SCBX
Weekly Tech
EU AI Act transparency rules take effect, fines reach 3% of turnover
New transparency obligations under the EU AI Act took effect on August 2, 2026, requiring deepfakes, AI-manipulated media, and undisclosed AI-authored public-interest text to carry clear, machine-readable labels. Anyone interacting with a chatbot, AI agent, or avatar must now be told they are not talking to a person. National market surveillance authorities, the European AI Office, and the European Data Protection Supervisor will enforce the rules, with penalties reaching €15 million or 3% of global annual turnover for non-compliant companies.
The Roundup— frontier moves this week
Wells Fargo to launch tokenized deposits for corporate clients this fall
Wells Fargo confirmed it will launch tokenized deposits for corporate and commercial clients this fall, letting them transfer, program, and settle bank deposits around the clock on its own blockchain platform. The initial service covers US dollars and British pounds for cross-border payments, with more currencies planned for 2027.
TakeawayTokenized deposits are moving from crypto-native pilots into mainstream corporate banking, pushing settlement toward always-on, programmable infrastructure.
PYMNTS · Aug 4→
Mistral releases Shieldstral, an open-weight AI safety classifier
Mistral released Shieldstral, a 3-billion-parameter open-weight model that judges text and images against moderation policies written as plain-language prompts, instead of a fixed, retrained category list. It covers 12 languages, runs on a single 16GB GPU, and reports safety scores matching or beating larger specialized classifiers. It ships under Apache 2.0.
TakeawayDefining moderation policy in natural language at inference time, not training time, could let small models match larger dedicated safety classifiers.
Heise Online · Aug 6→
Meta launches Muse Code, an AI coding agent
Meta launched Muse Code, a beta coding agent built on its new Muse Spark 1.2 model, aimed at long, complex coding tasks and positioned against Anthropic's Claude and OpenAI's Codex. It runs multiple sub-agents in parallel and logs its actions so it can resume after a crash. Pricing runs $1.25 per million input tokens and $4.25 per million output tokens.
TakeawayLong-horizon, multi-agent execution with crash-resilient state is becoming the baseline capability bar for coding agents across the major labs.
CNBC · Aug 5→
AI data-centre economics could cap the earnings boom
Investment strategist Joachim Klement argues in Reuters that AI data-centre economics may not support today's earnings expectations. A one-gigawatt data centre now costs an estimated $80-100 billion to build but is projected to generate only $10-12 billion a year, implying an 8-10 year payback before it turns a profit — a gap he warns could pull elevated corporate earnings back toward trend.
TakeawayUnless build costs fall or utilization revenue rises sharply, the earnings growth priced into AI infrastructure spending looks structurally hard to sustain.
Reuters (Joachim Klement) · Aug 5→
D-Wave and Nasdaq Verafin test quantum fraud detection
D-Wave and Nasdaq Verafin will evaluate quantum-hybrid annealing technology for financial-crime detection, starting with a proof-of-concept that could expand to pilot deployment. The work targets predictive models for fraud, scams, and money laundering by analyzing account activity and transaction patterns. The agreement is early-stage, with no production results yet.
TakeawayStill a proof-of-concept, not a deployment — an early signal quantum-hybrid computing is being explored for financial-crime detection.
The Quantum Insider · Aug 3→
Line OA
Connect with us
Connect with us
via LINE OA.
Stay updated with R&D. Scan the QR.
