This is going to sound a little (Ok, a lot) goonish ;) — first off, I love the LFM models. My first app in the App Store is built around one… but the vision models are VERY shy when it comes to NSFW images. I’m assume they simply weren’t shown any during training (which makes perfect sense) but it can be an issue if you’re trying to classify a bunch of images and need to separate out any that might be explicit.
@owenhagger@claudeai I’ve been using Fable on a cyber security harness build and been working for about 2 hours. I’ve used 4% of my Fable limit… and I have 3 fable resets banked.
@ihteshamali There is zero chance any 5 series Claude model will use this. Anthropic’s guardrails against cyber are incredibly sensitive. It can’t even read the description of my pentesting bot without dropping to 4.8.
At my day job, I often joke that companies like AutoDesk, Microsoft, Intuit, etc have a mission statement of “what are you going to do? Use something else?” — and increasingly the answer is “yes, yes I am”.
I got so sick of Wondershare constantly asking me for more money to upgrade my already paid product, and Adobe’s ridiculously priced subscriptions (and the other companies that want $400 for an app) that I built a complete Acrobat replacement. It works better and the buttons are exactly where I want them.
1000% agree. I was using LM studios and kept running into issues. So I said heck with that… and built my own. This is Sindri. A Llama.cpp wrapper. It has a “Sweep” mode that loads an MOE expert by expert until you hit the VRAM wall. It’s also a LOT faster than LM Studios… and when I find a new feature I want, I just add it :)
@MensHumor My Jeep is 27 years old and has 80k miles on it. I try to change the oil every year…. Usually that’s around 1000 miles. Last oil change was 3/14 and I’ve driven 822 miles since then.
Im probably not the target demographic for this discussion lol
:)
Oh, I didn’t fine tune, I cheated — sent my research bot out to look for existing fine tunes for writing, it came back with this one https://t.co/pbFInFXEMb
And it’s been my daily driver ever since.
(As an aside, if you need a good liar, try the Bonsai 27b, I was trying to use it for my AI assistant, but it would NOT stop improvising . I’d say “I need coffee”
And it would volunteer to go get it for me lol. So it’s great at roleplay :) )
A while back I built a harness to help me write… it basically stored the work I’d done and made it easy to keep track of who is where (my current novel is written in a Stand(ish) style… so multiple characters, multiple groups, and several story lines (so it’s easy to get confused) and it’s been super helpful …
So, then, as an experiment, I built an Auto-writer. You can give it a 2 paragraph hook and it will write an entire 150k word book. This was always just an experiment… can I make enough rules that a small local model (the Gemma4 26b, StyleTune) can write a novel that’s worth reading … and you know what? You can’t … lol ok, I kid, the work doesn’t outright suck, it’s pretty decent really. Would someone pay for one of them? Probably not… but it was a fantastic learning experience and quite a bit of fun.
Anyway, I had retired the app (technically it’s an agentic harness) but seeing stuff like this makes me want to flood them with AI slop novels … I can generate a 30 chapter, 150k word novel in about 2 hours lol :) set it to run overnight …
@sepo100@GuideboardLabs@liquidai So far, I haven’t tested any long form… but it’s definitely more creative on the story hooks. It’s going to be really difficult for it to beat the StyleTune finetune 26b I run … but I’m looking forward to testing it out :)
If anyone wanted to test the prompt as well :
Hyper realistic photo of a majestic, hooded Necromancer standing amidst a haunting graveyard. The figure possesses a muscular, bronzed torso with intricate glowing teal veins and is clad in ornate, dark bronze armor featuring skull motifs on the shoulders and greaves. A long, tattered crimson and gold cloak flows behind him, while his hands glow with ethereal turquoise energy, cradling a floating, flaming skull. He stands on a carpet of weathered skulls and jagged rocks under a dark, mystical forest canopy where bats flutter in the shadows. The lighting is dramatic and moody, dominated by a cool cyan bioluminescence that casts deep shadows and highlights the metallic textures of the armor and the fabric weave of the robes. Shot from a low-angle medium-full shot with a shallow depth of field, the composition emphasizes his imposing presence against the atmospheric, misty background.
Thanks to the fine folks at @krea_ai I got early access to the new Krea3 image model. Really really digging what Im seeing. I havent had a chance to get into the minutia of course, but here is an example (left image is Krea3, Right image is Krea2) - Just a LOT more detail shows up (and Krea2 was never lacking there :) ). This is a basic prompt, with no LORA or anything. The Krea2 was generated in ComfyUI locally on my PC.