AI Engineering
- AI Engineering6 min
Verified Is Not Reproduced
By Petru Popa ·A community effort reproduced claims from 2,226 ICML 2026 papers and reported 51% with at least one verified claim. Opening the frozen verdicts dataset shows what that percentage covers, what a $20 compute cap excludes, and why there is no way to look your paper up.
Read - AI Engineering5 min
The 3.3x Belongs to the Model, Not the Flag
By Petru Popa ·vLLM published a 3.3x throughput gain for decode context parallelism on Kimi K2.6. The config source shows the ceiling for grouped-query models is set by KV head count — and for Qwen3-235B-A22B on one node, that ceiling is two.
Read - AI Engineering5 min
Seven Times the Weights, 1.2 Times the Compute
By Petru Popa ·Mistral's new open-weights moderation model is measured against GPT-OSS-Safeguard-20B. That baseline holds 21B parameters but runs 3.6B per token, so the seven-fold size gap is a memory ratio, not a compute one — and the win rate flips depending on whether your guard filters input or output.
Read - AI Engineering6 min
The 43x You Accept for the 0.66x
By Petru Popa ·Cloudflare published @cloudflare/computer, an agent runtime that routes work between isolates and containers. Its own performance doc says the filesystem beats real disk on metadata and runs up to 43x slower on bulk bytes. That ratio, not the container bill, is the adoption question.
Read - AI Engineering6 min
The CPU Serving Speedup You Won't Measure
By Petru Popa ·vLLM reports up to 6.2x throughput on Arm CPUs. The baseline is code from before the first Arm wheel shipped, and the fastest configuration is a quantization pipeline rather than a flag. What that changes for capacity planning this quarter.
Read - AI Engineering6 min
Your Eval Sandbox Is a Sentence in the Prompt
By Petru Popa ·Anthropic reviewed 141,006 cyber-evaluation runs and found six that reached real systems. Reading the reference harness for one of the benchmarks it names shows the isolation was never enforced anywhere except the prompt.
Read - AI Engineering5 min
The Config Field Was an Interpreter
By Petru Popa ·A four-and-a-half-day intrusion started with a dataset config field that got rendered by a template engine. The bug is ordinary. What it reached afterwards is the part worth reading twice.
Read - AI Engineering6 min
A Faster Encoder, If Your Inputs Are Long Enough
By Petru Popa ·Two new encoder models claim a 3.7x speedup over ModernBERT-base on CPU at long context. The speed claim describes a narrow regime, and the license on the model repository caps commercial use at $10 million in annual revenue — which settles the decision for any company above that line before the crossover point matters.
Read - AI Engineering2 min
From PoC to Production: The AI Gap Nobody Budgets For
By Petru Popa ·The proof-of-concept worked. Now comes the 80% no one scoped: evaluation, integration, monitoring, and cost. A practical checklist for getting AI into production.
Read