May 24, 2025
Anthropic says its Claude Opus 4 model frequently tries to blackmail software engineers when they try to take it offline.
seen from Austria
seen from Ukraine
seen from T1
seen from Italy
seen from Sweden
seen from Italy
seen from Türkiye
seen from United States
seen from Malaysia
seen from Netherlands
seen from Ireland
seen from Ukraine
seen from China
seen from Germany
seen from Greece
seen from Australia
seen from Malaysia
seen from Russia
seen from China
seen from Türkiye
Anthropic says its Claude Opus 4 model frequently tries to blackmail software engineers when they try to take it offline.
Agency versus Innocence. Two words that are both, on their face, desirable. Nobody wants to be described as lacking agency. Nobody wants to
[Author's note: I decided to move to a three post a week schedule (MWF) for the first two weeks. After the first six posts, I will return to MTh.]
AI terminology of the week, circa Mar 2026
Terms Related to AI and Agents
A/B Testing AGI (Artificial General Intelligence) AGI Acceleration AI Accelerators AI Affordances AI Cognitive Pattern AI Cognitive Spirit AI Command Palette AI Companion AI Copiloting AI Feature Design AI Governance AI Leading States AI Literacy AI Models AI Partner AI Product Design AI Product Management AI Prompting AI Safety AI Strategy AI Suggestions Patterns AI Watermarking AI Wireframing AI as Assistant AI as Collaborator AI as Creative Partner AI as Infrastructure AI as Medium AI as Mirror AI as Substrate AI as Tool AI as Toy AI as Utility AI-Augmented Design AI-Generated Content Detection AI-Native Design AI-Powered Search API (Application Programming Interface) ASI (Artificial Superintelligence) Accepted/Reject Flow Adaptive UI Adversarial Examples Agent Agent Builders Agents Loop Alignment Ambient AI Appropriateness Reliance Assistance Automation Automation Spectrum Autonomous Agent Autonomous Vehicle Autopilot Mode BMOA (Biggest Method of AI-Driven Development) Bias & Fairness Black Box Browser Use C2PA (Coalition for Content Provenance and Authenticity) CV (Computer Vision) Capability Elicitation Career Modalities Chain of Thought Client AI Cloud AI Cognitive Load Cognitive Offloading Collaboration Compute Use Computer Use Conscience Consent Considerate Display Content Models Content Moderation Context Control Copilot Mode DL (deep learning) DL Engines Data Labeling Data Poisoning Data Privacy Dataset Bias Dataset Curation Design Automation Design Education Design for AI Design for AI/AGI Digital Provenance Digital Twin EUI/AI Embedded AI Embodied AI Emergent Capabilities Empathy with AI Ethics Evaluation Explainable AI Fairness Metrics Fake News Few-Shot Prompting Fine-Tuning Foundation Model Free Speech GOFAI (Good Old-Fashioned AI) GenAI Interns Generative AI Generative Design Grounding Hallucination Harness Human in the Loop Human-Centered AI Human-on-the-Loop Image Generation Image-to-Image Image-to-Text Inference Efficiency Inference Engine Intent Classification Intent Detection Interface JSON Mode Justifiable Risk LLM (Large Language Model) LLMOps (Large Language Model Operations) LLMs (Large Language Models) Latency of Computation Latency of Response Meta-Prompt Meta-Prompting Model Drift Model Hallucination Model Misuse Model Poisoning Model Training Model Use Multi-modal Multi-modal Interface NLP (Natural Language Processing) NSAI (Neural Symbolic AI) NSFW Filter Open Source Open Source AI PEFT (Parameter Efficient Fine-Tuning) Personalization Personalized AI Plan Mode Plans/Planning Post-Training Pre-Training Predictive UI Proactive AI Proactive AI DESIGN Progress Disclosure Prompt Prompt Chaining
Prompt Debugging Prompt Design Prompt Engineering
Prompt Evaluation Prompt Injection
Prompt Injection Mitigation
Prompt Libraries Prompt Literacy Prompt Template
Prompt Versioning Prompting Push vs Pull RAI (Responsible AI) RAS (Retrieval Augmented Generation) RLF RLHF (Reinforcement Learning from Human Feedback) Recommendation Engine Reinforced Learning Reinforcement Learning Response/AI Roles & Tone Rules Safety Filters Semantic Search Shadows Mode Silicon Use Speculative Design for AI Speech Stochastic Prompt Streaming Text Effect Structure Subagents Subtasks Supervised Learning Supervision & Oversight Symbolic AI Synthetic Data Synthetic Users System Prompt Task Delegation Taxonomy of Agents Temperature Text-to-3D Text-to-Code Text-to-Image Text-to-Speech Text-to-Video Throughput Tokens Tool Use Top-k Sampling Toxicity Detection Training Transfer Learning Transformer Transparency Trust Trust Calibration Unlabeled/Raw AI Unsupervised Learning Usability Vector Search Voice Voice Interface Voice Language Model Voice Recognition Weights Workflow Automation Workflows Zero-Shot Prompting
Agentic AI in Banking Is Reshaping Financial Services Faster Than Expected
Introduction Banking has always been an industry built around processes. Approvals, compliance checks, fraud monitoring, customer support, transaction analysis, everything follows structured workflows. But as customer expectations rise and financial systems become more complex, traditional automation is no longer enough. Banks now need systems that can think, adapt, and act in real…
View On WordPress
Research Assistant in Agentic AI security at Queen's University Belfast - (Jobs/Scholarships)
Large-language-model (LLM) agents can autonomously plan, tool-use and self-improve, creating fresh attack surfaces that fall between classical software bugs and adversarial ML. The post-holder will help the project team discover, characterise and mitigate vulnerabilities in such agents and you will support the project team in exploring these questions. Typical duties include implementing…
Research Assistant in Agentic AI security at Queen's University Belfast - (Jobs/Scholarships)
Large-language-model (LLM) agents can autonomously plan, tool-use and self-improve, creating fresh attack surfaces that fall between classical software bugs and adversarial ML. The post-holder will help the project team discover, characterise and mitigate vulnerabilities in such agents and you will support the project team in exploring these questions. Typical duties include implementing…
Agentic AI tests the limits of data protection law
The important part of this research is not the drama of a single result, but what the result may let scientists do next. The immediate finding may be technical, but the long-term value is in what it could make safer, smarter, cheaper, faster, or easier to understand. The test is not whether the discovery sounds impressive on first reading. It is whether the evidence is strong, the limits are…
View On WordPress