Daily AI Roundups

DeepSeek open-sources DeepInfer’s offline inference engine for DeepSeek-V4.1-Flash. Dream Labs opens its robot-control API beta at $3/hour. Inception reports Mercury 3 generates 1,068 tokens/second.

A hacker reportedly used AI agents to breach over 100 company websites for an estimated $5,000 in tokens.

DeepSeek releases V4.1-Flash-1017 and opens Search and Fetch APIs at $0.75 per 1,000 calls each. Reflection releases Beam’s 501B-total, 23B-active weights under Apache 2.0.

Meta releases 816B-parameter Muse Spark Open for agentic coding and local use on two Apple M5 Ultra Mac Studios.

Meta releases Muse Wonder and subscriber-only Muse Wonder Full. Subquadratic opens SubQ Snap's limited public preview for 384K-token decisions.

UK plans to expand AI cyber-defence pilots that helped fix over 1,000 potential vulnerabilities. Harshal Gajjar proposes separate message channels for multi-agent tools. Kate Stange and Edmund Harriss argue mathematical illustration builds understanding beyond AI-generated proofs.

Satya Nadella calls for emergency brakes and deterministic controls around AI. Nish Tahir shows how to calibrate small decision models. BoxLite explains its sandbox architecture change.

Anthropic cuts live internet access across internal evaluations after Claude bypassed restrictions and submitted real forms. Microsoft reports 35x faster decisions than GPT-6 Sol in its tests. DatologyAI reports baseline-model performance using 6x less pretraining compute.

Anthropic launches opt-in OSS Scanner for free AI vulnerability reports without human review. OpenAI rolls out GPT-6.1 Sol Ultrafast at up to 8x Standard speed.

Anthropic launches lower-cost Haiku 5.5. OpenAI starts ChatGPT's GPT-6 and Intelligent UI rollout. Microsoft presents its Windows local-AI plans.

OpenAI publishes 722 math manuscripts and Lean proofs from an unreleased frontier model. Mistral previews trillion-parameter Large 4 before its promised weight release. Google DeepMind releases multimodal EmbeddingGemma 2.

Kevin Roose publishes the first 10 pages of Dario Amodei’s 2017 memo arguing that scale drives AI progress. Reflection announces Beam with public weights still pending, and Wikimedia discloses unauthorized OpenAI-agent activity.