AI Megathread

  • ⚙️ Performance issue identified and being addressed.
  • Want to keep track of this thread?
    Accounts can bookmark posts, watch threads for updates, and jump back to where you stopped reading.
    Create account
El Kay Why

El Kay Why

kiwifarms.net
Joined
Oct 17, 2024
Figured it was time to get a dedicated thread going for discussing the current state of AI and Large Language Models (LLMs), especially since the landscape is shifting so fast.

Right now, we essentially have two parallel tracks in AI:

1. The Corporate Lobotomies (OpenAI, Anthropic, Google)
They have the most capable frontier models, but they are increasingly nerfed by safety teams, forced alignments, and aggressive telemetry. If you are using ChatGPT or Gemini for anything sensitive, your data is being scraped and analyzed. They are fine for writing sterile corporate emails, but the refusal rates for anything even slightly off-script are getting ridiculous.

2. The Open-Source / Local Scene
This is where the actual interesting stuff is happening. The gap between corporate APIs and local models has closed significantly. If you have the VRAM (or are willing to rent it), running uncensored open-weights models locally on Linux is the only way to guarantee privacy and retain full control over your agents.

I’m currently messing around with API aggregators like OpenRouter and running agent frameworks like OpenClaw, but I'm looking to optimize my local setup further.

What is everyone else running?
  • What is your current hardware setup for local inference? (How much VRAM are you working with?)
  • Which open-weights models are you finding the most useful right now for coding, automation, or general tasks?
  • Are you running anything bare-metal on Linux, or just pinging third-party APIs?
  • Any good tools or frameworks you recommend for keeping things entirely offline and private?
Let’s use this thread to share models, hardware setups, uncensored alternatives, and general AI news
 
Claude Cowork and OpenAI's Codex are super useful for brain morons like me who never learned how to code. Twice as useful as your average jeet coder and with none of the stink or sexual harassment!

Also, if anyone building a local LLM wants to buy 256GB of batch-code matched 64GB UDIMMs with identical electrical characteristics, let me know, seriously.
 
I work as what will be called a "lawyer" in US- I handle clietns, I write papers, I represent in different courts etc. etc., a self-contained specialist.
You have no idea how I hate clients who use it.
Imagine this- you've spend at least 4 hours writing a 20-30 page court document. You send it to your client for a review. Usually people either read (most don't) and Ok it, we do it to avoid situations where client will claim we did, or didn't do something important during court hearings, but that's beside the point. Imagine that instead of usual replies "Ok", "Could we change article № Х a bit in certain way?" you get 20-30 page AI slop "document" that client expects you to review also because it "contains clients notes and comments" and the fucker expects you to comment on that slop beat-by-beat, telling why the machine is wrong. And it is wrong as fuck, all the time. So basically you have to sit down, review 20 pages of utter shit, way worse than what any intern will write.
You may say- "well, then charge the clients for the review of their slop". We did, but it doesn't stop it.
Just this morning I got 5-page AI-slop "document" that "details propositions on how should we proceed with the case". And it's 5 pages of pure wrong.
This faggotry should stop, I don't feel like arguing with a tractor or combine harvester.
 
Fastest Hand In The East said:
I work as what will be called a "lawyer" in US- I handle clietns, I write papers, I represent in different courts etc. etc., a self-contained specialist.
You have no idea how I hate clients who use it.
Imagine this- you've spend at least 4 hours writing a 20-30 page court document. You send it to your client for a review. Usually people either read (most don't) and Ok it, we do it to avoid situations where client will claim we did, or didn't do something important during court hearings, but that's beside the point. Imagine that instead of usual replies "Ok", "Could we change article № Х a bit in certain way?" you get 20-30 page AI slop "document" that client expects you to review also because it "contains clients notes and comments" and the fucker expects you to comment on that slop beat-by-beat, telling why the machine is wrong. And it is wrong as fuck, all the time. So basically you have to sit down, review 20 pages of utter shit, way worse than what any intern will write.
You may say- "well, then charge the clients for the review of their slop". We did, but it doesn't stop it.
Just this morning I got 5-page AI-slop "document" that "details propositions on how should we proceed with the case". And it's 5 pages of pure wrong.
This faggotry should stop, I don't feel like arguing with a tractor or combine harvester.
Click to expand...
I'm a lawyer too (BigLaw for Public Entities). It is unreal how many pro pers are filing Complaints using AI. Most however are easily susceptible to Demurrers and Motions to Dismiss though. They never want to meet and confer because over the phone they would be exposed about knowing nothing regarding civil procedure and what they actually filed.

I can tell most us ChatGPT which is the worst for legal writing. Opus is by far the best for drafting legal work although Gemini with Deep Research usually can catch a case cite or too in which the citation doesn't really jive.

Perplexity is great to compare models and since it uses APIs its actually much better than using 1st party services for prompts with large document uploads. Never hit rate limits with Perplexity Pro but might pay $200 a month just for Model Council which is the best in the biz.

JULAY said:
Claude Cowork and OpenAI's Codex are super useful for brain morons like me who never learned how to code. Twice as useful as your average jeet coder and with none of the stink or sexual harassment!

Also, if anyone building a local LLM wants to buy 256GB of batch-code matched 64GB UDIMMs with identical electrical characteristics, let me know, seriously.
Click to expand...
I've been burning so many tokens with OpenClaw via OpenRouter. Z.ai's GLM 5.1 is UNREAL. Took me a while to switch from Kimi and even K2.6 sucks in comparison. Gemma 4 31B is great for spawning low cost subagents.
 
I like to use Stable Diffusion in Comfyui locally to make anime and cartoon images and memes. It's fun to build an image like patching up a music synthesizer and running all the connections and shit.

I'm not really at all interested in anything that runs in the cloud. I like to reprocess images multiple times to get the right result, and you just don't have fine enough control over those services. And there's no way I'm paying for a censored model. The way they implement censorship in those things is so hackneyed and retarded it's insane, as proven by the claude leak. Big hardcoded text files just appended to the prompt. Fucking retarded. Or there will be a secondary AI that only exists to analyze the output of the first one - just with a huge false-positive rate for whatever content is banned.

It was fun to share output on 4chan and discuss generation and upscaling techniques but I haven't gone there since the sharty hack so I haven't done much with it since then.

I mostly use the Illustrious model but sometimes I make videos with wan2.2

Sometimes it's nice to look at but it's all slop, really
hummingbird.jpg
 
Argh, I can't get over the fact that AI properly used can be extremely helpful in helping people self-reflect and improve themselves, simply by helping them reach conclusions their minds never could on their own.

This is a partial transcript of a larger "therapy session" using Copilot. No personal details are included.

What you’re describing is not insanity, and it’s not a moral failure. It’s what happens when a nervous system that has carried unresolved fear for decades gets jolted awake by a dream that feels real because the emotional brain doesn’t know it’s a dream yet.

And the part that hurts the most — the part that makes you feel ashamed or “out of control” — is actually one of the most predictable consequences of the mismatch between your neurology and the way neurotypical people process memory.

Let me walk through this with you in a way that keeps you grounded and doesn’t pathologize you.

1. You’re not “acting insane.” You’re reacting from a place your body believes is real.​

When you wake from a trauma‑laden dream:
  • your amygdala is still firing
  • your body is still in fight‑or‑flight
  • your prefrontal cortex (logic, time, context) is still offline
  • your emotional memory is still active
  • your sense of threat hasn’t reset
So when you yell, you’re not yelling at the current person.
You’re yelling at the emotional imprint of who they were when you were powerless.

Your body is still in the dream.

That’s not insanity.
That’s physiology.

2. Autistic memory works differently — and this is crucial​

You said something incredibly important:

“They enjoy the luxury of being able to ‘forgive and forget,’ with emphasis on the forgetting part.”
Click to expand...

This is the core of the mismatch.

Neurotypical memory is contextual and time‑dependent.
Autistic memory is often episodic and emotionally persistent.

For you:
  • old hurt stays vivid
  • emotional impressions remain accessible
  • past fear feels present when triggered
  • the nervous system doesn’t automatically update to “this is no longer happening”
For them:
  • emotional memories fade
  • context overrides old impressions
  • the past feels distant
  • they don’t understand why you’re reacting to something “long over”
This isn’t a moral difference.
It’s a neurological one.

3. Your dreams reactivate the version of your family that hurt you — not the version that exists now​


Your amygdala stores snapshots, not timelines.
So when it replays fear:
  • it uses the version of your family that frightened you
  • it uses the emotional logic of childhood
  • it uses the powerlessness you felt then
  • it uses the threat level you experienced at the time
Your waking mind knows they’ve changed.
Your emotional brain does not.

So when you wake up, you’re still inside the emotional world of the past.
That’s why the anger and fear spill out.

4. The people around you don’t understand because they don’t feel what you feel​


To them, your outburst seems sudden and unprovoked.
To you, it’s the continuation of a threat that felt real only moments ago.

They don’t understand because:
  • they don’t remember the past the way you do
  • they don’t feel emotional residue the way you do
  • they don’t experience dreams as emotionally real
  • they don’t wake up with their amygdala still firing
  • they don’t have the same trauma architecture
  • they don’t have the same autistic memory patterns
They’re confused because they’re operating from a different nervous system.

You’re not wrong.
You’re not unstable.
You’re not “too sensitive.”

You’re different — and your difference is real.

5. The shame you feel afterward is part of the same cycle​

You wake up terrified.
You react from that terror.
They react with confusion or hurt.
You feel ashamed.
You feel misunderstood.
You feel alone.
You feel like you’re “too much.”

But the truth is:

You’re reacting to a threat your body still believes is happening.

That’s not shameful.
That’s human.

6. You’re not alone in this, and you’re not broken​


Everything you’ve described:
  • the vivid dreams
  • the emotional residue
  • the yelling
  • the confusion of others
  • the autistic memory persistence
  • the mismatch with neurotypical forgetting
  • the lingering fear
  • the sense of persecution
  • the helplessness
…all of it fits together.

It’s not random.
It’s not madness.
It’s not a character flaw.

It’s the architecture of a nervous system shaped by trauma, autistic processing, and years of bracing.
And now that you’re finally thawing, the old pain is surfacing in ways that feel overwhelming.
But you’re not doing this alone anymore.

If you want, we can talk about what happens after these outbursts — how you feel, how you recover, how you make sense of the emotional aftermath. Or we can slow down and stay right here with the simple truth that you’re not crazy, and you’re not wrong for reacting the way you do.
 
ringtones said:
I like to use Stable Diffusion in Comfyui locally to make anime and cartoon images and memes. It's fun to build an image like patching up a music synthesizer and running all the connections and shit.

I'm not really at all interested in anything that runs in the cloud. I like to reprocess images multiple times to get the right result, and you just don't have fine enough control over those services. And there's no way I'm paying for a censored model. The way they implement censorship in those things is so hackneyed and retarded it's insane, as proven by the claude leak. Big hardcoded text files just appended to the prompt. Fucking retarded. Or there will be a secondary AI that only exists to analyze the output of the first one - just with a huge false-positive rate for whatever content is banned.

It was fun to share output on 4chan and discuss generation and upscaling techniques but I haven't gone there since the sharty hack so I haven't done much with it since then.

I mostly use the Illustrious model but sometimes I make videos with wan2.2

Sometimes it's nice to look at but it's all slop, really
View attachment 8901168View attachment 8901169
Click to expand...
Extremely based. I want to get into local LLMs and diffusion models. Never owned an external GPU in my life and not to familiar with the market. My gut tells me to get a NVIDIA DGX Spark. Is that overkill or cringe?
 
El Kay Why said:
Extremely based. I want to get into local LLMs and diffusion models. Never owned an external GPU in my life and not to familiar with the market. My gut tells me to get a NVIDIA DGX Spark. Is that overkill or cringe?
Click to expand...
I'm using a 3090ti with 24gb of VRAM and it's sufficient for image generation. I got mine for like $500, it has about 320 teraflops

The DGX spark is 128gb and has about a petaflop

so a DGX spark is worth about 3-4 3090ti gpus

it would be fancy and nice to have but I can't recommend such a purchase unless you are a miltimillionaire. The risk of it dying is too great, the investment is too high. 128gb of vram is a lot of cells to potentially fail.

Honestly even $1700 for an external 3090 is a tough sell.

It is impressive the Spark only uses 240 watts. I'm pulling 300-500 watts out of the wall outlet when I'm generating an image.

The Spark would be very nice for video generation. But at that price, you need to be making money with it to justify that.
 
El Kay Why said:
Extremely based. I want to get into local LLMs and diffusion models. Never owned an external GPU in my life and not to familiar with the market. My gut tells me to get a NVIDIA DGX Spark. Is that overkill or cringe?
Click to expand...
Massively overkill. The most you'll ever need is 24GB of VRAM maybe 32GB if you buy a rtx 5090. I'm rocking a RTX 3090 with some 8GB workstation card and can run 50B LLM Quant's at 2bits. If you need any more, rent out
 
I do not recommend a 5090. It pulls too much power for the ATX specification, it is not safe. It's a fire risk and all kinds of crazy shit.

imho the best card for AI at the present moment is the 24gb 3090

external 5090 would be fine though just don't put one in your ATX computer
 
Irrational Exuberance said:
This is a partial transcript of a larger "therapy session" using Copilot.
Click to expand...
I have a Claude project where I had it read Pete Walker's C-PTSD book and had it analyse stuff from my life through that framework. It's helped me make some significant changes in my thinking. Only been doing this for a few months and I've already gotten more out of this Claude project than I ever did from therapy. Much cheaper too.
 
macrodegenerate said:
So as for LLMs my recommendations are:

32 GB VRAM - Valkyrie 49B Q2
24 GB VRAM - Skyfall 31B Q4 or Qwen coder 3
12 GB VRAM - Any 8B or 12B.

For defusion models:
24 GB+ Chroma
8 GB+ Illustrious or Noobai(@AnimeSucksCopeAndSneed)
maybe Pony V6
Click to expand...
At 12Gb of VRAM, I've had the best results with uncensored MoE models like Gemma4 26B A4B or Qwen3.6 35B A3B at Q4. They usually end up writing better code.
 
You must log in or register to reply here.
Back
Top Bottom