general

[keen beacon]

what would you name this new retrained o3

o4 then benchmark it with the juice u used for o3... i guess?

[ocean vortex]

be explicit that it's not just "o3". I think that's fairly obvious, no?

[zinc ore]

They can't do that again because once they actually release people will test it and get vastly different results. People already picked up on what they did, so they risk harming their brand continuing to do so

So caught up on trying to generate hype compared to their competition that they just make highly misleading test results

But anyone perceptive can pick up on

[ocean vortex]

sure but the point is they knew to begin with that they aren't testing just the model o3. But rather their version of a "pro" setup which happens to be based on the new model

that was never gonna see the day of light as just "o3"

[keen beacon]

how would you do that though, its already a mess

the o3 in december is much less stronger than the retrained o3 with the new 4.1 base we have now, at least served with the same compute. so it does mean something

[ocean vortex]

o3-pro-preview, or whatever else...

[keen beacon]

still misleading since o3 pro is not gonna be served with that level of compute

it depends on the benchmarks though, if the o3 pro product they're gonna release performs impressively/or close compared to the initial o3 results, they could name it like that

[ocean vortex]

which would all line up if that version from december wasn't scoring higher, but it is... And that's the problem. It's worse but it was scoring higher, because it wasn't just o3... Like I really think this is obvious they didn't just accidentally do it LOL

[keen beacon]

at the time do u think they already had plans to retrain o3/etc

its misleading but i dont think its intentionally malicious

[keen beacon]

they could make it score higher if they used as much compute as they did before, but like the other guy said its a bad idea and this time it would be intentionally misleading imo

[ocean vortex]

then

now

[keen beacon]

arent they using cons@64 or smthing like that, the grey area which i assume the score is for?

[keen beacon]

this is pass@1 i think

[ocean vortex]

iirc it was pass@1 for o3 supposedly

[keen beacon]

whats the grey area then

[ocean vortex]

and cons@64 only for o1

[ocean vortex]

I'm not sure, maybe it was 🤔
this still doesn't add up though assuming preview wasn't using the new base model

[keen beacon]

even without the juicing it couldve just been doing even more massive and unreasonable chain of thought by default compared to the optimized new o3, to reach those scores

[fringe carbon]

i believe light grey is if it's reprompted

[tall summit]

What is cons@64, you might ask? Well, it’s short for “consensus@64,” and it basically gives a model 64 tries to answer each problem in a benchmark and takes the answers generated most frequently as the final answers.

[ocean vortex]

it can not be misleading if they are explicit that this isn't just o3. Like it's really not f'ing hard to do. You can literally name it "o3 + extended cot synthesis"

just like they are listing tools now, this isn't all that different

[keen beacon]

you want them to go back and edit all the videos/material after they changed plans? its just not realistic. they weren't in development far enough to know how they'd serve it so how would you make that distinction without knowing the future

they changed it now to o3-preview at least

[ocean vortex]

No I think my point is that now it's obvious they knew what it was and what it's going to be released as to start with. This wasn't "a mistake" lol

[keen beacon]

they knew they were gonna retrain o3 at the time lol?

[ocean vortex]

?? no they knew they were not gonna release o3-pro as "o3", and that the model to be actually released gonna be much closer to o1

[keen beacon]

but its not even o3 pro

[ocean vortex]

whatever, that is clearly not the point here lmao

I'm just calling it as such as it's the closest match to what it was

[keen beacon]

if they released that version of o3 without the juice and changed the reasoning effort levels would you be ok with it?

[ocean vortex]

I would be ok if it wasn't named "o3" in the graphs when they knew that model is nothing like "o3" to be released. And once again, you can't be serious thinking this was a mistake.

[keen beacon]

it isnt a mistake

bro?

[ocean vortex]

well then why are you asking if they should go back and edit it lol

[keen beacon]

it was intentional at the time but they didnt know the future and committed to it. it is now misleading in the present

[ocean vortex]

they did it deliberately, what's there to edit? 👀

[thorny drum]

its pretty lame but on the tier list of gaming benchmarks its only in the middle

i think anyone who cares a ton about benchmarks probably watched the stream and knew they used ungodly compute on them

[keen beacon]

its deliberate but they have no idea about how misleading it would be in the future lol

[thorny drum]

i personally think its cool to hear about these companies super powerful internal models

and i wasnt expecting an o3 that was $1/token

[ocean vortex]

Ok imagine yourself in their shoes. You have a model x, but instead of running it normally, you use pro-like system to boost the scores way up. But then still leave the name exactly as it was, model x, with no additions. Now in what universe would you think this can be not misleading? LOL

[keen beacon]

they were early in development they might have some idea but it could change drastically

[tall summit]

guys

it doesnt matter

just wanted to tell you

[raven void]

new o3 is pretty good, they've reduced the size while keeping similar performance to original o3

[keen beacon]

we like being stun locked 🤣

[tall summit]

pretty sure the consensus is o4 mini is better not "often" or even "sometimes" but more than just occasionally

[tall summit]

i love pointless arguments but when youre on the outside its easy to see how annoying they are

[ocean vortex]

nah I think it was obvious since the beginning this isn't competing directly with normal o1. Well for them at least, for us... even I didn't know they used something similar to pro for those scores. I was impressed back then but now I kinda feel scammed. Which is the whole reason of why I even wrote it.

[keen beacon]

theres nothing wrong with it lol

[tall summit]

^

[zinc ore]

Pro is just longer think for base o3 (from tech crunch)

[keen beacon]

its still impressive though for any system to reach that level even if it was juiced

[ocean vortex]

it's mostly hype building marketing. If you are gonna do something like that well then do what google did

AlphaGeometry

and the fact that they didn't even have new base (4.1) back then makes it worse, not better tbh
At least now with an improved model it is not 1 million miles away...

it's still an interesting good model, but yeah... the direction openai is heading and what have they became lately comparing to like 1.5 years ago is something I really do not like

don't get me started on the pricing too... they finally reduced the price but only for the new model, o1 is amazingly now more expensive than o3

[ember rapids]

We need Claude 4

That’s what I’m most hyped for

[ocean vortex]

with the releases of both 4-turbo and gpt4o, their pricing was essentially unmatched (good)

then o series started and all hell broke loose with them going to town on pricing lol

[thorny drum]

i feel like these new models r pretty good

whats wrong with them

[ocean vortex]

yeah it was a very novel idea. Preview had some bugs and was rough with later o1 being much better imo but this idea enabled good progress for sure

[ember rapids]

The most impressive aspect of o3 for me so far is its ability to generate good ideas

[hardy pecan]

Is o3, medium or high in the lmarena?

[glass arch]

base o3 is here

[dapper storm]

Yea but what level of thinking is that

[glass arch]

we have no freaking clue I guess

[keen beacon]

i like how o3 talks

it's got more of a personality than o1

[ocean vortex]

if it's not specified that's medium

[glass arch]

on chatgpt.com, it actually thinks with your name for some reason

[ocean vortex]

yeah I noticed that too. I found that odd since I have memory and CI disabled and it somehow still knows lol

[glass arch]

here is what it does for me

[elder rapids]

ye good vibes

[glass arch]

it also writes with a lot more emojis for some reason?

[elder rapids]

it still doesn't shine like 2.5 pro though tbh

[glass arch]

I think openai's models are the friendliest

gemini is quite annoying to talk to

[ocean vortex]

🧐

[elder rapids]

might be just getting used to

in my experience, the Gemini family is more of a customizable, but baseline neutral family

[ocean vortex]

it's their fine-tuning. They nearly fixed it but not entirely

[elder rapids]

I don't think it's a flaw lol

I think that's the intention

[glass arch]

it seems bad at understanding social norms
I hooked it up in such a way that me and some friends could speak to it in a discord call, and it is so hard to talk to

[elder rapids]

oh ikwym

[glass arch]

the openai models are better at social norms though

[ocean vortex]

"you are right to point that out!" - I hate these from gemini lmao

[elder rapids]

when I give it an example

and the example is using "you"

it thinks it needs to address it as the AI itself

rather than assuming it's a general example

[keen beacon]

i think 2.5 pro is less sycophantic

[ocean vortex]

yeah it's acceptable levels now. But sometimes still annoying

[glass arch]

if there was an AI model that I would trust to not destroy us if it got AGI, it would be chatgpt (any one of them tbh)

[elder rapids]

i wouldn't say Claude

it has a very good feel, but it's almost superficial

so yeah chatgpt prob

[ocean vortex]

it's the most stable behaving one for sure

[elder rapids]

to me Gemini has more of that meta intelligence, and that's what's deciding it's morality

rather than having instructed morals

[glass arch]

claude is actually very good at language (by this, I mean it can speak toki pona the best)

[elder rapids]

yep

[glass arch]

I did a few tests with the arena, and claude comes out on top for all toki pona conversations

[elder rapids]

very contained

[elder rapids]

ye that's what I look for in models

3.5 sonnet has always been clearly the best

but that's why I like 2.5 pro so much, because it does the same thing when you really want it to

[glass arch]

I want to run a test playing a game called "keep talking and nobody explodes"

I did it a few days ago with o3-mini-high

which of the new models should I pick?

[elder rapids]

keep talking and nobody explodes?

[glass arch]

yes

the one with the bomb

[barren prairie]

I can agree +++
But it improved a lot (I talk to Gemini flash thinking and it seems cute 😆😆) but chatgpt is more natural

[ocean vortex]

sounds like some nsfw rp lmao

[elder rapids]

I'm gonna explode

[ocean vortex]

💀

[glass arch]

nah, more like yelling at the AI for not understanding the maze puzzle

[elder rapids]

flash thinking was hella neurotic, but it was super adjustable

damn if y'all feel this way about Gemini

you guys gotta know the strats

theres some insane keywords that these models react to

[glass arch]

which

I gotta try now

[elder rapids]

Gemini and Claude

with "cut and cleanly"

or 4o, with "pragmatic"

[glass arch]

openai is reaally good at naming huh

4o

o4

[elder rapids]

feels like they're doing this on purpose

[glass arch]

what does the "o" in 4o mean again?

[elder rapids]

if they release o5 after releasing 5o

[elder rapids]

Omni

ngl when the first 4o was "anonymous-chatbot" it felt so fresh

like a clear step above the other models

[glass arch]

it was truly the best

[keen beacon]

anonymous chatbot is chatgpt 4o latest. are you talking about the i am good gpt2 thing or the first chatgpt 4o latest variant

[glass arch]

now I have to switch off of 4o because it's not even worth giving orders to

[elder rapids]

the first chatgpt 4o variant pre release

[keen beacon]

oh

[elder rapids]

I am good gpt2 was another pre release variant

of 4o

[keen beacon]

yea ik

[elder rapids]

and then "birdie" was athene 70b

which was lowk a secret

y'all didn't know about that one

the best model at that time behind 4o

[glass arch]

what is your guys' go-to prompt for these new models?

I ask for pacman in pygame

and o4 did really good at this

[elder rapids]

I usually give random puzzles

and then ask older models to solve the same ones

o4 mini is good at puzzles

o3 is horrible

o3 is super smart outside of these rigor subjects tho

[glass arch]

yeah for my use cases, these models are very optimized

[ornate stump]

Does anyone know if they’re planning to improve the voice mode? It’s still the least utilized and developed feature considering its potential.

[glass arch]

also, I don't think an AI is going to be fooling a human any time soon

[elder rapids]

I think until the competition gets stronger in voice

they're not gonna do that much

sesame is cool and all

[glass arch]

gemini is CRAZY at video

[elder rapids]

but the distribution is bad

since it's not a major distributor like Google

[elder rapids]

yep

[glass arch]

I streamed me and the boys playing the binding of isaac, and it was able to identify stuff

[elder rapids]

it's in its own league as far as understanding videos

[glass arch]

from what I see on x, the everything app, grok can understand videos pretty well too

[ornate stump]

Yeah, but I mean, I don't need the AI to flirt with me or be that assertive. But the voice mode in Gemini or OpenAI is stupid you not only can't have a conversation with it, but you also can't use it under any circumstances.

[glass arch]

it works for me mostly

it likes to end its turn too quickly though

well, not gemini

[elder rapids]

wym?

[glass arch]

gemini talks way too freaking much

[elder rapids]

since when can grok understand videos

[glass arch]

it seems like people can mention grok under a video and it will respond

[elder rapids]

I thought that was finding references

and transcriptions

rather than watching the video

[glass arch]

oh yeah probably

[elder rapids]

since in that case, it doesn't do very well

[glass arch]

I don't know what grok does

[elder rapids]

I've seen it used on memes and stuff

video memes

[ornate stump]

Gemini Live can’t search the web, though, right?

[elder rapids]

and it doesn't do very well understand what it even is

[glass arch]

it can't(?)

[elder rapids]

I think most recently

it may be able to

before no, now I think so

[keen beacon]

hmm sometimes 2.5 pro stops thinking on aistudio, especially with extremely long chats, not sure if its just a visual thing (it starts streaming with the same delay as if it were to start streaming a thought process), but if i prompt it to keep thinking itll usually pop it back up 🤣

[elder rapids]

Google might solve that soon

[elder rapids]

wym? like, the text just suddenly pops back in the message it just finished after you send something

[ornate stump]

I've just tried it and it can search the web now

[elder rapids]

oh ye this happens to me

same with it going outside of its thinking box

ye, I think it's a visual bug

[ornate stump]

Do you have any news/articles on this, or are you just speculating based on what they’re doing right now?

[elder rapids]

based on what they're doing right now

since they're focusing on multimodality

[real totem]

Bro

I just came back

And I see this

o4 mini and o3 and 4.1

[elder rapids]

the streaming delay?

[real totem]

Ain’t no way all of these released so fast

[elder rapids]

I'm confused

[real totem]

Which is the best

One

[elder rapids]

o3 for general tasks, o4 mini for coding tasks/puzzles, 4.1 is API only

[real totem]

Are they better

Than gemini 2.5 ro

Pro

[elder rapids]

and its good at small coding tasks

[ornate stump]

meh

[real totem]

Google’s next release

Is finna be crazy

[real totem]

It’s better

For me

Not by a lot tho so ye

[elder rapids]

o4 mini is a little better than 2.5 pro at pure coding tasks, from the benchmarks, o3 isn't really that much better, or even at all. It seems to fail things 2.5 is getting really easily, and 4.1 isn't a reasoning competitor

[real totem]

Oh

Makes sense

[keen fulcrum]

What is the difference between o4 mini and o4 mini high?

[elder rapids]

high/pro means it thinks longer

[keen fulcrum]

Simple to understand thanks

[ornate stump]

Gemini 2.5 got a lot of people to switch when they released it, not sure if it's the same with o3

[barren prairie]

Sometimes I think that Gemini 2.5 is overthinking ...sometimes , the Flash thinking make good answers while the pro sucks 🙂 I still donno why

[balmy mist]

https://x.com/koltregaskes/status/1912645196866793854

Kol Tregaskes (@koltregaskes) on X

ChatGPT Plus, Team & Enterprise subscribers get:
- 50 messages a week with o3
- 150 messages a day with o4-mini
- 50 messages a day with o3-mini-high

ChatGPT Pro subscribers basically her unlimited.

This is great but I'm questioning the value of Pro with such high Plus limits.

X · Apr 16, 2025, 11:12 PM

[elder rapids]

I think this is inherent to the prompt itself tbh

[real totem]

Yeah

They wont switch from gemini

Cuz its free

[elder rapids]

2.0 flash thinking is a true overthinker

[real totem]

And o3 is paid

[elder rapids]

at least from the months I used it

[ornate stump]

This is largely based on what you're doing, but if overthinking is an issue for you, maybe the O4 Mini will suit you better.

[keen fulcrum]

Gemini 3 will be groundbreaking yet again

[barren prairie]

It is just some QMC and questions

[elder rapids]

depends, the limits can be really restrictive, and o4 mini is cheaper than 2.5 pro

[keen beacon]

omg

o4 mini limits are pretty decent

[elder rapids]

although that's beneficial ONLY for coding

[barren prairie]

Nahhh at all

[elder rapids]

oh wait

[real totem]

Yeah

[elder rapids]

you're talking about on the apps

[real totem]

I just switch accounts

When I get restricted

[real totem]

The ai studio

[elder rapids]

ye

[real totem]

It’s 50 messages I think

It’s pretty good

[elder rapids]

if you're talking about AI studio vs chatgpt plus

[real totem]

Yeah

[elder rapids]

I don't think that's true tho

that's just for API I think

I've sent more than 50 per chat in an hour

ye the difference is big, thinks a lot

wonder how good 2.5 pro would be with that crazy length

[keen beacon]

hmm i was checking to see if the 2.5 pro thinking bug was visual: 5.5s to the first token (first token in thoughts) and 6.1s (2.5 pro immediately going into a response), it doesn't seem to be visual

[elder rapids]

would this not be directly visual

I'm confused on wym

[keen beacon]

if it were doing the thought process and skipping the thinking block

u would expect the first token to be delayed on the response

but its the same

delay with or without thoughts

so it doesnt seem to be visual

[elder rapids]

alr I'm lost but I hope you figure it out

sorry bro 🙏

[keen beacon]

its fine im explaining it confusingly

[leaden palm]

man how late is later...

[keen beacon]

u can select the models in direct chat if u want btw

both o3 andn o4 mini

[leaden palm]

o cool

time to give it my (only?) eval question

[keen beacon]

the mechanism if it's not visual i guess is because they exclude thinking blocks, i guess when u reach a certainn amount of turns the model's tendency to start with a thinking block isnt there since all the past and numerous amount of turns had no thinking blocks

[leaden palm]

o3 was... alright? the style was bad:

• manually wrapped its text (presumably trained on too much TeX)
• disregarded "mtok", took it to mean ktok, messing up the calculations
• called phi 4 multimodal "φ‑4‑mm"
  but once i got past that, i did enjoy its discussion of the math and attention and economics

meanwhile o4 mini just spreading misinformation

[keen beacon]

did u set it to 0 temp btw

[leaden palm]

i left it at default 0.7

[keen beacon]

yea try using 0 temp

it seems its very sensitive

[leaden palm]

i'll try again but i'd be surprised if that improved it more than regenerating would

what the hell is this 😭

okay it was definitely better

i still like the style of the o series

[elder rapids]

they're all o series

😭 🙏

[leaden palm]

4.1:

[leaden palm]

ehh actually it lost some of the math, so it might just be regeneration quality

[keen beacon]

it depends on how its ingested

but 2.5 pro i think is probably better

[elder rapids]

2.5 is absolutely better

it's not close

the Gemini models are basically made for that

tbh

as far as my testing goes

uploading anything to Gemini is far superior, it's audio understanding, it's video understanding

nah it goes for all the models

just try 2.5 pro with this stuff

it's seriously crazy

it's basically perfect

ye, but tbh its just different approaches

I like reading other models thinking

ye

can't say that's much of a problem tho cuz that's what benefits Gemini more in general

[glass arch]

I just played ktane with it

[plain zinc]

How do you like o3? o4-mini high?

[elder rapids]

o1 pro is more of a academic

[plain zinc]

Tell me, please.

[glass arch]

I am gonna upload my video of this session to #ai-creations when youtube finishes processing it

[elder rapids]

yep, I still like 2.5 pro the most tho

it's just the way I prompt it

and how it allows itself to be prompted

that's just super unique to me tbh

so the feel that I like in other models, I can definitely replicate in 2.5 pro

and then boom

no need for them

[drifting thorn]

Btw is the o3 using tools in the arena?

[glass arch]

ok I shared my video

[balmy mist]

nahh

[drifting thorn]

Sad

Waiting for Deepseek R2 that integrates MCP tools in chain of thought

[plain zinc]

Dragontail, riverhollow, and shadebook are gone.

Shadebrook

[balmy mist]

lol

[glass arch]

I like how chatgpt just takes the verbal abuse I give it

[balmy mist]

you better becareful abusing chatgpt

it got memory now

[glass arch]

I lobotomized that part

[late path]

does anyone else feel like o3's chatting style is strangely similar to r1...?

[elder rapids]

u sure?

that didn't seem to be the case

so it must've happened within the last hours

by the way, I'm not sure about o4 mini being cheaper in practice

seems to be p4p not on 2.5 pros level

same goes for o3

[glass arch]

when are we gonna be allowed to use 4.1 in the app

[elder rapids]

you're not I think

[glass arch]

!!!!

bruh

[keen beacon]

Chatgpt?

[elder rapids]

4o seems to be the absolute replacement

[glass arch]

yes

[keen beacon]

It's already live under 4o lol

[elder rapids]

wym?

in the chatgpt app?

[keen beacon]

They're continuing chatgpt 4o latest which is the chatgpt 4o model on the app, which is on the same base model as 4.1

[thorny drum]

4o != 4.1

[keen beacon]

It's mega confusing but yeah

Chatgpt 4o latest uses the 4.1 base model/has had it for a while before 4.1 released

The new base model where it was continued pretrained and has a newer cut off

[glass arch]

that is so stupid!????

[keen beacon]

Openai tries to make naming as confusing as possible

For some reason lmao

[elder rapids]

I mean not rly, as long as you consider -o as a variation

which in this case, it would still be 4.1 → 4 → 4 omni

[keen beacon]

There are small differences though the chatgpt 4o tune is slightly different though even if it is on the 4.1 base model, it is more human preference aligned. But model performance is largely the same

[elder rapids]

since 4o previously probably still used gpt 4, as a distill

so it would be 4.0 Omni

now it could be 4.1 Omni and therefore 4o

[glass arch]

can't wait for 4.5o

[elder rapids]

could probably still be called 4o

[glass arch]

will it be better at language?

[plain zinc]

Yes

I don't see them anymore.

I only come across 2.5 pro, 2.0 flash thinking

[elder rapids]

could be just unluckiness

[plain zinc]

Take a look for yourself too

You won't get them either.

It's strange that Legit didn't reveal anything about their loss.

[keen beacon]

dont really know who legit is but i guess he scrapes web dev arena

in the metadata there its still live

dragontail/etc

[plain zinc]

He's one of the main leaks of some cool stuff.

[plain zinc]

Did they turn it off for an update?

Did you see him just now?

Or recently?

[keen beacon]

idk if its in the arena or not, but its still in the metadata of web dev arena:

[keen beacon]

probably why his bot/etc didnt catch it

if it was removed off the main arena

[elder rapids]

just got River hollow lol

[keen beacon]

i just got dragontail in webdev arena btw: (so the metadata for web dev arena is right at least for web dev arena)

[plain zinc]

._.

It can't be that I'm unlucky._.

I know for a fact that there were none.

What the...

[elder rapids]

sometimes I get an extremely long period of not getting certain models

like, I haven't even gotten o4 yet

or o3

[plain zinc]

Is that what this is about?

Why is this happening?

Update?

[elder rapids]

seems like a balancing thing or smth

for the arena

could be that they actually did take them down

just for a little bit

for an update

we'll never know

[plain zinc]

and I get it._.

[elder rapids]

oh damn I just got o3 vs 2.5 pro

crazy

[plain zinc]

...

What...

[elder rapids]

😭

[plain zinc]

10 new models have just been added.

It was 96 and became 106

[elder rapids]

kinda excited for 2.5 flash

[keen beacon]

it was supposed to come out a while back

the model was added on the sdk

[elder rapids]

ye

[keen beacon]

0409 or smthing

[elder rapids]

that wasn't too long ago tho

[keen beacon]

with thinking budget

[elder rapids]

oh fr?

damn I hope it's cheap and smart

that would be crazy

[keen beacon]

yea despite the sdk name it didnt come on the 9th

they delayed it for some reason

[elder rapids]

if it's a jump from 2.0 thinking, it's gonna be good

[keen beacon]

oh 2.5 flash is gonna be good

[elder rapids]

although it was trash at coding

it was for some reason INSANE at some things

tbh bro

I keep thinking back

the old times

where athene 70b was the hidden gem

that pure RLHF model was sick when you prompted it for the right things

[elder rapids]

I hope the long context is somewhat maintained

if it's anything like 2.5 pro too

it's gonna go crazy

[hardy pecan]

https://simple-bench.com/

O3 beat out 2.5 pro

[plain zinc]

Yay!

And more...

HE's become insanely slow💀

[elder rapids]

damn I was expecting more

[plain zinc]

I had to wait 3-5 minutes for the code generation to finish.

[elder rapids]

53% for high

damn I said o4 was gonna get low too

and o3 was gonna get 50+

[hardy pecan]

Hey it shows how long 2.5 pro is too, since everyone is impressed by Geminis offering. 50s on that benchmark is great, and actually SOTA

[elder rapids]

but I thought it was gonna be higher than 2.5 💔

[hardy pecan]

yeah the smaller mini models tend to suck at that benchmark

[zinc ore]

Thing is how many tokens did 2.5 use?

[elder rapids]

there's no way for them to have optimized that

so it's gonna be baseline what you see

rather than a high or medium variant like o1, o3, and o4

[brittle tiger]

https://x.com/OfficialLoganK/status/1912690850292768951?t=odVru2H-4Tpr31466k_JcA&s=19

Logan Kilpatrick (@OfficialLoganK) on X

Gemini 🤔

X · Apr 17, 2025, 2:13 AM

[zinc ore]

Exactly

[brittle tiger]

Flash thinking tomorrow

[elder rapids]

it's the same reason why openAI models seem to dominate benchmarks

but then not do well in practice

[plain zinc]

Yessss

I already thought that Google wouldn't show anything this week and would give OpenAI all week.

[zinc ore]

I just wish I could more readily compare the tokens/thinking time being used

[plain zinc]

But they decided to go all in.🔥

Bold and fast

[zinc ore]

Makes it harder to tell if these little 3% gaps are meaningful or exaggerated

So I don't like there's just a persistent mystery now

[elder rapids]

ye

[thorny drum]

wdyt flash thinking means given pro is already a thinking model

[elder rapids]

but that's why it's still pretty much clear 2.5 pro is leading

[thorny drum]

or is this just flash

[zinc ore]

2.5 flash is a hybrid/unified model, so thinking model yes, once it releases

[elder rapids]

regardless of narrow task advantancment like o4 mini, or the general ability of o3

it seems like it's inherently a weaker model

due to the fact it literally has to think more

to achieve similar output

[brittle tiger]

I was really impressed by o3 but mostly for the tool use and UI. I think 2.5 Pro which was rushed out the door can match it and probably remains a dev fave. Think 2.5 flash will be better than o4-mini but just a hunch

[balmy mist]

https://x.com/OfficialLoganK/status/1912690850292768951

Logan Kilpatrick (@OfficialLoganK) on X

Gemini 🤔

X · Apr 17, 2025, 2:13 AM

no way

[elder rapids]

ye they've been doing this

for a while

it's a nice hint

[zinc ore]

Good guess is 2.5 flash, I think they might drop a coding model too

[elder rapids]

could, that would be pretty cool

that's what the discussion is about

[balmy mist]

i hope its nightwhisper

[elder rapids]

the fact it's so little above 2.5 pro while likely thinking so much more

going for all the other benchmarks as well

[elder rapids]

for example, this benchmark

[ivory schooner]

我正在学会等待Behemoth ......

[elder rapids]

o4 mini costs more than 2.5 pro

[keen beacon]

release tomorrow then

[keen beacon]

LOL

[elder rapids]

sorry I already predicted other things about these models today

I deserve the glaze

not you

[keen beacon]

buddy

who are you

[elder rapids]

2/2 you're just 1/1

[keen beacon]

do you have internal openai model access and connections?

no

pipe DOWN

🙄

[elder rapids]

I take a look inside without permission

I'm just like that

[keen beacon]

🤨

[leaden palm]

its just gonna be 2.5 flash...

[elder rapids]

sorry you need to be granted access

[leaden palm]

(although i'd be happy to be surprised)

[keen beacon]

2.5 flash is of course the minimum but there may be another model update

[elder rapids]

ye

[zinc ore]

I'm guessing they feel confident about 2.5 flash

[leaden palm]

maybe

[elder rapids]

that would be cool

[balmy mist]

you really called it lol

[leaden palm]

seems openai is the only one whos got tool use in thinking though which is EXTREMELY weird

like it's just a ui level thing

how you call your own inference api

[elder rapids]

exactly

it becomes so much less of a jump

[keen beacon]

it does feel kinda weak on their part to release 2.5 flash, a worse model, when openai have just mostly beat 2.5 pro

[elder rapids]

while 2.5 is base

[keen beacon]

i think they'll wanna flex

[balmy mist]

wait you have connects?

[leaden palm]

yes
it's just whether or not you start another <think> after a tool message

[elder rapids]

exactly, so they might just release a coding model

[zinc ore]

This is hilarious because they're just benching the same model over and over again, so naturally it'll get similar scores, just incrementally worse/better based on thinking time

[leaden palm]

what do you mean you don't know about that

[elder rapids]

@keen beacon is my connect

[brittle tiger]

If it clearly bests o4-mini they would def like that

[balmy mist]

google about to do openai dirty lol

[keen beacon]

yeah but i kinda doubt it

[balmy mist]

y they cant let openai have one week

[keen beacon]

o4 mini is very strong

[elder rapids]

ye

[late path]

this is our frontier model💀

[keen beacon]

in most cases it matches o3 and in some it beats it

[elder rapids]

but if 2.5 flash is narrow

[keen beacon]

i dont think flash has the chops for that

[elder rapids]

and anywhere near o4 mini

then o4 mini has to be cooked

the context + the price

at reasonable output

[balmy mist]

it would have to be nightwhisper right?

[brittle tiger]

I really dislike tests like this and token issues

[balmy mist]

they were waiting on this

like they probably been tesing o3 and o4 mini all day

[topaz peak]

really disappointing that o3 fails the bucket test damn

[balmy mist]

google is funny

i feel bad for open ai

they might have to releas o3 pro next week

[elder rapids]

alr chill

o3 pro is gonna be cracked, although I don't think it's going to be straight up better than the one we saw during shipmas

the direction and bases have changed

[zinc ore]

They might release an updated 2.5 pro as well, think it's been about 3 weeks since it dropped and last year they were doing updates on around 3 week gaps sometimes. Usually 4 at the latest.

[elder rapids]

maybe

but now that it's actually confirmed there's going to be SOMETHING

yep

this is a crazy advantage

the fact they said it's trained WITH the tools

as it's reasoning

ngl

I didn't even know HLE 2.5 pro was 18%

or I didn't pay attention to that

[late path]

o4mini's hallucination rate is unbelievably high. it scores worse than o1mini on OpenAI's own PersonQA benchmark

[elder rapids]

I wanna know how good 2.5 pro DR is with that

[elder rapids]

I do wanna point out tho, independent from benchmarks

it's smart

but it has unbelievable confidence in going in one route

[leaden palm]

do you have the data? can't find it in https://openai.com/index/introducing-o3-and-o4-mini/, https://openai.com/index/o3-o4-mini-system-card/, or https://openai.com/index/introducing-simpleqa/

[elder rapids]

o3 and 2.5 are very creative in their self reflectiveness, but o4 mini just really goes for it

super straightforward too

[leaden palm]

are you saying personqa is a subset of simpleqa

i dont think it is

[plain zinc]

Nightwhisper is coming out today!

[leaden palm]

ok its personqa

serves the same purpose i guess

[plain zinc]

dragontail cannot exit because it is still being tested

This may be the 2.5 flash model that will be released next week only.

[brittle tiger]

For ppl saying o4-mini is cheaper than 2.5 pro

[alpine coral]

yeah exactly.. i really don't think this is widely appreciated ha

[zinc ore]

Wow, nearly 20x price for o3 high

[alpine coral]

that would be v impressive if that's the case.. like kinda insane tbh.. dragontail consistently performs comparably to 2.5 Pro, sometimes does even better (in my experience)

[plain zinc]

Yes, I agree. Sometimes it's better.

[elder rapids]

ye I sent this earlier

[plain zinc]

Okay. This is 2.5 Flash

Not 2.5 pro modified version 🥲

[elder rapids]

too early to say that lol

since it could be

Gemini 2.5 flash preview

Gemini 2.5 pro 0417

[balmy mist]

yeah w/ the agentic reasoning built in

[plain zinc]

Is 2.5 flash really a nightwhisper._.

[elder rapids]

that would be insane

but honestly could make sense

[balmy mist]

then openai is cooked

[elder rapids]

if it really is

yeah

that would just be insane

[balmy mist]

i feel bad for openai again

[plain zinc]

2.5 flash with the maximum level of reasoning

[elder rapids]

like I would be in disbelief

[plain zinc]

2.5-flash-high

[balmy mist]

lets do a go fund me for openai

[elder rapids]

ye

but they all think for similar amount of times

they're capped

by lmsys

so whether it's thinking really fast and a ton, for the same amount of time

we'd never know

we don't even know if night whisper could even be flash 2.5

it could be those other checkpoints, why would the performance be so different

if not for a Gemini coder, 2.5 pro checkpoint, and then 2.5 flash checkpoint

[balmy mist]

do you work for openai?

they should hire you

how can google get you on their side?

notebooklm>

[elder rapids]

I see your point, but you probably think the image zoom feature that o3 uses is an example, but you're just wrong lmao, notebook llm? Google AI search (as in bard, from way before), native image gen flash? long context?

wym? this is exactly THE step forward these companies are trying to make, the tool usage, all that stuff basically came from bard lol

without search, hallucinations become a major problem again

also not to mention, project Astra

and still, the other things I mentioned

[alpine coral]

tbf they pioneered and open sourced the transformer architecture in 2017 - everyone else since (oai included) is arguably innovating on top of the status quo they established

[elder rapids]

even still tho

Google had been developing veo before open AI

had sora

they just announced it after

[thorny drum]

i really wonder who writes llm system prompts

ifl they have fun with them

[alpine coral]

yeah tbh i knew that's what you meant (was just thowing it in there - but in a way, kinda beside your point / taken for granted i know)

[thorny drum]

do you think they AB tested calling it 'yap score'

[brittle tiger]

i think you might be right about oai consumer stickiness if goog doesn't beat them by wide margin but they def innovate. invented self driving cars, solved protein folding, 2 nobel prizes in AI past year

[elder rapids]

what about context caching, Gemma, native audio understanding, learnLM, AI overviews

I honestly don't think that argument has any basis

even benefit of the doubt

[balmy mist]

@deep adder did you like 2.5 pro?

[elder rapids]

Google deepmind genuinely seem like the actual innovators, in and outside your direct chat bot interface

even mobile integration, Gemini has

how tho?

I would literally NEVER use it's o3s tool usage

if not for specific coding things

lmao

[plain zinc]

🤨

[elder rapids]

and btw

[plain zinc]

Don't like the model because it's a chatbot?

what?

[elder rapids]

no

read context

he's saying that openAI are the innovators

while naming things Google created

[elder rapids]

but it's not "actually" useful though?

you're saying it is

but that can't be the case if it's coding based

[plain zinc]

And completely replacing human labor

[elder rapids]

you can say native image gen

[keen fulcrum]

XAI, Grok 4 will crush

[plain zinc]

No, I don't want that kind of future.

[elder rapids]

but Google did that first

you can say vision

but Google did that first

you can say agents

[plain zinc]

I'm just for the AI to be smart enough to be an assistant and no more.

[elder rapids]

but Claude and gemini did that first

the only thing you can say, is "reasoning"

[brittle tiger]

i think talking to dolphins will be useful

[elder rapids]

but did Google not release a paper

explaining exactly what openAI did

[balmy mist]

who are the mods of this server? it would be funny to add tags to people name for if they support a specific company lol

[elder rapids]

like, right after?

[elder rapids]

ye I saw this too

😭 🙏

[balmy mist]

thats dope

[elder rapids]

why'd they do that tho

[balmy mist]

u dont wanna talk to dophins?

[elder rapids]

nah deepmind is cool, they just do some stuff

btw did we all just forget about genie

ye but openAI wouldn't be where it's at without the transformer architecture

they improved it for the same reasons Google knew

that's why Google was able to hop on board so early

ye

it's kinda crazy how one company

kinda sparked it all

or two ig

or 3

there's prolly more

but same idea

crazy how a couple companies are causing the craziest era

an invention probably greater than fire

ye but I'm pretty sure that would've been created regardless

things like that were bound to be created soon

not sure about large language models tho

cuz its actually pretty specific

like who knew bro

ye, but that seems more like the problem of antecedent improbabilities

[keen fulcrum]

Hi, here is an insane offer
https://www.lennysnewsletter.com/p/an-unbelievable-offer-now-get-one
For $240/year you get a 1-year Pro subscription of below items.

Bolt
Cursor
Lovable
Replit
v0
Linear
Notion
Perplexity Pro
Superhuman
Granola

It isn't

[ember rapids]

Logan tweeted the bat signal

Looks like we might be getting something tomorrow

[patent bane]

sybau

"iT IsN'T"

[drifting thorn]

Okay, just got the idea of this

It's basically dealing with the routing issue of having 1 million experts

but there's no companies(we've known now) that are using that much experts in a model

Deepseek has 256+1

Llama has 128/16 only

And other open-sourced models are dense models

[fleet lintel]

o3-high is extermely expensive

[elder solar]

i dont see gpt 4.1 on lm arena

[leaden palm]

in the leaderboard or in the arena?

[elder solar]

leaderboard

[leaden palm]

patience axel

[plain zinc]

They're probably collecting feedback from Gemini 2.5 Flash right now.

[elder solar]

flash??

theres only pro

[leaden palm]

you missed out on cloud next

(and the open secret that google tests many models and tunes on lm arena anonymously)

[elder solar]

but gemini 2.5 flash is not out

[leaden palm]

????

so what

it's being tested on lm arena anyway

[keen beacon]

screw openai

dont forget they offed their first whistleblower

[fleet lintel]

Coming out today... almost 100% sure from all the changelogs etc

[ornate stump]

but flash means just faster and cheaper, right? So why did it seem like it was actually a bit better in Arena than pro?

[oblique flint]

if flash 2.5 gets anywhere near o4 mini performance at flash 2.0 prices I'll be very happy

[fleet lintel]

who said it's better in Arena?

It's actually quite meh compared to bigger models

but given the cost, it's amazing

[fleet lintel]

o4 mini prices == 2.5 Pro prices. Despite name of "mini", it is equivalent for Pro models of Google

[fleet lintel]

Team Anthropic!! wohoo

[ornate stump]

idk I don't follow too much, but it was Dragontail or Nightwhisper or something like that, right? I've tried it a couple of times and it seemed like it had better structured output.

[alpine coral]

yeah if it's dragontail that is being referred to as 2.5 Flash, it's def not meh

[fleet lintel]

nah... Dragontail or NW are definietly not flash. If they are then Google has already won the game. I am ready to bet that they are not flash

[alpine coral]

which google anon model/s are you referring to then?

omg - could today be the day

google is finally unveiling euraka chatbot!!

[fleet lintel]

Riverholllow

[alpine coral]

oh yeah i can see that

in terms of performance

[fleet lintel]

or shadebrook.. They both behaved like flash

[ornate stump]

Ah, right that's what was confusing me.

[alpine coral]

yeah agreed. neither of them get close to 2.5 pro

[alpine coral]

i swear they've just forgotten about it.. been there for like more than a year now... next level stealth

[fleet lintel]

what eureka chatbot does?

[keen beacon]

Confuses me why they still host it lol

[alpine coral]

so if dragontail/nightwhisper isn't 2.5 Flash (which I think we can safely assume), i feel like there's a good chance that dragontail is 2.5 Pro with reasoning_budget set to high

[elder rapids]

ngl

even granting it, outside of speculation

that they'd do this

it wouldn't truly make sense

considering there's nothing about a whistleblower in openAi's case

that could be harmful in any way shape or form

or even openAI much as a company

[keen beacon]

..?

u dont believe companies off people

?

[elder rapids]

do I believe companies would kill people that would harm potential strategy?

[keen beacon]

Thinking openai killed that guy is wild tbh

[elder rapids]

yeah, that's inherent to the idea

lol

[fleet lintel]

i dont think tech companies do it "yet".

[elder rapids]

I'm not sure how that's really substantive

[keen beacon]

silicon valley is one of the most disgusting places

on earth

[elder rapids]

considering I said, let's grant the premise that he was killed

say he was killed, rationalize it

like, in chat

[keen beacon]

i dont think it's far fetched to say some companies would go to extreme lengths

[elder rapids]

and if you can't, why speculate

[keen beacon]

Iirc wasn't he whistle blowing something that was obvious, wasn't it about copyrighted content

[elder rapids]

ye

[hardy violet]

Does everyone remember the Ultra series? It was only used once. It's a great name. If a 2.5 Pro-High (something like that?) model actually exists, the name might be worth considering.

[elder rapids]

but even if we grant that too

that's not even crazy

[keen beacon]

its speculation true

[elder rapids]

nah but fr

of course speculation sometimes means something

but like, you don't gotta speculate, if on the grounds of accepting that premise, you still can't understand why

no dots are connected etc

[keen beacon]

nah cus ts pmo 😭 🙏🏽

[elder rapids]

then why is them killing the whistleblower unique

therefore it has to be the conclusion you're looking for

ykwim

[fleet lintel]

honestly, I think this is not the right place to discuss whether OAI assasssinated someone or not

[elder rapids]

ion think it's serious enough not to

just a little comment here and there

like now, discussion poof, dismissed

[elder rapids]

ye

it seemed to be a really really large model tho

like gpt 4

and that's really the only way they could've competed at that time

they limited usage too

[keen beacon]

is there a new gpt model dropping

feel like i hear about it everytime i open the chat

[elder rapids]

and that's surprising for Google

or maybe they allocated enough compute

during then and now

but who knows

if they gave 2.5 pro high compute

it WOULD dominate

[keen beacon]

https://techcrunch.com/2025/04/14/openais-new-gpt-4-1-models-focus-on-coding/

[elder rapids]

o3 full and o4 mini dropped

[keen beacon]

its alr dropped

Old news lol

[elder rapids]

bro is so behind

[fleet lintel]

i think there is no money or much real-world applications of Ultra models. Only benefit I see is to show how far ahead we are to the world (by benchmarks etc). That's why I think Google is not pursing them anymore

[elder rapids]

4.1 dropped like a little while ago 😭 🙏

[keen beacon]

they just dropped this monday

[elder rapids]

censor ts

not all of us w*rk

[keen beacon]

😭

[elder rapids]

ye but today

they released more models

o3 full and o4 mini

[fleet lintel]

I understand your sentiment. sometimes poeple behave as if 2 hour (or day) old news is like from few years back 🙂

[keen beacon]

Space moves very fast

[elder rapids]

ngl 4.1 is kinda bad outside of coding

has vibes

but synthetic asf

[keen beacon]

20$ a month

and if using free you only get like 30 back and forths

vs geminis free trial 1 month repeatable full advanced

[fleet lintel]

4.1 should have been just a tweet. but they hyped it like crazy.. I am still mad about it

[elder rapids]

not even that

straight up AI studio

give them your data, you get infinite usage

I know the data I give them allows them to progress btw

it's worth a ton

don't glaze me yet tho

[keen beacon]

i know

the thing about this is

skyvern and other things about to get full gemini support

[elder rapids]

whatever that means, I agree

🙏

[keen beacon]

https://github.com/Skyvern-AI/

[elder rapids]

niche territory right off the bat

[keen beacon]

its like the only computer vision project ik about

[elder rapids]

damn you either live under a rock or you don't

[keen beacon]

go look, that will be free? nah

https://github.com/browser-use/browser-use

[elder rapids]

yo what if Google releases an agent

like, a full on agent

[keen beacon]

you can try this but its made by LangChai or some ccp stuff

[keen beacon]

thats what i want bruh

they have screenshare but its bad

[alpine coral]

i think it was worthy of more than just a tweet.. hopefully one day it'll all be a lot clearer with hindsight / disclosures, and someone will create like an org chart or family tree, with a time axis.. it's hard to conceptualised / understand the constellation of models they've got atm, and how they all (or most) in one way or another kind of interrelate

[keen beacon]

sends like past 5 second worth of clip for ai to understand

[alpine coral]

we don't just want thinking models anyway imo

[elder rapids]

ye but now with their new understanding

and determination

[keen beacon]

i think that's what they'll do with copilot

but we'll see

[elder rapids]

seems like they can really create anything

they've just been upgrading the Gemini app nonstop

[keen beacon]

their funding

them mfs created a new material

[elder rapids]

was that even real tho

the Microsoft thing

[keen beacon]

https://deepmind.google/discover/blog/millions-of-new-materials-discovered-with-deep-learning/

[fleet lintel]

but hype like I can't sleep blah... was unwarranted

[elder rapids]

ohhh

yo wtf

why'd I not know this

bro picks and chooses which rock he lives under

[alpine coral]

yeah def wasn't a can't sleep moment - i'm with ya there ha

[fleet lintel]

impossible to be on top of everything in AI space. just focus on what is best for your professional career

[keen beacon]

https://www.capacitymedia.com/article/2e4ufsy6ss2sklukotgcg/news/article-googles-next-gen-quantum-chip

[keen beacon]

if this is real

like

[fleet lintel]

this is called marketing 🙂

[keen beacon]

still if its getting this fast and we're constantly advancing at this rate

[elder rapids]

nah but ask me anything I would've probably known

[fleet lintel]

dont worry too much about quantum computing for atleast 5 years. progress will happen but noting useable till thne

[keen beacon]

eventually things like AES will die out

[elder rapids]

just not that

which is crazy

how'd I not know that

I'm Lowkey disappointed I'm logging off

[keen beacon]

😭

[fleet lintel]

with waht?

[elder rapids]

got dealt a revelation

[fleet lintel]

oh take care!

https://www.reddit.com/r/singularity/comments/1k156qa/ig_google_has_won/

[hardy violet]

Okay, so here's the thing. But, like, O1 and O1 Pro, based on both their pricing and positioning, they don't really feel like the same model at all. If a 2.5 Pro-High (or a similar product) were released and named "2.5 Ultra"...@elder rapids@fleet lintel

[fleet lintel]

This is my team experience as well with o4/o3 models.. very expensive and not necessarily better

[fleet lintel]

makes sense. but it wont be called ultra. people have differnet understanding of what ultra means and launching a 2.5 Pro-High with ultra name would result in backlash

they might (and should) do more like Pro and flash models with different level of thinking budget (like OAI)

[hardy violet]

"Pro" and "Flash" are indeed clear and easy-to-understand product naming conventions. And yes, Google AI Studio needs a parameter to adjust the reasoning strength . Furthermore, I've noticed that many websites don't offer any option to adjust this reasoning strength, which is very inconvenient.🥲

[alpine coral]

yeah i don't think their model released will be called 2.5-Pro-High (though it might be.. just to hedge ha)

i think when they release the stable version, it will have a parameter for reasoning tokens (like sonnet.3.7-thinking is fixed at 32k; OAI models offer low, med, high). Perhaps it'll be a dropdown, low/med/high, but there's no reason why the end user couldn't specify the max number of tokens to be allocated for reasoning / thinking, like with max_output

and i think perhaps dragontail could be 2.5-Pro served with a high value set for reasoning_budget

but it won't be like separate model

[fleet lintel]

good guess.. i agree with you

https://x.com/GeminiApp/status/1912591827087315323?s=19

🤔 how useful is it?

Google Gemini App (@GeminiApp) on X

We’ve been hearing great feedback on Gemini Live with camera and screen share, so we decided to bring it to more people ✨

Starting today and over the coming weeks, we're rolling it out to *all* @Android users with the Gemini app. Enjoy!

PS If you don’t have the app yet,

X · Apr 16, 2025, 7:40 PM

[opaque adder]

is o4 mini or o3 better than gemini 2.5 pro yet

or is chatgpt still braindead

[fleet lintel]

o3 > 2.5 pro > o4 mini

but o3 is very expensive for relatively small benefit over 2.5 pro.

I am a bit disappointed by yesterday's OAI release.

But chatgpt is no way braindead 🙂

[opaque adder]

wtf

o3 is better than 2.5 pro now

surely nightwhisper is gonna come out

why are u disappointed tho

ah shet yeah

[hardy violet]

No, that's not quite right. The difference is minimal. However, in some o3 scenarios, the cost can be ten times higher or even more! And there are also cases, particularly in code writing, where it's less than ideal (certainly not as good as the scores might lead you to believe). While its Agent performs well, which is fair, the base model is just deeply unsatisfying.

[hardy violet]

I'm not sure what the state of English is like with this model (I'm using Chinese). I've tested its understanding of a few texts, and the results are far from ideal. At first glance, it seems impressive and insightful, but that's quickly revealed to be an illusion. Like DeepSeek R1, it tends to latch onto certain words and over-interpret them, and its language is quite flamboyant. If you like DeepSeek's writing style, then perhaps this model is a compromise. In comparison, Gemini 2.5 Pro demonstrates a remarkably solid, appropriate, and somewhat insightful understanding. Furthermore, if understanding is subjective, I also tested the application of translating Japanese song lyrics into Chinese, and it was terrible, very disappointing.

[keen beacon]

goooood morning

[balmy mist]

gm bro

[sonic tendon]

gm

[plain zinc]

I do not recognize o3-high as a competitor for 2.5 pro, because it is o3 with the MAXIMUM level of reasoning.

[opaque adder]

ur using a translator?

[opaque adder]

i didnt realise that this is ai

[opaque adder]

this one is sorta noticeable that it's ai translated because of
" and its language is quite flamboyant."
"but that's quickly revealed to be an illusion."

[alpine coral]

still 1000% better than old school translation (wouldn't even be a question if it was translated; there'd be a litany odd / seemingly misplaced words etc)

kinda interesting come to think of it.. like you give an LLM text to translate, and ig generally it both translates and polishes it? Like it will spit out the translation with properly formed sentences, correct punctuation etc, even if the original was sloppy (like typos, poor / no punctuation etc)?

[alpine coral]

also hopefully AI translated not AI generated (otherwise I think it like passed the turing test.. or we failed it lol)

[keen beacon]

https://x.com/legit_api/status/1912701325504020612

ʟᴇɢɪᴛ (@legit_api) on X

there's a chance that the next Gemini model to be released is 2.5 Flash

hopefully there's more after this

X · Apr 17, 2025, 2:55 AM

[balmy mist]

this si what o3 says about itself:
Think of me as a miniature A2A coordinator bundled with a suite of MCP‑ready tools.
Everything you saw in the video—agent discovery, delegation, precise tool execution—happens here on a smaller scale every time I answer. The protocols simply formalise and generalise what’s happening under the hood right now

so openai is pretty much doing what google is doing with A2A and anthropic is doing with mcps but built in

the only thing that will make o3 better is the ability add more tools or mcps to it overtime and allow o3 to create agents based on those tools

[drifting thorn]

omg another Chinese speaker in the chatroom

[tall summit]

how good is gemini 2.5 as a translator in those languages?

[drifting thorn]

Very good in "Chinese to English" I can confirm

[tall summit]

what lmao

paid tier for a discovery tool

[keen beacon]

either way people will just repost that stuff so you won't ever need to pay

[tall summit]

well yeah!

haven't looked into what he's doing but i do wonder

whether he has exclusive access to something (and what it is) or not

[balmy mist]

yeah to pay seems weird bc he posts it as soon as he gets news

thats what his twitter is for

[tall summit]

oh he'll surely stop posting it

[balmy mist]

like why he gets views lol

[tall summit]

once he makes it paid

but he also is saying he won't make it paid for now

clearly testing whether he can afford to

but nothings even happening i dont even know why im speculating

[balmy mist]

what woul dbe the benefit for people? like how useful is his tool anyway? it tells us what is most likely coming out but so does the companies

but for the webdev arena stuff its cool

like i never check webdev unless he posts about a new model

[tall summit]

he knows before the companies officially tell us

or at most immediately when they do since its a discord webhook and he gets notifications

[balmy mist]

not the gemini one

[tall summit]

?

[balmy mist]

yesterday

logan posted about a new model

then legit posted about it a few hours later

about the exact model

but at that point whocares

i can just wait for today

unless you in prediction markets

idk, i just dont see the need for buying that, especially when you have channels like this

the only thing about openai is their UI sometimes buggs out

i always have issues with their website

[narrow elbow]

TWEET LIKE TRUMP, “THIS IS GREAT TIME TO BUY”🤪

[balmy mist]

never get issues with studio, like maybe 1/10 times with studio and i usually know when it will bug out, but with chatgpt bruhh, its like 50/50

im gonna build an o3 clone that is adaptable where you can add new tools on the fly, its pretty much just a small scale ai ide or more specifically a augment or cursor or windsurf, but built into a model with specific mcps, like a starter package, i wonder if you built an ai ide with o3 as the brains, maybe that is where openai is going?

gpt5 is just o3 but access to more tools and trained on how to better use those tools and more agents built in

[ocean vortex]

Studio is a pain to use compared to chatgpt website. Good luck trying to branch the conversation from the middle of your chat… It’s really openai’s playground alternative rather than chatgpt. For that it’s alright

[balmy mist]

true but is openai playground free?

and tbh i dont like using openai playground tbh, its a little confusing for me for some reason

studio is very straightforward

[balmy mist]

the branching feature is more newer, and it works okay, but when you compare just fucntionality to what you can do in chatgpt studio works better

you cant even branch in chatgpt, so no point even mentioning that

[ocean vortex]

Are you joking, it works perfect. You just edit the message in the middle of your chat and your chat continues from that point. In aistudio it’s an absolute mess with it just leaving all the flood from previous branch below. It’s a playground and not a full functionality user friendly interface like cgpt

kinda by design

[balmy mist]

you can branch in playground?

[tall summit]

what kind of argument is "who cares" against the fact that it objectively does something most people don't have access to without doing what the bot itself does
though of course it's not worth it and i hope nobody pays for it if he ever makes it paid

[balmy mist]

what you are talking about is getting info an hour or sometimes minutes before people

[tall summit]

or days

[balmy mist]

depends

[tall summit]

honestly nobody KNOWS when anything is dropping

thats THE POINT

[balmy mist]

thats why to me its pointless

[tall summit]

i can see its use

though to me it doesnt much matter what will come out

its all just names

[balmy mist]

what would paying for it do for you?

[tall summit]

until we actually see what it does

which only ever happens on lmarena or private access (which nobody normal will have access to regardless) or the actual release which after that it doesnt matter

[tall summit]

??

[tall summit]

.

[balmy mist]

? i dont know how else i could write that

what would it do for you?

lol