I've been saying the same thing for a while now, we need to stop this Multi-Trillion parameter trend.
We certainly DO NOT need that much.
Andrej Karpathy:
“A 1B model can be much more effective than most current LLMs.
90% of current models waste a ton of capacity and money on remembering stuff that doesn’t matter.”
here is his 5 core thesis about Small LLM’s:
1. the intelligence was never the knowledge. Strip the memorized internet, keep the reasoning. What’s left is the cognitive core, and it’s small.
2. the core might fit in ~1B params. Maybe less. “Plenty of room at the bottom.” We’ve been paying for a library when we needed a brain.
3. size stopped being the bottleneck. What makes an agent work is the tools it can call and how tight the feedback loop is, not parameter count.
4. intelligence lives in the system, not the weights. A raw LLM is a CPU with no OS. Small core + good harness > giant model alone.
5. you can own it. A model on your hardware is one no one can rate-limit, price-hike, or deprecate on you.
in 12 months the gap between teams shipping on-device and teams paying per-token frontier bills won’t be capability.
TwIL an 8.8B model at 90%+ of frontier isn’t a benchmark flex.
It’s the thesis Karpathy’s been repeating all year, made real.
study more on this below ↓
30
I've been working on something even better 👀
#SOON
Substance Painter got cloned? 🤯
Meet SubCraft, an independent 3D texture-painting app that lets you paint directly into models. It's still an early-stage project, but definitely one worth keeping an eye on.
For more info check out our video below:
#3dart #substance #3D
1
1
45
ShinyXYZ retweeted
Replying to @ShinyStudio_XYZ @RaphaelLiu10
For people wondering how Comfyture handled it, I mean... Pretty nice isn't it ? 🥹 📷🐣
#OpenSourceHiggsfield
This might be private property.
Comfyture is PUBLIC property.
An open-source platform ran by baby pets. Trends, 3D, photoshoots, anime openings, whole films. MIT, built on Comfy.
That's my participation to the ComfyUI Challenge! 🐥
github.com/ShAInyXYZ/Comfytu…
#OpenSourceHiggsfield
Two years ago, I predicted a decline in demand for frontier models. I believe we're starting to see it happen now.
We don't need frontier models for everyday tasks.
The future will be a hybrid architecture: a core intelligence layer (the brain/orchestrator) that can leverage frontier models when necessary, combined with local models handling execution.
For most companies, this will eventually become the standard.
2027 will mark the beginning of the local model era.
Instead of relying on massive, general-purpose models that know a little about everything, we'll fine-tune smaller models to excel at highly specific, niche tasks.
Think about education: we don't teach every student every subject in depth. We structure curriculums so they can specialize and become experts in their respective fields.
Why shouldn't we apply the same logic to AI?
A chemistry student doesn't need to master military law or agronomy. Likewise, a model designed for a specific task doesn't need to know everything.
I think people abandoning local models in favor of API credits are underestimating how valuable local models can be for everyday tasks, private workflows, and automation.
The future isn't about having the biggest model. It's about having the right model for the right job with the right operative structure and harness around it.
a year ago, an interviewer called me naive.
all because i said customers would fine-tune open models instead of using frontier APIs. (true story)
this week, we released LFM3 at @liquidai:
small open models built to be specialized.
as a demo, I fine-tuned our 2B vision model to spot defects in drone photos of wind turbines.
before:
LFM3-VL-2B and GPT-6 Luna named the defect correctly in under 10% of images.
after:
a few hours of fine-tuning on one GPU made it jump to 67%.
frontier models are generalists.
your domain-specific problem needs a specialist.
how you can do it too:
1. pick your domain-specific, narrow task and evaluate if a frontier model is sufficient. If not:
2. choose the LFM3 model that fits your data.
3. curate your own examples from your own proprietary data
4. fine-tune in a few hours, even on a single GPU.
5. rerun the evals and deploy anywhere without having to call a cloud API.
Fine-tuning notebooks:
github.com/Liquid4All/cookbo…
Wind turbine project: github.com/Liquid4All/cookbo…
what task would you specialize a model for first?
1
1
91
New prompting technique: Jealousy
me:“Claude Code worked all day until everything was fixed.”
Codex :"Say no more"
1
63
Codex is SO SLOW !
I wonder what's going on at @OpenAI right now...
Two days ago it was perfect...
60
For anyone asking, Yes, Codex (GPT is down) with an error 401.
It's saying Unauthorized API Key provided for some reason ...
You can get my NCD from here to get notifications faster
github.com/ShAInyXYZ/NoClick…
1
344
Anthropic absolutely nailed it with Opus 5.5. I've been running 4 sessions simultaneously for over a day now and I've only used +15% of my weekly quota!
I'm not even sure I'll burn through my weekly limit! Which is a first,
Genuinely impressed. Hats off. 🎩
44
I kept losing a day to the same mistake: the architecture I planned and the one my coding agent built drifted apart.
So I built Dia-GramV: one diagram we both edit , the agent through MCP, me on a canvas linted on every change.
Open source, local, MIT 👇 github.com/ShAInyXYZ/Dia-Gra…
ShinyXYZ retweeted
AI4AnimationPy 1-Click Launcher --- Every frame is generated, not authored.
An animation engine that uses AI to GENERATE every next pose in realtime, instead of human animators having to build frames manually. And it's FAST.
Wrote a UI that runs on ALL machines, 1click launch.
We now have new demos in AI4AnimationPy for stylized biped and quadruped locomotion control. Besides PC you can also run them in your browser or on your phone. Try it out here🎮:
Web-Demo: paulstarke-ai4animationpy.hf…
GitHub: github.com/facebookresearch/…
Video: youtube.com/watch?v=qifmLpNz…














