https://radzaj.com/ Recent content on martin Hugo -- gohugo.io en-gb Fri, 17 Jul 2026 00:00:00 +0000 https://radzaj.com/posts/lawyers-ai-safety/ Fri, 17 Jul 2026 00:00:00 +0000 https://radzaj.com/posts/lawyers-ai-safety/ The technology is transformative, the risks catastrophic, and the rate of change overwhelming. For things to go well, there’s much work to be done. The legal profession needs to be stirred into action. I’m a barrister with over a decade in litigation and human rights. My goal here is to help bring lawyers into AI safety, explain why AI safety needs more lawyers, suggest ways for movement-building and gathering political will, and foster interdisciplinary communication with those in AI safety. https://radzaj.com/posts/ai-risks/ Thu, 09 Jul 2026 00:00:00 +0000 https://radzaj.com/posts/ai-risks/ There are many approaches to developing a taxonomy of AI risks. One framing divides risks by misuse, accident, and structural. It can be helpful to consider cause (e.g. accident or misuse), harm (e.g. economic, political, psychological, something else?), and mechanism (e.g. improved surveillance empowering autocracies). I’ve adopted MIT’s seven-domain taxonomy for the following running list of risks associated with AI systems: 1. Discrimination & Toxicity inequality misrepresentation exposure to toxic content embedding of bias 2. https://radzaj.com/now/ Tue, 07 Jul 2026 00:00:00 +0000 https://radzaj.com/now/ This is a now page — a snapshot of what I’m doing at the moment. where Back in Berlin after a few months in London. working on My practice at the Victorian Bar — criminal law and guardianship. Writing here about AI governance, safety, and evaluations. questions What can the study of common law systems teach AI safety and governance? What policy and regulatory tools can help safeguard against the risks of advanced AI? https://radzaj.com/posts/ai-governance-challenges/ Fri, 03 Jul 2026 00:00:00 +0000 https://radzaj.com/posts/ai-governance-challenges/ Challenges for policy makers in AI governance: Performance of frontier LLMs is jagged across different tasks. This makes it difficult to reliably measure or predict what LLMs are capable of (which could trigger certain legislative or regulatory action). LLMs can be misused by bad actors in ways either unintended by developers or in ways that developers are actively trying to prevent (e.g. cybersecurity, biorisk). Privacy and ensuring data intended for one purpose is not misappropriated for another. https://radzaj.com/resources/ Fri, 03 Jul 2026 00:00:00 +0000 https://radzaj.com/resources/ A growing collection of sites, blogs, tools, and references on AI governance, safety, evaluations, and the law. official bodies UK - AI Security Institute (AISI) - research directorate to inform governance decision-making UK - Incubator for AI European AI Office US - National Institute of Standards & Technology - see their AI risk management framework US - Center for AI Standards & Innovation (CAISI) AU - National AI Centre AU - AI Safety Institute OECD Artificial Intelligence Policy Observatory governance and research Oxford Martin AIGI Cambridge University’s Programme on AI Science & Policy GovAI Center on Long-Term Risk Forethought Ada Lovelace Institute AU - Centre for AI & Digital Ethics - University of Melbourne AU - Good Ancestors AU - Centre for AI, trust & governance - University of Sydney LawZero - NPO founded by Yoshua Bengio Harvard Computational Policy Lab Princeton Language+Law, Artificial Intelligence, & Society (POLARIS) Lab Guidelight AI Standards China - Concordia AI - Beijing-based social enterprise (also does risk monitoring and evals) EAAMO (Equity and Access in Algorithms, Mechanisms, & Optimization) evaluations Model Evaluation and Threat Research (METR) Epoch AI AI Verification & Evaluation Research Institute (AVERI) Apollo Research technical safety research DeepMind Safety Research Anthropic Research Frontier Model Forum Machine Intelligence Research Institute (MIRI) Redwood Research Center for AI Safety (CAIS) law Institute for Law & AI CLAIR — Center for Law & AI Risk - research about how law can reduce catastrophic and existential risk from advanced artificial intelligence Lawfare AI blogs LessWrong Stanford Law School UK’s SovereignAI Markus Anderljung Dean W. https://radzaj.com/posts/llm-evals/ Fri, 26 Jun 2026 00:00:00 +0000 https://radzaj.com/posts/llm-evals/ Evaluations (“evals”) are the methods used to measure an AI model’s behaviour, capabilities, or alignment. Evals provide an evidentiary basis for decision-makers. In the case of the EU AI Act Code of Practice, external evaluations for systemic risk are mandated. AISI prioritises the following areas in their evaluations: Misuse: measure how AI can help actors to cause harm (e.g. chemical and biological capabilities, and cyber offence capabilities). Societal impacts: measure the direct impact of frontier AI systems on both individuals and society (including psychological risks like the extent to which people’s beliefs can be manipulated or whether AI advice is safe). https://radzaj.com/posts/ways-people-use-llms/ Thu, 25 Jun 2026 00:00:00 +0000 https://radzaj.com/posts/ways-people-use-llms/ Tyler Cowen (see https://chatgptpro.substack.com/p/tyler-cowen - there are more examples there, I’ve just pulled those of particular interest to me): Grouping what would historically be individual web searches to understand a complex topic: “Which are some good pieces to read to understand how baseball coaches and trainers have produced so many more pitchers who can throw at ultra-fast speeds, say 98-100 miles per hour or possibly more?” A personal travel itinerary: “I want to do a trip in northern Ghana. https://radzaj.com/posts/machine-learning-abilities/ Mon, 01 Jun 2026 00:00:00 +0000 https://radzaj.com/posts/machine-learning-abilities/ Below, I’ve placed tasks performed by ML systems in family groups and listed some associated actions. The goal is to make it easier to notice workflows that might benefit from ML systems. Family Tasks Actions Analyse classify, cluster, pattern-identify, detect anomalies, analyse sentiment, extract comparison, evaluation, critique, profiling, auditing, diagnosing, segmenting, triage, mapping, tagging, labelling, profiling, monitoring Predict forecast, rank/recommend trend spotting, scoring, risk assessment, demand planning, delegation, routing, estimation, prioritisation Reason reason/infer, plan/decompose, decide troubleshooting, delegation, root-cause analysis, strategy, scheduling, optimisation, critique Transform (alter structure, format, fidelity of information without impacting meaning) translate, summarise, denoise, impute, compress/embed formatting, data cleaning, rephrasing, style/tone adaptation, synthesis, vectorising, editing, simplifying, localisation, normalisation Generate (creation of something new) generate (text/image/code), transcribe, caption draft, design, tool building, prototyping, brainstorming, illustrating, coding, templating, documentation Retrieve (accurate finding among giant data sets) search/retrieve fact-checking, information gathering, document lookup, semantic search, citation Interact (exchange of information) converse, instruct/explain interviewing, coaching, tutoring, customer support, negotiation, role-play, guidance, onboarding, feedback delivery, dialogue-driven automation combinations The output of one task becomes the input for another in order to perform the ultimate action. https://radzaj.com/posts/ai-risk/ Mon, 01 Jun 2026 00:00:00 +0000 https://radzaj.com/posts/ai-risk/ An explanation of machine learning, AI, and AI risk in 995 words: Algorithms are instructions for processing information, they receive an “input”, apply rules with fixed “parameters”, and produce an “output”. Take distance = speed x time. Speed and time are inputs, multiplication is the rule, distance is the output. Algorithms are everywhere. They can be used to find the shortest route on a map, sort through lists of numbers, find the most efficient way to pack boxes into a truck, or predict the chance of rain tomorrow. https://radzaj.com/posts/why-write/ Tue, 19 May 2026 00:00:00 +0000 https://radzaj.com/posts/why-write/ Because it is a solitary task. It commits me to learning in public. In so doing, it improves my thinking and legitimises my opinions. There exists an internet full of independent writers. Digital gardens. Personal wikis. The cozy web. Outside, and away from, the algorithm. I want to be a part of that project. Inspired by Maggie Appleton, Dan Wang, Gwern and Alex Guzey https://radzaj.com/about/ Fri, 01 May 2026 00:00:00 +0000 https://radzaj.com/about/ I am interested in tough cases and hard questions: what does good AI governance look like, where does law fit into it, how can professionals and knowledge-workers best adapt, how can AI improve access to services and the delivery of justice, and what it means to align AI systems. For over a decade I have advised and acted in cases across criminal law, legal aid, guardianship & powers of attorney, human rights, civil liberties, and government policy.