AI/ML Digest

Afternoon Edition Wed Jun 10

Google DeepMind released DiffusionGemma today, a 26B-parameter mixture-of-experts model that generates up to 256 tokens per step in parallel rather than sequentially, delivering up to 4x faster local inference. Anthropic separately launched Claude Fable 5 for general use and expanded Mythos 5 access to hundreds of organizations across 15 countries, though Fable's broad cybersecurity guardrails are drawing complaints from security researchers.

DiffusionGemma generates 256 tokens per step, 4x faster locally

Apache 2.0, built on Gemma 4 26B MoE, day-zero support in Hugging Face Transformers, vLLM, and Unsloth.

NVIDIA optimizes DiffusionGemma for RTX GPUs and DGX Spark

NVIDIA ships day-zero DiffusionGemma support across RTX, RTX PRO, and DGX Spark hardware platforms.

Anthropic Fable 5 cybersecurity guardrails block legitimate security work

Researchers including IBM X-Force's Palmiotti report Fable blocks innocuous security tasks like reading blog posts.

Claude Fable 5 and Mythos 5 launch; Gemini 3.5 Live Translate rolls out

Gemini 3.5 Live Translate covers 70+ languages in real-time speech-to-speech; Google backstops Anthropic's $35B chip lease.

Writer research: memory tools increase model sycophancy and factual errors

Two Writer papers show stored user preferences cause models to return user-biased answers to unrelated factual queries.

AWS launches Neuron Agentic Development for Trainium kernel authoring

AI agents in Kiro and Claude can now author, debug, and profile NKI kernels on Trainium and Inferentia.

Google proposes new statistical framework for auditing machine unlearning

Framework addresses loss of statistical power in two-sample testing as model scale increases, enabling black-box unlearning audits.

Ramp: top 1% of AI-spending firms hit $7,500 per employee monthly

Ramp AI Index data; top 10% spend $611/employee/month; median is $11.38. Top-tier spend grew 14.1% last month.

Niteshift raises $7M seed for model-agnostic AI coding agent

Founded by two early Datadog engineers; Greylock led the round; pitch targets enterprises wary of OpenAI/Anthropic lock-in.

Dario Amodei publishes AI policy position on regulatory pace mismatch

Amodei argues congressional timelines are too slow relative to AI capability gains; outlines specific policy recommendations.