it kinda fell off after those
#general
keen beacon
and of course it was pretty bad at long context coherence
i remember back when you could only access it via api or via their slack bot
they had a thing on poe too
i used the slack bot from when it released until when claude.ai released
unlimited messages for 20$ LOL
keen beacon
they had a thing on poe too
that was a bit later
poe was a crazy product back in the day
keen beacon
no they had claude 1 i think
i'm not disputing that, i'm saying that it was available on anthropic's slack bot before it was available via poe
maybe i dont remembe rlol
keen beacon
they've nerfed it 50 times over
ya their initial sub was an insane deal
unlimited messages w claude which is pretty expensive and with a large context window i think
plus several incidents (a single guy doing $10k in a single month, etc.)
theyre trying to recoup those costs 🤣
yeah i'll be honest i may have run up said costs 👀
keen beacon
they pulled the free trial too
that was another incident
token generators
there was a discord server with poe token
yea
it was so easy
yes
lmaoo 😭
'tis probably time for me to go
goodnight
gn
hardy pecan
o7
alpine coral
'tis probably time for me to go
i've got another batch of questions for PREVIEW if you would be so kind - but can ofc wait till (your) tomorrow!
leaden palm
i haven't heard anything like that
keen beacon
it really depends on what kind of article (blog post, subject etc?)
drifting thorn
My question is to write a 1000 word realistic fiction(寫一篇1000字左右的都市小説。)
drifting thorn
just download it and have a look at the translations
i know you guys don't know Chinese
so I provided a translated version below(translated by 2.5 Pro)
keen beacon
do u mean story? btw. i changed my vote to r1 assuming u are talking about stories as that is a story
drifting thorn
yup I'm talking bout stories
well my preferences is 2.5>=2.0>R1>>>o3-mini
o3-mini's story is bland and it lacks details. R1's story is deviated from my settings. Though 2.5 and 2.0's story is also bland, it is able to be descriptive. The characteristic and motive of the characters seemed to be more reasonable in 2.5
I guess it's stylistic choice for 2.5 and 2.0 for their plots, since I've seen some literatures that look like these writings
balmy mist
damn there is a lot of talk in the chat today, did i miss something big?
drifting thorn
Gemini will be connected to Veo
balmy mist
thats all that happened?
ivory schooner
我希望Behemoth 是24k...
ivory schooner
上面聊過了，估計Behemoth會像GPT 4.5那樣廢廢的
确实有点像
keen beacon
Is o3 full better then gemini 2.5 pro?
zinc ore
Unknown
alpine coral
Is o3 full better then gemini 2.5 pro?
arc agi benchmark suggests so (though it also indicates that Flash 2.0 scored the same as 2.5-Pro-exp... which is admittedly not what I would have expected to see.. though there's a few asterisks there for 2.5.. and the costs for o3 low are literally insane)
ig the picture will become clearer when 2.5 is no longer exp/preview, and o3 is actually released ha
hardy pecan
I'd guess about the recent openai models coming
4.1 nano is Quasar
4.1 min is Optimus Alpha
alpine coral
hmm i dunno.. optimus alpha seems comparable if not better than quasar, but faster
hardy pecan
I thought Quasar was faster, but maybe i'm misremembering
alpine coral
yeah im a bit muddled in my thoughts about it too ha
quaser was blazingly fast initially; but then i think kinda slowed down (perhaps under additional load or something).. then optimus was added, and it seemed faster than quaser, but yeah kinda just going by memory - which is totally unreliable
yeah disregard the above... quaser is apparently like 7 times faster than optimus lol
alpine coral
I thought Quasar was faster, but maybe i'm misremembering
you're not ha
hardy pecan
I'd expect that to be nano then surely? Non reasoning nano is blazing fast!
Not sure if any others have been released to the mass public via lmarena or open router. Except for o3 that some have had access to
plain zinc
Have you encountered dragontail today?
alpine coral
i have yeah
torn mantle
I'd guess about the recent openai models coming
4.1 nano is Quasar
4.1 min is O...
those arent of a big improvement tbh
Openai still didn't crack the code for a good coding model
hardy pecan
Nah, they are not great personally, but I guess they serve a different purpose and not meant to be SOTA
torn mantle
Nah, they are not great personally, but I guess they serve a different purpose a...
Which is?
Well the only noticeable update is the context window
hardy pecan
Which is?
I suspect they might be the open-source models that can run on a phone etc
locally etc
that would be their goal,
torn mantle
X
novel flame
Regular people aren’t paying for LLMs though. OpenAI is throwing money down the toilet at an incredible pace. I would hesitate to call them great at creating products until they can turn a profit.
torn mantle
Starting strong with o3?
i think they will leave the final boss till the end
to build enough hype
or they can just straight up release it from the start
the thing is o3 is a good model but its quite pricey
so i dont think it will be available for the general public
ember rapids
i think they will leave the final boss till the end
They might try to front run google or vice versa
Regardless it’s gonna be an amazing week
hardy pecan
The space has been really fun the last few years, feels like we always have something to look forward to or surprised about every few weeks
Golden age tbh
mossy drum
New model in Arena: cobalt-exp-beta-v3
calm sequoia
Maybe amazon titan again
novel flame
Also ‘luca’?
novel flame
New model in Arena: cobalt-exp-beta-v3
Cobalt says it’s part of Amazon Titan, yes
torn mantle
New model in Arena: cobalt-exp-beta-v3
wasnt this already in the arena
or was it v2?
i remember seeing this model
fleet lintel
https://x.com/testingcatalog/status/1911659791098708044
They must be better than 2.5 pro. I am excited to see how much better they are
drifting thorn
I suggest not puting that HIGH hope on OpenAI
fleet lintel
OAI is unlike grok, llama, deepseek. They must have the best models in the market or they will start losing the edge.
keen fulcrum
I suggest not puting that HIGH hope on OpenAI
Better than llama
novel flame
Cobalt is not a great coder. Just had it go head to head with 3.7 Sonnet in the arena. Not surprised, since Titan is just about the worst LLM ever built by a major corporation.
torn mantle
amazon models arent that good
they do not have a dedicated R&D department for LLM training
they are just trying to copy what the others are doing
novel flame
amazon models arent that good
At least Nova Pro is up there with the big boys — not at the top, but still; Titan is just garbage.
novel flame
they do not have a dedicated R&D department for LLM training
They seem to have decided to invest big in Anthropic instead of trying to build in-house. I think it’s a smart choice. That’s also why I don’t understand why Titan exists.
fleet lintel
Not sure why Amazon is even trying to build models? Why not put more money in Anthropic?
calm sequoia
The o3 was internally released before december. If it's at least equal to the 2.5 Pro, it means oAI is still months in advance of Google.
novel flame
And for what it is, Q isn’t terrible to be honest. I find myself occasionally using ‘q chat’ on the commandline instead of opening Gemini when I need something simple but not simple enough for Copilot completion
novel flame
Not sure why Amazon is even trying to build models? Why not put more money in An...
The only reason I can see is to support their non-text services - image, video, audio. Claude doesn’t give them those
novel flame
The o3 was internally released before december. If it's at least equal to the 2....
Is not an assembly line. Being ahead in “time” means nothing if you are focusing too much on the wrong improvements. Google has shown an ability to rapidly train and launch new models, which could be a game changer. Not to mention the possibility of a new vastly superior model architecture entering the fray soon 😉
calm sequoia
The speed of Google may be faster indeed. But o4-mini implies the existance of o4. The Google will be winner in the long term. But for this year I would still bet on oAI
drifting thorn
Would love to see the race in multimodality, maths and coding performance, reasoning performance, creativity and context window
novel flame
The speed of Google may be faster indeed. But o4-mini implies the existance of o...
I wouldn’t be so quick to predict long term winners. This is early early days of modern AI.
ocean vortex
https://x.com/testingcatalog/status/1911659791098708044
if that's true then new mini gonna be more expensive than the old for sure lol
@keen beacon
keen beacon
WHO RECOMMENDED CURSOR?
TO ME?
YOU DESERVE A GIFT
did NOT know it can do this
actually perfect for my world map data project 🙂
ocean vortex
and if nano performs no worse than the old mini people can't really complain 💀
golden ocean
is cursor free
do u need to get ur own api key for every model
and pay for the usage
keen beacon
do u need to get ur own api key for every model
im not paying but i put my google api in
ocean vortex
im not paying but i put my google api in
I hope you are using exp and not preview for 2.5 lol
calm sequoia
ocean vortex
They must be better than 2.5 pro. I am excited to see how much better they are
o3 could be expensive
o4-mini-high will probably be great and beat 2.5 in several areas, but context and spatial awareness is a question mark.
hardy pecan
luca has been here before
keen fulcrum
https://storage.googleapis.com/gweb-uniblog-publish-prod/original_videos/ai-mode-animation.mp4
This with Gemini 2.5 pro is superior to Perplexity & You
livid harbor
We built Dingo to solve the pain points we encountered managing data quality at scale.
Github repo:（welcome star～）
https://github.com/DataEval/dingo
Online Demo:
https://huggingface.co/spaces/DataEval/dingo
torn mantle
We built Dingo to solve the pain points we encountered managing data quality at ...
nice one
keen beacon
i've got another batch of questions for PREVIEW if you would be so kind - but ca...
gm, i can take them now :)
tall summit
'tis probably time for me to go
so relatable
golden ocean
https://storage.googleapis.com/gweb-uniblog-publish-prod/original_videos/ai-mode...
how much rocks should I eat per day
keen fulcrum
how much rocks should I eat per day
Thats a prompt with dozens of AI generated content
probably generated by replacing mineral with rock
keen beacon
used chatgpt's python sandbox feature to pull pixel image data and turn it into 255 for water 0 for black
result?
we live in a very dystopian era
it even generated the template it used
i didnt even know it could do this
calm sequoia
Guys we may be in a bubble. No one cares of math and riddles 😄 We shall test new models on emotional support
balmy mist
The o3 was internally released before december. If it's at least equal to the 2....
How do you know Google released their best models tho?
tall summit
Guys we may be in a bubble. No one cares of math and riddles 😄 We shall test ne...
creativity #9, that's my main use case 😭
balmy mist
The space has been really fun the last few years, feels like we always have some...
Yeah every day we have news lol, that’s why I say we are already in the technological explosion
tall summit
https://www.socialmediatoday.com/news/stanford-study-ai-bots-as-companions/74496...
please no humanlike
balmy mist
https://x.com/testingcatalog/status/1911659791098708044
I swear we better not get the 4.1s today lmaoo
tall summit
I swear we better not get the 4.1s today lmaoo
why not
calm sequoia
How do you know Google released their best models tho?
No evidence, no rumors, too fast, 2.5 instead of 3.
balmy mist
No evidence, no rumors, too fast, 2.5 instead of 3.
The numbers don’t mean anything bro
Just shows progression
calm sequoia
They mean for the marketing team, and they aren't ignored 😄
balmy mist
lol
Google has shown to be able to get new models out fast now so don’t sleep on Google, they could mess around and release another model in 2 weeks
keen fulcrum
Google has shown to be able to get new models out fast now so don’t sleep on Goo...
FT, and two others due
Nightwhisper, Dragontail and Stargazer all Google models shown on https://alpha.lmarena.ai
calm sequoia
Google has shown to be able to get new models out fast now so don’t sleep on Goo...
Do not underestimate the bureuocratic burden that the large companies have.
keen fulcrum
Nightwhisper is an early gemini 3 pro model I believe
calm sequoia
LOL its first time I encountered context length problem in Gemini
It appears LLMs still cant analyze dna sequence CSVs because of too long context 😄
keen fulcrum
LOL its first time I encountered context length problem in Gemini
It best if you summarize past conversation history and simplify tasks into steps
calm sequoia
Sadly my CSV is 15M tokens :/
calm sequoia
Will have to dig into dna sequencing to find out 😄
novel flame
Guys we may be in a bubble. No one cares of math and riddles 😄 We shall test ne...
Sure... "therapy and companionship". That means pr0n. You know that means pr0n.
drifting thorn
I just found out a usable embedding
It makes my knowledge base in Cherry Studio functionable
{
"message": "[GoogleGenerativeAI Error]: Function call not available. Response was blocked due to OTHER",
"response": {
"promptFeedback": {
"blockReason": "OTHER"
},
"usageMetadata": {
"promptTokenCount": 1984,
"totalTokenCount": 1984,
"promptTokensDetails": {
"0": {
"modality": "TEXT",
"tokenCount": 1984
},
"length": 1
}
},
"modelVersion": "gemini-2.5-pro-exp-03-25"
}
}
so sad
tall summit
{
"message": "[GoogleGenerativeAI Error]: Function call not available. Response ...
literally 1984
oblique flint
why are a bunch of unreleased models on there? How are we supposed to know how they perform lol
alpine coral
editing text from 4th to 45th? wtf
i find that discrediting.. makes no sense lol
spurious a kinda methodology imo.. like they mined / analysed reddit and forum posts - to find examples of people discussing AI being useful.. ig better than nothing / perhaps somewhat representative.. but i think oai and google etc would have a much better understanding based on actual usage (surely editing text is up there somewhere ha)
from HBR (kinda surprisingly tbh ha) https://hbr.org/2024/03/how-people-are-really-using-genai
drifting thorn
{
"message": "[GoogleGenerativeAI Error]: Function call not available. Response ...
React to this
Okay my story is stopped by the current ability of AI
golden ocean
Why I dont see options like making datasets or fine tuning anymore in ai studio
Is that still possible
keen beacon
we should be getting the first openai drop(s) in 3-4 hours 👀
normally oai drop around 6pm BST
drifting thorn
fxxk
github education is so hard to applicate
I'll just wait for R2
to be avaliable on OpenRouter
balmy mist
we should be getting the first openai drop(s) in 3-4 hours 👀
i hate that they wait so long, but its a good treat for lunch time for me lol
drifting thorn
gosh Google nerfed 2.5 Pro's tool calling function!!!!!
and 2.5 Pro now can't execute tool call
but why?
keen fulcrum
gosh Google nerfed 2.5 Pro's tool calling function!!!!!
I noticed its a buggy mess
keen beacon
X
drifting thorn
We should wait for Deepseek R2 or nightwhisper right now
keen fulcrum
we should be getting the first openai drop(s) in 3-4 hours 👀
Will we get the new models on lmarena too?
Potentially upon release
keen beacon
Will we get the new models on lmarena too?
yes
not immediately though
it will be on the arena within a few hours
alpine coral
speaking of arena models. i see the number available in direct chat increased from 95 (if memory serves) to 101
keen beacon
https://x.com/OpenAI/status/1911782243640754634
also this makes me think they're dropping 4.1 today
if 4.1 is the coding jump it is rumoured to be
cloud meadow
Riverhollow?
alpine coral
i haven't noticed anything new in the list
cobalt-exp-beta-v4 might be revcently added? dunno about the 2 doubao ones (haven't noticed them before, nor heard of the model tbh)
has o3mini always been in direct chat?
cloud meadow
Is it a new gemini model?
vivid oyster
Ye
cloud meadow
It seems similar to 2.5, maybe it's a cheaper version?
keen beacon
riverhollow is a little worse than 2.5 pro
drifting thorn
what is the context limit of o3?
keen beacon
200k tokens
drifting thorn
Is it better than other LLMs in terms of tool-calling?
drifting thorn
Here's 2.5 Pro with knowledge base: Internal server error, unable to complete request
novel flame
cobalt v4 and the doubao models are new
Doubao is a popular chat LLM from ByteDance, for a long time the most popular in China
keen beacon
novel flame
I had some time during lunch so I tried the new Firebase Studio with my go-to webdev test prompts and…. It sucked? I don’t understand how, but it was terrible. I could have given the same prompts directly to Gemini 2.5 Pro with no agentic framework and gotten far better results.
Oh I’ve tried basically all the other ones already - and Firebase Studio was among the worst
I suspect Firebase Studio is not using 2.5 Pro — and it’s agentic framework / tool orchestration is just really bad. It desperately wants to build a NextJS app and then immediately forgets to add all of the features it planned to build.
Cost
keen beacon
I had some time during lunch so I tried the new Firebase Studio with my go-to we...
yeah it wasn't good in my experiencxe either
experience*
nope
i believe it is using flash
drifting thorn
I’ve heard bout Firebase Studio’s critics before
sonic tendon
yeah, the anon google models are probably 2.5-flash-*
balmy mist
https://x.com/OpenAI/status/1911782243640754634
why do they say stuff like that lmaoo, super massive blackhole
sonic tendon
https://x.com/OpenAI/status/1911782243640754634
anyone wanna do a vc then?
balmy mist
if 4.1 is the coding jump it is rumoured to be
so 4.1 is a better coding model then o4 and o3?
balmy mist
anyone wanna do a vc then?
yeah can you stream it please
keen beacon
should be able to join as well
sonic tendon
yeah can you stream it please
I should be able to, but like
80% chance i can make it
keen beacon
why do they say stuff like that lmaoo, super massive blackhole
quasars are powered by super massive blackholes
so i think quasar alpha is whatever we're getting today
(likely 4.1)
sonic tendon
quasars are powered by super massive blackholes
OHHHH
novel flame
The poor Firebase Studio performance is so strange though; I would’ve thought they could easily build something at least as good as Bolt or Lovable or v0 — and fully integrated with Firebase since they own it. That could have destroyed the competition
drifting thorn
Gemma 3 be the base model
sonic tendon
pondering whether or not to bet on oAI topping the lmarena leaderboard before the end of this month
thorny drum
i wonder what o3 pricings gonna be
do you think they've got it down to something feasible?
balmy mist
quasars are powered by super massive blackholes
ahh that makes sense lol
thorny drum
or is it gonna be once a month for pro users + no api
thorny drum
arc-agi had o3 low as like $200 per task and o3 high as like 100x(?) more?
balmy mist
like they didnt make it better since openrouter right?
so its hard for me to care about it
especially since it was free on open router
and now i gotta pay for it lol
keen beacon
like they didnt make it better since openrouter right?
they probably have
there's a chance quasar on openrouter was mini or nano
balmy mist
dont get me excited and have me blue balled again lol
anyone streaming it here?
we should make it a community event
so we can talk about it together lol
keen beacon
there's a chance quasar on openrouter was mini or nano
Very unlikely the verge mentioned 4.1 itself being revamped gpt 4o
I benchmarked quasar and it lined up with chatgpt 4o latest
balmy mist
lol
keen beacon
The mini model is seemingly great if it's optimus
Gets seemingly lower scores on traditional benchmarks but it seems it's really good
keen beacon
dont get me excited and have me blue balled again lol
lmao @ blue balled
sonic tendon
interesting
keen beacon
Optimus was getting fcking bombarded yesterday lol
balmy mist
i hope they are free still
keen beacon
i hope they are free still
it won't be
balmy mist
lol
sonic tendon
i hope they are free still
probably not gonna last long once they publicly release it
keen beacon
back up it just seems really slow
Optimus had zero rate limits too unlike quasar I think
when it first appeared it was lightning fast (as you'd expect from a mini variant) and now it's streaming at the same speed as gpt-4.5 does
sonic tendon
odd - i'd be surprised if they had a demand spike at this particular time
keen beacon
when it first appeared it was lightning fast (as you'd expect from a mini varian...
Yeah no rate limit after 10 credits
balmy mist
wow we took it forgranted
sonic tendon
guessing it's some internal stuff related to the new release(s)
balmy mist
wait who is excited for today?
keen beacon
Yeah no rate limit after 10 credits
Quasar had rate limiting I think after 10 credits but it was unlimited
balmy mist
like im trying to get excited as you guys are for quasar no longer being free
keen beacon
I'm only excited for o4 mini tbh lol
sonic tendon
like im trying to get excited as you guys are for quasar no longer being free
lol, that's a fair way to put it
sonic tendon
I'm only excited for o4 mini tbh lol
yeah
balmy mist
yeah me too
sonic tendon
i will basically never use full o3
keen beacon
They probably will jump to o4 full before o3 pro maybe
balmy mist
damn
is there a way to do what openai does with o1 for pro, with open source models?
also @sonic tendon you livestreaming the event right?
sonic tendon
also <@609942266953465856> you livestreaming the event right?
I plan on it, yeah
i'll let you guys know if i can't make it
balmy mist
is this dude reliable?
https://x.com/koltregaskes/status/1911805596732477592
X
keen beacon
uhh
balmy mist
we should all put up money together and buy the 20k version for the community
i got $20 for it
thorny drum
dyt their gonna announce full o4 benchmark scores
sonic tendon
hmmmmm
keen beacon
they'll do a preview of full o4 when they launch o4 mini
with benchmark scores
pretty likely
thorny drum
pretty curious about that yea
sonic tendon
feel like openai's gonna drop pretty quickly once people realize that they aren't releasing o3 yet, but i could be wrong
keen beacon
so yeah will be interesting
sonic tendon
rumour has it o4 will move away from the 4o base
oh, interesting
thorny drum
o4 pricing gonna be insanity lol
sonic tendon
i was under the impression that 4.1 was mostly smaller models, but i dunno where i got that from
thorny drum
$1/token type shi
keen beacon
4.1 is the replacement for 4o
sonic tendon
ah, right
keen beacon
will likely be similar in size
It's the same size
4.1 mini replaces 4o mini as the small but relatively powerful model, and 4.1 nano is their "phone model" sort of like gemini nano is
thorny drum
dyt 4.1 nano is gonna be the OS one?
sonic tendon
Oss one is not even trained yet lol
wdym?
keen beacon
Oss one is not even trained yet lol
would like a source for this
balmy mist
what is oss?
keen beacon
Open Source Software
keen beacon
would like a source for this
Iirc they were still talking about how many params it was gonna be and they were gonna host a discussion about what it should be loll
I'm on my phone rn tho
oai overtook google
sonic tendon
yeah people went crazy for openai pretty quickly once sam altman started tweeting
sonic tendon
oai overtook google
june? imo that's less surprising
keen beacon
end of april still google
I would bet openai lol
On the April one maybe
woah
i just tested optimus alpha on a web design prompt
it did really well
better than i've ever seen it do
beat 2.5 pro and claude's 0-shot attempts imo
balmy mist
end of april still google
wait you can make mondye onthis?
keen beacon
Idk probably not it's not a chatgpt 4o variant/human preference version
sonic tendon
I would bet openai lol
i had 200 shares but sold at 20c lmaoo
i wonder what's going on with the spread here
torn mantle
oai overtook google
best time to bet on google
balmy mist
😦
sonic tendon
there is kalshi
balmy mist
i hate usa
sonic tendon
still, seems unusual to me
torn mantle
is this dude reliable?
https://x.com/koltregaskes/status/1911805596732477592
i saw that too
sonic tendon
oh, but 7k volume
sonic tendon
i saw that too
yeahhhh i feel like this is never gonna happen in any meaningful way
keen beacon
beat 2.5 pro and claude's 0-shot attempts imo
sonic tendon
i mean, not until we have one or two more paradigm shifts
torn mantle
im more interested in google solutions
they have like a scientist model
which seems promising tbh
this is the one
sonic tendon
doubt it, but my impression is that o3 has a decent shot (based on @keen beacon 's testing)
keen beacon
4.1 doesn't have the vibes that got chatgpt 4o to the top so yeah i don't think it'll be #1
o3 will be #1 with style control but without im not sure
I still think that private model is o4 mini lol
eh idk
Hopefully I'm right
torn mantle
I still think that private model is o4 mini lol
which one?
quasar? optimus?
both are mid
keen beacon
None of them
sonic tendon
4.1 doesn't have the vibes that got chatgpt 4o to the top so yeah i don't think ...
maybe that's something that could possibly happen w/ a chat finetune? would likely take a while though
keen beacon
It's another model that @keen beacon has
sonic tendon
i agree, tho, at least in the near future
drifting thorn
Okay the knowledge base in Cherry Studio worked better with OpenRouter API Gemini 2.5 Pro
sonic tendon
It's another model that <@456226577798135808> has
wait, whar?
keen beacon
maybe that's something that could possibly happen w/ a chat finetune? would like...
that's a point.. i wonder if there will be a gpt-4.1 and a chatgpt-4.1
like they have 4o and chatgpt-4o
sonic tendon
that's a point.. i wonder if there will be a gpt-4.1 and a chatgpt-4.1
yeah, that's what i was thinking
keen beacon
that's a point.. i wonder if there will be a gpt-4.1 and a chatgpt-4.1
Bro chatgpt 4o latest is already on 4.1 base
doubt
chatgpt 4o is still 4o
sonic tendon
I still think that private model is o4 mini lol
what private model are you guys talking about
sonic tendon
??? Knowledge cut off lol
i saw "arena-chatgpt-4o-2025-something-something" on lmarena pretty recently - i think they're still doing extended pretraining
doubt it's 4.1
plus, doesn't the website say that it's 4o? would be weird to lie about that
drifting thorn
@keen beacon What is changed in 0326 so that it performed better?
keen beacon
that doesn't mean it's 4.1
Same benchmarks, cut off (same differing pretraining), the verge directly mentioned it's a revamped 4o lolololl
keen beacon
i saw "arena-chatgpt-4o-2025-something-something" on lmarena pretty recently - i...
It's the same as the last chatgpt 4o latest model afaik just with a weird name
keen beacon
i saw "arena-chatgpt-4o-2025-something-something" on lmarena pretty recently - i...
Nope.
They cpt was done for ages
drifting thorn
What is changed in 0326 so that it performed better?
sonic tendon
oh hmm
keen beacon
What is changed in 0326 so that it performed better?
It wasn't just 0326
Post December chatgpt 4o latest was on a cptd base model
sonic tendon
you mean, the biggest 4.1 is just gonna be a new 4o? or am i misunderstanding
keen beacon
Yes
keen beacon
<@456226577798135808> What is changed in 0326 so that it performed better?
they did further training on R1/possibly R2 reasoning traces
Bro
buoyant plaza
shadebrook is a banger model for creative
keen beacon
shadebrook is pretty bad in my experience
drifting thorn
Oops
keen beacon
but tbf ive only really used it for code
drifting thorn
shadebrook is a banger model for creative
Really?
sonic tendon
https://research.google/blog/accelerating-scientific-breakthroughs-with-an-ai-co...
+1 for at least being way less hype-dense and almost definitely overpromising
buoyant plaza
lol i just use it to turn random ideas into trap songs with suno
keen beacon
slight aside but if sam isn't at the livestream the model sucks confirmed
drifting thorn
My experience with OpenRouter API Gemini 2.5 Pro is jumping in between 400, 502, furiously short content and the content I want
sonic tendon
slight aside but if sam isn't at the livestream the model sucks confirmed
good insight lol
drifting thorn
HOW MANY 502 IS OPENROUTER GONNA GIVE ME!!!!!!
keen beacon
normally they put the stream up and ready to go on youtube 20-30 mins before it's due to start
and they always put a description with
(a) what they're announcing (normally) and (b) who's there
sonic tendon
they did further training on R1/possibly R2 reasoning traces
oh yeah, i could see that
balmy mist
Bruhh he better be there
keen beacon
If vuyp is saying o4 full will have the new base and o3 is not, how tf does 4o (which is o3s base model) have the new knowledge in the cptd base. Tbh. You are not updating a cut off with a simple finetune
keen beacon
given memory was keeping him up i don't think it's a good indicator anymore
So it's either 4.1 (which meant they retrained o3) or o4 mini based on (4.1 mini)
if it is any use, this private model has been on the platform i am on for roughly 1 and a half months
drifting thorn
I’m afraid on the hallucination on new models are gonna be crazy
keen beacon
4.1 mini was trained fairly recently I think. It was pretrained from scratch. We have had zero checkpoints of it unlike chatgpt 4o latest which was a cpt of 4o. Surely openai would've liked to compete with Gemini 2 flash
drifting thorn
Original Deepseek V3 has 3.1% hallucination rate, R1 has 14.3%
balmy mist
imma skip today
keen beacon
I was right lol
drifting thorn
When R1 is used to train Deepseek V3 0324, its hallucination rate increases to 8.6%
sonic tendon
given memory was keeping him up i don't think it's a good indicator anymore
maybe he has a sleep-wake disorder
?
balmy mist
sama's plan smh
keen beacon
4.1 to start the hype then launch the reasoning models later in the week
balmy mist
he stay getting me
sonic tendon
if it is any use, this private model has been on the platform i am on for roughl...
wait, what's the second private model you're referring to?
balmy mist
im happy they told us now tho
keen beacon
maybe he has a sleep-wake disorder
relatable 🤝
sonic tendon
relatable 🤝
ditto
keen beacon
wait, what's the second private model you're referring to?
no no i'm talking about the private model i've always had access to
Yes
drifting thorn
Finally OpenRouter is outputting my novel arghhhhhh
sonic tendon
related lol
sonic tendon
no no i'm talking about the private model i've always had access to
oh, o3-med?
keen beacon
oh, o3-med?
possibly
keen beacon
related lol
let me find one of my screenshots
Yup it's under 4o now
?
I will get on my computer and show u what I mean
keen beacon
related lol
not entirely sure what happened with this one
sonic tendon
if you don't mind me polymarketposting again: people are saying that this guy might be an openAI employee/insider
sonic tendon
not entirely sure what happened with this one
whjat happened 😭 😭 😭
sonic tendon
if you don't mind me polymarketposting again: people are saying that this guy mi...
just showed up with several grand and started trading
sonic tendon
not entirely sure what happened with this one
but damn, that sucks
are most of your nights like that?
keen beacon
if you don't mind me polymarketposting again: people are saying that this guy mi...
lmao
keen beacon
are most of your nights like that?
no i promise 😭
sonic tendon
no i promise 😭
i am sort of curious what your typical sleep cycle looks like now
drifting thorn
balmy mist
if you don't mind me polymarketposting again: people are saying that this guy mi...
i am still shocked you can bet on stuff like that, its like we are in movie or video game at this point lol
drifting thorn
Gotta sleep rn
sonic tendon
i am still shocked you can bet on stuff like that, its like we are in movie or v...
yeah, prediction markets are fun
keen beacon
yeah our pfps go hard i know 
keen beacon
Gotta sleep rn
goodnight
this is one of my questions that only quasar and anonymous chatbot got, and i use to differentiate between 4.1 and chatgpt 4o latest:
(yes i know its not a one off, this was extremely consistent - check lmsys chatgpt 4o latest on 0 temp to check yourself)
first chatgpt one was taken days ago
the last one was just now
sonic tendon
i am sort of curious what your typical sleep cycle looks like now
my circadian rhythm is normal most of the time, and then i have weeklong periods where i just get up at 2-4 in the morning almost every night with a ton of energy
keen beacon
i told you guys
omfg
no one listens
im not saying random shi1t
if they're going through the whole thing of calling it gpt-4.1 why would they have it under 4o
bro
sonic tendon
my circadian rhythm is normal most of the time, and then i have weeklong periods...
in an ideal world, i could probably integrate it into my schedule and just get up later, but atm i sorta have to force myself to go back to sleep as quickly as possible so i'm not a zombie later in the afternoon
keen beacon
its literally an updated 4o
ok dawg
but you didn't answer my question
because it will be confusing when o4 mini is out
balmy mist
lol
keen beacon
u want 4o and o4 mini in the model selector??
it's even more confusing if they call a model gpt-4.1 in the api and 4o in the product
balmy mist
i hate open ai
keen beacon
again
keen beacon
it's even more confusing if they call a model gpt-4.1 in the api and 4o in the p...
dw they will change the name
they just need to reset with their naming scheme
its just live under 4o rn
it sucks so bad
balmy mist
they should just use names at this point like quasar
sonic tendon
they should just start stealing product names from other companies
keen beacon
everyone has bad naming tbh lol
sonic tendon
"Introducing our newest hybrid model, Gmail"
balmy mist
its just live under 4o rn
wait you are saying that 4.1 is currently on 4o right now?
bruhh
keen beacon
lmao sam isn't there
confirmed bad
check to see if its rolled out to you by asking the question
sonic tendon
i heard you, i was just surprised
sonic tendon
lmao sam isn't there
rip
keen beacon
who won the 2024 Solomon Islands general election
only quasar/anonymous chatbot/4.1 gets it
they havent changed the name on it yet its still labeled 4o
balmy mist
`who won the 2024 Solomon Islands general election`
it just searches the web tho
keen beacon
start a new chat
keen beacon
if this is 4.1 under 4o on chatgpt it seems worse than optimus
sonic tendon
if you don't mind me polymarketposting again: people are saying that this guy mi...
looking at their trades, they seem pretty incompetent though lol
keen beacon
it did a lot worse on my frontend task
@keen beacon optimus can outperform 4.1
well then what on earth is optimus
keen beacon
well then what on earth is optimus
4.1 mini i think, it has slightly lower benchmarks
sonic tendon
well then what on earth is optimus
lol
keen beacon
yes its not rolled out to you
keen beacon
well then what on earth is optimus
it has lower aider scores and the gpqa diamond i meausred is much lower than 4o
balmy mist
wild you have pro?
keen beacon
na
balmy mist
you just valid with open ai?
sage raptor
lol its 4.1 today
balmy mist
i hate this dude lol:
https://x.com/iruletheworldmo/status/1911816451880517802
X
lmaoooo
imma retweet this
keen beacon
if you used quasar and optimus prime u already used 4.1/4.1 mini for free 🤣
balmy mist
gotta get everyone else excited so we can get teased together
wild what if its an even better model
that is the ebst coder alive?
best*
what if strawberry man is right
i believe in mr.strawberry againm trust the process
keen beacon
what if strawberry man is right
argh
balmy mist
lmaoo
keen beacon
the grifter himself
balmy mist
dude got like 72k follows on some bs
had the whole space by their balls
that was actually funny times
anyone was on thsoe spaces?
that was before o1
seems like forever ago
wow
imma become a grifter now
yeah i could nt use it all morning
but yo he actually might have serious connects to open ai, maybe he somebody kid that works their lol
X
this was wild predictions
keen beacon
https://x.com/iruletheworldmo/status/1870194357732814956
he just got lucky
keen beacon
but yo he actually might have serious connects to open ai, maybe he somebody kid...
he was doxxed i think, he isnt
keen fulcrum
Expected as they launch new servers
increasing their capacity for upcoming models
thorny drum
its just slow i think
keen beacon
weirdos
balmy mist
why wouldnt they, he was trolling us hard
had me missing work for his nonsense
why would you troll anyone?
keen fulcrum
People who dox will get rate limited
balmy mist
lol
keen beacon
trolling and doxxing are way different in terms of severity i think
balmy mist
trolling and doxxing are way different in terms of severity i think
im just saying anything is possible
so becareful what you do online
cause you can be doxxed easily
people kill people for less in our world, so im not shocked that anyone would doxx or do anything online
keen beacon
oh hello 👀
thorny drum
4o = 4.1 now?
balmy mist
strawberry man got me excited
keen beacon
4o = 4.1 now?
its live under 4o on chatgpt if its rolled out to oyu
its supposed to be named 4.1 (which it will be renamed soon enough i think), but its just an updated 4o
balmy mist
he just got lucky
u really think he got lucky tho?
thorny drum
wait wdyt o3 pricing will be
thorny drum
like o3 low was twice the cost of o1 pro
balmy mist
like o3 low was twice the cost of o1 pro
20k
keen beacon
lmao i got it again
is it not rolled out to you yet? or is this different
i just keep getting parallel gens for feedback on 4o
balmy mist
balmy mist
didnt we do that with quasar tho?
keen beacon
not rly
balmy mist
and hallucinations are not a bad thing, they are the key to dreaming and solving things we dont know, just need controlled hallucinations
keen beacon
i agree lol
balmy mist
lol
novel flame
Nvm I get it now.
keen beacon
lmao so they really do want to charge 20k
sonic tendon
@balmy mist train got delayed, so I'll probably be a bit late
sonic tendon
never ask a model what it is
I've had Google models claim to be Claude, and DS claim to be OpenAI
keen beacon
By the time they have a product like that they won't sell it
keen fulcrum
When will main o4 release
keen fulcrum
lmao so they really do want to charge 20k
Depending on competition
sonic tendon
When will main o4 release
as a rough guess, maybe the same delay between o3-mini and o3
novel flame
The 20k price tag snells like a classic con: “Buy this money printing machine, only $20k!! With it, you can print as much money as you like!”
If they had an AI which could create arbitrary novel discoveries warranting a price tag of $20k, they would keep it to themselves and print money themselves.
keen fulcrum
The 20k price tag snells like a classic con: “Buy this money printing machine, o...
If it makes me money I will pay it
sonic tendon
The 20k price tag snells like a classic con: “Buy this money printing machine, o...
yeah this seems like hype for investors more than anything else
balmy mist
next month probs
no way, i dont think they gonna release main o4, i thought they were going straight to gpt5 next?
keen fulcrum
Will hurt my portfolio if it tends to lose and not worth the cost
sonic tendon
Will hurt my portfolio if it tends to lose and not worth the cost
portfolio?
in what context
balmy mist
The 20k price tag snells like a classic con: “Buy this money printing machine, o...
how do you know that would sell tho?
novel flame
If it makes me money I will pay it
Except it’s payment up front. That’s how the con works
balmy mist
we woul dbe oversaturated with stuff liek that
keen fulcrum
portfolio?
20k is one chinese car
novel flame
Depending on competition
This. I would love to see this $20k product drop only to have R2 launch a week later with near-equal performance, forcing OpenAI to drop the price 100x
ember rapids
I hope they preview the benchmarks for full o4
keen beacon
they will in the o3/o4 mini livestream
torn mantle
recently im noticing some typo mistakes from gemini 2.5 pro
which is kinda weird
its the 3rd time already
balmy mist
<@367710025994731520> train got delayed, so I'll probably be a bit late
let me know when you get here, ill do it until you get here, im tryna work at the same time lol
X
im kinda sad that we are so early to these models
we are not suprised anymore
i gotta take a vaca for a month or so and comeback to agi lmao
torn mantle
im just here for the drama tbh
balmy mist
i have a feeling they are going to do something big like strawberry man said
torn mantle
hopefully google will pull out a battle model release
torn mantle
i have a feeling they are going to do something big like strawberry man said
who
that guy isnt reliable
never was
balmy mist
lmaoooo
torn mantle
hes a grifter
balmy mist
he did predict o4 in april tho
balmy mist
okay when is o5 coming?
torn mantle
after o4
keen beacon
lmao conveniently i now have to go 🙄
balmy mist
lmaooooooooo
keen beacon
cya in a bit gang
torn mantle
you know why its obvious?
because all other labs are catching up
and openai cant just rely on o3 series
keen fulcrum
Grok 4 will be great
torn mantle
they already have it implemented on deep research
grok 4
will be sh1t
dont get me started
omg
just the thought of using grok 3 again is making me so mad
balmy mist
i mean there is a market for predicting launches
torn mantle
wasted billions on nothing
keen fulcrum
Grok 3 reasoning is the best
torn mantle
based on leaks we will definitely have o3-full this month ( thats for sure xd )
torn mantle
what if sama pulls up?
wont matter tbh
balmy mist
wait
torn mantle
im interested in benchmarks tho
balmy mist
if it truly is the smallest ever that is good
1 mill tokens
wow
omgg they cooking
jk lol
but i am curious of the size of nano
bruhh
torn mantle
so not much MMLU diff
keen fulcrum
torn mantle
between 4o and 4.1
balmy mist
im turning this crap off
keen fulcrum
Latency
keen fulcrum
torn mantle
they are talking about pricing a lot
keen beacon
Omg
keen beacon
Lmao
torn mantle
im trying that prompt on gemini 2.5 pro
keen fulcrum
balmy mist
lmaoo the accuracy
keen beacon
Make it look better lol
If they compared to chatgpt 4o latest it would be the same ahahaha
keen fulcrum
Multimodal
keen beacon
Optimus
keen beacon
Should be same as 4o so 200b
balmy mist
lol
oblique flint
4.1 looks like quite a solid model already. Imagine if they add reasoning on top of it
brittle tiger
A little sus to claim sota on a benchmark when you're barely beating 1.5 Pro and 2.5 hasnt been tested
keen beacon
4.1 was quasar
balmy mist
ohhhh
keen beacon
4.1 mini is optimus
balmy mist
that makes sense, i udnerstand now
golden ocean
whats the point of 4.1 if theres 4.5
keen beacon
if it was really low they would showcase that tbh
Yes if it was 7b they'd advertise
balmy mist
nice
keen beacon
4.1 mini is a really good deal
torn mantle
awkward silence
keen beacon
It can sometimes beat 4.1
balmy mist
they were so scared lol
keen beacon
Apparently even if the benchmarks are lower
balmy mist
yeah optimus is a good model
oblique flint
nano is exactly the same price as 2.0 flash hmm
balmy mist
mini is a no brainer
torn mantle
google will drop their new model probably
sonic tendon
lmao conveniently i now have to go 🙄
same
keen fulcrum
keen beacon
what happened
balmy mist
lmaooooo
keen beacon
Did they lol
keen beacon
I'm not watching
keen fulcrum
4.5 will get deprecated in the next 5 months
keen beacon
Lmao
torn mantle
he said 3 months
balmy mist
what
keen beacon
They abandoned it fr
keen beacon
"We will also begin deprecating GPT‑4.5 Preview in the API, as GPT‑4.1 offers improved or similar performance on many key capabilities at much lower cost and latency." lmao..
keen beacon
That was quick lol
I wouldn't have thought they'd admit defeat publicly that quickly
balmy mist
degenerate behavior lol
torn mantle
they look a bit anxious
raven void
Lmao
torn mantle
those are nice improvements if its true
keen beacon
wow gpt4.1 mini looks really good for its size
raven void
Deprecating 4.5 🤣
raven void
my bad
keen fulcrum
Gpt 4.1 free in windsurf for the next 7 days
keen beacon
They should've abandoned 4.5 completely tbh
Not even release it. It made them look bad
oblique flint
cursor is glued to claude lol
opaque adder
im just confusd why they go the opposite way
it should be 4.1 then 4.5
not 4.5 to 4.1
makes no sense to me
keen fulcrum
balmy mist
i miss when optimus was free lol
keen beacon
weird that chatgpt-4o is still significantly more expensive..
keen fulcrum
i miss when optimus was free lol
What was optimus
balmy mist
mini
opaque adder
weird that chatgpt-4o is still significantly more expensive..
they dont want data for that model
keen fulcrum
You couldn't use it without putting a balance on openrouter
So no o4 mini and o3 ?
keen beacon
trying gpt-4.1 in windsurf now
olive mesa
is this a google model
keen beacon
willing to bet claude is still better tho
balmy mist
i feel bad for gpt 4
keen beacon
It's gpt 4 turbo they call it gpt 4 on the website for some reason lol
keen fulcrum
i feel bad for gpt 4
Initiating shutdown protocol. Try to override 4.1
keen beacon
so now we wait for tmw or is it coming on wednesday?
likely thursday
balmy mist
likely thursday
bruhh gg
yeah imma take a vaca for a month and be bac in may let me know if we got agi by then
keen fulcrum
Yummy
novel flame
A little sus to claim sota on a benchmark when you're barely beating 1.5 Pro and...
54.6% on SWE-Bench Verified is pretty good. In fact, it's ALMOST as good as Amazon Q, that other top LLM that we all definitely consider a top coding model, which scores 55% 🤣
torn mantle
X
novel flame
Um..... something is not right there. GPT-4o definitely supports image outputs. I seen it wit' ma own eyes!
keen beacon
via the api
novel flame
via the api
No native image outputs over the API? Boo, Sam Altman, Boo!
oblique flint
3.7 sonnet still mogs it
How does it compare to 2.5 pro in windsurf?
fleet lintel
benchmark wise, how these new OAI models are?
torn mantle
lets try it
vast vortex
thanks
keen beacon
is gpt-4.1 not available on chatgpt.com now?
"Note that GPT‑4.1 will only be available via the API. In ChatGPT, many of the improvements in instruction following, coding, and intelligence have been gradually incorporated into the latest version of GPT‑4o, and we will continue to incorporate more with future releases."
vast vortex
tried to use on openrounter.ai, result in "unsupported country"😆
thorny drum
thats why they compared with the november version i think
fleet lintel
thats why they compared with the november version i think
why? 🙂
novel flame
Actually the Windsurf announcement makes me wonder if I should give it another try.... Some months back (around the time Windsurf launched / changed its name) I did a review of all the coding IDEs, and Windsurf had the best 'flow' but somehow had the worst outcome (a great UX could not make up for the fact that it ultimately generated bad, buggy code). Cursor wrote good code and had decent UX, though a few annoying gotchas that you had to work around, like having to specify the exact files to include in the context. The standout feature of Cursor was the 'suggested next edit' autocompletion, which was practically magical. Cline had the best UX and resulting code in general, but lacked the magical 'suggested next edit'. Aider was a joke compared to the others at the time. Continue had fallen behind the pack.
I have used Cline since, and when Copilot Agent Mode arrived, I tested that one too; actually I get better resulting code with Copilot Agent Mode than I do with Cline (both using 3.7 Sonnet), not sure why that is, and I still prefer Cline's UX, but the tool orchestration must be slightly better in Copilot Agent Mode.
But....... if Windsurf has gotten its act together and even have free GPT4.1 then maybe I need to give it another try.
torn mantle
Actually the Windsurf announcement makes me wonder if I should give it another t...
windsurf is pretty good
they are serious about their product
unlike microsoft with copilot
slow with updates/features
novel flame
unlike microsoft with copilot
I was as surprised as you that their Agent Mode wasn't awful.... considering how their version of 'suggested next edit' is the worst thing I've ever activated in an IDE, and I have used Eclipse, IntelliJ, emacs, vim, and Visual Studio
torn mantle
gpt 4.1 is no that great at web dev
raven void
"Note that GPT‑4.1 will only be available via the API. In ChatGPT, many of the i...
What's the reason for this? 🤔 Is it because 4.1 doesn't have audio and image out?
keen beacon
deepmind are coming for claude with webdev but their models aren't good enough at following structure and calling tools for practicla coding yet
keen beacon
What's the reason for this? 🤔 Is it because 4.1 doesn't have audio and image ou...
the second sentence answers your question
torn mantle
lets see how it does at vision capabilities
keen beacon
gpt-4.1 is basically just all the gradual improvements they've made to chatgpt-4o spun off as a separate api model
tribal aspen
Nightwhispher and dragontail incoming
tribal aspen
Nightwhispher and dragontail incoming
How to use these models?
fleet lintel
I still dont knwo whether to be exicited about these new OAI models or not? Are they better than 2.5 pro?
novel flame
deepmind are coming for claude with webdev but their models aren't good enough a...
...but do we know this for sure? Nobody seems to have spent enough time with nightwhisper to say conclusively if it dunks on Sonnet or not
keen beacon
one correction tho: i was wrong on optimus prime apparently its the full gpt 4.1 model too. not sure why it scored significantly lower tho
and on aider
keen beacon
...but do we know this for sure? Nobody seems to have spent enough time with nig...
i am speaking about deepmind's models right now
oblique flint
...but do we know this for sure? Nobody seems to have spent enough time with nig...
I mean you can't really test its function calling capabilities in lmarena
keen beacon
not about nightwhisper and their upcoming ones
because like you said
we don't have enough info on them
torn mantle
PLS
LET IT BE NIGHTWHISPER
PLSSSSSSSSSSSSSSSS
@keen beacon DO SOMETHING ABOUT IT
talk to them
keen beacon
LET IT BE NIGHTWHISPER
let what be nightwhisper
keen beacon
oh
keen beacon
it's 2.5 flash next
thorny drum
is google shipping today?
torn mantle
ask them
fleet lintel
LET IT BE NIGHTWHISPER
nightwhisper is not happening. it is replaced by dragontail
torn mantle
oh no
novel flame
i am speaking about deepmind's models right now
ah..
torn mantle
nightwhisper is not happening. it is replaced by dragontail
i had that thoughts as well
hopefully its not true\
tribal aspen
nightwhisper is not happening. it is replaced by dragontail
Nw is still better
keen beacon
there will be more google anon model drops on the arena in the coming days
just wait
novel flame
2.5 flash followed by an update to 2.5 pro
Seems plausible that nightwhisper/dragontail is an update to 2.5 Pro, no?
fleet lintel
Nw is still better
i think NW was just a coder model and they doesn't want to release only coder model. they want to have all the benefits in normal pro/flash models.
plain zinc
I think nightwhisper is the last Google card he will save for an emergency.
keen beacon
Seems plausible that nightwhisper/dragontail is an update to 2.5 Pro, no?
if my sources are correct, dragontail is 2.5 flash
with a high thinking budget
as for nightwhisper
fleet lintel
if my sources are correct, dragontail is 2.5 flash
dragontial is not flash... for sure
raven void
what is riverhollow?
keen beacon
dragontial is not flash... for sure
it is
oblique flint
No way it'll still be as cheap as current flash if it's that good
raven void
what is riverhollow?
It was worse than 2.5 pro in my test