The Scientist

Daily Dose of Data Science detailed how vLLM keys each 16-token KV block on a hash of the parent block plus the token IDs inside it, so two RAG requests retrieving the same chunks in a different order share zero blocks. TLDR Data reports block-paged allocation reaching roughly 96% cache utilization and 2–4x more concurrent users per GPU with bit-identical attention outputs. For you, cache warmth now differs per A/B arm, so per-arm latency and cost stay confounded until you log cached versus cold input tokens.

CSO Update reports InjecMEM durably altering an agent's future responses after the session that planted the instruction ended, plus a DNS-rebinding flaw in NemoClaw that gives any visited browser tab unauthenticated access to a local Ollama server. The Hacker News reports GPUThor defeating ECC on NVIDIA RTX A6000 GDDR6 memory and escalating to host root. Both failures sit past your measurement boundary: an eval harness that flushes context between trials, and a checkpoint hash checked only at load time.

The Pragmatic Engineer assembled four published LLM migrations: Bun ported 530,000 lines from Zig to Rust for $165K in API spend, roughly $0.31 per line, against Sentry's 2021 manual JS-to-TypeScript conversion at about $21–42 per line. Airbnb published the only usable throughput curve — 75% of 3,500 files in four hours, 97% over four days, the final 3% a human week. The consequence for your backlog is that items parked as uneconomic under labor pricing now turn on whether you can write an accept function stronger than 'the suite is green'.

Morning Brew details the product spec inside Meta's up-to-$18B state settlement: a two-hour daily cap for users aged 13–17, a midnight-to-6am lockout, no push notifications between 8am and 3pm, and likes hidden by default for minors. Read as engineering requirements, those terms right-censor time-on-app at 120 minutes for one cohort and set some notification propensities to exactly zero. Your IPS and doubly-robust estimators will clip those infinite weights silently and return a plausible, invalid number.

Paul Smalera traces roughly $27B of NVIDIA capability purchases in nine months without a single acquisition: about $20B to Groq in December 2025 for a non-exclusive inference license plus substantially all assets, and about $7B to Poolside last week, with offers to the 109 engineers who built its Laguna model. GroqCloud now serves traffic on its own processors alongside NVIDIA hardware. A pre-2026 TTFT benchmark therefore measured a system that no longer exists, and heterogeneous backends behind one endpoint produce bimodal latency that means hide.