
How to Decide: Loop Transformer Blocks or Add More Layers
Loop transformer blocks or add more layers? Reusing weights cuts training compute 6.8-18% at equal loss, but forward passes and KV cache never shrink.
Author
Software Engineer · AI Builder · Writer · Expedia Group
Software engineer writing about AI, Claude Code, LLMs, OpenAI, Anthropic, and developer tooling. 133 articles on AI engineering, production systems, and the tools shaping modern development. 5+ years at Expedia Group, Tekion, and BYJU'S.
Google Search · Preferred sources
If you already read this writing, add umesh-malik.com as a Preferred Source. Google can then highlight it with a preferred badge in Top Stories, AI Overviews, and AI Mode — for you, not as a site-wide ranking boost.

Loop transformer blocks or add more layers? Reusing weights cuts training compute 6.8-18% at equal loss, but forward passes and KV cache never shrink.

How to stop a yo-yo DDoS attack: Read the Docs held 5.5M requests/minute for 10 days using JA4 fingerprinting and edge caching, not IP bans.

Run Kimi K3 locally on a MacBook by streaming 1.45TB of experts from four SSDs — the 1 tok/s number, and why doubling drives doesn't double speed.

Post-quantum TLS migration checklist: X25519MLKEM768 cut Cloudflare's handshake retries from 52% to 3.7%. How to check, enable, and verify on your own origin.

How to stop AI scrapers overloading your server: git.kernel.org now loses 14-16 of its 90 CPU cores to bots, and proof-of-work only bought a few months.

Javascript proxy reactive state in under a kilobyte: an 80-line get/set trap that tracks reads, batches writes, and quietly breaks on one property name.