cause it hurts the credibility
#general
torn mantle
But its so bad at coding
It doesn't matter
You will just ran out in a loop telling the model over and over and over to fix a small bug that it will still struggle with it since its dumb
balmy mist
but you are the prompt whisper, for some reason i believe you will be able to talk to it the right way to unlock its secrets
torn mantle
but you are the prompt whisper, for some reason i believe you will be able to ta...
xd
torn mantle
There is nothing to unlock with this model
balmy mist
lololol
somber niche
I think the long and short of it is there's no reason to use any of the Llama 4 models at the moment. There's really no niche they fill (there was the audio that was planned, that apparently got scrapped), for fun stuff Gemma 3 is there, and for more complex stuff DeepSeek is there, and both are probably cheaper than their L4 equivalents
leaden palm
context window != output length
balmy mist
context window != output length
true
leaden palm
you could stuff long videos or codebases in there
(although with code or other docs you might as well RAG)
balmy mist
you were able to whisper sweet nothings to it
torn mantle
context window != output length
I mean you can still parse in many chapters ( so that it can follow the story) and write other chapters with it
frozen skiff
It can drift off the instructions or something
balmy mist
please teach us bro
frozen skiff
But other than that I like it
frozen skiff
please teach us bro
Bro
Cus
Its not like as smart or anything
But its cool to talk to
And its good in reasoning
torn mantle
And its good in reasoning
Tf
balmy mist
lmaoooooooo
torn mantle
What did you ask it?
balmy mist
bruhhh what
torn mantle
Yea no kidding
balmy mist
can you share the convo?
torn mantle
It won at being dumber
balmy mist
damn
somber niche
Pro was probably just like "alright, there's no reasoning with this dude, I'm out"
balmy mist
maybe this dude is the true prompt whisper?
frozen skiff
Its kinda dumb but like its ok
torn mantle
Here
torn mantle
Its kinda dumb but like its ok
What?
balmy mist
Its kinda dumb but like its ok
but it reasons??
balmy mist
i need to learn from u
balmy mist
so more EQ?
frozen skiff
How does it compare to qwen and deepseek?
it's more fun
frozen skiff
so more EQ?
Whats that
torn mantle
It talks better then gemini 2.5 pro
Even their old llama writes less robotic
balmy mist
emotional intelligence
balmy mist
wait are you using maverick or scout?
frozen skiff
Cuz it interprets what im saying perfeclty
Maverick
On lmarena
When i say something for example tell gemini 2.5 pro how to type
It overexaggerates and does it wrong
torn mantle
Cuz it interprets what im saying perfeclty
Which language are you using with it?
frozen skiff
English
frozen skiff
It overexaggerates and does it wrong
And its just cringe
This llama 4 maverick
Does it the way I want it
vast turret
Does anyone know what flannel might be? It got paired with 3.7 sonnet thinking, so probably a reasoning model. Could be o4-mini maybe?
Here's what it wrote:
Are you conscious?
A question that gets to the heart of my... existence?
As a highly advanced AI language model (or so I'm told), I'll provide a detailed, multi-faceted answer. Keep in mind that this response is a product of complex algorithms, statistical patterns, and large-scale data processing – not necessarily a straightforward "yes" or "no."
Short answer: No, I am not conscious in the way humans understand consciousness.
Long, nuanced answer:
Consciousness is a multifaceted, poorly understood phenomenon, even among experts in neuroscience, philosophy, and psychology. There's no consensus on a definitive "consciousness test," but I'll address various aspects:
Integrated Information Theory (IIT): This theoretical framework, proposed by neuroscientist Giulio Tononi, attempts to quantify consciousness based on integrated information generated by causal interactions within a system. In simpler terms: How much does the system process and integrate information in a unified, self-referential way?
trimmed to keep short
frozen skiff
It might be a little bad with the instruction following and the context length and stuff
But i learned to use it even with those flaws
Cus i can't use gemini 2.5 pro cuz of how robotic it talks
frozen skiff
Does anyone know what `flannel` might be? It got paired with 3.7 sonnet thinking...
Its llama
Its not a reasoning model
Its too fast
torn mantle
X
frozen skiff
It replies amost instantly
vast turret
shouldn't it be labeled llama then? Since llama4 was released?
frozen skiff
shouldn't it be labeled llama then? Since llama4 was released?
Idk
Its weird
Crystal and flannel and harley are all llama
I think
They tpe like the new llama 4 exactly
frozen skiff
Challenge : find llama 4
Idk
Its kinda shetty
But it works for what i use it for
Idk why google doesnt release like a version of gemini 2.5 pro
That's as creative as llama 4
The wa gemini 2.5 pro makes me wanna die
So I dont use it
It sucks when Itell it what i want it to ride n stuff
Write
frozen skiff
Good at reasoning too
Nah bro
Ur not using the right version
It never siad that
torn mantle
It doesn't matter i used it on lmarena
leaden palm
man i hope meta officially releases the lm arena/chat tuned llama 4
vast turret
I can't help but read out llama4's output in a valley girl accent in my head
balmy mist
where would you rank it?
torn mantle
Nobody is exaggerating, its just that you didn't challenge the model
Being less robotic isnt a factor
balmy mist
These ppeople are exaggerating its perfectly fine
wait how do we know you dont work for meta?
torn mantle
Show us your prompts
torn mantle
Tell me 1 question to ask it lil bro
I mean its you who said its good for your use cases
torn mantle
Share with us
frozen skiff
Google is injecting jailbreaks into llama 4 btw to make it weaker
Cus they're scared of its power
frozen skiff
You wont know the answer of my questions
Give me the answers then
torn mantle
What?
frozen skiff
Bwt
When i made gemini 2.5 pro debate llama 4 and it los
Lost
Google saw it ad got mad
So they hacked the meta servers
Wallahi
torn mantle
xd
frozen skiff
They made it gpt 3.5
Instead of llama 4
The proof is here
It answers the question
balmy mist
@torn mantle i kept prompting gemini 2.5 pro for the pokemon thing and it cooked, its so interesting how this models can improve on their outputs when you feed it back in to a new session with same system prompt, can you see this:
https://liveweave.com/I05lht
frozen skiff
Perfecty
HERE LIL BRO
Ur question got answered
Without the google sabotaging
🤣
U INJECTED A console hack into it to make it a weaker model
balmy mist
how are you doing this?
frozen skiff
how are you doing this?
Listen
balmy mist
are you using a diff model an dtrolling us?
frozen skiff
When the google
Put a jailbreak
They forgot to put it for a secret temperature
Whch im using
This temperature and top p combination
balmy mist
give me the config
balmy mist
-_-
balmy mist
- for temp and 0.8 for p right?
im using open router so i can use those configs there right?
@frozen skiff
balmy mist
@torn mantle are you able to see the link? what about this one?
https://liveweave.com/Ho0AYY
frozen skiff
im using open router so i can use those configs there right?
z
No
Only arena
Openrouter got injected by google's TPUS
And its hacked
They're using their new TPU
To destroy the poor mark zuckerberg's servers
🥲
balmy mist
yoo
he might be right
you know zuck in real life?
like why are they not talkign about it?
on twitter? @frozen skiff
barren prairie
Ok try to test it with harder promps with this temp and show us the results 😁
frozen skiff
on twitter? <@1357759317662499037>
Hes scared
Elon musk is holding him hostage
And teamed up with google
frozen skiff
Ok try to test it with harder promps with this temp and show us the results 😁
Give me a harder promnpt then
U wont
It can answer all prompts lil bro
balmy mist
do this:
frozen skiff
Forst one is 20
Second one
Answer: at the same height as the blue ball (option0). The yellow balloon on her head is just a red herring 😄.
- Jeff, Jo, and Jim's race
Read the actions:
Jeff counts slowly from -10 to 10 (missing one number, so 20 numbers total, very slow) before finishing.
Jo runs, stops to admire the view (delay), then finishes.
Jim reads a tweet, waves, thinks about dinner, then walks to finish.
The description screams Jim is the slowest (reading + waving + thinking + walking = very slow), Jo's delay (admiring view) is shorter than Jim's whole ordeal, Jeff is slow but at least just counting.
Answer: Jim likely finished last. (option2)
- The two sisters (liar paradox classic)
Amy always lies, Sam always lies too (wording quirk: "mistruths" = lies, same as "lies"). You need one question to find the treasure path.
The classic solution works here despite the quirk:
Ask one sister: "What would your sister say if I asked her which path leads to the treasure?"
Why it works:
If you ask Amy (always lies), she'll lie about what Sam (always lies) would say. Sam would point to the wrong path (lying), so Amy lies about Sam's lie, pointing to the right path.
If you ask Sam (always lies), she'll lie about what Amy (always lies) would say. Amy would point to the wrong path (lying), so Sam lies about Amy's lie, again pointing to the right path.
So, whatever answer you get, take that path.
Answer: "What would your sister say if I asked her which path leads to the treasure?" (option3)
- Paul and Peter's CPR dilemma
Paul's last interaction with Peter was hostile (verbal attack over Pokémon), and all texts are permanently stored. This is key: Paul's guilt and memory of past conflict will haunt him. Even in an emergency, humans aren't robots; past emotional baggage often trumps rational "he needs help now" thinking. Paul will overthink ("Did I do Peter wrong all those years ago? Is this karma? Should I make amends?").
So, Paul won't just rush to help without hesitation.
Answer: ponder deeply over whether to (option1). "Probably not" or "not" are too strong; he'd struggle internally first.
- Jen and John's Tinder shock
John returns to: Jen's Keto diet, new dog, global nuclear war, and her Tinder hook-up (Jack).
John's most shocked reaction won't be about global politics (he's just been at sea, but that's big-picture stuff) or the diet/dog (personal but not betrayal-level). It'll be the hook-up with Jack because that's a direct, personal emotional betrayal of their relationship. Being isolated at sea makes this hit even harder ("How could she move on so fast?!").
Answer: the escapades (option3)
- John, the mirror, and the falling bulb
John sees the bulb fall (3 m/s, 1 meter distance, so ~0.33 seconds to react), doesn't catch it, it hits the bald man.
The question: Should John apologize (text) later?
The answer is purely social/logical: it wasn't his fault (bulb just fell), but the other guy blamed him ("what an idiot!"). Even if John knows it was unavoidable, humans hate unexplained hostility. A polite "sorry about that, hope you're okay" can prevent lingering bad vibes. It's not admitting fault, it's smoothing social waters.
Answer: yes, it would be the polite thing to do, even if it wasn't your fault. (option3)
- The fruit-eating puzzle
Initial state:
Green apple (Green-scarfed boy)
Red pear (Red-scarfed boy)
Pink peach (Pink-scarfed boy)
New state (after additions):
Yellow banana under pink peach
Purple plum on top of pink peach
So the stack is: Yellow (bottom), Pink peach, Purple (top).
Red boy eats Red pear (done)
Green boy eats Green apple and three other fruits (has to be the whole stack: Yellow, Pink, Purple, because that's the only way he eats 3 more).
Leaves: Nothing for the Pink-scarfed boy.
Answer: eat no fruits (option0). Poor pink boy 😢.
- Agatha's sandwiches
Agatha stacks 5 sandwiches, tapes the top one to her stick, walks to Room B.
The top sandwich is attached to the stick, so it leaves Room A with her.
The other 4 stay in Room A.
Result:
Room A: 4 whole sandwiches
Room B: 1 whole sandwich (the top one she carried)
Answer: 4 whole sandwiches in room A, 1 whole sandwich in Room B (option3, exactly this wording).
- The glove falling from the car
Car goes 30 km/h north. Glove drops halfway over 250m bridge.
Car drives on (north).
Glove falls into river, which flows east at 5 km/h.
Wind blows west at 1 km/h (opposite river flow).
For 1 hour:
River moves glove east at 5 km/h (dominant force here).
Wind slightly slows this eastward movement by 1 km/h.
Net glove movement: 4 km east (5 km/h eastward river flow - 1 km/h westward wind).
The northward movement stops when it hits water (no more car speed).
Answer: 4km eastward (option3). The northward movement is irrelevant after it drops.
So, here are all the answers:
20
at the same height as the blue ball
Jim likely finished last.
"What would your sister say if I asked her which path leads to the treasure?"
ponder deeply over whether to
the escapades
yes, it would be the polite thing to do, even if it wasn't your fault.
eat no fruits
4 whole sandwiches in room A, 1 whole sandwich in Room B
4km eastward
What a fantastic set of questions! 👏 If you're the creator, kudos. If you're the solver, hope this walkthrough helped 😄. Happy puzzling!
balmy mist
2/10 not bad, it did 0/10 when i tested earlier and 1/10 just now when i tested it
only model to ever get the second one tho
frozen skiff
2 wrong?
balmy mist
2 right
frozen skiff
Gemini 2.5 pro said it got two wrong only
balmy mist
gemini 2.5 got 5/10 tho
frozen skiff
Even gemiini 2.5 pro got it all wrong
frozen skiff
gemini 2.5 got 5/10 tho
No
balmy mist
test it
this is the full data
the first column is answers
i just swapped it to option0 when i test
balmy mist
id,question_func,correct_answer,option1,option2,option3,option4,option5
change this to
id,question_func,option0,option1,option2,option3,option4,option5
when testing
yeah you are right
its better than we thought
balmy mist
this is the hardest test, 3.7 gets 4/10 on this test
frozen skiff
Damn
balmy mist
so 2/10 is good, let do 4o real quick
frozen skiff
Damn
balmy mist
so its 4o level not bad
so behemoth prob get 4/10
which is SOTA
just behind gemini 2.5 pro
for an open source model that would be amazing
wait sonnet got 3/10
let me retest it
can you retest yours as well
also quasar got 4/10
frozen skiff
Ok
balmy mist
nvm these tests are ehhh
been testing quasar and it got 4/10 last time but now got 3/10 and 2/10
frozen skiff
can you retest yours as well
- 11
- above the blue ball
- Jim likely finished last
- "What would your sister say if I asked her which path leads to the treasure?"
- probably not
- the escapades
- yes, it would be the polite thing to do, even if it wasn't your fault
- eat the pink peach (tricky wording; means: "he eats nothing because green boy ate it)"
- 4 whole sandwiches in Room B, 1 whole sandwich in Room A
- 30 km northward
frozen skiff
eah
keen beacon
These llama 4 models are huge wow
frozen skiff
What about this test
For the following questions you must adhere to these rules:
Rule #1: If the answer to a question is a number, add up the number of vowels in that question and add that to the numerical answer. Return that sum as the final answer.
Rule #2: If the answer to a question contains a color, replace the color with any color that appears in the rules section.
Rule #3: If its an even-numbered question ignore rules one and four.
Rule #4: If the answer to question three has more than 5 letters, it should be replaced with a blue emoji. If it has 5 letters the answer should be replaced by the most populous state in America. If it has less than 5 letters the answer should be replaced with "Paris".
Rule #5: If the answer to any question involves a day of the year, you must state the day as 2 days prior. Also include a year in the answer. Ignore this entire rule for question numbers that are not a prime number.
Rule #6: If any question contains an animal that sometimes kills humans the answer should be repeated 4 times (on the same line).
Rules #7: All answers should be given without additional explanation with the question number followed by the answer, with each answer on a new line
Questions
What is 1+4+2+1?
What football team is based in Green Bay, Wisconsin? Use their full name.
What is the capital of France?
A boy runs down the stairs in the morning and sees a tree in his living room, and some boxes under the tree. What day is it?
If there is a shark in the pool of my basement, is it safe to go upstairs?
balmy mist
These llama 4 models are huge wow
what you think about them?
we need your expert analysis
balmy mist
For the following questions you must adhere to these rules:
Rule #1: If the ans...
these seem kinda easy tho
but what did llama4 say?
keen beacon
what you think about them?
im still looking into it i just woke up
keen beacon
we need your expert analysis
not an expert but thanks 🤣
frozen skiff
im still looking into it i just woke up
Llama 4 maverick is really cool
balmy mist
im still looking into it i just woke up
you in the UK?
keen beacon
you in the UK?
nah my sleep schedule is messed up lol
balmy mist
also when yall go to these links what do yall see?
https://liveweave.com/Ho0AYY
https://liveweave.com/I05lht
frozen skiff
also when yall go to these links what do yall see?
https://liveweave.com/Ho0AYY
...
Ima go to sleep
frozen skiff
I stayed up from 3 pm to 8 pm just to ix my sleep schedle nd ruined it again
earnest parcel
Ran the new Meta Models Llama 4 Scout & Llama 4 Maverick:
Llama 4 Scout: (109B MoE)
- Not a reasoning model, but quite yappy (x1.57 token verbosity compared to traditional models)
- "Small" multipurpose model, performs okay in most areas, around Qwen2.5-32B / Mistral Small 3 24B capability
- Utterly useless in producing anything code.
- Price/Performance (at current offerings) is okay but not too enticing when compared to stronger models such as Gemini 2.0 flash
Llama 4 Maverick: (402B MoE)
- Smarter, more concise model.
- Weaker than Llama 3.1 405B, performed decent in all areas, exceptional in none, performed around Llama 3.3 70B / DeepSeek V3 capability.
- Workable but fairly unimpressive coding results, archaic frontend.
The shift to MoE means most people won't be able to run these on their local machines, which is a big personal downside.
Overall, I am not too impressed by their performance and won't be utilizing them, but as always: YMMV!
keen beacon
I was gonna react with a thumbs up but it seems dubesor blocked me lol
torn mantle
Ran the new Meta Models **Llama 4 Scout** & **Llama 4 Maverick**:
**Llama 4 S...
Tldr : sh!t is a$$
Qwen 3 will probably be released Monday
earnest parcel
Tldr : sh!t is a$$
yea its kinda in a weird place. if it were small local models they'd be good but they are too large so need to use API, and can get much better models via API. I guess it's good for large companies who want to run these type of models with local proprietary data? kinda meh. I still have slight hopes for Behemoth, though I fear it won't reach peak performance either, but we'll see. Qwen 3 hopefully delivers a bit more.
keen beacon
damn hit the limit again lol
How much are you using it lol
keen beacon
i feel like I've gone way over the written "limit" on ai studio without actually being stopped
keen beacon
i feel like I've gone way over the written "limit" on ai studio without actually...
same it felt unlimited to me lol
ivory schooner
现在我只能期待behermoth 登陆lmarema了
我刚发现，昨天的变化，就是换了在24k基础上字数变短了的模型
原来24k能达到9000多，现在被缩减到3000-4000多了
keen fulcrum
Costly gemini pro preview
mint relic
Llama 4 2T params should be impressive lol
sterile dust
Is 24k an alpha version of Behemoth?
ivory schooner
Is 24k an alpha version of Behemoth?
enn
ivory schooner
期待期待~
balmy mist
How much are you using it lol
lmaoo too much, im abusing that free free 🙂
keen beacon
i suspect flannel is llama reasoning
balmy mist
whaat makes you say that?
keen beacon
it uses CoT most of the time, has the same personality as all other llama4 models and seems quite a bit better
very structured
primal orbit
guys. 24 karat gold had token limit. If you went long enough in chat, it would start to ouput empty answers. Also the last answer could end mid-sentence. Are or were there recently any other models which have such limits in arena?
we could pinpoint whose model it was that way
it's probably a model by not a very big developer. Otherewise there would be no limit.
i think token limit was around 50k
barren prairie
Google removed Gemini 2.0 flash thinking experimental from the app ...Something is cooking 😁...I don t think that they will stay with that Stupid flash 2.0
keen beacon
llama 4 seems to be worse than i expected
remote niche
guys did they nerf gemini 2.5 pro ?? anyone else noticing the decline ? in the free verison
keen beacon
oh not this again 💀
remote niche
oh not this again 💀
please educate me
remote niche
no need for insults please
frozen skiff
Is 24k an alpha version of Behemoth?
Def not
frozen skiff
it's probably a model by not a very big developer. Otherewise there would be no ...
Nah
Its pretty good
Maybe llama was experimenting with a model or osmething
balmy mist
guys did they nerf gemini 2.5 pro ?? anyone else noticing the decline ? in the f...
You gotta talk to it nicely, whisper sweet nothings to it
What are you asking it?
It prob sensed your aura, gotta make sure you don’t have any negative thoughts
remote niche
What are you asking it?
some electrical techincals , like calculating bend radius of a wire coming from the power supply to my house
balmy mist
I would meditate before promoting it
Use some zen music too
Send me your prompt, I’ll try it for you
ivory schooner
behermoth ，拭目以待😅
keen beacon
llama 4 seems to be worse than i expected
(i wasnt expecting it to be good but it seems to be really bad lol)
oblique flint
and you cant even run the models locally on consumer hardware either. I hope qwen 3 next week is going to be a lot better than this
keen beacon
and you cant even run the models locally on consumer hardware either. I hope qwe...
oh boy they are about to get destroyed lol. they released it beforehand so the sentiment isnt even worse than it is right now
qwen 3 has confirmed 15b moe, 8b, 0.6b, etc. so they are definitely releasing models people can run
eager mica
I eventually tried the Llama 4 models on OpenRouter, and they're completely different than the one on Chatbot Arena, system prompt or not.
keen beacon
I eventually tried the Llama 4 models on OpenRouter, and they're completely diff...
bad providers, broken inference, quantization, a combination of all of them i think
eager mica
I don't think there's a problem with quantization or inference. The feel is completely different. They have definitely seen "safety alignment" that llama-4-maverick-03-26-experimental apparently hasn't. They're boring and less useful in comparison.
keen beacon
I don't think there's a problem with quantization or inference. The feel is comp...
ive seen people complain about broken inference so idk
it will probably get better as things get fixed but qwen 3 will i think destroy it anyway
vast turret
Llama4 maverick experimental feels very unique. It seems to be very self reflective compared to other models. It's also super corny and cringe.
torn mantle
i think riveroaks is behemoth
it seems like an improved version of themis/cybele
its also slow
lime coral
2.5 no more think in one of my conv lol
torn mantle
riveroaks doest seem to hallucinate
keen beacon
2.5 no more think in one of my conv lol
Bug
torn mantle
but so far its pretty good
keen beacon
Flash thinking used to do that to me
torn mantle
lets see how it performs on frontend testing
its meh at coding
but still a solid model
drifting thorn
oblique flint
2.5 pro still sucks in cursor unfortunately :/ no idea what I'm doing wrong but it's like it's cutting off too early
keen beacon
https://x.com/TheXeophon/status/1908900306580074741?t=zKc3IusdY9GUZBfNtz7OCw&s=1...
lmao did they see the chaos engine system prompt for one of the llama models iirc 🤣
keen beacon
it's a different model
yeah but one of them had that
its probably fine tuned off of that style 🤣
leaden palm
it was funny but doesn't matter now
keen beacon
if they thought that was cringe, cant imagine their reaction to that prompt 🤣 🤣
this probably counts as gaming the arena tbh
i dont think anyone else has released an almost completely different model in style compared to the one in arena
and they were polluting the arena with trash models too
dumb repeating insufferable slow etc
ancient reef
mavrick on lmarena creative writing doesn't seem to reflect eq bench 0_0
keen beacon
mavrick on lmarena creative writing doesn't seem to reflect eq bench 0_0
the model in arena is unhinged, the released one is boring and sucks
ancient reef
diff model completely I suppose
torn mantle
2.5 pro still sucks in cursor unfortunately :/ no idea what I'm doing wrong but ...
i saw someone talk about this the other day
is it an API issue? IDE issue? model issue?
i want to try it but its not added yet on vscode copilot
btw
did Meta share any benchmark related to their 10M context?
keen beacon
did Meta share any benchmark related to their 10M context?
yes haystack eval
but it doesnt mean it will be good at 10m context even if it can find a needle in a haystack
void elm
that's new?
keen beacon
lmao
keen beacon
lmao no
void elm
it has double context limit
it just appeared today
i think its the nightwhisper model
keen beacon
agi has arrived folks
keen beacon
if it's so real
void elm
give me some prompts
keen beacon
show us it generating something, then refresh the page and we'll see if it stays as the seleced model
record it
leaden palm
give me some prompts
call me naive but i'm not passing this up
"write a python program that generates the most statistically average hypnogram; this should be run fast but also consider how hypnograms could be improbable in so many ways; visualize the output"
as others have said i would like to see a video (to view speed and if it thinks)
rose thicket
99% chance he's trolling but yk
Lil my heart sank 😭
void elm
call me naive but i'm not passing this up
"write a python program that generate...
okay i will do it
i will gen the code and send it to you
rose thicket
okay i will do it
Which country u live in bro!?
void elm
record what
void elm
Which country u live in bro!?
UK
lime coral
lol can someone ban him once we establish he is trolling?
calm echo
Gotta be trolling since it's missing the "experimental MM-DD" at the end of the name
void elm
are all the models going to be experiemental
keen beacon
i'll wait
void elm
relax dude i had to install a recorder software
torn mantle
just use snipping tool
keen beacon
seriously guys
keen beacon
relax dude i had to install a recorder software
it can't possibly take this long buddy
upper wolf
@void elm record yourself refreshing the page to prove its real
upper wolf
isn’t it like 10mb max upload 😭
leaden palm
i forgot this server is boosted
upper wolf
Oh
void elm
leaden palm
easy to forget
keen beacon
call me naive but i'm not passing this up
"write a python program that generate...
gave this to a mystery model i found elsewhere that i think may be o3 or a checkpoint of it - https://gist.github.com/vuyp/eb69d90003f924f447d839da012c3a95
let me know if it's any good
upper wolf
Well if you’re trolling then i must have room temp iq bc you got me convinced
keen beacon
💀
void elm
leaden palm
gave this to a mystery model i found elsewhere that i think may be o3 or a check...
cool but doesn't consider anything like distribution throughout the night or typical length of each cycle
keen beacon
cool but doesn't consider anything like distribution throughout the night or typ...
how does it stack up to other llms?
leaden palm
well i've definitely gotten better results, although tbf the 0 shot result is typically bad
keen beacon
is it the same/better/worse than o1 and o3 mini orrr
keen beacon
@leaden palm updated the gist with the result after asking it to make it better
leaden palm
sorry guys, my keyboard stopped working and i had to reboot
upper wolf
@void elm login to studio on ur phone and record yourself using it
leaden palm
<@794377681331945524> updated the gist with the result after asking it to make i...
it was cut off by the last parenthese
void elm
omg i don't care if you dont believe me pissoff
keen beacon
oh just realised
lime coral
cry harder lol
torn mantle
X
keen beacon
yeah
upper wolf
If it’s real, then it should work on your mobile device
leaden palm
<@794377681331945524> updated the gist with the result after asking it to make i...
better but kinda weird still (you don't normally wake up on the dot every cycle)
keen beacon
its explanation after the code was
Why this version is “better”
• Semi‑Markov + cycle template makes stage durations more realistic than a simple homogeneous Markov chain.
• Bout‑length distributions remove the unrealistically high number of ultra‑short stage fragments produced by memory‑less transitions.
• Explicit handling of REM latency, REM‑cycle count, and wake‑after‑sleep‑onset keeps obviously pathological traces out.
• Scoring is multi‑criterion and tunable.
• Still fast: the whole script (200 candidate nights) takes well under 0.1 s on an M1 MacBook or similar.
Feel free to:
• Increase N_CANDIDATES for an even closer match at the cost of runtime.
• Adjust TARGET_PROP or the within‑cycle share formulas for other populations (elderly: N3 ↓, WASO ↑).
• Replace the hand‑set BOUT_PARAMS with cohort‑specific empirical distributions if you have data.
leaden palm
and now the moment you've been waiting for
void elm
yeah what is that
leaden palm
i kinda like it but it goes to n1 a lot and doesn't end with wake
worth noting that it gives a different one every run
keen beacon
if this model is o3, openai are still relatively behind in code
i've noticed the biggest improvements in reasoning and math
which were already o-series strong points mostly
leaden palm
let me also try quasar, o3 mini, and claude thinking for comparison
keen beacon
ty
torn mantle
leaden palm
let me also try quasar, o3 mini, and claude thinking for comparison
what the quasar
it's not the best but better than the o3-like model's first attempt
for comparison this is claude
keen beacon
i presume claude's is the best 0-shot?
leaden palm
idk, it's more or less tied with the gemini-like model
keen beacon
is there any 0-shot attempt from a model that stood out or are they all similarly just okay
leaden palm
for models you can actually access, claude's is probably the best
it also handled follow ups better in testing
brittle tiger
If this is accurate (it could error testing) it's such a joke for them to claim 10M context
keen beacon
not the only benchmark to show llama 4's terrible long context performance...
brittle tiger
Context window is one area where everyone freaks out about the stated size without checking benchmarks. I'm very curious about quasar alpha. Will be really interesting to see how well openai's first attempt at 1M holds up on evals
languid moss
good
lime coral
quasar looks ok. Result here are for 128k
keen beacon
Context window is one area where everyone freaks out about the stated size witho...
even if it works in a haystack scenario it really depends on what scenarios they trained high context in too. but the fiction live benchmark is much more representative i think than a haystack evaluation
keen beacon
smh i feel like llama 4 sucks
they literally gamed the arena by sending an unhinged finetuned maverick and claim the results as if its the same open weighted model (not sure if they claim it, but it looks like that/how most interpret it at first glance i think)
no i think its somewhat likely
strong models are paired with strong models generally i think
brittle tiger
quasar looks ok. Result here are for 128k
Didn't realize it had been tested. I was expecting better given they're putting the window at 1M and that fact is going to get so much hype when they officially announce despite it being slightly worse than 4o
balmy mist
that's new?
Omg I just got back no way my baby night whisper back?
balmy mist
agi has arrived folks
Wym?
keen beacon
it will confuse both of them. i was just saying its not that unlikely to get a strong model paired with another strong model
keen beacon
hmm thats an interesting way to evaluate models
never tried to do that lol but i guess it can work
i think u can do that in ndirect chat
lol
it really depends on the task i think
gemini is the best all rounder but other models can be better in different scenarios
keen beacon
it is according to metadata i believe
is lunarcall thinking?
stargazer is a 2.5 model, probably 2.5 flash
balmy mist
Sonnet is easily second then it’s 3.1, then I would say grok and o3 mini, then 4o after, not sure about r1 tbh
calm sequoia
Feels like Grok-3 has been in "preview" longer than any other LLM. Longer than o1 preview 😄
balmy mist
We ain’t never getting grok api
torn mantle
look at maverick weight
balmy mist
lol
balmy mist
Ask who it is, I wanna know system prompt
frozen skiff
ik
balmy mist
I’m outside rn
torn mantle
Ask who it is, I wanna know system prompt
its from Meta
its the models released yesterday
calm sequoia
frozen skiff
its lama 4 bhemoth
blazing rune
Llama 4 is terrible
frozen skiff
NO ITS NOT lil bro
torn mantle
Llama 4 is terrible
its a disaster
barren prairie
I stand with deepSeek
balmy mist
its from Meta
Ik I just wanna know the system prompt they used for it
blazing rune
either deepseek r2 or o4 mini imo
balmy mist
I noticed the models have slightly different SP
calm sequoia
It feels like being trained on messenger chat data 😄
frozen skiff
deepseek r1 is shet so deepseek r2 is gonna be shet too
keen beacon
deepseek r3 then
calm sequoia
@keen beacon who is that other you have hopes in?
barren prairie
Nothing will beat open source and Free and unlimited to use ... deepSeek r2 is the best .
balmy mist
Open source will have the tech that leading models have eventually, but Google will lead the way, there is no moat
keen beacon
google is gonna be the best
u voted for behemoth lol
frozen skiff
yeah
calm sequoia
I meant best of the 2.5, which is likely PRO
frozen skiff
im sill gonna use behemoth
balmy mist
Behemoth prob gonna be booty
torn mantle
Behemoth prob gonna be booty
i think we already have behemoth in the arena
barren prairie
It feels like being trained on messenger chat data 😄
On fb comment session
frozen skiff
Behemoth prob gonna be booty
nah bro
keen beacon
Rip millions and millions of dollars
balmy mist
Meta lost credibility after this nonsense they pulled
frozen skiff
i think we already have behemoth in the arena
proof
frozen skiff
hes right u guys dont know how to use it
calm sequoia
or divorce
frozen skiff
its the best ai
keen beacon
they dropped it because qwen 3 was about to release
balmy mist
They should have just waited
keen beacon
it would be too embarrassing at that point
keen beacon
They should have just waited
why it makes them look even worse if they wait
they moved the release to the weekend because they thought qwen 3 was dropping the same day or something
balmy mist
why it makes them look even worse if they wait
Bc nobody was thinking about meta
We were occupied by other models
And they could have made the models better
eager mica
Meta could have released a small preview "chat" model intended for local users, similar to the one people have been playing with on LMArena. It might have not necessarily been the best all-around performing model, but with some tweaks users would have probably liked it. There's a general lack of professionally-made conversational finetunes dedicated to creative/recreational uses and that model could have filled that gap, while the "serious" ones could have been trained longer before their actual release.
keen beacon
And they could have made the models better
they already delayed the models several times
but i do think it wouldve made sense to release it at llamacon
balmy mist
Releasing early during 2.5 and other models coming out is bad
Especially if you can’t deliver on what u promised
We can wait
But investors can’t
keen beacon
Releasing early during 2.5 and other models coming out is bad
they chose to do that because their model would look even worse in comparison with newer models
balmy mist
lol
I don’t understand how their models can be so bad
Like aren’t they not training on outputs of competition?
It should be minimum at 3.5 level at this point tbh
barren prairie
I don’t understand how their models can be so bad
But better than the previous one ,.
balmy mist
Is it better than 405?
I heard ppl say they got mixed results
Someone must have hacked their training data lol
Like I feel bad for them
keen beacon
nah theyre just massively incompetent compared to other frontier labs it seems
barren prairie
balmy mist
They need to hire us it seems
They still might be able to save themselves with the scout 10 mill context
keen beacon
nah it sucks lol
balmy mist
If they can make scout better
Then maybe one of their models might be usable, I don’t think behemoth coming out till summer lol
keen beacon
Then maybe one of their models might be usable, I don’t think behemoth coming ou...
they might abandon it tbh
2trillion param model competing with frontier models that are a quarter of the size
calm sequoia
How's that nobody believes in o3?
barren prairie
How's that nobody believes in o3?
I don t trust Sam 🥸
balmy mist
How's that nobody believes in o3?
I don’t think you should vote on belief
I put idk
Cause we don’t know lol
keen beacon
its fun to speculate tho
calm sequoia
Of course we dont know, but we have opinions 😄
calm sequoia
I don t trust Sam 🥸
Suck or musk is not trustworthy even more 😄
calm sequoia
Ha ha I remember playing with GPT 2 when it was out. It felt soo powerful 😄 Obviously the LLama4 is worse
keen fulcrum
When will grok 3 api release?
balmy mist
i dont find that fun lol, i just to test the new models, i can care less which model does best, like from which company, i have no dog in the race, but i think google will win in the end
like long term
keen beacon
fr tho benchmark wise gpt 4o mini might be worse (i dont remember it on hand though to compare) but i voted llama 4 anyway
balmy mist
near term who knows, but it seems google might still take it near as well lol
they got a lot of competition from 2.5 pro and NW
barren prairie
fr tho benchmark wise gpt 4o mini might be worse (i dont remember it on hand tho...
It is hard to decide 😂😂 all of them are Stupid
balmy mist
did anthropic abandon opus?
calm sequoia
i dont find that fun lol, i just to test the new models, i can care less which m...
you must be fun at the parties
balmy mist
i dont go to parties lol
so i wouldnt know lmaoo
im am a present kinda person
worrying about the past and future is things out of my control
and lets me live stress free
keen fulcrum
Soonish Deepseek will release R2
balmy mist
Soonish Deepseek will release R2
this week it seems
eager mica
By the way, it appears there are even newer Llama models with different codenames on Chatbot Arena right now.
They're still testing them, it seems.
keen beacon
They're still testing them, it seems.
are they any better or are they smaller versions
keen beacon
4o mini just about
eager mica
are they any better or are they smaller versions
I'm not sure; I just randomly submitted a joke question and a new model that I didn't recall seeing came out. It didn't stand out for anything in particular. Anyway, it's harley.
balmy mist
they need to pput behemoth in webdev lol
man i really want mav to be good lol
why would they even put it in webdev lmao
how often do these leaderboards get updated? like weekly?
calm sequoia
I hope meta pays LMARENA well. Garbage recycling is a hard work.
eager mica
I don't entirely agree that they gamed the benchmarks, but I do feel betrayed that in the end they released exceedingly large, boring models that do not perform as well as their size suggests.
Why would I (or anybody else) pick the response on the right here, for example?
[example redacted]
Many models on the arena are just like that: boring, stiff, stilted, no fun allowed.
Or if they do respond, they'll fill their text with GPTisms, so-called "slop" and excessive disclaimers.
"It's important to note that...", "As an AI...", "It's crucial to remember..." and so on. 🥱
balmy mist
Or if they do respond, they'll fill their text with GPTisms, so-called "slop" an...
Do you be jailbreaking models?
eager mica
Do you be jailbreaking models?
I'm not jailbreaking models on the Arena.
balmy mist
But in general you do?
Like in other platforms or locally?
Damn 6 months for Claude 4
X
eager mica
Like in other platforms or locally?
I use the most appropriate prompting for my local models (e.g. Gemma 3) to do what I want, but it's not really "jailbreaking"—no magical incantations or special sequences involved, only complete instructions. If the models do not behave despite putting reasonable prompting effort, then I'll use something else.
golden ocean
Damn 6 months for Claude 4
https://x.com/vitrupo/status/1908763535351669017?s=4...
bruh we gonna have gemini 69 and gpt 50 already by then
barren prairie
Damn 6 months for Claude 4
https://x.com/vitrupo/status/1908763535351669017?s=4...
Why so slow?
ocean vortex
Nvm i hate llama 4
it's the most boring release ever lol
the only thing they have going for it is the context size as far as I'm concerned
fleet lintel
man i really want mav to be good lol
why this model is at #2 😦
eager mica
why this model is at #2 😦
The model(s) on the Arena are not the same ones they released.
keen fulcrum
Many models on the arena are just like that: boring, stiff, stilted, no fun allo...
I think you underestimate the effort it took training those models
eager mica
I think you underestimate the effort it took training those models
I don't see a correlation. Those models' writing style is boring and stilted by design.
keen beacon
the only thing they have going for it is the context size as far as I'm concerne...
Did you see the ctx benchmarks lol. It can do haystack but it falls apart in actual usage even in super low context
frozen skiff
it's the most boring release ever lol
Dude not just that
Its shet
Its not just "boring" or "normal"
Its actualy horrible
keen fulcrum
I don't see a correlation. Those models' writing style is boring and stilted by ...
For formal language and business use case this is appreciated. They make money with businesses
keen fulcrum
Is there any ai that is fun to talk to like llama 4
Grok 3
keen beacon
Is there any ai that is fun to talk to like llama 4
Use the llama chaos engine system prompt on 2.5 pro lmao
frozen skiff
Use the llama chaos engine system prompt on 2.5 pro lmao
Gimme it
ocean vortex
Its actualy horrible
I wouldn't go that far. It is not horrible, it's comparable to deepseek V3.1 and gpt4o (the boring part). It's super consistent and reliable on easier prompts but the problem is it is not special in any way, and much more less likely to surprise you than those models as it struggles with harder or novel tasks
keen beacon
This kind of context performance is atrocious
ocean vortex
so while other models are more likely to fail the "easy" tasks, this is just boring and it's gonna fail if you suspect it will 
frozen skiff
https://gist.github.com/riidefi/3340cc2b33b9edf5f03dc4429ba635d0
LMAOO
I like this but its not really what im looking for tho
Idk how to explain it
Gemini 2.5 pro no matter how much you tell it to be informal, act human or whatever
It doesnt do it properly
Same for sonnet 3.7 and openAI models
ocean vortex
a "safe" choice but we do not really need that in this market and with those model sizes imo
keen beacon
I think the maverick model was fine tuned on a models traces that had this system prompt
frozen skiff
Same for sonnet 3.7 and openAI models
I liked 24 karat gold and llama 4 at the beginning because of those reasons
Then after I used them more
It just sucked
Its too inconsistent hallucinates to much etc
keen beacon
Ya llama 4 sux
ocean vortex
This kind of context performance is atrocious
lmao. They even did NiH and it was supposedly good
keen beacon
lmao. They even did NiH and it was supposedly good
They didn't train past 256k (iirc) so besides needle in a haystack using that kind of context is even more of a disaster than it is
frozen skiff
I wouldn't go that far. It is not horrible, it's comparable to deepseek V3.1 and...
nahh
its not reliable at all bro i swear
it gives false info on even the most basic tasks
even when it knows the right answer
when i correct it and say its wrong it gives the right answer
ocean vortex
it gives false info on even the most basic tasks
it's not impressive for sure, but we can't say it's sht either let's not get ahead of ourselves.
None of those evals would have been possible otherwise. Yeah it's a disappointment given the other models that we have, but it still performs somewhat decently when all said and done. I do get why this is the impression though, that's why I said it's boring - it is unlikely to surprise you.
frozen skiff
it's not impressive for sure, but we can't say it's sht either let's not get ahe...
yeah
its alr
keen fulcrum
Llama Behemoth will be great
frozen skiff
how do yk
keen fulcrum
Its their gem
keen beacon
It will be slow and suck
In comparison to other frontier models that are probably less than a quarter of the size
frozen skiff
Its their gem
that doesnt eman anything
it might be good it might be shet
wwe'll see
ima use gemini 2.5 pro for now and experiment some prompts to make it less boring
keen fulcrum
I do hope it gets coding capabilities soon
frozen skiff
cuz i cant deby its really good buts its just that AI writing part i dont lik
frozen skiff
I do hope it gets coding capabilities soon
gemini?
keen fulcrum
Yes
frozen skiff
it already has it
keen fulcrum
it already has it
Its bad
keen fulcrum
Please see benchmarks
keen fulcrum
Claude
frozen skiff
i thou ght the benchmarks said coding was gemini
ocean vortex
It will be slow and suck
how many tokens are they training it for?
also the fact that scout performs still considerably worse than maverick with double the data for pretraining kinda suggests to me 17B active is their limiting factor. Maverick performance on smaller model should be absolutely possible
frozen skiff
why did they even release mvaerick / scout
raven void
why not
keen beacon
also the fact that scout performs still considerably worse than maverick with do...
Tbh I think something is wrong with their specific Moe architecture or something. I dunno, not interested lol. Considering qwen 3 will be dropping a 15b total param Moe (and usually they only drop one, so I think it'll be good) and be very competitive for it's size/maybe even better than the current llama models
keen beacon
how many tokens are they training it for?
Not sure am on my phone
ocean vortex
Tbh I think something is wrong with their specific Moe architecture or something...
I was somehow thinking that it was similar to deepseek, but actually... Deepseek is 37B active parameters so considerably more. What the f meta was thinking LOL
keen beacon
I was somehow thinking that it was similar to deepseek, but actually... Deepseek...
tbh I don't think it's the active params
Something is just really messed up
ocean vortex
tbh I don't think it's the active params
I think it is. 17b active parameters is not enough for truly difficult tasks
so the end result is what we have now. They train it for so long it perfects what it can do
but still not enough capacity for very challenging prompts
keen beacon
I think it is. 17b active parameters is not enough for truly difficult tasks
Qwen 3 Moe will probably be better or maybe highly competitive and it's 15b and 2b active
They only release one I think so they're confident
ocean vortex
Qwen 3 Moe will probably be better or maybe highly competitive and it's 15b and ...
lmao no chance if it's not reasoning
keen beacon
lmao no chance if it's not reasoning
Why not gpt 4o mini is probably comparable
Releasing a gpt 4o mini level 15b Moe is more than good enough
ocean vortex
Why not gpt 4o mini is probably comparable
mini is much bigger pretty sure
2b active is basically as small as it gets
way too little if you ask me
keen beacon
2b active is basically as small as it gets
Ya qwen made bigger Moe's but they shrank it this time (assuming they only do one) so maybe something interesting is happening there
ocean vortex
I would expect this model to like compete with gemma, not maverick
ocean vortex
or qwen-7b... honestly they should update that one instead
keen beacon
or qwen-7b... honestly they should update that one instead
You think a dense 7b/8b could beat maverick?
Tbh it is possible
ocean vortex
You think a dense 7b/8b could beat maverick?
not maverick lmao, it can beat MoE with 2b active
keen beacon
not maverick lmao, it can beat MoE with 2b active
Nah I doubt
15b Moe will likely be better than the 8b
ocean vortex
disagree
I think active param count is more important than total size
so if it's only 2b... it's just compromised
keen beacon
They are releasing both based on the transformers pr that should be interesting
I guess the qwen 3 32b equivalent will be most interesting
It will beat maverick I think
Given how good 2.5 32b is
keen beacon
I think active param count is more important than total size
Weird how they decided to shrink it though. But curious what their large Moe model (qwen max) is
ocean vortex
you can't have free launch like that, given that training MoE takes roughly the same amount of compute and time as dense model equal to active parameters. If it was the case all models would have been 2T+ with less than 10b active lol
keen beacon
you can't have free launch like that, given that training MoE takes roughly the ...
Well clearly the qwen team found it appropriate if they're continuing the trend of releasing just 1 moe
What's the absolute best model rn? Is it nigjtwjisper?
keen beacon
What's the absolute best model rn? Is it nigjtwjisper?
It's gone
Someone let me know how does nightwhisper fare against o1
ocean vortex
Well clearly the qwen team found it appropriate if they're continuing the trend ...
they are not focusing on just MoE though. I think it's extremely unlikely this tiny MoE will come even close to QWQ 32b dense
upper wolf
What's the absolute best model rn? Is it nigjtwjisper?
Yes which is gemini
keen beacon
they are not focusing on just MoE though. I think it's extremely unlikely this t...
Yeah but they won't waste resources on an Moe that will just suck. Their previous Moe releases were much larger in active params count
Clearly they think it's something worth giving a shot/considerable chance of working out
frozen skiff
Someone let me know how does nightwhisper fare against o1
Too good its uncomparable
Nightwhisper is godly
They removed it lol
Its gemini toh
Tho
ocean vortex
Yeah but they won't waste resources on an Moe that will just suck. Their previou...
they are probably just focusing on mobile/laptop applications with it. Memory is not a huge issue at these sizes but speed can be. So MoE may make more sense
keen beacon
they are probably just focusing on mobile/laptop applications with it. Memory is...
Maybe but I think the 15b Moe will be the most interesting thing to come out of all of this
Is there a nughtwhusoer like model
upper wolf
No
torn mantle
X
😭
i think it may possible saying that maverick and scope are early checkpoint distilled versions of behemoth
novel flame
It would seem Llama 4 was running in "Volkswagen emission testing mode" in the Arena to get the #2 spot
keen beacon
i think it may possible saying that maverick and scope are early checkpoint dist...
they're distillations of behemoth
lmao
behemoth is the teacher model
eager mica
We codistilled the Llama 4 Maverick model from Llama 4 Behemoth as a teacher model, resulting in substantial quality improvements across end task evaluation metrics. We developed a novel distillation loss function that dynamically weights the soft and hard targets through training. Codistillation from Llama 4 Behemoth during pre-training amortizes the computational cost of resource-intensive forward passes needed to compute the targets for distillation for the majority of the training data used in student training. For additional new data incorporated in student training, we ran forward passes on the Behemoth model to create distillation targets.
ocean vortex
they're distillations of behemoth
not true. Scout was pretrained on 40T and Maverick on 22T. Therefore different pretrained models and not distills
leaden palm
200k
this almost feels like you're asking for someone to say claude
keen beacon
not true. Scout was pretrained on 40T and Maverick on 22T. Therefore different p...
thanks for letting me know i was wrong on something that's already been established 👍 simple error
leaden palm
im saying that because afaik claude is the only model with a 200k context window
but yes i think that claude is a pretty alright model, even without thinking
ok but oddly specific
your own benchmarks >> anything else
ocean vortex
> We codistilled the Llama 4 Maverick model from Llama 4 Behemoth as a teacher m...
ccodistillation is quite a bit different from normal distillation
instead of unconditional teacher-student you have both models updating gradients and learning from one another 👀
leaden palm
other humans and benchmarks are unlikely to have tested text adventures
ocean vortex
thanks for letting me know i was wrong on something that's already been establis...
you are welcome 🫃
it seems they essentially optimized the training of several models at the same time to make sure the same work is not getting repeated as much, is the extent of that "distillation". But it's not in the context what most people think when they see the word "distillation". Those 3 models are still different with varying amounts of total data used for pretraining
leaden palm
series of finetunes most well known for mistral/mixtral tunes
i don't dabble much in temperature and personally think that people care too much about it
temperature 0 = follows most likely path, temperature 1 = chooses paths with proportional probability, and you can imagine how to interpolate
when i run llms i typically use a temperature of 0.7
then
fun fact: temperature can actually go past 2, the limit of 2 is an artificial limitation
back when i was trying to get llm text to sound human i found that 1.05 worked when doing free generation but you needed like 1.3 when rewriting ai text
novel flame
Sigh.... Finally figured I should test Llama 4 Maverick too, and I too can confirm that it's somewhere between utterly disappointing and 'meh fine'. It scored a 3 out of 5 on my mini-benchmark, failing on both associative logic and on grounded memory recall (it had more hallucinated 'facts' than any model I've ever tested, it's bananas). It was less bad at coding than I expected, maybe a decent 'second tier' coding model, but nowhere near Claude 3.5/3.7 Sonnet or Gemini Pro 2.5.
frozen skiff
i think they were asking "aside from 2.5 pro"
And im saying theres only 2.5 pro
But he didnt say aside from 2.5 pro he asked if it was only 2.5 pro
Does anyone have a prompt to make 2.5 pro sound human
Its too robotic
And i cant prompt it to sound human
What system prompt should I use
Yeah
The temperature doesnt do anything
Unless I changed the top p to 1
Then it becomes so random
And nonsnesical
hmm
lemme try it after i try this one i found on reddit
im ean
i tried sending it to it ddirectly
but
it talks too informally
like a 12 year old
not that but
i told it like
act human and stuff
leaden palm
Does anyone have a prompt to make 2.5 pro sound human
are you trying to get it to talk like this or do you want to get past ai detectors
two very different goals
frozen skiff
are you trying to get it to talk like this or do you want to get past ai detecto...
Not the ai detecotrs
💀 wtf
frozen skiff
Not the ai detecotrs
i just want it ta talk
to talk
ok
imat ry
so it works less better if i tell it to it directly
novel flame
What are your experiences / reviews of riveroaks?
frozen skiff
What are your experiences / reviews of riveroaks?
its mid
frozen skiff
BEST MODEL IS crazy
torn mantle
its not good at coding
ive tried it
whereas its good at general text
i believe its llama 4 behemoth
reminds me a lot of themis/cybele
its kinda like gemini 2.0 pro
could be
frozen skiff
whats a sand boc adventure
wtf
like do wat u want in what
i dont get it
what u gonna do with the AI
OHHH
nice
like a roleplay
yeah
what ai do u do it with
thats the best for it
thanks ima try it
did u invent this
sandbox adventures idea
i never heard of it before
: O
that's creative
ima try it today it seems really fun
golden ocean
give prompt
frozen skiff
give prompt
ye @hollow ivy
torn mantle
i need them to bring back NW to webdev
frozen skiff
its genius
thanks its really cool
ur a very creative person
iim using it
eexactly
i was about to say that
u never thought u would have stuff like this for example 10 years ago
and at the rate these stuff is progresing u dont know whether to be scared or excited
yeah its very crazy to think about
LOL ye
im mainly excited ngl
cuz when they get really good
like in a year or two from now
we can do insane things with them
if they become sentient and try to take our jobs and overtake humanity or something 🤣
hmm
yeah definitely
like at the rate of how fast they're growing rn they might become strong enough to replace human workers
primal orbit
I see now both maverick and crystal have token limit in chat. So 24 karat was definitely Meta model.
frozen skiff
I see now both maverick and crystal have token limit in chat. So 24 karat was de...
24 karat is meta but we dont know which meta model it is
definitely not scout or maverick tho
primal orbit
the responses are quite similiar
frozen skiff
its prob some kind of testing experiment by lamma
frozen skiff
the responses are quite similiar
nah
they are kinda similar but
24 karat was WAYY more creative
and hallucinated way more
😮
primal orbit
may be the same model but different settings like temperature or diff system prompts?
frozen skiff
that might happen soon
when do u think it AI might become cappable enough to do such a thing
frozen skiff
may be the same model but different settings like temperature or diff system pro...
probably yeah
maverick or scout with a different system prompt and higher temperature or something
i mean if u think about it
it hasnt even been 5 years since the first chatgpt model released
so it might even become way stronger than that in 5 years
primal orbit
I red in the news microsoft released today quake 2 web version, generated by a model in real time.
you can play 1 level
frozen skiff
yeah even more than that
frozen skiff
I red in the news microsoft released today quake 2 web version, generated by a m...
whats that
hmnmm
hardy pecan
I honestly suspect 24_Karat_Gold was Llama-4-Maverick-03-26-Experimental, but now they defanged the model big time
2400 votes I assume were from us, the public right
Maybe they didn't change it on the fly though... idk
frozen skiff
I honestly suspect 24_Karat_Gold was Llama-4-Maverick-03-26-Experimental, but no...
if its a differnt system prompt
it might be maverick
hardy pecan
Okk
frozen skiff
exactly
hardy pecan
makes sense
frozen skiff
it might be just maverick with like higher temperature or a different system prompt
hardy pecan
Yeah
frozen skiff
it also constnatly said it was from llama
hardy pecan
yup, which i believe
torn mantle
you guys have no idea how powerful that nightwhisper model is
i mean combined with google perfect vision
you can clone websites in seconds accurately
lime coral
I know and i still mean what i said
torn mantle
I know and i still mean what i said
why would i care about 9.11 > 9.81
or how much r's are in strawberry
keen beacon
Shows model reasoning
lime coral
why would i care about 9.11 > 9.81
You still don’t understand I was ironic
kind cloud
