AI Engineering Signal #166
Fable 5.2 tops SimpleBench at 94.9%, beating Gemini 4 Argon and Carbon
Signals
Fable 5.2 tops SimpleBench at 94.9%, beating Gemini 4 Argon and Carbon
rerun model routing evals; Gemini 3.x-era baselines are now stale for production decisions.
Web
Microsoft Research backdoors agent via fewer than 10 neurons
neuron-level tampering is documented; audit all fine-tuned or third-party agent weights before deployment.
Web
Deleting weights post-training measurably improves LLM output
add weight pruning to your post-training pipeline before serving; low-cost gains available now.
ArXiv
OpenAI builds hard kill switch for research models
factor model-level shutoff requirements into research-tier API contracts before signing.
Web
New York AI safety law activates, covering OpenAI and Anthropic
review API usage agreements for liability pass-through clauses before your next renewal.
TechCrunch
DeepMind releases AlphaProof Code
formal proof generation now has a public artifact; relevant for pipelines touching code verification or mathematical reasoning.
Web
The Take
The attack surface for deployed agents just expanded from prompt injection to weight-level backdoors, while kill switches and compliance laws arrive in the same week. Audit infrastructure and model provenance tracking are now the critical path, not capability benchmarking.
Subscribe
Related Signals