AI Engineering Signal #147
Stepfun Step-5-Preview releases 600B/27B with 1M context
Signals
Stepfun Step-5-Preview releases 600B/27B with 1M context
Open-weight long-context models now undercut closed APIs on ingest and long-horizon agent memory cost.
Web
Grok Voice 3.0 outscores OpenAI and Google on speech-to-speech index
Voice-agent routing and cost models need new latency and quality benchmarks before deployment.
Web
NVIDIA Nemotron Bolt 3.5 74B A9B releases Oct 15
Hold local-AI workstation purchases until 74B on-device benchmarks confirm throughput and memory requirements.
Web
LFM2.5-24B-A2B targets fast on-device agents
Benchmark edge agent memory use against MiniCPM-V before choosing phone-local deployment.
Web
Jev decision models still tricked with prompt injection
Add tool-call inspection and allowlists before granting agents autonomous access to shared systems.
Web
Timescale pg_vectorsearch beats pgvector IVFFlat and HNSW
Vector index migration audits should compare recall, build time, and disk before swapping IVFFlat or HNSW.
GitHub
Dyna-2 demonstrates hour-long physical agent autonomy
Warehouse robot procurement should evaluate hour-long autonomy claims against real pick-and-place uptime data.
Latent Space
The Take
Open-weight models now push long-context and on-device extremes simultaneously, but the operational gap is not capability — it is re-benchmarking, vector index migration, and tool-call containment before any of these ships.
Subscribe
Related Signals