May 29, 2024, 6:35 PM

#

Yep, otherwise you have to pay $99/year as a dev, just to sign your apps...

fading moth May 29, 2024, 6:36 PM

#

So, it makes no sense. OpenAI putting out an Apple only thing that I would love to try.

#

Given Microsoft backing.

#

The assistant was like my most interested thing during their announcement.

tidal sail May 29, 2024, 6:37 PM

#

I wonder how windows 12 will work with all this ai hallucination. Will that thing break their system?

fading moth May 29, 2024, 6:38 PM

#

Well, MS has said W11 is the "last" iteration and it will instead just be patched and improved as an ongoing product

tidal sail May 29, 2024, 6:38 PM

#

Chatgpt alone uses 6GB of my ram space currently lol

fading moth May 29, 2024, 6:38 PM

#

Chrome and Firefox eating up mah RAM too

tidal sail May 29, 2024, 6:39 PM

#

fading moth Well, MS has said W11 is the "last" iteration and it will instead just be patche...

MS says everything

fading moth May 29, 2024, 6:40 PM

#

Ye, that's why I said said :p

#

Instead of W11 is the last

#

The only people who know are above my paygrade.

sleek vortex May 29, 2024, 6:42 PM

#

near imp yeah, but that's OpenAI, but indeed, I was suprised considering they literally h...

"just buy a copilot+pc instead"

sleek vortex May 29, 2024, 6:42 PM

#

fading moth Well, MS has said W11 is the "last" iteration and it will instead just be patche...

they said that for 10

#

then the ceo changed

fading moth May 29, 2024, 6:42 PM

#

Copilot+ spooks me

tidal sail May 29, 2024, 6:43 PM

#

I remember once when microsoft said windows 10 would be the last windows version lol

fading moth May 29, 2024, 6:43 PM

#

And I would need to be an entire laptop to use it

sleek vortex May 29, 2024, 6:43 PM

#

recall is a cool rag idea until they start selling my data

#

why does it take screenshots

#

surely better way exists

fading moth May 29, 2024, 6:43 PM

#

Ye

warm cave May 29, 2024, 6:44 PM

#

Maybe the images is for the users benefit, and what is stored is the output from Phi 3 vision

fading moth May 29, 2024, 6:45 PM

#

GPT assist doesn't do that stuff like copilot+ ... The memory for the assistant may be less "accurate" in the long term but I really just need it to help with my current questions instead of asking if it remembers what I was doing three months ago.

#

Data is taken either way, but it's the amount that is so vastly different

tidal sail May 29, 2024, 6:46 PM

#

fading moth Chrome and Firefox eating up mah RAM too

If ai will be integrated in windows12, you would need atleast 32GB ram to operate

agile jay May 29, 2024, 6:46 PM

#

tidal sail If ai will be integrated in windows12, you would need atleast 32GB ram to operat...

Yep, the lowest you can get on a windows 12 device is 32GB

north magnet May 29, 2024, 6:46 PM

#

why is the api so bad it doesnt even seem like its online

fading moth May 29, 2024, 6:46 PM

#

Well, the new laptops have it integrated, I believe?

#

Like you can't turn it off

tame current May 29, 2024, 6:47 PM

#

north magnet why is the api so bad it doesnt even seem like its online

which model are you using?

north magnet May 29, 2024, 6:47 PM

#

the api isnt using the same model as the webchat clearly

tidal sail May 29, 2024, 6:47 PM

#

agile jay Yep, the lowest you can get on a windows 12 device is 32GB

Ooh my! That will cost alot

north magnet May 29, 2024, 6:47 PM

#

i want my money back

agile jay May 29, 2024, 6:47 PM

#

tidal sail Ooh my! That will cost alot

Not really, 32GB ram has been cheap for a long time.

#

Not Apple where adding 8GB more ram increases the price by a few hundred dollars...

north magnet May 29, 2024, 6:48 PM

#

tame current which model are you using?

LLaMa 3 i beleive

#

anyone else here using the api?

tame current May 29, 2024, 6:48 PM

#

the sonar one? with -online suffix?

#

it's unusable, i've built my own api with other model

#

i can provide you source code

north magnet May 29, 2024, 6:49 PM

#

its litreally unusable

tidal sail May 29, 2024, 6:49 PM

#

agile jay Not really, 32GB ram has been cheap for a long time.

Are you also on windows?

north magnet May 29, 2024, 6:49 PM

#

tame current i can provide you source code

im intrested

sleek vortex May 29, 2024, 6:49 PM

#

agile jay Not really, 32GB ram has been cheap for a long time.

unrelated but can i ask you a go question

fading moth May 29, 2024, 6:49 PM

#

32GB of ram is pretty standard for even prebuilt these days. The RAM speed might be crap, but it's still 32.

sleek vortex May 29, 2024, 6:49 PM

#

tame current i can provide you source code

do you run it on local gpu?

#

or cloud?

tame current May 29, 2024, 6:50 PM

#

i'm using it on openrouter currently, but i will later adjust it to be able to run locally

north magnet May 29, 2024, 6:50 PM

#

i just dont understand why it gives such bad respones then the web app? it doesnt say anywhere when signing up for the api that its using a unusable model

tame current May 29, 2024, 6:50 PM

#

it's work-in-progress but searches well and is better than perplexity api already

fading moth May 29, 2024, 6:50 PM

#

Make sure you have a spicy GPU if you want to run locally

agile jay May 29, 2024, 6:50 PM

#

sleek vortex unrelated but can i ask you a go question

Sure

tidal sail May 29, 2024, 6:50 PM

#

fading moth 32GB of ram is pretty standard for even prebuilt these days. The RAM speed might...

So that means, 5 years from now, people will go upto 128GB ram?

agile jay May 29, 2024, 6:51 PM

#

tidal sail Are you also on windows?

Yep, I use windows and linux.

north magnet May 29, 2024, 6:51 PM

#

tame current it's work-in-progress but searches well and is better than perplexity api alread...

does it give simmlar responeses to the web app?

#

the thing im confused about is i was under the impression perplexity was just a wrapper, so how come the api doesnt give the same respones

agile jay May 29, 2024, 6:52 PM

#

north magnet does it give simmlar responeses to the web app?

Nope, they are not compareable at all.

north magnet May 29, 2024, 6:52 PM

#

the api model doesnt even seem online...

fading moth May 29, 2024, 6:52 PM

#

tidal sail So that means, 5 years from now, people will go upto 128GB ram?

If we think back to the nineties and consider people thought using 1gb of data was impossible to fill up for a non-commercial user... CoD is 200gbs. What we think is excessive now just becomes standard later.

north magnet May 29, 2024, 6:52 PM

#

yeah but they dont state that anywhere @agile jay

#

i want my money back

#

@signal hamlet

mystic ivy May 29, 2024, 6:53 PM

#

Hey sorry if this is wrong channel but who can I contact to get higher api limits?

fading moth May 29, 2024, 6:53 PM

#

There is sort of a soft cap with speed at a universal level when we start dealing with quantum computation at a consumer level.

sleek vortex May 29, 2024, 6:53 PM

#

open router has phi3 for free??

#

wait what

agile jay May 29, 2024, 6:53 PM

#

mystic ivy Hey sorry if this is wrong channel but who can I contact to get higher api limit...

Maybe try in #🧪│api-general

sleek vortex May 29, 2024, 6:53 PM

#

i could try this instead of llama 8b for sources...hmm

fading moth May 29, 2024, 6:53 PM

#

mystic ivy Hey sorry if this is wrong channel but who can I contact to get higher api limit...

Nobody gets higher limits. :s

mystic ivy May 29, 2024, 6:53 PM

#

fading moth Nobody gets higher limits. :s

I know some that did :((

tame current May 29, 2024, 6:54 PM

#

north magnet the api model doesnt even seem online...

perplexity app uses different model than the one on api, and it's half-baked after some change

mystic ivy May 29, 2024, 6:54 PM

#

I want to be part of the happy few

fading moth May 29, 2024, 6:54 PM

#

mystic ivy I know some that did :((

Businesses undoubtedly

north magnet May 29, 2024, 6:54 PM

#

there is something shady going on here ngl

mystic ivy May 29, 2024, 6:54 PM

#

fading moth Businesses undoubtedly

It's for my business too, a 100k+ users

sleek vortex May 29, 2024, 6:54 PM

#

pplx api maybe only got released to look good

#

for vc

mystic ivy May 29, 2024, 6:54 PM

#

mystic ivy It's for my business too, a 100k+ users

actually 500k+

sleek vortex May 29, 2024, 6:54 PM

#

but they really might only have like 2 employees on it

#

idk

north magnet May 29, 2024, 6:54 PM

#

they make it seem like its using the same model

#

legit a scam

agile jay May 29, 2024, 6:55 PM

#

sleek vortex pplx api maybe only got released to look good

It only has one endpoint...

sleek vortex May 29, 2024, 6:55 PM

#

yeah its nothing like the pplx frontend

fading moth May 29, 2024, 6:55 PM

#

mystic ivy It's for my business too, a 100k+ users

Then you might be in luck. There is contact info on Perplexity's site somewhere about enterprise stuff.

sleek vortex May 29, 2024, 6:55 PM

#

at most it might be the same as what free search gets...?

north magnet May 29, 2024, 6:55 PM

#

im going to bring some friends here and dig into there business model

agile jay May 29, 2024, 6:55 PM

#

/chat/completions

north magnet May 29, 2024, 6:55 PM

#

something seems wrong

mystic ivy May 29, 2024, 6:55 PM

#

fading moth Then you might be in luck. There is contact info on Perplexity's site somewhere ...

alright will try to find it

agile jay May 29, 2024, 6:55 PM

#

north magnet something seems wrong

You can say that again

tame current May 29, 2024, 6:55 PM

#

north magnet im intrested

will send you in 20min on dm, i fked up something in code

north magnet May 29, 2024, 6:56 PM

#

i was under the impression perp was a wrapper for other models, yet they cant make that model api accsaible?

#

seems like they are stealing and using something they dont want us to find

sleek vortex May 29, 2024, 6:56 PM

#

yeah no pplx api is only for their own models

#

idk

north magnet May 29, 2024, 6:57 PM

#

so confusing man

sleek vortex May 29, 2024, 6:57 PM

#

i dont think they have the infra to actually host api scale other models

north magnet May 29, 2024, 6:57 PM

#

whats the best api you guys are using

sleek vortex May 29, 2024, 6:57 PM

#

i think they just have some good code and a few cloud gpus on it

sleek vortex May 29, 2024, 6:57 PM

#

north magnet whats the best api you guys are using

groq is good for small models

#

idk

#

about it really

north magnet May 29, 2024, 6:57 PM

#

really

sleek vortex May 29, 2024, 6:57 PM

#

well

#

it has llama 8b and 70b

north magnet May 29, 2024, 6:57 PM

#

i never used grog

sleek vortex May 29, 2024, 6:57 PM

#

fastest and cheapest out there

north magnet May 29, 2024, 6:57 PM

#

groq

sleek vortex May 29, 2024, 6:57 PM

#

they dont use gpu

tame current May 29, 2024, 6:57 PM

#

openrouter is the best, has the most models for cheap

agile jay May 29, 2024, 6:57 PM

#

Groq for speed and price.

sleek vortex May 29, 2024, 6:57 PM

#

they use their own type of chip (LPU)

sleek vortex May 29, 2024, 6:57 PM

#

tame current openrouter is the best, has the most models for cheap

Surge limit: By default, all users are subject to a maximum rate limit of 200 requests per second to defend against denial-of-service attacks. Contact us in Discord or using our support@ email address if you need a higher limit.

Free limit: If you are using a free model variant (with an ID ending in :free), then you will be limited to 20 requests per minute and 200 requests per day.

#

hmm

fading moth May 29, 2024, 6:57 PM

#

Oh. Speaking of Groq, what y'all think of the xAI thing musky was talking about?

agile jay May 29, 2024, 6:58 PM

#

Open router and vercel for other stuff.

agile jay May 29, 2024, 6:58 PM

#

fading moth Oh. Speaking of Groq, what y'all think of the xAI thing musky was talking about?

Different groq

sleek vortex May 29, 2024, 6:58 PM

#

thats grok right?

north magnet May 29, 2024, 6:58 PM

#

i need acruate update responses for my app thats why i was planning to use perp till i realised how dogshit the api is

agile jay May 29, 2024, 6:58 PM

#

There is Groq the infrastructure company, and Grok the x.ai model.

north magnet May 29, 2024, 6:58 PM

#

got it

sleek vortex May 29, 2024, 6:58 PM

#

north magnet i need acruate update responses for my app thats why i was planning to use perp ...

do you think this sort of api is in demand?

agile jay May 29, 2024, 6:59 PM

#

sleek vortex do you think this sort of api is in demand?

Always

fading moth May 29, 2024, 6:59 PM

#

agile jay There is Groq the infrastructure company, and Grok the x.ai model.

Then what was the big announcement I was reading about this week?

north magnet May 29, 2024, 6:59 PM

#

im guessing yes

agile jay May 29, 2024, 6:59 PM

#

SInce people always want upto date info

sleek vortex May 29, 2024, 6:59 PM

#

fading moth Then what was the big announcement I was reading about this week?

that was grok x.ai (twitter)

agile jay May 29, 2024, 6:59 PM

#

fading moth Then what was the big announcement I was reading about this week?

Probably they are rolling out the next version of Grok

fading moth May 29, 2024, 6:59 PM

#

OH

sleek vortex May 29, 2024, 6:59 PM

#

not groq LPU inference

fading moth May 29, 2024, 6:59 PM

#

Gotcha

north magnet May 29, 2024, 6:59 PM

#

perp was giving perfect repsones in the web app, im so heartbroken the api isnt the same

fading moth May 29, 2024, 7:00 PM

#

Well, ~~Groq~~ Grok was crap on Twitter, I don't see why anyone would want to run it locally

sleek vortex May 29, 2024, 7:00 PM

#

north magnet perp was giving perfect repsones in the web app, im so heartbroken the api isnt ...

ive been working on something similiar

#

nothing prod ready tho

#

idk how i would scale it to a full api

north magnet May 29, 2024, 7:00 PM

#

yeah man keep me updated 100%

sleek vortex May 29, 2024, 7:00 PM

#

fading moth Well, ~~Groq~~ Grok was crap on Twitter, I don't see why anyone would want to ru...

grok*

fading moth May 29, 2024, 7:00 PM

#

My bad

north magnet May 29, 2024, 7:00 PM

#

i have 1,000 of customers and would pay top dollar for this

#

s

sleek vortex May 29, 2024, 7:00 PM

#

right now i have working web search kinda model as decent as perplexity frontend

north magnet May 29, 2024, 7:01 PM

#

i must be missing something

#

how come we cant archive the same results as them if they aren't using there own model, i get it has harcoded prompts but how is it getting arcuate information via web search?. couldn't we jsut build the same thing?

#

or is that what you are doing @sleek vortex

sleek vortex May 29, 2024, 7:02 PM

#

yeah im doing that

#

doing it better (i think)

fading moth May 29, 2024, 7:03 PM

#

A secret sauce that perplexity has on their web app?

sleek vortex May 29, 2024, 7:03 PM

#

well ive built that in less than a week

north magnet May 29, 2024, 7:03 PM

#

something shady man

sleek vortex May 29, 2024, 7:03 PM

#

something equivilant to pplx web app

north magnet May 29, 2024, 7:03 PM

#

i dont think its using LLM

#

ngl

sleek vortex May 29, 2024, 7:03 PM

#

i can explain what i think pplx has

#

if you want me to

north magnet May 29, 2024, 7:03 PM

#

yes please

agile jay May 29, 2024, 7:03 PM

#

Yep, it's not hard to make a perplexity like app. The harder part is getting VC money.

sleek vortex May 29, 2024, 7:03 PM

#

yeah lmao

agile jay May 29, 2024, 7:03 PM

#

To scale to the moon.

north magnet May 29, 2024, 7:03 PM

#

i have funding

#

personal funding

#

i dont need vc's

agile jay May 29, 2024, 7:04 PM

#

So do I, but most of the competition will likely get nuked by openai when they release their search.

#

The path to AGI is full of dead startups...

fading moth May 29, 2024, 7:05 PM

#

Wasn't Gemini 1.5 Pro with the 1M window also supposed to be a nuke?

agile jay May 29, 2024, 7:05 PM

#

Who actually needs 1M context?

sleek vortex May 29, 2024, 7:05 PM

#

north magnet yes please

my theory is something like this:

they take the user's query
in copilot mode, they send this to a small finetuned llm which returns a bunch of searches to make on google/bing/their own indexer
in non-copilot mode, they just use keyword extraction or search your query as is on google/bing/their own indexer

they may have layers in the backend that summarises the sources or uses embeddings, but im not entirely sure - if they have their own indexer then they may be running this in the background but i really doubt pplx is doing this

they then take the top N results and fit as much as they can into the LLM's context and make a response

north magnet May 29, 2024, 7:06 PM

#

i have users without web seacrh, web search will only make the respones in my app 100x better which should = more growth

sleek vortex May 29, 2024, 7:06 PM

#

ive been trying to make something similiar that can also do multistep reasoning

north magnet May 29, 2024, 7:06 PM

#

maybe im missing something

agile jay May 29, 2024, 7:06 PM

#

The longer the context the slower the response, and the more entropy to the output.

north magnet May 29, 2024, 7:06 PM

#

sleek vortex my theory is something like this: - they take the user's query - in copilot mod...

this makes alot of sense

sleek vortex May 29, 2024, 7:06 PM

#

agile jay The longer the context the slower the response, and the more entropy to the outp...

well they need that 2.8million context to show their investors how you can upload a 2hr movie and ask it about a random frame

sleek vortex May 29, 2024, 7:06 PM

#

north magnet this makes alot of sense

do you want me to give you details of what ive been doing?

agile jay May 29, 2024, 7:08 PM

#

Yep, the more difficult part is making the model answer the way the user wants.

#

Which is what multi step reasoning is useful for.

sleek vortex May 29, 2024, 7:08 PM

#

yeah

#

but right now my issues/todos are

first moving the codebase out of python local
somehow scaling it
and then i need to maybe build my own indexer/cache layer backed on s3 or some form of cheap storage
and then frontend, ofc

north magnet May 29, 2024, 7:10 PM

#

sleek vortex do you want me to give you details of what ive been doing?

honestly its going to sound stupid but I'm just making a wrappers for different llm's that provide specifc responses for a targeted group and then proving them user friendly UI/UX to interact with and charging a subscription for the amount of tokens they use

#

perp skipped half of this for me

agile jay May 29, 2024, 7:10 PM

#

sleek vortex but right now my issues/todos are - first moving the codebase out of python loca...

Yep, I think the best options are to make the backend in Go.
To change the infra to use a producer/consumer model for easy scaling.
To pre-cache a lot of popular sites using katana.

sleek vortex May 29, 2024, 7:10 PM

#

i mean yeah half of ai is just making it acessible and useful to each user's own circumstances

north magnet May 29, 2024, 7:10 PM

#

providing up to date responses

#

but the api is useless

#

so i cant use it inside my apps

sleek vortex May 29, 2024, 7:10 PM

#

why did i get flagged bruh

#

who deleted my message

fading moth May 29, 2024, 7:11 PM

#

Try sending as txt

agile jay May 29, 2024, 7:11 PM

#

sleek vortex why did i get flagged bruh

No idea, but they can't hide it from me...

sleek vortex May 29, 2024, 7:11 PM

#

📎 example_query.txt

agile jay May 29, 2024, 7:11 PM

#

The ipad pro m4 pricing one

sleek vortex May 29, 2024, 7:11 PM

#

this the sort of thing ive been building so far

north magnet May 29, 2024, 7:11 PM

#

ooo

sleek vortex May 29, 2024, 7:12 PM

#

accurate up-to date information using multiple llms and custom searching pipeline

north magnet May 29, 2024, 7:12 PM

#

for e-com?

sleek vortex May 29, 2024, 7:12 PM

#

and it can get faster than 21 seconds

#

my parallemism needs a rewrite

agile jay May 29, 2024, 7:12 PM

#

Yep, python can be janky when trying to make it concurrent.

sleek vortex May 29, 2024, 7:12 PM

#

north magnet for e-com?

no just something like pplx

north magnet May 29, 2024, 7:12 PM

#

you should try and self fund this, i would be intrested

sleek vortex May 29, 2024, 7:12 PM

#

thats what i am sorta trying to do :d

#

i have no real money of my own so

agile jay May 29, 2024, 7:12 PM

#

You don't need much funding for it.

sleek vortex May 29, 2024, 7:12 PM

#

using what i can get

north magnet May 29, 2024, 7:12 PM

#

dm me details

sleek vortex May 29, 2024, 7:12 PM

#

for free

agile jay May 29, 2024, 7:13 PM

#

sleek vortex for free

Go fund me, lol

sleek vortex May 29, 2024, 7:13 PM

#

nobodys going onto gofundme for an ai project

#

if i built out the whole project and platform i could probably get users from just promoting it

agile jay May 29, 2024, 7:14 PM

#

Or start a patreon for people to support you if they want.

sleek vortex May 29, 2024, 7:14 PM

#

then make a consumer subscription

#

like pplx

#

cheaper since i dont have opus

#

have every model except opus

agile jay May 29, 2024, 7:14 PM

#

sleek vortex if i built out the whole project and platform i could probably get users from ju...

Yep, there's a lot you can improve on with the current model.

north magnet May 29, 2024, 7:14 PM

#

you never know man

#

gofund me might actually help xd

sleek vortex May 29, 2024, 7:14 PM

#

i plan to add code interpreter and other tools too idk

agile jay May 29, 2024, 7:14 PM

#

Such as making it more agentic, since you have steps and intent prediction.

sleek vortex May 29, 2024, 7:14 PM

#

agile jay Such as making it more agentic, since you have steps and intent prediction.

yeah

#

right now dont have dependencies really working tbh

agile jay May 29, 2024, 7:15 PM

#

Yep, since you don't have a clear schema for it.

north magnet May 29, 2024, 7:15 PM

#

are you working on it alone?

agile jay May 29, 2024, 7:15 PM

#

In my case, I just use gRPC to combine them, and have them as seperate services.

#

Yep, and by sharing progress so I and a few others can help with dev suggestions.

agile jay May 29, 2024, 7:16 PM

#

sleek vortex yeah

You could probably make more by just focusing on making an API for other devs to use, lol.

sleek vortex May 29, 2024, 7:16 PM

#

north magnet are you working on it alone?

yeah ive been working on it here in this server for like a week or two

agile jay May 29, 2024, 7:16 PM

#

And then increasing the margins.

sleek vortex May 29, 2024, 7:16 PM

#

from scratch

north magnet May 29, 2024, 7:16 PM

#

lmao the fact we are talking about building a competitor in there discord cracks me up

sleek vortex May 29, 2024, 7:16 PM

#

agile jay You could probably make more by just focusing on making an API for other devs to...

yeah

#

yeah lmao ive had that thought at the back of my mind

tame current May 29, 2024, 7:17 PM

#

north magnet lmao the fact we are talking about building a competitor in there discord cracks...

fixed the code, accept invitation

sleek vortex May 29, 2024, 7:17 PM

#

agile jay You could probably make more by just focusing on making an API for other devs to...

yeah like even the fire scraping forgot-the-name that somebody else sent

agile jay May 29, 2024, 7:17 PM

#

Fishy.ai

sleek vortex May 29, 2024, 7:17 PM

#

like thats what just a project turned into a nice credits/api

#

and then people use it

#

because convenience, right?

agile jay May 29, 2024, 7:17 PM

#

sleek vortex yeah like even the fire scraping forgot-the-name that somebody else sent

Yep, probably the way to go, for the dev facing side.

sleek vortex May 29, 2024, 7:18 PM

#

agile jay Fishy.ai

my 50 opus queries can come up with a better name

sleek vortex May 29, 2024, 7:18 PM

#

tame current fixed the code, accept invitation

?

north magnet May 29, 2024, 7:18 PM

#

i wonder if the vc's know how much money they are missing out on bc the api doesnt work

#

might have to let them know

sleek vortex May 29, 2024, 7:19 PM

#

perplexitys product is

#

the frontend facing product

agile jay May 29, 2024, 7:19 PM

#

But probably pplx also doesn't want an easy API, since then someone can easily just make a mirror site that just uses the API...

sleek vortex May 29, 2024, 7:19 PM

#

why do you think theyve got so much funding from telecoms

agile jay May 29, 2024, 7:19 PM

#

And make it cheaper than the subscription...

sleek vortex May 29, 2024, 7:19 PM

#

integration with korea telecom this that

#

its all a consumer focus

agile jay May 29, 2024, 7:19 PM

#

Yep, SKT and Softbank partnerships.

sleek vortex May 29, 2024, 7:20 PM

#

ill wait for the vodafone partnership so i dont have to pay for pplx pro...

#

¯_(ツ)_/¯

#

until then!

agile jay May 29, 2024, 7:21 PM

#

Yep, the question now is how to make citations a lot better in search.

sleek vortex May 29, 2024, 7:21 PM

#

citations...hm

agile jay May 29, 2024, 7:21 PM

#

Since currently just adding a super long list of source numbers is probably not the way to go...

meager sparrow May 29, 2024, 7:23 PM

#

sleek vortex May 29, 2024, 7:23 PM

#

we would need the mini models to push forward the used sources

meager sparrow May 29, 2024, 7:23 PM

#

What is going on here??

sleek vortex May 29, 2024, 7:23 PM

#

wha

#

this was ages ago

#

agile jay May 29, 2024, 7:23 PM

#

Just pages.

meager sparrow May 29, 2024, 7:23 PM

#

They are telling me “please react to the channel” with that and then they showed me I have to then access this channel

#

But I have no idea what channel it is

sleek vortex May 29, 2024, 7:23 PM

#

#🅰│web-alpha-feedback

#

meager sparrow May 29, 2024, 7:24 PM

#

sleek vortex this was ages ago

Ok my guy I’m Not god I don’t know everything neither am I up to date on everything…

#

It is not my fault

sleek vortex May 29, 2024, 7:24 PM

#

meager sparrow Ok my guy I’m Not god I don’t know everything neither am I up to date on everyth...

nono, i dont mean it rudely :d

#

but yeah its a new feature they were testing

agile jay May 29, 2024, 7:24 PM

#

meager sparrow Ok my guy I’m Not god I don’t know everything neither am I up to date on everyth...

Maybe... Ask perplexity?

sleek vortex May 29, 2024, 7:24 PM

#

it isnt the best - i think theres still a decent amount of issues with it

meager sparrow May 29, 2024, 7:24 PM

#

sleek vortex nono, i dont mean it rudely \:d

Did they initiate the feature? Could I access it?

sleek vortex May 29, 2024, 7:24 PM

#

but its not bad either

meager sparrow May 29, 2024, 7:24 PM

#

sleek vortex it isnt the best - i think theres still a decent amount of issues with it

Could I see for myself?

#

I love experiments

agile jay May 29, 2024, 7:25 PM

#

@sleek vortex katana is pretty good for crawling sites btw.

sleek vortex May 29, 2024, 7:25 PM

#

yeah it is

meager sparrow May 29, 2024, 7:25 PM

#

agile jay Maybe... Ask perplexity?

Many times they don’t respond to me at all

sleek vortex May 29, 2024, 7:26 PM

#

meager sparrow Could I see for myself?

i got access by doing some form ages ago

#

https://perplexity.typeform.com/pages-beta

Typeform

Perplexity Pages - Beta Access

Turn data collection into an experience with Typeform. Create beautiful online forms, surveys, quizzes, and so much more. Try it for FREE.

#

not sure if theyre still checking it though

meager sparrow May 29, 2024, 7:26 PM

#

sleek vortex i got access by doing some form ages ago

But it is not like a new update or a feature everyone can access

sleek vortex May 29, 2024, 7:26 PM

#

i think they were beta testing it

#

so they made it like gated

#

but then did they give up or something

#

as there hasnt been feedback for a few weeks

sleek vortex May 29, 2024, 7:27 PM

#

meager sparrow Could I see for myself?

if you want to see it right now, i could put in a query for you

sleek vortex May 29, 2024, 7:27 PM

#

agile jay <@957611986835898441> katana is pretty good for crawling sites btw.

yeah, but we'd have to roll our own parser ontop

#

what i have right now is pretty good

#

but yeah as you said id like to be able to deal with sites like apple too

agile jay May 29, 2024, 7:27 PM

#

sleek vortex yeah, but we'd have to roll our own parser ontop

It's also easily accessible in Go, since it's a go package...

#

agile jay May 29, 2024, 7:28 PM

#

sleek vortex but yeah as you said id like to be able to deal with sites like apple too

I was thinking of a technique to easily remove a lot of junk data.

meager sparrow May 29, 2024, 7:29 PM

#

sleek vortex if you want to see it right now, i could put in a query for you

Yes Sir

agile jay May 29, 2024, 7:29 PM

#

Basically compare all the pages of a site and remove the duplicate elements.

sleek vortex May 29, 2024, 7:29 PM

#

what on earth is this price

agile jay May 29, 2024, 7:30 PM

#

There are multiple layers in place.

meager sparrow May 29, 2024, 7:30 PM

#

@sleek vortex I clicked on the link you sent and it told me to insert an email and it said “we will send you a message shortly”

sleek vortex May 29, 2024, 7:30 PM

#

agile jay Basically compare *all* the pages of a site and remove the duplicate elements.

interesting

agile jay May 29, 2024, 7:30 PM

#

Which is probably why.

meager sparrow May 29, 2024, 7:30 PM

#

So alright

sleek vortex May 29, 2024, 7:30 PM

#

meager sparrow <@957611986835898441> I clicked on the link you sent and it told me to insert an...

Yeah i got accepted after like a day or two

agile jay May 29, 2024, 7:30 PM

#

sleek vortex interesting

I would imagine that would remove a huge amount of junk.

#

And work for nearly every site.

sleek vortex May 29, 2024, 7:30 PM

#

yeah then we could combine with convert to markdown or something

#

markdown is quite good because it preserves title weights/significance from articles

agile jay May 29, 2024, 7:31 PM

#

Yep, makes life so much easier.

sleek vortex May 29, 2024, 7:31 PM

#

where to start with go

agile jay May 29, 2024, 7:31 PM

#

sleek vortex markdown is quite good because it preserves title weights/significance from arti...

I use markdown a lot, so i know why it is good lol...

sleek vortex May 29, 2024, 7:31 PM

#

oh i was asking you a q before

#

this probably is basics but

#

agile jay May 29, 2024, 7:31 PM

#

Yep, you didn't say it though.

#

Go by example?

sleek vortex May 29, 2024, 7:31 PM

#

yeah

#

is s an array, or a slice referencing the array?

agile jay May 29, 2024, 7:32 PM

#

Basically you can think of slices as the default arrays. Since it's pretty uncommon to use an array, which has a fixed length.

#

But they are using a slice in that one, since they didn't specify the length of it.

sleek vortex May 29, 2024, 7:33 PM

#

But howcome when they do

#

s = s[:0]

meager sparrow May 29, 2024, 7:33 PM

#

@sleek vortex what can it do anyways???

#

I hate is it even good at doing ?

#

Pages

sleek vortex May 29, 2024, 7:33 PM

#

sleek vortex `s = s[:0]`

right after they extend it?

meager sparrow May 29, 2024, 7:34 PM

#

I have experimented with various Language models and some Ai powered search engines #

sleek vortex May 29, 2024, 7:34 PM

#

so is the slice like undelying reference to the array
which is why its able to be expanded all of a sudden and re-adapt the elements

agile jay May 29, 2024, 7:34 PM

#

sleek vortex so is the slice like undelying reference to the array which is why its able to b...

More or less. and the [:0] is just a slice, which python also has.

sleek vortex May 29, 2024, 7:34 PM

#

meager sparrow I have experimented with various Language models and some Ai powered search engi...

basically it asks your query to some small llm which splits it into 3 titles, then its the same as going and asking perplexity free to answer each of those title queries

#

its really not the best imo

meager sparrow May 29, 2024, 7:34 PM

#

So I already have a lot of knowledge and understanding in these models

sleek vortex May 29, 2024, 7:35 PM

#

agile jay More or less. and the [:0] is just a slice, which python also has.

hm, okay

agile jay May 29, 2024, 7:35 PM

#

Slicing is pretty useful, do you not use it when limiting the model input?

sleek vortex May 29, 2024, 7:35 PM

#

no i do but

#

go is like a bit different in how you can re-expand the slice

#

thats my main question

meager sparrow May 29, 2024, 7:36 PM

#

sleek vortex basically it asks your query to some small llm which splits it into 3 titles, th...

Huh? So you ask a question and then it answers them?

#

Not sure what you mean

sleek vortex May 29, 2024, 7:36 PM

#

they do s = s[:0] but then it's re-expanded with the same s?

sleek vortex May 29, 2024, 7:36 PM

#

meager sparrow Huh? So you ask a question and then it answers them?

let me screen rec a demo

agile jay May 29, 2024, 7:36 PM

#

sleek vortex go is like a bit different in how you can re-expand the slice

Oh, yes, you can change the length of the slice, but you normally don't;

meager sparrow May 29, 2024, 7:36 PM

#

sleek vortex let me screen rec a demo

K

sleek vortex May 29, 2024, 7:36 PM

#

agile jay Oh, yes, you can change the length of the slice, but you normally don't;

yeah like in python you dont do that right

#

once youve sliced it you lose the rest

#

same with like js

agile jay May 29, 2024, 7:37 PM

#

sleek vortex yeah like in python you dont do that right

It does do it, but under the hood.

#

some_list = [1, 2, 3, 4, 5]

some_list = some_list[:0] # gets rid of all items in the list
some_list[25] = 25 # now i've added it to the 25th index, even though there are no values between 0 and 24

#

So it's something you can do, but you rarely ever see it in code.

meager sparrow May 29, 2024, 7:39 PM

#

@sleek vortex you said you were going to screen record a demo

#

💀

agile jay May 29, 2024, 7:39 PM

#

He's busy learning some Golang

#

To become a Gopher/Goblin

meager sparrow May 29, 2024, 7:41 PM

#

Ok..

sleek vortex May 29, 2024, 7:41 PM

#

compressing it

#

im not lying

#

macos screen recorder outputs huge files

#

so im running ffmpeg on it (slowly)

#

frame= 3033 fps= 73 q=31.0 size= 8960kB time=00:00:50.51 bitrate=1453.0kbits/s dup=27 drop=0 speed=1.21x

#

nearly done

#

slightly long but yeah there you go!

agile jay May 29, 2024, 7:44 PM

#

Yep, ffmpeg can take a while, if you're doing CPU encoding.

meager sparrow May 29, 2024, 7:47 PM

#

sleek vortex

It looks really professional

agile jay May 29, 2024, 7:47 PM

#

But can you be sure it didn't hallucinate?

meager sparrow May 29, 2024, 7:47 PM

#

Looks like something you can use to get 100% on a whole assignment

agile jay May 29, 2024, 7:48 PM

#

meager sparrow Looks like something you can use to get 100% on a whole assignment

Not really. It's more like a overview of a topic, rather than an answer to an essay.

meager sparrow May 29, 2024, 7:48 PM

#

agile jay But can you be sure it didn't hallucinate?

Well if you verify the information that it is giving you by means of sources, then they is a very low chance it hallucinated

agile jay May 29, 2024, 7:48 PM

#

meager sparrow Well if you verify the information that it is giving you by means of sources, th...

You would be surprised...

north magnet May 29, 2024, 7:55 PM

#

how do you have the exact same UI as perp?

#

did you rebuild it

#

or get access to the source code

agile jay May 29, 2024, 7:55 PM

#

north magnet how do you have the exact same UI as perp?

pages is perp

north magnet May 29, 2024, 7:56 PM

#

im confused

#

can someone explain please

halcyon coral May 29, 2024, 8:08 PM

#

north magnet im confused

The "Pages" feature is currently in closed beta.

sleek vortex May 29, 2024, 8:16 PM

#

north magnet im confused

check dm

half venture May 29, 2024, 8:32 PM

#

#

Nvidia single handedly is doing the heavy lifting for the entire US economy at this point

sleek vortex May 29, 2024, 8:34 PM

#

2.82T now

#

going to surpass apple so soon wth

#

apple is 2.92T

#

what the hell

half venture May 29, 2024, 8:34 PM

#

Yeah

#

Went from the 94th position

#

To 2nd

#

Within a year

#

When this bubble pops

#

Oh boi

sleek vortex May 29, 2024, 8:35 PM

#

i mean

#

will it pop

#

i didnt see any insane company growth like this in failed hypetrains like crypto/web3

half venture May 29, 2024, 8:36 PM

#

It definitely will unless

#

Openai comes out

sleek vortex May 29, 2024, 8:36 PM

#

gpt5

half venture May 29, 2024, 8:36 PM

#

Yeah maybe that

#

But more like

#

Massive rollout worldwide free gpt 4 voice for the normies

sleek vortex May 29, 2024, 8:36 PM

#

yeah

#

idk

half venture May 29, 2024, 8:37 PM

#

I don't think you understand but

#

Most people only use chatgpt 3.5

#

And that's it

sleek vortex May 29, 2024, 8:37 PM

#

did they go up after copilot+pcs

sleek vortex May 29, 2024, 8:37 PM

#

half venture Most people only use chatgpt 3.5

yeah

#

most people use that

#

some might be touching google gemini and others

half venture May 29, 2024, 8:37 PM

#

And now gpt 4o is free including vision browsing etc as a small limit

sleek vortex May 29, 2024, 8:37 PM

#

but id assume the paid model population is really low in actual consumer adoption

agile jay May 29, 2024, 8:38 PM

#

Yep, but now they all have access to 4o with all its features.

half venture May 29, 2024, 8:38 PM

#

Definitely going to hype the markets

sleek vortex May 29, 2024, 8:38 PM

#

the average consumer, doesnt know what they could do, or at least thats what i think

agile jay May 29, 2024, 8:38 PM

#

But it also means that the next model should come out soon, for the plus users.

#

Otherwise, what's the point.

half venture May 29, 2024, 8:38 PM

#

Now the revolutionary thing will be if.....

#

They roll out voice mode

#

For free

#

As well

agile jay May 29, 2024, 8:38 PM

#

Yep

half venture May 29, 2024, 8:39 PM

#

If some boomer in middle of nowhere

#

Gets to use the voice

#

I bet he will.take this ai stuff more seriously

agile jay May 29, 2024, 8:39 PM

#

Yep, all those retired boomers with no social life will likely use it a lot.

sleek vortex May 29, 2024, 8:39 PM

#

i wonder how google would change

#

if they released project astra tommorow

#

but after their already bad situation...

#

no clue

agile jay May 29, 2024, 8:40 PM

#

Guess AGI for president will be more realistic since it will have the retired voters votes.

half venture May 29, 2024, 8:40 PM

#

Yeah definitely

agile jay May 29, 2024, 8:40 PM

#

sleek vortex i wonder how google would change

Google has no quality control...

#

And by far the most AI devs

half venture May 29, 2024, 8:41 PM

#

Didn't google raise the price for gemini flash

sleek vortex May 29, 2024, 8:41 PM

#

bruh their teams literally invented transformers

half venture May 29, 2024, 8:41 PM

#

Right after bragging that it's cheap

agile jay May 29, 2024, 8:41 PM

#

half venture Didn't google raise the price for gemini flash

Yep, doubled more or less

sleek vortex May 29, 2024, 8:41 PM

#

they have their own insane tpus

half venture May 29, 2024, 8:41 PM

#

agile jay Yep, doubled more or less

Ugh 💀

sleek vortex May 29, 2024, 8:41 PM

#

i honestly dont know why they arent first...

#

so stupid

#

they have the whole internet

#

they have all the compute ever

#

what are they missing?????

agile jay May 29, 2024, 8:41 PM

#

Because they are bad at making new products.

half venture May 29, 2024, 8:41 PM

#

sleek vortex i honestly dont know why they arent first...

You know you could ask the same about American government

#

Or any organization

#

The answer is

agile jay May 29, 2024, 8:42 PM

#

They are only good at going into a current field and improving it.

#

Can't think of a field made by google.

half venture May 29, 2024, 8:42 PM

#

The beuracratically stuck in a limbo

sleek vortex May 29, 2024, 8:42 PM

#

agile jay Can't think of a field made by google.

true

#

they didnt invent search

#

but they won in the end

#

or they have at least

#

the future, maybe not

half venture May 29, 2024, 8:42 PM

#

They didn't invent YouTube

sleek vortex May 29, 2024, 8:42 PM

#

yeah

#

brought it

half venture May 29, 2024, 8:43 PM

#

Actually I am more pissed off at Google for that

#

They have fuc k i ng YouTube

agile jay May 29, 2024, 8:43 PM

#

Yep, it's probably the reason why they have such a large graveyard compared to other companies.

half venture May 29, 2024, 8:43 PM

#

Make something better than Sora

#

Like come on

sleek vortex May 29, 2024, 8:43 PM

#

yeah...

#

they have the whole of the internet on google images

#

why is their latest model not dalle 9 level

half venture May 29, 2024, 8:43 PM

#

They are just being a wuss

agile jay May 29, 2024, 8:43 PM

#

Because they have too many devs.

sleek vortex May 29, 2024, 8:43 PM

#

why is SGE powered by gemma 0.1b

#

like ...

half venture May 29, 2024, 8:44 PM

#

sleek vortex why is SGE powered by gemma 0.1b

LMAO

sleek vortex May 29, 2024, 8:44 PM

#

no wonder youre being flamed about eating rocks

agile jay May 29, 2024, 8:44 PM

#

Doesn't matter if you have the most compute, if it's shared with a large dev team.

sleek vortex May 29, 2024, 8:44 PM

#

just throw some godamn compute at it

#

then finetune a model later

#

like bruh

half venture May 29, 2024, 8:44 PM

#

The issue is

#

Search is so profitable for.them

#

And so cheap

#

They want genai to be just as cheap

#

But it's not

sleek vortex May 29, 2024, 8:44 PM

#

bruh then finetune a model that incorporates ads

#

genai expensive entry

agile jay May 29, 2024, 8:44 PM

#

Maybe they are making AI summary sh*t on purpose.

sleek vortex May 29, 2024, 8:44 PM

#

but then make it cheap

#

then what

agile jay May 29, 2024, 8:44 PM

#

To make people less likely to use it.

sleek vortex May 29, 2024, 8:44 PM

#

release ai summary 2.0

half venture May 29, 2024, 8:45 PM

#

LLMS even a 1 billion parameter when deployed at a scale of billions actually is very expensive

sleek vortex May 29, 2024, 8:45 PM

#

agile jay To make people less likely to use it.

but then why roll it out to the whole world...

half venture May 29, 2024, 8:45 PM

#

Google is in scale.of billions

#

Microsoft in millions

#

That's the difference

agile jay May 29, 2024, 8:45 PM

#

sleek vortex but then why roll it out to the whole world...

To make people not trust it, before a good version even comes out.

half venture May 29, 2024, 8:46 PM

#

Also it's been almost 3 weeks

#

No voice

#

And they are treating chatgpt.free users better

agile jay May 29, 2024, 8:46 PM

#

Yep, likely because of sky drama

half venture May 29, 2024, 8:46 PM

#

And not even mentioned us plus opens