Try out our narrow agents on #SpaceKit https://t.co/2vkB2JZtXl a demonstration of decentralized #AI
Sources for #Agents will be released next week along with the CLI to train, merge and infer!
DM any comments or suggestions
SWTCH Labs has executed the original vision of building a decentralized media network.
#SpaceKit is meant to empower #creators and #developers in the new digital economy by providing the foundation to create safe apps!
Smart contracts can call #AI natively on #SpaceKit.
Train a #Growformer brain, deploy as a Fact Package, invoke inference from any contract, same import surface as ERC-20.
No GPU needed.
Sub-second inference on CPU.
Open-source SDK.
👉 https://t.co/hhL4gFazo6 · https://t.co/GbQUmN25nQ
#SpaceKit#News
Take a look at what we have "built"/"shipped", with the CLI now preparing to be released to security researchers and the public.
Initial access will be gated during testnet phases, due to the size of the testnet being limited to 1000.
https://t.co/dB5vwNb88j
SpaceKit Update!
Managed to provide a unified interface for the #Growformer Neural Substrate via #SpaceKit CLI!
Here is the workflow minus training the causal brain in 3 steps once you have the project setup.
% spacekit agent train --project crypto/crypto-sentiment-analysis.gf.toml
% spacekit agent merge --brain crypto/agent/crypto-brain.bin --overlay-brain causal/agent/causal-brain.bin --brain-output crypto/agent/crypto-causal.bin
% spacekit agent infer --brain crypto/agent/crypto-causal.bin --prompt "I love that Bitcoin is above $100K , this cycle feels different and I'm finally in profit"
Results from cli from --infer
% Brain loaded: crypto/agent/crypto-causal.bin · Crypto DeFi Sentiment Analysis (2 groups)
POSITIVE (strong) — I don't have enough information to explain this with confidence. Grounded in the user's own words: "I love that Bitcoin is above K , this cycle feels different and I'm finally in profit"
QA'ing now and preparing the spacekit-projects documentation to provide a "Getting Started with SpaceKit" section.
@IanCutress I've focused my research to take advantage of CPUs, specifically for training and inference.
Example sentiment agents here, we have more contributors adding more soon for the public domain: https://t.co/2vkB2JZtXl
I posted https://t.co/f5M2TNZbYN on LinkedIn and X in the past 24 hours.
That single post, without paid promotion, without any campaign, drove roughly 4,300 SpaceKit-JS VM Node spinups in Canada alone, plus an additional 1,000+ across 18 other countries on https://t.co/FihFZ8m7uF
The infrastructure activates instantly because each page load is itself a deployment, not a signup.
We have a new website to check out
https://t.co/btqYsZjsaJ
We are focused on #SpaceKit testnet which is live now with a growing list of our #Growformer models.
Our latest smart contract we are adding next is #RouteKit which employs all the #SpaceKit VM features.
Stay Tuned!
We have launched the https://t.co/ywDTTd3e78 to test drive our work on the https://t.co/9A1keWseGv narrow models.
- Fintech Sentiment Analysis Agent
- Cryptocurrency Sentiment Analysis Agent
Our models run in the browser, #IoT, servers and mobile!
#NarrowAI#Agents
https://t.co/9A1keWseGv now has a runtime!
- Growformer exports a narrow agent brain
- Growformer runtime loads the agent brain
- Growformer runtime works on mobile/desktop
Our QA efforts over the weekend have shown a viable solution to running #AI on mobile.
Generative Demo for https://t.co/9A1keWseGv brain/agent, training and then rendering what it learned at record speed!
#Investors https://t.co/9A1keWseGv is raising a round to continue on our Roadmap, DM for more details.
https://t.co/SXGgS0zVIJ MNIST Generative Demo
The demonstration learns 2k MNIST samples and then generates the learned sample numbers in 5 seconds on CPU.
Stable Diffusion (GPU, 50 steps): 3-8 seconds
Stable Diffusion (CPU, 50 steps): 2-10 minutes
DALL-E API: 3-6 seconds (cloud GPU)
Growformer Brain (CPU, 1 pass): 5 seconds
We're matching GPU cloud generation speed on local CPU trained on 2k images.
DM for more info, we are training #NarrowAI #Agents now with tool calls based on this architecture with zero-forgetting and hallucinations. Try Us!
#AI #NeuroAI #ContinousLearning #generativeart
Just hit MNIST train+inference in 31.77ms on CPU.
Full brain: 109MB. Zero forgetting.
Our AI substrate handles vision and language, no GPU at any stage:
- 97.7% MNIST accuracy. Up from 97.3%.
- 90% conversational accuracy.
- 25-minute CPU training.
- 106MB deployment for multiple-subjects.
Everyone else is still building bigger LLMs that take seconds per token.
We grew a brain that thinks in real time. Real agency will now be in motion.
#AI #NeuroAI
Our new #AI at @swtch_labs grows neural structure like a biological brain.
No backprop, No transformers.
- Neurons compete, specialize, and lock in as engrams
- Single‑pass text generation
- Inline tool execution
- Zero forgetting by design
- Fully specialized codegen on extended holdouts
This is real continual learning.
The entire agent test model is 18.9MB with sub‑second CPU inference.
More data density, more learning, larger file sizes, but it beats loading gigabytes by a long shot.
DM me for more information and early access.
#NeuroAI #ContinualLearning
"micro-brains", "micro-agents", "task driven" specialized "expert-agents" with a narrow vision, these are a much better premise than having 3 or 4 large AI SaaS providers.
Zero forgetting. Confirmed.
Fresh continual MNIST run → saved checkpoint → reloaded and re-tested all 5 tasks.
Result: 97.3% average accuracy.
Zero degradation. Zero hallucinations from interference.
This is a continually learning substrate that actually remembers.
Completed full continual MNIST, 5 sequential tasks, 10 digits, no retraining
0vs1: 98.3%
2vs3: 95.8%
4vs5: 98.0%
6vs7: 98.3%
8vs9: 97.3%
Average: 97.5%
One growing organism. Zero catastrophic forgetting. CPU-only.
This is a continually learning substrate on the path to AGI.
Our organism is actually learning like a living brain.
- XOR: 100%
- Spirals: 94.5%
- Concentric Circles: 97.9%
Trained on spirals → saved checkpoint → adapted to circles with zero full retraining.
We don’t code intelligence.
We grow it.
#NeuroAI#ContinualLearning#AI
Our "grown intelligence" is already solving classic problems:
• XOR: 100%
• Spiral Classification: 75.4%
Not bad for a system that builds its own architecture instead of being hand‑designed.
#AI#NeuroAI#AIBiology