The AI That Tutored Another AI Overnight
NVIDIA showed off a neat little experiment: hand the coding agent Codex two prompts, and it post-trained a small model, Cosmos 3 Nano, from 54.41% to 93.35% accuracy in a single day. One AI, quietly tutoring another into competence. That loop, an AI making an AI better, is what people call recursive self-improvement, and post-training just means taking a finished model and sharpening it a bit more. Paxis and Metis watch the whole self-improvement marathon unfold beside a waterfall.
Source: RT @kimmonismus: NVIDIA says Codex post-trained Cosmos 3 Nano from 54.41% to 93.35% accuracy in one day - with two prompts · twitter
What this means for ThakiCloud
The catch with a self-improving loop isn’t the accuracy, it’s the meter. Every round of getting smarter is another stretch of GPUs running flat out, and on a rented cloud that stretch shows up as a line item. Run the loop on someone else’s hardware and the model improves while the invoice does too. ThakiCloud’s Metis keeps training and inference inside your own walls (on-prem), so both the smarter model and the compute that made it stay under your control. Let Paxis’s agents refine the model and Metis absorb the training, on your racks, not a landlord’s counter.
An auto-generated comic riffing on this week’s industry news.