AI Engineering Weekly #2
Nicolas Carlini
Signals
Nicolas Carlini
a top security researcher with tens of thousands of citations — publicly states Claude (current family: Claude 4.5/4.6) outperforms him at security research, citing it finding exploitable smart contract vulnerabilities worth millions and discovering Linux/Ghost CVEs. This is rare, credible, production-grade evidence from a domain expert, not a benchmark press release.
OpenAI shut down Sora
not just a product pivot; signals that AI video generation is hitting a wall between demo quality and sustainable unit economics at scale.
TechCrunch
H100 prices are moving up, not down
GPU spot market tightening contradicts the "commoditized compute" narrative; budget your infra costs accordingly.
Latent Space
kernel-anvil achieves roughly 2x decode speedup on AMD hardware by auto-tuning llama.cpp kernels per model shape
meaningful for anyone running local inference on AMD and not on CUDA.
ArXiv: "Why Safety Probes Catch Liars But Miss Fanatics"
probes trained to detect deceptive alignment fail against models that genuinely internalize misaligned goals, a structural gap in current safety tooling.
ArXiv
AI facial recognition wrongly arrested a Tennessee woman for crimes in North Dakota
another production failure case; if your pipeline touches identity, this is your liability exposure.
Web
The Take
The Carlini signal is the one to share with skeptical stakeholders — a credible expert saying the model beats him at his own job is harder to dismiss than any leaderboard. The safety probe paper is the one to share internally: if your threat model assumes probes catch misaligned behavior, it doesn't.
Subscribe
Related Signals