고정된 트윗
Storm Clan Demo just dropped — deck-building bullet hell on the alien ocean ⚡
Free on Steam now → store.steampowered.com/app/4617620/St…
#DeckbuildersFest #Deckbuilding #Roguelike #BulletHell #IndieGame #Steam #Wishlist
English
Studio Zamudio One
1.2K posts
Studio Zamudio One
@StudioZamudio
Studio Zamudio One -- Forging unforgettable and innovative game experiences and building the next generation.
가입일 Temmuz 2025
77 팔로잉132 팔로워
@quimedesu @LottoLabs I don't think you can train a 12B on a 3090.
English
90% of pretraining is just getting nonslop datasets and curating by hand
You can train a 1B model over a couple weeks on a 3090
Once you have the data you can plug and play different architectures
Just start bro
(Also another reason I like the 3090 for inference, you can actually train with it too)
English
@peterom Any company who thinks they have the right to piss all over free-thinking is 100% purely evil and dangerous.
English
My view on Anthropic is that they're fundamentally good, well-meaning people trying to do their best.
But the telling part isn't that they fixed the invisible safeguards - it's that 'tell the user' wasn't the default in the first place. Nobody in the room thought people were owed notice that their outputs were being quietly degraded.
This is the product of a belief, mostly unspoken externally, that they're the adults in the room - that everyone else is either misguided, a threat, or lemmings who need to be led.
And it's driving them, without meaning to, to dismantle the principles of open knowledge and open science that made AI progress (and their own company) possible in the first place.
The unspoken premise underneath it all: only they can be trusted to reach AGI first, to constrain everyone who'd do otherwise, and to walk us down the golden path.
History's verdict on people who believed that has never been kind.
POM@peterom
English
@caps_raunak AI mean that tens of thousands of people right this moment are building their own OSs. Most will fail, but many will easily exceed what already exists. Even if most people make sloppy code with AI, statistics also demand far superior code.
English
@StudioZamudio Humans already struggle with messy code from Microsoft and Google. AI won't make it worse we're used to fixing big confusing systems!
English
Unpopular prediction
In 3 years there will be a category of developer called "AI collapse engineers." Their entire job will be fixing systems built entirely by AI that nobody on the team understands anymore.
They will be the highest paid engineers in the industry.
The companies who fired their senior developers to save money will be paying 10x to hire them back under a different job title.
English
@DabsMalone Millions of people are already building their own. There's no cutting off the AI right now, no matter who demands it. The only way to cut them off is literally destroying the power-grid. I've made many of my own NNs. There's no way for a government to stop that.
English
Learn to code.
Because the day AI coding tools get regulated, censored, paywalled, or switched off, everyone who outsourced their skills is going to find out they never had them.
We’ve already watched platforms kill products people depended on. It happens every year.
If you can’t build without permission from a trillion dollar company, you don’t own the capability.🤝
English
@__tinygrad__ That's about the dumbest thing I've ever heard anyone say. Should BIC pens be liable for something that a person writes? Should contractors that build sidewalks be liable if a robber uses it before robbing a store?
English
While we are doing lobbying for AI laws, I propose this. All providers of closed source models are 100% liable for their outputs. Copyright, libel, conspiracy, fraud, etc... If you want safe harbor from this liability, release the weights. This is real AI safety legislation.
English
@babayagatwt That's all I do is think about the projects. There's a substantial amount of planning and execution involved. AI is just the tool for coding, that's it.
English
If you are in #SteamNextFest let's see your Indie Game. We'll be sharing a selection of games all week from this list and posting some unique content for 10 games that have caught our attention. Good luck! #IndieGame
English
@cloneofsimo That's a pretty strange speculation. Have you never used GLM or Kimi? They are ahead of most if not all American frontier models. Considering Fable's suspension, I would say they're ahead of every American model! You don't code much.
English
@TomDavidsonX You've got to be kidding... 🤣 You think that the NSA and DoD haven't already engineered the worlds most superior AI, and that they didn't start in the 1970s? They know well what they are doing, better than anyone else.
English
Seeing some ppl interpreting the Fable shutdown as positive news on govt intervention
But i don't think we'll execute a pause well*, or use the pause effectively, if the govt has NO idea how ai tech works and just wades in arbitrarily without consulting experts
it's def evidence for "govt might do lots of random extreme stuff" -- and that's prob good if you're p(doom) is v high and you want to fuck shit up and shake up the game board. but still, if you think alignment is v hard then ultimately we need sane ppl in charge to handle it
*executing a pause well is hard! You need to figure out effective verification, strike a deal with China, avoid extreme power concentration, catch rogue projects, maintain safety standards even as many unsafe projects catch up to the frontier
English
@10_X_eng I hate those crimp rings. I really don't like them.
English
You gotta be kidding me. You really can't just go once to Home Depot or Lowes
English
@80s_channel I just AI generate all my music. I can have exactly what I want, and how I want it. I can even write the lyrics!
English
@loftwah People who cannot produce, blame the AI in arrogance. It isn't the AI, it is them.
English
Irony is that you claim to be a software developer and you can’t get production ready out of AI but a non technical person can.
Will@wrhensel
@loftwah I’ve used AI, it cannot produce anything “acceptable” in terms of a production-ready application on its own.
English
@KimKatieUSA It's a LOT higher than, lol, 4.5 in New York City!
English
Tokyo
Population: 37 million
Homicide rate: 0.2 per 100k
New York City
Population: 19 million
Homicide rate: 4.5 per 100k
Why can’t America have cities that are as safe, clean, and orderly as Japan’s?
English
@AsianDawn4 Imagine what America would look like right now if we had lost WWII??? That's the first time I've ever thought this.
English
Banking is getting ridiculous. I logged into my account to transfer a certain sum of money from savings into checking. Said sum was apparently too large and they are holding it for up to 48hours to clear.
English
@Tablesalt13 Lots of people around the world are having children. Certain specific genetic populations have become far less productive. This is engineered. It's genocide.
English
‼️This BLEW MY MIND
this guy just figured out how come almost NO ONE around the world is having kids....
My generation didnt even have to THINK about this.
English
@ItsmeAjayKV @googlegemma Diffusion generation of code is just in it's infancy. It may never come to fruition, but it is a VERY interesting concept. Yeah, that model is just too small for code, as a diffusion model. Amazing none the less.
English
Is it just me or is @googlegemma Diffusiongemma broken, no matter what i do, i cannot get it to produce good outputs.
This is also why just trusting raw benchmarks is not enough.
anyone else on same boat?
12GB 3060
model diffusiongemma-26B-A4B-it-Q4_K_M.gguf
diffusion steps: 30
AJ@ItsmeAjayKV
Spent the past two days learning and testing new @googlegemma DiffusionGemma 26B on the new WIP llama.cpp diffusion branch (PR #24427)on my 3060. Unlike traditional LLMs that generate one token at a time, DiffusionGemma generates a whole block of tokens and repeatedly refines it over multiple denoising steps. Setup: - RTX 3060 12GB - diffusiongemma-26B-A4B-it-Q4_K_M.gguf @UnslothAI quant - `-ngl 15` (anything higher OOM) - Peak VRAM: ~10.4GB Prompt: "Write a Python quicksort implementation" 512-token generation (2 diffusion blocks): 12 steps -> 12.1s (42.4 canvas tok/s) 18 steps -> 17.9s (28.6 canvas tok/s) 24 steps -> 24.2s (21.2 canvas tok/s) 32 steps -> 32.0s (16.0 canvas tok/s) 48 steps -> 37.2s (13.8 canvas tok/s) A few observations: - Lower step counts were noticeably worse in quality - Denoising cost stayed almost perfectly flat at ~0.5s/step - Peak VRAM stayed around 10.4GB across all runs - It's not that bad even with offloading most to cpu - Diminishing returns started appearing around the 24–32 step range Still early days, but it's pretty wild seeing a 26B diffusion language model running locally in llama.cpp
English
@LexnLin They're there to clean up after the UFC fight at the White House tonight?
English
it's not looking good for eu anyways.
If we are lucky, people in the US get access to fable again the next week.
I guess I gotta wait for a few weeks/months tho😭
Rob Hallam@robj3d3
Update on Fable 5: > Anthropic staff have flown to Washington > Ongoing talks are happening with Trump administration > Both sides are eager to solve the dispute Something will probably come out tomorrow. This is also likely the reason we haven't heard from @AnthropicAI yet despite the "24hrs" promise.
English