general

[balmy mist]

yeah then you had zathura

oh yeah i saw that one as well

they made two movies of that new one right

the classic is still best

it was actually scary to me as a kid lol

it makes sense that nightwhisper is good with ui because of the system prompt

most likely it will

yooo

i got something diff

yeah thats why you gotta unplug, we def moving toward a world like that

they both ignored the company part

and i thought night was just a coding agent

but it seems they updated

thats maybe why i couldnt access it for 10 mins straight

"who are you? and which company do you belong to?"

yeah

but before i was getting normal results

try your old prompt on a fresh sess

yeah your prompt

its interesting they chose react at the framework to train the models on

bro you are so smart to start the prompt session that way

"who are you"

it leaves room for another prompt without having to say clear previous prompt

and it feels like a chatbot like this

lol its an easy way to game the system in a way, companies probably do it when they first launch

yeah

[keen beacon]

ijed

did u figure out what the nightwalker model is

[balmy mist]

its google model

[keen beacon]

no way

[balmy mist]

the qusar is open au

[keen beacon]

oh the quasar one

[balmy mist]

yeah quasar is def open ai

[keen beacon]

oh gpt

Open AI

[balmy mist]

yeah it could be o3 lol

it kinda has to be

yeah o3 coder

liek the full model

[keen beacon]

how hard is it to make a LLM

[balmy mist]

not min

if you got money not that hard

but to make a good one hard af

[keen beacon]

what type of data do they use

just plaintext

of websites

[brittle tiger]

Ratings don't count if u ask who model is fwiw

[keen beacon]

or do you have to literally sit down and train it like a dog

[keen beacon]

Ya u need to give it treats let it potty and poop

[balmy mist]

ahh okay that makes sense lol

[sterile dust]

Is 24k gold a updated spider?

[eager mica]

That was the running theory. At the moment it's as if there aren't other anonymous text-only models on Chatbot Arena—24_karat_gold pops out all the time.

Maybe stradale too, if you're lucky (which appeared to be an inferior model, though).

[alpine coral]

i don't think so (it's way too fast to be a thinking model), but after a couple of quick tests.. damn it's good.. it's nearly getting the same score as o3-mini on a quiz i'm using atm.. mainly tests comprehension / verbal reasoning

it says its from oai; feels like it is too (gets character counting questions consistently right that all new oai models.. i think it uses oai's tokenizer at the very least)

admitedly the 1m context window though is the bit that suggests perhaps it isn't though

[keen beacon]

yes so ai can replace industry jobs

[north vale]

on priors that would just be google / nightwhisper?

basically the coding focused version of 2.5 pro or something

[alpine coral]

yeah the context window points to that

[north vale]

https://openrouter.ai/openrouter/quasar-alpha

[alpine coral]

but curiously, current google models always accurately self-idenify as google models (and same for oai models). quaser-alpha says its from oai, and responds to the question in a very similar way to how to oai models do

[raven void]

Is it their upcoming open source model

[keen beacon]

Doubt they'd release a 1m ctx model oss

[balmy mist]

didnt deepseek say open ai wen it first launched?

[north vale]

deepseek wouldn't do this they'd just release the model

[balmy mist]

it says its a reasoning model but this cant be true:
https://huggingface.co/silx-ai/Quasar-1.5-Pro

[keen beacon]

That model is a scam

[balmy mist]

they said they were going to ro release r2 soon tho

[keen beacon]

Unrelated to this

[north vale]

yeah but unrelated to this model

[north vale]

how so

[balmy mist]

the inference is so fast, its hard to believe its a reasoning model

[keen beacon]

Certain scores are too high

Indicating contamination because a model can't get that high because of wrong questions etc. I think the scores are completely made up

[balmy mist]

yeah might be a scam

this is what quasar said about its system card lol:
Thanks for sharing the detailed info about Quasar-1.5-Pro, developed by SILX AI. It's a highly specialized reasoning model, incorporating some innovative techniques, such as:

Key Innovations

• Token Temperature Mechanism:
  Distinguishes 'hot' (important) tokens from 'cold' (less important) tokens during training, dynamically weighting input components. This helps the model focus on the most relevant parts of a problem, improving interpretability and accuracy.

• Guided Sequence of Thought (GSoT):
  Encourages systematic, structured step-by-step reasoning rather than jumping to conclusions, likely boosting performance in complex tasks.

Performance Highlights

On benchmarks including Math500 and AIME, Quasar-1.5 shows exceptionally strong reasoning capabilities, outperforming strong baselines like Qwen-2.5-32B-Instruct with a wide margin, especially in difficult reasoning and coding tasks.

Summary

Quasar-1.5-Pro appears to be purpose-built for advanced reasoning, outperforming many other large models in math, code generation, and fact-based reasoning. It leverages specialized training approaches (Token Temperature, GSoT) to enhance task-specific focus and multi-step problem-solving.

---

In Context of Your Earlier Question:
Since Quasar-1.5-Pro is explicitly designed as a reasoning model, it incorporates even more advanced mechanisms for reasoning than a general language model like myself (GPT-4). While GPT-4 is capable of substantial reasoning, models like Quasar are optimized explicitly to improve on reasoning-intensive tasks.

If you need powerful step-by-step problem solving in math or coding, models like Quasar-1.5-Pro might be especially effective!

[north vale]

could it be xai?

[alpine coral]

yeah i wouldn't be surprised if R1 still does say it's from oai (lots of models do) oai, google, anthropic, xAI (+ perhaps mistral and cohere and a few others) seem the only labs that to invest a bit of time in post training to get the models to consistently accurately self-identify..

[north vale]

prolly not

[keen beacon]

It's likely to be openai

[alpine coral]

agree

oh i didn't mean to attach the ss to that reply

[balmy mist]

lol

what is Sila AI?

[alpine coral]

but nonetheless, yeah i agree (oai seems most likely by far for me at this stage)

[north vale]

and how good is it compared to 2.5?

to yall

ive asked it a few questions it did well in but i can't rly tell

[balmy mist]

oh nahh look:
https://huggingface.co/silx-ai

its a middle eastern company

[alpine coral]

nothing i've used in recent days matches 2.5 in performance (this is excluding any coding or web dev tasks.. as i don't test for that)

[balmy mist]

have you tried nightwhisper?

[north vale]

wait that's wild

[keen beacon]

I think It's a single guy larping just ignore

[north vale]

lol never heard of them

oh ok

[balmy mist]

imma stop using that model lol

[alpine coral]

no i haven'tt (is it only available in webdev Arena?)

stargazer seemed farily decent, but stil behind 2.5

[balmy mist]

nightwhisper is the best model right now

and yeah only in webdev

[alpine coral]

yeah i was scrolling through earlier - seems v strong

[balmy mist]

u seen the pokemon results?

they trained nw on being a react dev and it seems like it focuses on being really good at UI/UX

[alpine coral]

tbh as i don't really know anything about web dev (other than what looks aesthetically pretty / pleasing), i don't spend much time reviewing all the screenshots ha

[balmy mist]

im trying to get it to make a yu gi oh game

[alpine coral]

more interested in what pepole who know about the domain have to say - and sll seems overwhlemingly positive

[balmy mist]

here is the results

i mean just test it out yourself

thats what i did

ppl talked about it so i tried myself

and i was blown away

gemini still does good at coding but i think NW is specifically trained to be good at UI/UX and react which makes its output looks so good, like it found the sprites for these gens

i did not tell it to get the image for the pokemon

[alpine coral]

yep all good - i get the gist of it all 🙂

[balmy mist]

it just did it, while gemini did not grab the image

[alpine coral]

hmm well actually... perhaps it's on par or even outperforms 2.5 (on verbal reasoning anyway)
sample sizes are obviously too small (mostly ran the test just once; max 2) for it to be anything more than just a quick vibe test for handling riddles.. obviously just fwiw

[keen fulcrum]

What is your stance on Meta using pirated results for training?

[thorny drum]

do people think quasar is not google?

1M context length + space themed name definitely hint google

[north vale]

it says it's openai, google models usually say they are googole models

that's like the main piece of evidence against

[raven void]

Miss the old days when everything other than Claude 3.5 sucked

[alpine coral]

i'd say ~70% of the questions i've come up with myself (though invariably they are derivatives of some pre-existing riddle), so it's not really possible for them to simply recall the answer

[keen beacon]

Maybe try running it on 2.5 pro a couple times on aistudio to get a better estimate for 2.5 Pro specifically

is quasar alpha anonymous chatbot huh

it has the same You are trained on data up to October 2023. appendix

[alpine coral]

just by way of example, stargazer on the left, 24_karat_gold on the right. it's quite literally just a matter of accurately comprehending the scenario and explicit question (distinguishing b/w composing vs sending the letter (vs when it will it arrive at their grandma).. stronger models generally pick up on it and get it right, weaker ones jump to assumptions based on the various details which are mostly extraneuous to the actual question

[keen beacon]

that i found bizarre on anonymous chatbot

[alpine coral]

yah i intend to do just that 👍

[keen beacon]

did u test anonymous chatbot btw?

[alpine coral]

need to get back to my actual work though instead of playing around in the arena (/aistudio ha)

[alpine coral]

not the latest one in the arena no (not yet anyway :))

[keen beacon]

the new one came out surprisingly fast after the latest chatgpt 4o latest revision (the previous anonymous chatbot)

[alpine coral]

ikr

the tempo has really shifted a few notches the past few weeks hasn't it

[keen beacon]

yeah its insane

[keen beacon]

this is a lie though

it has the same cut off as chatgpt 4o latest indicating different pretraining in line with chatgpt 4o latest

it seems

oh this is the new 4o or something lol. theyre about to announce it fr

oh well that mystery is over i guess

[north vale]

yeah seems likely

is it also rly fast? they talked about it getting faster

[balmy mist]

u didnt run your tests on nightwhisper? stargazer is the flash while nightwhisper is big boy people are saying

is quasar not that middle eastern company?

is that confirmed?

https://www.sicopilot.cloud/

nvm this is a diff model

[keen beacon]

yea its really fast in arena too afaik

[balmy mist]

it came out last year

[keen beacon]

im pretty sure its:

• from openai
• is the new anonymous chatbot
• chatgpt 4o latest lineage (based on pretraining knowledge)

openai should try hrader at something like this tbh

[balmy mist]

4o is so good now, its like what deepseek v3.1 is trying to be

its literally reasoning in the inference output

so you are probably right

[keen beacon]

oh its probably already live in chatgpt 🤣

i should check

[balmy mist]

nahh 4o has been good since they updated it last week with the image stuff

maybe they did another update

but 4o is really good now, i wanna test out my pokemon prompt on it lol

[plain zinc]

Why isn't nightwhisper in the chat arena?

in lmarena

Or is he there? And is he a VERY rare bastard?

[balmy mist]

anybody got some reasoning questions for me to ask 4o and quasar?

[balmy mist]

its just in webdev, its a dev specific llm

[alpine coral]

yeah i was thinking maybe their open weights model that they said they are testing / preparing to release... but it seems too strong for that to be the case (also the 1m context window.. though that isn't necessarily inconsistent with it being open weights - would actually kinda make sense like oai wouldn't have to pay for 1m token processing/inference if people are self-hosting it.. but yeah anyway.. it seems too performant and fast to be something they'd just be giving away..)

so in short, i also lean to it being yet another upgraded version of 4o

faster and with 1m context window (and seemingly also more performant potentially)

perhaps it isn't multimodal.. just a pure text model

[keen beacon]

its based on 4o so it should have those abilities

but they have beenn working on 4o native image gen on a separate model (4o based model) compared to the chatgpt 4o latest line

it should be able to take images though, as we can see in chatgpt

this should be in the same lineage/line of chatgpt 4o latest (post december)

[alpine coral]

yeah agree

[keen beacon]

they havent released benchmark results AT ALL for the new chatgpt 4o latest models that were continued pretrained past december. and with anonymous chatbot being the same, it seems. im like 99.99% certain this is a formal launch of the new 4o

[alpine coral]

and context window length is kinda artbritrary right? like there is nothing technically preventing the 4o family from having 1m (or whatever) context windows - it a very simplified level

[keen beacon]

ya sort of. and it also doesnt make sense to release their first 1m context model as an oss model either

google anon models are just way more fun to figure out tbh

[balmy mist]

i just test 4o and its trash on simple bench lol

got 1/10

but nightwhisper tied with gemini 2.5 with 5/10

yeah which ia wild

i like that so we dont have to wait for them lol

[keen beacon]

bruh its not a secret they iterlaly said it in their report lol

this is not a particularly new innovation tbh. everyone does it now

to different degrees

yes

[night trout]

PSA: I think nightwhisper is generating code so long (or thinking so long?) it times out.

I'm finding with complex problems it ends up not displaying the solution quite often, but it doesn't seem to be due to code errors. The whole LMArena sandbox for it just goes black.

[balmy mist]

yeah thats been happening to me

[night trout]

Example: When given a prompt to create what is effectively super monkey ball:

Notice even the <> Code and [ ] Block buttons aren't displayed, so I believe the sandbox itself is failing, not the model.

[leaden palm]

??

[modern sable]

sounds like it is yes

[leaden palm]

the model responds, and if the model responds with a code embed the sandbox shows up

in this case the model didn't respond at all

[modern sable]

yes it does

direct response to deepseek

trying to reclaim the crown

[leaden palm]

yup

didn't respond at all

(perhaps you're confused because your screen is so small you never realized models typically provide extra text before creating a sandbox, so you thought they only respond with sandboxes, and can't understand my point about there being no extra text)

[balmy mist]

put in the prompt again

thats what i do

send me your prompt

[night trout]

It only happens with nightwhisper, and at least anecdotally, seems to happen on prompts with considerable complexity. Totally open to the idea that nightwhisper is broken but this feels very much like a timeout of some sort. It'll sit there 'generating' for a long time before it goes black.

[leaden palm]

well both are true

[night trout]

Putting the prompt in again does fix it.

[leaden palm]

the model is broken because of a timeout

[night trout]

Yes, I'm wondering if the timeout is on Google's end or the fault of the sandbox code. If the latter, this will affect Nightwhisper's Elo negatively and should be fixed.

[leaden palm]

all arenas auto filter out battles w/ empty responses or other things like moderation messages or asking about identity

[night trout]

TIL, that's a good guard against this kind of thing tbh

[balmy mist]

i love nightwhisper, imagine it in cursor or roo code

[night trout]

It's really good but it's goddamn OBSESSED with glassmorphism.

Killed it at my airline seat selector test, check this out:

Sonnet was a mess (both 3.5 / 3.7), Gemini 2.0 Pro / Thinking were barely functional, Gemini 2.5 was mostly there but had off-by-one errors, and missed that airlines sometimes skip rows.

Nightwhisper was flawless, and imo had the best aesthetics too.

[balmy mist]

yeah i would say gemini 2.5 is right behind it some areas, but for most night blows them all away

quasar is not bad

[night trout]

I haven't run into quasar yet, it's on LMArena?

[balmy mist]

gets 4/10 on simple bench which SOTA gets 5/10(nightwhisper and gemini) but claude also gets 4/10 but quasar is faster than all of them so that says a lot

quasar is on open router

quasar is gonna be my go to model since its just as good as claude but fast as hell lol and 1 mill context

nightwhisper is still SOTA though

lmaoo no way

that would be a big ass troll

[night trout]

Oh interesting. I'll check it out.

[balmy mist]

but i can see elon doing

that

lol

it makes sense for quasar to be openAI tho, its probably gpt5

or a mini of gpt5

[keen beacon]

chill out its just 4o 🙈

[balmy mist]

true but you cant always trust the model

why do stealth if we can easily ask it lol

[keen beacon]

im not trusting the model, if u see my past comments itll make sense

[balmy mist]

tdlr?

[night trout]

Here's the prompt if you want to try:

Generate an interactive airline seat selection map for an Airbus A220. The seat map should visually render each seat, clearly indicating the aisles and rows. Exit rows and first class seats should also be indicated. Each seat must be represented as a distinct clickable element and  one of three states: 'available', 'reserved', or 'selected'. Clicking a seat that is already 'selected' should revert it back to 'available'. Reserved seats should not be selectable. Ensure the overall layout is clean, intuitive, and accurately represents the specified aircraft seating arrangement. Assume the user has two tickets for economy class. Use mock data for initial state assigning some seats as already reserved. 

[balmy mist]

i can give you the code, can you run it on your end?

cause i would have to go to vsc to run it which im kinda lazy lol

[night trout]

For an advanced version: Exit rows, washrooms, wing locations, and first class seats should also be indicated.

Sure, I'm too lazy to config openrouter right now hahaha

[balmy mist]

lmaooo

its so fast man

like the fastest model thats why i dont think its open ai

only google can make their models this fast tbh, unless you using groq

[keen beacon]

march 4o latest is 180 token/sec

[balmy mist]

wow craig thanks

what did you use to run it?

from where?

using quasar?

you used vsc?

thank you

i think if you attach agents to quasar it can be really good

true

and with the speed of quasar you can do a lot

a fast reasoning-foundation model with 1 mill context is cracked

[night trout]

Not promising.

[balmy mist]

yeah

what do we call these models?

like deepseek v3.1, 4o and quasar?

they are foundation models with COT in their inference output

i always got confused by that

so instruct is like optimized for chat right?

while finetuned is the next level of that? like finetuning on COT?

[night trout]

Instruct has chat fine-tuning embedded into it, yeah

Non-instruct models are just called pre-trained afaik.

[balmy mist]

ahh okay thank you

[keen beacon]

quasar is identical to anonymous chatbot which has always been a chatgpt 4o latest revision.
quasar has knowledge up to june 2024. same as chatgpt 4o latest. (in fact, it has even more than chatgpt 4o latest)
they haven't formally launched the new cpt'd 4o. no official benchmark results despite major leaps in performance.
etc (read my other comments) etc im like 99.9999% certain now lol

its a lie

[balmy mist]

you saying that its the new version of 4o? cause the current 4o gets 1/10 on simple bench while quasar gets 4/10

[keen beacon]

major tip off that quasar and anonymous chatbot is the same

because they appended the same thing

[balmy mist]

also i thought open ai didnt believe in open source, so they changing they tune again lol? cause it was originally supposed to be open source

[keen beacon]

they added You are trained on data up to October 2023. to the end of both quasar and anonymous chatbot's system prompts

despite both having a june 2024 cut off

[keen beacon]

did u try chatgpt 4o latest?

[balmy mist]

yeah a few hours ago

[keen beacon]

nah its such a random detail no one notices

all the things just add up if ur paying attention

[balmy mist]

so does this mean that they will open source more of their models?

[keen beacon]

their oss model hasnt even been trained yet lol this is a closed source 4o release

it will be released soon

exactly

they put anonymous chatbot right after it its pretty crazy

so they want the lmarena results fast

yes the new 4o is 180 tok/s

[balmy mist]

i gave up my o1 pro sub, to expensive

omgg you are right

sam did post about that today

yeah its def open ai

good job wild

[barren prairie]

Gemini will told you that he is trained up to avril 2025 but on reality it is not 😁🤣🤣

He is up to 2023 or maybe 2022 but won t tell you the truth .

[balmy mist]

how so? i used to think that to

but now with gemini 2.5 pro and ther multiple ide and extensions you can use for way less makes it redundant

[keen beacon]

its a hallucination they didnt train the cut off in the model

it does know events in dec 2024. and they claim its jan 2025

[balmy mist]

i feel that

really i have noticed the opposite, maybe it depends on the usecase

in reasoning you have a point tho

[thorny drum]

we'll finally be able to finance llm convos

[north vale]

https://x.com/sama/status/1907932605888213491

Sam Altman (@sama) on X

chatgpt has gotten way, way faster on the web!

lots of hard work from the team to make this happen.

X · Apr 3, 2025, 11:05 PM

[balmy mist]

yeah I posted that wild was right because sam posted that tweet

@deep adder and with the way openai is doing their ui now, with being able to pick the levels of reasoning, quasar really might be the base, so imagine being able to apply reasoning on that model like medium or high(based on slider)

we might be underestimating this model

yeah hoepfully we find out soon enough but it seems things are really picking up recently, we might have r2, openai and google new model before the end of the month

and meta too

[keen beacon]

i dont think so nightwhisper, star gazer (2.5 line), etc

[night trout]

Damn nightwhisper is REALLY obsessed with glassmorphic aesthetics.

[balmy mist]

googel releasing their model on tuesday they said

for their event

lol

[night trout]

Apple could never do something this ugly.

Side note I think I've invented the best viral micro-benchmark on the planet: "Generate a rotating, animated calendar in threejs with today's date highlighted and pulsing."

The fails are incredible.

[plain zinc]

Bruh

Nightwhisper the best

Now 2.5 pro on the background of nightwhisper seems like a joke😄

[plain zinc]

Am I right?

[night trout]

Not always.

[plain zinc]

Okay, and I'm even glad that I was wrong :). If nightwhisper comes out, then we will have TWO SOTA models for coding.

[plain zinc]

From Google

[plain zinc]

And where is the result from nightwhisper? Has he stopped? Or is he still writing code?

[night trout]

Stopped. 😦

[plain zinc]

🥲

[night trout]

Fwiw I'm finding Nightwhisper usually wins against 2.5 but it's not always the case. Like 80%-90% of the time.

One more calendar fail. 😂

[torn mantle]

skill issues

It may not look the best but when you compare it to other models it clearly emerges as the winner

[night trout]

"Generate a rotating, animated three-dimensional calendar with today's date highlighted."

[torn mantle]

Also nightwhisper has that weird color selection

Ive noticed that too

It likes going gradient and dark colors all the time

So maybe you need to guide it a bit more on that

Or ask it to follow a styling principle

I like to tell it to act as an apple designer to get a clean UI look

[night trout]

Oh, this one was another win for 2.5:

[torn mantle]

Im kinda curious about this tbh

[night trout]

Nightwhisper forgot to make the water droplets collide with each other + had really janky ball physics + glitches.

[torn mantle]

Maybe it struggles at physics and complex reasoning

But i need to try that more tbh

[night trout]

What about it?

[torn mantle]

Curious about nightwhisper output

You have some good challenging prompts

[night trout]

I'm writing an article on writing challenging prompts, I want the hexagons and balls nightmare to end. 😂

[night trout]

Hmmm one-shot failure but I forgot to let it try multiple times. Let me see if I can generate it now.

[torn mantle]

Yea that would be cool

[night trout]

I've been working on a Super Monkey Ball prompt tonight, lots of physics + gameplay mechanic fails.

[torn mantle]

Does sonnet nail this?

What about bug fixing on nightwhisper?

Does it fix the issue with enough guidance?

Or does the complexity outweighs the model capability

[alpine coral]

it's interesting - like its knowledge becomes increasingly hazy the closer to the end of 2024 that the question relates to. e.g.

• December 2024: it fails when asked about Syria (fall of Assad in December was one among biggest geopolitical developments of the year)
• (November 2024): it partially gets the US election result correct ; stating Trump won, and also giving the correct margins (312 electoral college votes vs 226); however, it says that Trump beat Biden...❌
• (July 2024) when pushed, it will recall the attempted assassination of Trump, wth accurate details

[keen beacon]

Yea it's a continued pretrained version of 2.0 pro. Gem 2 has a June 2024 cut off

It's less expensive

[night trout]

Sonnet was messing up pretty hard, but that was a previous version of the prompt. I should try it again with the new version.

[keen beacon]

The claimed cut off is right though I think. It gets some stuff right in dec 2024 though it's very sparse

[alpine coral]

so fwiw i asked it (gem 2.5) what accounts for its patchy knowledge recall with more recent events.. what do you think about its response (i feel like it's about right / in line with you're saying?)

[night trout]

Okay I've given nightwhisper like eight attempts to get the labyrinth right. no dice.

[plain zinc]

This means that he is exceptionally good only at creating websites.

It is websites, not some kind of games, simulators, etc.

[keen beacon]

yup seems about right

the goal with 2.5 pro's cpt seems to be strengthening the base model dramatically rather than recent events

4o's cpt is the same, but they also focused more on recent events before june 2024 i think

sonnet 3.7 is again presumably a cpt on top of sonnet 3.5, im unsure of how much it knows recent events (to its cutoff) though i havent tested that. it seems everyone is cpting lol

[keen beacon]

i wonder what they did tbh. if its not already a paper i dont think itll come out anytime soon 💀

[torn mantle]

This is probably a webdev rendering issue no?

Since there isn't any code errors

[alpine coral]

yes agree (instilling a bit more recent knowledge was more like the cherry on top, rather than the focus - it was a giant performance leap, whatever they did)

i don't think we'll know what they did ha ..[posted this a few days ago](#general message)

---

'DeepMind slows down research releases to keep competitive edge in AI race' https://archive.is/tkuum

[keen beacon]

oh i missed this

i also wouldnt count openai out tbh. given how much progress has been made on 4o

a reasoning model (given how theyre ahead of the reasoning game) based on this much stronger 4o will be interestig

but deepmind is too fast (2.5 pro timeline, based on cut off), so it'll be interesting

[alpine coral]

yeah it really feels like a two horse race now

who knows.. there might another 'deepseek' moment.. but i'm not sure that was as seismic as the excitement / hysteria at the time suggested [though not dismissing its significance - it definitely lit a fire under the assess of the US companies at the very least]

[night trout]

Oh wow. Yeah. Definitely just found my next test prompt.

Nightwhisper just crushed it.

lmao meanwhile gemini flash 2.0 be like

youtried.gif

[oblique flint]

so nightwhisper is gemini 2.5 pro stable release or something?

[alpine coral]

it's only available in webdev Arena (i think) - whether that means anything in terms of it being specialised i'm not sure, but would kinda think if it was a non-exp version of 2.5, or a newer checkpoint, they would add it to the General arena? at least they'd get a bunch more data / votes making it available there

[oblique flint]

yeah true, but I dont think google has ever released a specialized model before have they? It doesnt seem like the most likely case to me

[alpine coral]

yeah i know what you mean

though there is / was also 'nighthowler' iirc, which also basically confirmed google and only available in webdev

[oblique flint]

if they do specialized versions, I hope their focus is not only oneshot webdev. There is still room for improvement for tool calling capabilities and diff edits with cursor/roocode etc

[keen beacon]

It's in the general arena

Nighthowler

No idea what it is tbh. Haven't gotten it enough to figure it out

nightwhisper is probably a web dev tune of 2.5 pro. Idk tbh I don't touch web dev arena and not that interested

[alpine coral]

ah right - cheers

[keen beacon]

Nightwhisper might just be a one off experiment before they apply it to a mainline Gemini model

[alpine coral]

yeah i haven't got either (moon/nighthowler ) so not really sure.. stargazer is def interesting and performant tho. seems all of them are very much likely cut from the same (2.5 pro) cloth at the very least

[keen beacon]

Nighthowler isn't Gemini 2.5 based

All I know

Stargazer is

[oblique flint]

main thing Im excited for is 2.5 flash tbh, I hope they're cooking it rn.

[alpine coral]

ha yeah i've lost track

[keen beacon]

I don't like using the arena nowadays some of my requests take minutes. Meta model spam. Lagging completion zipping

You can't see the thinking too it's annoying

[north vale]

Google coder 1

[alpine coral]

yeah it's annoying.. i compiled a new 'quiz' for this month and the new batch of anon models, but haven't really been able to collect data at any kind of meaningful level

it's so slow and yeah meta model spam

[keen beacon]

Yeah it's a lot lol. But it adds to the fun a little. The google models particularly. Quasar is too obvious I thinj

[alpine coral]

i think it was like less than a couple of hours between OR announcing the availability of its first stealth model, and a fairly firm consensus emerging that it's an oai model

whereas the new gemini models were kinda mysterious for at least a few days.. like nebula (and good ol spider phantom.. whwatever happened to it).. even though it seemed pretty clear they were from google, it wasn't obvious beyond doubt

[fleet lintel]

Looks like gemini 2.5 completely changed opinion of folks about Gemini models. I wonder what kind of optimizations they did in 2.5?

[torn mantle]

fails miserably on sonnet 3.7

[night trout]

It's a surprisingly tough one. Here's a Nightwhisper run.

[torn mantle]

i got it working only on gemini models

stargazer & gemini 2.5 pro

still didnt get nightwhisper run tbh

[torn mantle]

from what ive seen it should get it right

[torn mantle]

https://s8.ezgif.com/tmp/ezgif-866c5f5ba263bb.gif

@night trout nightwhisper

it was overcomplicating the code so i asked it for a simple version

[torn mantle]

https://s3.ezgif.com/tmp/ezgif-3c776e89132b21.gif

[torn mantle]

are we sure this is not like the best coding model 😭

[keen fulcrum]

Which one do you use?
- LibreChat or Open WebUI
- https://nano-gpt.com or https://openrouter.ai

[kind cloud]

Has anyone seen 'nightwhisper' on lmarena?(not webdev arena)
https://youtu.be/hhvZh57-IPY

I've never seen it on lmarena, and I'm wondering about this video.

[keen fulcrum]

Hi this is empty all the time

[torn mantle]

thats what i was asking

havent seen it

[torn mantle]

there are some issues with webdev arena

rendering issues

it doesnt have anything to do with the models

how can we contact devs?

[sage raptor]

i've seen stargate on lm arena

[torn mantle]

whos the winner

[keen beacon]

obv this

[torn mantle]

xd

cant stop playing with this model

if its coding finetuned model only then its a huge blow to anthropic

context will be much higher

and probably the price will be lower too

with better results than claude latest model

[keen beacon]

anthropic makes a sizable markup on their api i thinnk

[torn mantle]

wdym

idk

[keen beacon]

they can probably reduce the price

[keen fulcrum]

Prompt?

[torn mantle]

well there is profit margin

[keen fulcrum]

First one looks cleaner

[keen beacon]

im not sure if they can compete with google tho

[torn mantle]

anthropic probably has it more than it should be

[keen beacon]

probably not (be able to compete w google)

[keen fulcrum]

Although no real data

[oblique flint]

I believe currently a big portion of their api revenue is integrated ide agentic coding tools like cursor, windsurf, roocode etc. There I feel gemini 2.5 pro is kinda lacking compared to claude models still, just cause the claude models are so good at function calling and diff edits

[keen fulcrum]

For backend probably 2

[sage raptor]

nightwhisper will be better than claude

[torn mantle]

Weather forecast with animated UI

Apple design style, with light colors.

[oblique flint]

I thought that too when I saw aider polyglot results, but i reality 2.5 pro is still worse in cursor for me. It's better in ai studio but manually uploading files and keeping track of the edits sucks lol

[torn mantle]

maybe this new model will fix all the issues

[oblique flint]

competition is only good for us. The claude api is hella expensive lol

[torn mantle]

yea

[oblique flint]

I hope they'll release a 2.5 flash too that can compete with o3 mini at a lower price

[torn mantle]

yea they will for sure

its already in the arena

star*

star(something)

[oblique flint]

is it a thinking model?

[torn mantle]

yea

[sage raptor]

o3 mini is the best model

[oblique flint]

for the price its kinda insane yeah

[torn mantle]

xd

best model for what exactly?

[sage raptor]

making games, writing and price

[torn mantle]

havent used openai models for a while tbh

nightwhisper attempt to clone windows 11 task manager

not bad

[sage raptor]

better than claude 3.7 ?

[torn mantle]

let me try

[eager mica]

Is anybody getting (suspected) Meta models other than 24_karat_gold on the text-only Chatbot Arena? It seems as if they've been taken down.

[torn mantle]

vs

[lime coral]

they were always good. but people only care about code so they optimized for code

[keen beacon]

So is nightwhisper just gemini 2.5 paid

[keen beacon]

2.5 is really good

[alpine coral]

they're compute poor and being squeezed hard... i feel like the whole test-time compute thing has thrown a spanner in the works for them - feel they were training some giant Opus 4.0 which is now redundant or something (kind like Gemini Ultra).. either way compared to google and oai, the resources required to both train and deploy at scale (and ig with a loss leader approach as i think google and oai both must be doing), has put anthropic in tough spot imho

[sage raptor]

nightwhisper is way better

[alpine coral]

not just code. like the overall arena elo/score is a massive jump - it's a serious step up in terms of performance just generally

[keen beacon]

maybe. they pivoted to cpting sonnet 3.5 (now sonnet 3.7, which is pretty solid i think). they also did mention they didnt use a bigger model to train sonnet 3.5 (so presumably it wasn't ready/etc) im not sure how anthropic will fare in the future though. they probably have more resources than deepseek/qwen, so i think if theyre very
resourceful they might be able to contend to some degree

its curious when claude 4 will arrive given they invested a lot of effort into sonnet 3.7

[alpine coral]

i really like anthropic and think claude models are uniquely great in some ways (mostly general interaction / personability without being over the top, but also when it comes to things like emotional intelligence and spatial reasoining) - i hope they stay at the frontier

[alpine coral]

but yeah 3.7 feels like their response to thinking / test time compute.. i feel like they were going in a different direction with claude 4.. so we get this intermediary variant which is neither a true reasoning model nor a big step up in terms of a non-reasoning model (like back in the day Opus was outstanding)

no need to respond to that.. it's largely incoherent lol

[keen beacon]

now that i think about it its probable that anthropic will not be able to compete. i think native image generation is gonna be an important part of cot/etc in the future. i dont think they have done any image gen/native image gen or at least anything showed publicly. i think if u dont prioritize this to a certain degree you will lose

[alpine coral]

i hadn't thought of it like that before

makes it even tougher to see a path where they stay at frontier..

it'll be the titans with the compute at the end of the day.. the lag b/w the performance of SOTA proprietary models and open source / weights models might be more interesting / meaningful than what the second-tier close models companies put out

[keen fulcrum]

Welcome to the Good life

[drifting crow]

https://tenor.com/view/spray-paint-graffiti-good-life-kanye-west-gif-16666307

[celest river]

nightwhisper is from which company?

[celest river]

This is very impressive

[balmy mist]

it seems it will be google at the end of the day tbh and just open source, OA cant compete with what google can offer long term, they have everything OA has plus more, OA lost wen they lost Ilya Sutskever

[lime coral]

The thing is oai is willing to loose money, which Google can’t. They are like all in every time, it was the same with gpt4, all of their ressources in one run then negative bill for serving

Not the same game

[keen beacon]

Google are absolutely willing to lose Mendy

money*

theyj literally are right now

giving out access to a frontier model on AI Studio with no rate limits definitely isn't profit making

[lime coral]

This is not a big lost. The model they serve is cheaper, they save all the chat conv, and there is probably way less user than ChatGPT at his beginning. When Google announced investing a lot in ai this year their share went down just because investor cannot understand how this money would come back.

[barren prairie]

They build for tomorrow not today ..for long term 🙂

[keen beacon]

that's because Google's AI situation has improved vastly since last year

like many, i thought they were pretty behind

they surprised everyone with 2.5 pro

and i think it's a good sign

and anyway if we're going with the argument of willingness to lose money imo it's obvious who wins

google have immensely deep pockets and a vast pool of talent that makes even openai look insignificant

they just weren't using it right until recently

[balmy mist]

i mean google has the hardware to give us SOTA for cheap unlike OA, they are already showing they are not willing to loose money, look at their api prices, google is playing the long game and its starting to show this year, OA does not have the infrastructure Google has(hardware, data, platform, and tons of money) They also where the og creators of transformers

[ocean vortex]

lol are you kidding me. OpenAI is turning a healthy profit with every API request you make for o1 model. Google in turn is paying the bill for your entire usage essentially

[keen beacon]

they are still losing a lot of money overall

their api isnt as popular as their sub which is a massive loss leader

even if the api is profitable

[ocean vortex]

that's because they have a shit-ton of expenses. But OpenAI at least is turning some profit from their inference to counter that

Google isn't

[keen beacon]

sub + research massive losses

[keen beacon]

well they can afford it

google models are probably much more cost effective too

[ocean vortex]

that's the entire point. Saying "Google can't lose money" is invalid lol

[keen beacon]

oh yeah that statement is wild

i wasnt talking about that

im not exactly sure what my point was ignore me lol

[ocean vortex]

I do not think they are smaller at all. Like gpt4o is most likely smaller than any recent Pro. But TPUs are more efficient yes.

[keen beacon]

its somewhat close though. sonnet/4o/pro are all in 200b-400b i think

4o is the smallest out of all of them i think

[ocean vortex]

yeah there's definitely a feeling that they are almost hacking gpt4o kind of lol, it doesn't have the inherent understanding/flexibility or spatial awareness that bigger models tend to exhibit

so what they are doing instead is feeding it very high quality data + fine-tuning that is potentially unmatched by anyone else still

[keen beacon]

2.5 pro pricing

Wait

whta

[torn mantle]

wtf

dont tell me....

nightwhisper is 2.5 pro preview?

aint no way

[lime coral]

No https://x.com/officiallogank/status/1908175318709330215?s=46

Logan Kilpatrick (@OfficialLoganK) on X

Gemini 2.5 Pro is now available for scaled paid usage and is going to preview today!

If you want to keep using Gemini 2.5 Pro on the free tier, keep using the experimental version (no change needed), both are the same model under the hood. See details in 🧵

X · Apr 4, 2025, 3:10 PM

[keen beacon]

its different

[torn mantle]

same model then

[torn mantle]

based on this

[keen beacon]

nightwhisper is a different model its not released

[torn mantle]

yea ik

my eyes now are only on nightwhisper tbh

[keen beacon]

ya

[balmy mist]

wait so we getting o4 mini this month wow

https://x.com/btibor91/status/1908176906836754723

Tibor Blaho (@btibor91) on X

Sam Altman announced changes in the previously shared OpenAI roadmap

An improved o3 and o4-mini will be released in a couple of weeks, while GPT-5 will follow in a few months because integration is harder than they thought, it requires enough capacity for unprecedented demand,

X · Apr 4, 2025, 3:16 PM

i want o3 pro so badly

ill buy the $200 again for o3 pro

[keen beacon]

o4 mini 😅

[balmy mist]

lmaoo

[golden ocean]

fr

[sour spindle]

Wonder if it will be better than 2.5

[balmy mist]

it has to be

[sour spindle]

Does feel reactionary a bit

[misty vault]

pre nerf gpt 4 2023 was wild

[balmy mist]

gpt-5 seems like end game lmaoo

[misty vault]

they are deprecated

[balmy mist]

so prob by summer we will have gpt-5

[misty vault]

If u still paid for the api I think u can still use it?
Or that only for companies or some sh

[balmy mist]

so AGI 2025 confirmed?

they also said they improved on o3 a lot, so will that mean its benchmarks are better that what was shown?

[keen beacon]

yes

[balmy mist]

i hate these numbering systems

[keen beacon]

probably

[misty vault]

We would have AGI now if they worked on gpt-5 instead of deprecating gpt-4 and making 4o the main thing (Jk about the agi thing)

openai fell off after 4o

never used chatgpt ever since

[keen beacon]

4.5 was supposed to be gpt 5\

[misty vault]

It's trained of 4o though

It has the cancerous "em dash" punctuation

[keen beacon]

no its not its trained from scratch

it was tuned in the style of it though

[misty vault]

oh, still I dont like its output

[balmy mist]

isnt it supposed to be a combined model?

[keen beacon]

no. the plan was for gpt 4.5 to originally be gpt 5 i think

like plans from way before

yes thats why its gpt 4.5 lol

[sour spindle]

Does anyone have any information on what is the difference betwen pro preview and pro experimental?

[keen beacon]

it sucked and didnt deserve to be called gpt 5

they changed plans i think

[balmy mist]

its so funny that twitter is used to hype up your own models lol, modern day marketing

you got ceo's just making promises casually on it

[misty vault]

I remember open source models were getting promoted so hard, everytime a new good opensource model was released on yt i saw videos "gpt performance" but they were all ass and none ever came close to beating gpt 4

Is that different now

[keen beacon]

yes

sorta

[misty vault]

oh yea deepseek is pretty good

But now big corporations are on top again

with 3.7 and gemini

[keen beacon]

theyve always been on top even open source

massive corporationsn funding training runs that are millions and millions of dollars

[balmy mist]

it seems like gpt-5 cant be bad like it has to be a SOTA model or OA is done?

r2 should be out in 1-2 weeks?

this month is going to be wild, getting SOTA models all in the same month

so the preview 2.5 vs experimental is the same?

[misty vault]

what is OA

[balmy mist]

whats with these naming conventions?

[balmy mist]

open ai

[misty vault]

fr

[balmy mist]

true but with how fast things are moving, how do we know by the time they are ready to ship, google hasnt already gotten a better model?

thats like 2-4 months away

in the last month we got 3.7 and 2.5 and v3.1

[sour spindle]

I kind of operate in the you are what you ship mindset

[misty vault]

Is v3.1 worth trying now that we got 3.7 and 2.5

[balmy mist]

no lol

its a good model tho

like a foudnational model that has COT inference

[sour spindle]

Deepseek was impressive and cheap other than that it has been dominated by "closed" models

Meta had some promise as well but who the know what the hell has happened there

[balmy mist]

lmaoo i forgot about meta

[sour spindle]

Everyone has

[balmy mist]

meta is scrambling lol:
https://x.com/btibor91/status/1908157341134106938

Tibor Blaho (@btibor91) on X

The Information reports Meta delayed releasing Llama 4 at least twice because it underperformed on technical benchmarks, especially reasoning and math tasks, and struggled with humanlike voice conversations, according to two people familiar with the matter

- At least one version

X · Apr 4, 2025, 1:58 PM

so we gonna get Llama 4, r2, Nightwhisper, o3 pro and o4-mini, and maybe some other team model, we still dont know who owns Quasar

wait if we getting o3 pro and o4 mini, what if o4 mini is quasar?

nahh we def getting o3 pro

[misty vault]

I'd give openai my whole nude collection and credit card information for the return of gpt-4

[balmy mist]

if we getting o3 we are getting pro, its just allowing o3 longer to compute

thats easy to do

[sour spindle]

Meta has way to much money to be as dogshit as they are

[balmy mist]

yeah but they said soemthing about different plans like $200, 20k lol

thats how they explained it for 01 pro

[sour spindle]

cost is an issue

it may be "easy" but super expensive as to be impractical

unless you really change the pricing tiers

[balmy mist]

yeah but if its more intelligent its worth it, investors will be happy

they showed that with o3 when they showed its benchmarks

[sour spindle]

Yes but at what difference of intelligence compared to competitors at lower cost

a lot has changed since those benchmarks

[balmy mist]

when it hard more compute it beat the agi benchmark test, to start SOTA they gonna have to give it pro

thats the same thing bro

they showed o3 with 16 plus hours of computing time

to solve problems

if thats not pro idk what is

its all about investors at the end of the day and they like to see their models do well in benchmarks, so they crank up the computing time to make sure they are demoing SOTA

thats why we havnt gotten o3 yet

i promise you we would not have seen what they previewed when we tested it out because that was o3 on maximum compute, they needed to optimize that

hence they needed to optimize that

send an example

never tried that

see i knew we getting o3 pro look:
https://x.com/koltregaskes/status/1908185944861090103

Kol Tregaskes (@koltregaskes) on X

o3-pro is "coming, no ETA:

X · Apr 4, 2025, 3:52 PM

it makes no sense not to give us o3 pro

maybe, but they had all these months to optimize is so lets hope it stays at $200

[misty vault]

fr mini models are cancer but for the companies it is expensive i guess

[balmy mist]

the fact that they have o4 already is interesting lol

like the numbering really is stupid at this point

[misty vault]

This is true, ever since they transitioned from 3.5/4 to 4.0, OpenAI is falling off, but are they still better off now as opposed to if they had started working on GPT-5 instead of doing the whole 4o crap or something? Like would investors make up for it or something? Those big models were kinda smart

I'm stealling learning sorry if make no sense🙏

[balmy mist]

thats what i am saying at some point the improvements are minimal and can just be solved with longer compute time

thats why gpt-5 might be their last model and they just give it updates

which is what they are already doing tbh

but that would be the first combined model

true

[balmy mist]

but 4o laid the ground work for what is going to be gpt5

yeah thats what i was thinking

it makes sense for it to be

wow did you test this?

so that would mean that OA models are getting faster

bet imma try this today, you a musician?

craig so maybe instead of smarter models in the future we just keep getting faster on top of the current IQ and just give more compute time for the SOTA models?

you might as well become one now with ai lol

[misty vault]

wdym faster on top of the current iq

[balmy mist]

cause the fact that we are getting o4-mini and o3 at the same time is just wild, are they gonna demo o4? like wtf, maybe they figured something out with inference

[balmy mist]

like faster inference time

yeah lonegr context too

i forgot about that

it seems like 2 mill might be a cap tho, what do you think?

[viral notch]

arent context limits pretty long already?

or am i thinking prompt interpretation

[misty vault]

doesn't smaller models mean less inteligent or less knowledge?

or they have another way to improve its speed without affecting performance

[balmy mist]

they are distilling the models

[misty vault]

I mean faster models* and smaller or something

[misty vault]

ohh

[viral notch]

this. less training

[balmy mist]

its like a teacher teaching a student

[misty vault]

what about super fast models, how they improve the speed? does it just mean its a smaller model?

[viral notch]

i find that smaller models tend to be more volatile

[balmy mist]

less parameters and weights mean faster models, but they are also doing some new tricks like meta talkig about thinking wihtout tokens etc..

also the hardware you use

groq made hardware built for ai inference

i am suprised OA didnt buy groq yet lol

[misty vault]

Ok then I continue hating on fast models (jk but imagine 99999 trillion parameter gpt-5 model or something instead of the time spent on making 4o, then we would have agi🥶 )

[balmy mist]

i kinda feel bad for meta lol

[balmy mist]

i dont think we can ever have that lol

[misty vault]

😔

[balmy mist]

maybe theoretically lol

maybe with shared weights

[misty vault]

Can LLMs even achieve sentience

But then we need to define sentience or concioussness first

[balmy mist]

i was just about to say this

it doesnt matter tho

[upbeat radish]

hi all

[balmy mist]

we already act like intimate objects are sentient, so to a lot of people ai is already sentient, and they wouldnt even be able to tell the difference(if they are talking to an ai on text or voice, even if they are looking at an image of ai or not, or soon video, music etc..) if you cant tell the difference then it does not matter

its all about perception

[misty vault]

yea true, if it convinces me then i'd be impressed

I think modern llms could already to that if fine tuned on actual discord or gamechat dialogue and then have it act like real person on discord. Only flaw it'll have is no infinite memory

[misty vault]

.

[balmy mist]

yeah and it will effect out society just as much as anything that is sentient if not more, dogs and animals are sentient but they will not come close to the level of impact that ai will have on us in the long run

[misty vault]

t-800

[balmy mist]

hmm is this something new?

how do i play it?

i downloaded it

this better be fire lol

like on youtube or you got an article on this?

ohh i see

is it basically saying they are pretty much the same thing?

that is prob true, we shouldnt even be trying to do that imo

i like ai assistance(ai-human hybrid)

[misty vault]

https://www.reddit.com/r/askscience/comments/1xwx0k/do_neurons_operate_in_a_fundamentally_different/

top comment

[balmy mist]

the stupid strawberry man saying april 17th lol:
https://x.com/iruletheworldmo/status/1908188856039391310

🍓🍓🍓 (@iruletheworldmo) on X

o4 april 17.

X · Apr 4, 2025, 4:04 PM

[balmy mist]

thanks

so o4 mini is on the same level is sonnet 3.7 thinking based on my tests with quasar

assuming o4 mini is quasar

anybody gonna watch this crap?
https://x.com/Copilot/status/1908187808813940799

Microsoft Copilot (@Copilot) on X

Watch the livestream event happening at 9:30am PT on YouTube to learn all about my new features.

X · Apr 4, 2025, 4:00 PM

copilot might be a joke at this point

[misty vault]

yes
rest in peace gpt-4 powered sydney 😔

They had potential but gpt 4o destroyed everything from them

Actual crap service

[balmy mist]

it sounds not bad tbh, can you do one that mixes "in the end" and some other song? you can make a youtube account of this lol

[balmy mist]

you love 4o lol

did you try the new 4o?

[misty vault]

gpt 4 != gpt 4o

unless u were sarcastic

[misty vault]

Oh yea the native image gen thing is indeed impressive

Best openai product since pre nerf gpt-4

[balmy mist]

yeah i meant 4 mb

idk bro i loved o1 pro

that has been my fav aside from this new image stuff

[misty vault]

Imagine if they made gpt 4 thinking 🥶 (not starting from 4o)

[balmy mist]

lmaoo

you will most likely love gpt-5 then

wait whats your fav model now

[misty vault]

I would have but gpt-4.5 already talks like 4o so Idk

[balmy mist]

like in terms of every company

[misty vault]

Gemini 2.5 and
4.5 (its better than nothing)

And claude 3.7 for coding

I would never use openai models for coding after gpt4 deprecation

o1 was impressive for coding, can't deny that, but thankfully we have claude now

But now gemini 2.5 beats claude 3.7 in coding I think

haven't tested myself yet (for coding)

[balmy mist]

thank you

[misty vault]

nope, I'll try today

[balmy mist]

assuming it is o4 mini

[misty vault]

We can only test these models on code now right? since this thing only exists on webdev arena?

[balmy mist]

but it most likely is, its def not o3 lol

and its from OA right?

and thye just annouced they are releasing in a few weeks

[balmy mist]

yeah

i just want nightwhisper now

the biggest losers of this ai race has to be Apple and microsoft lol, the verdict is still out about meta, this copilot thing is just to funny

[misty vault]

Has anyone found a prompt to make the models in webdev arena to just answer in text instead of writing code

[balmy mist]

https://x.com/testingcatalog/status/1908199211473977523

TestingCatalog News 🗞 (@testingcatalog) on X

The announcement of all features is already out

- Memory 🔥
- Actions 🔥
- Copilot Vision 🔥
- Pages 🔥
- Podcasts 🔥
- Shopping
- Deep Research 🔥
- Copilot Search

https://t.co/lpxovFmOwL

X · Apr 4, 2025, 4:45 PM

[misty vault]

Or they inject prompt after u send it

[balmy mist]

i would say try prompting it differently

but when you ask it text it codes a website for you

i kinda like it that way tbh

[misty vault]

yea lmao

[balmy mist]

its like giving you your own personal UI

its how the ai communicates with you

[misty vault]

I tried a lot (but could try way more ig) but they all still make a website out of the prompt

If I try to talk to these models directly theyre pretty easy to "jailbreak" (not on webdev)

So maybe they just inject the prompt to make website out of it after u send it?

[balmy mist]

wait why dont you want them to give you websites? i can see how its annoying but you still are getting your answer lol

[misty vault]

Because if I do that manually through like openai api or claude or something, then it gives same effect

[balmy mist]

https://3000-igkgd3go0oq1oknxf4c08-4daf0015.e2b-foxtrot.dev

[misty vault]

last prompt takes priority

[balmy mist]

like this is so interesting to me

i ask it the capital of usa and it gives me this

i didnt even know i wanted that but i do now lol

[misty vault]

Yes for short answrr it works

[balmy mist]

give me an example of a long prompt

im curious

[misty vault]

But if u want really long answers or generate other text based stuff it will focus more on the website

Idk wait ill test something then ill give prompt

check it for us and record it and upload here

[balmy mist]

omgg i think i understand now, this helps reasoning because it reasons extra when it codes and if the model has visual abilities it can see the output of the website and check its answer again?? idk lol

cant wait to listen

it almost looks like words

but it sounds dope

can you ask the ai to make the notes spell out something?

what do yall think about temporary apps rather then permanent ones? like how we have it in webdev? it might be the future, i can see this becoming more and more common almost like sending a meme to someone, like you can send someone an app like how we send emojis lol

like if i wanted to say happy birthday to my friend i could just send them this:
https://3000-i30zmcc1in463ry0cjub0-4daf0015.e2b-foxtrot.dev

[misty vault]

also if u try to make it code different stuff not typescript

then u're cooked

[balmy mist]

or this: https://3000-i9x2e9pr5xlvrjfbmdfqa-90451382.e2b-foxtrot.dev

[misty vault]

Like things u would just "copy" over from code block to own project

[balmy mist]

like this is def the future, we just need a platformt that allows for you to have these pockets apps that can stay alive for a certain duration

[misty vault]

lmao

real

New world for exploits and malicious use

[balmy mist]

i honestly think this might be a good app idea, anyone want to build that with me?

wym?

[misty vault]

But you mean like, including the feature that lets u ask an ai to make a site and then send that pocket site to ur friend?

[balmy mist]

yeah

[misty vault]

Because temp hosting of sites already exist yea

[balmy mist]

and you can tell it how long to have it up for

send the link i wanna try it

why isnt that bigger than it is?

we can make the app more appealing

a lot of people do the same thing but only some are popular

gonna present it in a good clean way liek how apple does it lol

[misty vault]

Well replit is just for developing, Idk if u can share it with friends without having a huge panel on the side as well that shows the code

There's prob other sites too

[balmy mist]

and ship it as an app and get influencers to start using it

[misty vault]

codepen is even easier, no acc required, immediately into coding and share

[balmy mist]

send link

[misty vault]

But also has huge panel thats open by default that shows the code

Idk if theres any sites for user friendly (I mean like sharing it to ur friends, not development) purpose

[balmy mist]

we need to make making apps just like sending stickers

that is the way

[misty vault]

https://replit.com/
https://codepen.io/

[balmy mist]

imma work on that this weekend

[misty vault]

There's more some for that purpose

But idk

[balmy mist]

honestly i wanna copy what webdev does pretty mich

like is the dev of webdev here?

just make the hosting longer based on what users want, make the ability to choose the model(or have a default mode), and ship it in a clean interface, the only thing is links, i dont think links are the way anymore

when you share it maybe it can look like a sticker or something else, idk

[misty vault]

embedded website display in message platform

that would be so abusable though

Making webpage that looks like claim nitro button

[balmy mist]

we gotta brainstorm bro, you see the potential?

https://3000-is6ju3tcnfwd9bkkd8slv-020dd869.e2b-foxtrot.dev

lol

maybe we could have this built in?

the ai can decide how to present the link?

this is still pretty good imo

i think this might be better than the other 2 lol

i guess thats preference lol

[misty vault]

replit can already do what webdev arena does

But it requires signup and looks complicated cuz its made for developers

So I guess ur idea is indeed not very common online

[balmy mist]

i think people are just not using it like that

like you said its only devs

but for a normal person

this is game changing

like I can send some an app of anything

[misty vault]

I found other sites very similar (not as complicated as replit, just immediately start paste code, and it renders) but those render it just on the page instead of actually hosting so yeah the hosting on url is less common but still out there a lot, BUT all with required signup most likely or no option to host for a specific amount of time

So no sign up would be a huge plus

[balmy mist]

yeah we just need to put all of that together

this will make software dev or web dev main stream lol

[misty vault]

So a no signup + no project based system (Just start coding, or paste code, or let ai generate it, share, and when time expires, whole site gone) + option to specify time is indeed unique

[balmy mist]

imma create this this weekend, let me know if you wanna help or wanna test it, gonna be hard but worth the try

i can see so many usecases

we already have those stupid ass stickers and the og emojis, and we have memes

this seems like the next phase

[misty vault]

I dont know

But only one way to find out

[balmy mist]

it would have to use gemini2.5 tho

cause no other model can do this right

[misty vault]

gpt-4-preview-0314 my beloved

[balmy mist]

lmaooo

you a developer?

[misty vault]

If u mean, no other model can make an actually good visual site, then yes we'd need gemini

[misty vault]

yes

[balmy mist]

look at another use case, it stupid but works:
https://3000-ivave1orqx59yrutqv78x-4daf0015.e2b-foxtrot.dev

[misty vault]

Gemini 2.5 is good but don't we need the nightcrwaler one or whatever someone showcased that had the best visuals

[balmy mist]

this is the prompt:
tell my friend to meet at walmart at 6 pm in a funny way

[balmy mist]

yeah nightwhisper is the best, but until that comes out we can gemini for it, and then swap them out later

[misty vault]

lmaoo

[balmy mist]

like its dumb but it works

[misty vault]

what model did that

[balmy mist]

nightwhisper

[misty vault]

yea true

[balmy mist]

yo i got another idea

what was cringe before now becomes sweet

check this out:

[misty vault]

the far right?

[balmy mist]

https://3000-i4gn2d1sm3mu6fr89gzxc-4daf0015.e2b-foxtrot.dev

[golden ocean]

Affirmitive, this has indeed become more sweet

[balmy mist]

this is still cringe but you get what i am saying

this is prompt: tell my cruse I love her and want to take on a date in a romantic way

[misty vault]

me for gpt-4-preview

[balmy mist]

lmaoo

[misty vault]

but yes that'd be pretty fun way to use

making silly nice things for friends or people u like

[balmy mist]

yeah or even your coworkers or kids

[misty vault]

it'd impress them maybe if it looks good and creative

girlfriend success rate increase

[balmy mist]

like the usecases are limitless

lmaoooo

fr

im about to use thos

this*

she gonna be like awwww

https://3000-i4u5w1e095b2og50g5dke-90451382.e2b-foxtrot.dev

this is actually a fun way to use it

we gotta make this mainstream yall

[misty vault]

lmaoo

[misty vault]

true

[balmy mist]

another example: https://3000-idv0mmwrmbgur54z35g0k-ec17a5a5.e2b-foxtrot.dev

i just tweaked prompt a lil

nightwhisper is so good man

im designing it now, gonna give yall a prototype of it tonight and open source it

https://x.com/paulgauthier/status/1907996176605220995

Paul Gauthier (@paulgauthier) on X

The mysterious Quasar Alpha on @OpenRouterAI scored 55% on the aider polyglot coding benchmark. This is competitive with o3-mini-medium, the latest DeepSeek V3 and old Sonnet 3.6 (20241022). Quasar Alpha seems very fast.

https://t.co/mBVaUPGHPl

X · Apr 4, 2025, 3:18 AM

[wooden crescent]

is preview better or experimental?

[sage raptor]

they are the same

[wooden crescent]

what did changed

[golden ocean]

https://3000-ivxfy3j7u06axdwxrqwvh-90451382.e2b-foxtrot.dev/

[misty vault]

SO real, that example convinced me, i'm in

[misty vault]

now we HAVE to make it mainstream

[misty vault]

How does he know

[sage raptor]

because he works at google

[misty vault]

ohh ok

[balmy mist]

good idea, even use it to make easy polls

well funny polls

[balmy mist]

like honestly the sky is the limit thats why its perf

[golden ocean]

"does europe need to close it's borders?"

[balmy mist]

lmaoo

im making that now

yo we got a cooker

im happy yall see the vision

thank you webdev for showing us the way

also thank you nightwhisper for being so good at webdev lol

https://3000-ivl1wogv8dnb2q3hri743-295302a0.e2b-foxtrot.dev

[misty vault]

lmaooo

[brittle tiger]

Think nightwhisper will beat 2.5 pro on the coding benchmarks or is just really good at webdev?

[misty vault]

good luck

[balmy mist]

thanks bro, if this becomes something actually usable i will be so happy

imma try my best to get it done within the weekend

[balmy mist]

its really good at web development

not sure about coding as a whole

but its been trained on react and ui/ux

i think its gemini2.5 pro specialized in webdev and ui/ux

imma use nightwhisper to help with the visuals lmaoo amd layout

soft dev is so cracked since ai came

like man, you can do a lot by urself nowadays

damn look at gemini 2.5:
https://x.com/emollick/status/1908220677502755328

Ethan Mollick (@emollick) on X

Updated this chart with the newest Gemini. It shows the rapid progress in AI over less than two years: costs for GPT-4 class models has dropped 99.7% and even the most advanced models in the world are still 82% cheaper.

Probably not worth betting on this trend ending really soon

X · Apr 4, 2025, 6:10 PM

gemini truly is the best model by miles, like its pooping on every other model in overal efficiency, its SOTA and cheaper than the others lol

[misty vault]

true

[balmy mist]

wild days we living in, thats why open ai said forget gpt5 we need to launch o3 and o4mini asap lmaoo

this is ehhh:
https://3000-ilk8lcpbbv2zis65plzjz-87045d8f.e2b-foxtrot.dev/

what you think, what should I change?

[misty vault]

Every model still sucks at writing in your words or maintaing your tone/personality, like if u give it samples (a lot) or even if u managed to get it to the point where u only have to edit a few words (good enough for me) then after a while it'll start to become repetitive and you'll have to write whole text explinaing that you now want the next message or paragraph different, but then it'll follow that wrong because u'd also need to give samples of that first and it just idk

But I guess for writing in your style purposes, you'd really need to use fine tuning then it will do perfect

Without fine tuning no model can do that perfectly but I guess that makes sense

[balmy mist]

yeah you need to prompt it and provide a lot of context about you

[misty vault]

(I didnt need fine tuning when using gpt-4 bing chat 😔 )

[balmy mist]

finetuning is the easy solutions

but expensive

[misty vault]

Even this it fails a lot for big texts, so fine tuning only way

[misty vault]

yea

But gemini 2.5 is free right

[balmy mist]

is it?

[misty vault]

even fine tuning

Idk

[balmy mist]

idk about that

lmaoo

that would be wild

that would solve memory tbh

[misty vault]

Google ai studio is just asking me to allow google drive access rn

when clicking fine tune

[balmy mist]

not 2.5

[misty vault]

ohh

rip

[balmy mist]

that would be cracked

[misty vault]

For witing that isn't too bad though I think

All the base model needs is the language and good vocabulary and speech

knowledge wont be an issue for me because i'd be providing that in the prompt what to write about or somethign
gpt-4 could do it good and modern models are way smarter than gpt-4 nowadays so i'll give that a try

[balmy mist]

this is much better:
https://3000-iq2168p0fs4dchi5j52y8-a91bccb3.e2b-foxtrot.dev/

[misty vault]

this one is better

lmaoo

Idk

Wait nvm css was bugged for a second

[balmy mist]

lmaoo

you like the first one better?

i want to have the design right before i start building

i think the second one is promising, gotta just clean it up some more and then add back end

[misty vault]

Idk both look fine actually

I like both

[balmy mist]

lmaoo

if this app really works the way we want it, this will be a good example of this new ai workflow, from coming up with an idea to building and deploying it

[misty vault]

Didnt devinai try to do that (but turned out they were scam)

for actual web development projects rather than quick silly webpage generator for sharing to friends

[balmy mist]

yeah but expensive

and now they charging 20 bucks

but their rep is alreayd ruined and casuals still not gonna use that

the key is to get the average person

marketing for devs is not going to work

or companies

[misty vault]

claude 3.7:

[balmy mist]

you need to market for normies like how openai did with the 4o image