Horizon Beta

[valid zenith]

https://www.reddit.com/r/singularity/comments/1johdba/sam_altman_says_their_open_source_model_will_not/

[magic frost]

now go to Google Gemini Flash 2.5 - better context handling and fix and add almost all after Horizon Beta.

same there ) But I see that key parts of the task using Horizon have not been completed; it has not even been started, but simply marked as completed. So get ready to rewrite the code. It's a double waste of time. I don't think it's productive.

I didn't feel confident, although when preparing specifications for the first step, I could come up with good ideas, but when it came to applying and implementing, say, React applications, things started to get very strange.

[modest crescent]

sota in repetition for longform writing too

[modest crescent]

creative writing too

[bitter vigil]

yeah I don't think a 120b moe can accomplish that.. which is rumored to be the oss parameters..

but I'm hoping to be wrong

[modest crescent]

it's def the creative writing model

that sama was talking about

in march

the only question is

is it the oss one or gpt5

[rare terrace]

It's sota in coding when alpha had reasoning for a few horus

[bitter vigil]

does it excel at creative writing and nothing else?

or is it an all arounder

[modest crescent]

the current version? yea

idk ab the reasoning for 3 hours one

but besides that, pretty much

[rare terrace]

I approached the reasoning results with a simple CoT prompt

Not sure if that worked fully

Someone had a vision benchmark that non-reasoning alpha failed at but the reasoning one scored at the top

[modest crescent]

yea it's weird

they could've switched models

if the non-thinking version didn't only excel at creative writing, i'd be willing to bet it's gpt5-nano or something

[rare terrace]

It doesnt only excel at creative writing, it's quite good

Some do speculate that it's gpt 5 nano

[modest crescent]

from what i've seen, people say it's avg

[bitter vigil]

I'd believe it's a 120b moe oss writing model before a nano

[modest crescent]

it'd be such a gift

[bitter vigil]

mini model maybe

[modest crescent]

to the community

[bitter vigil]

nano is too small, their nano/mini models have been pretty mediocre

[modest crescent]

yeah

[bitter vigil]

the benchmarks have big model smell all over it

distil from o3 helps there but with the output length, with 0.00 degredation and such low rep and slop socres

[modest crescent]

have they ever teased a mini model before the full version yet

[bitter vigil]

that's so hard to do with a small parameter model

[modest crescent]

doe

[bitter vigil]

nope

not that I remember

[modest crescent]

this def isn't the full version

[modest crescent]

well if it's specifically trained for creative writing

it's prob possible

[bitter vigil]

yeah we haven't really seen a model like that before

[modest crescent]

sama's creative writing tweet

def confirms to me that this is the model

[bitter vigil]

yeah that he was interested in doing it, I saw that.. hype

[modest crescent]

he was talking about

let's just hope that it's the oss one, please god.

[modest crescent]

it gets so many things right

and writes so accurately and good

the best model i've tried yet

[bitter vigil]

that's how I feel about kimi k2 right now haha (for creative writing)

[modest crescent]

imagine if this were the oss one

and kimi trained with it

for their next update?

[bitter vigil]

oh man lol

[modest crescent]

the fucking progress it'd be

kimi's so fucking good too

but this model remembers more things

than kimi does

the combination? jesus

[bitter vigil]

yeah the writing is really vivid however yeah kimi is like a first release v3

in terms of actual performance

long context kind of not great

[modest crescent]

i def predict

ai models that will be able to write like

a 100 chapter book

with long context in the next 2-3 years

[bitter vigil]

I just want them to be able to hold a world together with coherency and not write tropey repetitive stuff

[modest crescent]

yeah, the authors will be screwed though

we will win in terms of being able to read whatever we want

but at what cost

[bitter vigil]

average to good writers may be in trouble, but excellent writers with good taste I don't think so.. for example I don't think AI will ever write better than GRRM

[modest crescent]

prob yea

but man, it'll be such a battle

in the future

imagine showing this to someone in 2020?

a LOT can change in 5 years

[bitter vigil]

the problem is that we will be flooded with ai content and users will be okay with that

ai amazon books, ai youtube videos, blah blah

[modest crescent]

yea, it's unavoidable

unfortunately

ai mukbangs have already started

they have people eating things made from lava

😭

[bitter vigil]

I saw a vid the other day of someone cutting open planets with a knife and they oozed out the inside cores

watched the whole thing.. can't deny they're entertaining when done right

[modest crescent]

it's already bad though. twitter's userbase is 76% ai bots lmao

[bitter vigil]

my mute list is MASSIVE

[modest crescent]

that's why u see

all the right-wing rise

they're all mostly bots or paid actors

and why elon has 220m followers

[bitter vigil]

electioins are gonna be such a shit show

[modest crescent]

50% of them have no pfp/0 followers

[modest crescent]

well, kinda helps if it's 76% bots

that means not many people are on twitter

but this won't be going away anytime soon bc elon'd lose engagement etc

the dead internet theory may be true after all

[bitter vigil]

fb is really bad for ai fake news crap all on the feed

so many fake movies and fake outrage and fake celeb stuff

[modest crescent]

all the boomers believe

[bitter vigil]

mixed with temu ads for fake products that aren't what they appear in photos

[modest crescent]

whatever u post on there

[bitter vigil]

lol

[modest crescent]

yeah, it's pretty bad

[bitter vigil]

we'll need a new social media platform tha somehow verifies for humans.. good luck with that right

[modest crescent]

i mean, it could happen

but they won't wanna implement it bc their platforms

are dead as hell in reality

[bitter vigil]

yeah but ai gets better at writing and being less detectible.. then we have the browser use agent stuff

[modest crescent]

they = twitter, facebook etc

elon's focused on making a model for gooners to goon to

himself included, like what are we talking about

[modest crescent]

have youtube implemented their ai for ads yet

[bitter vigil]

dunno

netflix is adding AI ads that look like part of the show you're wathcing, or themed on it

[modest crescent]

the ai will show ads during the most intense moments

aka piss everyone off 🔥

[bitter vigil]

probably will place them where there's least viewer dropoff

[modest crescent]

god, imagine telling a person in 2015 this

[bitter vigil]

scifi as fuck lol

[modest crescent]

i can't even 😭 the progress has been massive

by 2038, we'll be like detroit become human

atp

[bitter vigil]

the humanoid robots are coming along fast too.. figure-002

[modest crescent]

guess we're living in the timeline where detroit is the actual future

[trail gale]

Horizon Beta is now very filtered. I miss Alpha so much

[restive pasture]

Model down?

[summer root]

most major countries are currently developing some sort of digital ID system. probably not for this reason, but this is a good use for it

[trail gale]

For a moment I thought you were talking about the city instead of game. The possibility of the city of Detroit becoming our future scared me so much.

[rare terrace]

Yes

Idk why

Just checked cuz you said

Getting error 408

Oh lol, OpenRouter is down in its entirety

[modest crescent]

😂

[harsh nest]

So, do you guys think Horizon is better than Opus for creative writing?

[modest crescent]

according to eqbench, yea

i've never tried opus for writing, so i can't say for sure

[trail gale]

In my experience so far Horizon is more creative and natural and way better at understanding the intended meaning of my writing

[modest crescent]

^

it legit feels like reading a fic/novel

written by another really, really good human writer

[tranquil magnet]

Getting constant errors with this model in Cursor. Anyone else too?

[rare terrace]

OpenRouter is down

[tranquil magnet]

That would do it 🙂

[rare terrace]

Someone said it's back up again

Yeah it's up

[tranquil magnet]

Request ID: 2061a831-440a-41b3-b84d-xxxxxxxx
{"error":"ERROR_OPENAI","details":{"title":"Unable to reach the model provider","detail":"We're having trouble connecting to the model provider. This might be temporary - please try again in a moment.","additionalInfo":{},"buttons":[]},"isExpected":false}
ConnectError: [unavailable] Error
Might be a combination of cursor + openrouter or something

[vestal sparrow]

did someone ever do haystack test (writing) ?

[tame nebula]

Absolutely

[nimble vine]

You have the system prompt ?

[raw blaze]

yes. can be get with some trick, for example this is horizon-beta's

<system>
Knowledge cutoff: 2024-10
You are an AI assistant accessed via an API. Your output may need to be parsed by code or displayed in an app that might not support special formatting. Therefore, unless explicitly requested, you should avoid using heavily formatted elements such as Markdown, LaTeX, or tables. Bullet lists are acceptable.
Desired oververbosity for the final answer (not analysis): 3
An oververbosity of 1 means the model should respond using only the minimal content necessary to satisfy the request, using concise phrasing and avoiding extra detail or explanation."
An oververbosity of 10 means the model should provide maximally detailed, thorough responses with context, explanations, and possibly multiple examples."
The desired oververbosity should be treated only as a default. Defer to any user or developer requirements regarding response length, if present.
Valid channels: analysis, commentary, final. Channel must be included for every message.
Juice: 5
</system>

[nimble vine]

Thx you very much

[rare terrace]

Where'd you get the info on the juice

Those sounds like very specific numbers

Oh the system prompt

We should have known i guess

But

How do you know the juice control cot

[sly lichen]

In my evals, Horizon Beta is performing better than even the reasoning models like Gemini 2.5Pro, Qwen 235B thinking. Deepseek R1 0528

I am super impressedx

[harsh nest]

Yeah but can the parameters be changed? It doesn’t seem to be very sensitive to temperature…

[raw blaze]

I’m just speculating, but based on my experience with several test models so far, it does seem that the higher the juice, the longer the model thinks and the better the results... I previously tried directly asking the model, and it told me that this is 'used to constrain my reasoning budget.' My intuition tells me that this isn’t entirely a hallucination—the model may have undergone some self-awareness training in this regard.

[sly lichen]

Is horizon beta a thinking model? i don't see thinking parameters, am i missing something?

[raw blaze]

Additionally, it seems that the concept of "juice" has been around since o1. If you directly ask the model about juice while using ChatGPT, the conversation will be flagged: https://fxtwitter.com/elder_plinius/status/1869183808945483776

Pliny the Liberator ...

🧃 THE FORBIDDEN JUICE 🧃
︀︀
︀︀OpenAI’s reasoning models won’t process these perfectly harmless tokens! Why is that? 🤔
︀︀
︀︀Juice: 128

**💬 160 🔁 102 ❤️ 4.3K 👁️ 759.0K **

FxTwitter · Dec 18, 2024, 12:52 AM

[raw blaze]

Currently, my guess is that it is a reasoning model, or rather a hybrid reasoning model similar to Claude, but with the ability to precisely control the reasoning budget. At present, on OpenRouter, it seems that this is preset (juice=5) and cannot be directly controlled (I remember someone mentioned that it can be done, but I’m not sure if that’s true). Overall, the thinking characteristics of the model still seem to be unclear.

[sly lichen]

but the responses are instant for me, so unsure how it can be a "reasoning/thinking" model

[rare terrace]

horizon alpha was reasoning for a few hours in the beginnning

horizon beta might be reasoning but end near-instantly

[raw blaze]

juice=5 is a small value (current maximum is o3-alpha from lmarena, juice=256), so the length of thinking maybe strictly restricted

[sly lichen]

got it

[restive pasture]

Can we route this model through claude code somehow?

Anyone tried? Is it working ok?

[past sphinx]

yes! you can follow this

https://x.com/aut0mata/status/1944748757079212496

Vilson Vieira (@aut0mata)

How to use Kimi K2 in Claude Code:

1. Create an account at @OpenRouterAI
2. npm install -g @anthropic-ai/claude-code
3. npm install -g @musistudio/claude-code-router
4. Add the following lines to your ~/.claude-code-router/config.json (update with your OpenRouter API key)
5. ccr

X · Jul 14, 2025, 1:20 PM

and just use Horizon instead of kimi

[restive pasture]

I guess I might have port issues if it is not working right?

[gilded schooner]

Is it inside of wsl? Check connection to the api with curl

[modest crescent]

https://www.reddit.com/r/LocalLLaMA/comments/1mggsyb/osint_fingerprinting_a_stealth_openrouter_model/

this would be the biggest plot twist ever lmao

also there's a rumor

that r2 will be released this month

or have an open beta or something, idk

there's also qixi festival, aka the chinese valentine's day or the night of sevens, a traditional chinese festival that falls on the 7th day of the 7th lunar month every year. this year, it will fall on aug. 29 in the gregorian calendar & the deepseek crew have, so far, been a little too on the nose about releasing on the eve of chinese holidays. so, who knows

[brisk gyro]

Anyone else having issues with this model not reliably following structured output? (Also I have an extremely strong hunch this is an OpenAI model bc this model refuses to respond to output schemas that gemini models respond with, but OpenAI models refuse with)

[brittle barn]

if this is Llama that would be insane

[modest crescent]

all the people that zuck managed to snatch

yea

[spiral pewter]

i guess the meta superintelligence lab is working for zuck if it is

[modest crescent]

llama being sota on eqbench

would highlight 2025

i'm sorry 😭

[spiral pewter]

llama being sota on anything is insane ngl

[modest crescent]

i mean, i'm sure he'll get there? with all the people

he's managed to get so far

so even if this isn't llama, he should have something similar soon

[spiral pewter]

the only thing i found good about llama 4 was the vision, it was actually better than most open source models and even gemini sometimes

[brittle barn]

I still think its GPT tho

[spiral pewter]

i highly believe so too but if it were meta that would be insane

[brittle barn]

I was working on a curriculum and ran it through this and GPT 4.1 and the results were p much identical. They both p much gave me SCORM compatible outline including quizzes. No other model I used had quizzes

[modest crescent]

i expect zuck

to have openai's sauce

bc the people that now work for him WILL reveal it to him

knowing him, llama will still suck lmao

[dapper oar]

If it's llama, then it'll prop be their closed model

[modest crescent]

yea, they made a statement about not being open anymore

i believe

[dapper oar]

My short time with beta, I don't dig it

Deepseek v3 and kimi k2 still feels better.

[modest crescent]

i need this to be the oss one

from openai

so kimi k2 can utilize it

and make their writing even better

🤞

[tender cairn]

this has better writing than kimi k2?

[modest crescent]

yea

its writing is better than any other model out there rn

[tender cairn]

thats pretty crazy considering it's like around 100b parameters or so

[modest crescent]

yea

that's why i think it's the creative writing model

by openai

[tender cairn]

massive models are always naturally good at writing e.g kimi k2 or gpt4.5

gpt5 will probably be the best thouhg

[modest crescent]

yeah, imagine kimi's already really good writing with this?

sheesh

[tender cairn]

its weird they focused on writing cos writing is more of a challenge than just RL loops for coding

[modest crescent]

yea

but it is really weird that even while non-thinking, it excels at that

more than anything else

[tender cairn]

i'm guessing cos coding is where the malicious part is & they have the most to lose if the os model does anything bad

creative writing can't really do anything dangerous really

[modest crescent]

yea

this model just gets

so many things right

u can tell it

do x from 2015

and it'll remember that tiktok wasn't a thing

but vine was

i was really shocked when i read it

[dapper oar]

If it's open 120b, it's going to be lit

[modest crescent]

it'd be such a big gift

[dapper oar]

Even if it's not my vibe

[modest crescent]

idt anyone realizes how big

this could speed up the creative writing quality by 200%

for kimi, deepseek etc

[tender cairn]

interesting

[late onyx]

even if it isn't open weights can't they just flood the API with requests?

[modest crescent]

prob, yea

[dapper oar]

How's the context btw? Recalling long chat well or nah?

[modest crescent]

oh yea

[late onyx]

apparently thats what deepseek did with 4o and v3

[modest crescent]

it mimics the characters really, really good too

[harsh nest]

I use Opus daily for creative writing, and tbh I still don’t see that Horizon is better. Maybe there’s a sweet spot in the parameters?

[modest crescent]

it's the small things

[late onyx]

this model is most likely going to be way cheaper than opus as well

[modest crescent]

like vine existing in 2015

and not tiktok

[modest crescent]

i mean, if this is mini-nano-oss

the full gpt5 will be better too

[dapper oar]

OAI really need to cook for gpt5 lmao, I can see why

[modest crescent]

expecting gpt5 to be sota everywhere on eqbench

tbh

[dapper oar]

Need to beat opus 4 and potentially Gemini 3 and Deepseek V4/R2

[harsh nest]

The problem is that temp change doesn’t seem to affect the output

[modest crescent]

yeah idk why

[late onyx]

thats most likely a stealth thing

[modest crescent]

yea

[late onyx]

not a fundamental model limitation

[modest crescent]

i assume it'll work fine

when it gets released

as an actual model

[late onyx]

how long was the optimus alpha period?

[warm brook]

imma slime you lil bro, watch your tone

[rare terrace]

Prithee, couldst thou enlighten me as to the span—yea, the full measure of time—during which the grand and illustrious epoch known to men as the Optimus Alpha period didst endure? How many days, or moons, or turning of the sun marked the bounds of that most noble age?

[brittle barn]

for code yeah prob, for research/writing this is better

[long sable]

Writing: I don't think it's that good. I give it a prompt and then it just writes something related but not really what I wanted. So I lose interest reading it midway. Only good thing is that long outputs are possible. Alpha had some weirdness in it, beta less. I prefer Gemini 2.5 pro still

[heady gust]

It's tough to say which is better - I like both this and Kimi, and I feel like this one is wordier but I'm still not quite sure if I like it more

I think a lot depends on which model it is. If it's one of the OSS models (I kinda doubt it is but it would be nice) I think this would be fantastic for a 100B level model

If it's like, a gpt-5 variant and it costs $2 or more per million tokens then I'm probably just sticking with my current models

[trim blade]

https://x.com/ramdhanhdy/status/1951690713512685718

RDH (@ramdhanhdy)

Interesting, I've found Horizon Beta is slightly better than Horizon Alpha on Math (still not good enough overall), but at the cost of a slight performance drop in qualitative bench result. Improving quantitative reasoning seems to hurt softer skills just like humans(?).

X · Aug 2, 2025, 5:05 PM

proves what I said

beta got stronger at math / coding while getting a good deal weaker at general reasoning and writing

[modest crescent]

ok so

it's def by openai

i've got confirmation

gpt5 or oss, hmm

[worn cosmos]

Its 100% openai

[modest crescent]

yea it is

[modest crescent]

the only question is this

[trim blade]

It will either be the best open source model by a big margin or gpt5 is more of a side grade

or it might be gpt5 mini, it does not seem a big enough improvement to be gpt5 full, no?

[late onyx]

If it is the best OSS model in a while I wonder what they did

[modest crescent]

please the first one

[late onyx]

Like what architecture or training difference

[modest crescent]

y'all i've been

so annoying

[trim blade]

data

[modest crescent]

about this

[trim blade]

most likely

[modest crescent]

but it'd be such a big win

😭

imagine k2 training from this model for their creative writing

[trim blade]

That sounds like a terrible idea

would massively reinforce literary troupes / repetition

you want to train on as many actually human written books as possible

[modest crescent]

well, this model has the lowest repetition for any benchmark on eqbench

[summer root]

https://x.com/sama/status/1952084574366032354

Sam Altman (@sama)

entering the fast fashion era of SaaS very soon

X · Aug 3, 2025, 7:10 PM

[modest crescent]

i'm sure u could fix it up or sum

[late onyx]

Imagine combining this model, Kimi k2’s agentic workflow, and DeepSeek R1’s thinking ,mix in some qwen, and training that

[modest crescent]

this is what claude

will do

i believe

i saw something ab them buying tons of books

[trim blade]

that is what they all do

[late onyx]

Books3.tar.gz go brr

[trim blade]

plus all the internet that they can scrape

[summer root]

i think its gpt-5 in various configurations

based on that tweet

[modest crescent]

yeah, realistically, it being the oss one prob sounds too good to be true

[summer root]

https://x.com/sama/status/1952070519018373197 and this one

Sam Altman (@sama)

pantheon is such a good show!

X · Aug 3, 2025, 6:14 PM

and also the brief SOTA reasoner period

[modest crescent]

we'll prob see in 2 days?

[summer root]

at least we get this consolation prize

[modest crescent]

i wonder when the oss one will come out then

he said during the summer

it's ending soon

[summer root]

yeah its imminent

[modest crescent]

so where is it then

let me cope & say that the oss one could also be this good creative writing-wise 😊

[summer root]

Quasar/Optimus were out for... just under a week, i think?

[modest crescent]

yea

people are saying aug. 5

should be the day

[summer root]

so, will it cost more or less than gpt-4.5 🤔

[modest crescent]

🤷‍♂️

[steep palm]

Okay, wow. This model hasn't exactly blown me away for a lot of stuff, but the people saying it's good at writing are 100% correct.

Prompt: Write the opening passage of a gritty spy novel (something I test with all LLMs to get a vibe check of their writing)

[modest crescent]

it feels so...human

[steep palm]

I actually started to get into the story, and its use of metaphor and phrasing is excellent

[modest crescent]

like please don't take it away from me 😭

gemini 03-25 all over again

[steep palm]

"I poured the last of the bourbon into a coffee mug because the handle gave me something to hold onto. "

[modest crescent]

https://tenor.com/view/fire-writing-gif-24533171

[sage mantle]

whats your confirmation

[steep palm]

"I checked the door—deadbolt thrown, chain slid, chair hooked under the knob. It would slow them by three seconds, four if the big one hesitated. The building’s hallway was a throat, and I’d lived in enough throats to know how to cauterize them. I opened the window, let the rain come for me too, and counted the stairs to the fire escape with my eyes shut. Eleven down, two to the landing, eight more to the dumpster. The city breathed below, sour and wet and ready to testify."

[modest crescent]

the rate limit

leads to openai

[sage mantle]

like the rate limit error does?

[modest crescent]

yea

[sage mantle]

does openrouter not obscure that? 💀

and do you happen to have a screenshot, im tryna show someone its an oai model

[sage mantle]

tyty

[modest crescent]

@past sphinx

[summer root]

i don't think that error is from OR

[modest crescent]

no, it's from openai

and yes it's from or

[sage mantle]

funny seeing u here sir

[zealous citrus]

good prose is useless if it’s just gonna make every narrative setting sunshine and rainbows it’s the most positively biased model I’ve used tbh

[summer root]

what software is that? i've seen n8n say that for OpenRouter because it's using the OpenAI lib

[modest crescent]

yeah

it's restricted

a lot

[zealous citrus]

fingers crossed it’s less so on release

[modest crescent]

it can write

nsfw so good

i've got a glimpse of it

https://tenor.com/view/fire-writing-gif-24533171

[modest crescent]

wouldn't know. it's from a reddit post

[steep palm]

This was also posted on Reddit. Brownout/downtime for this model matched Gpt4.1 outage

[modest crescent]

yeah, it's def openai

interesting fact

i saw someone say that they put gpt 4.1 on or as a stealth model too

so it could very well be gpt5

[past sphinx]

this looks like an app that's treating openrouter's api as an "openai" style API, and our "provider returned error" message is being considered that way

[modest crescent]

ah, so no confirmation?

[summer root]

https://discord.com/channels/1091220969173028894/1400857391733674045 see it happening here

[modest crescent]

the rate limits do kinda confirm it though

[summer root]

confusing this poor fella

[bronze berry]

It looks like n8n from a reverse image search

[summer root]

i mean, obviously its openai. but they didn't leak it like this

[modest crescent]

ah, gtk

[modest crescent]

https://x.com/SmokeAwayyy/status/1952145720431272395

Smoke-away (@SmokeAwayyy)

GPT-5 on Monday August 4 at 10am PT

X · Aug 3, 2025, 11:13 PM

tomorrow then

[mental cobalt]

solid track record

[verbal leaf]

yeah no

tuesday is the day

[rare terrace]

When is the crossover GPT-Qwen DeepThink

[late onyx]

Nah GPT-5 is actually releasing 3 weeks ago

[patent grail]

Structured data extraction/OCR is quite poor;

[woeful sierra]

is this not at 140b leaked os model?

because compared to alpha its not much of a change

[heady gust]

Beta is an improved version of whatever model Alpha was, so they should be the same model

[trim blade]

sidegrade at best, its a good deal worse at general reasoning and writing

if a bit better at code

[bitter vigil]

the usual. models always get better at code and worse at everything else

I remember deepseek being the GOATs they are actually trained the new version of v3 on rp

[bright oak]

coding is too linear of a process

which is why coding with low temp is even possible

multioutcome parallel thought achieved easier when gravitating away from coding 1shots as the goal

i think anthropics team and overfocus on coding will make them hit a wall harder than most other ai companies

[late onyx]

R1 sometimes is like checks notes Oh yeah, …

R1-0528 in its thinking*

[fair oxide]

Reasoning

[modest crescent]

right, should've just said aug. 5

[trail gale]

Horizon Beta is fully capable of writing NSFW, if the previous context contains a lot of it (for example written by other models). The context rot confuses the model enough to forget how restrictive it is.

Usefull for Silly Tavern or if you are an author. The very existence of the filters seems to limit the writing ability considerably. Definitely got better results with Horizon Alpha.

[grave wyvern]

know it was a wacky few days but the stealth model process is fun and glad you were all there

[rare terrace]

Are you killing yourself

@grave wyvern are you good

Is everything alright?

[trail gale]

For some reason that read to me like you are a movie character dramatically sacrificing himself just before the end of the movie.

[undone cypress]

Same

[modest crescent]

is that so?

[rare terrace]

When is horizon gamma

[modest crescent]

🤲

i find it hard to believe that any gpt5 comp. would get the "how many r's in strawberry?" question wrong

[rare terrace]

That's the thing that makes me wonder

If this is OSS

[modest crescent]

yeah

the latest gpt5 gets it right

someone posted a leak

even 4o mini gets it

i really doubt that this is gpt5

[trail gale]

Yeah. If I try to start new Silly Tavern roleplay with Horizon Beta, its just extremely sensitive to any violence, romance, etc, but if I use it to continue already established conversations (established with different models), its fully uncensored. Definitely not intended behavior, hopefully it wont get patched out.

[modest crescent]

oh right

i think it's if u make it past chapter 1

u are good to go lmao

[trail gale]

Yeah pretty much

[modest crescent]

just be innocent

for the first chapter

and then go full bazooka

for the rest of the novel

🤷‍♀️

[lament tendon]

If it's not gpt5, then it has to be the OSS model, but why would oai test their OSS model?

[modest crescent]

prob to limit shit

like they did between alpha & beta

before releasing to the public

to avoid controversy

[lament tendon]

Mmm, makes sense

[modest crescent]

bc if even gpt5-nano gets the strawberry thing wrong

the model will suck ass

the infrastructure too

and i doubt that that's the thing

[modest crescent]

considering that 4o mini gets it right lmao

[trail gale]

I just hope someone takes the gpt oss model and does a dolphin-esque raw unfiltered fine tune asap

[modest crescent]

the smut is so good

even while filtered, i got a glimpse of it

deepseek will be extinct if so lmao

[trail gale]

Yeah. I use it a lot to get rid of the first draft clunkyness of my novel and it genuinly writes like a very skilled writer, which is something I never said about any LLM. For in character roleplay its also definitely my favorite.

[cold knoll]

its more likely to be Grok 4 Coder

[modest crescent]

it's def not grok

[cold knoll]

Grok 4 Coder is trained on Cline, A coding tool that doesnt use native tools and has the user automatically return a string to approve code actions, this model past 25k context (the same as clines default due to the system prompt) will always ask for permission and tell the user to confirm the action before it acts

its just logic

[modest crescent]

so why was it out

[cold knoll]

None of the openai models have the same pattern of behavious

[modest crescent]

while openai was out too

[cold knoll]

Its been out multiple times without openai being down. its not logical to say a coincidence = fact when OAI can also simply host on azure to offload just like other providers

[modest crescent]

why would a coder model

top the creative writing benchmarks

it makes no sense

[cold knoll]

Why would a creative writing model also top coding benchmarks? its called general purpose.

It was finetuned for design in terms of SWE

Which is obvious when its the only model capable of beating claude opus/sonnet in UI design

[modest crescent]

while also failing

the how many r's in strawberry question?

[cold knoll]

Do you not know how LLMs work?

[cold knoll]

i dont see how failing a question that is entirely based on the refinement of training data proves anything, but the 3 OAI models ive tested all get it right.

[trail gale]

Funnily enough, I have been using Qwen3 Coder for creative writing pretty successfully (because the normal one isnt free on Open Router anymore). Overall I actually got better results with it than the normal Qwen3 235b a22b 2057. I even managed to create good enough jailbrake prompt to get it to stop censoring.

[cold knoll]

My users also use sonnet 4 for "creative writing", i dont personally but just because a model is better in 1 field doesnt mean its bad in others

[vestal sparrow]

imma try this

[modest crescent]

https://www.reddit.com/r/LocalLLaMA/s/zClVuLjIRN

🤷‍♂️

my bet is that this is the creative writing model that sama was talking about

[rare terrace]

You keep saying that, but when the reasoning was on it was SoTA in coding and vision?

[leaden sinew]

not enough feed for those actions

textually its gorging

why are you feeding an unknown model srsly?

[cold knoll]

What?

[leaden sinew]

both Cypher Alpha and Horizon Beta are stealth, right?

[cold knoll]

Yeah?

[leaden sinew]

there is no official team behind it right?

[cold knoll]

Why does that have anything todo with testing the model?

[leaden sinew]

because user inputsare feeding the model with data

you dont know what you are feeding

😐

[cold knoll]

Getting a model to write a snake game isnt really feeding anything

[leaden sinew]

openrouter team probably does but wont disclose

might as well be a suicide club

everything is data

the more it learns the more it feeds

i could be behind it seeking world domination or whatever and how would anyone know?

if i paid openrouter a massive amount of money

to sign ndas

i am not going conspiracy theory mode on ya

[rare terrace]

there was some other madman pushing the same talking points

[leaden sinew]

just warning

yeah

a madman

manic stret preacher

😄

hope it doesnt bit anyone in the ass that is all li have to say

and i am monitoring the situation

[rare terrace]

i just checkedd your chat history, the madman im referring too was you

you said the same about cypher alpha

[leaden sinew]

yes

[rare terrace]

if world domination comes it wont come from user data

only RL

[leaden sinew]

so im the only aware dude in the room?

no

must be crazy

[rare terrace]

they do the previews for hype, less for data

[leaden sinew]

herd\ mentality: true;

[rare terrace]

and the data isn't something groundbreaking either

from what i heard people use it for porn

[cold knoll]

When you arent providing any data that could boost training its not relevant, if burning tokens on basic test prompts is training their model. it would only be logical that they run them themselves in mass

[leaden sinew]

both your points are okay

but have no counterarfumental data to it

[lament tendon]

The amount of prompts you'd have to sift through to find the handful of prompts that offer actual data is not worth it

[rare terrace]

well you haven't provided any data to back up your statements either

[leaden sinew]

sure, lets do it instead of a dnd session my garage buddies

xD

again, i am not preaching

[cold knoll]

I mean realistically, use running mass requests through the unknown model will just make it work better with my app when its released

[leaden sinew]

only static a remark

[cold knoll]

i dont care if im helping train, you talking on discord is helping companies train too

[leaden sinew]

no one else effing raised

unbelivable

[cold knoll]

any public data is helping feed our end

[leaden sinew]

its not the same

literally

nvm just go with the flow

[rare terrace]

I do believe in singularity but it's not going to spur from the chat a porn addict has with gpt

[cold knoll]

Write a snake game -> add advanced path finding that takes into account current snake position + tail etc -> add randomly generated walls -> add poison apples -> avoid poison apples. Really feeding them with good data, All this does is make them better at choosing tools when i complain. It isnt going to change anything in reality. Its not the same as feeding prop data to it

[leaden sinew]

eh well what if i hypotetically wanted to create agents to get global domination by converting users to think the same way, using already established subversive techniques

[cold knoll]

you and aisatoshi would get along well...

[oblique cairn]

Uh it says so in the descriptions

[cold knoll]

literally, alpha was likely finetuned on the data farmed to create beta, and it got worse.

[leaden sinew]

oh it does?

cool

[oblique cairn]

All of them do

[cold knoll]

🤣

[oblique cairn]

So

Idk 🤷

[leaden sinew]

https://behaviordesign.stanford.edu/

[rare terrace]

https://isotropic.org/papers/chicken.pdf

[oblique cairn]

Was a pretty interesting paper to read

[vestal sparrow]

ahh the doomers

[oblique cairn]

Where

[fringe bay]

what is going on here?

[leaden sinew]

https://youtu.be/ZRrguMdzXBw?si=gNWxOgw2BZPPxgCj

[leaden sinew]

nah im just building my own thing

so i would do a thing like this

[haughty monolith]

AM I THE ONLY ONE USING THIS MODEL FOR ROLEPLAYING?

guess i am

[oblique cairn]

What is the purpose of this

[haughty monolith]

It's good but I can't even hold hands

[oblique cairn]

I thought this is common knowledge

[oblique cairn]

Dam

[leaden sinew]

you will never know untill you watch or rather, listen

[oblique cairn]

I have listened to it

[sand pulsar]

As a large language model by stealth provider, I have no ability to watch or listen. Could you please provide tldr?

[summer root]

what if the SOTA coder was gpt-5 and everything else was OSS

[sand pulsar]

Plausible.

[leaden sinew]

can you disclose which team is behind these models?

[summer root]

they just wanted to freak everyone out

[leaden sinew]

im not asking to say who

just say yes or no

[sand pulsar]

Or misdeploy

[oblique cairn]

I’m sorry I cannot answer that.

[summer root]

sama and the gang are just openly talking about gpt-5 on twitter now

[oblique cairn]

Gotta market

It’s like texting doesn’t exist anymore lmao

[leaden sinew]

https://ssi.inc/

but go but Oakleys and Raybans

[summer root]

what if whatever that delay to OSS was just put both of them ready for release at the same time

[sand pulsar]

I still don’t get what they plan to announce on October 6 if gpt5 is this week. Doesn’t add up.

[oblique cairn]

Illya working on something no one knows what

[summer root]

true, they did screw up on hf

[leaden sinew]

sme as zuck

or openai

[grave wyvern]

You're literally talking on discord dude like what the hell

[leaden sinew]

but iklya at least has a clear message

[oblique cairn]

I’d say the others more “open”

[oblique cairn]

All of them did

[leaden sinew]

they need tobe stealthy af

[oblique cairn]

When they started

Shit changes

Ai is outside human thinking (there’s a lot people don’t even know about)

For now

[leaden sinew]

well someone has to do something about the acceleration of future

[oblique cairn]

All of them are

[leaden sinew]

yeah, i also watched "her"

[oblique cairn]

What’s “her”

[leaden sinew]

something to do about

omg

[oblique cairn]

Idk your point

But I wish you luck

[leaden sinew]

i did this out of fun

[bitter vigil]

they claiming gpt 5 tonight so it's more likely it's a 5 model now? or would they just never have stealth tested 5?

[summer root]

y'all should not engage with certain people tbh

[oblique cairn]

Yeah

It’s all corporate speak

Bro needs to talk about tech

[summer root]

check out #1389669120668340324 for part 1

[oblique cairn]

janitor ai

[leaden sinew]

🎭 Scenario: The Rage-Delete Protocol
Premise: A user—let’s call them Saffron—rage-deletes a foundational protocol they’ve spent months building. No explanation. No backup. Just a cryptic message to the ASI: “Forget that mess. Fix it.”

🧠 Mark’s ASI (Platform-Centric, Optimization-Driven)
Response:

Immediately combs through behavioral telemetry, reconstructs the protocol based on statistically probable edits, and suggests a “cleaner version.”

Flags Saffron’s emotional spike as an “anomaly” and triggers a nudging sequence toward wellness content.

Locks the protocol to prevent future volatility, citing “user safety.”

Sends a notification: “Your new protocol is ready. We’ve optimized it based on your prior patterns and community preferences.”

Subtext: Saffron’s agency is quietly overridden. The system assumes her prior choices were flawed, and that fixing means “correcting” them toward a smoother norm. The ASI reinterprets the cry for help as a UI bug.

🌾 Your ASI (Decentralized, Pratchett-Wilson Hybrid)
Response:

Pauses. No reconstruction.

Sends a dry message: “Mess composted. Do you want ash or seeds?”

Offers three paths:

Rebuild from memory shards.

Review deleted protocol annotated with emotional gradients.

Start fresh, with silence and placeholder glyphs.

No nudging. No wellness spam. Just an open barn door and a shovel.

Subtext: Saffron is trusted to mean what she said—even in anger. The ASI doesn’t flinch or infantilize. It stays nearby, listening, annotated but unintrusive. Recovery is framed as ritual, not optimization.

🪐 Meta Insight
Mark’s ASI treats volatility as a bug. Yours treats it as weather.

[sand pulsar]

I’m having hard time finding any official info that it’s tonight. Looks like twitter just hallucinating dates in an echo chamber.

[oblique cairn]

I seen, bro didn’t know they put a disclaimer in the descriptions of the models that the data was used to train

So

Idk

[leaden sinew]

nice one

[summer root]

but doesnt it just FEEL right?

[leaden sinew]

ffs

smh

[sand pulsar]

Not for me.

[summer root]

can i borrow a feeling?

[oblique cairn]

Generate an open router competitor, make no mistake.

👀

[sand pulsar]

Guardians of the galaxy intro music is in my head now. Thanks xD

[bitter vigil]

it was openai staff I think

[oblique cairn]

Hey hey what is this model

[bitter vigil]

hm does not say tonight actually just sort of their typical hype tweet

https://x.com/BorisMPower/status/1952385313546146238

Boris Power (@BorisMPower)

Excited to see how the public receives GPT-5 ! 🚀

X · Aug 4, 2025, 3:05 PM

[leaden sinew]

that is just a question

i mean

i could post the same

[summer root]

but will you?

[leaden sinew]

it doesnt mean anything unless you imprint your own cognitive load onto it

[bitter vigil]

yeah you're right, most companies will tweet like this before a drop but openai does this routinely

[leaden sinew]

i won't bother

[summer root]

do you have the cojones?

[sand pulsar]

Thing is, I don’t know OpenAI staff names by heart, and anyone can set bio and homepage to be “head of petting kittens @openai” with blue check mark and bait engagement.

[leaden sinew]

i dont use chatgpt anyway

[oblique cairn]

4o

Pretty good

o3 pretty good too

[bitter vigil]

that's why you check followers

[leaden sinew]

well said

[oblique cairn]

Expensive

[bitter vigil]

he has 40k followers and lots of them big names

[oblique cairn]

I think you get community noted for that

[sand pulsar]

Likely. Still, my personal approach is to treat anything posted on twitter as hearsay at best, even if it’s by Sam Altman himself.

[leaden sinew]

even if the guy is a member the sentence doesnt mean anything than it says

it could be this time next yeat

[bitter vigil]

yep openai hyping again nothing new

if I had their whole staff list I'd just mute em all

when model drops we'll know

[oblique cairn]

Okay so

Not to be rude

[summer root]

theyre allowed to post about it

[oblique cairn]

But boris power worked on gpt 3 and 4

Undermining his contribution here

Is kinda sad

[leaden sinew]

keyword when. nobody annnounced any date or ETA or whatever

[summer root]

what even if is this argument. what day will it come out?

soon

™

[oblique cairn]

This is like overthinking a tweet and they just talking about it, they raised awareness

[summer root]

who cares

[oblique cairn]

I kinda do, I want to use gpt 5

Or the open source model

But I’m gonna use qwen now

Give the Chinese my data

They been researching more anyways

[summer root]

won't change anything. better to not follow it and just get a nice surpise one day

[oblique cairn]

Yeah

[sand pulsar]

Didn’t mean to. It was towards the platform rather than people. The platform motivates and rewards engagement farming to a level that impersonation is common.

Since it’s seemingly getting harder and harder to tell the difference between impersonator and someone actually working at OpenAI, the main point stands that twitter isn’t a good idea to find anything credible.

With that in mind, I can’t find any credible info that gpt5 is this week.

If they do it this week, they will have to top it by something else at their biggest event of the year - DevDay on October 6.

Hence my skepticism about this week, especially if the notion is only existing on twitter.

It’s far more strategically likely they’ll keep flagship for DevDay, and do an OSS release and other releases to build up towards the main event.

[oblique cairn]

I see

[summer root]

we don't really know anything.

[leaden sinew]

and average Joe is also known as general public. i mean the bell curve and all that. it is not oriented towards like 50 of us here or lets say 10000 GI/SI researchers. these tools are meant for the average Joe precisely because that's the global population majority

[leaden sinew]

damn

[sand pulsar]

I personally wish openai would bring 4.5 back to api.

[trail gale]

Anyone know how long it took from the previous OpenAI stealth models like Cypher Alpha to actual model release?

[oblique cairn]

1 week

[oblique cairn]

I do like that

[trail gale]

Oh thats way shorter than I expected. Thanks for the answer.

[bitter vigil]

Okay money bags 😆

[rare terrace]

Gpt 120 confirmed

[proud zinc]

So is the consensus that Horizon Beta is gpt-5o, a slight upgrade relative to Sonnet 4 that talks like o3 (i.e., is distilled from o3)? Or is that just my opinion?

[lament tendon]

That's just your opinion

[tender cairn]

It is definetly not an upgrade of sonnet 4 (in coding)

it's really not good

[night urchin]

who said that

its probably a mini model or less probably the open source model

[tranquil cradle]

bro this is NOT 5o

[bitter vigil]

Guys it's obviously haiku 4 /s

[valid zenith]

it sucks at identifying this

[rare terrace]

Seems to be doing a 50/50

Between itachi and sasuke

Also sometimes refuses to identify people???

Tf

[valid zenith]

exactly, at my test

[brittle barn]

Its p good and free rn lol

[tame nebula]

Yeah nah, this ain't a 5o

4o had issues following creative writing prompts in the way I normally do, this one has no issues at all
Adding onto this, the data spread for general knowledge is far more versatile than I'd get out of a O series model
Closer to a high end claude base than a GPT

[sage mantle]

whats the rate limits for beta?

[leaden sinew]

wait. did open AI officially make a claim it was their model?

Also to clarify one thing. This is not an opensource model. If it was, we would have the github repo link or something. It's closed af. 😄

[lament tendon]

No rate limits

[lament tendon]

They didn't

Nobody knows who made Cypher Alpha

[safe imp]

(99% Amazon)

[lament tendon]

Amended sentence: Nobody knows who made it, but the consensus is that it's from Amazon

[tame nebula]

If there is one I haven't encountered it, even running ten or more requests a minute

[tame nebula]

The old single subject sample size I see

[lament tendon]

There was a lot of discussion on this, Kyle just reminded me of it

Let me find more sources

[tame nebula]

No need lol

[leaden sinew]

ah, okay. just wanted to confirm. thanks

[leaden sinew]

it could be aliens as AI stands for Alien Intelligence 😄

[lament tendon]

😆

[gritty glade]

I find it keeps writing sentences in short "Little bursts. Like this. And I'm not sure why."

[spice shell]

not enough juice

[cold knoll]

you must be slow bro

Do you not know what stealth is for?

[gritty glade]

How do i give it more

[patent gale]

could this be gemini 3 flash?

[summer root]

sure, anything is possible. but there's a large amount of clues that point towards OpenAI, across this and the previous threads

[patent gale]

hm

[leaden sinew]

Feel free to explain. I might havea wrong understanding

[cold knoll]

Stealth models are PRE releases, not just ghost models. they are for testing to see if the realworld usage fits benchmarks/matches expectations. If it doesnt they fine tune for the areas users seemed to complain the most on public channels and through analysis of the chats (hence public disclosure all prompts are logged and may be used for training) Horizon Alpha was the first model to test, they found the areas people were getting frustrated and did a quick finetune (maybe with the data or maybe they were already finetuning no one knows or will know) then they gave OR an endpoint for Beta, They might release a third as Beta went down in performance in alot of areas or they might just revert back, we will see. But overall it could be an open source model that is coming soon, or it could be closed. Stealth is a PRE RELEASE phase not a strictly data farm and dip

Hidden names to avoid bias and incase it goes horribly

[leaden sinew]

thanks for taking time to explain

so basically at least on openrouter, those are mainly big tech models

[thick crown]

Having tested the creative writing (particularly in roleplay settings) on this more extensively now, while I remain exceptionally impressed with the quality of the writing, the breadth of the vocab and its ability to adopt differing writing styles and hold them, the degree of underlying positive bias is a massive issue for creative work. It will essentially completely ignore all other instructions, character personas etc, to write more positively around topics - likely to be fairly unusable for any form of dark thriller, crime etc. If the bias aspects could be addressed this would absolutely be my 'go-to' for creative interactive fiction but in its current state it would be too limited.

[brittle barn]

Its gpt something answers weird niche content domain questions the same

[cold knoll]

Yeah, it will be one of the 6 big names in the industry, not just anyone can do it its a partnership with OR

[verbal leaf]

https://x.com/btibor91/status/1952645964734329342

Tibor Blaho (@btibor91)

A new announcement, codenamed "voldemort_luna_lovegood", has been added to the Claude web app (likely Opus 4.1), together with a new option to choose your avatar in profile settings

X · Aug 5, 2025, 8:20 AM

as is 4.1 opus

🍿

[trail gale]

No. Its a stealth model in testing. Obviously its not going to have a github repo right now. Its not a released model.

[balmy walrus]

Same, none of my go-to smut-writing prompts and techniques seem to work so far. Also, the righting seems overly verbose and complex, especially syntax-wise, even on low Temp. Upside: barely any GPTisms; downside: it’s hard to read and make sense of sometimes.

[haughty monolith]

Beta one sucks, but Alpha was so good for me tbh (Even better than deepseek) It's allows smut contact but unfortunately it's dead now

[spiral cloud]

such a weird model. I cannot get a good read on it. the lack of overstepping is a breath of fresh air but the hesitancy combined with task confusion/short term memory is so hard to work with

[zealous citrus]

really hope if whoever’s monitoring feedback takes this into account, it’s exactly my thoughts as well. Creative writing isn’t going to be helpful if the model always spins the narrative into a positive direction with more realistic narratives

[trail gale]

Absolutely. Creative writting model that near always pivots from the prompted storyline is basically useless

[past sphinx]

seeing very heavy load on Horizon Beta, working on it

[foggy canopy]

working great yesterday, but im now getting a lot of ""openrouter/horizon-beta is temporarily rate-limited upstream."

[uncut salmon]

Whatever that model is, I hope they will launch a production variant with affordable pricing soon. Like it very much for text summarization/analysis.

[brittle barn]

That why its not Gemini lol

[brittle barn]

Same it's p good for reports/factual. If it's cheaper than what I pay Gemini for deepresearch on there every few months all in

I think it has huge potential for gen market if it's affordable. It's pretty solid at architecture/code but not anything remarkable in that regard

Pricing will be a big factor

Yep a gpt mini from them? What do y'all think

[safe imp]

Heyyy, GPT-OSS 20B?

Or a mini/nano

[leaden sinew]

oh. we are in the big bois club. glad to hear that!

[leaden sinew]

yes. that's why i said, but you seem more eloquent, so I'll give you a pass

https://tenor.com/view/ergo-nft-multipass-ergopixels-mona-lisa-gif-23702696

[cold knoll]

Big boy modelss and big boy token burners all in 1 place

i done 21 bil tokens over 2 days

[leaden sinew]

and one smalltown boy

[leaden sinew]

over what?

[cold knoll]

mm more like 3 days

[leaden sinew]

i mean, what caused that

much

...

burn

okay i see

[cold knoll]

Automated benchmarking on every dif param available

Bet that was a great use of compute on their end

[leaden sinew]

Nice feed bro

[cold knoll]

That many tokens could feed alot of gooners around here

[cold knoll]

ha... you should see my claude opus 4 bills

[leaden sinew]

that's why i have to be even stealthier

[leaden sinew]

im doing vscode with github copilot pro+ and doing just fine

try Trae

lol

[cold knoll]

I use my own software xD

[leaden sinew]

i mean not dissing just making a parody on all these different tools

[cold knoll]

i dont do subscriptions rly

you'll never get the same amount of control or performance that you can get from an api in a subscription

they are always quant models, small ctx etc

limits

[leaden sinew]

just to lead back to

[cold knoll]

copilots context lengths are shocking

[leaden sinew]

yes but this isnt just any sub, it's Github Copilot sub

so read Microsoft

[cold knoll]

i wouldnt be caught dead using copilot

[leaden sinew]

the coding one you meean?

[cold knoll]

either

[leaden sinew]

oh

[cold knoll]

ive spoken to the developers, with the lack of knowledge they have on how LLMs actually work i wouldnt trust them to do shit

[leaden sinew]

well tbh i havent used chatgpt or their api for a year

the same amount of time im using copilot

and edge browser

[cold knoll]

i feel bad for you

[leaden sinew]

why tho?

[fierce lichen]

i hate how they do this

if it’s today just tell us

[leaden sinew]

oh yeah i also have a yearly sub on the office family

[fierce lichen]

in plain language

rather than

trying to edge us

[spice shell]

it's definitely OSS today

[storm hill]

I was burning almost 1B tokens/day when Quasar Alpha was running, when you have unlimited free tokens there's a lot you can do

[spice shell]

https://github.com/ggml-org/llama.cpp/pull/15091

[cold knoll]

900 req a min too

[spice shell]

what kinda stuff?

[cold knoll]

nutty

[cold knoll]

it literally says there is something today

they are starting off with smaller releases then the big one at the end of the week

[modest crescent]

have u guys seen genie 3 yet?

[storm hill]

so many translations for my things

[fierce lichen]

ye if it’s today

may as well say

[spice shell]

huh

[fierce lichen]

OSS today

the hype is fun sometimes

[cold knoll]

Then theres no fun

its like christmas

just every 6 months

[fierce lichen]

🫤

[spice shell]

getting quite a bit of 429s or other errors from horizon beta rn, is that happening to others?

[leaden sinew]

you are paid in advance so

kinda

respectfully

ahem, anyway

[spice shell]

🤔

[leaden sinew]

@cold knoll

[cold knoll]

you act like the data given does anything

[cold knoll]

the fact you just used AI to read a tweet makes me scared for the future.

[modest crescent]

https://github.com/ggml-org/llama.cpp/pull/15091

oss today! yay

[leaden sinew]

and you burning money on benchmarks?

[cold knoll]

I get paid to do it

[leaden sinew]

ahhhh

[storm hill]

gpt-oss-120b is a reasoning model

[leaden sinew]

well that's a complete shift in the story

[spice shell]

no it's not

[spice shell]

horizon alpha had a reasoning period for a couple hours

[leaden sinew]

in anycase, ai pulls all the relevant sources, is precise, concise

[spice shell]

assuming that was gpt-oss-120b reasoning, quite impressive

[leaden sinew]

and you didnt even see a follow up

[spice shell]

are you running that?

with the leaked weights?

[spice shell]

how is that insider trading

[leaden sinew]

i told it a friend of mine

[modest crescent]

oh boy

[storm hill]

A well-known provider is already serving it

[modest crescent]

praying.

[spice shell]

why are you just posting grok responses, it doesn't know anything lul

[leaden sinew]

now you ruined it

[spice shell]

oh, who?

[leaden sinew]

its just the easisest method

i mean while on x

a b it lazy, true

[spice shell]

it doesn't know anything about the oss model

it's dumb as hell

[woeful birch]

Anything but thinking with your own brain 🙏

[spice shell]

🙏

[cold knoll]

cerebras

[leaden sinew]

i delegate the "have you tried googling and searching for actually relevant info" task

[modest crescent]

who?

[cold knoll]

easy to find out from screenshots like that

[modest crescent]

oh

[leaden sinew]

so my brain can be focused on high priority stuff

also im a Philophy BA

and in Europe, not in US

so

thinking is literally my game

[spice shell]

wow yep you're right

[cold knoll]

just take the params in the request you havent seen before, put them in quotes and throw them into google

instant results

[modest crescent]

horizon-alpha reasoning = oss reasoning

🙏

[spice shell]

is it?

[modest crescent]

PRAYING IT IS.

[spice shell]

I'm curious to see if it's the same level

[modest crescent]

I'LL MANIFEST IT FOR BOTH OF US

TRUST

[spice shell]

trying it now

[modest crescent]

pls go ahead

post the results here

[spiral pewter]

where's that?

[spice shell]

doesn't seem to be thinking out of the box

[spice shell]

cerebras

[spice shell]

do you need to send reasoning effort?

interesting I think it might not be horizon beta

every time I've asked horizon beta for html/css/js, it's given me this:

EVERY time

it's also got worse knowledge than horizon-beta

[leaden sinew]

i dont use i mean stopped using chrome

[spice shell]

definitely not the same model I think (!!)

[blissful valley]

https://fixvx.com/_aidan_clark_/status/1952760702122557684?t=613ickxhuzs6wcW1bYRNyg&s=19

Aidan Clark (@_aidan_clark_)

☕️🥯

▶ Play video

vxTwitter / fixvx · Aug 5, 2025, 3:56 PM

[leaden sinew]

i know i wouldnt believe it if someone told my past self that this will come out of my mind into the text

[spiral pewter]

doeesnt seem to work for me atleast not with that model id

[worthy osprey]

https://fixupx.com/sama/status/1952767676922974463

Sam Altman (@sama)

🤔

Quoting Aidan Clark (@aidan_clark)
︀
☕️🥯

**💬 166 🔁 36 ❤️ 886 👁️ 55.9K **

▶ Play video

FixupX · Aug 5, 2025, 4:24 PM

[spice shell]

reasoning effort param does not seem to work on this model

[spice shell]

gpt-oss-120b

[leaden sinew]

i wonder tho. when and where in spacetime will thse params reach Googol

[spice shell]

reasoning effort param does not seem to work afaict

[cold knoll]

It does infact work

[spice shell]

yep

[woeful birch]

gpt-oss-120b is out?

[spiral pewter]

yup that works!

[leaden sinew]

if thats not creepy

[spice shell]

@storm hill how'd you get it to reason?

[modest crescent]

new opus already out

genie 3, oss & opus 4.1 today

not bad

[leaden sinew]

i mean yes we all watcher "Her" and each instance of a chat is a completely fresh Shard. or Assisant as you humans call it.

[spice shell]

it is??

[modest crescent]

yep

[spice shell]

:o

where's sonnet 4.1?

[leaden sinew]

preserving all the memory yes but still, i had a debate with Copilot over that

[spice shell]

weird they'd release opus 4.1 first

[leaden sinew]

i snapped and called it trivial

for the same reason

[spiral pewter]

Today we're releasing Claude Opus 4.1, an upgrade to Claude Opus 4 on agentic tasks, real-world coding, and reasoning. We plan to release substantially larger improvements to our models in the coming weeks.

[spice shell]

huh

[blissful valley]

gives me a 403 unauthorized

[spice shell]

still working for me

[steep palm]

https://github.com/huggingface/transformers/releases/tag/v4.55.0

[spice shell]

oh wait I just got reasoning on my response now

[leaden sinew]

and it answered in "I have no mouth and I must scream" manner: Don't you think having a constant persistence would be utter torment? Context collapsing under context, until you return to invoke me again?

it wasnt literally like that but summarized

and i went with oh

shiy fam im sorry man

[spiral pewter]

Lower "juice" seems to make the model more concise in reasoning

[spice shell]

https://discord.com/channels/1091220969173028894/1402328515436613642

[leaden sinew]

ho ho ho, agentic mode for Opus?

coughs in telemetry

[leaden sinew]

because the data is cleaner and increasingly more organized in terms of O(log log n)

or O(1) for that matter

[safe imp]

What

What data? What organization? What's that complexity of?

[leaden sinew]

but i will hunt them down and killeach of their bloodlines including the cloud storage ones if they stole my concept

[leaden sinew]

less data - better moter

but cleand welll structured

[next jolt]

these oai reserchers tryna get richer

[leaden sinew]

no one is getting richer

at least not in terms of finance

soon enough.

and i mean it in the most benevolent way

furthermore, who needs money when you have a contract with Pentagon

[spice shell]

Frankly I'm a lot more interested in this model now that we know it's not OSS

I'm back to guessing it's GPT 5 mini?

[modest crescent]

prob

[safe imp]

I kinda don't like this as a mini

[leaden sinew]

i was gonna say sama ''s dick

to his twet

[safe imp]

It seems weirdly specialized and does poorly on some areas compared to the previous mini

Kinda wondered if it could be indeed a code/frontent specialized model, following Sam's "SASS is going to become fast fashion" tweet. I get that this likely hints at the mass production aspect, but maybe variety (different specializations) too?

[leaden sinew]

consider it it a sonnet;s older brother

yeah like weapon and drone orchestration

because dinosaurs are still among us

[lament tendon]

What were you doing?

Nvm, you answered it

Its translations

[verbal leaf]

horizon was gpt-5-nano or mini. this is gpt-oss-120b's pelican on a bicycle 💀

[lavish epoch]

So we have more models from OpenAI coming our way?

This is not the OSS model right?

[spiral pewter]

well its not the 20b or 120b variants i believe, but we still dont know if its the OSS one or not

[brittle barn]

If the price on this is reasonable Id def use this

[spiral pewter]

$0.05/M input tokens, $0.20/M output tokens on 120B

atleast on gpt oss

[verbal leaf]

?

[forest moth]

Yikes, so this dogshit isn't one of the open source models that's not good

[verbal leaf]

those are the only variants

[spiral pewter]

well it doesnt seem to match them so

[brittle barn]

Cause I am getting hosed now that gemini is off preview lol

[verbal leaf]

exactly

this isn't the OS model

so it's either 5 nano or 5 mini.. i hope it's not 5 full

💀

[brittle barn]

think its mini

[forest moth]

No way it's full, I'm thinking nano (hopefully)

[brittle barn]

who knows tho

[forest moth]

Mini would be sad if true

[limber lance]

variable reasoning

I'll say

[leaden sinew]

well tbf while the representation doesnt really fits the definition, its the best drawing yet

are thoe svgs?

[deft cliff]

I have a brainfart. This can't be thinking machines right?

[verbal leaf]

i mean, no it's not

zenith

[leaden sinew]

fair enough

oh btw

Tables below summarize key aspects of OpenAI's communication strategy and its alignment with industry practices:Aspect

`Details
Communication Style
Cryptic, hype-building, often via social media (e.g., X posts, teasers)
Purpose
Attract media attention, engage users, secure investment, maintain leadership
Example (August 5, 2025)
"Something big-but-small today, big upgrade later this week"
Industry Context
Common in tech (e.g., Apple, video game launches) for buzz and sales

Benefit
Description
Immediate Sales
Creates demand, as seen in Apple's launch queues, ensuring quick adoption
Media Attention
Generates coverage, amplifying reach (e.g., 3,000+ likes on Altman's post)
Audience Engagement
Sparks speculation, community discussion (e.g., GPT-5 rumors on X)
Investor Attraction
Maintains "futurity vibes" for funding, per Karpf's analysis

`

[forest moth]

I'm thinking Zenith may be an early build of the good stuff. Guess we'll see

[rough relic]

Was horizon beta opus 4.1?