Today’s top line: Nvidia’s CEO admitted an affair under US espionage scrutiny while AI firms secure power and cut agent costs – Anthropic bought $3.4B in turbines and Meta released a low‑cost Muse Ember model for agent execution.
Nvidia CEO admits affair amid US intelligence espionage allegations
Summary: Nvidia founder and CEO Jensen Huang acknowledged an extramarital affair that US intelligence has alleged involved a person recruited by China’s Ministry of State Security. Nvidia’s board opened an independent investigation and Huang said he will cooperate; the company’s shares fell more than 6% in early trading.
Digest: The disclosure raises national‑security and export‑control scrutiny for a CEO who has lobbied for broader chip exports to China; expect congressional and regulatory attention and scrutiny of Nvidia’s China business.
Source: Ars Technica
Anthropic buys $3.4 billion of natural‑gas turbines for power
Summary: Anthropic is spending $3.4 billion to buy Caterpillar natural‑gas generators capable of producing up to 1.7 gigawatts of electricity. The purchase follows a pattern of compute firms securing power and comes as Anthropic disclosed in its S‑1 that it pays SpaceX subsidiary xAI $1.25 billion per month through May 2029 for compute services.
Digest: AI companies are increasingly owning energy infrastructure to guarantee supply, exposing them to construction backlogs, grid‑connection delays and fuel‑price risk – a shift firms and customers should watch for cost and sustainability implications.
Source: TechCrunch
Meta releases Muse Ember, a low‑cost model for agent execution
Summary: Meta’s Superintelligence Lab released Muse Ember, its lightest Muse model, positioned for agent execution tasks and trained to handle text, images and audio. Meta says Ember costs 10 cents per million input tokens and is cheaper on composite benchmarks than some rivals; it is intended to pair with larger Muse models where the larger model plans and Ember handles high‑volume execution.
Digest: Lower per‑task costs and multi‑modal capability make Ember attractive for high‑volume agent workflows (customer support, routing, extraction); businesses running agent fleets could cut operating costs by using a planner + lightweight executor architecture.
Source: TechCrunch
Claude Code settings files leaked secrets to public GitHub repos
Summary: Researchers at Apiiro found over 100,000 public GitHub repositories that included Claude Code’s local settings file “.claude/settings.local.json”; about 1,100 files from 63 organisations contained secrets such as API tokens, database passwords or private keys. The file can record approved shell commands, and although Claude Code adds it to a global .gitignore, repo‑level ignores are not always set; earlier bug reports dating to August 2025 were closed without action.
Digest: AI developer tooling can persist sensitive approvals into source control; organisations should run secret scanners, audit repos and demand safer defaults from vendor tools (Anthropic is “looking into” fixes and recommends secret scanning and an auto mode).
Source: Ars Technica
WorldParticle: one model that simulates many materials
Summary: Researchers from MIT, UBC and partners introduced WorldParticle, a deep‑learning system that represents materials as particles and predicts their motion step‑by‑step to simulate phenomena from sliding sand to protein folding. The team says the model learns underlying dynamics rather than being hand‑coded to specific equations; the work will be presented at ACM SIGGRAPH Asia and is published open access.
Digest: A single, general simulator could speed engineering and research workflows that today require many specialised solvers; industrial adoption will depend on accuracy and scaling to real‑world problems.
Source: MIT News
Nemotron‑Personas‑Finance‑Japan: a million‑record synthetic finance dataset
Summary: Elith and NVIDIA released Nemotron‑Personas‑Finance‑Japan, an open persona dataset at million‑record scale pairing Japan’s official statistics with household finance and financial literacy data, built with NeMo SDG‑PGMs and NeMo Data Designer. It is released under a CC BY 4.0 licence for commercial use and was developed with input from Japanese banks and insurers.
Digest: The dataset gives Japanese financial teams a privacy‑safe way to train and test customer simulations and conversational agents without using real customer data; expect uptake in local fintech and insurer model development.
Source: Hugging Face
Quick hits
- SpaceX says it is now an “infrastructure company” – A SpaceX VP described rapid AI adoption across engineering workflows and automation of visual inspections, claiming AI tools improved defect detection from about 94% (humans) to 99.9% (Grok) on some tasks (Ars Technica).
- Medium launches TK editor with built‑in AI assistant – Medium’s new TK app embeds an AI assistant in the editor to help drafting, style and provenance notes and lets authors disclose how AI was used (TechCrunch).
- HOPE pruning improves MoE expert removal – A new higher‑order pruning method (HOPE) accounts for pairwise expert interactions in mixture‑of‑experts models and shows up to +6.1% gains on agentic coding tasks at 50% pruning (Hugging Face).
- MIT’s FLEX helps robots infer intent from natural language – FLEX combines a large language model with a robot motion‑planning pipeline so robots generalise beyond literal commands (MIT News).