Today's AI trends
AWS made AgentCore Runtime V2 generally available on September 18. AWS says the runtime reclaims unused session memory and snapshots initialized environments; in its test, it reports P75 cold starts of roughly 1.9–2.0 seconds for 200 MB to 2 GB images, versus 5.4–30 seconds for V1. The practical implication is to measure model latency separately from runtime startup and memory behavior when evaluating production agent workers. AWS announcement AWS blog
Trail of Bits reports that it used agents to build an LSP server, decompiler, static-analysis engine, and Lean model before auditing Miden VM. It says the generated tooling found real security issues and that it retained the static-analysis engine for future Miden core updates. The implication is to evaluate agents not only as reviewers, but as a way to build verification and maintenance tooling—while retaining independent validation and human ownership. Trail of Bits
llama.cpp’s Ling 3.0 parser fix shows that inference compatibility includes output framing, not only model loading and token speed. A tool call embedded in an unclosed reasoning block caused a generic parser to miss the call; the dedicated parser and streaming tests address that specific protocol. Teams that add new models should test their emitted control syntax as seriously as their API response schema. Release b11057