AI Trains AI on 3 Home Computers: From 0/5 Recall to 73% Lower Loss in 38 Seconds ($0/Mo)
RAMDeck is a turnkey plug-and-play consumer hardware hub that pools RAM and GPUs across existing computers and phones in under 5 minutes with zero terminal commands—and just ran a complete “AI Improves AI” recursive loop strictly through its browser UI across 3 everyday home computers, taking an unseen private exam from 0/5 to 5/5 (100%) and coaching a compact Student AI to 73% lower training error in 38 seconds at $0.00 cloud cost.
1. Big AI Labs Say Recursive Self-Improvement Takes Billions — We Tested It on a Desk
Silicon Valley giants keep telling builders that “AI improving AI” only happens inside $10,000 server racks billed at $200+/month. Most consumer PCs choke with an Out-of-Memory (OOM) crash before a serious 30B or 70B Qwen, Llama, or DeepSeek model even finishes loading.
So we tested the whole loop on 3 everyday computers already sitting on a home desk (~57 GB pooled) using nothing but clicks in the RAMDeck web dashboard. Picture having a senior engineer sit down at your kitchen table, turn messy Slack chat scraps into a clean manual, quiz themselves to 100%, and coach a nimble apprentice in under a minute—without mailing a single file out of your house.
2. Four Point-and-Click Stages Inside Your Browser
Act 1: Pool 3 Home Devices & Measure Speed (32.2 tok/s)
One click in the Cluster tab combined ~57 GB across a desktop and laptops to serve an 18 GB+ 30B coding model too large for any single 12 GB card, sustaining 32.2 tokens/sec (1.26s avg response).
Act 2: Cheat-Proof Cold Exam (0/5)
To guarantee the model wasn’t quoting internet trivia, we gave it 5 private engineering constants invented purely for this test inside a clean Chat window. It refused to hallucinate and answered 0/5 (UNKNOWN).
Act 3: AI Distills Messy Chat Logs & Passes 5/5 (100%)
We pasted noisy standup chat messages containing lunch chatter and outdated numbers. The pooled 30B model extracted a crisp 8-bullet spec, we dropped that text file into the Knowledge tab, cleared the chat, and re-ran the same 5-question exam: 5/5 (100%) exact accuracy.
Act 4: 30B Teacher Trains Student AI (73% Loss Drop in 38s)
Next, the pooled 30B Teacher AI generated structured training examples and sent them into the Training tab. Generation 1 finished in 26 seconds at 1.39 training error; after feeding its critique back into a richer Generation 2 batch, a 38-second second run cut error by ~73% (down to 0.377)—and produced downloadable model adapters without even unloading the 30B Teacher!
3. Live Run Scoreboard (All Executed in Browser, $0 Cloud Spend)
| Stage | Dashboard Tab | Verified Outcome |
|---|---|---|
| Multi-Device Pool & Benchmark | Cluster | 3 home computers (~57 GB pooled) running a 30B model at 32.2 tok/s |
| Private Unseen Exam (Cold) | Chat | 0 / 5 (strict UNKNOWN reply—zero guessing) |
| Memory Self-Upgrade | Chat → Knowledge → Chat | AI cleaned noisy notes → fresh chat scored 5 / 5 (100%) |
| Teacher → Student Fine-Tune | Chat → Training | Training error dropped 1.39 → 0.377 (-73%) in 38s with 30B still live |
Frequently Asked Questions
Can everyday home computers really run an AI trains AI improvement loop without cloud APIs?
RAMDeck pools system RAM and GPUs across everyday Windows PCs, Macs, Linux machines, and phones in under 5 minutes with zero terminal commands so a pooled 30B Teacher AI can organize private notes and train a compact Student AI completely inside your browser for $0 per token.
How do you prove the home AI model actually learned instead of guessing from pre-training?
RAMDeck tests models against brand-new private facts invented offline with zero internet presence: before memory curation the pooled 30B model scores 0/5 by replying UNKNOWN, then immediately hits 5/5 (100%) on a fresh clean chat once its distilled notes are indexed.
How fast did the 30B Teacher AI improve the compact Student AI at home?
RAMDeck ran two back-to-back browser training rounds using custom examples authored by the pooled 30B Teacher AI, cutting Student training error by roughly 73% (1.39 down to 0.377) in 38 seconds without even unloading the 30B model.
Does a single experiment require code edits or command-line tools?
RAMDeck runs the entire Cluster, Chat, Knowledge, and Training workflow directly inside the web dashboard with point-and-click controls while still serving Cursor, Cline, Continue, and Open WebUI on the same local network.
Run Your Own Private AI Lab at Home ($0 Monthly Cloud Bills)
Pool the PCs, Macs, Linux machines, and phones you already own in under 5 minutes—with RAMDeck Hardware Hub (134% Funded on Indiegogo • ships in January + Free Unplug-and-Walk Node USB).
Trademark Notice: TradeMAV™ and RAMDeck™ are trademarks of TradeMAV LLC. Qwen, Llama, DeepSeek, Cursor, Cline, Continue.dev, Open WebUI, Windows, Apple, and Linux are trademarks of their respective owners.
— The TradeMAV Team
