AI Builder Stack Hits Four Fault Lines

The most important signal is not one breakthrough but a shift in where engineering risk lives: weights, training data, serving infrastructure, and tool permissions now matter as much as prompting.

• –MIT’s results are promising but preliminary; independent analysis questions some weight-count claims and stresses that several experiments used tools or model feedback.
• –Revolut’s $0.30/M figure is a blended internal cost, enabled by 56 B200 GPUs, tuning, caching, utilization, and routing—not a generally available GLM-5.3 price.
• –The backdoor paper is the most operationally serious finding: 0.1% poisoned pretraining data produced roughly 92.7% triggered attack success in a 4B CLI agent after clean post-training, though frontier-scale confirmation remains open.
• –Opus 5.5’s blueprint demonstrations show workflow acceleration, not automatically code-compliant or permit-approved architectural documents; professional review remains essential.
• –Builders should treat model provenance, sandboxing, least privilege, and inference economics as first-class product features.