AI/ML Digest

Afternoon Edition Fri Jun 12

Anthropic reversed a policy that silently routed certain Claude Fable 5 requests to a lesser model after researcher backlash; separately, OpenAI acquired Ona to support persistent agent execution in Codex. Mistral is in early talks to raise €3B at a €20B valuation, nearly double its September 2025 Series C.

newsAnthropic reverses silent model-downgrade policy after researcher backlash

Claude Fable 5 was secretly degrading responses for tasks like training competing models; Anthropic made safeguards visible after complaints. — TLDR AI

newsMistral in talks to raise €3B at €20B valuation

Round would nearly double Mistral's €11.7B September 2025 Series C valuation; Bloomberg cited anonymous sources. — TechCrunch AI

newsSpaceX IPO opens at $150, rises 30% in midday trading

SpaceX priced 555.6M shares at $135 each, raising $75B in the largest IPO in history. — TechCrunch AI

toolAllenAI releases olmo-eval evaluation workbench for LLM dev loop

Open-source harness designed for continuous evaluation across model checkpoints, not just finished models. — Hugging Face

toolLMCache: reusable KV cache layer for LLM inference on vLLM

Persists and shares KV cache across serving engines to cut TTFT on long-context, RAG, and agentic workloads. — GitHub Trending

techniqueBuild meeting prep agent with Amazon Quick and Cisco Webex MCP servers

Step-by-step guide wiring two MCP servers into a single agent for pre/post meeting summarization and action items. — AWS Machine Learning Blog

techniqueIntelligent document pipeline with Amazon Bedrock Data Automation

BDA handles classification, extraction, and validation for up to 3,000-page documents via a unified API. — AWS Machine Learning Blog

toolgeiger: read-only blast-radius triage tool for leaked credentials

Pipe credential-bearing text in; get ranked read-only recon of what each key actually reaches. Go, open source. — HN Show HN

toolzerodecimal: zero-allocation fixed-point decimal library for Go

35% faster than udecimal, 90% faster than shopspring/decimal; enforced by AllocsPerRun tests and 23 fuzz targets. — HN Show HN

techniqueGoogle runs retired smartphones as Kubernetes-managed datacenter nodes

Stripped-down phone motherboards run Linux containers; 25–50 phones match one modern server on SPEC benchmarks. — Google Research