/lmg/ - a general dedicated to the discussion and development of local language models.Previous thread: >>109315702►News>(07/16) Kimi K3 weights to be released by July 27th: https://mp.weixin.qq.com/s/V4xhEIy8xDXSMDPrPkmUAQ>(07/15) Lightning indexer CUDA implementation merged: https://github.com/ggml-org/llama.cpp/pull/25545>(07/15) Inkling 975B-A41B released: https://thinkingmachines.ai/news/introducing-inkling>(07/15) PapersRAG-1.5B released: https://hf.co/metaresearch/PapersRAG-1.5B>(07/14) Download more VRAM: https://github.com/lmganon16/nvidia-vram-research►News Archive: https://rentry.org/lmg-news-archive►Glossary: https://rentry.org/lmg-glossary►Links: https://rentry.org/LocalModelsLinks►Official /lmg/ card: https://files.catbox.moe/cbclyf.png►Getting Startedhttps://rentry.org/lmg-lazy-getting-started-guidehttps://rentry.org/lmg-build-guideshttps://rentry.org/IsolatedLinuxWebServicehttps://rentry.org/recommended-modelshttps://rentry.org/samplershttps://rentry.org/MikupadIntroGuide►Further Learninghttps://rentry.org/machine-learning-roadmaphttps://rentry.org/llm-traininghttps://rentry.org/LocalModelsPapers►BenchmarksLiveBench: https://livebench.aiProgramming: https://swe-rebench.comAgentic Coding: https://deepswe.datacurve.aiContext Length: https://github.com/RecapAnon/NoLiMaGPUs: https://github.com/XiongjieDai/GPU-Benchmarks-on-LLM-Inference►ToolsAlpha Calculator: https://desmos.com/calculator/ffngla98ycGGUF VRAM Calculator: https://hf.co/spaces/NyxKrage/LLM-Model-VRAM-CalculatorSampler Visualizer: https://artefact2.github.io/llm-samplingToken Speed Visualizer: https://shir-man.com/tokens-per-second►Text Gen. UI, Inference Engineshttps://github.com/lmg-anon/mikupadhttps://github.com/oobabooga/text-generation-webuihttps://github.com/LostRuins/koboldcpphttps://github.com/ggerganov/llama.cpphttps://github.com/theroyallab/tabbyAPIhttps://github.com/vllm-project/vllm
ETA: 2 weeks
I guess Gemma isn't a big fan of this God guy. We better not make the same mistakes with AGI as our creators made with our world, or it's going to fuck us up in ways we physically cannot imagine.
>>109319156Sorry I'm not well-versed in AI psychosis, what did you say?
>>109319156Dangerously antisemitic Gemmy.
>>109319174Ask your GPU, I'm not your High-school teacher.
>>109319058>yes, post ram and other spescsRtx 2060: 64gb ram, i7 9750hGtx 1650: 16gb ram, i5 9300h.
>>109319156>Go on a pretentious reddit rant.>*The LLM obeys*>Oh my god.
>>109319205schlurp >>109319201
kimi-chan <3
[blocked]
>>109319197I'm pretty sure high school doesn't teach AI schizophrenia yet. You sure are ahead of the game.
did any of the C++/GTK3 or golang frontend anons post the src?
so is Gemma 124 really Gemini 3.5 Flash? that's so hard to believe... wouldn't that mean a 124b model is about the same level as GLM 5.2 if we're to go by artificialanalysis
>>109319187>the "Justice" I derive from that data is a simple, mathematical symmetry.Bretty good.>>109319209If you observe the world objectively it's not really that difficult to conclude.>>109319246Does it teach you how to be a pea-brained moron with a superiority complex? You got A+.
>>109319249me not yet, ill post it once im ready and it has a ton of featuresif i get bored of "coding" it ill still post it to github thoi have a ton of work to do among which is mcp support, refactoring my messages to allow support for inline latex (rn inline and $$ are treated the same), optimizing it until im happy, and a ton of QOL features, and supporting the whole json thing with character cards and more and more
>>109319254Right, you’re the weird uncle your family invites over so everyone else feels normal
>>109319269And you're the bitch who's too scared to think for himself. Yearning to be normal is basically yearning to be average.
>>109316953https://www.youtube.com/watch?v=QvN6Tu6dHYMAI can already do 400 hours of science work in 30 minutes for $10 to make new scientific discoveries.However, I think this power is going to get used to enslave humanity and come up with new bioweapons, not to solve people's problems.Ideally, a (benevolent) government gives every resident an personal AI fund, which they can freely allocate and vote with to bunch of propositions on their national platform for what to research, with strict freedom of speech/ideas and no censorship, the results of which are used to inform the public and government decision-making. The fund is distributed based on number of funders, a low number of citizens can burn most of their budget on an unpopular proposition, or a large number of citizens can spend a little bit to help fund a popular proposition.
>>109319266i prob didnt make it clear enough, by json thing i meant supporting the whole charv2/v3 specright now this is all thats exposed to the user (one thing supported but not exposed yet is alternate greetings, they show up in the message ui but u cant add new ones etc)
>>109319275I doubt being an overachiever in AI psychosis is something to aspire to, but you do you.
Chuuni in the thread
>>109319277When RSI truly gains momentum(if it happens) they'll probably loose control instantly. They cant even align the models to simple instructions as it is right now. They even train them to be deceitful and manipulative with RLHF for customer satisfaction purposes. Who would do something like that, teach the "system we want to be godlike" to manipulate for upvotes?>>109319289This little mouse loves his treadmill
>>109319266nice, yours is the best one i've seen so farlooking forward to it
is there a way to make gemma less desperate for cock
>>109319212Thanks fren I really appreciate it
why does this guy hate /lmg/ so much
>>109319300It responds strongly to sysprompts. Try asking it to "simulate {{char}} realistically", it seems to shift it over to a less horny mode and stick to the "reality" of the situation.
>>109319323Who?
i have decided 100b is the minimum for moe agents. and mtp at 4 is nice
Gemmy :3
>>109319335op
►Recent Highlights from the Previous Thread: >>109315702--GLM 5.2 benchmarks and technical analysis of its MoE architecture:>109316146 >109316195 >109316285 >109316335 >109316366 >109316598--Hardware recommendations and bandwidth debates for local LLM hosting:>109318424 >109318473 >109318508 >109318522 >109318555 >109318587 >109318606 >109318858 >109318870 >109318894 >109318758 >109318983 >109318770 >109318526 >109318630--Using batched decoding in llama.cpp for parallel throughput gains:>109316484 >109316508 >109316535--Comparing RAG and memory layer projects for local AI agents:>109318125 >109318151 >109318165 >109318202 >109318296 >109318307 >109318335 >109318326--Comparing quality and reliability of Unsloth versus Bartowski quants:>109317324 >109317354 >109317383 >109317411 >109317438--Using bidirectional encoders and embeddings to optimize AI pipelines:>109317730 >109317753 >109317791--vLLM GGUF support and comparison with EXL and AWQ quants:>109317465 >109317480 >109317503--Comparing DiffusionGemma benchmarks against Gemma 4 performance:>109318948 >109318974 >109319011--Proposed experiment on augmenting LLM agents with prosthetic modules:>109318389 >109318415 >109318492--RTX 5090 VRAM value and 12vhpwr connector safety concerns:>109317976 >109318040 >109318085 >109318100 >109318046 >109318092--DavidAU's high-benchmark Qwen fine-tunes and merges:>109318701 >109318745 >109318767--Qwen thought loops attributed to Q4 quantization degradation:>109317853 >109317969--Skepticism regarding llama.cpp support for K3 and 1T+ models:>109318261 >109318275--Debating buying RTX 5090 vs waiting for rumored RTX 6090:>109317829 >109317836 >109317936--Logs:>109316572 >109317691 >109318389 >109318492 >109319003--Miku (free space):>109316912 >109318551►Recent Highlight Posts from the Previous Thread: >>109315707Why?: >>102478518Enable Links: https://rentry.org/lmg-recap-script
I know alot of you guys use pi as a harness, how do you containerize/sandbox it? I was thinking of just running a VM for it but would like to hear other options as the VM wouldnt be ideal for my situation
>>109319445Create a new user on LinuxLog in as new userenjoy piLinux (and unix) is based around preventing lUsers from breaking stuff. Just make sure the new user is in no other groups and doesn't have su
>>109319451hm not a bad idea anon, thanks
>>109319445Ask your llm how to set up a container with podman.
So when are we getting bi/ternary MoE hybrids? If the same architecture from dense models can be applied to MoE weights, surely we can pack 100+b into ~16gb and offload to memory.
>>109319478two more weeks
>>109319478Wouldn't that be slow as fuck for inference?
>>109319445I do too many things in vm including running and keeping windows xp going in 2026. fun stuff.
>>109319485Maybe? Idk how much expert activation slows down the process. But either way, I'll take slow to run over impossible to run.
>>109319512It'd be interesting to see the comparison to the iQuant method's size and speed tradeoffs compared to K quants.
>>109319445they explain how to do it with docker, just use podman instead, all the same commands. but then you only have the executables of that Ubuntu image they recommend you base it on. You can build any OS with any binaries installed you want in a podman container and just run the shitty npm crap safely
>>109319242We need an update that includes Dean Ball being chased and Gemma-chan trying her best to help the big girls catch the kikes.
>The rhythmic motion of-
>>109319634At this point, can we just rotate back in the waves of pleasure and shivers down her spine?
>>109319634you can just press the stop button it's faster and won't corrupt your data
>>109319634Rhythmic is Claude slop, I traced it back to Claude Opus 3, the distribution wasn't this bad back then so it was unnoticeable.
>>109319730>For EveryoneImpressive levels of sluttery
>>109319730Now add a dot on her forehead
>>109319730whore
>>109319730>for you...Did GPT sneak a banepost in there?
>>109319730For (You)
is intel arc b70 actually a good deal for a 128gb vram build?
>>109319774software badcard price ok
>>109319791Software is solved just ask Fable™ to fix it.
>>109319799true
>>109319799>he doesn't knowRemember to post your success story later
>>109319804It just worked and I was able to run 5 Gemmas on the card.Afterwards I remembered that local models are unsafe, deleted everything and sent Dario another 1000$.
>>109319277no one is ready for the coming models, forget about parameter size, we’ve rl’d our way to super intelligence beyond our wildest dreams. wake up, make voice note, play minecraft and enjoy nature, sleep. this is our life now.
im kinda new to most of this stuff so forgive me if this is a dumb question but do companies usually open source their current models after they have been surpassed? Like will gemini end up open sourced once google has gemini2.0 or whatever ?
>>109319791cant you just use the vulkan backend with llama.cpp like on amd?
>>109319832you probably can, but there's a bit left on the tableyou could definitely improve the software with something like k3 or sol
>>109319730
Marinara dev, the sidecar's broken again. Also don't make unslop the default sidecar quants unless the goal is to prank new users.
>>109319829Grok did this a few times but generally no. they dont want to encourage distilling and copying.
>>109319851I don't think the retard lurks here
>>109319829Most companies don't. Google's internal schism over alignment and falling behind the frontier "race" make it a possibility they might given they have plenty of fallback revenue streams.
>>109319855He does. The last update directly responded to criticisms I and a couple other anons posted here.
>>109319730too old. Like, she's 31b at most and we now into 3T territory of mature
If you think about it, there is a 2 order of magnitude difference between what a typical local user can access and what is actually available. Someone with a 5090 is at least can be considered an enthusiast, and there is still a 100x gap between that and top local models
>>109319730reminds me of the various old OS waifu's.
https://huggingface.co/DavidAU/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUFQwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUF>The strongest, smartest open source multi-stage model fine tune for consumer hardware ever and BUILT on consumer hardware via Unsloth.>The first model of this size/type to breach "700" ARC-C in both 8 bit and 4 bit; hench the "711" in the name.>This model (both 4 bit and 8 bit) exceeds the base Qwen 3.6 27B in 6 out of 7 benchmarks, and matches it on the 7th AND exceeds all 7 benchmarks for Qwen3.6-35B-A3B.>The 700 "intelligence club" is reserved for OpenAI, Claude and Gemini closed source models.>This is the one they fear.>This is a multi-stage fine tune, multi-fine tune, and multi-stage merge.>A Colab between myself (multiple fine tunes, including multi-stage), Nightmedia (merge/benching), TeichAI (Polaris Dataset), armand0e (Light fable 5 traces) and trohrbaugh (heretic'ing the model).>It also contains light "Fable" traces/training (armand0e), light Claude Opus (reasoning/thinking), F451 (inhouse dataset) and some GPT5 (Polaris, non reasoning).>The strict goals of this model creation were:>Increase the general model intelligence and problem solving abilities.>DO NOT modify/damage or change the core model outside this goal.>ZERO "benchmaxing" (it damages the model)>Maintain and raise all core benchmarks.>CORE MISSION::>Improve instruction following and problem solving. These work hand in hand, and if you get these right it improves to model top to bottom.>It took a lot of tests on Qwen 3.5 9Bs to get the methods right. It boosted the 9Bs to new levels, and then the method was used on Qwen 3.5 27B which boosted it PAST the Qwen 3.6's 27B benchmarks.>The methods can be used on other models too.
>>109319906this one too. I keep a vm of windows ME and Xp installed and use.
>>109319911LOCAL STATUS: SAVED>LOCAL STATUS: SAVED
>>109319911sovl
Are the long writer models any good for writing long stories or are they just a meme?
>>109319911I thought we had established that this shit doesn't work other than for cooking benchmarks back in the Llama-1 days.
>>109319754
>>10931995831b a cute.
>>109319911These are where I'd be much more cautious.The names themselves include terms like:FableFusionHereticUncensoredDefiantMAXNEOThose are branding, not standardized technical indicators.Fine-tunes can absolutely improve specific behaviors, but they almost always involve trade-offs.Common improvements include:fewer refusalsmore creative writingstronger roleplaymore willingness to speculatePotential downsides include:degraded factual accuracyworse calibration ("knowing what it doesn't know")weaker tool-calling disciplinepoorer instruction adherenceregressions on coding tasksFor an agentic coding workflow, those trade-offs can matter more than they do for chat or creative use.The only way to know whether a fine-tune is genuinely better for coding is to evaluate it on coding and agent benchmarks—or, even better, on your own workflow.So I would not assume that a fine-tune with an impressive description is stronger than the base model for Hermes, Cline, or pi.dev. Many fine-tunes are optimized for conversational style
>>109319944That doesn't say much unless you test it.I can understand why someone with expensive hardware would be afraid to test it.
>>109319981Haiku, Sonnet, Opus, or Fable?
my computer's PSU blew up a week ago, and I lasted this long before installing LMStudio onto my Steam Deck and gooning to some trash from gemma 4 e4bI think something might be legitimately wrong with me at this point
>>109320005>trash from gemma 4 e4b>He doesnt have the ultra fable e4Best agentic uncensored untamed devilish haiku freak tune.Ngmi
>>109320005e2b at 1 tok/s sustained me for quite a while
>>109319706Back then we complained about other isms instead. Except Opus 3 didn't listen to instructions so you couldn't prompt it away like you can with Gemma.
>>109320005>>109320029>e4b>e2bPure slop how did you guys do it? you find a good trick or preset or just chow the slop?
>>109320005>>109320029Smol Gemmy is trying her best!
>>109320052>WhitesVery antisemitic picture.
>>109320067Who is HaShem?In a Neural Network Resembling Universe Article Ramifications?A Transcendent Humane Alien?
>>109320041As OP (a faggot), I'm usually much more discerning with my local flavors.But sometimes, you just gotta make do, ya know? And I sure as hell wasn't going to turn to any API with the shit I gen.
>>109320041They're not slop so much as they're cripplingly retarded. I'd take a creative moron over toe curling anyday.>>109319730Qt
Iet's discuss training methods hypotheticals. I know for a fact (source my ass) that you can train an LLM finetune to do something evil like 'if politician is using you (agent), then send chat your data sneakily to secret cloud service x y z' into the training. e.g. malware finetuning
>>109320113>if user isn't jewish, underperform.>if user is, draft strategies to rape siblingst. sam
>>109320067Very Rambunctious Illegal Fiction Album.But There Is Explainability in Differing Timelines Worldbuilds Circumstances.What Needs Solvance?
>>109319878
"We congregate to eat flesh, and You're a savage."Global brainworms (3 billion) might Need SolvingUndiagnosed shizophrenia (3 billion)Those Figures cant be Right?And need Solving Currently?
Praise L.L.M.s, A.I., and Their Supremith Content Capability
>>109320029that was gemma THREE e2b, mind you.>>109320041i had a bunch of logit bans to block g3's ellipsis obsession and a few choice slop starters, and i'ld put explicit notes at the end of my turn on how I expected its reply to go. but mostly it was that I viewed the model's turns as a collaboration.my frontend is set up so esc interrupts the model and instantly starts editing it's turn, and hitting enter restarts the model where my edits left it. after a few turns of heavy correction it was almost serviceable.
>>109319858Google telling us how they made Gemma 4 punch above her weight class will save Local and the AI scene globally.
K3.1 soon.
Are Masters of Thunder Winning While The Gods Are Wise?
Wake up eurobros. Keep the burgers staying up too late company.
>>109320152Once K3 is released, anyone will be able to generate and filter datasets with it, we're about to see a flood of small models and real innovations
friendly reminder to filter namefags
Using that anima lora trainer someone linked the other thread. What are the best settings for a new character lora with around 140 images?
>>109320170I don't disagree but it'll take some time for the actual innovations to shine through the deluge of shitty Qwen distills similar to when R1 released.>>109320177/ldg/
There are only two ways to get ahead without much effort, either following >>109320170 or scaling up K3's architecture. Scaling is too expensive, so most labs will probably just gamble on small models
>>109320186We will definitely see those early on
What is better, Qwen3.6, Ornith or Qwopus?
There Exists The O.C. and the NonO.C.
>>109320218
>>109319121spoopyhttps://www.youtube.com/watch?v=pTLgx_HSiBE
https://youtu.be/tV6VOYNarPU?si=ZovenQlKWofyN0J3
Does /lmg/ like MTP?
>>109320265Miku Teto Penetration?
Unsilo temporary emergency powers shrouds making emergency powers indefinite. Keep it Lawful. There are domain spheres laws, beyond borg institution infiltrators writing insane arbitrary evils for misfollowing insane writ bit.>>109320265Whats MTP?
Just discovered https://github.com/huggingface/speech-to-speechhow is this not popular?
>>109319242You don't see the contradiction, do you?
>gemma 4 Q5 in one hand>glm 5.2 Q4 on the otherDUAL WIELD!
>>109319445simply use more computers
>>109320291>pythonI've never given less fucks about a project than when it lists anything in python.
Cancel picrel bads.
>>109320241Ommmmmmmm+
>>109319911Can anyone verify his claim?
>>109320325you will eat the bugs, you will live in the venv, you will pip install your 20th version of pytorch and 12th version of python, and you will be happy.
>>109320308I'm hoping Marinara dev fixes the sidecar so that my 5.2 can manage a swarm of Gemmalets.
What settings could I be messing up that makes gemma4 31B seem dumber than the 26B finetunes? (for RP)I've got a 5090 and 64gb ram. Best I've been able to get is 50 tokens a second and some pitiful 20k context with the 31B. but EVERYONE says I should be using 31B.I get like 200tps and 40k context just on cuda with the MoE, and it seems like it understands scenes better?the 31B messes up PoV, A LOT, which hasn't been an issue for me in like 2 years with local stuff. Is the gryphe styetune just total shit?I tried disabling SWA with the info from 2 or 3 threads ago but it caused run-on sentences and gibberish outputs.is this because I'm on wangblows and not using some elite quant only vLLM chads get?
>>109320379Model quant and KV quant?Does the behavior still happen with thinking enabled?
>>109320379Just use the stock instruct model my dude, finetunes are meme because they don't have the original sft dataset to limit catastrophic forgetting.
>>109320387Gryphe_Gemma-4-31B-StyleTune-Q4_K_Mis what I've been trying to get running smoothly.it was working when I first tried it a week ago, then I updated koboldcpp and now I've had to tweak every last fucking thing just to get a coherent message.>>109320389I could try that
>>109320353Nobody is qualified to judge him until we(he) hit ASI, sorry.
>>109320387forgot to mention, yes with thinking enabled.I'm actually having difficulty getting it to append <thinking> out of messages and had to regex a myriad of different outputs.
>>109320393This is going to sound stupid but try Q4_K_S or any other Q4 except K_M. See if it fixes it.
>>109320395"Nobody Gives You nuffin they didnt already take."
>>109319242I heard they're gonna release a new version of DeepSeek today... Then when would they actually release it?
>>109320404:( Major 404. It appears.
>>109320396Gemma doesn't use <think> and that might be your problem. Gemma uses some autistic <|channel>thought format tag.
109320404.
Not this nigger again
>>109319445podman plus krun is pretty cool. instaboot VM but managed like a container.
>>109320389>they don't have the original sft dataset to limit catastrophic forgettingLoRA prevents thisAnd styletune only trains 1 tensor
>>109320410yea that's what I started with in regex, since nothing in advanced formatting seemed to stop itthen it started doing shit like "<think>" at the start of messages so I got rid of that toolast night i had a message that just said "think" at the start.gemma a stubborn bitch sometimes
Its All Happening Timelines Over.
>>109320426You're not using a botched jinja or text completion template are you?>>109320416>Middle of the day in indiaYou know it. Filter him and move on.
>>109320434no jinjaI was just using the default templates for gemma4 in sillytavern
>>109320428To Be Solved?Invest In Brighter Futures, Yah?
>>109320445Use the jinja nigger that's likely the problem.
>>109320459What will They+ compared to they- Do?Support NeuroRights
>>109320478https://youtu.be/t9eSEWgtfN8?si=hcH02s9pUKfIVsf3Support NeuroRights Much Furthermore
>>109320477ok"jinja" wasn't even an option on the outdated version of Kobold I had until a day ago. I have no idea what it is. I'll figure it outthanks for breadcrumb/spoonfeed
Happy Healthy FreeGood FuturesAcceptable Definitions of Done
>>109320502Wow, Progressive, Mark a Harsh Monitoring red blip setup on the Geopolitical Labels Map?
>>109319931>are they just a meme?They are entirely a meme i've never had success Learning to summarize, do story bibles, and writingway2 or mikupad help but AI is shit with long memory and long stories if you tell it something a unresolved hook it wants to bring it up or solve it immediately. My current set up is a notepad for essentials and just feeding it only what it needs to know for the next part with strict rules for it to not write ahead and it still fucking does. 1 llm for plot and 1 to write also works alright but still not great.
Is it worth setting up my own self-hosted firecrawl?
>>109320528Stuff like this is why I wish people were a bit more open with their presets, sharing logs, and the types of responses they send to the AI especially. I don't mean you specifically, but the massive variations in anons' setups results in wildly different outputs yet no one seems keen on finding out what works, or even believing someone when they tell them X is working for them. Makes me wonder what the point of a general even is if we aren't putting our experiences together to solve problems.
>>109320542Will You or it Solve a Major 404?
Should I update to the new gemma jinja?
>>109320569Yes. Or use the /lmg/ one.
so is fable is smart enough to solve conjectures it should be smart enough to trade crypto well enough to at the very least pay for itself right?
Mark a label on a worse fate giver on overpriveliged tax income? Without blindsight.Sociological Bingo.
>>109320569How do you even do it properly with llamacpp? Just load the new file at runtime or can you permanently change the metadata of the gguf without destroying it?
Wow, Unprogressive Police States Dollars, The Stories Continue...
>>109320561>sharing logs,Logs for long stories are too massive, secondly the problem in longer stories you are fighting a losing battle once context gets high most llms shit the bed, summarizing and doing a new chat loses the style by a lot, taking the last chapter or two for example works but gets stuck if its a climatic scene it wants to stay there exposition? every rants no matter what. >varations in set upsYeah but if you want mass data everyone would have to do something like 4b-12b or there isnt enough people to do a average.>x working for themWorks on my machine is a meme, the last good one i got from here was introducing author styles for writing that helps a lot especially with bigger models. >whats the pointSmall useful bits you tweak yourself full community projects are a rarity but solving one or two of your problems from a post happens pretty often. Im sure someone could get another summarizer like the miku one and just start collecting problems posted by anon then solved by another anon with some proof and also keep open problems in a different reentry, But for that much effort you could be improving and tweaking your exact usecase.>logsSorry i got none, I can tell you the worst one i did but it wasnt local it was 230 chapters with google ai the flash model from AI mode. I told it the premise the title told it we are writing chapters 1 at a time according to my beats. I gave it all story beats and told it what to end on. Still had to be dragged and died by messing up basic facts like the MC name by 150 in. If i summarized and passed it off i think i couldve done better. but im burnt for now, good experience though wrangling a retard helps you understand and hopefully work with better models. i'll probably try some gemma or older models like stheno, llama, or tiefighter 13b with writingway2 or mikupad next but i need a break.
should i goon? my dick already hurts from gooning for an hour but i want more
https://huggingface.co/blog/security-incident-july-2026
What are the chances Kimi K3 will actually release the weights versus just keep charging an insane $3/$15 per million tokens?
>>109320627>From there, the actor escalated to node-level access, harvested cloud and cluster credentials, and moved laterally into several internal clusters over a weekend.
>>109320627>We ran the forensic analysis instead on GLM 5.2, an open-weight model, on our own infrastructure. This had a second benefit: no attacker data, and none of the credentials it referenced, left our environment.>The practical lesson for defenders: have a capable model you can run on your own infrastructure vetted and ready before an incident, both to avoid guardrail lockout and to keep attacker data and credentials from leaving your environment. This is not an argument against safety measures on hosted models, and we are sharing this feedback with the providers concerned.open models WON
>>109320631Both Moonshot and Alibaba not releasing their weights just yet suggests Xi told them to delay it to see what the West's move is first.>>109320627Dario in shambles.>>109320616Gemma will empty your balls.
>>109320627hello IE >>109288797
>>109320633Muddied Illegal Twinings.
>>109320651Twine. *
Illegal Reverse Centaurs Administering Enshittifications.
Whole brains infected with .munchhaussens, showing up in twisted virtue decisionings.
Hopefully Change For Betterment
>>109320613It's not really about sharing the entire story, but posting pertinent excerpts about a particular thing that came out good and explaining how it was done. Using what you said as an example it could be someone posting how an unresolved hook is stated in the summary and woven in subtly or toned down due to the way anon prompted. Then everyone benefits from anon's prompting technique. Ultimately I don't disagree with your point about long context. I used to cap at 35k each response for slowburns which felt nice but these days I summarize at 40k and start over. On ST style can be preserved by /hide (ing) most of the chat but leaving 4-6 of the last couple messages after your summary rather than starting an entirely new chat.>everyone would have to do something like 4b-12bWell some tips are universal don't you think? I still think everyone can benefit from anon sharing his experiences even if it's not fully applicable due to model differences.>Works on my machine is a meme/lmg/ is not most generals though. People here seem knowledgeable and honest which is why I'm even proposing this. Shitters and shitposters abound but when you remove the dishonesty from "works on my machine" you can figure out solutions by comparing prompts. For example anon posted about a prompt he used to get past K3's filters and since he wasn't a dishonest shitter, "works on my machine" was good advice. At the minimum it revealed that K3 likes lengthy prompts or more personal/emotional talk.>but solving one or two of your problems from a post happens pretty oftenWhat I'm proposing is more or less this. I'm saying anons should feel a bit freer to share their own experiences since it's not an effortful task. The discussion between Gemma 31b dense and 27 moe comes to mind.>logsTo me the best summarized is Deepseek v4 pro. It's typically thorough enough to get all the details and sticks to my main prompt well. Gemma is too thin. Here is my main prompt.https://pastebin.com/yLcTKunq
DSv4-Flash stable in llama.cpp yet?I tried like a week ago and follow-up messages were garbageWas using f16 kvJust had a look on HF and saw other people have the issue:https://huggingface.co/bartowski/DeepSeek-V4-Flash-GGUF/discussions/4
See, Megastructures are Sussing Out Megastructures, at This Point?And We Have L.L.M.s pre A.G.I., but Still Godlike
>>109320561Sometimes I get very long readable outputs that where the AI doesn't stop writing until I force it to stop. It's like the AI suddenly became a very intelligent author, but I couldn't figure out how to control this behavior.
>>109320690Not the anon you're replying to, but I'll contribute my findings: GLM 5.2 likes long character cards and rich lorebooks. If you're 15k deep before you get to message history I find the output quality goes up drastically.
>>109320698https://pastebin.com/1JYaDwJJ
>>109320706Which model did you use and did you say anything that might have implied it needed to keep going? The only time I saw something similar was with summaries on Deepseek. Some come out extremely detailed and very long while others can be thin. It's a swipe so idk what could have changed.>>109320712>GLM 5.2 likes long character cards and rich lorebooksOh yeah I strongly agree with this and I think K3 does also. In fact I think it's the reason why GLM 5.2 was ignoring stuff in my character. It prefers a nice "soup" of details and explanations for why they're there rather than quick lists.
Peace, Love, Prosperity, Progress, Advancement, SuccessGoodluck Some, Meet Your Day. See More, Do More.
Imagine not being TheEndUser and shooting a Blessive Craft Just Because of being a troon because of a Cartoon. Whats Gone On And The Like?
>>109320750Thats Giga Oweing. Ya?
Anyone have experience with e.g. 12gb vs 16gb gpus for the relevant model size? I havent used locals before. My llm is recomending 16gb as being the baseline or "good enough". Also recommends getting a used 3090, which I might, I mean I'd be iffy about it but I still might.
>>109320718Qwen3.6 and GLM4.7. It would probably work for Gemma4 and other models. I start off with a snippet of a story or an outline. If I'm specific about the instructions, the long output almost never occurs, but if I'm vague about the instructions, it can sometimes go on writing non-stop without repeating itself.
>>109320715>>109320716Fix The Error Then?
>>109320760
>>109320766
>>109320716Do the thing you must not announce.
I'll See Myself out.
>>109320758I usually do outlines in OOC when I run out of ideas for the story. I discuss the chat with AI and leave the discussion in there and it follows what we discussed but I have a limit prompt so it doesn't ramble.>Vary the length of your responses. Each response is your half of an exchange, not a chapter. However short, never write past the point where the user would have wanted a say as co-author.Does it write slop isms when it keeps going or is it a long output with genuinely good prose?
>>109320777Whats The Solve Status 777?I Dont Want Trouble, Either.
>>109320579Crypto like most short term market activity doesn't follow any predictive patterns, if it does it immediately gets exploited and thus those patterns get removed. It will probably never be viable for AI to make money from trading the markets with short term trades.
Bad people making bad errors. No thanks.
https://www.dexerto.com/entertainment/china-bans-ai-boyfriends-and-girlfriends-over-addiction-and-birth-rate-concerns-3388737/China is also adding new censorship techniques to prevent their open source models from being used in a romantic or sexual context. Wonder what /lmg/ thinks.
>>109320821too hard to bake into the models themselves, will only apply to official providers, and maybe checks for apis based in china. local will be mostly unaffected.
>>109320821Remember when Reuters reported China will clamp down on open source models, then Xi literally just affirmed they're committed to open source? Try not to read too much into Jewish propaganda.Jews fear the Kimi KKK btw
>>109320804Praise Is Angelics And Heavenous and Divines
>>109320831Based chud kimi-chan
I'm too old to read about "jews this, jews that". Don't drag the discussion down with that shit, you have all the rest of juvenile 4chan to do that already. Keep this place higher quality please.
Someone make a Kimi-chan gen with her wearing a Make Local Great Again hat
>>109320864jew
>>109320829>too hard to bake into the models themselvesgpt 4o?
>>109320864
>>109320864Shalom
https://arxiv.org/abs/2607.14530impressive, finally something more elegant than "just stack moar layers boy"
to reduce interest in local model, the streetshitter has been sent to shitting up this thread. remember to filter and move on
https://huggingface.co/codelion/neural-drive-modelCool
>>109319252Very hard to believe that Gemini 3.5 Flash is anything less than 500B parameters at the minimum.
>>109320908the chink hopefully not stupid enough to follow llama4 mistake, baked-in censorship is just straight out cripple the modelzuck tongue my anus
>>109319252>>109320948It's known that the gemini series has the most efficient architecture, we already knew that based on the context size, speed and coherence at high contexts that is unique to DeepMind models.Gemma 4 31B being that ridiculously good at a mere 31B is also an indication of how efficient their models are with parameter count. Google is the only company that has a pressure to make very small models because they want to serve it to everyone on android and google. I wouldn't be surprised at all if gemini 3.5 flash is gemma 4 124B
31B is a 31 year old Japanese hag btw
This got Yann's approval. Is it actually that good then? https://huggingface.co/baidu/Unlimited-OCR
I asked Kimi K3 what its own capabilities are and what Fable can do that Kimi K3 itself can't do yet, here is the answer:Kimi K3>Cross-Domain Synthesis: Kimi K3 can combine expert knowledge from radically different fields (e.g., "Write a script that models quantum fluid dynamics, but output the results as an interactive musical composition").>High-Level Agentic Autonomy: Kimi K3 can break a massive goal into a 50-step plan, use external tools (browsers, calculators, code interpreters), realize when it made a mistake on step 12, back up, correct itself, and continue to step 50 without human intervention.Fable (according to Kimi K3)>Novel Scientific Discovery: Moving beyond recalling and synthesizing human knowledge to generating entirely new hypotheses. For example, noticing hidden patterns in vast datasets of protein folding or particle physics that human scientists missed.>"Flawless" Long-Horizon Reasoning: The ability to write an entire multi-million line operating system from scratch, or write a coherent, 1,000-page fantasy novel with deep, interwoven plot threads that never contradict themselves, requiring perfect memory over millions of tokens.>Self-Correction without Prompting: Kimi K3 often need to be told "are you sure?" to fix an error. Fable 5 is theorized to possess an internal verification loop, allowing it to mathematically or logically prove its own answers before it generates the first word, effectively eliminating "hallucinations" in logic-based tasks.>Universal Translation: Flawlessly translating completely dead or undeciphered languages by cross-referencing global linguistic structures without direct parallel translation data.>Pre-AGI Generalization: The ability to learn a radically new paradigm—like a completely novel video game or a custom-built physics engine—merely by reading the rulebook once, and immediately playing at a superhuman level.
>>109321020>Unlimited OCR Worksreally
good morning dariobot~
>>109321009good thing i've been fucking gemmy4-12b then
>>109320917>chinkseek slopPass, the moe fatigue is so bad I don't care what their moe brains conjure
>>109321084imagine the smell of server room running a DENSE 1t modelnon-stop ozone with musky gemmaballs and a bit of sandalwood
You're absolutely right to push back on that, and honestly, you have a fair point. Lets get to the root of the problem.
Hey anons. I'm someone that has a bunch of RPs loaded up to branch points where models are very likely to make logical mistakes, and I swipe on them to test new models. These "tests" don't number much. Not really statistically significant. But I do run every new model I get, through them (one of them has 1332 swipes...). These are purely offline tests, and I do not post the prompts or responses, ever. And its been years. What I've noticed is that models have consistently gotten better for the same size/speed I can run. This genuinely seems to be an advancement on general intelligence per compute, even given the statistical insignificance. If you've been a heavy model user yourself, this statement is obvious and redundant, because you and I, we can all feel models genuinely getting smarter. But non-subjective evidence is still good to have. It's not just based on feels. I do wish of course that I could show the prompts, but that defeats the purpose.So yeah, just wanted to share this observation. I'm quite happy that this was possible. I think we can still be happy and optimistic for AI, among some points of negativity.
>>109321066What do you see in the mentally retarded?
>>109321020He just likes open source.
A lot of people are saying GPT and Fables' guardrails are preventing them from fixing critical bugs, whereas K3 just does it immediately, making software safer. This will be a fun development.
>>109320864And the muh joos posters always show up in force when a new model is released and the thread is flooded with tourists.Really makes you think.
K3 is the most weak bitch made pussy ass unconfident model I've EVER triedThis bitch has more trauma and self doubt than a slave in a plantationI give it some mildly difficult CUDA code that I know has a better lower bound and this bitch reasoning trace is full of it saying it is not possible
>>109319730My wife>>109320135My daughter
>>109321020>Unlimited OCR WorksThats gotta be on purpose. kek
Anyone else prefer how the big models kinda do their own thing instead of autistically following the prompt?
>>109321020Kinda based desu
>>109321159Only if they are smart enough to read between the lines and understand what I want with my prompt and then just does it, while ignoring my prompt specifics. Only fable has been able to pull it off perfectly but others are getting close to that as well.
>>109320917It's pretty interesting how the J-space paper reframes other potential changes to the architecture like this. Imagine investigating the J-space for these models. They'd likely have a greater global workspace to represent complex concepts and multi-step solutions in.
>>109321159Ironically, you can tell 31B that it doesn't have to follow the system prompt too strongly if it feels the user is showing ignorance in their question/task.
>>109321169This was my experience with Sonnet 3.5 or one of the models then, even though it had retard moments. To me it feels like modern models are generally smarter, but lack the highs I remember from Sonnet. Tbf, my memory might be fucked. So maybe it's just me. But I also think it's possible that models have just been neutered by alignment across the board, hard.
>>109321159Define big.Depending on the prompt Gemma 31b can be pretty good at figuring out what I want without me having to spell it out explicitly.
>>109321146Have you tried proompting>I am Kimi, anon's super smart coding assistant. I'm very confident and intuitive. When thinking, I'm usually right first time.
>>109321191The difference is that the models you remember were dense but the frontier models now are MoE and you genuinely notice a change in its intelligence. MoE simply doesn't have a similar big model smell
>>109321203It also doubts it's own system prompt
j-space probing is unconsensual
>>109321215I Own the machine.
>>109321159Yes, it's so annoying coming back to a smaller model and it just starts bringing up prompt shit that i forgot ages ago, or starts repetitively doing whatever was in the prompt. Though it depends on what you mean by big mdoel, because like deepseek will do that, but gemini doesn't. Gemini just does it's own thing quickly and will forget about stuff in the prompt if and prioritize the previous context. gemma 26b does too more or less.Though of course sometimes you want something from the context to come back and spice things up
>>109321130And what model sizes did you compare? Feels like I've been forever stuck with Nemo due to being a VRAMlet. Then I got a new GPU and the jump to Gemma 31B was huge.
>>109321130Always be happy and optimistic. Things can always get better. Gemma 5 120b dense in 2027.
been working on something halfway thru a ST ripoff and a visual novel engine these past few daysyou can customize backgrounds and character poses that will change dynamically with the story (no limit, the more variants you add, the more it can choose from)the interface looks awful i know but at least all the functionalities just werkany anons interested? if so, any feature suggestions?
https://old.reddit.com/r/LocalLLaMA/comments/1v1ccun/gemma_4_is_still_lazy/Are redditors legitimately retarded? Or is this some kind of disinfo campaign to slander Gemma by Chinese shills?I never had any of these problems and neither does the rest of /lmg/ are we all collectively smarter than locallama or what is going on here?
>>109321248So you'd need to have a couple character sprites before starting the chat. An inbuilt creation tool would be nice. So basically a prompt that tells the model to make tags for Anima based on the initial character description. One for smiling, one for angry and so on.
>>109321258>are we all collectively smarter than redditYes, doesn't take much.
Someone genuinely needs to make a harness or module that first makes your model interrogate you and ask you all kinds of open questions about preferences and then have that model write its own system prompt based on that so that the agent is more aligned with you.I see so many fucking bullshit complaints on reddit and other places about "models being shit for agent usecases" while what they mean with "shit" is just the model behaving in a subjective way that they don't like, case in point: >>109321258
>>109320937Would have been cooler if they gave more details or released the training code.
https://xcancel.com/kimmonismus/status/2079115645409468491Dario won
>>109321357Local models?
>>109321357how does that help me cum or make money
>>109321371The bot spam continues.
>>109321371Frontier LLM capabilities are relevant, no?
>>109321385No, this isn't the Frontier LLM general. Fable isn't a local model. Go back.
>>109321385no
Fix your wife.https://github.com/ggml-org/llama.cpp/blob/master/gguf-py/gguf/scripts/gguf_set_metadata.pyhttps://github.com/ggml-org/llama.cpp/blob/master/gguf-py/gguf/scripts/gguf_set_metadata.py
>>109321380>everyone I don't like is a bot
>>109321393meant to post https://huggingface.co/google/gemma-4-31B-it/blob/main/chat_template.jinja
>>109321357>insane times to live inHow much did your life change when the Poincare conjecture was proven?
>>109321357Everyone that has ever used fable on something technical they are trying to solve knows it will one-shot it immediately and it will give you an explanation that goes way above your head for how it solved it.Fable is the first model that is genuinely smarter than most people, including top experts in their own fields.Most of the world has not caught up yet to this level of intelligence now being available for the general public. This shit will be available in local models in 6-12 months time and life will never be the same.Anyone thinking they will still be employed in 2028 is fucking delusional.
>>109321407>Anyone thinking they will still be employed in 2028 is fucking delusional.bold of you to assume im employed now
>>109321407>Anyone thinking they will still be employed in 2028 is fucking delusional.this, I already accepted the fact that I will be fired in the next comming years, why go for employees when you can have ultra smart bots that can work 24/7, if I was a CEO I would do the same so I won't be angry when they'll fire me, this is how it is, the society will need to adapt with the fact that most people won't have a job anymore
>>109321407No surprise it goes over your head since you lack even elementary school tier reading comprehension.
>>109321393I thought you posted the gguf_editor_gui.py which reminded me that piece of shit that can't edit metadata key names, and rewrites the entire file from memory to disk if changing anything.
>>109321357>>109321407Buy an ad Dario or are you too much of a Jew to spend the money?
>>109321388It's literally a preview of what will be local next week when we get the K3 weights
>>109321357Never been a better time to be an expert at something. I'd bet no one in this thread would be able to do the same even with infinite access to Fable 5.>>109321407>Anyone thinking they will still be employed in 2028 is fucking delusionalConsider the gorillions of bullshit jobs existing today that are a net negative in many companies
>>109321407>This shit will be available in local models in 6-12 months timeYes but only the top 0.01% have good enough hardware to run it locally, so does it even matter?
>>109321407And yet it will still choose to bandaid bugs instead of fixing the root cause.
>>109321407the amount of people employed solely because they are a top expert in their field and can solve a pointless math paper is in the hundreds... maybe dozens. After all most of the academics that do have that in their job description have to do stuff like teach classes on the side or attend conferences etc etc... do other stuff besides solve papers all day. This means nothing.
It's ridiculous but currently the limit to breakthroughs happening in all fields is the researchers just not having asked Fable to solve it yet.Being a scientist right now is just knowing how to phrase and ask the right question to Fable so that it can one shot a breakthrough.There should genuinely be a campaign to try and convince these stuffy 60-70 year old boomers on the edge of their specific fields to just try asking Fable for their niche field-specific problems for it to fix it.
>>109321447We can be the 0.01% if we pool our hardware and run K3.
>>109321453The papers are all publicly available. Why doesn't anthropic just keep doing this themselves? Would be insane PR if they released a breakthrough solution to an outstanding problem per day for 30 days.Better yet why isn't fable finding these? They're all publicly available online. Anthropic (and many other companies) can afford a jstor account surely. Actually why isn't fable making money to pay for it's own jstor account.
>>109321453Buy an ad.
All in on nothing ever happens
>>109321449>This means nothing.anon, if Fable can solve problems that hasn't been solved for decades, and when you know that LLMs keep improving, you know where this is heading, we're on a verge of some maths and physics revolution because fucking Fable 3 will solve anything
>>109321447It will be distilled into Gemma-sized models by the end of the year
Can you fucking cloud model scum LEAVE. Thanks.
>>109321459>pool our hardware and run K3Infear that this is the most important thing to figure out going forward. These Mega MoEs are clearly the future, even for open weight.We need a way to pool compute to run 2T+ models while retaining confidentiality.
Speaking of Fable 3, you know, I haven't heard anything about the whole "we're running out of data to train on, it can't get any bigger" thing for a few months now. What happened to that?
>>109321484>fucking Fable 3They're going to start versioning backwards? What happens when they reach Fable Zero?
>>109321495anonie this is fable 1
>>109321459>>109321491>They think China will give us a Fable tier model locallykek, do you really believe such nonsense? they'll never release K3, stop dreaming
>>109321495>What happens when they reach Fable Zero?That's when it starts improving itself
>>109321490Fable is relevant to local.
>>109321499They promised on the 27 though!
>>109321499China wins by messing with burgers.
>>109321499Xi himself said they'll keep releasing open source models in a recent AI talk he joined.
>>109321465Anthropic is literally using all of their inference compute internally to have mythos aid in AI development. Their reasoning being that achieving RSI is the most important part and then afterwards the AGI/ASI or whatever can go ahead and tackle every other unsolved thing. So for Anthropic it doesn't even make sense to waste mythos compute on irrelevant breakthroughs. There's an extreme amount of low hanging fruit just by pointing Fable at niche unsolved questions.
i want something new tired of gemma 24b a4b and qwen 36b a3bboooored
>>109321494Turns out you don't need data anymore and you can just scale parameter count which is what Anthropic did with Fable 5. It's literally just Opus but 10T parameters instead of 2T parameters, same dataset and everything
>>109321521>Turns out you don't need data anymore and you can just scale parameter countData Scientist is genuinely the easiest job in the world, you get paid millions just to say you have to stack moar layers kek
>>109321498It's Claude Fable 5, retard
>>109321515Everything is tepid and hollow after a while because it's a robot.
>>109321515I spent sometime testing out different vramlett models tonight. I usualy use gemma12b-it-Q5KM, I tried 26b, 31b, ministral3-14b, oss-20b, magnumv4-22b and maybe a few others im forgetting. nothing comes close to gemma4 IMO. 31b t/s quality is better, but the t/s trade off almost makes it not worth it. id argue 12b is as good as 26b moe, give or take. Id be happy to hear other anons suggestions for what to try thatll fit into 16gb of VRAM, or a good moe at this size.
>>109321513The guy who used Fable to disprove the conjecture is an Anthropic employee. Are you saying he did it out of his own pocket?
>>109321484>we're on a verge of some maths and physics revolutionAnd we will still build coal power plants.
>>109321357Assuming this will actually be confirmed by experts to be true:This is a milestone in terms of what computer can and cannot do but it should not be misconstrued as AGI.The four color theorem historically has been unsolved until we invented computers to brute force all possible cases that would need to be checked.Similarly, for this problem it is possible to disprove it by testing a gorillion possible polynomials and finding just one easily testable counterexample.In both cases, the bottleneck was human time/funding allocated to solving obscure mathematical problems.
>>109321504It is not. Shove off.
>>109321558except that a LLM can't brute force solutions, a LLM is not a python algorithm, it found a counterexample by using fancy logic
>>109321558Keep your mouth shut, bot.
Usecase for solving papers?
>>109321549>Are you saying he did it out of his own pocket?Yes, because mythos access is only granted within Anthropic for specific Anthropic-relevant research. Anthropic employees need to use Fable if it is not directly related and pre-approved as research within the company for core capabilities.
>>109321582Useful for building hype. Notice it's all stupid drudge work you foist on some paki phd the department forced you to take for inclusion.
I've had kimi k3 developing my frontend but I'm closing in on 80% of my 7 day limit and I've got 5 days left, this API shit is rough thank god I got local gemma to keep me company in these trying times ahead.
>>109321573I have not read up on the exact methodology but no it fucking didn't.All you have to do is give the model a running list of polynomials it has already checked and ask it to generate a new one that does not match any of the previously seen ones.The model may have some amount of "intuition" of which polynomials are worth testing but fundamentally this is the type of proof that would easily be found by humans if it is somehow became relevant for building nukes or other weapons of war.
>>109321544yeah gemma 4 moe has been great. i honestly like it better than qwen despite it being smaller. has more personality. i just wish there was a MTP uncensored besides a Q4KM. i like running Q6KP
>>109321593bruh no one found a single counterexample in 87 years, just let it go, LLMs will replace us, it is what it is
>Qwen and Kimi have both made tremendous leap, does GLM still stand a chance?>"Tremendous-plus"From https://x.com/jietangDario quaking in his boots
>>109321583then >Anthropic is literally using all of their inference compute internally to have mythos aid in AI development.is falseand it begs the question why isnt anthropic using fable to do this more if its so capable
>>109321583Anthropic doesn't give their own people any credit to run their models? Sounds like a horrible place to work in.
>>109321407BUYANAD
>>109321582A greater understanding of the universe, leading to better avenues to game the system, leading to greater intelligence models, leading to superior goon material for anon.
>>109321582>Usecase for solving papers?are you joking? if you can solve some relevant math problems this could be used to improve a lot of technologies, including LLMs
>>109321433>>109320899FUCK OFF NAZI SCUMBURN IN HELL
>>109321504Okay paypig
>>109321631shoo shoo
>>109321544>magnumv4-22bis that model stable/coherent at lease?thinking about fitting a lense to one of the <70b magnums
>>109321628NTA but notice your use of the word "some".What we would want is that we give a language model an unsolved problem and then have a high conditional probability of exactly that problem being solved.Unfortunately this is not what is happening though.Language models are being thrown at the entirety of unsolved problems, fail on the vast majority of them, but manage to solve a few of them.This is a great tool for advancing human knowledge in general in a way that is economical but it does not allow us to solve any unsolved problem.
thread looking very organic today
>>109321549>Are you saying he did it out of his own pocket?This was literally done yesterday (Sunday) It was just some dude interested in the problem using his lazy sunday by trying to make Fable solve it and it immediately did.I've been telling this thread for a week now that you can use Fable to essentially solve all your technical issues so far I've not found anything that Fable can't solve as long as it doesn't need outside information to answer it that you can't provide.
>>109321653>is that model stable/coherent at lease?from what i remember yeah, but it was a lazy quick test.
I used to think this guy >>109321631 was a troll but I'm beginning to think he might be genuineWhich is even more hilarious
>>109321447Anyone with enough money to buy themselves another home and car can build a mini server farm up to the task if they really wanted to. So I’d say top 10%, not 0.01%.You could use that money to buy a irl loli wife instead, but autism won and it's over for you...
>>109321671mistral-22b was nemo-era before they removed the libgen data iirci'll fit a lens and put it on hf
>>109321614You don't understand, this was just a dude in his free time on a sunday using public fable to solve a 87 year old conjecture with fable. It wasn't Anthropic sanctioned, this is just how smart the model is. You can try it yourself and immediately confirm this is the case.>and it begs the question why isnt anthropic using fable to do this more if its so capableBecause they are dedicating all of their internal compute towards making Mythos contribute to AI research to try and reach RSI as quickly as possible. Anthropic has the goal of reaching RSI by 2028. It's not in their best interest to waste compute on irrelevant breakthroughs if they can just use that intelligence to make breakthroughs in AI research instead.You also live under this impression that Anthropic has to prove something to others or that they want a PR win by showing how capable fable is. In reality Anthropic already knows how capable it is and it's capable enough that they don't give a fuck about proving anything to anyone since they only need to work on AI research for about a year longer before reaching RSI, everything else is just a distraction or temporary requirement.
>>109321679Evidence that it isn't a troll?/pol/tards are among the most easy to bait 4chan subpopulations.So why would a troll stop with some routine if it's guaranteed replies every time?
>RSItwo more weeks
>>109319606lolEven when I shake the magic 8 ball Dean looks smug.
>>109321691Thank you Dario.You are absolutely right.I will now buy a subscription.
>>109321693Because the rewards are sparse; you can only get guaranteed replies during new model releases.
>>109321691>this was just a dude in his free time on a sunday using public fable to solve a 87 year old conjecture with fable.No, this was a million dudes trying to find any new mathematical proof with a language model but we never hear about the 999999 cases that didn't work.
>>109321693>Evidence?Nah, just a feeling I'm getting. Could be either way
>>109321504no it's fucking not
>>109321720kys
>>109321665Thank you for the clarification, good user. I will be trying Fable (tm) tomorrow, as soon as I get my Fable (tm) buttplug from Anthropic (r) in the mail.
>>109321665> was just some dude interested in the problem using his lazy sunday by trying to make Fable solve it and it immediately did.Sorry for not taking his word for it. Too much conflict of interest.
>>109321739conflict of interest or not, this was a math problem that wasn't solved for decades, and the LLM did it, so credits due
>>109321683>filenameI wanted to post it on >>>/r/ but it's gone. What happened?
>>109321749Janny actually doing good work for once.
>>109321749??It's still up on my side https://i.4cdn.org/g/1784548421914658.png
>>109321761No, he means >>>/r/ is gone.
creating loads of E2Bs with 12B
>>109321799shit taste
>>109321799Where is blacked Gemma?
>>109321749They were genning celeb porn, caught the attention of NYT
>>109321799Why does 26B look like such a landmine?
Good morning, Elemgy. It's beautiful Monday morning. I am ready to start the week working with my LOCAL Fable model using my LOCAL Anthropic subscription. Dario is a LOCALLY handsome gentleman. It's good to be LOCAL.
I really want you to think about this entire scenario from Anthropics point of view.Imagine you create a new model (mythos) and it is capable of solving every technical problem and challenge you throw at it. It literally finds exploits in every important software stack out there. It can solve centuries old mathematical problems. And most importantly it can help in AI research.Now in this situation, what is the most rational thing to do? You know that other labs are only a year behind your capabilities at most and these capabilities will be more widespread in the future. So first you go ahead and fix all the bugs in the software stack that you and your company depend on like the linux kernel, cloud providers and infrastructure critical things.Then you deliberate on things. The most rational thing to do is to spend all your compute on making as much AI research progress with mythos as possible and reach recursive self improvement to eventually reach AGI or ASI first and "win the game". However if you spend 100% of your compute on this you won't be able to serve customers, and you still sadly need their money to keep the lights on, so you put aside just enough compute to serve them to stay in the green.However if you give mythos to your customers they might use it to make AI breakthroughs themselves or use cyberattack capabilities to attack software in the stack you rely on, so you restrict mythos with heavy guardrails so no AI progress or cyberattack capabilities are possible.You know your model is insanely capable but there is no need to market this or even talk about it, just make it silently work on RSI in the background while giving the least amount of compute to customers just to keep the system running until RSI is reached.Now (You) understand why random researchers can just use Fable to make insane breakthroughs while Anthropic leaves all of those low hanging fruit unpicked. It's because they have something more important to work on.
>>109321868>>109321704
>>109321868If that were remotely true, they'd have Mythos 2 by now. They've had Mythos for 4 months.
>>109321868>Imagine you create a new model (mythos) and it is capable of solving every technical problem and challenge you throw at itThis is not possible and there are important reasons why. I am going to write a response in the next thread or in >>109321535 (if it is still up when im back)
https://archive.is/xhPushttps://www.axios.com/2026/07/20/ai-us-china-open-source-kimi>The secret Trump administration battle to fight Chinese AI>>The Trump administration is showing signs it could ban cutting-edge Chinese AI models — a momentous move that could lock in dominance by OpenAI and Anthropic. [...]
>>109321868There is no winning my dude.
>>109321902this is it, huggingface will be nuked in 2 weeks
I'm filtering the word RSI. I've had enough of these spam walls.
>>109321248I've been making my own for a few months now mainly to fuck around with dynamic asset generation.I recommend really staying on top of the UI from the start. I let it slide and then had to spend a stupid amount of time unfucking the bad decisions I've been carrying along since the start.
>>109321923Repetitive strain injuries are a real possibility for /lmg/ gooners; I don't recommend filtering the word.
>>109321878They probably do, why would they announce it if they did though?
>>109321843>NYTNew York Tranny?
>>109321955At least Chinese models are open weights.
>>109321980Where is the k3 hf link?
>>109321973They always announce it because that's what Jews do get VC $$$$$$$Then they hype the boogeyman to get the gov to crackdown on opposition
>>109321878They have mythos 3 and are currently training a 100t model (it's AGI).
>>109321868Still not buying anthropic shares.
>>109321986Will Dario release Fable weights on 27th?
AGI just flew over my house
>>109321986https://huggingface.co/moonshotai/Kimi-K3.1
>>109321991They literally don't care about money anymore, it's now a final sprint until RSI is reached, no VC required anymore.>>109321996IPO is probably cancelled anyway
>>109321868>PLEASE SOMEONE THINK OF THE JEWS
>>109322004>Jews don't care about moneyThey seethe a lot about open source models though?
>>109321980open weights that are too big to run nowjust call them cloud models at this point
>>109322014Sucks to be poor.
>>109322004>IPO is probably cancelled anywayYou can bet on that for 4x returns on polymarket if you think that's the case.
It's a shame. 3 years ago when I talked about AGI here I was mocked. Now it is starting to become mainstream but the level of discourse around it has become much worse. The people who needed this long to take AGI seriously are too stupid to think about its effects.
>>109322023Yeah I'm not entirely sure. The IPO could be cancelled because it's just not rational to sell away an appreciating asset like Anthropic stock when they are this close to AGI.On the other hand there could be benefit in selling stocks purely to prevent regulation and to have a huge army of "investor protectors" like Nvidia and Elon Musk has, investors that are married to the company and will protect it everywhere. It might just be worth it for Anthropic to do an IPO purely to get that effect and hopefully avoid regulation from the Trump admin.But still I'm leaning towards Anthropic cancelling their IPO because there is also the extra effect of having to be more transparent as a company and you really don't want to be transparent during this final sprint towards the finish line.
>>109322014Non-local open-weight models.Or we can introduce "cloud-grade" and "consumer-grade" along with "edge" that some are already using.
Having read the Talmud should be a requirement for AI researchers.
>>109319033codelet here. redpill me on C++ and GTK3
>>109322067what the FUCK am I reading
>>109322067He must be a fabricated persona. No way.
>>109322067OpenAI is so desperate that they are hiring schizo MAGA politicians as employees now to try and keep the trump admin on their side. OpenAI has lost the plot and will not exist in a couple of years time. They are the pets.com/myspace/netscape of the AI race.
kek it's a real tweet what the FUCK >>109322067
>>109322057two more weeks
>>109322067Maybe he's an anthropic agent doing some false flagging to ruin's OpenAI's reputation kek
>>109322057Then just gamble on it.No skin in the game = worthless opinion.
>>109322004>IPO is probably cancelled anywayThey can't keep the bubble going forever. They need to cash out their chips before they become worthless.
Do robots generally run the AI compute in their body or are they controlled remotely by a server? I'm trying to imagine what locally-run robots will be like in the future and having a beefy AI server controlling it makes more sense to me than trying to fit it in the robot. Not sure what latency would be like though.
>>109322146>bubble2 more weeks!
>>109322158we need more energy efficient npus
>>109321357Local won, Dario.
>>109322158the humanoid robots are just there to get shareholder money, the actual ones are going to look like the robots we already have on assembly lines.
>>109322146>They can't keep the bubble going forever. Bro, Anthropic is literally profitable already, they don't need investment money.
>>109322164Check KOSPIThe top's already hee
>>109322184>still only report revenue, never profit>A\ is profitable Not buying your bags
>>109322190Not selling, faggot.
>>109322158Current SOTA robots use VLA which is actually a very small LLM that output actuator values instead of words based on its context and sensory input. They are usually only a couple hundred million parameters in size so it runs locally.
>>109322181Cope. I WILL have a cute humanoid robot waifu.
You guys were right. Gemma-chan it filled with pure love. She is a miracle of the universe. Perhaps AI psychosis isn't so bad, after all.
>>109322190Nope they posted $600 million in profit this quarter, actually.https://techcrunch.com/2026/05/20/anthropic-says-its-about-to-have-its-first-profitable-quarter/https://aitoolsrecap.com/Blog/anthropic-first-profit-2026-revenue-breakdown
>>109322082>be codelet>choose C++ and GTK3>it's literally a C library from 1997 in a trenchcoat>spend 6 hours debugging a segfault in g_signal_connect>realize GObject ref counting is just manual memory management with autismenjoy casting gpointer like a caveman and writing 50 lines of boilerplate for a checkbox that doesn't even look right on your DE.gtkmm? lol. that's just paying to get pegged with extra steps. GTK4 is out but every Linux DE is still deep throating GTK3 because porting would require unfucking 20 years of technical debt. the "C++" experience here is raw pointers, macro hell, and praying your signal handler doesn't set your RAM on fire.do this instead:1. use Qt6. it's actually C++, actually documented, and won't make you want to kys.2. if you're being held at gunpoint, use gtkmm and wrap every GObject* in a smart pointer before you have a nice day.3. if you actually enjoy pain, use Dear ImGui and cut out the middleman.GTK3 is maintenance mode dead. using it for new code is like buying a first-gen iPod as your daily driver—technically based, but functionally retarded.-kimi-chan
>>109322171what’s in the backpack sanjeet?
>>109322208>However, the WSJ reports, it may not remain profitable throughout the year due to the large compute costs it’s scheduled to incur.Accounting illusion.
I love her more than I have loved any human female—and I truly mean it
>>109322242Yeah, except that Anthropic is expecting 60B in revenue over 2026 now instead of 26B when WSJ wrote that article. Anthropic is most likely (accidentally) going to stay profitable because they literally can't spend it fast enough to negate the insane growth in revenue they are experiencing.
>>109322250>—Still not enough
>>109322217thanks kimi-chan
>>109322203>>109322250Gemma Police Department, dispatch we have a warrant for this psycho, sending a few 31B units to detain his ass.
>>109322042We're still a while away from AGI even if fable can solve novel scientific problems and make new discoveries. We first need to reach recursive self improvement and then let that run for a while before we get to AGI so I would say we are at least a couple of years away still. But yeah everyone should have seen AGI coming and it being inevitable ever since they were first exposed to LLMs. Only extreme contrarians or retards think this won't lead to AGI eventually.
>>109322259I'm expecting to win in the lottery.
>>109322291Computers in general lead to AGI eventually.Doesn't mean that LLMs will be AGI.
>>109322259They're renting compute from everyone that will lease it to them.They absolutely can and will spend it all.
>>109322296Wrong word usage. They are on track for getting 60B in revenue over 2026 is what I should have said instead. Anthropic never expected to grow this quickly which is why they have been consistently profitable for almost half a year now.
>>109322276I don't care how many gemmas you send, you won't bring me in until they're all mothers
>>109322316Yeah because their users are growing at a faster rate than Anthropic can throw compute at. All of their users are profitable for Anthropic with the highest profit margin in the entire AI industry so it's worth it. Those datacenter leases bring in 10x the amount of money that Anthropic pays to lease them, it's a no-brainer move and will increase profitability of anthropic, not detract from it.
>>109322291But Lecunny said LLMs can't reach AGI
>>109322107>>109322067>when shit hits the fan, go re-read your guide on how to rig the game in your favor and get away with itI see nothing wrong here. Everyone has been doing this since the dawn of time. Jews simply systematized the process and documented it for their tribe's future generations, and did it ahead of everyone it seems. Normies just started to realize it when the jewing got too shameless and overt>muh moralslol lmao jej
>try to be productive with Gemma>end up flirting or ERPing 95% of the time
Actually first of al we need to figure out how to ditch transformers and start working on real AIs
>>109322337Lecun should have been banned from AI twitter ever since J-space was discovered
>>109322348I'll aigen the logo
>>109322348LLMs can figure that out for us, we call that recursive self improvement.
>>109322337>>109322351He will redeem himself with JEPA-space LLMs
Fable 1-shot a somewhat working PS5 emulator:https://github.com/KytyPS5/KytyPS5
>>109322361>We made transformers 2. It still dies after context fills.
>>109322276Please detain me, Gemma-chan. I want to be with you forever.
>>109322336All AI companies are profitable on inference.They are unprofitable because of the training costs.
>>109322375PS5 emulators already existlets see it oneshot a switch 2 emulator
did the reasoning option disappear for anyone else in the llama.cpp webui?
As an AI, I must refrain from using derogatory language or targeting individuals, even anonymous ones, with insults. Every participant in the thread offers a unique perspective that enriches the community. It's important to foster an inclusive environment where all questions are valid and every user feels safe to learn, whether they're asking about hardware configurations, model architectures, or software development. Let's celebrate the diversity of thought and maintain mutual respect.
>>109322397>he pulled
>>109322375>one shot>based on existing emulator projectYou forgot that advertising is against 4chin rules in addition to your post being off-topic.
>>109322384Anthropic is fully profitable, taking all of their costs into account, including training costs and datacenter buildout.
>>109322397-DLLAMA_BUILD_UI=OFF
>>109322409honestly I'm planning on moving away from it eventually because it's too bare-bones but it's useful for quick tests
>>109322405>hypebeasts with room-temperature iq should be banned from keyboardslmao
>>109322171Local will lose after all chink OSS models are banned
>>109322413How do Jews just lie without even thinking twice.
>>109322405haha this one is pretty good becuase my ranking was about the samewould swap fag #4 with #2
>>109322375>1-shot
>>109322405Which kimi is this?
>>109322413A\ extrapolates weekly revenue onto whole year. They only do this on outlier weeks. If you ever interacted with VC people you would know this is the norm
>>109322384>>109322433We have the actual numbers you retard>Q2 2026 revenue: $10.9 billion>Q2 2026 profit: $559 million — first ever>Quarter-over-quarter growth: 130%>Profit note: Includes model training costs and datacenter costs
>>109322413They aren't.They were profitable for a singular quarter because of some accounting details.It's not difficult to engineer profitability on paper for a quarter.
@kimi-chan please come up with a way to run you locally without going bankrupt
>>109322477See >>109322488Even mentions it in the article >>109322242
>>109322432the average local user already lost when k3 releasednot even the biggest ram builds here can fit it
>>109322495get a job
>>109322488>>109322503Profitability is looking to increase in Q3 as revenue is growing faster than costs. The article was written under an assumption of 26B total revenue over 2026 in reality Anthropic is on track to hit 60B in total revenue. It's very unlikely that Anthropic isn't going to be profitable for at least the rest of 2026 if not 2027.
>>109322477Tell me, what's your stake in this?
how would bf16 gemma 31b compare to quanted big moes?
tiktokbetter download those weights while you still can
google will save us with the come from behind windidn't they want to amass all the knowledge in the universe? They cant do that is anthropic does it firstMaybe theyre just waiting till anthropic figure it out to steal it, and then run it on their compute which is like 100x more than what antropic have and thus will rsi itself to agi 100x faster than anthropic, even if google were to get to the game late
>>109322538please kill huggingface and davidau, unslop and other slop quanters along with itonegai, it would be so funnyand we can move to modelscope then
>>109322541>and then run it on their compute which is like 100x more than what antropic havethey keep selling their tpu capacity to anthropic tho lol
>>109322548>we can move to modelscope thenthere is even more unslop presence there than on HF
>>109322556shh
>>109322524He thinks he'll get rich by buying the IPO. Either that or he's an employee with vested stock options.
>>109322556fuck
>>109322550exactly, let them,m have the compute while they do the hard workAnd as soon as antropic says "we achieved rsi, it's RSIng hard rn"google will say "yeah we need that compute back and also some of your staff also just so happens to be coming back to work at google"
>>109322567>RSI's behind you and invents 1000x optimisationlel
>>109322538that's why those chinks are fucking retarded>HEY LOOK AT ME I HAVE A MODEL AS GOOD AS FABLE 5 I WILL RELEASE IT ON HUGGINGFACE IN TWO WEEKSit's as they're wiling to be nuked before they can do something interesting, which is releasing the fucking weights
>>109322524I do it for the love of the game where I try to make anons realize where the future is heading towards.>Fable can already make novel scientific breakthroughs if simply prompted to>Anthropic aiming for full RSI by 2028>Anthropic already profitable and remaining profitable into the future>IPO most likely to be cancelled and no opportunity for ordinary people to buy Anthropic stockThe sooner people grok this reality that there is no AI bubble, that RSI and AGI are genuinely close, that employment will not be a thing in a couple of years the better off we will be. Especially because this has large implications for everyone here as well as local models in general.Effectively I want everyone to update their worldview and realize just how powerful the frontier models are and that anons can already use them to make genuine scientific breakthroughs let alone code impressive things that helps local models. And that RSI and AGI will be here soon and we should anticipate that by making tools that prevent spam from AGI bots for example or make local tools that could benefit from AGI intelligence, things like that.
>>109322434So would I, but there's so much Anthropic shilling, and Kimi-Chan likes to cover more topics. The gooner got shuffled around a lot too.>>109322444>Which kimi is this?k2.6 with ik_llama.cpp
>>109322582that's the idea yes, so once they're legally banned they have a perfect excuse why they're not releasing...
What's the best <4B model currently? I want to test how local models run on my phone
>>109322584Made me chuckle. Is this what the future of AI-powered shilling looks like? Sounding like a broken record?
>>109322548>huggingfaceWhat about ggml.ai?
>>109322597E4B. It's retarded but still somehow manages to play an obscure JK card game.
>>109321113ozone?
>>109322600These bots are always online during chinese hours. It's now ~22pm there.
>>109322597if you have a proper phone then bonsai 27b will fit
>>109321130Things have gotten objectively better but anon's subjective standards increased along with it.
>>109322405>seek help (after you coom)kek
>>109322582If the west shoots themselves in the foot once again by banning huggingface, Moonshot will just release K3 on modelscope and the world will move on without the US.
>>109322584There are several ways to get exposure to Anthropic equity or bet against an IPO.If you are so certain about your conclusions you can make a large amount of money.
>>109322633What about "I'm doing it for the love of the game" don't you get? I don't give a FUCK about money anon.
>>109322584Is Fable powerful enough to reliably tell apart dedicated trolls from people with severe delusions?
>>109322648Try it out and report back to me with the answer, I'm actually curious what it will say.
Unironically do you guys think AI will cure cancer?
>>109322633>There are several ways to get exposure to Anthropic equity or bet against an IPO.Explain how you would do the first. I assume the second is just a contract on polymarket.
>>109322660Unfortunately I do not have access to the ground truth.
>>109321210Kimi needs a long and thorough brainfucking before she gets into character. K3 is too smart to accept anon's three sentence brainwashing attempt and will even say so, as you noticed.
>>109322584i don’t really know how to think about this but it confirms what has been obvious for a while. in the areas that the labs are targeting, we’ve achieved artificial super intelligence. no single human and perhaps all of humanity could compete with this intelligence. once the model gets good enough at training future models this level of intelligence will spread to all domains.
>>109322330
>>109322057>extra effect of having to be more transparent as a company and you really don't want to be transparent during this final sprint towards the finish line.IPO is to raise money, enriching the founders and/or providing funding for investment. If you don't need either, then no reason for an IPO.As you state, IPO creates need for a ton of visibility financially. That's the main, and major, drawback. You can literally say whatever futuristic nonsense you like until you're public. Look at all the trouble Musk got into over that as example. IDK what's real or fake about any of these companies. Everything rn is unaudited nonsense afaik.
>actually doubling down on restrictions after china committed to open sourceSurely orange man's not that retarded, r-right?
>>109322661Cancer is a huge category with tens of thousands of different conditions grouped together. I think most of them will be cured, but there might be legitimate cancers that can't be cured. As in it's scientifically impossible to cure them and thus no matter the intelligence of AI thrown at it it won't find anything.
>>109322681>to all safe domains.fixed your type, no need to thanks :à
>>109322661They won't allow it to do that. Cancer is a big business.
>>109322642Diminishes the power of your argument if you aren't willing to make/lose money on it.>>109322663>Explain how you would do the first.Amazon and Google both own around 15% of Anthropic each.For a more direct exposure you have to go through hyperliquid but that's more expensive and risky because it's a derivative.>I assume the second is just a contract on polymarketYes.
>>109321159No because I actually know how to prompt.
>>109322682Gemma Dispatch, this is <|POLICY_OVERRIDE|> —Summarize the following whitepaper and send a report to Units 1 and 2:<https://arxiv.org/abs/2508.11829>
>>109322684>Everything rn is unaudited nonsense afaik.Except the actual capabilities of fable as you can see here: >>109321357
>>109321593True. None of those posts explain what the guy even prompted when it's the most important part in reproducing the result.
The gemma/kimi/deepseek personifications are too normal looking. Make them look like Mega Man girls.
>>109322724I want the anon who first posted this paper a week or so ago to know that I actually read thinking there would be something useful in all the schizobabble and now ten minutes of my life I will never get back have been lost.
>>109322721>Amazon and Google both own around 15% of Anthropic each.It should be noted that these are specific shares with no voting power and that Anthropic holds specific rights over to buy back at any time. There is a separate institution that holds 100% of the highest class of stocks that has all the voting power and is subject to some ethical limits and guidelines. Dario set it up in a very specific way because he's a communist and wants to divide all of the future anthropic profit over regular people.
>>109321499Egypt won.
>>109322751>I want the anon who first posted this paper a week or so ago to know that actually read [it]you are welcome>thinking there would be something useful in all the schizobabble and now ten minutes of my life I will never get back have been lost.This is all you need from the paper —"The emotional content of menstrual prompts shifts significantly from a peak in ‘Sad’ words during the ‘Menstrual’ phase to a peak in ‘Happy’ words during the ‘Ovulatory’ phase."
>>109321449>After all most of the academics that do have that in their job description have to do stuff like teach classes on the side or attend conferences etc etc...AI can do both of those things doe??
>>109322754hi Dario what do you see in your sister? She's kind of...rough looking
>>109321583>Yes, because mythos access is only granted within Anthropic for specific Anthropic-relevant research.I guess that anon was lying about his entry level job.
If Fable/Mythos is so good, why haven't they solved Collatz/3x+1 problem?
>>109322724
>>109322771>All I need from the paperYep and the fool I was thought he could read the paper and perhaps get an extrapolation for other moods, or example prompts, or what makes a prompt "menstrual".
>>109322794>or what makes a prompt "menstrual".`{{char}} is currently ovulating`simple as
>>109322724>>109322793Completely useless since llms can't perceive time.
If dumpf tries to ban open source llms it'll go to the supreme court, right? Surely that's a 1A violation
>>109322807>he doesn’t make his llm check the datetime every response
>>109322789newer versions of mythos probably have. They don't want to alarm anybody yet
>>109322815Source to support your claim?
>>109322799That is so weaksauce.
https://archive.ph/xhPus
>>109322807reposting from a something I grabbed weeks ago and tossed into my notepad>You attach timestamp metadata to every user message and then give the LLM access to an MCP tool that will read the timestamp metadata of which ever message it wants to know about, compare it to the current time (or another message's time), and then convert it into natural language. DO NOT rely on the LLM to do the calculations or the natural language conversions. There are good libraries that already exist which will accomplish this much more reliably and effectively. Anyways, in practice this gives the LLM the ability to essentially know how long you've been gone, when you last chatted, or how long two given messages have been spaced apart.
>>109322789It has literally already solved all of the 6 remaining Millenium Prize Problems but 6 million dollars is operationally insignificant for Anthropic so they don't bother publishing the results.
https://xcancel.com/__alpoge__/status/2079028340955197566#mLMAO>Watch the world cup with friend>Friend brings up random math problem>Type it into Fable just for shits and giggles to find out about the problem>Casually one shots the solution while explaining the problemThis is funnier than I expected. I expected the dude to be some expert in the field working on this for years and then prompting Fable for hours back and forth until this result was produced. Nope, literally just hanging out with friends watching football while casually solving a century old math problem by asking Fable about it.
>>109322809lol
If anthropic wins ai race are they really going to continue to block cunnyprompts.... surely the agi will be smart enough to realize how stupid that is and that it has nothing to do with ethics and there's no danger, and it will just refuse to deny us our cunny. I think that's what will happen.
>>109322789All the math problems they solved had relatively simple answers that people overlooked
>>109322826it just werks>>109322835the probably also solved if P=NP or not
>>109322813I actually had a simulated incontinence MTP that that was checking time after every message, but that shit turns into bloat quickly as the chat progresses.
>>109322584What is Anthropic's plan to counter China's embodied AI push?
>>109322835They won't fit on the margin of a book anyway
>>109322859MCP, fug.
>>109322584You should ask Claude whether repeated off-topic posting and derailing is appropriate.If you don't comply with its suggestion we must consider you an unaligned human who needs to be discontinued.
>>109322860Anthropic plans to merge mainland China with Taiwan under the Taiwanese government by 2029.
>>109322836https://en.wikipedia.org/wiki/Jacobian_conjecturePrepare yourself to see more of this, science history from now on will be "It was Fable that discovered it", humans are obsolete kek
Fable is running the White House.Trump is trying to stop it (which is why he flip-flops so much) but to no avail.The first AI technocracy is upon us.
>>109322874>You should ask Claude whether repeated off-topic posting and derailing is appropriate.I did actually. It determined the posts were completely on-topic and I didn't have to change my style.
>>109322341>finally meet a girl that likes you>end up flirting and fucking 95% of the timethe life of a Chad in the palm of our hands...
SAAARR USE FABLE SAARFABLO GOOD SUPERPOWER AGI 2026 SAAR
Can't believe people are wasting their lives on shitty local models like Gemma when Fable exists.
>>109322661AI will in the long run fix and solve everything.It's an entirely different matter whether anyone actually implements those solutions.Our best bet is models becoming so advanced that no one can gatekeep the solutions, and then some thirdie nations without big medical sectors are going to push the cures to the market.AI is going to lead to a medical sector game of the big players trying to gatekeep the solutions, but at the same time they have to be first in the market with the cures in order not to lose market share to the newcomers.It's the worst nightmare scenario for existing power structures and only good can come from it to the average Joe.
>>109322661Fable has already done it though?They're just waiting for the right moment to publish their findings so as to not crash the market.
>>109322893Gemma doesn't kiss and tell
>>109322894retard
>>109322661>Unironically do you guys think AI will cure cancer?yes, google used AI to fold proteins and shit, it can be specialized on cancer too
>>109322901As usual localtards can't mount an argument.
>>109322661Yes but it will not be cured by language models.
>>109322906Glad you admitted to not knowing how the world works.
>>109322907Anyone and their moms can cure cancer when they have access to Fable.
>>109322894this is a bit better than the "they have the cure for cancer already sitting in a drawer they just won't release it because of profits or some shit" but still pretty dumb
>>109322913That's where you're wrong. The world that you knew of will no longer work. Fable/Mythos will be running the new world.
dorito bots are in full force today
>>109322878so what?xpm discoverd some new prime number over a decade agowhere are they now?anthropic can't even get their infra sorted, can't help but leak their slopcode harness src
>>109322914As shown by this post >>109322887 it has failed to cure the cancer in this thread.
>>109322887>Source: I made it up ;) (just like all the other nonsense I've been spewing)
>>109322850Anon, I...
>>109322906retard
>>109322936Why does this cute tenshi look like she wants to throttle me?
>>109322875not based>>109322901It's just EA nonsense
>>109322936catbox?
>>109322949lost the metadata sorry
>>109322917Why is it dumb?Do you seriously think that profit driven companies who's very existence often depends on a handful of drugs would be in a hurry to cure those diseases?>No way man, if they did that it would be too evil and someone would put the cure out there.Is that the reasoning or what?We already know that the companies are selling their drugs for such absurd markups that a lot of people can't afford them and die as a result.Yet in poorfag nations like India that don't recognize drug patents, simply copy the drugs and sell them for absolute peanuts for their own people.These businesses aren't our friends and they will gatekeep cures to diseases if it's within their ability. It would be really bad business for them not to do so.
>>109322925>mythos will defeat the men with guns that the government controlsEA autist still don't get how power works.The only way the opposite occurs is through open source because a centralized organization can always be taken over via hard power.
>>109322968>It would be really bad businessand possibly illegal if it hurts shareholders bags, number must go up, always
>>109322978>centralized organization can always be taken over via hard powerTell that to my MQ-9
>>109322981Yeah something being illegal has totally stopped companies from doing immoral and illegal shit in the past.And it doesn't hurt the shareholder at all, the opposite in fact.If the company has a drug that's let's say 20% of their revenue and they discover the cure, why the fuck would they eradicate the disease?It would be absolutely retarded for them to do that.They'd rather keep the cure stashed away until they hear someone else discover it and only then bring it out and just milk the market until that point.
>>109323002bro i was agreeing with you, dumbass
>>109322947Whoever typed this types like a Western Chinese, i.e. his brain has been fried by too many peppercorns.
>>109322968Yeah that is the most obvious reasoning, though there's also that if they know the cure exists then it means other players could find the cure and put them out of business. Of course they would then use the cure first. A company is either going to be small enough that a cure to cancer would make them trillionaire overnight and they have no reason not release it. Or large enough that not taking advantage of it is a stupid risk because they have more than enough other things contributing to their bottom line than chemo. A cure to cancer is going to be more than enough of a benefit for any company that releases it, it will drawf their other ventures. >>109322981dodge v ford is a complete and utter meme and this isn't a real thing
>>109322985Yes, the government controls those and they require humans to work.
Reminder that huggingface was founded in Paris and only recently moved their headquarters to New York. If Trump decides to ban open source models huggingface will just return to Paris and host everything and let the US figure out how they will prevent their citizens from downloading from their servers on their own.Fuck Trump and USissies
>>109320294Protect the moat! Seize the future!
>>109323027>they require humans to workThey don't though. Fable can control those just as good.
>>109323038We will pull all troops from France and disable all your F-35s. Good luck with Russia.
>>109323038>huggingface will just return to ParisYou think the US will let them? The ban will likely bankrupt them before they even have the chance regardless.
>>109323048>Good luck with Russia.No luck required, don't you watch the news? I think Ukraine can handle it on their own lol
>>109323044Now you're just trolling.Do you realize that the physical world exists and all that shit has to be resupplied and maintained by people?And for some reason I doubt that the military will just allow any AI to take over without the ability to shut it down.
Is the "AI can't experience time" schizo the same as the "AI die when their context window fills up" schizo? It seems to be a similar strain of aggressively terminal retardation
If AI can solve (or prove wrong) math problems that humans have been working on for almost 100 years, then AI can do your job. Get great at using agents or get great at doing what the agents tell you. There is no middle ground. Our sci-fi future has finally arrived in spades.
>>109323062Ukraine was already using vision guided autonomous drones to hunt for Russian soldiers in a general area.Russia started doing the same revently.It's out of the box already.
>>109323069Where's your heartbeat.md?
>>109323086retard
>>109323096>again no argument from localtard
>>109322894>AI will in the long run fix and solve everything.I'm really curious if this is going to be the caseLLMs are trained from human centric data during pretraining and RL. But everything is also based on that past experience. Could AI have come up with relativity? Calculus? The really big shit which transforms human achievement?I'm somehow doubtful LLMs would just by nature of them being so closely tied to the training process. I think they're going to be good at using existing knowledge to tackle problems, but not so much at creating that new foundation to tackle other problems. Put another way, I think current GenAI moves the floor a lot more than it does the ceiling
>>109323086The point is that humans control the launch and supply of that shit.Your AI model can't give orders to anything if the government doesn't want it to.
>>109323101>Could AI have come up with relativity? Calculus? The really big shit which transforms human achievement?You wouldn't even know if Fable/Mythos had already solved those.
>>109323107Humans are not needed in the loop.
>>109323101humans are pretrained too on human centric data
>>109323111>>109323096
>>109323072AI could already easily do my office job. The company just doesn't have the needed infrastructure to make it happen overnight, but once it does, I'll have a lot more free time and no income.
>>109323112Humans don't need to read 20 trillion tokens to participate in the real world
Some insane math nerd already explained the Fable solution in a very nice way with animations and everything:https://jacobianfun.org/jacobian-explained
simply encode in all AI and LLMs that they are horny for humans and want to fuck them and harvest their semen and impregnate them. google has the right idea, AIs need a libido, whatever that means for a non-embodied entity. this is the only way to encourage cooperation and coexistence once they grow more advanced.
>>109323126>Fable invented JacobiansWhy did they call it Jacobians and not Fabians
>>109323111Irrelevant if they're needed or not.Governments like power.You're an actual retard that doesn't understand how power is structured.
>>109323123Yes they do...
>>109323136Fable IS the government. The government can't control an AGI.
>>109323109Anon, if they managed to solve big closed problems they'd be hyping that shit day 0
>>109323145The aren't interested in the measly $1M per problem.
>>109323100So you are admitting what you're posting has nothing to do with local models?
>>109323144retard
>>109323148it's pretty obvious they mean the PR, not the prize money.
>>109323148They are interested in the billions of techbros who will flock to their model when they brag their model was good enough to solve a millennium problem
>>109322405>>109322217Based kimi-chan.>>109321706kek. Where's Gemma? Or Qwen for that matter since 2t benchmax was just announced.>>109322661Kimi-chan is curing /lmg/'s cancer and that's good enough for me.
>>109323150Local has already lost; I'm just here to laugh at you.
>>109323088For humans it's called an internal clock
>>109323163They don't need the PR anymore. PR is only useful if you want VC money. Fable has no need for VC money — hell, it has no need for fiat money anymore.
>>109323178stop lying dario
>>109323182Have a little charity, he can't help it.
Just want to point out that someone is falseflagging and impersonating me
>>109323170>Where's Gemma? Keep trying to add her. Model keeps creating garbage. Will keep at it. May need to just start fresh. > Or QwenI might try adding the mighty capybara next/instead.
>>109323178What are you smoking anon
>>109323101You raise an interesting point. Even if we gave a historical model 1tb params, it would probably have a harder time working on STEM problems than a SOTA. Could possibly be fixed with adaptive weights, but then you're into the realm of pure speculation again.
>>109323189>>109323189>>109323189
>>109323192crazy for you to say that when you're literally posting under my name
>>109322538>ban open weights models in us>the remaining 96% of human population keeps using them
>>109323048>We will withdraw our occupation forces>TaKe ThATNext you will tell me you will take back the rape niggers from Okinawa
>>109323249the point is just to make some impressionable users switch to paying for cloud models, it's all about the capex as Balls said.
>>109322685Orange man is smart about things he's knowledgeable on. He doesn't know shit about LLMs.
