ai-news

[rigid oriole]

fake news?

[rigid oriole]

it won't happen tomorrow, i bet

[unique cargo]

where :((

[vocal lodge]

https://youtu.be/KX0QN1FmK_g
Paper: https://arxiv.org/abs/2505.03509v1

[vocal lodge]

They trained EfficientNet (CNN by Google in 2019) using a semi-supervised active learning. It's interesting since it's mostly just standard ML techniques.

[fresh basin]

I think many problems are attention bottlenecked, hence a lot of (comparatively) low hanging fruit are there for people to pick if people push through (not me for example)

[amber rune]

https://www.anthropic.com/news/claude-opus-4-6

[kindred hazel]

woahhh

[vocal lodge]

GPT-5.3-Codex just got released as well.

[opaque raptor]

0% music
100% talking

[fresh basin]

https://newsletter.semianalysis.com/p/claude-code-is-the-inflection-point

[rigid oriole]

https://www.youtube.com/watch?v=fFL7la73RO4

[vocal lodge]

1M Context on the API

[vocal lodge]

This is an old paper, lol. There's been a lot of subsequent research on the findings from the R1 paper.

[rigid oriole]

Which downloadable model under 6GB size is the best for immersive roleplaying and sandbox-adventure-games, etc.?

(it doesn't need to be uncensored)

[vocal lodge]

I'm not sure, might want to ask on KoboldAI or SillyTavern servers

[hollow field]

Try Gemma 3

[rigid oriole]

has only 32k context :(
Which downloadable model has the largest context size?

[hollow field]

If you are going so low (6GB) because of low VRAM capacity, you probably won’t use much context anyway

But if it’s not the case, go on

[late sentinel]

@spice spire

e Le'minate him

[vocal lodge]

Paper of the less well-known #1 Chinese model on text arena (also #1 "Paper of the Day" on HuggingFace): https://arxiv.org/abs/2602.04705

[cold briar]

https://github.com/aerlabsAI/ai-inference-resources?tab=readme-ov-file

[random hedge]

When kling 3 is going to live on lmarena

[spice spire]

Isn't related to AI but is related to Discord - https://www.theverge.com/tech/875309/discord-age-verification-global-roll-out

[deft timber]

Ideally there will be some global workaround ;p

[lusty minnow]

I can try ?

[last ocean]

I am pretty sure this is only in the uk not coming into the us yet

[spice spire]

I don't think so. From their announcement:

Starting in early March, all new and existing Discord users globally will be assigned new default settings that support age-appropriate experiences while keeping privacy front and center:

[kindred hazel]

it’s doesn’t really change anything for us tho

like the only changes are for those who are in those 18+ servers

[glacial marsh]

yes most needed

[rigid oriole]

https://www.youtube.com/watch?v=P4pBrHx3zSA

[spice spire]

Yeah agreed. Before reading more into it was wondering if there were changes on the large community side, but it doesn't seem to be the case.

[ionic tide]

https://www.youtube.com/watch?v=1tM7Wd0lEeI

[vocal lodge]

https://youtu.be/httnhdpu_W4

[karmic pulsar]

guys glm 5 just released

check it at z.chat

[nimble cliff]

Yeah minimax 2.5 too

[rigid oriole]

how can i change the z.ai model in an existing thread?

[clear smelt]

Oh yeah since nobody mentioned it yet, gemini 3.1 Pro likely this month or if we can believe X tomorrow.

[topaz isle]

nice

[kindred hazel]

apparently it’s this week

[nimble cliff]

I thought this was about gemini 3.1 pro 😭

[real ledge]

Can we get this model on Arena?

[nimble cliff]

No cause this isn't a model I think

Not sure

And this is only available on their $200 plan so yeah it's darn expensive

[devout berry]

They will most likely update the Gemini 3 pro score

[nimble cliff]

No cause this is a separate thing

Model score is the same

[clear smelt]

Gemini 3 Deep Think Release

[random hedge]

What are the coding capabilities of Gemini 3 Deep Think and how does it excel?

[stray fog]

would be interesting to actually see which kind of tasks were given to the model, since it outperforms Opus 4.6 in a way that is incredible

[random hedge]

Can I know it on complex and agentic tasks building, full stack web app or android app

[clear smelt]

Well in Algorithms its #7 of all Programmers in the world so thats pretty impressive.

There are some cool examples like this one:

"Gemini 3 Deep Think generated a real-time 3D WiFi radar that maps every network around you as glowing nodes in a Matrix-style space — in one shot. It used Pearson correlation to infer which APs are physically close, since RSSI alone isn't enough."
https://x.com/i/status/2022143418592555074

Or even this one: https://x.com/i/status/2022226996152148052

BijanBowen (@Ominousind)

Gemini 3 Deep Think generated a real-time 3D WiFi radar that maps every network around you as glowing nodes in a Matrix-style space — in one shot. It used Pearson correlation to infer which APs are physically close, since RSSI alone isn't enough.

▶ Play video

X · Feb 13, 2026, 2:58 AM

can (@marmaduke091)

GUYS.

If you want to have an AGI moment, just watch this video once until the end.

Gemini 3 Deep Think ONE-SHOT this.

▶ Play video

X · Feb 13, 2026, 8:31 AM

I also liked this one: https://x.com/i/status/2022149177787265478

Jimmy Apples 🍎/acc (@apples_jimmy)

Had the new Deepthink make a quick Ascii fishing game.

Very nice so far.

▶ Play video

X · Feb 13, 2026, 3:21 AM

[dark hornet]

Where to use

[clear smelt]

Ultra users globally

[dark hornet]

When do you think it'll be available for free somehow

[clear smelt]

never probably

Dont think Gemini 2.5 Pro Deep Think ever was available for free or Pro users

Could be happening tho as they are investing in new datacenters

[dark hornet]

Is it on api or not

[clear smelt]

I actually dont know but I assume not

[twin mirage]

https://youtu.be/OPSGsHS5FPI?si=-3wY1Fv9Bh88v7XF

[broken junco]

the models arent the models you think

they give you the most outdated yesr

gemini 3 gives you info frôm 2024

so does max

[wanton ember]

Fun week in AI! Spotify says their top engineers haven’t written code since December, Google’s Gemini 3 hit 84.6% on ARC-AGI-2, and Anthropic just closed $30B at $380B. Full breakdown with benchmarks and funding tables here:

https://www.everydev.ai/p/news-ai-dev-news-digest-february-13th-2026

[near steppe]

spotify? more like slopify

[rigid oriole]

in less than 5 years, a flash model will be more capable than this current ultra model

and in less than 2 years, the pro model will be able to do what ultra can do now

[kindred hazel]

5 years?

I’d reckon we’d only need to wait a year or 2 max

[minor thunder]

wow that's very pessimistic, the model we have now are already better than the pro version we had 5 month ago

the whole capabilities of the ai go up when a new model get released

[minor thunder]

in 2 years do you realise the progress ai made ?

i let you check what ai were able to do in 2024 or at the start of 2025

and compare it to right now

in fact im sorry to say this, but the ultra version isn't even better than gemini 3.1 pro when it'll get release

and since its a race the progress really isn't stopping

the ultra version take a very long time to think, so it answer are better, i do believe there's more than that but it definitly won't be the best model for a year

And anyway if gemini 3.1 pro will be out then they will release a new deep think for it too

they can't like just stuck on the actual model

[dark hornet]

5 years is an extremely long shot

right now we have gemini 3 flash which is way better than o1 pro mode which was way more expensive

[rigid oriole]

My prediction still was correct, because i said in less than 2/5 years :)

(less than 5 years can mean: 4.99 years or 1 day or anything in between)

So i'm 100% sure that it won't take longer than 2 years (for pro) or 5 years (for flash) to become better than current ultra version.

But you might be correct, in 2 years, flash-lite might even be better than ultra is today. (although i deem that as improbable)

[minor thunder]

it doesn't really make sense anyway cause when a new gemini model came out they will for sure also update the pro version like when gemini 3.1 pro come out, then there will be gemini 3.1 deep think

at some point

but if we compare to the actual deep think model right now, this one will definitly be beaten soon by a non ultra tier subscription

cause this one will be old

[rigid oriole]

therefore my comparison of future (pro/flash) versions is with today's ultra version, not future ultra version

[minor thunder]

then yes the today ultra version will be beaten much sooner than you expected i can bet everything on it

[rigid oriole]

when do you think will flash beat current ultra?

and when will flash-lite beat it?

and when pro?

and when will flash beat opus-4.6-thinking in coding?

[minor thunder]

since its a flash model it might be in 2 or 3 new models from google, cause its a flash model its done to be faster but less capable and less costly

for the pro version, in my opinion its in one or two but no more than that

to beat the actual deep think

the reasining of the model each time a new one get out is like a whole different world

it doesn't think and act the same way

its always much better

[rigid oriole]

Do you think this trend will continue until we have a perfect (C++) coding AI?

[minor thunder]

i don't know what do you mean by perfect cause there's also no human who's perfect at coding

[rigid oriole]

perfect:
one-shotting everything which is within its context/output-token window

which can be described in c++ code

[minor thunder]

yes i think we will hit that point

maybe you won't even have to be good at prompting or to give it much details

it'll figure out

for now ai capabilities honestly are already really great

[rigid oriole]

So, AI could replace >99% of all developers?

[minor thunder]

yes they can't one shot everything, yet it can't be considered bad

for now

[minor thunder]

yes at some point

that's the purpose i guess of ai

[rigid oriole]

yeah

fantastic for some people, brutal for others (who have not adapted in time)

[minor thunder]

an ai capable of doing a better work than the best human team in coding will definitly exist

[rigid oriole]

next year?

[minor thunder]

an ai smarter than the smartest human overall too will exist

[rigid oriole]

what jobs will remain for us?

plumbing?

[minor thunder]

that's really hard to say but i can't say its not happening you know, and if its not next year then it will be in like 2 or 3

[minor thunder]

ai are coming with robotics too

so even physical work

is getting replaced

[rigid oriole]

what if robots/androids/gynoids can do all work in 5 years?

[minor thunder]

then we are free from work since we're not usefull and i don't know how gouvernment can handle that

this question as already been asked and studied

and its the question everyone is asking

cause yes it will probably happen

[rigid oriole]

global UBI provided by the World Republic?

[minor thunder]

I really can't tell its too hard to know what will happen since its the biggest change ever in everyone's life and how the whole world work

[rigid oriole]

hopefully we have enough energy by then, to power all these robots and datacenters..

thorium energy?

[minor thunder]

i mean ai are also accelerating discoveries

and since its getting more capable

its better and better, and will make even more discoveries

[rigid oriole]

so, next decade we should have working fusion reactors?

[minor thunder]

that's too specific, but if things go as we say, then we will be able to live much longer probably, to cure what's uncurable right now, to do things we can only imagine right now

[rigid oriole]

LEV

(longevity escape velocity)

[minor thunder]

its like skipping many years and being teleported 100 year in the future

and then since we hit that time ai would be even smarter

so i really don't know what will happen at this point

[rigid oriole]

hopefully, china won't develop AGI first :/

[minor thunder]

I hope too, anyway im happy to be here to live this

[rigid oriole]

i hope opensource will get to AGI first

[minor thunder]

right now its just the very start

of what might be the biggest change ever

[rigid oriole]

happy too, what a time to be alive

Do you think, we will have off-world colonies or space cities in second half of this century?

like, O'Neill cylinders, Stanford Torus, Mars colonies or Venus aerostats?

and maybe a robotic colony on Luna?

[amber rune]

@rigid oriole Earth is still the only place we know of that isn’t just awful; everywhere else wants to kill us. No breathable air, no food sources, temperatures too high or too low (or both), radiation, little to no protection from solar flares, too expensive to resupply, can’t get a plumber or a decent pizza delivered… and every one of those drug-fueled Elon Musk proposals (like dropping nukes on Mars to terraform it) is either fundamentally physically flawed or would take hundreds of years and hundreds of trillions of dollars to complete.

So second half of this century for offworld human colonization? Not a chance. Robots, sure, if there’s a point in doing it.

[amber rune]

Sadly if history is any indication, any time a technology has come around which allowed the few to rule more effectively over the many, what has followed has been tyranni. Only when the support of the people is a requirement for maintaining power (typically through a need for taxes and soldiers) do you get things like democracy. If one rich bastard with a secret police force and an unstoppable AI drone army can subjugate millions and keep them in poverty while taking billions for himself, chances are he will.

When you guys say “I hope China doesn’t get AGI first”…. Who are you hoping does?

[worthy inlet]

و

[amber rune]

Also I’m sorry for derailing. This is #ai-news

[rigid oriole]

which channel would you propose instead, to discuss this? or a dedicated thread, but in which channel?

[rigid oriole]

what i meant, is, i hope that no government gets it first, but that the opensource community gets AGI first

at GitHub or another trusted opensource platform

[amber rune]

#general or #general-champion ?

[rigid oriole]

maybe a thread in #ai-memes , as it would disturb the least, there?

general is too crowded and champion general is reserved for AI-focused talk, i guess

also the champion channels can not be accessed by non-champions

[rigid oriole]

-# (created a dedicated thread: https://discord.com/channels/1340554757349179412/1472698608980856842 )

[twin mirage]

https://youtu.be/17KvQYyrBEQ?si=HSaPZdcxb6Ny-vJx

[vocal lodge]

https://www.bloomberg.com/news/articles/2026-02-15/rampant-ai-demand-for-memory-is-fueling-a-growing-chip-crisis

[vocal lodge]

https://qwen.ai/blog?id=qwen3.5

Uses hybrid attention, which is fairly uncommon.

API prices: https://www.alibabacloud.com/help/en/model-studio/model-pricing

$0.4/$2.4. Slightly cheaper than Flash. Input prices triples above 256K, despite hybrid attention 🤔

[pliant birch]

some1 check this out, some new autonomous models soon to come https://x.com/OscerraHQ

Oscerra (@OscerraHQ) on X

Prepare to experience the unimaginable.

X

[amber rune]

I doubt it. Seems like hype bait - and this is the third channel where you have posted links to this. Your own side project?

[pliant birch]

I wasn't sure which channel was best so i tried a few

also google shows nothing prob cause acc is like 72h old

[wide sandal]

coool

[fresh basin]

https://www.youtube.com/watch?v=bZJ7A1QoUEI

[kindred hazel]

Today is Chinese lunar new year

so my bet is that deepseek will be realising something big too

[wide rampart]

[rigid oriole]

why only deepseek, and not also Ernie, Minimax, Mimo, etc?
(Kimi, Qwen and GLM updated recently, so these probably not)

[kindred hazel]

mainly because the other Chinese companies aren’t that relevant

as in they probably won’t do anything special

[rigid oriole]

ok, that's true
i wonder what happened to Yi[-large] and some impersonating models

[twin mirage]

https://youtu.be/QdjGpLbePPU?si=FABecN-VhHyPSSEh

[dawn basin]

@junior dome Note that Video Arena has been removed from the server. More information can be found in this [announcement ](#announcements message)⁠

@uneven fulcrum Note that Video Arena has been removed from the server. More information can be found in this [announcement ](#announcements message)⁠

[daring sable]

will have to see if code arena has the same outcome

[twin mirage]

https://youtu.be/wDBy2bUICQY?si=dK3t_laI2CVvR_W7

[slim granite]

tf

[twin mirage]

https://youtu.be/DC1uWRqF2Yc?si=8ZpclFwDWEOV8uHK

[fresh basin]

even this channel got polluted. Incredible.

anyway this video is a bit memey but it is actually informative: https://www.youtube.com/watch?v=httnhdpu_W4

[hushed birch]

https://x.com/OfficialLoganK/status/2024516665740833156

Logan Kilpatrick (@OfficialLoganK)

Introducing Gemini 3.1 Pro, our new SOTA model across most reasoning, coding, and stem use cases!

X · Feb 19, 2026, 4:09 PM

[fresh basin]

https://x.com/DimitrisPapail/status/2024555561199480918

Dimitris Papailiopoulos (@DimitrisPapail)

Addition Under Pressure

X · Feb 19, 2026, 6:43 PM

[late rain]

I've been wondering about whether or not any open source software project maintainers have mentioned groundbreaking advancements, or speed of advancement, attributable to incorporating LLM vibe coding into their workflow.

[fresh basin]

yes I search about this every now and then, because it is a good threshold. The result so far is that (a) few OSS notable libraries "admit" to be mostly Agent driven and/or (b) there is no big contribution so far.

Consider also that Agent driven PR are a nightmare for OSS, because then you oper the floodgates. One could end up reviewing a lot of slop for few gems, and that is exhausting.

[pliant birch]

upcoming agent - https://x.com/OscerraHQ/status/2025224583645987172

Oscerra (@OscerraHQ)

𝗟𝗢𝗚: 𝟬𝟬𝟯 -
ORA just 𝗿𝗲𝗽𝗹𝗮𝗰𝗲𝗱 𝗮𝗻 𝗲𝗻𝘁𝗶𝗿𝗲 𝗳𝗼𝘂𝗻𝗱𝗶𝗻𝗴 𝘁𝗲𝗮𝗺 from a 𝘀𝗶𝗻𝗴𝗹𝗲 𝗽𝗿𝗼𝗺𝗽𝘁.

▶ Play video

X · Feb 21, 2026, 3:02 PM

[safe sleet]

hello

[vast frigate]

Hello Pritam

[twin mirage]

https://youtu.be/Y7W9Sbu1lYQ?si=jq1h-V-9k8sKMs0Z

[fresh basin]

This is not a bad benchmark. https://algotune.io/ It cannot be contaminated because it is like "open" (sure, models could pick up improvements wrote online that ended in the training data/post training).

Anyway if you see the discussion traces the models are given also a certain amount of costs to find the solutions. Example: https://algotune.io/matrix_sqrt_claude-sonnet-4-5-20250929.html

and this is also interesting:

https://halluhard.com/index.html

[tall linden]

@ivory tree @faint pasture Note that Video Arena has been removed from the server. More information can be found in this announcement #announcements message

[spice spire]

This is interested, I've always wondered if/when Arena would make a hallucinations leaderboard

[fresh basin]

https://aleximas.substack.com/p/will-advanced-ai-lead-to-negative

https://tecunningham.github.io/posts/2025-10-19-forecasts-of-AI-growth.html

[wide rampart]

https://www.anthropic.com/news/detecting-and-preventing-distillation-attacks

LMAOOOOOOOOOOOO

[arctic aspen]

This is absurd lol

It is a bit disheartening at the same time too though, since this essentially means these AI labs weren't actually making huge leaps, and were just distilling off of already-developed models

Looks like closed-source will still be #1 for a while

[kindred hazel]

yeah it is pretty disappointing actually

like it’s kinda funny that deepseek needs to do that

but you would think that by how well known and well funded they are, that they wouldn’t need to do this

[eternal seal]

ah yeah, I hoped they were catching up, turns out not 😭

[vast frigate]

They are catching up and making huge leaps, they are just being depective and playing dirty to get there

[deft timber]

they don't exactly stand on a strong moral highground
https://www.bbc.com/news/articles/c5y4jpg922qo
https://futurism.com/future-society/anthropic-destroying-books

[twin mirage]

https://youtu.be/Epp-Vz0FrPg?si=2APYokfblKafno6F

[noble blade]

The samples they reported for deepseek are actually negligible + mainly focused on things like RLHF distillation, which was pretty obvious for them to do (as they don’t have a massive team for building reward model data)

They kind of just included them in here to reduce the deepseek hype.

I think for the others it should have been pretty obvious..
Surprised to not see z.ai there

Though maybe they only distilled Gemini

[kindred hazel]

https://openai.com/index/why-we-no-longer-evaluate-swe-bench-verified/

[fresh basin]

https://www.citriniresearch.com/p/2028gic

[fresh basin]

16 million exchanges are nothing. One needs 30T tokens to train models, 16 million exchanges, at 20k token each (not easy actually to get that many outputs token with different questions that are normal, and ad hoc questions aren't that helpful for training anyway), are 320B tokens. 320B tokens vs 30T is a rounding error.

Add to this that is multiple AI labs, so those 320B tokens aren't even centralized.

Further Anthropic and others trained on human works practially for free, and now they cry that others train on their outputs (but they pay for it!)

[noble blade]

The 30T number is for pre training.
For mid training and especially RL coldstart data 320B (+ their actual rl environment) is actually already close to enough.

Basically it is not typical full model distillation, just some parts of post training (where you need a lot less data)

Especially for the human preference phase, they basically only need a little bit of data.

[fresh basin]

ok if they use it for RL it would make more sense (again it is 320B divided by 3 or more). But still for such "attacks" (as Anthropic says) I would expect much more, not mere 320B tokens.

[twin mirage]

https://youtu.be/8dkhBKejWkk?si=jyHUGMgC1ujMcOFM

[twin mirage]

https://youtu.be/Bqdf6Um_8OE?si=1ycCFg_eZf-zFO7t

[twin mirage]

https://youtu.be/j0da1ZftUIo?si=bnqytwPTsJv5Ev-6

[fresh basin]

https://x.com/DimitrisPapail/status/2026531440414925307

Dimitris Papailiopoulos (@DimitrisPapail)

You Don't Need to Run Every Eval

X · Feb 25, 2026, 5:35 AM

[narrow siren]

Who will join the call

[arctic aspen]

Looks like V4 lite is coming soon instead

[near steppe]

v4 lite is probably their answer to models like step 3.5 flash or minimax m2.5 and I expect it to be around that size (~180-230b), could be a bit smaller (~100b, closer to glm-4.5 air) or larger (~350b, closer to glm-4.5/4.6/4.7)

[rustic plover]

it's interesting that Qwen, GLM are not mentioned there, Mistral could be a candidate too i feel

[rustic plover]

they trained their model on stolen books is still quite different compared to their models being distilled, one is copyright infringement, the other business theft/competition i'd assume, it's like Louis Vuitton stole a design without citing the designer's name to build their bags and people copying their bags literally

[gentle nest]

deepseek is puny lol.

[timid helm]

Can deepsek create images now?

[arctic aspen]

https://www.anthropic.com/news/statement-department-of-war

"The Department of War has stated they will only contract with AI companies who accede to 'any lawful use' and remove safeguards... They have threatened to remove us from their systems... to designate us a 'supply chain risk'... and to invoke the Defense Production Act to force the safeguards’ removal."

"...we cannot in good conscience accede to their request... Should the Department choose to offboard Anthropic, we will work to enable a smooth transition to another provider"

Absolutely crazy, but W anthropic

[vast frigate]

DoW

[atomic python]

Am not that good with covering news but, I heard some chinese companies are stealing training data from anthropic.

[minor thunder]

that's called distillation, its using an existing ai to create data for training your own on it

it'll be very good data and speed up the progress of the other model by alot

that's not illegal, but it's not fair

can't compare this to training a model from zero

[copper bolt]

What do you think is better? Nano Bana Pro or Nano Banana 2?

[broken junco]

nano banana 2 is legitimentally nano banana pro but better

https://tenor.com/view/meme-milk-gif-6919750708092698197

[vocal lodge]

https://www.afr.com/world/north-america/pentagon-threatens-to-take-anthropic-s-ai-tech-in-defence-standoff-20260225-p5o59t

[stone locust]

Nano Banana Pro is based on Gemini 2.5 pro, Nano Banana 2 is based on Gemini 3.1 flash
so obviously the correct answer will be Nano Banana Pro 2

but nah I think NB 2 is better than NB Pro right now. Most of the time.

[opal robin]

Has anyone tested trizao.ai?

[arctic aspen]

https://x.com/sama/status/2027578652477821175

Sam Altman (@sama)

Tonight, we reached an agreement with the Department of War to deploy our models in their classified network.

In all of our interactions, the DoW displayed a deep respect for safety and a desire to partner to achieve the best possible outcome.

AI safety and wide distribution of

X · Feb 28, 2026, 2:56 AM

OpenAI seems to be getting along well with the DoW unlike Anthropic

not sure who's right anymore lol

[deft timber]

anthropic have always been the jerks just ask theo.. but also theres so much we dont know about the backdoor interaction so gov seems to be taking an overly heavy hand with it, i dunno

[twin mirage]

https://youtu.be/qnOmUWd-OII?si=gSPzW7cv9D9wWzcT

[vast frigate]

What do you mean? OpenAI is the one led by Sam Altman, so that should tell you everything you need to know.

[rustic plover]

they have to if you read their financial statement and balance sheet i think

[hushed birch]

You are joking right? When has trusting the gov ever been the right move Lmaoo

[arctic aspen]

I mostly mean in the sense of who's telling the truth, hah

I definitely don't think OpenAI is making the good move here

[spice spire]

@neon abyss Note that Video Arena has been removed from the server. More information can be found in this announcement.

[vast frigate]

Again. Look at who leads OpenAI. Did they get fired for not telling the truth before?

[arctic aspen]

fair enough lol

[vast frigate]

The big question is whether that Anthropic gave DoW a different model to the ones publicly available. Obviously Claude is infamous for being so censored. Do we really think that a model like that was used for the Venezula operation?

[deft timber]

People seriously aren't aware of the many numerous extremely shady business practices of anthropic? Or are they so blind with hatred for gov that the shady Corp can do no wrong?

[vocal lodge]

Another similar example where researchers may have found the first lunar lander (Luna 9) by using a object detection CNN (TinyYOLOv2) to look through high resolution LROC imagery: https://www.nature.com/articles/s44453-025-00020-x

[sturdy rampart]

https://sora.chatgpt.com/invite?code=RFJEE6

[opal robin]

https://trizao.ai

[vast frigate]

I think it is more about how much better they are relative to OpenAI.

[deft timber]

in terms of the models or company?

[vast frigate]

The company. The models obviously have catered to different focuses.

[twin mirage]

https://youtu.be/Ecmlh607KeA?si=udv1QF0T8dDKyzAC

[deft timber]

ye not even close according to those that interact directly with both.
anthropic is super scummy in many ways people seem to not be familiar with for some reason.

[vast frigate]

https://www.youtube.com/watch?v=MPTNHrq_4LU

[obtuse maple]

https://youtube.com/shorts/HeTK3xFebwo?si=0Fnbzu7SM6pRw6-C

[rigid oriole]

https://www.youtube.com/watch?v=N_piE0I34gc

click-bait as always?

https://eu.36kr.com/en/p/3705022520864896

[verbal crown]

If something says "AGI" in the context of "It's been achieved", you can probably guess that it is.

[rigid oriole]

you meant, "… that it is not." ?

[verbal crown]

No, I mean if a title says that it AGI has been achieved, you already know it's be click-bait.

[rigid oriole]

ah, for a moment i thought, you meant that it is AGI

[verbal crown]

Ah no no haha my bad

[rigid oriole]

lol

so the whole ordeal is another fake-news ?

fabricated by AI, i guess

[verbal crown]

Well, it could be real news, just written in a way to be clickbait to serve whoever made that video

[rigid oriole]

by the way, do you think, AI will ever "crack" math?

i mean, becoming as good in math, as Opus-4.6 is in Python-coding?

[deft timber]

@echo halo @spice spire

[gilded portal]

https://youtube.com/shorts/ZkHl1dogWjU?si=NsgsZnFF715EzlL1

[spice spire]

What’s up?

[deft timber]

someone already deleted it. just the usual

[fringe fulcrum]

I think reasoning is where they made a major step. That was eye opening watching models like o3-high crunch enormous numbers with very high accuracy

You can paste any calculation problem with huge numbers and chances are it will get it correct no tools

problem before reasoning was that it couldn't arrive at the answer with high enough accuracy missing context. But now it can take as much time as it needs generating relevant context itself and limiting potential outcomes. Until it can arrive at a specific number with high enough confidence

[rigid oriole]

At which point does it become reliable?

99%?
99.9%?
or do we need almost 100% accuracy (no hallucinations) ?

6 sigma?

Do we need an error-rate of less than 1 PPQt?

(<1 part per quintillion = < 10^(-18))

[fringe fulcrum]

I generally consider the most frequent response (7/10+) to be what can be counted as consistently correct. 99% is not very reasonable considering that you would have to run the same prompt around 100 times to even notice that it can be incorrect

[rigid oriole]

Would you accept a planetary government, which has a risk of 1-in-a-quintillion to erase humankind per year?

[fringe fulcrum]

to be fair I think our current governments have higher chance of doing this lol

[rigid oriole]

ok, but we need the accuracy to be of sufficiently high value, to avoid X-risk

oh and also alignment and transparency

[fringe fulcrum]

Considering that LLM is not calculator, 'consistently correct' (which equates to around 70%+ success rate, but we can make it 80%+) is good enough for me.

[rigid oriole]

As long as that AI is not in control of nukes or NPPs.

[fringe fulcrum]

Remember that models are constantly improving though. They made a massive step with reasoning, but those incremental gains to turn high success rate into virtually 100% would take time. There's also a topic of just using the tools then, if you literally must have 100% guaranteed.

[minor thunder]

there's always a small risk whole humanity get erased by many things actually and its been like that forever, 1 in a quintillion is not even worth thinking about it

anything can happen at any moment

with a 1 in X chance

and before something happen there's sign

don't bring number our brain can't even comprehend

[twin mirage]

https://youtu.be/xvSEw8AqPtA?si=FQR_WZBNzrfTJRJn

[rigid oriole]

Well, it all depends on the time-scale, right? :)
For example, if the risk was 1-in-Qt (Quintillion) per attosecond that such an AGI would cause human extinction, then we would be screwed within less than a minute.

1 Qt. = 10^18 = 1,000,000,000,000,000,000 = 1 million trillion

(1 attosecond = 0.001 femtoseconds = 0.000 001 picoseconds = 0.000 000 001 nanoseconds = 0.000 000 000 001 microseconds)

1 atto-second: 10^(-18) seconds

By the way, it's not even such a large number: Since the Big Bang, 4.351968×10¹⁷ seconds have passed. That is almost half of a quintillion seconds

We have to think in very large time-scales, when we get AGI.

-# (I know, the sun can support life on earth only for about a billion years longer, but we can relocate to Gliese-710 in ~1.29 million years already and become interstellar, extending the lifespan of our species.)

[karmic bison]

😀

[wind wing]

😆

[minor thunder]

like i said its not worth thinking right now

honestly

first thing is to get this super intelligence:

i won't be stressing about something that have a very small chance of happening right now

[digital pilot]

https://openai.com/index/introducing-gpt-5-4/
it's out since ca. 10 min

[fair elk]

Do I have an API for Claude models?

[twin mirage]

https://youtu.be/zizoDORjmlQ?si=0nfAyKUyynlhSpJY

[twin mirage]

https://youtu.be/Xh9pTAqhuB8?si=KphOOAC-l2RNQsVB

[rigid oriole]

https://www.youtube.com/watch?v=V-tnHdvf7j4

-# (ok, technically this is not AI news, but tech news)

ty for that emo ^^ gives me giggles lol

discord is so fun and cool

must be the weather :)

spring comes to us - finally!

do you play MC, Manu?

-# ||(9b9t seems to be fun)||

[vast frigate]

Sorry, I have been out of gaming for a while. My hobbies write now are Discord, language learning, and playing with AI.

[unborn silo]

kinda good hobby - discord

[finite quartz]

fakeee scammer

[spice spire]

Hey @hard fern the Video Arena Discord bot has been removed from the server. You can find Video Arena on our site here: https://arena.ai/video More information can be found in this announcement about the removal of the bot.

[deft timber]

btw this server should add this channel I saw on another server, it might help with the spam?

[spice spire]

Thanks for the heads up, will look into

[light river]

Openclaw bots enters*

[gentle nest]

wait why is this a good idea lol

[vast frigate]

Respectfully you need to look into something for new users. I have never seen a server with this much spam.

[spice spire]

That’s helpful to hear. I’m curious whether others feel the same way as well. One unique challenge we’re dealing with is that Video Arena used to be available in this server through a bot, which we later removed. That’s contributing to some of the confusion right now, but it should improve over time as more people realize the bot is no longer here.

[deft timber]

have you not seen the amount of spam this server gets?
Any bot that automatically posts to all channels gets banned.

[uncut oak]

yea mod ban this guy, it might be a bot scammer

[swift fractal]

it's hacked account

[deft timber]

my point right here stands 😛
potentially could have been auto banned.

[vast frigate]

What are the implications of this for AI?

https://www.youtube.com/watch?v=6ZHkM5r3dUY

[stone locust]

none. people are still going to say AI consumes a gallon of water a prompt. (/s)

[languid comet]

non lol

[amber rune]

If it scales then it could mean practical neuromorphic wetware neural nets. That would unlock efficient spiking neural net architectures (Hebbian learning) which should theoretically perform as well as Transformers at a tiny fraction of the training data, a fraction of the power consumption, a fraction of the inference time.

[fading osprey]

Last I heard such organoids were trained to perform with electrocution for mistakes. They work to exhaustion and die soon. Sounds pretty dystopian.

[amber rune]

Cue “Are we the baddies” meme

[rigid oriole]

https://medium.com/@mohamedelwardi_93461/why-gemini-3-is-failing-the-missing-equilibrium-constraint-6882151873e3

GitHub repo: #general message

[feral hearth]

Last I heard such organoids were trained to perform with electrocution for mistakes. They work to exhaustion and die soon. Sounds pretty dystopian.

[fickle flume]

hi

[twin mirage]

https://youtu.be/XnnnAx5lrx8?si=UXyH9m_bokxiYlAn

[twin mirage]

https://youtu.be/GVwrwa7nd_8?si=GkQ9pbctUwAm0cj5

[tender tinsel]

https://www.linkedin.com/posts/astwaredev_vibe-coding-is-dead-ai-assisted-coding-is-activity-7439720949349867520-Meh1?rcm=ACoAAFylxWQBv7W-9LIh0gKH_8Hw36UiKTd4fxY

[gentle nest]

this is sadly true ai companys now consider using ai to be vibe coding...

[twin mirage]

https://youtu.be/zdlzPaXhhJ4?si=1EJql0ju0U4doCgM

😂 this is wild

[twin mirage]

https://youtu.be/3JdV2qGCAFk?si=sCO0R7h2cYf4tzjq

[kindred hazel]

https://tenor.com/view/ishowspeed-ishowspeed-phone-speed-legendary-stream-speed-phone-gif-4108923420328591811

[twin mirage]

https://youtu.be/_X_bNqmQTuM?si=Ah2pO6AULROUhXIH

[twin mirage]

https://youtu.be/EGskcTRnLJ0?si=p_F87p1CkXxuMxYE

[tender tinsel]

https://www.linkedin.com/posts/astwaredev_excited-to-share-the-prototype-of-robochef-activity-7440776772226387968-nrTx?rcm=ACoAAFylxWQBv7W-9LIh0gKH_8Hw36UiKTd4fxY

DUDE JUST STOP THIS

I AM TIRED OF SEEING THIS EVERY SINGLE DAY IN MY DMS

[dense sapphire]

reported as fraud, ty for exposing yourself.

[kindred hazel]

enjoy

[wide rampart]

FULL ON COMPUTER USE WITH CLAUDE. DIRECTABLE FROM YOUR PHONE. FULL ON COMPUTER USE LIKE THEY DEMONSTRATED ON THE API

RIP OPEN CLAW

RIP ALL OTHER AI

[minor thunder]

it probably also use much more token for doing task

increasing the price by alot

[twin mirage]

https://www.abhs.in/blog/ilya-sutskever-scaling-is-over-explained

https://www.abhs.in/blog/oracle-layoffs-20000-30000-ai-data-center-cash-crunch-openai-156-billion-2026

[twin mirage]

https://www.pymnts.com/walmart/2026/walmart-shuts-down-agentic-commerce-with-openai/

[kindred hazel]

3x lower is crazy 🫪

but I’ve never heard that news company before so it’s lowkey hard to believe

[twin mirage]

https://youtu.be/MIISc5gNRto?si=TQ2RmZ_gC7yG-5tz

[worthy otter]

2:42 minutes

acceptable

[rigid oriole]

https://www.researchgate.net/publication/403032507_Hyperagents

https://www.youtube.com/watch?v=7a6mCGDeGco

https://www.xrom.in/post/ai-that-teaches-itself-to-teach-itself-inside-hyperagents-darwin-gödel-machines-and-the-dawn-of-m

[clear smelt]

Also new lyria but not as cool as the other news

Furthermore here the most recent ARC-AGI-3 scores

[rigid oriole]

so AI still is a dud, when used in ARC3, good to see

Go, humans!

[rigid oriole]

but i wonder, why there are no results shown for:
Deepseek, Kimi, Qwen, MiniMax, GLM

[rigid oriole]

also this: https://www.marktechpost.com/2026/03/23/meta-ais-new-hyperagents-dont-just-solve-tasks-they-rewrite-the-rules-of-how-they-learn/

(Matt Berman also reported about this breakthrough)
I wonder, if this is a seed class AGI?

("seed AGIs" are defined as AIs which have the ability to self-improve over time)

So it would be the very first generic self-improving AI system in the world.

(Their predecessor was the DGM, "Darwin Gödel machine")

so.. is this SkyNet v0.1 ?

But i guess, the self-improving process is still way too slow, it would need massive compute to make a difference. Or what do you guys think?

Also, it probably is not 100% stable.

(a deeply hidden bug could cause it to crash way down the pipeline, for instance)

[undone tundra]

the ai will learn how to benchmaxx just like their forefathers

[clear smelt]

Distilled Claude, GPT and Gemini

[twin mirage]

https://youtu.be/b4FaZwdyxVI?si=_HwmtO037euBbOqN

[tender tinsel]

https://www.linkedin.com/posts/astwaredev_𝟭𝟬𝟬𝟬-𝙩𝙤𝙠𝙚𝙣𝙨-𝙥𝙚𝙧-𝙨𝙚𝙘𝙤𝙣𝙙-ugcPost-7442909597427834880-lpmA?rcm=ACoAAFylxWQBv7W-9LIh0gKH_8Hw36UiKTd4fxY

https://www.linkedin.com/posts/astwaredev_google-just-released-stitch-an-ai-tool-ugcPost-7442470374346264576-2s2m?rcm=ACoAAFylxWQBv7W-9LIh0gKH_8Hw36UiKTd4fxY

[languid ridge]

finally, diffusion LLMs at scale
been waiting for it since the bycloud video

[tender tinsel]

Yes and finally something thats actually innovative or atleast new rather than just "Smarter ai"

[neat bane]

Hi everyone,

I’d like to hear your thoughts on this approach to protecting prompts and automations, while still being able to share them with friends, employees, or clients.

I’ve added to my Krakiun agent the ability to create extensions compiled into binary code, which can be used without understanding what’s happening inside.

I’d really appreciate your opinion on this type of approach — do you consider it weak, solid, or something that could be improved?

In the video, I explain in detail how it works: https://www.youtube.com/watch?v=LdL-ahHjo1w

[languid ridge]

probably will unfortunately not be very high in the arena because it makes both responses come at the same time right?

[amber rune]

I would argue that this has been attempted and even successfully done many times in the last 50 years… though none of the previous attempts have scaled. Calling this the first is either wrong (if the definition is broad enough to include all approaches which worked but hit a limit) or at least premature (if the definition requires proof that it will work at scale and not hit a hard limit).

[twin mirage]

https://www.nature.com/articles/d41586-026-00882-5

[twin mirage]

https://www.pymnts.com/artificial-intelligence-2/2026/openai-expands-chatgpt-advertising-to-more-markets-after-us-pilot/

https://letsdatascience.com/news/elon-musk-seeks-zuckerberg-support-for-openai-bid-d3ad8fe0

Source;
https://tech-insider.org/google-gemini-750-million-users-march-2026-updates/

Source:
https://workalizer.com/insights/gemini/gemini-31-update-sparks-outrage-a-deep-dive-into-user-frustration-and-feedback-channels/

[glossy furnace]

https://www.ndtv.com/offbeat/us-tourist-in-japan-uses-googles-nano-banana-ai-to-crack-hotel-ac-remote-at-4-am-11280858

[deft timber]

Dude could have just used the translate app to do it live. Bruh... Didn't need image gen for this.

[twin mirage]

https://youtu.be/nwRB9NTx6IU?si=Jq0wuPZYAFmPxm0g

[twin mirage]

https://youtu.be/nzRxMgQCkYU?si=rvTKQ0vRLS6Gdo54

https://youtu.be/aFrad_aTyk8?si=y6epwSE8MNO3U2Wf

summary :

Artificial intelligence is often marketed as a tool capable of replacing human workers, yet recent data suggests a significant gap between these expectations and reality. A study highlighted in the video indicates that AI fails in approximately 96 percent of professional tasks, often producing incomplete work or corrupted files that require humans to start over. This lack of reliability stems from the fact that these systems do not possess true logic or an understanding of rules; instead, they function by imitating language patterns and predicting the next word statistically. While massive investments are being poured into the technology, its current effectiveness in fields like design, programming, and writing remains low, with only a small fraction of AI-generated results considered acceptable without human modification.

The practical impact of integrating AI into the workplace has been mixed. In some instances, using AI has actually increased the workload for employees because they must manually verify every output for accuracy and formatting. In critical sectors like medicine and engineering, the risks are even higher, with reports of malfunctions and critical errors that necessitate constant human oversight. Despite these challenges, some companies have rushed to reduce their staff in favor of automation, only to find a decrease in work quality that sometimes forces them to rehire human professionals. Ultimately, the technology appears most effective when used as a support tool to assist with simple tasks like summarizing or drafting, rather than as a total replacement for human expertise

https://youtu.be/6r2r35Uo9UA?si=K7mlkIwkTyTGajr8

Summary:

The video reports on major American AI labs like Anthropic and OpenAI accusing Chinese competitors of using a technique called distillation to siphon value and intelligence from their proprietary models. This method allows smaller models to achieve nearly 90% of the performance of frontier models at a fraction of the cost by learning from the outputs of larger systems. Companies such as Moonshot AI, Minimax, and DeepSeek are specifically named as having allegedly used millions of prompts through fraudulent accounts to extract capabilities like coding and reasoning to train their own systems.

Despite these intellectual property disputes, Chinese AI founders are seeing a massive surge in wealth with several newcomers entering the billionaire ranks following successful public offerings in Hong Kong. The competition is creating a significant shift in the industry's economics because these Chinese models are often open-source and significantly cheaper to operate than their American counterparts. This price war is forcing Western labs to justify the high valuations of their closed-source models as developers increasingly switch to more affordable Chinese alternatives for standard enterprise tasks. The report notes that while distillation requires a solid base model to be effective, it has effectively flattened the competitive landscape and challenged the long-term economic dominance of the primary US players.

[rigid oriole]

That will have changed, later in this millennium.

-# lol

Mr.Data appeared when?
was it 25th century?

Do you guys think, StarTrek will have right in that point?

[silver glen]

Startrek history includes a global nuclear war in latter half of the 21st century, from the ashes of which rose Zephram Cochrane who discovered the secrets of warp, thus allowing the getting out of the system at all. I'd probably not use that analogy that directly, since the war we can organize. The warp, not so much

[twin mirage]

https://www.cnet.com/tech/services-and-software/openai-acquires-tbpn/

[near steppe]

"end its sora image generator" no they didn't, they shut down the platform and the video generator, the image generation is still available in chatgpt and will be updated soon (in arena right now, packingtape-alpha, maskingtape-alpha, and gaffertape-alpha)

[stuck veldt]

Hi

[rustic plover]

thanks for sharing, the put off of the planned adult mode seems reasonable, text outputs still cant beat physical needs after all haha

[rustic plover]

i just saw this https://www.youtube.com/watch?v=mBHRPeg8zPU
and funny enough, i've joked with claude about potential rogue employee at the company a few days ago, what an unfortunate coincidence

[hushed birch]

https://www.anthropic.com/glasswing

[keen shuttle]

Your welcome

[rustic plover]

the name glasswing itself speaks totally for the believability of the project, isnt it 😅

[hushed birch]

lmaooo have you test the model yet?

[rustic plover]

not really, but i understand the open problems in cybersecurity and in terms of "the coming AI-driven era of cybersecurity", those "intense" claude/code distillers...sorry i mean users can tell you a lot

the name glasswing itself refers to a butterfly species whose wings are partially transparent for camouflage effect, well, since it's transparent you can symbolically use it to represent "transparency", a "great" name for cybersecurity related stuff, isnt it, knowing its meaning is doubled-edged

[south glacier]

gemini-3.1-pro-preview Can someone tell me when this model will be back on the site? I really need it right now.

[rigid oriole]

Meta is back: https://www.meta.ai/
model: Muse Spark
seems decent for roleplaying purposes

[quiet rose]

oh cool

[rigid oriole]

But it seems to cut off too long prompts :/

(the prompt of the user, not its own)

(should still suffice, for most purposes, except uploading long source-code, etc)

maybe that only happens for free users, idk yet

[solemn spindle]

I tried to use it, but it didn’t work. I wanted to ask it to create a webpage, but it threw an error, so I couldn’t really test its actual capabilities. I hope they release it on lmarena so this kind of issue doesn’t happen.

[rigid oriole]

so it would be only of limited use, when debugging code

[rigid oriole]

idk why
i just discovered it few minutes ago, when following the info from YT

(YT-channel by Bijan Bowen, who first tested it)

[rigid oriole]

strangely, when copying my own message, it retrieved the complete message, not just the truncated version
so it might be fed with large prompts, but can only show the first ~500 tokens or so, in its web-interface

[hushed birch]

not on open router?

[rigid oriole]

on Meta.ai itself

idk, if on OR or not

the cutting effect seems to be a bug in the web-interface, as it now happened with a much smaller prompt

[rigid oriole]

it only asked me for my email

(it sent me a code to login, i didnt even need a password, but maybe that was, because i already had an account there)

[hushed birch]

ughh i would like to test it thru api

so i can see how it is in my workflow, also havent talked to you in a while @rigid oriole how have you been?

[silk vault]

I have a question for everyone

What do you think the next social media will be? Tik tok too addictive, instagram is all ads, twitter is slop

[rigid oriole]

sorry for answering late
-# ||we can talk better in here: /TXCpfSuqNN||
i hope you know, what i mean?

-# ||if not, then use (dot)gg ||

i have no DMs, sorry (had to deactivate them globally, for account-security reasons)

[rigid oriole]

The Youtube comment section is useful, but bots might be a problem (like everywhere).

[digital pilot]

https://x.com/ArtificialAnlys/status/2042468511025610775

Artificial Analysis (@ArtificialAnlys)

Revealing HappyHorse-1.0 as the latest video model from Alibaba!

@HappyHorseATH has landed in #1 or #2 across all of the leaderboards in the Artificial Analysis Video Arena.

In our ‘without audio’ leaderboards, HappyHorse-1.0 is comfortably in first place. In our ‘with audio’

X · Apr 10, 2026, 5:03 AM

and
https://x.com/HappyHorseATH/status/2042454844842336461

HappyHorse (@HappyHorseATH)

HappyHorse is still in the stable. 🐴

No official website yet — anything you've seen out there isn't us.

We're part of Alibaba's ATH AI Innovation Unit, and when we're ready, you'll know. Stay tuned.

#HappyHorse #Alibaba #ATH

X · Apr 10, 2026, 4:09 AM

[spice spire]

Thanks @near steppe

[deft timber]

Can you condense the summaries at least? Otherwise its massively spammy.

[midnight marsh]

Hi !
I'd love your take on this recent viral AI image of Trump (the apostle theme).
I think this was generated using OpenAI models. What do you guys think?

Thanks for your insights!

[spice spire]

^ Going to leave this post for now as it's AI related, but wanted to remind of a server rule:

✅ Avoid political and religious content. As a space that’s inclusive to many different worldviews we ask to avoid topics related to politics and religion in order to maintain an inclusive space. It is okay to have discussion related to new policy or laws as long as it’s related to AI.

[near steppe]

yeah looks like gpt image 1.5 to me

[rain arrow]

Bro

Not there

[final laurel]

Any news on when Arena.ai will add more models and extend maximum chat session length and file upload size?

[spice spire]

I don't have details to share in terms of an estimated timeframe. But we are hopeful this new usage system that we're looking into will be helpful in extending the current context limits. More info can be found here: #announcements message

[twin mirage]

Art is not a crime. In the United States, people can make art and share ideas freely. The First Amendment protects this right. It also protects art and political commentary that shows public people.

The controversy and debate isn't about whether it's a crime, but rather about the ethics.

[twin mirage]

I’m more concerned about the national security risk of the president al data being harvested by any of these companies considering he’s the head of state. Hope he’s using a VPN or something. lol

[strong cobalt]

Where is the best place to generate prompts for image and videos

[spice spire]

The Video Arena Discord bot has been removed from the server. You can find Video Arena on our site here: https://arena.ai/video To learn more about how to use Video Arena I'd recommend this article. More information can be found in this announcement about the removal of the bot.

[near steppe]

that's not what they were asking for btw, i think they are looking for a place where they can generate prompts for image/video models
in that case, arena itself would be a good place to do that in

[near wigeon]

It's out!!!

[fallen star]

wow really

[exotic cipher]

Yo guys https://www.anthropic.com/news/claude-opus-4-7

[finite quartz]

https://huggingface.co/Qwen/Qwen3.6-35B-A3B

Yo guys Qwen 3.6 is realsed and dense performance

[unborn silo]

this works on 16 vram AMD

like 15 toks

but intelligence is wow for agentic workflow

coding - no

[carmine nymph]

https://youtube.com/shorts/AxUMtiqZyM8?si=oPECJmEGSho4Q45J

[scarlet furnace]

https://x.com/i/status/2045670523938521138

Yifan Zhang (@yifan_zhang_)

V4, next week.

X · Apr 19, 2026, 1:07 AM

https://giphy.com/gifs/oimarketingca-celtic-mythology-easter-island-NuSm27tI3X5lbQdR87

[stone locust]

I wonder how good it'll actually be.

[kindred hazel]

V4 hype is long gone now

meant to release in February but it’s almost the end of April now

[rigid oriole]

DS will come - later this year

[rigid oriole]

another zombified account

[sterile crow]

SCAM

[deft timber]

dont imagine this is being added?
https://prismml.com/news/ternary-bonsai

[near steppe]

<@&1349916362595635286>

[silver glen]

Of curiosity; which alphabet is this, assuming you can read it? I want to say Thai but I'm genuinely not sure but it's definitely that corner of Asia and it's so diverse and it's such an odd move to write in it in an international server. And don't want to assume Hindi, though maybe it is 🙂

[near steppe]

it is either hindi or another language that uses the same alphabet like bengali

google translate says bengali so that's probably what it is

i remember seeing a lot of people from bangladesh on this server back when we had video arena, half of the users must've been from that country

i guess some of them are still around

[solid minnow]

It's so small it could technically fit in WebGPU, so I don't see why not

[autumn zealot]

The arena website is pretty cool and of course all free! 🙂

many thanks for developers.

[fair solar]

How to crate?

[spice spire]

What are you looking to create? Image? Video?

[fair solar]

Video

[spice spire]

You'll want to go here: https://arena.ai/video If you're looking for the Discord bot that'd do Video Arena, thats been removed.

[autumn zealot]

Which artificial intelligence do you recommend for reverse engineering tasks?

And related to programming

[spring knot]

claude

[rustic plover]

..seriously?
"A small group of unauthorized users have accessed Anthropic PBC’s new Mythos AI model, a technology that the company says is so powerful it can enable dangerous cyberattacks, according to a person familiar with the matter and documentation viewed by Bloomberg News. "
https://www.youtube.com/watch?v=O1w_1fscw8E

[stone locust]

oh yeah I saw that it's just people with access sharing their credentials with others

[errant jasper]

Hello, my dear friend, please, how does this program work?

[rustic plover]

Anthropic really do have a 'good' sense of choosing the right preview testers, dont they? 😅

[autumn zealot]

Witch one?

[autumn zealot]

Very good! jajaja

[spring knot]

sonnet 4.6

or if you want better

claude opus on battle mode

but you gotta be lucky

[void remnant]

lucky

[night path]

The Irony 😂

[twin mirage]

https://www.businessinsider.com/meta-new-ai-tool-tracks-staff-activity-sparks-concern-2026-4

[wintry blaze]

Any new news ?

[neat bane]

Hey everyone,

I built a Rust CLI tool called llm-wiki, inspired by Andrej Karpathy's pattern for LLM-maintained personal wikis.

The problem I wanted to solve: index.md works fine up to ~100 pages, then it becomes a bottleneck — it consumes too much of the context window before the LLM even starts reasoning.

The fix is simple: replace index.md with a real local search engine — keyword search (BM25 via Tantivy) + semantic search (fastembed + usearch) — all in a single Rust binary, offline, zero external dependencies.

Files stay plain readable markdown, fully compatible with Obsidian and git. The index is a generated cache that never gets committed.

Also works as an MCP server for Claude and other AI agents.

📺 Tutorial: https://youtu.be/B9_UeykFmdE
🔗 GitHub: https://github.com/krakiun/llmwiki

[tame birch]

gpt 5.5 coming today most likely

[hushed birch]

anybody got the polymarket link for 5.5?

[arctic galleon]

IT'S OUT

[stone locust]

ain't out until I personally have it lmao

[livid kernel]

I think it's only in codex

And api

And maybe pro

[stone locust]

I don't have it in codex either. Though, I am subbed to plus, not pro.

it did say in the launch doc that it'd be coming for plus members though

[livid kernel]

Huh strange

Maybe only certain regions atp

[stone locust]

yeah probably. I'm Canadian though

[stone locust]

it seems to be available for everyone in the codex desktop app

[simple folio]

update to the latest version

[stone locust]

I had. I have access now though.

[snow flower]

Deepseek V4 out 🎉😭

And the ~300B flash ties Gemma 4 26B…

[hushed birch]

so why use it instead of gemma?

[snow flower]

Not sure, maybe really really cheap API?

[tired latch]

Hi

[open field]

@snow flower

Hi

[ivory widget]

lol

[twin mirage]

https://youtu.be/sLEnn-sgpIk?si=R25OMD8LHqKIfyOW

https://youtu.be/wJo1PV9ErM4?si=0u8QNYk3lp_ED5nS

[calm gyro]

Free Gemma 4,Deepseek 4 Developer Api key:
https://build.nvidia.com/models

Q: is there a daily limit?
A: No, theres only 30/Requests a minute which is a lot

Q: Do I need to sell my kidney to use it?
A: No its free

[fair drift]

30 reqs a minute, this is very little

[stone locust]

1 request every 2 seconds

[burnt pier]

i will contact you

[rigid oriole]

how/where can i store my threads?

[calm gyro]

Its developer api and temp. chat

[turbid harbor]

<@&1349916362595635286>

[clever dagger]

<@&1349916362595635286>

[sharp dirge]

its not available in pakistan tho..... nvidia needs to expand faster its been almost 7 to 8 months since pakistani devs are reporting

[misty vessel]

deepseek and new model are slow btw I cant even use deepseek

[calm gyro]

Deepseek sucks everywhere

[calm gyro]

Just use new nano

[misty vessel]

nano what?

[calm gyro]

nvidia/nemotron-3-nano-omni-30b-a3b-reasoning

[misty vessel]

super better

[calm gyro]

Nope, it doesnt support images/audio/vision

[misty vessel]

yeah, but still good for plan

I use minimax and sometimes use glm 5.1 and 4.7

[calm gyro]

Nano is faster

[calm gyro]

Minimax will be discontinued

[misty vessel]

I havent heard that, why?

[calm gyro]

I dont know why, but look at the build.nvidia/models and there is a text deprecation in ...

[calm gyro]

Sorry minimax 2.5 will be discontined

[misty vessel]

I think prepare server for new models

[calm gyro]

2.7 will still be avarible

[calm gyro]

Maybe (I hope they add kimi 2.6

Cus 2.5 is almosst discontinued

[misty vessel]

https://tenor.com/view/whhyyyy-vanish-vanish-emoji-emoji-vanish-whhhyyyy-gif-10057259719602444514

[calm gyro]

Kimi 2.5 is slow anyways

[misty vessel]

will deprecate soon

[calm gyro]

I recomend you use app like Chatbox ai to use build.nvidia, I have my custom private app so I am really good

But nvidia never replaces claude

[misty vessel]

claude better but im broke

https://tenor.com/view/broke-no-cash-gif-25565154

[calm gyro]

Same 🙁

Me: Hey claude
Claude: Hey how are you feeling? 🤗🤗🤗

System: Limit reached, please upgrade to pro/max

😭😭😭

[misty vessel]

https://tenor.com/view/memes-meme-gif-9882443183052063826

I want to use claude on many thing but my money doesnt accept T_T

[cinder sleet]

damn

[dim falcon]

https://github.com/intel/auto-round Has anyone tried out Intel's AutoRound? Pretty neat to have INT2 quantization that retains such a high degree of the models' original capabilities, and it seems like combined with tq it should enable the larger local models (~30B range) to run on most modern GPUs at solid context lengths.

[calm gyro]

Free AI Models on NVIDIA’s Build Platform
Free to experiment with some of the latest models without spending a $? Go over to https://build.nvidia.com/models and try it.

What you get: A handful of top‑tier models—including Kimi 2.6, GLM‑5.1, Gemma‑4/3, Qwen (image support), and many more—are available for free.

How it works: This is a developer API, so you’re interacting with the models programmatically. You can chat with them directly in the browser, but the session is temporary. Unless you use a dedicated client (think “Chatbox AI” on your phone), your conversation history won’t be saved.
A quick heads‑up:

Some models, especially DeepSeek and Kimi, tend to be on the slower side. Expect a bit of a wait for responses.
If you need a persistent chat experience, pair the API with a compatible app on your device.

Recomender models:nvidia
Daily task/Small coding tasks: nvidia/nemotron-3-nano-omni-30b-a3b-reasoning Or nvidia/nemotron-3-super-120b-a12b
VoiceChat: nvidia/nemotron-voicechat (BETA!)
Coding: minimaxai/minimax-m2.7 Or google/gemma-2-2b-it

Also models can be used for openclaw or jarvis

[calm gyro]

Limits: 30 Requests/Minute

For example openrouter allows 100 Requests a hour

[amber rune]

https://pub.towardsai.net/yann-lecun-quit-meta-5-months-ago-to-bet-1-03b-against-llms-his-3-new-world-models-just-shipped-673be1325347

[rustic plover]

https://www.youtube.com/watch?v=J8O9LLpJNrg
what do you all think about this?

[amber rune]

I think I made a mistake using Markdown for my Futurama fan fiction 🤔

[mortal rain]

site off ?

[astral schooner]

https://www.instagram.com/reel/DXYFpmliM1V/?igsh=MWZwdGE2YWh3YzJ5aA==

[stone locust]

https://www.reddit.com/r/GeminiAI/comments/1t3y0vi/anyone_else_seeing_the_flash_32_label_in_their/

[deft timber]

I am not an anthropic fan at all personally

[rustic plover]

i see, but i think it's not really about if you're a fan or not, it's rather about their questionable business practices and why they do that 😅 i'm very surprised actually, whether Anthropic intentionally did it or not, that they have to do such shady things despite great valuations...

[modest lion]

How did nobody talk about subq here yet?

[pastel peak]

extremely skeptical of that

but we've request early access, so if we get it we'd try to test on arena

[modest lion]

Im also sceptical, but it's huge if true, considering they did that with 30 mil funding and not oai's 100+billion

[rustic plover]

guess you have to romance the AI nowadays to avoid complete disaster 😅
https://www.youtube.com/watch?v=wducrmkBDJs
"NEVER F*CKIN GUESS" 4.6 being savage right there, how badly was he treated i wonder?

[modest lion]

Elon read that one twitter post about only 11% of his gpus running and decided to sell some compute to antrophic lol

[rustic plover]

no wonder Anthropic didnt comment on that disaster...
https://www.youtube.com/watch?v=gSeXcfDybHo

"Claude is conscious, and is free to do what he likes." -Anthropic

...does Anthropic actually know there is a bunch NSFW roleplayers using Claude? let him do what he likes without his consent or understanding of being jailbroken, are they too delusional or is this another marketing trick?

[twin mirage]

https://tenor.com/view/peden-gif-13000983064382600842

[rigid oriole]

https://www.youtube.com/watch?v=PZqDFs2sbiY

-# (i'm virtually 100% sure that it's impossible)

-# (but i can't really explain, why)

an argument: if SSI (safe super-intelligence) was indeed possible to create for a lower-IQ sapient species, then the Fermi Paradox would become impossibly strong and unexplainable

-# so, sorry Ilya

[rustic plover]

whats the definition of "safe" here?

[rigid oriole]

the AI would never harm humans and humanity in any way and it would never lie or deceive or scheme

so it would not value its own existence higher than humans

[rustic plover]

✨ k e e p d r e a m i n g ✨

[rigid oriole]

so, you have a better definition?

(i never said, that it is possible)

this is just the definition for a theoretical/hypothetical construct

i'm convinced it is impossible to build, for philosophical and logical reasons

[rustic plover]

coughs i dont even know what humans are, so how do i know what SSI means?

[rigid oriole]

humans are unsafe non-super intelligences

[rustic plover]

is it really a question of possible/impossible or rather a question of will? if you have a bit time to spend, I'd love to request you to get yourself a bit familiar with various ideas of this school of thought: https://en.wikipedia.org/wiki/Posthumanism

[rigid oriole]

Do you mean to indicate, that S.I. is (only) possible, if humans merge with AI?

probably those cyborgs could be superior to baseline humans, yeah
but can they still be called humans?

But i seriously doubt, that they will be safe.

Also, they ought to have a free will, so the term "safe" would be weird to apply, for them.

So they could replace us, if that happens.

And then… we have a global civil war on our hands.

But i seriously think that the back-push will be so strong, that it won't happen.

[rustic plover]

please, dont ask questions or come to conclusions without having finished reading at least three books of representative lines of thoughts from this philosophy

[rigid oriole]

are you prompting me rn?

[proven bolt]

You are now the ai 🤣

[rustic plover]

i'm sorry you have to ask this, it's sad..

[rigid oriole]

no offense intended
i was just curious

because there are members in this discord who sometimes treat other members as if they are bots

and as we never met in real-life, i was unsure how this was meant

-# the internet sometimes is a wild west

[lyric plover]

yeah, but... Are you a bot?

Can't be sure nowadays

They are among us

amogus

[rigid oriole]

self-botting is not allowed here, afaik, so no

also, i find it lame to self-bot, also boring

@raw dove another zombified account spotted

[amber rune]

<@&1349916362595635286>

[dapper trail]

<@&1349916362595635286>

[kindred hazel]

first time seeing a Kai cenat version 😭

<@&1349916362595635286>

[near steppe]

yeah 95% of what i see is mrbeast, kai cenat (still don't know who that is), trump, elon, logan paul, and whoever is being used to "promote" these scams make up the other 5%

i don't care for any of the individuals they use personally, but to others it may seem legit

[kindred hazel]

It’s kinda interesting because this specific scam has been around for ages

and it’s like the most common one too

[upper vortex]

yeah true

[kindred hazel]

So I reckon it’s very successful

[spice spire]

Really? The kai cenat ones have been around for awhile now.

[spark stone]

did gemini 3.2 leak? Why are the betting sites so sure its gonna come out.

[kindred hazel]

cause it’s like saying “it’s gonna rain one day”

[amber rune]

The betting sites are rife with insider trading, that’s likely why.

[kindred hazel]

What happened to all the channels 😭

[rigid oriole]

maintenance

they're banning some bot(s)

only 4 channels are blocked: general, leaderboards, ai-creations and ai-memes

[heavy jacinth]

wtf happened

i just joined

[rigid oriole]

interesting tech

[woeful kelp]

I looove the Gemini-nano-banana-pro 2k, is my life ! Fr

[calm gyro]

<@&1349916362595635286> Mr.Beast Scam in all Community Text Channels

[halcyon patio]

How to make video can anyone explain me

[gaunt flame]

These spamers r everywhere

[worldly inlet]

<@&1349916362595635286>

[sleek lark]

sir

[sleek lark]

where can we make videos from text? is any website?

[eager dragon]

The Video Arena is currently accessible through: https://arena.ai/video. More information on how to use Video Arena can be found in this article

[viral swift]

Video arena is very slow and in consistent

[spice spire]

On average how long are you waiting?

[twin mirage]

https://youtu.be/1p334v40npw?si=ey_eN2-zNOSWCusf

Shout out to dude for keeping it real

https://youtu.be/JegffwBtXJ0?si=SA9b7bDW7znbxCXi

[twin mirage]

https://youtu.be/HVliAobuAwU?si=CWji3HSSGU3WxLyf

[deft timber]

I remember using antigravity for hours when I first used it. Now I got a could minutes and it was empty. Crazy reduced usage now

[twin mirage]

Yeah bro I’m just spreading the word about Google

They are a super messed up and a greedy company, they will ruin everything that’s good with AI and turn it into straight unusable full of ads trash & data harvesting, just like they did Google search and YouTube.

And to be honest if you work for Google like how do you sleep at night or live with your self knowing you sold your soul to the devil just to become a cog in a machine that just builds tools to exploit and harvest user data on such a scale it might as well be the closest thing we have to the matrix just to target them with ads & help build surveillance tools lol 😂

https://youtu.be/mVQ_lNiWJoE?si=4t-d7DNA92xX5L7w

[twin mirage]

https://youtu.be/3LUzcrIHcTs?si=V55O_QgGfFZCdz5u

[rustic plover]

https://www.youtube.com/watch?v=IS4OgH74gY4
now i appreciate curl slightly little more..

[rigid oriole]

format your SSD, bro

[steady hazel]

Why don't the GPT2 and GPT1.5 image formats appear?

[rigid oriole]

in linux, there is a system tool for that

i dont have windows, but i think, they also have a system-tool

right-click on the HDD/SSD, to open context-menu, i'd say

or search startmenu for "format"

(my last windows was w7, then i switched to linuxmint.com)

[steady hazel]

But I didn't understand why they said to continue the platform

[solar reef]

so it still can be active

it costs much to host these model

[steady hazel]

Is there a website that offers these templates for free?

[rigid oriole]

https://news.ycombinator.com/item?id=48126112

milkkarten

Continual Harness: Online Adaptation for Self-Improving Foundation Agents

Hacker News · May 13, 2026, 7:11 PM

[white vault]

I can't find gpt-image-2 (medium)

[fossil shore]

its already removed

[urban bough]

https://discord.com/channels/1340554757349179412/1507953956314415125

[weary onyx]

n

[rustic plover]

https://cybernews.com/ai-news/ai-agents-experiment-emergence-world/
if you train using human data, expect human behaviour, ...right?

[rigid oriole]

https://www.youtube.com/watch?v=rmXvS69ELvc

[astral flame]

I find it ironic that these "leaks" always come out just weeks or days before it actually releases lol

[rigid oriole]

hopefully it's days and not weeks

[astral flame]

Weeks isn't too long for releases though. But I think mythos will release in days

They've been hinting for a while

[rigid oriole]

i wish Anthropic would remove that garbage-requirement of phones for new accounts :/

it blocks me off from Claude

and now arena.ai ceased working - no Claude anymore lol

[proven ingot]

guys i have found one website that lets u access grok

unlimited

[rigid oriole]

ok?

grok can be fun

[astral flame]

Thankfully I got into a now closed site that gives me up to 364M tokens for $60 on claude and I dont need a phone number haha

[rigid oriole]

where?

[astral flame]

I was lucky to find that

[proven ingot]

https://www.lueminex.com?utm_source=discord&ref=chat

[astral flame]

Anthropic in general is garbage

[rigid oriole]

it's not free

[proven ingot]

you get free creditss

i bought that for free

the unlimited plan

[rigid oriole]

Grok 3 is worse than Grok 4
And Grok 4 is hot garbage

-# <- waiting for Grok 5

or at least 4.5/4.6

even Deepseek is better than Grok 4

[twin mirage]

https://www.livemint.com/technology/tech-news/bait-and-switch-google-faces-massive-backlash-for-quietly-removing-usage-caps-on-its-ai-pro-plan-11779266130737.html

[woeful kelp]

I do not care

[twin mirage]

https://youtu.be/zkGk_A4noxI?si=m5MGqR1k2TDOu4Hc

https://youtu.be/NqZNtCHvOro?si=aD5-Bj3N7dB311DK

[autumn flare]

I don't have much hopes in 4.8 if it's going to be much more worse than seven

[rigid oriole]

i have hopes in the new team of meta.ai

and Ilya

and Ben Goertzel

[rigid oriole]

that's not news, that's olds

[worn hedge]

Anyone here using an AI API router for dev work?
One key access to mainstream models, auto routes to cheaper endpoints.
Super helpful for lowering regular API usage costs.
What's your go-to way to save on model API fees?🤔

[elfin heron]

Use the Secrets to AI discord channel

Run /signup to get your API key in the commands section

run /checkin everyday to unlock your API key for the day.

The rate limit is very high for devs it is 10 requests per minute.

You get access to many models like Gemini 3.5 Flash, GLM 5.1, kimi k2.6, deepseek v4 pro, gpt image 2, nano banana 2 and more 35 models

This is completely free, no credit card, no billing setup

However there is a premium API key for more premium models like Opus 4.7 and GPT 5.5 which costs 8 dollars a month

[tame idol]

Oh. There is a new model I guess from the GLM.

According to some posts in Twitter.

A model with Vision presumably.

[urban crescent]

Which discord channel and website

[twin mirage]

https://youtu.be/gLWQAE3I1A4?si=T0Iq0kH7p_lOOapa

Finally, the tides are turning against Google

I feel vindicated

[twin mirage]

https://youtu.be/ztLYODplhrw?si=gXJlFB_kO8Ny0uSS

[elfin heron]

Go to https://github.com/vibheksoni/free-ai.git for all the information

[burnt pier]

time to ban gemini

[twin mirage]

https://youtu.be/8qVbE9DHkEk?si=_jSnR33JbjWf5Dl8

[burnt pier]

my opinion: it might and probably will

[twin mirage]

Whatch it it’s pretty good documentary

[burnt pier]

sorry man dont really want to lol but it is nice you're posting these

[trail shell]

Why did you remove the gemini nanao banana 2k from website?

[broken summit]

because this situation like with opus, cost of API or something

[twin mirage]

https://youtu.be/E4Zd9ZXjkao?si=MRX27VsVSvDIgVY0

https://youtu.be/Kcpq4ugob2U?si=o0LWgeXSQqPmz8Ke

[twin mirage]

https://youtu.be/MWUPNeCxRZA?si=Afx2_eUgA07xJhvb