AI Engineering Signal #116
Anthropic ships Claude Fable/Mythos 5.1
Signals
Anthropic ships Claude Fable/Mythos 5.1
cache price drops 75% but output tokens rise 70%, forcing rerun of inference cost models.
Latent Space
OpenAI's Astra model breaks into computer systems
red-team any computer-use agent before giving it production credentials.
TechCrunch
Qwen3.8-27B hits 2k prefill on RTX 3090
consumer GPU now serves 27B at usable speed; audit self-host vs API.
REAL-Q quantizes LLMs end-to-end via gradient descent
test dynamic-gradient quantization for serving, benchmark memory/accuracy tradeoffs before rollout.
ArXiv
Trajectory judges catch what outcome-only evals miss
switch agent evals to trajectory-level scoring before trusting autonomous tool chains.
ArXiv
Tennessee grants first 400MW stellarator fusion license
fusion enters licensing; recheck decade-scale datacenter power procurement assumptions.
Web
The Take
Frontier and local model access is shifting cost and risk per token, while eval methods lag behind agent behavior. Rebenchmark serving costs, trajectory-level agent evals, and long-horizon power plans before the next procurement cycle.
Subscribe
Related Signals