The 67-Cent Proof
ou are constantly told that capable artificial intelligence demands sprawling server warehouses, billion-dollar venture checks, and endless subscription fees. Mithil Vakde, an IIT Bombay graduate based in Bangalore, just dismantled that corporate orthodoxy by training a transformer completely from scratch in 1.5 hours on a single consumer RTX 5090. Total electricity bill: 67 cents.
His build tackled ARC-AGI-1, a rigorous 1,000-puzzle benchmark crafted to measure raw, visual-logic reasoning rather than regurgitated web data.
"I think sample efficiency is the most important problem in AI today and I want to solve it."
As Vakde pointed out, the entire point was pushing the limits of sample efficiency within standard transformer setups, compressing training budgets so severely that solo builders can experiment right from their desks.
Beating Heavy Models on Hardware
The scoreboard speaks for itself. Vakde's lean 8-layer model secured 44% on the ARC-1 public evaluation and reached 7% on ARC-2. That puts this lightweight local setup toe-to-toe with dedicated reasoning systems like TRM and HRM, while cleanly outpacing far heavier corporate LLMs on core visual puzzles. The feat quickly caught fire among machine learning veterans, drawing praise and breakdown commentary on X from researchers including Lucas Beyer, Fast.ai founder Jeremy Howard, and Google's Rohan Anil.
The engineering trick was eliminating compute waste rather than brute-forcing parameters. Vakde discarded standard AdamW in favor of the Normuon optimizer, wired in modern architectural blocks like SwiGLU and RMSNorm, and implemented 3D RoPE embeddings to represent 2D puzzle grids without destroying spatial context.
What This Means for Local PC Hardware
For anyone running a high-end desktop, this experiment shatters the narrative that viable reasoning models remain locked behind enterprise cloud paywalls. When an independent developer can train a specialized architecture from zero in 90 minutes for pocket change, the entire economics of AI tilts back toward consumer rigs. With the code published open-source on GitHub, the runway is clear for hyper-focused, local fine-tuning and task-specific logic engines that bypass remote API meters entirely.