AI Engineering Signal #128
Anthropic says Claude is helping build the next version of itself
Signals
Anthropic says Claude is helping build the next version of itself
Expect AI-generated R&D patches as a dependency; gate each with evals and provenance.
Web
Self-generated prompt injection hits compaction summaries
Treat compaction output as attacker-controlled input when building agent memory and audit trails.
Simon Willison
Bonsai 2 27B runs in-browser under 6GB
Rerun edge deployment and WebGPU capacity plans for quantized 27B-class models.
Web
Uber publishes retry storm protection patterns
Add token-bucket and jitter to internal service clients before agent traffic spikes.
Web
Keep vLLM prefix caches warm across agent turns
Cut multi-step agent prefill cost by retaining KV cache between calls.
Web
Huawei AI chip demand outpaces supply
Qualify Huawei accelerators now if China-region latency or data-residency requirements apply.
Reuters
3D nanomembrane Legos enable photonic integration
Watch photonic interconnect prototypes entering advanced packaging roadmaps.
Web
The Take
Agent and coding systems are getting productionized through the unglamorous layer: cache warmth, retry discipline, compaction trust boundaries, model compression, and accelerator supply. Teams treating that plumbing as the product are the ones cutting cost and failure rates.
Subscribe
Related Signals